ADVANCED CHARACTERIZATION OF

BATTERY CELL DYNAMICS

 ADVANCED CHARACTERIZATION OF
BATTERY CELL DYNAMICS

By MARVIN MESSING, M.A.Sc, B.Eng.

A Thesis Submitted to the School of Graduate Studies in Partial Fulfilment of the

Requirements for the Degree of Doctor of Philosophy

McMaster University

© Copyright by Marvin Messing, September 2021

DOCTOR OF PHILOSOPHY (2021), McMaster University, Hamilton, Ontario

(Mechanical Engineering)

TITLE: Advanced Characterization of Battery Cell Dynamics

AUTHOR: Marvin Messing, M.A.Sc, B.Eng.

SUPERVISOR: Professor Saeid R. Habibi

NUMBER OF PAGES: xx, 223

- ii -
Lay Abstract

Replacing conventional gasoline/diesel powered cars with battery powered vehicles is part of

a solution to the climate crisis. However, the initial costs paired with range anxiety stops many

from switching to electric cars. Both cost and range are related to the battery pack. To achieve

the best possible range for the lowest possible cost, battery packs must be carefully controlled

by sophisticated algorithms. Unfortunately, battery range or health cannot be measured

directly, but must be inferred through measurable indicators. This thesis explores battery

behavior under different operating conditions and develops improved methods which can be

used to determine battery health and/or range. A powerful method usually used only in

laboratory settings is studied and improved to make it more suitable for implementation in

electric cars. In this work it is used for accurate battery health determination. Furthermore, a

strategy for improving battery range determination at low temperatures is also proposed.

- iii -
Abstract
Battery Electric Vehicles (BEV) are gaining market share but still must overcome several

engineering challenges related to the lithium-ion battery packs powering them. The batteries

must be carefully managed to optimize safety and performance. The estimation of battery

states, which cannot be measured directly, is an important part of battery management and

remains an active area of research since small gains in estimation accuracy can help reduce

cost and increase BEV range.

This thesis presents several improvements to battery state estimation using different methods.

Electrochemical Impedance Spectroscopy (EIS) is receiving increased attention from

researchers as a method for state estimation and diagnostics for real-time applications. Due to

battery relaxation behaviour, long rest times are commonly used before performing the EIS

measurement. In this work, methods were developed to significantly shorten the required rest

times, and a State of Health (SoH) estimation strategy was proposed by taking advantage of

the relaxation effect as measured by EIS. This method was demonstrated to have an estimation

error of below 1%.

At low temperatures, the accuracy of the battery model becomes poor due to the non-linear

battery response to current. By using an adaptive filter called the Interacting Multiple Model

(IMM) filter, the next part of this work showed how to significantly improve low temperature

State of Charge (SoC) estimation. Further reduction in estimation errors was achieved by

pairing the IMM with the Smooth Variable Structure Filter (SVSF), for SoC estimation errors

below 2%.

The work presented in this thesis also includes the application of Deep Neural Networks (DNN)

for SoC estimation from EIS data. Finally, an extensive aging study was conducted and an

- iv -
accelerated protocol was compared to a realistic drive cycle based protocol using EIS as a

characterization tool.

Keywords:

Batteries, State Estimation, Electrochemical Impedance Spectroscopy, Adaptive Filtering,

Battery Aging, Deep Neural Networks

-v-
Acknowledgement

First, I would like to thank my supervisor at McMaster, Dr. Saeid Habibi as well as my

supervisor at Cadex, Dr. Tina Shoa who worked together to provide this opportunity for me

and supported me throughout. I would also like to thank my supervisory committee members

Dr. Gillian Goward and Dr. Ryan Ahmed for their feedback.

My thanks also to the Natural Sciences and Engineering Research Council of Canada

(NSERC), Ontario Graduate Scholarship (OGS) program, Queen Elizabeth II Graduate

Scholarship in Science and Technology awards program, and Cadex Electronics for financial

support.

Special thanks also to our industry partner Cadex Electronics for allowing me the use of lab

space and equipment as well as support my projects with technical guidance.

At the Center for Mechatronics and Hybrid Technologies (CMHT) I would like to thank Cam

Fisher, Jessica Petrunti, Zeina Tawakol and my fellow grad students with special thanks to Sara

Rahimifard for her friendship and support, and Ben Miethig for welcoming me with open arms

on my first day and his continued friendship. I would also like to thank Dr. Phil Kollmeyer for

sharing his battery testing knowledge.

I am forever grateful for the support and encouragement from my family and friends, with

special thanks to my partner Erin for her love and support, and for enduring my absence for

many months while I was splitting my time between Vancouver and Hamilton.

- vi -
Co-Authorship

This thesis has been prepared in accordance with the regulations for a sandwich thesis format or as
a compilation of research papers stipulated by the faculty of graduate studies at McMaster
University. This thesis consists of the following papers:

Paper I

M. Messing, T. Shoa and S. Habibi, “Lithium-Ion Battery Relaxation Effects,” Published in

2019 IEEE Transportation Electrification Conference and Expo (ITEC), 2019, pp. 1-6, doi:

10.1109/ITEC.2019.87904491.

Marvin Messing conducted the literature review, designed experiments, designed and

built test equipment, conducted experiments, analyzed results, and wrote paper I.

Dr. Shoa and Dr. Habibi provided research guidance and edited the paper.

Paper II

M. Messing, T. Shoa and S. Habibi, “Electrochemical Impedance Spectroscopy With

Practical Rest-Times for Battery Management Applications,” Published in IEEE Access, vol.

9, pp. 66989-66998, 2021, doi: 10.1109/ACCESS.2021.3077211.

Marvin Messing conducted the literature review, designed experiments, designed and

built test equipment, conducted experiments, analyzed results, and wrote paper II.

Dr. Shoa and Dr. Habibi provided research guidance and edited the paper.

Paper III

M. Messing, T. Shoa and S. Habibi, “Estimating Battery State of Health using

Electrochemical Impedance Spectroscopy and the Relaxation Effect,” Published in Journal

of Energy Storage, vol. 43, Article 103210, 2021, doi: 10.1016/j.est.2021.103210.

Marvin Messing conducted the literature review, designed experiments, designed and

built test equipment, conducted experiments, analyzed results, and wrote paper III.

- vii -
 Dr. Shoa and Dr. Habibi provided research guidance and edited the paper.

Paper IV

M. Messing, T. Shoa, R. Ahmed and S. Habibi, “Battery SoC Estimation from EIS using

Neural Nets,” Published in 2020 IEEE Transportation Electrification Conference & Expo

(ITEC), 2020, pp. 588-593, doi: 10.1109/ITEC48692.2020.91615231.

Marvin Messing conducted the literature review, designed experiments, designed and

built test equipment, conducted experiments, analyzed results, and wrote paper IV.

Dr. Shoa, Dr. Ahmed and Dr. Habibi provided research guidance and edited the paper.

Paper V

M. Messing, S. Rahimifard, T. Shoa, and S. Habibi, “Low Temperature, Current Dependent

Battery State Estimation using Interacting Multiple Model Strategy,” Published in IEEE

Access, 2021, doi: 10.1109/ACCESS.2021.3095938.

Marvin Messing conducted the literature review, designed experiments, designed and

built test equipment, conducted experiments, analyzed results, and wrote the majority of paper

V.

Sara Rahimifard provided SVSF-VBL Matlab modules, drafted the background theory

sections on SVSF, and provided paper edits.

Dr. Shoa and Dr. Habibi provided research guidance and edited the paper.

- viii -
Paper VI

M. Messing, T. Shoa and S. Habibi, “EIS from Accelerated and Realistic Battery

Aging,” Published in 2021 IEEE Transportation Electrification Conference & Expo (ITEC),

2021, pp. 720-725, doi: 10.1109/ITEC51675.2021.94900911

Marvin Messing conducted the literature review, designed experiments, designed and

built test equipment, conducted experiments, analyzed results, and wrote paper VI.

Dr. Shoa, Dr. Ahmed and Dr. Habibi provided research guidance and edited the paper.

1
In reference to IEEE copyrighted material which is used with permission in this thesis, the IEEE does not
endorse any of McMaster's products or services. Internal or personal use of this material is permitted. If
interested in reprinting/republishing IEEE copyrighted material for advertising or promotional purposes or for
creating new collective works for resale or redistribution, please go to
http://www.ieee.org/publications_standards/publications/rights/rights_link.html to learn how to obtain a License
from RightsLink.

- ix -
 Permission to Use

In presenting this thesis in partial fulfillment of the requirements for a Postgraduate degree

from McMaster University, I agree that the Libraries of this University may make it freely

available for inspection. I further agree that the permission for copying this thesis in any

manner, in whole or in part for scholarly purposes, may be granted by the professors who

supervised my thesis work or, in their absence, by the Head of the Department or the Faculty

Dean in which my thesis work was conducted. It is understood that any copying or publication

or use of this thesis or parts thereof for financial gain shall not be allowed without my written

permission. It is also understood that due recognition shall be given to me and McMaster

University in any scholarly use which may be made of any material in my thesis. Requests for

permission to copy or to make other use of material in this thesis, in whole or part, should be

addressed to:

Head of the Department of Mechanical Engineering

McMaster University
Faculty of Engineering
1280 Main Street West
Hamilton, Ontario L8S 4L6
Canada

-x-
 Table of Contents
Chapter 1: Introduction ................................................................................... 1
1.1 Lithium-Ion Batteries ...................................................................................... 2
1.1.1 Aging Mechanisms & Stressors .............................................................. 5
1.1.2 Impedance .............................................................................................. 7
1.1.3 Relaxation Effect .................................................................................... 8
1.2 Electrochemical Impedance Spectroscopy ..................................................... 11
1.3 Battery Modeling .......................................................................................... 15
1.3.1 Impedance Models ................................................................................ 19
1.3.2 Kramers-Kronig Transform and The Voigt-Circuit ............................... 23
1.3.3 Optimization for Parameterization ........................................................ 23
1.4 Battery State Estimation ................................................................................ 24
1.4.1 State of Charge ..................................................................................... 25
1.4.2 State of Health ...................................................................................... 27
1.5 Fractional Order Calculus.............................................................................. 28
1.6 Deep Neural Networks .................................................................................. 31
1.7 Battery Testing & Aging ............................................................................... 37
1.8 Custom Battery Cell Tester ........................................................................... 40
1.8.1 Battery Adapter..................................................................................... 48
1.9 Research Contributions ................................................................................. 50
1.9.1 Hypotheses ........................................................................................... 50
1.9.2 Contributions ........................................................................................ 52
1.10 Thesis Outline ............................................................................................... 56

Chapter 2: Lithium-Ion Battery Relaxation Effects........................................ 58

Abstract.................................................................................................................. 58
2.1 Introduction .................................................................................................. 59
2.2 Experimental ................................................................................................. 62
2.3 Results and Discussion .................................................................................. 62
2.3.1 Impedance Model ................................................................................. 63
2.3.2 Relaxation Effect .................................................................................. 66

- xi -
 2.3.3 Impedance Maps ................................................................................... 70
2.4 Conclusion .................................................................................................... 74
References ............................................................................................................. 74

Chapter 3: Electrochemical Impedance Spectroscopy with Practical Rest-times for

Battery Management Applications .............................................................................. 76
Abstract.................................................................................................................. 76
3.1 Introduction .................................................................................................. 77
3.1.1 Motivation and Technical Challenges ................................................... 77
3.1.2 Relaxation Effect and EIS Measurement ............................................... 80
3.1.3 Contributions ........................................................................................ 82
3.1.4 Paper Outline ........................................................................................ 82
3.2 Short Term EIS Measurement Method .......................................................... 83
3.3 EIS Characterization and Experiments .......................................................... 90
3.4 Relaxation Effect Analysis Method ............................................................... 93
3.5 Results & Discussion .................................................................................... 94
3.5.1 Relaxation Effect and C-Rate ................................................................ 94
3.5.2 Relaxation Effect and Voigt-Fit MSE.................................................... 96
3.5.3 Relaxation Effect and Δ𝑍 ...................................................................... 98
3.6 Conclusion ...................................................................................................101
Acknowledgement ................................................................................................102
References ............................................................................................................102

Chapter 4: Estimating Battery State of Health using Electrochemical Impedance

Spectroscopy and the Relaxation Effect .....................................................................105
Abstract.................................................................................................................105
4.1 Introduction .................................................................................................106
4.1.1 Motivation and Technical Challenges ..................................................106
4.1.2 Relaxation Effect and SoH Estimation .................................................107
4.1.3 Contributions .......................................................................................109
4.1.4 Paper Outline .......................................................................................109

- xii -
 4.2 EIS Characterization and Aging Experiments ...............................................109
4.3 Relaxation Effect and Impedance .................................................................113
4.4 Battery Modeling .........................................................................................116
4.5 SoH Estimation Method ...............................................................................120
4.5.1 Empirical SoH Model ..........................................................................120
4.5.2 Model Fitting Procedure ......................................................................122
4.6 Results and Discussion .................................................................................124
4.7 Conclusion ...................................................................................................132
Acknowledgements ...............................................................................................133
References ............................................................................................................133

Chapter 5: Battery SoC Estimation from EIS using Neural Nets ...................136
Abstract.................................................................................................................136
5.1 Introduction .................................................................................................136
5.2 Experimental ................................................................................................138
5.2.1 Characterization Tests and Data ...........................................................138
5.2.2 Battery Model ......................................................................................140
5.2.3 Network architectures ..........................................................................141
5.3 Results and Discussion .................................................................................143
5.4 Conclusion ...................................................................................................151
Acknowledgment ..................................................................................................152
References ............................................................................................................152

Chapter 6: Low Temperature, Current Dependant Battery State Estimation using

Interacting Multiple Model Strategy ..........................................................................154
Abstract.................................................................................................................154
6.1 Introduction .................................................................................................154
6.1.1 Motivation and Technical Challenges ..................................................155
6.1.2 Interacting Multiple Model Filter .........................................................157
6.1.3 Smooth Variable Structure Filter ..........................................................158
6.1.4 Contributions .......................................................................................159

- xiii -
 6.1.5 Paper Outline .......................................................................................159
6.2 Theory .........................................................................................................159
6.2.1 Battery Modeling & Parameterization ..................................................160
6.2.2 Extended Kalman Filter .......................................................................161
6.2.3 Smooth Variable Structure Filter ..........................................................163
6.2.4 Interacting Multiple Model Filters ........................................................164
6.2.5 Vehicle Modeling ................................................................................167
6.3 Low Temperature Drive Cycle Testing .........................................................172
6.3.1 Experiments and Experimental Setup ...................................................172
6.3.2 Drive Cycle Datasets............................................................................174
6.4 Results & Discussion ...................................................................................176
6.4.1 C-rate Specific Equivalent Circuit Models ...........................................176
6.4.2 Filter Initialization and Baseline...........................................................180
6.4.3 Performance Comparison .....................................................................182
6.4.4 IMM Voltage Modeling & Mode Probabilities .....................................186
6.5 Conclusion ...................................................................................................190
Acknowledgment ..................................................................................................190
References ............................................................................................................190

Chapter 7: EIS from Accelerated and Realistic Battery Aging ......................193

Abstract.................................................................................................................193
7.1 Introduction .................................................................................................193
7.2 Methods and Theory ....................................................................................196
7.2.1 Electric Vehicle Modeling ...................................................................196
7.2.2 Battery Aging ......................................................................................198
7.2.3 Battery Modeling .................................................................................201
7.3 Results and Discussion .................................................................................203
7.3.1 Voltage and Temperature Response .....................................................203
7.3.2 EIS Characterization ............................................................................206
7.4 Conclusions and Future Work ......................................................................211
Acknowledgment ..................................................................................................211

- xiv -
 References ............................................................................................................211

Chapter 8: Summary, Conclusions and Recommendations ...........................213

8.1 Research Summary ......................................................................................213
8.2 Recommendation for Future Research ..........................................................217

References .................................................................................................................219

Table of Figures
Figure 1-1: Lithium-ion battery components and operation ........................................... 3
Figure 1-2: Overpotentials Caused by Impedances ........................................................ 8
Figure 1-3: Battery Voltage Relaxation. ...................................................................... 10
Figure 1-4: Battery Relaxation Effect Process Long Term (a), Short Term (b). ........... 10
Figure 1-5: EIS Signal and Analysis ........................................................................... 12
Figure 1-6: A discrete time domain sine wave (a) transformed in to the frequency domain using
the Fourier Transform (b). ................................................................................... 13
Figure 1-7: Nyquist plot showing EIS results in frequency domain [28]. ..................... 15
Figure 1-8: ECM to model battery internal processes separated for anode and cathode (a) and
combined (b). ...................................................................................................... 17
Figure 1-9: RC battery model of order n ..................................................................... 18
Figure 1-10: EIS Nyquist plot (a), with RC model fit (b), with fractional fit (c). .......... 20
Figure 1-11: Impedance model used to model EIS response ........................................ 22
Figure 1-12: Representation of CPE element with RC approximation or FOC. ............ 29
Figure 1-13: Common impedance models ................................................................... 31
Figure 1-14: Neuron ................................................................................................... 32
Figure 1-15: Simple Neural Network .......................................................................... 33
Figure 1-16: Experimental OCV-SoC curve results for new and aged batteries. .......... 38
Figure 1-17: Simple Vehicle Model ............................................................................ 40
Figure 1-18: Custom battery test system diagram (only one channel shown) ............... 41
Figure 1-19: Multiplexer diagram for 2 channels......................................................... 43
Figure 1-20: Custom battery test system software architecture .................................... 44

- xv -
Figure 1-21: Graphical user interface for custom battery test system ........................... 45
Figure 1-22: Database architecture for custom battery test system ............................... 48
Figure 1-23: Custom battery adapter for custom battery test system. 1 – battery, 2 – adapter
base, 3 – screw terminal connector, 4 – nickel tab. .............................................. 49
Figure 1-24: Battery welding fixture ........................................................................... 50
Figure 1-25: Contributions Overview .......................................................................... 52
Figure 2-1: Nyquist plots for 100%, 70% and 30% SoC at 25°𝐶 and 40°𝐶. ................. 63
Figure 2-2: Impedance model used to model relaxation effect a), and model fit to relaxation
data at 25°C and 90% SoC b). ............................................................................. 64
Figure 2-3: Impedance model fitting error with different SoCs over different rest times.65
Figure 2-4: Modelled relaxation effect at 90% SoC and 25°𝐶 a), and 40°𝐶 b). ............ 66
Figure 2-5: Percentage change in impedance model parameters from values at 420 minutes at
25°C for a) 100% SoC, b) 90% SoC, c) 70% SoC and d) 30% SoC. .................... 68
Figure 2-6: Percentage change in impedance model parameters from values at 420 minutes at
40°C for a) 100% SoC, b) 90% SoC, c) 70% SoC and d) 30% SoC. .................... 70
Figure 2-7: Impedance maps showing values for a) Qp, b) Qd and c) Rp for different SoCs and
rest times at 25°C. ............................................................................................... 72
Figure 2-8: Impedance maps showing values for Qp, Rp, and Qd for different SoCs and rest
times at 25°C (a-c), and 40°C (d-f). ..................................................................... 73
Figure 3-1: Three repeats of EIS measurements (set 1, 2 and 3) for different rest times after a
5C discharge pulse. ............................................................................................. 83
Figure 3-2: Effect of drift compensation with 1-minute rest. ....................................... 84
Figure 3-3: Voigt Circuit............................................................................................. 85
Figure 3-4: Real and imaginary EIS data with Voigt fit at 5-minute rest (a), EIS gain and phase
with Voigt fit at 5-minute rest (b), Voigt fit residuals at 5-minute rest (c), Voigt fit MSE
for various rest times and datasets (d) all after 5C discharge. ............................... 87
Figure 3-5: Voigt filter smoothing result. .................................................................... 89
Figure 3-6: Short rest time EIS measurement methodology. ........................................ 90
Figure 3-7: Experimental protocol showing where EIS measurements are taken with respect to
discharge pulses of different C-rates. ................................................................... 91
Figure 3-8: Schematic of experimental setup and battery cell. ..................................... 92

- xvi -
Figure 3-9: ΔZ(ω) for different frequencies and rest times. ......................................... 94
Figure 3-10: Evolution of Nyquist plots measured at different rest times after different C-rates.
........................................................................................................................... 95
Figure 3-11: Voigt-fit MSE for different C-rates and SoCs (a), MSE for different C-rates and
temperatures (b), mean and SD of combined conditions (c). ................................ 97
Figure 3-12: ΔZ for 90%, 70% and 50% SoC compared to reference at 25°C and 90% SoC (a).
Nyquist plots for 90% and 50% SoC after 30 minutes of rest (b). ........................ 99
Figure 3-13: ΔZ at 25°C, 35°C and 10°C compared to reference at 25°C and 90% SoC (a).
Nyquist plots for the same temperatures (b)........................................................100
Figure 4-1: Schematic of experimental setup and battery cell. ....................................110
Figure 4-2: Three repeats of a 5C, 15s discharge pulse at 90% SoC, 100% SoH. .......112
Figure 4-3: Evolution of Nyquist plot with rest time and SoH. ...................................114
Figure 4-4: Short-term relaxation effect impedance evolution tracked with EIS data (plots
highlighted according to rest time) and conceptually pictured with battery half-cell
diagrams. ...........................................................................................................116
Figure 4-5: Nyquist plots of low and high SoH EIS measurements with model fits (a),
fractional order impedance model (b). ................................................................118
Figure 4-6: Evolution of 𝑹𝒆𝒒 with rest time at 100% SoH (a), log-linear trend of 𝑹𝒆𝒒 for rapid
relaxation region (b). ..........................................................................................121
Figure 4-7: Procedure to obtain SoH estimate from EIS data at different rest times. ...124
Figure 4-8: Short term evolution of Req with rest time (a) and semi-log-linear fits to the initial
rate of change of Req (b). ....................................................................................126
Figure 4-9: Req short term relaxation evolution slopes and SoH. ...............................127
Figure 4-10: Estimated SoH vs. True SoH and slope repeatability (a), SoH Estimation Error
and RMSE (b). ...................................................................................................129
Figure 4-11: 𝑹𝒆𝒒 slopes for different SoCs and Temperatures for battery at 100% SoH.130
Figure 5-1: Nyquist plots for 100%, 70% and 30% SoC at 25°C and different SoH (a). Nyquist
plots for 95% SoC, 90% SoH at 25°C after 15 second, 5C discharge pulse measured
different rest times (b). .......................................................................................139
Figure 5-2: Impedance model used to model relaxation effect (a), and model fit to relaxation
data at 25°C and 90% SoC (b). ...........................................................................141

- xvii -
Figure 5-3: Network structures for raw EIS input (a), and EIS model parameters (b)..142
Figure 5-4: RMSE evolution during training epochs for DNN using raw EIS data (a), and DNN
using EIS model parameters (b)..........................................................................145
Figure 5-5: RMSE evolution during training epochs for DNN using raw EIS data (a), and DNN
using EIS model parameters (b)..........................................................................147
Figure 5-6: RMSE for validation and training for different number of network layers (a), and
different number of nodes per layer (b). .............................................................149
Figure 5-7: RMSE for each SoC. ...............................................................................150
Figure 6-1: 3rd order battery equivalent circuit model. ...............................................160
Figure 6-2: OCV-SoC Curve for different temperatures. ............................................161
Figure 6-3: IMM Algorithm Diagram. .......................................................................166
Figure 6-4: Current profile samples from EPA drive cycles for low C-rates (a) medium and
high C-rates (b). Validation drive cycle current profile (c). .................................171
Figure 6-5: Experimental setup diagram (a), picture and custom control software screen capture
(b). .....................................................................................................................173
Figure 6-6: Drive Cycles at 0°C and 100% SoH for parameterization of the high C-rate model
(a), the medium C-rate model (b), the low C-rate model (c), and for validation (d).176
Figure 6-7: Performance for L1 (a), M1 (b), and H1 (c) c-rate specific models at 100% SoH
and 0°C. .............................................................................................................179
Figure 6-8: Change in model resistances with C-rate at 0°C. ......................................180
Figure 6-9: Baseline filter performance at 100% SoH, 25°C (a) and 0°C (b). .............182
Figure 6-10: IMM SoC Estimation and error at 0°C for 100% SoH (a, c) and 90% SoH (b,d),
C-rates for reference (e, f). .................................................................................184
Figure 6-11: Impact of current bias on RMSE for 100% SoH (a) and 90% SoH datasets.185
Figure 6-12: Voltage and voltage error of the three models (L1, M1, H1) used in the IMM-
SVSF at 100% SoH and 0°C. .............................................................................187
Figure 6-13: Mode probabilities for IMM-EKF at 100% SoH (a), 90% SoH, (b) and IMM-
SVSF 100% SoH (c), 90% SoH (d). C-rates shown for reference in (e) and (f). ..189
Figure 7-1: Aging test load profiles in units of C-rate for drive cycle aging test (a) and
charge/discharge aging test (b). ..........................................................................200

- xviii -
Figure 7-2: Aging test procedure showing the usage of different drive cycle protocols, different
charge rates, and EIS steps. ................................................................................201
Figure 7-3: Equivalent circuit battery model fits to Nyquist plots (a), fractional order model
circuit (b), RC-based circuit without SEI layer branch (c)...................................203
Figure 7-4: Drive cycle aged and charge/discharge aged voltage profiles for fresh cells (a) and
cells at 85% SoH (b). .........................................................................................204
Figure 7-5: Comparison of temperature profiles for fresh cells and cells at 85% SoH during
drive cycle aging protocol (a) and charge/discharge protocol (c). .......................206
Figure 7-6: Nyquist plot evolution at various SoH targets with fractional order model fits for
drive cycle aged EIS data (a) and charge/discharge aged EIS data (b). ................208
Figure 7-7: Battery model parameter evolution with SoH for drive cycle and charge/discharge
aging for ohmic parameter R_0 (a), SEI layer parameter R_1, and charge transfer
parameter R_2 (c). .............................................................................................210

- xix -
 List of Tables
Table 1-1: Custom battery test system parts list........................................................... 42
Table 3-1: Summary of Minimum Rest Time for Different Conditions.......................100
Table 4-1: SoH and Test Conditions for 14 Battery Cells at 90% SoC and 25°C. .......111
Table 4-2: Model Parameter Bounds and Example Values at 90% SoC, 100% SoH, 25°C and
15 minutes Rest ..................................................................................................122
Table 4-3: Semi-log Fit (Equation 5) Results and Statistics ........................................126
Table 4-4: Double Exponential Model (Equation 6) Fit Results and Statistics ............127
Table 4-5: Repeatability of the SoH Estimation Method.............................................127
Table 4-6: Method Comparision ................................................................................132
Table 5-1 Network Hyper Parameters ........................................................................143
Table 5-2: Network Performance ...............................................................................145
Table 6-1: Simple Vehicle Model Parameters [29], [34], [35] ....................................169
Table 6-2: Battery Model Parameters .........................................................................178
Table 6-3: Filter Initialization Parameters ..................................................................181
Table 6-4: Filter RMSE Comparison for validation cycles at 0°C ...............................184
Table 7-1: Vehicle Model Parameters for Tesla Model 3 Long Range ........................198

- xx -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

Chapter 1: Introduction

To have one or several vehicles standing at the ready at any time to quickly get people to

near locations within tens or even hundreds of kilometers is one of the great luxuries

enjoyed by many since the late 20th century. Although public transportation is well

developed in many countries, there is nothing quite like the feeling of freedom that comes

with owning a vehicle. However, the majority of vehicles still emit exhaust gases which

contribute to two major issues: 1) air pollution and 2) the climate crisis. The findings of a

recent study on global air pollution highlight again the need to remove sources of pollution

from densely populated areas [1], where people and vehicles are found in large

concentrations. For example, the study found that over one million people die each year

due to air pollution in China alone. In addition, it found that air pollution is causing a

reduction of almost two years in the global average life expectancy. The sources and impact

of the climate crisis are detailed in a United Nations (UN) report, which predicts further

increases in flooding, droughts and hurricanes around the world, if Greenhouse Gas (GHG)

emissions are not reduced drastically [2]. Global transportation comprises around 15% of

all GHG emissions [3]. In the United States, this percentage is almost double (29%) and

almost 60% of that is due to light duty vehicles [4]. Therefore, the automotive industry is

heavily investing in Electric Vehicles (EV) as part of the solution to the global air pollution

problem and climate crisis.

Battery packs in EVs are overdesigned to carry more capacity than they use. This is done

to ensure that the battery operating range stays within safe charge and discharge limits [5],
-1-
 Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

[6]. However, accurately determining those limits over a wide range of operating conditions

and variations in battery designs is a significant challenge [7]. EV battery packs are

continuously monitored by Battery Management Systems (BMS) which contain advanced

algorithms to achieve optimal and safe operation. Extensive research is currently carried

out find new ways to maximize the useable battery capacity. One of the most challenging

aspects of battery management is the estimation of hidden battery states. This thesis

explores state estimation for lithium-ion batteries and EV applications. The findings of this

thesis contribute to the improvement of battery state estimation through in-depth research

into impedance spectroscopy, adaptive filters, and machine learning methods. In this

Chapter, background material is reviewed in Sections 1.1 to 1.8 including introductions to

Lithium-Ion batteries, impedance spectroscopy, state estimation, neural networks and

battery testing. In Section 1.9 the research hypotheses and contributions are described.

Section 1.10 provides an outline of the thesis.

1.1 LITHIUM-ION BATTERIES

Lithium-ion chemistries dominate the market for secondary batteries, especially, for all-

electric vehicles [8]. A lithium-ion battery consists of an anode (negative electrode),

electrolyte, a separator, and a positive electrode (cathode), as shown in Figure 1-1. During

the charge process, electrons flow into the anode and combine with lithium-ions to

intercalate lithium into the electrode. During discharge, electrons flow out of the anode and

into the cathode to intercalate there [9]. The porous separator ensures that electrons cannot

pass from anode to cathode without an external circuit. There is a tendency for electrolyte

-2-
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

reduction at the negative electrode (Anode) due to its low potential. This causes the

electrolyte to react with the electrode and leads to the formation of the Solid Electrolyte

Interphase (SEI) layer. After the SEI layer has formed, it protects the electrode from further

undesirable reactions with the electrolyte. On the positive electrode a similar layer is called

the Cathode Electrolyte Interphase (CEI) [10]. The CEI is thinner since there is only a

small, or no driving potential on the cathode.

Figure 1-1: Lithium-ion battery components and operation

There are several variants of lithium-ion chemistries with a broad range of properties.

Lithium cobalt oxide (LCO) batteries are the most common type in many applications.

However, the high cost, poor durability and low availability of cobalt makes LCO an

unfavourable choice for large battery packs [11,12]. Instead, nickel cobalt aluminum oxide

(NCA) cells can eliminate some of the drawbacks of LCO, but safety concerns remain.

Specifically, the risk of thermal runaway is high in NCA batteries which has lead to several

EV fires [12]. Nickel manganese cobalt oxide (NMC) lithium-ion cells have high energy

-3-
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

density [12] and are, therefore, chosen by some car manufactures for EV battery packs.

However, the thermal runaway risk is still present, although only at higher temperatures

compare to NCA. Examples of catastrophic battery failure also exist for NMC battery packs

[12]. Lithium-ion phosphate (LFP) and LCO with spinal lithium titanium oxides (LTO)

anodes deliver the best results in terms of safety, but energy density is reduced [8]. Thus,

the selection of presently commercialized lithium-ion batteries is a trade-off between

energy density (which translates to EV range) and safety.

There is growing concern that conventional lithium-ion batteries (LIB) cannot meet the

increasing energy storage demands for mobile applications. Therefore, improvements to

lithium-based and alternative chemistries are constantly explored. One promising

improvement to lithium-ion chemistries is the use of Solid-State Electrolyte (SSE) instead

of organic solvents. The use of SSE allows for increased energy density, increased battery

safety as well as reduced volume due to bi-polar cell stacking [13], [14]. However,

challenges concerning the transport properties of SSEs as well as the scalability of

manufacturing processes must be addressed before SSEs become a viable solution.

Nonetheless, Solid State Batteries (SSB) are presently regarded as the next leap forward in

battery technology.

Some advances have been made in the development of sodium-ion batteries (NIB). In [15]

the progress and challenges are discussed. NIBs show promising performance; however,

more studies are required to form a complete picture of the properties of NIBs with regards

to lifetime, safety, and cost.

-4-
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

1.1.1 AGING MECHANISMS & STRESSORS

The useful lifetime of a lithium-ion battery is a function of its operating conditions. To

achieve maximum lifetime, the battery should be kept at room temperature and charged or

discharged at low current, strictly between optimal voltage limits. However, for automotive

applications, for example, the operating temperatures can range from -30°𝐶 to +52°𝐶 [16],

charge rate is high, to reduce charging time, and discharge profiles depend on the driving

habits of the operator. The impact of deviating from optimal operating conditions and the

resulting battery degradation, are studied extensively in literature, for a range of automotive

battery types. Gao et al. studied voltage limits as well as charge rates of an 18650 LCO

battery and found a drastic increase in degradation rates, for voltages above 4.2 V and

charge rates above 1C [17]. Su et al. applied different stressors and found that charge

current, charge voltage, temperature, charge current, and discharge current contribute

heavily to battery degradation of an 18650 NCA battery [18]. They separate degradation

due to temperature into low-temperature (below 25°𝐶) degradation and high-temperature

(above 25°𝐶) degradation. This difference in temperature degradation mechanisms was

also found by Waldmann et al., for an 18650 NMC battery [19]. A prismatic NMC cell was

studied by Waag et al. to show the change in battery impedance (see Section 1.1.2) with

temperature, SoC, and aging; however, the aging procedure is not specified [20]. Leng et

al. used a prismatic LCO battery to study the effect of temperature on the parameters of a

combined electrochemical-electric model [21]. They found the parameters changed

significantly with increased operating temperature between 25°𝐶 and 55°𝐶 and related each

-5-
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

parameter to the degradation of components inside the battery. In addition, they showed the

degradation of the electrodes to be the main contributor to aging.

Another form of battery aging is calendar aging which refers to degradation that occurs

while the battery is at rest over long periods of time. Calendar aging was investigated by

[22–25], but no clear consensus exists. This may be due to a combination of different

behaviours between chemistries as shown by [25], who showed different aging behaviors

with temperature for LMO, NMC, LFP, and NCA. The lack of consensus may also be due

to additional degradation introduced by testing procedures as mentioned by [22] who argue

that some types of characterization tests introduce additional aging. Since calendar aging

happens slowly over a long period of time, it is often difficult to distinguish it from

degradation caused by diagnostic testing.

Finally, some studies show that vibration also causes battery degradation [26,27]. However,

a recent study by Hooper et al. [28] shows only a small, yet statistically significant, increase

in ohmic resistance after an equivalent vibration profile of 10 years of EV driving, for

18650 NCA cells. This study included control samples to differentiate degradation related

to vibration from calendar aging.

A detailed review of lithium-ion battery aging mechanisms is given by [11]. The main

mechanisms are the growth of SEI, the buildup of other surface films on the electrodes, and

the break-down of electrode material. The initial formation of the SEI is important for the

operation of the battery as it electrically separates the electrolyte from the electrode.

However, over time it grows by consuming lithium-ions which causes capacity fade.

-6-
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

1.1.2 IMPEDANCE

The components of the battery as well as the electrochemical reactions taking place during

battery operation cause several overpotentials (or losses). Figure 1-2 shows a typical

voltage drop which occurs when current is drawn from the battery. The voltage profile

shows three drop regions which represent different overpotentials. These overpotentials can

be grouped as follows:

1. Overpotential 𝜂𝑂 manifests as the immediate voltage drop observed within a few

milliseconds after current is applied. It is caused by the ohmic resistances of the cell which

include current collectors, the solid parts of the electrodes, electrolyte resistance, and

interface resistances between current collectors and electrodes.

2. Overpotential 𝜂𝐶 causes a further voltage drop within a few seconds and is a combination

of charge transfer losses 𝜂𝐶𝑇 , SEI losses 𝜂𝑆𝐸𝐼 and contact resistance losses 𝜂𝐶𝐶 .

𝜂𝐶 = 𝜂𝐶𝑇 + 𝜂𝑆𝐸𝐼 + 𝜂𝐶𝐶

(1-1)
3. Overpotential 𝜂𝐷 is a result of diffusion losses from lithium inside the electrodes (solid

state) as well as from lithium ions in the electrolyte (liquid state) and occurs during

sustained current draw.

Each of the overpotentials occur because of some form of opposition to current flow

through the battery. This current flow opposition is measured by impedance, a combination

of resistance and reactance (capacitive and inductive) effects. The impedance of a battery

changes depending on the battery chemistry and materials used, the excitation signal, the

SoC, the SoH and the battery temperature. In Section 1.2 a method is introduced that can
-7-
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

identify the impedance of the different battery components in a non-destructive, in-vivo

way. Studying battery impedance under different operating conditions can provide insight

into battery dynamics and battery states. In Section 1.3 battery models are described which

imitate the impedance behavior of batteries. Accurate battery models are a crucial part of

battery state estimation. In the next section the battery relaxation effect is discussed which

is closely related to the impedance characteristics of a battery.

Figure 1-2: Overpotentials Caused by Impedances

1.1.3 RELAXATION EFFECT

When a load is applied to a battery its voltage drops due to overpotentials which occur

because of impedances as described in the previous section. If the load is removed the

battery voltage does not instantly rise back to a constant Open Circuit Voltage (OCV) but

instead recovers slowly, stabilizing over time. This is known as the battery relaxation effect.

Figure 1-3 shows battery voltage behaviour during a discharge pulse followed by a rest
-8-
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

period at OCV. At the start of the rest period, the voltage rises immediately to recover the

ohmic losses. During the first few minutes of the rest period, the voltage continues to rise

as lithium ions diffuse back into the electrodes as shown in Figure 1-4b. The relaxation

effect continues for hours after that with the voltage increasing slowly over time while

lithium is diffusing inside the electrodes to equalize micro potentials as shown in Figure

1-4a [29], [30]. The relaxation effect has several implications for battery characterization

and management:

• Battery models must reflect relaxation behaviour. The accuracy of battery models is
crucial to the performance of state estimation algorithms, so they must capture both
load conditions and rest conditions accurately. Additional battery model terms may
be necessary to capture the time constants associated with the long-term effects.

• The OCV of a battery is a useful quantity for SoC estimation. However, true OCV
only occurs when the relaxation effect has slowed sufficiently. This means long rest
times are required before accurate OCV can be measured.

• Rest times must be included for most test procedures to ensure consistent response
of the battery. A relaxed battery will respond differently to loads than a battery which
was recently excited. Therefore, rest times of 1 to 4 hours are usually observed in
between excitation tests.

-9-
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

Figure 1-3: Battery Voltage Relaxation.

Figure 1-4: Battery Relaxation Effect Process Long Term (a), Short Term (b).

In the next section, an impedance measurement technique is described for which the impact

of the relaxation effect on measurement results must be considered.

- 10 -
 Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

1.2 ELECTROCHEMICAL IMPEDANCE SPECTROSCOPY

One of the major areas of research presented in this thesis concerns the results obtained

using Electrochemical Impedance Spectroscopy (EIS). EIS can reveal a detailed snapshot

of the impedance of the internal components of batteries, which can be interpreted to

indicate the health of various components, the battery state, temperature, and fault

conditions. EIS is based on small signal excitation at different frequencies. For example, in

Galvanostatic EIS (GEIS) a sinusoidal current signal is applied to the battery at different

frequencies and the corresponding voltage response is measured. The current signal must

be sufficiently low for the battery to remain in its linear response region, but high enough

to maintain adequate signal to noise ratio. The current value which satisfies these

requirements depends on the battery type and size as well as hardware characteristics of the

measurement equipment. A linear response of the battery is usually ensured if the battery

voltage stays within +/- 10 mV throughout the EIS test. A voltage-controlled version of

EIS, called Potentiostatic EIS (PEIS), can guarantee this voltage range by applying voltage

and measuring the current. PEIS is primarily used in the work described in this thesis. The

measured EIS data can be converted to frequency domain using the Discrete Fourier

Transform (DFT) [31] to produce a characteristic plot called the Nyquist plot as shown in

Figure 1-5.

- 11 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

Figure 1-5: EIS Signal and Analysis

Given a sine wave signal 𝑓(𝑖) in time domain with discrete points 𝑖 from 0 to 𝑁 − 1, the

DFT 𝐹(𝑛) at point 𝑛 in the frequency domain can be calculated for each 𝑛 from 0 to 𝑁 − 1

using Equation 1-3.

1 2𝜋𝑛𝑖
𝐹(𝑛) = 𝑁 ∑𝑁−1
𝑖=0 𝑓 (𝑖 ) exp (−𝑗 ) (1-2)
𝑁

The DFT 𝐹(𝑛) produces complex numbers (real and imaginary pairs) for each 𝑛 which

correspond to frequencies from 0 to the sampling frequency 𝑓𝑠 . The exact frequency 𝑓𝑛 for

each 𝑛 is calculated using Equation 1-3.

𝑛
𝑓𝑛 = 𝑓𝑠 (1-3)
𝑁

However, the highest frequency 𝑓𝑚𝑎𝑥 which can be observed is half of the sampling

frequency according to the Nyquist theorem as shown in Equation 1-4.

1
𝑓𝑚𝑎𝑥 = 𝑓𝑠 (1-4)
2
𝑁−1
Therefore, only the first half of the DFT results from 𝑛 = 0 to 𝑛 = produce meaningful
2

results, beyond that the results repeat. The following is an example of the DFT calculation:

- 12 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

2𝜋𝑓1 𝑖 𝑓𝑠
Assuming 𝑓 (𝑖 ) = sin ( ) with 𝑓1 = 10𝐻𝑧 and 𝑓𝑠 = 0.5𝑘𝐻𝑧, 𝑁 = = 50 points are
𝑓𝑠 𝑓1

obtained for one sine wave period as shown in Figure 1-6a. The Matlab implementation of

the Fast Fourier Transform (FFT), a more computationally efficient version of the DFT,

can be used to obtain the transformed sine wave data. Figure 1-6b shows the magnitude of

the FFT output for frequencies from 0 to 𝑓𝑠 . The first peak corresponds to the frequency of

the time domain sine wave signal with the frequency domain index 𝑛 = 1, which can be

found using Equation 1-3. The complex valued FFT result at 𝑛 = 1 is 𝐹(1) = 1.55 −

j25.18 with an absolute value of |𝐹 | = 25.23.

Figure 1-6: A discrete time domain sine wave (a) transformed in to the frequency domain
using the Fourier Transform (b).

The signal in the example above could be one frequency of the EIS signal applied to a

battery, a sine wave with 1 Ampere amplitude for example. The impedance can then be

- 13 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

calculated using Equation 1-5. Here, 𝐹𝑖𝑛 (𝑛) is the FFT (or DFT) of the input signal and

𝐹𝑜𝑢𝑡 (𝑛) the FFT of the output signal, the voltage response of the battery in the case of this

example. The frequency domain index corresponding to the signal frequency 𝑓𝑛 is 𝑛 = 1.

𝑉
Finally, the impedance 𝑍 is obtained using Ohm’s law 𝑅 = , or in this case, the FFT of
𝐼

the voltage signal at index 𝑛 divided by the FFT of the current signal at index 𝑛 is equal to

the impedance as shown in Equation 1-5.

𝑍 = 𝐹𝑜𝑢𝑡 (1)/𝐹𝑖𝑛 (1) (1-5)

The above method assumes a signal which is low in noise and has an integer valued number

of periods per sine wave frequency. The former depends on the quality of the

instrumentation and the latter on the signal generation. If the signals are noisy, the

measurements can be averaged over several sinewave periods to produce improved results.

If the signal contains incomplete sine wave periods, windowing can be applied to the FFT

to remove the incomplete signal parts. In laboratory settings, potentiostats are frequently

used to perform EIS tests. These devices contain precision circuitry to generate clean sine

waves and accurately measure the response of electrochemical devices such as batteries.

Most modern portentiostats also include on-board signal post-processing features.

The impedance 𝑍 for frequency 𝑓𝑛 represents a point on the Nyquist plot, where the real

part of 𝑍 is plotted against the imaginary part. By convention, the imaginary part is

multiplied by −1. Figure 1-7 shows an example of a Nyquist plot with real and imaginary

pairs plotted for EIS excitation signals between 700 Hz and 0.4 Hz [32]. EIS, Nyquist plots

- 14 -
 Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

and how the data can be used for battery modeling are discussed more in Chapters 2, 3 and

5.

Figure 1-7: Nyquist plot showing EIS results in frequency domain [32].

EIS is used heavily in laboratory settings for battery characterization. However, studies

have indicated some success in implementing feasible setups to perform EIS on-board a

vehicle for real-time diagnostics. Abedi et al. [33] reviewed state-of-the art methods for

implementing online EIS and found that efforts are centered on utilizing parts of existing

balancing circuitry to generate the EIS signal. Further real-time EIS solutions are discussed

in Chapters 3 and 6.

1.3 BATTERY MODELING

Battery models fall into three categories: Electrochemical Models (EM) (based on

fundamental equations), empirical white box models, and empirical black box models.

Black box models are usually used to model not just the measurable behaviour of a battery,
- 15 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

but also the hidden states. Battery states and black box models are discussed in sections 1.4

and 1.7. EMs have the potential to be highly accurate since they can closely reflect the real

behaviour of batteries. However, to increase accuracy, more complex processes must be

modeled, increasing the overall complexity of the model. In particular, differential

equations are frequently used for EMs, requiring significant computational resources. In

addition, constants associated with the physical processes must be parameterized by

performing extensive laboratory experiments [34]. Therefore, EMs are impractical for real

time applications, such as battery management systems in electric vehicles (EV) [35]. In

contrast, empirical models can be simple and still provide high accuracy, but only for a

limited range of conditions. Furthermore, accuracy is often hard to interpret since it depends

on the quality of the underlying data and experiments [36]. Still, empirical models can

provide a suitable trade-off between accuracy and complexity.

The most popular white-box models are equivalent circuit models (ECM). ECMs model

batteries as circuits of ideal electrical components (resistors, capacitors) to approximate

processes inside the battery. For example, the overpotentials described in Section 1.1.2

could be modeled using the circuit shown in Figure 1-8a [37]. In this circuit, series

resistances are used to represent the ohmic resistances at the anode and cathode, and

resistor-capacitor (RC) pairs model impedances of the interface layers (SEI and CEI),

electrode/electrolyte charge transfer, and electrode diffusion. However, the values for each

capacitance and resistance must be determined experimentally, and simple, non-destructive

battery tests cannot distinguish between anode and cathode overpotentials. In addition, the

SEI impedance is much more dominant than the CEI impedance, so CEI is often omitted.
- 16 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

Instead, the anode and cathode circuit elements are combined into a series resistance

followed by RC pairs as shown in Figure 1-8b. Depending on the application, additional

RC-elements can be added to capture battery behaviour more closely. For example,

dynamic load profiles of EV applications may require additional RC element with small

time constants. Stationary battery applications, where load profiles are constant for longer

periods of time, may required larger time constants.

Figure 1-8: ECM to model battery internal processes separated for anode and cathode (a)
and combined (b).

Figure 1-9 shows the fundamental ECM which consists of RC parallel pairs repeated n

times in series. Since RC pairs only loosely represent the physical processes, 𝑛 can be

increased to model experimental data more closely. Many different circuits have been

proposed to increase accuracy with fewer circuit elements, each with their own advantages

and drawbacks [38]. However, there are an infinite number of circuit configurations
- 17 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

possible for any experimental result, which questions the relevance of fitted parameters.

Each circuit element must be carefully related to physical phenomena inside the battery to

give ECM models more relevance.

Figure 1-9: RC battery model of order n

Nonetheless, ECMs are used in EVs for battery state estimation due to their simplicity. First

and second order RC circuits were compared by [39], showing a reduction in SoC

estimation error for the higher order models. In [40], a third order RC circuit was used,

showing further reduction in estimation error. ECM parameters are usually obtained using

a discharge pulse method, where the characteristics of the voltage response can be related

to resistance and capacitance values. A common protocol is the hybrid pulse power

characterization (HPPC) pulse [41] which is used by the US Department of Energy (DoE).

ECM parameters can also be obtained by fitting the circuit to the characteristic frequency

response of a battery using EIS (Section 1.2). EIS gives a much more detailed picture of

the battery response and shows the limitations of RC based circuits as shown by Farmann

et al. [7]. They showed how RC-based circuits can only roughly approximate the shape of

the battery impedance response. In the next section, impedance models are introduced, an

improved version of ECMs which can model EIS data more closely.

- 18 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

1.3.1 IMPEDANCE MODELS

The impedance profile of a battery holds useful information about its internal condition.

Impedance is the combination of reactance and resistance and comes from the interaction

between different materials inside a battery, the material characteristics, and from chemical

reactions [42] as was introduced in Section 1.1.2. A detailed picture of the battery

impedance can be obtained using EIS as described in Section 1.2. A Nyquist plot obtained

from EIS data of a fresh lithium-ion cell is shown in Figure 1-10a. The ECM shown in the

previous section (Figure 1-9) with 𝑛 = 2 can be used to fit to the Nyquist plot. For this

data, with 𝑛 = 2 (i.e. two RC-elements) the charge transfer process and the diffusion

process are captured. The SEI impedance is negligible here, due to the high SoH and

specific properties of the cell used. The ECM fit is shown in Figure 1-10b. As is apparent

in the figure, the ECM fit does not follow the data very closely. This is because processes

like charge transfer in a battery are non-ideal and, therefore, exhibit a distribution of time

constants, rather than a single time constant modeled by one RC-branch. Increasing 𝑛 can

capture more time constants and improve the fit, however, the model complexity grows and

the meaning of model parameters is obscured.

- 19 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

Figure 1-10: EIS Nyquist plot (a), with RC model fit (b), with fractional fit (c).

Alternatively, the ECM can be refined by introducing a non-ideal circuit component: The

Constant Phase Element (CPE) [31]. The CPE can model the capacitive behaviour of non-

ideal processes with only a single additional model parameter 𝛼, the constant phase

exponent. The CPE impedance is slightly modified from the ideal capacitor impedance
1
(𝑍𝐶 = 𝑖𝜔𝐶 ) as shown in Equation 1-6.

1
𝑍𝐶𝑃𝐸 = 𝑄(𝑖𝜔)𝛼 (1-6)

- 20 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

Here, 𝑄 is related to the capacity in units of (𝐹 𝑠 𝛼−1 𝑐𝑚−1 ). The exponent 𝛼 has a range

of 0 ≤ 𝛼 ≤ 1, and for 𝛼 = 1, 𝑍𝐶𝑃𝐸 = 𝑍𝐶 (with 𝑄 = 𝐶), and the CPE becomes an ideal

capacitor. The CPE models a non-ideal capacitor which has an imaginary component as

well as a real component. Physically this means the capacitor is “leaking”, i.e. experiencing

a current drain. The CPE can also be used in parallel with a resistor and the combined

circuit is called a Zarc element. Figure 1-11 shows a circuit which uses a Zarc element to

model the charge transfer processes (anode and cathode combined) and a CPE to model

diffusion processes (also anode and cathode combined). In addition, an inductor is added

to model the inductive effects resulting from cell connections and measurement cables at

high frequencies. Such models are often referred to as impedance models to distinguish

them from RC circuits [43]. The fit of this model to the EIS data is shown in Figure 1-10c.

This model can accurately reflect the impedance behaviour of the battery. The Zarc element

(𝑄1 , 𝛼1 ||𝑅1 ) results in a depressed semi-circle due to the use of the CPE. This better reflects

the multiple time constants associated with the charge transfer processes, which is non-

ideal due to surface roughness and porous electrodes. The CPE in series creates an angled

line, where the angle is defined by the exponent 𝛼. If 𝛼 = 1, the angle is 90° resulting in a

vertical line representing an ideal capacitor. This series CPE (𝑄2 , 𝛼2 ) models the non-ideal

diffusion processes inside the battery electrodes.

- 21 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

Figure 1-11: Impedance model used to model EIS response

The complex impedance 𝑍, of the impedance model in Figure 1-11 changes with frequency
1⁄
𝛼
𝜔 according to Equation 1-7, where the time constant 𝜏 = (𝑅𝑝 𝑄𝑝 ) .

𝑅𝑝 1
𝑍(𝜔) = 𝑖𝜔𝐿 + 𝑅𝑠 + + (1-7)
1+𝑅𝑝 𝑄𝑝 (𝑖𝜔𝜏)𝛼 𝑄𝑑 (𝑖𝜔)𝛽

A special case of the CPE is the Warburg element given by Equation 1-8. This equation is

of a similar form as the impedance for a CPE, but with 𝑎 = 0.5, resulting in a line at a 45°

angle on the Nyquist plot.

1
𝑍𝑤𝑏 = 𝑊(𝑖𝜔)0.5 (1-8)

As batteries age, the SEI layer increases and starts to contribute significantly to the battery

impedance at frequencies around 0.5kHz to 0.1kHz. To model the SEI layer impedance, a

second ZARC element can be used or a traditional RC element.

Impedance models are frequently used throughout this thesis to model EIS data. In

particular, impedance models proved to be suitable to capture battery relaxation effects, as

shown in Chapters 2, 3, and 4. Chapter 3 employs a useful property of RC models to

validate EIS data. This is discussed in the next section.

- 22 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

1.3.2 KRAMERS-KRONIG TRANSFORM AND THE VOIGT-CIRCUIT

As was stated in the previous section, an 𝑛𝑡ℎ order ECM such as shown in Figure 1-9 (also

known as a Voigt-Circuit) can still fit to EIS results with low error if 𝑛 is increased

sufficiently. A near perfect fit can be obtained with 𝑛 = 20 or higher, but the model

becomes impractical and meaningless. However, the Voigt-Circuit model can still be used

as part of a validation procedure for EIS results called the Kramers-Kronig (KK) transform.

The KK transform makes use of the fact that the real components 𝑍′ of the EIS results can

be obtained from only the imaginary components 𝑍′′ (and vise versa) using Equation 1-9

to calculate 𝑍′′ and Equation 1-10 to calculate 𝑍′ at signal frequency 𝜔.

2𝜔 ∞ 𝑍 ′ (𝑥)−𝑍 ′ (𝜔)
𝑍 ′′ (𝜔) = − ( ) ∫0 𝑑𝑥 (1-9)
𝜋 𝑥 2−𝜔 2
2 ∞ 𝑥𝑍 ′′ (𝑥)−𝜔𝑍 ′′ (𝜔)
𝑍 ′ (𝜔) = 𝑍 ′ (∞) + 𝜋 ∫0 𝑑𝑥 (1-10)
𝑥 2−𝜔 2

The KK transform itself is difficult or impossible to implement for real datasets due to the

infinite limits of the integrals. However, any linear circuit is KK-transformable, which

means that if a dataset can be approximated using a linear circuit, such as the Voigt circuit,

the dataset is also KK-transformable [31]. This means that if a good fit to the Voigt-Circuit

can be obtained, EIS data can be shown to be KK-transformable, and, therefore, valid. This

method is used in Chapter 3 to ensure the validity of EIS results. Furthermore, the method

is extended such that it can be used to filter out noise from the EIS data.

1.3.3 OPTIMIZATION FOR PARAMETERIZATION

Impedance models are parameterized by optimizing an objective function such as the mean

square error (MSE) shown in Equation 1-11:

- 23 -
 Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

∑𝑛 ̂𝑖 )2
𝑖=1(𝑦𝑖 −𝑦
𝑀𝑆𝐸 = (1-11)
𝑛

where 𝑦𝑖 is a measurement data point, 𝑦̂𝑖 is the model prediction, and 𝑛 is the number of

data points. For impedance models with equations in frequency domain, the most common

optimization method used to minimize the MSE is the Levenberg-Marquardt (damped

least-squares) algorithm [22,44,45]. The genetic algorithm (GA) is frequently used for time

domain optimization [46–48]. The particle swarm optimization (PSO) algorithm is used by

some for time domain battery model parameterization [7,40]. The PSO algorithm is also

used for EIS model fitting in the work presented in this thesis. Chapter 4 includes a set of

upper and lower parameter bounds suitable to fit EIS data obtained from Samsung

INR2170-50E cylindrical battery cells using the PSO algorithm.

1.4 BATTERY STATE ESTIMATION

The State of Charge (SoC) and State of Health (SoH) of a battery are fundamental quantities

which must be estimated for safe and efficient battery operation. Other states, such as State

of Power (SoP), which indicates how much power can safely be drawn from a battery, are

derived from these states. SoP is usually calculated by taking into account the predicted

battery voltage after applying current over a small time-delta of 1 to 20 seconds [7].

Farmann et al. [49] reviewed SoP methods and grouped them into two categories:

characteristic maps (CM) and equivalent circuit models (ECM). CMs are maps which store

the power value that can be drawn from a battery under different SoCs, temperatures and

currents. ECM based methods are used in the same way for SoC estimation which is

discussed in the next section.

- 24 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

1.4.1 STATE OF CHARGE

The most important information a battery user requires is how much longer the battery can

be used until it must be recharged. This depends on the application specific battery usage

and on the amount of charge remaining inside the battery. State of charge is a quantity to

indicate the remaining charge of a battery, defined as the percentage of charge available

with respect to the maximum available charge. SoC is also used to keep the battery within

a safe operating range during operation. Unfortunately, SoC cannot be measured directly,

but must instead be estimated from measurable quantities such as voltage, current, and

temperature. SoC estimation can be grouped into three categories: direct methods, model-

based methods, and data-driven methods.

Battery current (𝐼) can be used directly to calculate SoC from the charge added to, or

removed from the battery. The known initial charge (𝐶𝑖 ) is required as a starting point from

which the accumulated charge over time is integrated. The result is divided by the

maximum available capacity (𝐶𝑛 ) as shown in Equation 1-12 [38]. This method is

frequently referred to as coulomb or amp-hour counting.

∫ 𝐼𝑑𝑡+𝐶𝑖
𝑆𝑜𝐶 = 1 − (1-12)
𝐶𝑛

The accuracy of the coulomb counting method depends on the accuracy of the current and

time measurements since measurement errors are amplified during the integration. In

addition, the required starting reference charge (𝐶𝑖 ) will drift over time [50] because of the

measurement errors.

- 25 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

OCV can be used to estimated SoC by using a battery characteristic OCV-SoC relationships

which must be determined offline beforehand. Since true OCV rarely occurs during normal

battery operation, battery models are used to estimate OCV under a given load. The

accuracy of the OCV-SoC method depends on the accuracy of the OCV-SoC data, the

accuracy of the battery model, as well as voltage and current measurement accuracies [51].

Kalman filters (KF) are used to achieve optimal SoC estimation in the presence of

modelling, measurement, and process noise [52]. Usually, battery model equations are non-

linear, but can be linearized and used in the Extended Kalman Filter (EKF). The EKF is

described in detail in Chapter 6.2.2. Sepasi et al. [53] use the EKF with a 2nd-order RC

model to estimate SoC for new and aged batteries, by updating the model parameters

online. The linearization in the EKF can reduce the accuracy of the estimator, especially if

the modelling and measurement equations are highly non-linear. Therefore, improvements

were made to the EKF to retain estimator accuracy with non-linear equations. Pan et al.

[54] used a 1st-oder RC model with an EKF and a grey model to avoid linearization. A

fractional order model was used with an unscented Kalman filter (UKF) by Mu et al. [46]

to find more accurate estimates using the unscented transform. The accuracy of the EKF or

UKF is highly dependant on the accuracy of the system model. Since it is often hard to find

accurate models for real applications, efforts have been made to improve estimation

robustness for inaccurate models. The Smooth Variable Structure Filter (SVSF) is a robust

filtering method, described in detail in Chapter 6.2.3. The SVSF was used by Afshari et al.

[40] and Ahmed et al. [47] with a 3rd-order RC model and electrochemical model,

respectively. With the SVSF, a boundary can be defined for uncertainties and noise levels
- 26 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

of the system which ensures convergence of the estimate to a value within the boundary. In

Chapter 6, the EKF and the SVSF are used together with the Interacting Multiple Model

(IMM) filter for low temperature SoC estimation. The IMM is used in cases where several

models are possible, but the correct one is unknown (see Chapter 6.2.4). The IMM uses the

likelihood of each model to compute a blended state estimate. EKF, SVSF, and IMM are

also discussed in detail in Chapter 6.

Data driven methods use advanced machine learning techniques to predict SoC from a

given usage profile. The accuracy of such models depends on the amount, diversity, and

quality of the data used to train the model. Tong et al. [55] used three neural networks to

predict SoC for idle, charge, and discharge conditions. The networks were trained with

drive cycles and validated with discharge pulses. Chemali et al. [56] used deep neural

networks to estimate SoC at different temperatures. Neural networks are introduced in

detail in Section 1.6 and applied to SoC estimation from EIS data in Chapter 5.

1.4.2 STATE OF HEALTH

As discussed in the previous section, SoC is an important quantity which must be estimated.

However, as the battery ages, SoC estimation algorithms must adapt to account for changes

in the battery behaviour. This can be accomplished if the state of health is known. As was

the case with the SoC, SoH is a hidden quantity which cannot be measured directly. Similar

to SoC estimation, SoH estimation can be grouped into categories: direct measurement

methods, indirect analysis methods, adaptive filtering methods, data driven methods, and

physical modelling methods. Most of these are reviewed by Xiong et al. [57]. Li et al. [58]

- 27 -
 Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

developed a single particle electrochemical model which can predict SEI growth and crack

propagation at different operating temperatures and track SoH this way. Pastor-Fernandez

et al. [59] compared results from EIS and IC-DV (incremental capacity and differential

voltage) data to identify degradation mechanisms. To overcome barriers of using EIS in

real time applications, Mingant et al. [60] proposed a quasi-electrochemical impedance

spectrum which requires less precise equipment. Tian et al. [61] used a fractional order

impedance model in combination with incremental capacity analysis to account for SoH.

The SVSF used by Afshari et al. [40] for SoC estimation can estimate SoH from a chattering

parameter of the filter. Li et al. [62] applied a Gaussian filter to incremental capacity

analysis to identify features of interest for SoH estimation. Yang et al. [63] identified time

constants during the constant voltage (CV) charge region and related them to the SoH of

the battery. Eddahech et al. [64] also investigated the CV region during charge, but related

it to degradation mechanisms during calendar aging. Hu et al. [65] used HPPC tests, sparse

Bayesian predictive modelling, and the concept of statistical sample entropy to model the

change in the pulse response with SoH. Further SoH estimation techniques are reviewed in

Chapter 4, where the battery relaxation effect (Section 1.1.3) is combined with EIS (Section

1.2) to estimate SoH.

1.5 FRACTIONAL ORDER CALCULUS

The impedance models introduced in Section 1.3.1 can often achieve higher voltage

modeling accuracy. However, to use the impedance model for state estimation in a similar

way as ECMs (for example, with Kalman filters), the model equation must be in time

- 28 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

domain. Translating a complex impedance equation such as (1-7) into time domain is a

non-trivial task. One approach is to simply approximate non-ideal elements such as the

CPE using a string of RC-pairs as shown in Figure 1-12. However, this adds a lot of

additional model parameters which is undesirable. A relatively new approach uses

fractional order calculus (FOC) [44,46,61,66–68] instead to convert fractional order

impedance elements into time domain, also shown in Figure 1-12.

Figure 1-12: Representation of CPE element with RC approximation or FOC.

FOC involves the use of the fractional order operator 𝐷𝑡𝛼 which is defined by Equation 1-

13, for a time domain measurement 𝑥𝑘 at discrete timestep 𝑘 with sample rate 𝑇𝑠 . Here, 𝛼

is the fractional order exponent from the CPE impedance, and (𝛼𝑗) is the binomial

coefficient calculated using Equation 1-14 by making use of the Gamma function Γ.

1
𝐷𝑡𝛼 𝑥𝑘 = 𝑇 𝛼 ∑𝑘𝑗=0(−1)𝑗 (𝛼𝑗) 𝑥𝑘−𝑗 (1-13)
𝑠
𝛼! Γ(α+1)
= Γ(𝑗+1)Γ(𝛼−𝑗+1)
(𝛼𝑗) = {𝑗!(𝛼−𝑗)! (1-14)
1
Figure 1-13 shows impedance models which have been converted for time domain use.

Mauracher et al. [69] transformed the Warburg impedance in the model shown in Figure

- 29 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

1-13c into time domain by first approximating the element as a string of RC pairs and

assuming each RC pair to have equal values. Wang et al. [70] as well as De Sutter et al.

[42] used FOC to convert the impedance of a CPE to time domain and parameterized the

model shown in Figure 1-13a using time domain discharge pulses. Eddine et al. [67]

translated the Warburg impedance into time domain using FOC for the model shown in

Figure 1-13b, parameterized the model using time domain discharge pulses and validated

the results using EIS data. Ideally, EIS should also be used to parameterize impedance

models as it provides the most complete impedance profile. To obtain the model parameters

for all discharge currents, EIS can be collected under a DC bias current which is possible

only for lower currents. At higher currents, the battery would discharge significantly over

the duration of the EIS measurement, changing the result, or fully discharging the battery.

Xu et al. [43] parameterized a Warburg element with EIS data in frequency domain and

used FOC to convert the impedance to time domain. This was possible because the portion

of the Nyquist plot modelled by the Warburg element does not change significantly with

discharge current. Kollmeyer et al. [71] used time domain pulses to obtain ohmic resistance

values under different discharge currents and used the results to scale EIS data. They then

used the scaled EIS and the RC pair conversion method to parameterize the model shown

Figure 1-13c.

- 30 -
 Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

Figure 1-13: Common impedance models

1.6 DEEP NEURAL NETWORKS

Machine learning techniques are used in a large variety of applications and battery state

estimation is one of them. In Chapter 5 deep neural networks are used for estimation of

SoC from EIS data and from battery models parameterized with EIS data. This section

serves as a general introduction to neural networks and deep neural networks.

The concept of Artificial Neural Networks (ANN) is based on early models of human brain

function. A basic ANN consists of three layers, an input layer a hidden layer and an output

layer. The number of elements of the input layer and output layer depend on the application

and available data. For example, in the case of binary classification from a set of n

measurements, the input layer would have n elements, one for each measurement, and the
- 31 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

output layer would have two elements, one for each class. The hidden layer consists of one

or more neurons, the basic building block of ANNs shown in Figure 1-14. A neuron 𝑗 of

hidden layer 𝐿 takes in inputs 𝛼0𝐿−1 from the previous layer (for example the input layer) to

compute an output, or activation, 𝛼𝑗𝐿 via an activation function 𝜎(𝑧 𝐿 ) from the weighted

input 𝑧 𝐿 . The weighted input 𝑧 𝐿 is computed from weights 𝑤𝑖 and biases 𝑏𝑖 using Equation

1-15.

𝑧 𝐿 = ∑𝑛𝑖=0 𝑤𝑖 𝑎𝑖𝐿−1 + 𝑏𝑖 (1-15)

A step function could be used for the activation function 𝜎 to achieve threshold-like

behaviour of the neuron. However, the fact that the step function is discontinuous presents

a challenge for training procedures introduced later in this section. Instead, the Sigmoid

function is used, a smooth, differentiable version of the step function shown in Equation 1-

16.

1
𝜎 (𝑧 𝐿 ) = 𝐿 (1-16)
1+𝑒 −𝑍

Figure 1-14: Neuron

Several neurons can be combined to form a network structure as shown in Figure 1-15.

Neural networks with large number of layers and neurons have proven to be capable of

- 32 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

solving complex problems and referred to as a Deep Neural Networks (DNN). By tuning

the weights and biases, the network can be used to, for example, perform classification. A

given set of inputs will propagate through the network and produce a probability of the

inputs belonging to a certain class, based on the activation of the different neurons and the

interaction between them [72].

Figure 1-15: Simple Neural Network

Tuning of the weights and biases requires an optimization algorithm. For optimization

problems in general, an algorithm should modify the tuning parameters and then evaluate

the performance of the system to be optimized by minimizing a const function 𝐶. In the

case of multi-layer neural networks, a complication arises from the fact that the tuning

parameters of the hidden layers are not directly connected to the output, which means the

desired output values of the hidden layers are unknown. However, the cost evaluated at the

output layer can be backpropagated through the hidden layers. Therefore, neural networks

are optimized, or trained, using the backpropagation algorithm with the gradient descent

optimization method. The method works as follows [73]. In general, the error 𝛿 𝐿 at the

output of layer 𝐿 can be calculated using Equation 1-17, where ∇𝑎 𝐶 is the gradient of the

- 33 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

cost function, ⨀ denotes the Hadamard product (element wise vector product), 𝜎′ the

derivative of the activation function and 𝑧 𝐿 the weighted input at layer 𝐿 (Equation 1-17).

𝛿 𝐿 = ∇𝑎 𝐶⨀ 𝜎 ′ (𝑧 𝐿 ) (1-17)
The error 𝛿 𝐿 can then be related to the error of the previous layer 𝛿 𝐿−1 using Equation 1-

18, where (𝑤 𝐿 )𝑇 is the transpose of the weights of layer 𝐿.

𝛿 𝐿−1 = ((𝑤 𝐿 )𝑇 𝛿 𝐿 )⨀ 𝜎′(𝑧 𝐿−1 ) (1-18)

Finally, the gradient at each layer can be calculated using Equation 1-19, where 𝑤 is the

weight of a neuron with error 𝛿𝑜𝑢𝑡 and 𝑎𝑖𝑛 the activation of the previous neuron. 𝑏 is the

bias of the perceptron, and 𝛿 is the error evaluated at the same neuron as the bias.

𝜕𝐶 𝜕𝐶
= 𝑎𝑖𝑛 𝛿𝑜𝑢𝑡 , =𝛿 (1-19)
𝜕𝑤 𝜕𝑏

With the equations introduced above, the gradient between any two nodes connected

between adjacent layers can be calculated and used in the gradient descent algorithm to find

the direction of possible minima of the cost function 𝐶.The gradient descent method itself

includes a tuning parameter 𝜂 called the learning rate, which is related to the step size the

algorithm takes in its search of solutions. Parameters other than the weights and biases of

the network are referred to as hyper parameters and must be carefully chosen. In the case

of the learning rate, 𝜂 must be selected to be large enough such that the gradient descent

algorithm learns fast enough, but not too high for the algorithm to overshoot solutions.

The data used to train the DNN should be split into training data and validation data as a

basic way to protect against overfitting. During training, the inputs of the whole training

set are fed to the network and the cost function is used to compare the network output

against a ground truth. This repeated until a set number of cycles, or Epochs, are completed.
- 34 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

The number of epochs can potentially be large, especially if the network starts with poor

initialization of parameters. Several methods exist to improve the performance of DNN

training. The selection of the cost function plays an important role in the speed with which

the training converges to a minimum. The quadratic cost function shown in Equation 1-20

for example may be a desired choice, where 𝑛 is the number of training samples, 𝑥 is the

training sample, 𝑦(𝑥) is the ground truth evaluation given 𝑥 and 𝛼 𝐿 (𝑥) is the DNN output

given 𝑥.

1 2
𝐶 = 2𝑛 ∑𝑥‖𝑦(𝑥 ) − 𝛼 𝐿 (𝑥 )‖ (1-20)
However, the partial derivates of the quadratic cost function are proportional to the derivate

of the activation function 𝜎′, which becomes small far away from the threshold. As a result,

learning is slow for values far from the solution. The speed of learning can be improved by

using the cross-entropy cost function shown in Equation 1-21 instead.

1
𝐶 = − 𝑛 ∑𝑥[𝑦(𝑥) ln(𝛼 (𝑥 )) + (1 − 𝑦) ln(1 − 𝛼(𝑥 ))] (1-21)
The partial derivates of the cross-entropy function turn out to be proportional to the

activation function 𝜎, instead of its derivative. This means that a badly initialized DNN can

still produce large enough gradients to learn faster.

Several other modifications and hyper parameters exist and those relevant to the results

shown in Chapter 5 are introduced next:

Stochastic Gradient Descent: The speed of the gradient descent algorithm can be

improved by randomly sampling the dataset into small batches and using the batches for

training instead of the entire dataset. The gradients computed from the batches serve as

- 35 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

close approximations of the gradients of the whole dataset but take less time to compute.

This method is called Stochastic Gradient Descent (SGD). In addition to the learning rate,

SGD requires the size of the batches as another hyper parameter.

L2 Regularization: Increasing the number of neurons and layers can improve the

performance of the DNN. However, if training sets are limited in size, the DNN is likely to

overfit to the training data as the number of parameters grow. To protect against overfitting,

L2 regularization can be used. This method adds a weighting factor to the cost function as

shown in Equation 1-22. Here, the first half of the equation is the regular quadratic cost

function, and the second term is the L2 regularization factor. This factor is the sum of

𝜆
squares of the weights of the network scaled by where 𝑛 is the size of the training set
2𝑛

and 𝜆 is a new hyper parameter.

1 2 𝜆
𝐶 = 2𝑛 ∑𝑥‖𝑦(𝑥 ) − 𝛼 𝐿 (𝑥 )‖ + 2𝑛 ∑𝑤 𝑤 2 (1-22)
Using L2 regularization forces the learned weights to be small, only allowing bigger

weights if significant C reduction is achieved. Avoiding large differences between weights

promotes generalization of the network.

Pruning: The large number of model parameters required to run a completed DNN can be

impractical of constrained embedded applications. However, not all model parameters are

equally important. In fact, a large number of neurons can be removed without significantly

reducing the network performance. Pruning works by choosing a threshold and removing

all neurons with weights below this threshold. This process can be repeated until the highest

network performance with the desired number of parameters is achieved.

- 36 -
 Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

1.7 BATTERY TESTING & AGING

In this section, some of the battery testing methods used throughout this thesis are

introduced. The following tests were used:

1) Break-in Procedure

The first time a battery cell is tested, it needs to undergo a break-in procedure consisting of

several charge and discharge cycles to stabilize the batteries’ response. All batteries used

in the experimental parts of this thesis were first cycled 10 times for this reason.

2) Capacity Test

The capacity test determines the SoH of the battery by coulomb counting during discharge.

For this test, the battery is first fully charged according to manufacturer specifications,

followed by a rest, followed by a complete discharge during which the coulomb counting

is performed. The total coulombs counted during the discharge step (assuming an initial

charge of zero) are the measured capacity of the cell.

3) OCV-SoC Test

The OCV-SoC test provides the data used for the OCV-SoC lookup tables for state

estimation algorithms. For this test, the battery is first fully charged according to

manufacturer specifications. After a rest period, the battery is discharged at very low

current until the lower cut-off voltage is reached. Following another rest, the battery is

charged at the same low current until the upper voltage cut-off is reached. Since low current

is used for this test, the voltage measured during discharge and charge is close to OCV.

Lower current will result in more accurate OCV measurements, but the test duration will
- 37 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

increase. Figure 1-16 shows example OCV-SoC datasets for a new (100% SoH) and an

aged (80% SoH) Samsung INR2170-50E cylindrical battery cell. The shape of the OCV-

SoC curve depends on the type of battery used. The cylindrical NMC batteries used in this

study show a relatively steep OCV-SoC curve slope. Other chemistries can exhibit OCV-

SoC curves with a nearly horizontal section between 80% and 20% SoC. Obtaining a

distinct SoC value from OCV becomes less accurate for near horizontal parts of the OCV-

SoC curve. The steeper slopes shown in Figure 1-16, on the other hand, allow for more

accurate readings since the points on the curve are more clearly separated. As the battery

ages, the OCV-SoC behavior changes slightly, mostly at high SoC (>80%) and at low SoC

(<20%). In the middle region the results are similar for fresh and aged cells. Also shown in

Figure 1-16 are polynomial fits for both curves. Polynomial orders of 10 th or higher are

typically needed to fit OCV-SoC curves. The OCV-SoC test procedure is described in more

detail in Chapter 6, where it is used as part of SoC estimation algorithms.

Figure 1-16: Experimental OCV-SoC curve results for new and aged batteries.

- 38 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

4) EIS Test

EIS provides information about the battery impedance behaviour and can be used to

parameterize impedance models. The EIS test is performed using a potentiostat. Since this

is one of the most used tests in this study, the test details are described repeatably in method

sections of Chapters 2, 3, 4, and 7.

5) Charge Discharge aging

The most common form of aging is cycle aging where batteries are continuously charged

and discharged to full charge and full discharge depth. While this method can be relatively

fast, the way the battery ages is different from realistic conditions in EV applications. As a

result, battery behaviour after cycle aging may be different from its behaviour after field

operation, particularly when considering impedance profiles. The cycle aging procedure is

described in detail in Chapter 7 where it is compared to realistic drive cycle aging.

6) Drive cycle aging

Realistic battery aging is a time and resource intensive process by which battery cells are

artificially aged in controlled laboratory settings. Instead of full charge and discharge

cycles, standardized drive cycles are used to age batteries. These drive cycles are provided

for example by the US Environmental Protection Agency (EPA) and consist of speed vs.

time data for different driving scenarios. To be able to use EPA drive cycles in battery

aging, the speed vs. time data must be converted to battery load demand. This can be

achieved using a simple vehicle model as shown in Figure 1-17.

- 39 -
 Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

Figure 1-17: Simple Vehicle Model

This model considers forces acting on a vehicle such as drag due to air (𝐹𝐷𝑟𝑎𝑔 ) and rolling

resistance (𝐹𝑟𝑜𝑙𝑙 ) from the tires on the road (assuming flat roads, zero road angle). These

forces must be overcome to achieve the speeds demanded by the EPA drive cycle, using

the power provided by the battery. The battery losses are approximated with a constant

efficiency number (𝜂𝑏𝑎𝑡𝑡 ). Further losses due to drivetrain (transmission and motor)

inefficiencies are approximated by 𝜂𝑑𝑟 . While not suitable for exact range simulations, the

simple model can provide representative load profiles to use for battery aging. The details

of this model including model parameters and equations are provided in Chapter 7. Chapter

7 also describes an aging study which compares charge/discharge cycle aging with realistic

drive cycle aging.

1.8 CUSTOM BATTERY CELL TESTER

The research reported in several parts of this thesis required experimental data from

lithium-ion batteries under various operating conditions. A custom battery test system was

designed to perform different types of battery experiments. The design of the test system
- 40 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

was subject to the following three key requirements: 1) the voltage and current sensing

must be accurate enough to achieve an amp-hour counting accuracy of less then 0.5 %, 2)

all experiments shall run autonomously while ensuring battery operation stays within safe

limits, and 3) the system shall be open to extension of functionality. Figure 1-18 shows an

overview of the components and connections of the test system and Table 1-1 lists each of

the devices used.

Figure 1-18: Custom battery test system diagram (only one channel shown)

Any Windows 7 (or higher) PC can be used to run the system, so no PC is specified. The

PC hosts custom software to control the devices used to perform various battery tests. To

charge and discharge batteries a power supply and a load were used, respectively. Both

power supply and load support 4 individual channels, allowing the system to test four

individual cells simultaneously. The load is controlled via RS232 serial communication,

and the power supply via Ethernet. The battery voltage is measured using the NI9239, 4

channel analog input module which is housed in the National Instruments Data Acquisition

- 41 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

(NIDAQ) chassis. Battery current is measured via the voltage drop across a shunt resistor.

This voltage drop is measured by the NI9219, 4 channel analog input model, which is also

housed in the NIDAQ chassis, and controlled via the Universal Serial Bus (USB)

connection. The Ethernet controlled potentiostat used in this work has only a single

channel, however, a custom designed multiplexer allows for automated switching of the

potentiostat between each of the four channels. Four mechanical relays are used to switch

each of the two load lines (positive and negative) as well as each of the two sense lines

from the potentiostat to a battery adapter.

Table 1-1: Custom battery test system parts list

Name Manufacturer Model Description

Potentiostat BioLogic SP150 1 channel
Load Agilent N3306A 4 channels, 20V/15A/300W
Power Supply Agilent N6773A 4 channels, 60V/60A/300W
National
NIDAQ cDAQ-9178 8 slot chassis
Instruments
National 4 channel NIDAQ module to
Analog Input NI9239
Instruments sense battery voltage
National 4 channel NIDAQ module to
Analog Input NI9219
Instruments sense shunt voltage
Thermal Thermotron T8200 -10°𝐶 to +40°𝐶
Chamber
Micro Multiplexer control and thermistor
Arduino Uno
Controller sensing
RSN-100- 1𝑚Ω resistance for current
Current Shunt Riedon 100B measurement
Thermistor Semitec 103AT-11 One thermistor per battery
Multiplexes the potentiostat to 4
Multiplexer Cadex Electronics Custom
channels
Battery With Kelvin connection (Section
Cadex Electronics Custom
Adapter 1.8.1)
Figure 1-19 shows a diagram of the multiplexer illustrating one active channel and one

disconnected channel (grey connections). Each set of four relays is controlled by an

- 42 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

Arduino microcontroller, through its Digital Inputs/Outputs (DIO). The same

microcontroller also measures the battery temperature via thermistors. The microcontroller

is attached to the PC via USB. A thermal chamber houses each of the four battery adapters

and is able to maintain a stable environment. The thermal chamber is controlled via

Ethernet. The battery adapters are discussed in detail in Section 1.8.1.

Figure 1-19: Multiplexer diagram for 2 channels

The different parts of the system are orchestrated by custom software running on a host PC.

The software was developed in Python 3.7 with an architecture as shown in Figure 1-20.

- 43 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

Figure 1-20: Custom battery test system software architecture

The software consists of two separate processes, several threads, and many classes.

Processes are isolated from each other since each uses its own Python interpreter instance.

Threads run code in parallel within the same processes. Classes contain groups of methods

to perform the different functions of the system.

A watchdog program, Figure 1-20 block (1) runs independently of the main program as a

separate Python process such that the watchdog can shut-down and re-start the test system

in case of malfunction, or if safety parameters are exceeded. The watchdog starts the

Drivecycle aging system process (2), which opens up the Graphical User Interface (GUI)

(16) shown in Figure 1-21. Pressing the “Connect” button starts a thread to establish

communication with the different hardware devices (3). Next, the channel worker (4) is

- 44 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

stated which contains four threads, one for each channel. The scripts class (5) contains

functions which represent different experiments which can be performed using this system.

Pressing a “Start” button next to a channel row on the GUI (Figure 1-21) will prompt the

user to select from a list of experiments as well as select a battery ID.

Figure 1-21: Graphical user interface for custom battery test system

Battery IDs must first be registered in the underlying Microsoft SQL Server relational

Database (DB), which is controlled by the DB handler (17). After entering the required

information, the experiment is started via the event handler (14). The event handler will

also stop the experiment if a stop button is pressed. While an experiment is running,

different devices will be used by the scripts class (5) via the device handler class (6). The

simplest devices to control are the thermal chamber (simple read/set of temperature) and

the Arduino microcontroller (set multiplexer channel/read thermistor temperature). To be

able to run custom current profiles such as drivecycles, the load (12) and power supply (13)

must be controlled via its own thread, the cycler worker (10). Cycler refers to the

- 45 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

combination of load and power supply. The Potentiostat is not directly controlled by Python

code, instead a C# API library is used. This is because well designed C# example code was

already available from the manufacturer of the potentiostat. Fortunately, C# code exported

as a Dynamically Linked Library (DLL) file can be integrated into Python via the Pythonnet

library. A similar approach was used for the NIDAQ, where the direct control of the

NIDAQ is also performed in C#. This was done due to the high sampling rates required by

the NIDAQ modules, which are too fast for Python to handle. Via the C# layer,

measurements can be buffered and down sampled to be made available to Python at a much

slower rate. The data is sampled periodically, requiring another thread, the NIDAQ worker

(9). The data is also independently monitored by the data observer (11).

During the execution of experiments, the GUI displays select measurements, such as

voltage, current, temperature, charge (in Ampere-hours), and aging cycles (if applicable,

depending on selected test). The active test step is also displayed along with the battery ID

for each channel. Data and events are communicated to and from the GUI to the rest of the

system via a publisher/subscriber signal architecture, facilitated by the signal handler class

(4). A subscriber (for example a part of the GUI) can be attached to a publisher (for example

the data observer) without the need for the two parts to be aware of each other. The

publisher does not know which are all the subscribers, it simply sends out its message. This

is illustrated in Figure 1-20 via the dotted lines creating weak connections between the GUI

and the rest of the system. This publisher/subscriber architecture is a common Object-

Oriented (OO) design pattern. Another OO design patter which was used is the Model-

View-Controller (MVC) pattern. The MVC pattern further simplifies the decoupling of
- 46 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

GUI (front-end) and “system” (back-end) by sharing a single “data model” object between

front-end and back-end. The data model is the data carrier where experiment data, shared

settings, as well as the back-end status is stored. The publisher signals are then used to

notify different parts of the system of relevant changes to the data model.

Finally, the data collected from the experiments is fed into the DB, again via the DB handler

(17). The DB architecture is shown in Figure 1-22. Batteries are stored in the “Batteries”

table, each with a reference (foreign key) to a battery model from the “BatteryModels”

table which in turn references a manufacturer from the “Manufacturers” table. When a new

experiment is started, a new entry into the “MasterExperiments” table is created using the

battery ID provided by the user via the GUI. Each experiment can contain one or more

steps, which are recorded separately in the “Experiments” table. The data obtained from

the experiments comes in two forms, time domain data (voltage, current, temperature, etc)

and EIS results (real, imaginary, frequency). The “ExperimentLogs” table contains the time

domain data and the “EISDatas” table the EIS data. Since many different drive cycles are

frequently used in this work, they are also stored in this DB in the “DriveCycles” and

“DriveCycleDatas” tables. Specific drive cycles can also be linked to specific experiments

via the “ExperimentDriveCycleLinks” table. The data stored in this DB can be accessed

directly from MATLAB, where scripts can be written to automate data retrieval. The

system described in this section enabled accurate, repeatable, time efficient and reliable

battery experiments, and was instrumental in obtaining the various results presented in this

thesis.

- 47 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

Figure 1-22: Database architecture for custom battery test system

1.8.1 BATTERY ADAPTER

The test system includes a custom adapter to attach batteries to the devices using four-point

measurement (Kelvin connection). A Kelvin connection separates current carrying (load)

wires from sense wires. If a sense wire shares the same conductor as a load wire, the

measured voltage includes the battery voltage as well as an additional voltage drop across

- 48 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

the shared conductor. This additional voltage drop also varies with current, which makes it

difficult to predict for calibration. The adapter shown in Figure 1-23 implements a Kelvin

connection for a cylindrical battery cell to avoid any additional voltage drop. Nickel tabs

welded to the positive and negative terminals of the battery extend outwards to allow

connections on both ends. The adapter also includes screw terminal blocks to connect larger

load cables from the load and power supply. The side with the terminal blocks is the load

carrying side, and the opposite side is for sense lines. The adapter base was machined from

a block of ABS plastic.

Figure 1-23: Custom battery adapter for custom battery test system. 1 – battery, 2 –
adapter base, 3 – screw terminal connector, 4 – nickel tab.

Figure 1-24 shows a 3D printed welding fixture which was used to align the tabs on both

sides of the battery during the welding processes. The figure shows the fixture with the top

of a battery visible in the center, and a nickel tab placed in a slot on top. Once both tabs are

welded on the top and bottom of the battery, the tabs can be gently folded up and the battery

can be extracted using the clearance slot. Without this fixture, alignment cannot be

- 49 -
 Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

guaranteed. If the welded tabs are not parallel to each other, the battery cannot be properly

attached to the adapter.

Figure 1-24: Battery welding fixture

1.9 RESEARCH CONTRIBUTIONS

In this section, the hypotheses which formed the basis for this thesis are stated, an overview

of the project contributions is provided, and primary and secondary contributions are

presented.

1.9.1 HYPOTHESES

The following hypotheses formed the basis for the results shown in this thesis:

Hypothesis 1: EIS Measurement duration can be shortened by reducing the rest time
requirement through understanding the battery relaxation effect.

EIS is a powerful technique with a lot of potential to significantly improve diagnostics in

BMSs. However, the long rest times required to ensure stable conditions create a barrier to

- 50 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

the successful implementation of EIS for real time, EV applications. Rest times are usually

required to avoid measurement drift and other instabilities. Some methods exist which can

compensate for measurement drift, however, the impact of shorter rest times on EIS

measurements under different battery operating conditions is not well understood.

Hypothesis 2: The battery relaxation effect contains information about the battery
SoH which can be extracted using EIS.

The battery relaxation effect is known to correlate well with SoH. However, proposed SoH

estimation methods based on the relaxation effect require fully charged or fully discharged

batteries or other conditions difficult to achieve in real time applications. EIS can be used

to rapidly characterize the relaxation effect and estimate the SoH in the process.

Hypothesis 3: The IMM can be used to improve SoC estimation at low temperatures
by considering C-rate specific models.

The IMM is a useful filter to blend several independent system models into a unified state

estimate. At room temperature or above, a single ECM can accurately model the response

of a battery independent of current. However, at low temperatures, the battery response

changes as current increases, reducing the accuracy of the ECM. In the context of state

estimation, the updating of model parameters in real time is only possible for slow changing

conditions such as temperature, SoH and SoC. Current in EV applications can change

significantly within seconds, making state estimation un-observable. Instead, multiple

battery models each dedicated to specific regions of current magnitudes can be used and

unified with the IMM.

- 51 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

1.9.2 CONTRIBUTIONS

Figure 1-25 provides a graphical overview of the different parts of the thesis project. The

major parts of this project, which are published as journal papers, are highlighted with red.

Conference papers are marked with black. Finally, non-published contributions in the area

of battery testing are also included. Next, the primary and secondary contributions are

listed.

Figure 1-25: Contributions Overview

1.9.2.1 PRIMARY CONTRIBUTIONS

The primary contributions of this work are itemized and briefly described in this section.

- 52 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

1. Characterized the short-term battery relaxation effect to reduce EIS rest time

requirements and shorten testing times (Chapter 3).

This first contribution is further to Hypothesis 1 and based on in-depth research into the

battery relaxation effect and its impact on EIS measurements. Several hours of rest are

usually required before EIS can be measured to simply avoid the rest-time effect. In Chapter

3 methods and results are presented to help understand the relaxation effect and to

significantly reduce the rest time requirement to 5 minutes under controlled conditions.

2. Used the battery relaxation effect together with EIS to develop a new method for SoH

determination (Chapter 4).

The battery relaxation effect changes as the battery ages. However, as shown in literature,

this property of the relaxation effect is difficult to capture with time-domain methods. In

Chapter 4 a method is proposed to use impedance modelling and EIS to capture the battery

SoH information contained within the relaxation effect. Further to Hypothesis 2, this

method can determine SoH within an error of 2%.

3. Developed an improved low temperature SoC estimation strategy using the IMM with

SVSF and C-rate specific battery models (Chapter 6).

Battery SoC estimation shows increased errors at low operating temperatures, due to the

increased non-linear response of the battery with current (or C-rate). In Chapter 6, a method

is proposed to blend three battery models, each specialized for a different C-rate range, into

a single SoC estimate. This is achieved through the used of the IMM-SVSF filter, which is

shown to have an SoC estimation error below 2%. This is in response to Hypothesis 3.
- 53 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

4. Characterized the long-term battery relaxation effect with EIS and showed its impact on

impedance model parameters (Chapter 2).

As part of the general research into the battery relaxation effect, the long-term relaxation

effect was characterized with EIS. In Chapter 2, the EIS results were modeled with an

impedance model to show how model parameters change with relaxation as well as SoC

and temperature.

5. Demonstrated SoC estimation with deep neural networks from frequency domain EIS

data as well as impedance model parameters (Chapter 5).

The large EIS dataset obtained as part of the research into the relaxation effect proved to

be complex due to its many dimensions (rest time, SoC, temperature). DNNs lend

themselves well to such complex data and were used in Chapter 5 to estimate SoC from

EIS data.

6. Conducted an aging study to compare accelerated aging to realistic drive cycle aging

using EIS to show differences in impedance characteristics at different SoH (Chapter 7).

The impedance characteristics of aged batteries are dependent on how they are aged. To

understand the impact of different aging pathways, an extensive aging study was

conducted. In Chapter 7 the results of this study, comparing charge/discharge aging to drive

cycle aging, are presented. The impedance characterization throughout the aging tests give

insight into the impact of each aging method.

- 54 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

1.9.2.2 SECONDARY CONTRIBUTIONS

A number of additional contributions are listed and briefly explained in this section.

1. Developed Voigt-circuit based filtering method to improve EIS data analysis (Chapter

3).

Reduced rest times prior to EIS measurement can introduce drift and noise into the EIS

data due to the battery relaxation effect. To smooth out EIS results, a Voigt-circuit was

used as a filter. The use of this filter allowed for consistent comparison of EIS

measurements as well as tracking of the relaxation effect. The filter is described in Chapter

3.

2. Compared the performance of the SVSF and EKF as part of the IMM filter for battery

SoC estimation applications (Chapter 6).

In Chapter 6 the IMM-EKF was shown to significantly improve SoC estimation at low

temperature. In addition, the IMM-SVSF was shown to be able to further reduce the SoC

estimation error due to its adaptive properties and robust estimation.

3. Developed simple EV model to convert speed vs. time driving schedules to drive cycle

current profiles (Chapter 7).

An EV model based on Tesla Model 3 was developed using basic vehicle dynamics and

constant powertrain efficiencies. The model is described in detail in Chapter 7 and used for

realistic battery aging.

4. Published case study on impedance model fitting from EIS data.

- 55 -
 Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

Methods for impedance model fitting to EIS measurements are not widely accessible.

Fundamentals and equations can be found in literature, but few tutorials and examples exist.

As part of this research project, a case study was made available online demonstrating

impedance model fitting with real EIS results using MATLAB as a framework. The case

study can be found online [32].

5. Designed and implemented a multi-channel battery test system (Chapter 1).

To perform accurate battery characterization as well as realistic drive cycle aging tests, a

custom battery test system was designed. This test system is capable of performing fully

automated tests on four cells at the same time. The detailed implementation of the test

system is described in Section 1.8.

1.10 THESIS OUTLINE

Chapter 2 provides a literature review on the long-term battery relaxation effect with focus

on the impact of the relaxation effect on battery modeling. Results are presented showing

how model parameters change during the relaxation effect for batteries at different SoCs

and temperatures.

Chapter 3 provides a literature review on characterization of the battery relaxation effect

with EIS. In addition, the technical challenges regarding EIS measurement and the

relaxation effect are discussed. A filtering strategy is proposed to obtain consistent EIS

measurements during the short-term relaxation effect. Finally, results are presented to show

how different conditions impact the relaxation effect and EIS measurement with short rest

times.
- 56 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

Chapter 4 provides a literature review on the use of the relaxation effect as well as the use

of EIS for SoH estimation. In this chapter a new method for SoH estimation is proposed

combining the relaxation effect and EIS measurements.

Chapter 5 provides a literature review on SoC estimation techniques with special focus on

machine learning methods. Furthermore, the chapter presents SoC estimation results for a

deep neural network trained on EIS data.

Chapter 6 provides a literature review on low temperature SoC estimation techniques and

presents the technical challenges related to current dependant battery models for SoC

estimation. The IMM-SVSF low temperature SoC estimation strategy is presented in this

chapter.

Chapter 7 presents and discusses results of the accelerated and realistic aging studies. The

vehicle model equations are also introduced in this chapter.

Chapter 8 provides the conclusions and discusses future work.

- 57 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

Chapter 2: Lithium-Ion Battery Relaxation Effects

Marvin Messing1,2, Tina Shoa2, Saeid Habibi1

1
Department of Mechanical Engineering, McMaster University, Hamilton, ON, Canada,
2
Cadex Electronics, Richmond, BC, Canada
This paper is published in 2019 IEEE Transportation Electrification Conference and
Expo (ITEC), 2019, pp. 1-6, doi: 10.1109/ITEC.2019.8790449. This paper is republished
here with permission2.

ABSTRACT

The accurate estimation of the state-of-charge (SoC) of lithium-ion batteries is crucial for

safely operating electric vehicles. One way to obtain information about SoC is to utilize

battery impedance profiles. Effects of temperature, SoC, and state-of-health (SoH) on

impedance have been studied using Electrochemical impedance spectroscopy (EIS) but the

effect of relaxation period following a charge or discharge cycle requires more attention.

In this study EIS results are obtained with respect to relaxation period at different SoCs and

temperatures. An impedance model is fit to the data and the change of model parameters

with relaxation is analyzed. The results show that the behaviour of the model parameters is

in good agreement with electrochemical theory. Furthermore, it is found that changes in

some model parameters are significant when compared to changes in SoC. This highlights

the need to account for the relaxation effect when measuring battery impedance.

2
In reference to IEEE copyrighted material which is used with permission in this thesis, the IEEE does not
endorse any of McMaster's products or services. Internal or personal use of this material is permitted. If
interested in reprinting/republishing IEEE copyrighted material for advertising or promotional purposes or
for creating new collective works for resale or redistribution, please go to
http://www.ieee.org/publications_standards/publications/rights/rights_link.html to learn how
to obtain a License from RightsLink.
- 58 -
 Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

2.1 INTRODUCTION

Lithium-ion batteries (LiB) are the current choice for many applications due to their energy

density, and long life-span [1]. However, charge and discharge rates, voltage levels, and

temperature have to be carefully managed to ensure the safe operation of LiBs [2]. The

quality of battery management depends on the accuracy of the estimation of battery states

such as state of charge (SoC), and state of health (SoH). These states have to be estimated

since they cannot be measured. Instead, battery voltage, current, and temperature are

measured, and used in estimation strategies [3]. The most important estimate is the SoC

which indicates how much energy is left in the battery. The SoC changes non-linearly as

the battery is discharged and depends on the discharge profile. All charge is depleted once

the lower voltage limit – which is determined by the manufacturer – is reached. One factor

determining the accuracy of SoC estimation is the SoH of the battery. The SoH affects the

maximum capacity that the battery can supply at any point in its life, relative to the initial

rated capacity. The SoH changes due to aging mechanisms inside the battery [4], which

depend on the operating conditions during the lifetime of the battery. Therefore, SoC must

be estimated using measurable signals such as voltage, current, and temperature, but also

by factoring in how the behaviour of the battery changes as a function of SoH.

The management of LiBs is particularly important for battery packs in electric vehicles

(EV). A wide variety of parameter estimation techniques are available for EV applications,

collectively covering most operating conditions and individually showing reasonable

tracking accuracy [5].

- 59 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

However, the challenge remains to find a more accurate, reliable, universal, and feasible

estimation strategy.

Battery states are also related to battery impedance. Battery impedance holds useful

information about the internal condition of the battery. Impedance is the combination of

reactance and resistance. In batteries, impedance comes from the interaction between

different materials, the material characteristics themselves, and chemical reactions [6].

Impedance can be modeled using fundamental electrical circuit components such as

resistors and capacitors. However, imperfect circuit elements must be used to increase the

modeling accuracy. Models containing such imperfect elements are referred to in this work

as impedance models, to distinguish them from ideal equivalent circuit models (ECM).

Electrochemical impedance spectroscopy (EIS) is a method frequently used to characterize

the impedance of batteries. In this method, the battery response to a small, sinusoidal signal

at multiple frequencies is measured. Using the Fourier transform the time domain response

can be converted to frequency domain. From the frequency domain data, a characteristic

Nyquist plot can be constructed and used to gain insight into the state of the battery,

including its internal resistances, capacities and time constants. To accomplish this, the

impedance response of a battery is fit to ECMs or impedance models to mimic electrical

systems that result in a similar shape of the Nyquist plot, in response to EIS. Any elements

used in ECMs or impedance models must relate to internal characteristics of the battery to

be meaningful [7]. Hardware to implement EIS is not usually found on-board battery

management systems (BMS) in EVs because of the added cost and complexity of the

- 60 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

electronics [8]. However, some recent studies have shown EIS implementations utilizing

existing electronics [9], [10]. If hardware barriers can be overcome, EIS could provide

valuable measurements that can be used in SoC estimation algorithms. Another concern

with EIS is that measurements are dependent on a variety of factors such as SoC, SoH,

temperature, and time of measurement with respect to other battery excitation

(charge/discharge) – known as relaxation. Therefore, the relationships between these

variables and the shape of the Nyquist plot must be studied and understood. Many have

studied the impact of SoC, SoH, and temperature [11]–[14], however, few have shown the

impact of relaxation. In [15], discharge pulses were used to measure the change in lithium-

ion battery impedance with relaxation. They found that many ECM parameters change with

relaxation for different SoCs and discharge currents. In [16] the change in OCV with

relaxation time was analyzed. Here, time constants of ECM parameters were related to

SoH. Relaxation was investigated by [17]–[19]. The relaxation effect was acknowledged

by [15] as part of an impedance characterization study and changes in the Nyquist plot were

shown to 40 hours. They concluded that impedance changes due to relaxation are small

compared to impedance changes due to temperature and SoC but significant enough to

require the relaxation effect to be accounted for when comparing impedance results. In [18]

the Nyquist plots were shown to stabilize after relaxation of 4 hours under constant SoC

and temperature. In [19] relaxation effects were shown to still be visible after 10 hours, and

a range of SoCs and temperatures were tested. Both [18] and [19] tested different lithium

ion chemistries and battery formats, and developed models to explain the cause of

relaxation.
- 61 -
 Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

The proposed study aims to provide insight into the variation of EIS in relation to

relaxation effects for a range of battery conditions for a state of the art commercial,

cylindrical cell. In addition, the impact of relaxation on the parameters of an impedance

model is analyzed.

2.2 EXPERIMENTAL

A custom battery test bench was designed to perform unattended relaxation experiments by

integrating a BioLogic SP150 potentiostat with Agilent loads and power supplies (N3306A,

N6773A respectively) and a Testequity 1007C thermal chamber. In this study a new

Samsung INR21700-48G, 4.8 Ah, cylindrical lithium ion battery was used. The cell was

conditioned by 10 full charge discharge cycles. The cell was charged as specified by the

manufacture’s datasheet starting with a constant current (CC) phase at 0.3C 4.2V followed

by a constant voltage (CV) phase to 0.02C cut-off current. Full discharge was done at 0.2C

to 2.5 V in CC mode, and SoC targets between 100% and 30% were obtained at 0.2C as

well using ampere-hour counting. EIS data was collected between frequencies of 30 mHz

to 30 kHz with a voltage amplitude of 5 mV, 6 points per decade and 5 sine wave periods

per frequency. EIS data was measured immediately after reaching the target SoC (via

discharge only), and re-measured at 30, 60, 180, 300, and 420 minutes. The relaxation test

was performed at 25 °C and 40 °C cell surface temperature.

2.3 RESULTS AND DISCUSSION

The EIS response of a battery was measured as described in the previous section to capture

the relaxation effect. The relaxation test was repeated three times to ensure the repeatability

- 62 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

of the impedance data. It was found that relaxation effects could be distinguished from

repeatability errors. The average standard deviation of the impedance was found to be 4.6e-

05 mΩ. Figure 2-1: Nyquist plots for 100%, 70% and 30% SoC at 25°𝐶 and 40°𝐶. shows

Nyquist plots for impedance data at 24 °C and 40 °C and how the impedance changes from

100 % to 70 % to 30 % SoC. In the following subsections the impedance data is fit to a

model and the change of model parameters with respect to the relaxation effect is analyzed.

Figure 2-1: Nyquist plots for 100%, 70% and 30% SoC at 25°𝐶 and 40°𝐶.

2.3.1 IMPEDANCE MODEL

Figure 2-2a shows an ECM adopted from [22] and used in this work. The complex

impedance 𝑍 of this ECM changes with frequency 𝜔 according to (2-1), where the time
1⁄
𝛼
constant 𝜏 = (𝑅𝑝 𝑄𝑝 ) .

𝑍(𝜔) = 𝑖𝜔𝐿 + 𝑅𝑠 + 𝑅𝑝/(1 + 𝑅𝑝𝑄𝑝(𝑖𝜔𝜏)𝛼) + 1/𝑄𝑝(𝑖𝜔)𝛽 (2-1)

- 63 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

Here, 𝑖 is the applied current magnitude, 𝐿 is the inductance due to cables used for

measurment, and 𝑅𝑠 is the ohmic resistance of the battery. 𝑅p and 𝑄p, 𝛼 define the

polarization resistance and a constant phase element (CPE) for capacitive effects from the

electric double layer. Together they form a ZArc element. Finally, 𝑄d, 𝛽 defines another

CPE to capture the solid-state diffusion process. The impedance model can fit the EIS data

well as shown in Figure 2-2b for impedance data after discharge to 90% SoC with no rest

and 7 hours of rest. The impedance model parameters were optimized using a combination

of non-linear least squares and particle swarm optimization algorithms.

Figure 2-2: Impedance model used to model relaxation effect a), and model fit to relaxation
data at 25°C and 90% SoC b).

- 64 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

However, neither algorithm was able to produce fits with consistent fitting error. Figure

2-3 shows how the fitting error changes for different SoCs and relaxation times. The model

fit has greater error at zero rest and stabilizes to a constant value for data at 60 minutes and

after. This fitting error must be considered when analyzing the relaxation results.

Figure 2-3: Impedance model fitting error with different SoCs over different rest times.

Figure 2-4 shows how the fit evolves with relaxation time at 25 °C and 40 °C. The change

of the Nyquist plot between 5 and 7 hours is small but still present and is smaller at higher

temperature when compared to the lower temperature. It can be observed that after 5 hours

the change in battery impedance has slowed significantly. However, it is unclear when

exactly it has slowed significantly enough such that any further impedance changes are

negligible. Changes in impedance may be deemed negligible if the change has minimal

impact on the fitting of an impedance model. This is because ultimately the impedance

- 65 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

model may be used to further analyze the battery behaviour and, therefore, the accuracy

and consistency of the model becomes important.

Figure 2-4: Modelled relaxation effect at 90% SoC and 25°𝐶 a), and 40°𝐶 b).

2.3.2 RELAXATION EFFECT

To understand how the parameters of the model shown in Figure 2-2a change with

relaxation time, the percentage change of each parameter P from its value at 7 hours (420
- 66 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

minutes) 𝑃420 was calculated using (2-2). Pt is the value of a model parameter (one of L,

Rs, Qd, Qp, Rp, a, or b) at relaxation time t. The datasets for each parameter were fit with

exponential functions to model the rate of change during relaxation.

𝑃t – 𝑃420
Δ420 = (2-2)
𝑃420

Figure 2-5 shows Δ420 at 25 °𝐶 for each model parameter at three SoCs, 100 % a), 90 %

b), 70 % c), and 30 % d). The inductance L and the ohmic resistance Rs stabilize within 10

minutes in all cases. The changes of L and Rs before 10 minutes are inconsonant and likely

due to the error introduced by the fitting process as discussed in the previous section. At

100% SoC the parameters with the longest relaxation time are Qp and Rp, i.e. two

components of the ZArc element. The depression constant a (or α) for the first semi-circle

does not change. The depression constant for the second semi-circle b (or β) shows a

smooth decay at a fast rate. The second CPE parameter Qd on the other hand shows noisy

data points. This may be explained by little or no lithium diffusion at 100% SoC due to low

availability of intercalation sites on the anode. At 90% SoC, where more intercalation sites

are available and the relatively high cell potential accelerates diffusion, Qd behaves similar

to Rp. At 70 % Qd stabilizes rapidly again which may indicate a point where the cell

potential (driving force) and the SoC (available intercalation sites) work together to

equalize concentration gradients [21]. This is reflected by the lower initial values of Δ420

for Qp and Rp. However, while the difference is lower initially, it takes longer for

parameters to stabilize due to many available intercalation cites for the lithium and,

therefore, potentially longer travel paths [19].

- 67 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

Figure 2-5: Percentage change in impedance model parameters from values at 420 minutes
at 25°C for a) 100% SoC, b) 90% SoC, c) 70% SoC and d) 30% SoC.

Finally, at 30 % SoC the cell potential is low, such that Qp, Rp and Qd equalize slowly

despite the large number of available interalation cites. In fact, the relaxation is slowest at

30 % for all three parameters. Lithium slowly diffuses into and through the electrode to

find intercalation cites during relaxation. At low SoC this process is slow because of the

reduced driving potential. The exponential fit to the Qp dataset at 30 % SoC (Figure 2-5d)

suggests that at 420 minutes (7 hours) the cell has not yet stabilized.

- 68 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

At 40 °C (Figure 2-6) the model parameters change in a way similar to that at 25 °C. The

parameters stabilize faster at higher SoC and slower at lower SoC, with 30 % showing the

slowest rate of stabilization. The polarization parameters Qp and Rp stabilize at a slightly

slower rate at 40 °C compared with 25 °C. This is unexpected because of higher reaction

kinetics at higher temperatures. However, the decrease in rate is small, such that additional

data will be needed to understand this trend. In contrast, Qd does stabilize faster in all cases

as expected. At 100 % SoC, both diffusion parameters exhibit noisy behaviour again. At

40 °C the exponential fit for Qp and Rp at 70 % SoC (Figure 2-5c) and Qp at 30 % (Figure

2-6d) do not reach steady state, suggesting again that further rest is required.

- 69 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

Figure 2-6: Percentage change in impedance model parameters from values at 420 minutes
at 40°C for a) 100% SoC, b) 90% SoC, c) 70% SoC and d) 30% SoC.

2.3.3 IMPEDANCE MAPS

Figure 2-7 shows how the dominant model parameters Qp, Rp, and Qd change with time and

SoC at 25 °𝐶. The relaxation effect manifests mostly at the edges of the three-dimensional

plots. These plots visualize the difference in magnitudes of the parameter change due to

relaxation and due to SoC. At the middle range of SoCs (80 % to 40 %) the parameter

changes due to relaxation become insignificant when compared to the parameter changes

due to SoC. However, the relaxation effect does cause significant parameter changes at the

edges of the SoC range (> 90 % and < 40 %). This is evident in Figure 2-7a and c for Qp
- 70 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

and Rp at 25 °𝐶. At 40 °𝐶 the same can be seen in Figure 2-8c for Rp only. The magnitudes

of the parameter values also change significantly between 25 °𝐶 and 40 °𝐶. The diffusion

CPE parameter Qd exhibits a linear trend with SoC from 90 % SoC for both temperatures.

This is a useful property of SoC estimation.

- 71 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

Figure 2-7: Impedance maps showing values for a) Qp, b) Qd and c) Rp for different SoCs
and rest times at 25°C.

The need for rest time during experiments can be eliminated with the help of impedance

maps shown in Figure 2-7 and Figure 2-8, since measurements at 0 minutes can be
- 72 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

extrapolated to rest time values. This is especially important for real time applications for

EIS measurement where rest times are impractical.

Figure 2-8: Impedance maps showing values for Qp, Rp, and Qd for different SoCs and
rest times at 25°C (a-c), and 40°C (d-f).
- 73 -
 Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

2.4 CONCLUSION

In the work presented in this paper the relaxation effect of a commercial lithium ion battery

was characterized using EIS and impedance modelling. A suitable impedance model was

found in literature and used to model the relaxation effect. The change of each model

parameter with relaxation time was analyzed and found to be consistent with

electrochemical theory. Results at 40 °𝐶 indicate a small decrease in relaxation rate which

must be investigated further as it is contrary to previous literature findings. Impedance

maps show that the changes in model parameters due to relaxation are significant at certain

conditions when compared to the changes in parameters due to SoC and temperature. This

should be validated by assessing the impact of the change in model parameters due to

relaxation on the accuracy of the model. This work highlights again the need for careful

consideration of the relaxation effect. The study will be expanded to longer rest times as

well as different temperatures, charge/discharge rates, battery types, and impedance model

in future work.

REFERENCES
[1] M. A. Hannan, M. S. H. Lipu, A. Hussain, and A. Mohamed, “A review of lithium-
ion battery state of charge estimation and management system in electric vehicle
applications: Challenges and recommendations,” Renew. Sustain. Energy Rev., vol.
78, no. August 2016, pp. 834–854, 2017.
[2] R. Xiong, J. Cao, Q. Yu, H. He, and F. Sun, “Critical Review on the Battery State of
Charge Estimation Methods for Electric Vehicles,” IEEE Access, vol. 6, pp. 1832–
1843, 2018.
[3] L. Ungurean, G. Cârstoiu, M. V Micea, and V. Groza, “Battery state of health
estimation : a structured review of models , methods and commercial devices,” no.
July 2016, pp. 151–181, 2017.
[4] M. M. Kabir and D. Demirocak, “Degradation mechanisms in Li-ion batteries: a
state-of- the-art review,” Int. J. energy Res., vol. 41, no. April 2017, pp. 1963–1986,
2017.
[5] M. U. Cuma and T. Koroglu, “A comprehensive review on estimation strategies used
in hybrid and battery electric vehicles,” Renew. Sustain. Energy Rev., vol. 42, pp.
- 74 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

517–531, 2015.
[6] M. Schönleber, C. Uhlmann, P. Braun, A. Weber, and E. Ivers-Tiffée, “A Consistent
Derivation of the Impedance of a Lithium-Ion Battery Electrode and its Dependency
on the State-of-Charge,” Electrochim. Acta, vol. 243, pp. 250–259, 2017.
[7] R. Cottis and S. Turgoose, Electrochemical Impedance and Noise. Houston: NACE
International, 1999.
[8] J. Meng et al., “An Overview and Comparison of Online Implementable SOC
Estimation Methods for Lithium-Ion Battery,” IEEE Trans. Ind. Appl., vol. 54, no.
2, pp. 1583–1591, 2018.
[9] E. Din, C. Schaef, K. Moffat, and J. T. Stauth, “A scalable active battery
management system with embedded real-time electrochemical impedance
spectroscopy,” IEEE Trans. Power Electron., vol. 32, no. 7, pp. 5688–5698, 2017.
[10] X. Wei, X. Wang, and H. Dai, “Practical on-board measurement of lithium ion
battery impedance based on distributed voltage and current sampling,” Energies,
vol. 11, no. 1, 2018.
[11] P. Kollmeyer, A. Hackl, and A. Emadi, “Li-ion battery model performance for
automotive drive cycles with current pulse and EIS parameterization,” 2017 IEEE
Transp. Electrif. Conf. Expo, ITEC 2017, pp. 486–492, 2017.
[12] C. Pastor-Fernández, K. Uddin, G. H. Chouchelamane, W. D. Widanage, and J.
Marco, “A Comparison between Electrochemical Impedance Spectroscopy and
Incremental Capacity-Differential Voltage as Li-ion Diagnostic Techniques to
Identify and Quantify the Effects of Degradation Modes within Battery Management
Systems,” J. Power Sources, vol. 360, pp. 301–318, 2017.
[13] U. Westerhoff, T. Kroker, K. Kurbach, and M. Kurrat, “Electrochemical impedance
spectroscopy based estimation of the state of charge of lithium-ion batteries,” J.
Energy Storage, vol. 8, pp. 244–256, 2016.
[14] B. Fridholm, T. Wik, and M. Nilsson, “Robust recursive impedance estimation for
automotive lithium-ion batteries,” J. Power Sources, vol. 304, pp. 33–41, 2016.
[15] H. Wang, M. Tahan, and T. Hu, “Effects of rest time on equivalent circuit model for
a li-ion battery,” Proc. Am. Control Conf., vol. 2016–July, pp. 3101–3106, 2016.
[16] P. S. Attidekou, C. Wang, M. Armstrong, S. M. Lambert, and P. A. Christensen, “A
New Time Constant Approach to Online Capacity Monitoring and Lifetime
Prediction of Lithium Ion Batteries for Electric Vehicles (EV),” J. Electrochem.
Soc., vol. 164, no. 9, pp. A1792–A1801, 2017.
[17] W. Waag, S. Käbitz, and D. U. Sauer, “Experimental investigation of the lithium-
ion battery impedance characteristic at various conditions and aging states and its
influence on the application,” Appl. Energy, vol. 102, pp. 885–897, 2013.
[18] A. Barai, G. Chouchelamane, Y. Guo, A. McGordon, and P. Jennings, “A study on
the impact of lithium-ion cell relaxation on electrochemical impedance
spectroscopy,” J. Power Sources, vol. 280, no. Special Issue, pp. 74–80, 2015.
[19] F. M. Kindermann, A. Noel, S. V. Erhard, and A. Jossen, “Long-term equalization
effects in Li-ion batteries due to local state of charge inhomogeneities and their
impact on impedance measurements,” Electrochim. Acta, vol. 185, pp. 107–116,
2015.
[20] J. Schmitt, A. Maheshwari, M. Heck, S. Lux, and M. Vetter, “Impedance change and
capacity fade of lithium nickel manganese cobalt oxide-based batteries during
calendar aging,” J. Power Sources, vol. 353, pp. 183–194, 2017.
[21] T. R. Jow, S. A. Delp, J. L. Allen, J.-P. Jones, and M. C. Smart, “Factors Limiting
Li + Charge Transfer Kinetics in Li-Ion Batteries,” J. Electrochem. Soc., vol. 165,
no. 2, pp. 361–367, 2018.
- 75 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

Chapter 3: Electrochemical Impedance Spectroscopy with

Practical Rest-times for Battery Management
Applications

Marvin Messing1,2, Tina Shoa2, Saeid Habibi1

1
Department of Mechanical Engineering, McMaster University, Hamilton, ON, Canada,
2
Cadex Electronics, Richmond, BC, Canada
This paper is published in IEEE Access, vol. 9, pp. 66989-66998, 2021, doi:
10.1109/ACCESS.2021.3077211. This paper is republished here with permission.

ABSTRACT

The State of Charge (SoC), State of Health (SoH), and State of available Power (SoaP) of

Lithium-Ion Batteries (LiB) are critical quantities which cannot be measured but must be

estimated by Battery Management Systems (BMS) instead. A technique known as

Electrochemical Impedance Spectroscopy (EIS) provides a non-destructive way of

measuring battery impedance, offering detailed insight into the battery state of operation.

Several challenges must be solved in to utilize EIS as part of the BMS, including the

defining of operating conditions at which to perform the sensitive EIS measurement. In

laboratory conditions, several hours of rest are used to ensure a stable response of the

battery, but such rest times are impractical for BMS applications. This paper proposes a

methodology combining drift compensation and a Voigt-circuit used as a filter to obtain

valid EIS data with short rest times under different operating conditions. Extensive tests

were conducted on lithium nickel manganese cobalt (NMC) batteries to evaluate the

methodology and show how the different operating conditions impact the rest time required

- 76 -
 Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

for valid EIS measurement. The results show that rest times between 5 and 30 minutes can

be used to obtain useful EIS data for a wide range of operating conditions.

3.1 INTRODUCTION

In this paper, we address some of the challenges pertaining to Electrochemical Impedance

Spectroscopy (EIS) measurements for applications in Battery Management Systems

(BMS). Specifically, the impact of the battery relaxation effect on EIS measurements is

examined to determine the shortest possible rest required before valid EIS results can be

obtained under various operating conditions. Extensive experiments have been conducted

on Lithium nickel manganese cobalt (LiNiMnCoO2 or NMC) batteries. This paper

proposes a methodology that combines drift compensation and Voigt-circuit based filtering

to produce valid EIS results. In addition, different battery relaxation behavior is induced

using discharge pulses of different current rates (C-rate) to understand the relationship

between battery discharge and relaxation. The discharge pulses are performed at different

SoCs and temperatures, and EIS measurements are obtained for each case at different rest

times. This section discusses the motivation and technical challenges, summarizes recent

relevant literature, presents the main paper contributions, and provides an outline of the rest

of the paper.

3.1.1 MOTIVATION AND TECHNICAL CHALLENGES

Battery management systems (BMS) perform multitude of functions that impact the

performance and the safe operation of batteries in advanced applications. Lithium-ion

batteries (LiB) are currently the preferred choice for energy storage in Electric Vehicles

- 77 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

(EVs) as well as many other applications, primarily due to their high energy density and

their relatively long cycle life [1]. The BMS must estimate, in real-time, the battery’s State

of Charge (SoC), State of available Power (SoaP) and SoH, but this continues to remain a

challenge [2], [3]. SoC provides an indication of how much driving range remains before

the battery must be charged. However, as the SoC decreases, the battery voltage also drops,

which means more current is required at lower SoC to achieve the same power compared

to high SoC. SoaP provides a way of predicting the available power output within the

voltage and current limits of the cell in the battery pack. SoH determines the total available

capacity of a battery and gives insight into the battery behavior as it changes with age.

Inaccuracies in SoH estimates lead to further inaccuracies in the SoC and SoaP estimates,

and, as a result, battery packs in EVs are over engineered to account for these uncertainties

[4]. A variety of methods exist to estimate SoC, SoH, and SoaP for BMS and EV

applications including model-based filters and observers [5]–[12] machine learning

methods [13], [14] and methods based on direct measurement of charge/discharge behavior

of the battery [15] as well as several others [16]. In laboratory settings, advanced

instrumentation and exhaustive test methods are used to extract additional information from

the battery and infer SoC and SoH. One such method is EIS, which measures the impedance

of a battery [17]. In EIS, a sinusoidal voltage or current signal of small amplitude is applied

to the battery at different frequencies. Using the Fast Fourier transform (FFT) of these

signals, the impedance of the battery is obtained by calculating its amplitude and phase

responses for each frequency. The real and imaginary component pairs of the complex

impedance can be plotted to provide the Nyquist plot. The battery impedance and its
- 78 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

representation of Nyquist plot changes depending on the state of the battery. In general, the

high cost of EIS hardware as well as length of test times make EIS characterization

unfeasible for integration into BMS [18], however, efforts are being made to change this

by developing more cost-effective ways to generate EIS [19]. The hardware requirements

depend on the desired frequency range and the range of the internal resistance of the

batteries used. Furthermore, the hardware should be able to maintain a sufficiently high

signal to noise ratio, while only exciting the battery within its linear operating range. Using

potentiostatic EIS (PEIS) allows for the control of the excitation voltage such that linearity

is maintained. However, the lower the battery internal resistance, the higher the output

currents that must be measured. To avoid high output currents, galvanostaic EIS (GEIS)

can be used, where current is the input signal and voltage is measured. For GEIS the current

must be carefully chosen such that the signal to noise ratio of the voltage measurement is

sufficiently high, but the linear operating range of the battery is not exceeded and the

excitation does not cause the battery to heat up significantly. Din et al. [20] combined active

balancing circuitry and a control strategy to produce sinusoidal battery cell excitation. They

used a switched inductor ladder configuration which allows dissipation of power from one

cell to another, reducing energy loss. Their results show close agreement when compared

to lab-grade measurements. Wei et al. [21] proposed a hardware solution for exciting 12

battery cells in series with a single current signal, and measured the voltage response of

each cell. They use a DC to AC converter to generate the pack input signal. Lohmann et

al. [22] proposed a method for extracting impedance spectra from conventional driving

cycle data. In this scenario, the electric motor of the EV generates the signal. However, and
- 79 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

as a result, the frequency range is limited to that of the drive cycles. The studies reviewed

above provide possible answers for practical EIS as part of a BMS, however, consideration

of relaxation effects and required rest times before measurement are largely omitted. In the

next part of this section, studies which discuss relaxation effects are reviewed.

3.1.2 RELAXATION EFFECT AND EIS MEASUREMENT

EIS varies according to the battery’s SoH and SoC amongst other factors. To use EIS in

practical applications, the impact of different measurement conditions must be investigated.

SoC, SoH, and temperature were investigated in several studies [23]–[27] in laboratory

settings for the purpose of battery characterization. In most EIS studies, a rest time period

of 1 to 24 hours is used after charge or discharge and before the EIS measurement [28].

This is done to ensure that only the small signal EIS excitation is causing the battery

response measured by EIS. During normal discharge, the battery response is highly non-

linear, and once the discharge is stopped, the battery voltage rises rapidly (relaxation

effect). Over time, the voltage stabilizes, and EIS is usually measured once the relaxation

effect has slowed significantly, to avoid contribution of the relaxation effect to the EIS

measurement. Kindermann et al. [28] presented a thorough investigation of the long-term

relaxation effects under different conditions and for different types of batteries. They

showed that battery relaxation still impacts EIS results after 40 hours, and low SoC and low

temperature slow the rate of relaxation. Barai et al. [29] also studied the long-term

relaxation effect as well as its impact on cell resistance and capacitance, showing that ohmic

resistance is independent of relaxation, but capacitance follows a logarithmic trend. Waag

- 80 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

et al. [30] acknowledged the relaxation effect and demonstrated changes in the Nyquist plot

after 40 hours. They concluded that the impact of the relaxation effect is small compared

to the impact of SoC and temperature, but significant enough to account for it when

interpreting EIS results since the relaxation effect can change the EIS measurement by 5 to

10%. The relaxation effect was also investigated without the use of EIS by Wang et al. [31],

who used a current pulse followed by different rest times to extract rest time dependent

parameters for a second-order equivalent circuit model. They found robust trends for

changes in model parameters with rest time, discharge current and SoC. Zinth et. al. [32]

investigated changes in graphite anode electrodes using neutron diffraction to track the

impact of different discharge rates on the relaxation effect. They showed long lasting

effects for low temperatures beyond 11 hours. Deverakonda et al. [27] studied the impact

of the relaxation effect on the parameters of the Thevenin equivalent circuit model after

different discharge and charge currents for lead acid batteries. In our previous study on

relaxation [33], we investigated how fractional order impedance model parameters change

with the long-term relaxation effect at different SoCs and temperatures.

The above studies all focus on the behavior and implications of the battery relaxation effect

over long periods of time, conclude that the relaxation effect is still present after days of

rest, and largely use 1 to 3 hours before performing EIS measurements. However, no

solutions exist for obtaining EIS results with short rest times (<1 hour). Short rest times are

desirable to reduce measurement times for BMS applications, but validity of EIS

measurements must be ensured by filtering out measurement contributions from the

relaxation effect. The objective of the present study is to address this gap by investigating
- 81 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

the short-term relaxation effect of different severity induced by discharge pulses for

different SoCs, and temperatures to define operating points for EIS measurement with short

rest time.

3.1.3 CONTRIBUTIONS

This paper presents the following original contributions and conclusions: (1) A

methodology using drift compensation and a Voigt-circuit as a filter to convert short rest-

time EIS measurements to valid EIS data is proposed (2) a method comparing EIS

measurements obtained during relaxation effects is developed (3) it is shown that EIS

measurements with short term relaxation times are valid and useful under specific

conditions (4) the effect of discharge pulse depths is only visible up to 30 minutes of rest

for high SoC, and temperatures at and above 25°C.

3.1.4 PAPER OUTLINE

The paper is organized as follows. Section 3.2 discusses EIS validity, drift compensation,

Kramer-Kronigs relations and then introduces a methodology which combines these

methods with a Voigt-circuit used as a filter for obtaining valid EIS results with short rest

times. In Section 3.3, the experimental methods used for investigating the relaxation effect

are described. Section 3.4 describes a method to simplify analysis of the relaxation effect.

Section 3.5 discusses the general variation in behavior of the relaxation effect under

different experimental conditions including changes in SoC and temperature, evaluates

short rest time EIS validity, and shows how the combination of discharge pules, SoC, and

- 82 -
 Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

temperature impacts the relaxation effect and EIS results. Concluding remarks are provided

in Section 3.6.

3.2 SHORT TERM EIS MEASUREMENT METHOD

When a discharge pulse is applied to a battery, the rate of the relaxation effect changes

depending on the C-rate. Higher C-rates, for example 5C, cause more pronounced and

longer relaxation behavior. Figure 3-1 shows Nyquist plots after a 5C discharge pulse at

different rest times for three tests (sets 1 to 3) conducted on different days, but with constant

operating conditions (90% SoC, 25°C). The results show that the EIS measurements are

repeatable enough to clearly distinguish different rest times, suggesting that the relaxation

effect yields repeatable results when measured with EIS. This repeatability of the short-

term relaxation effect suggests it is worth measuring EIS with short rest times if the effects

of relaxation are understood and properly compensated.

Figure 3-1: Three repeats of EIS measurements (set 1, 2 and 3) for different rest times after
a 5C discharge pulse.

- 83 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

An EIS measurement is considered valid if it is applied to a system that satisfies the criteria

of linearity, stability and causality [17]. EIS can be applied to non-linear systems such as

batteries only if they remain within a piece-wise linear operating region, normally

following long rest periods and requiring small amplitudes of input excitations. Collecting

an EIS spectrum with short rest times after a high current discharge will cause drift in the

measurement as shown in Figure 3-2 for 1 minute of rest after a discharge pulse, thereby

violating causality and stability. Nonetheless, EIS can still be used in this case by applying

drift compensation and by using the Kramer-Kronigs transformable Voigt circuit as a filter,

which are discussed in this section.

Figure 3-2: Effect of drift compensation with 1-minute rest.

Drift in the EIS measurement can be compensated by following the procedure described by

Zahner [74], which is implemented in most modern potentiostats. In this method, multiple

periods of the EIS excitation signal are measured at each frequency to determine the non-

periodic (under drift), mean DC component. The mean DC component is then subtracted

from the signal to obtain a drift-corrected signal. This signal still contains non-periodic

components, which are identified and eliminated in the frequency domain after applying
- 84 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

the FFT. The resulting EIS measurement, after drift compensation, is also shown in Figure

3-2.

The Kramer-Kronigs transform calculates the imaginary part (Z’’) of the impedance from

the real part (3-1) or the real part (Z’) from the imaginary part (3-2) given a frequency ω as

follows.

2𝜔 ∞ 𝑍 ′ (𝑥)−𝑍 ′ (𝜔)
𝑍 ′′ (𝜔) = − ( 𝜋 ) ∫0 𝑥 2−𝜔 2
𝑑𝑥 (3-1)
2 ∞ 𝑥𝑍 ′′ (𝑥)−𝜔𝑍 ′′ (𝜔)
𝑍 ′ (𝜔) = 𝑍 ′ (∞) + ∫0 𝑑𝑥 (3-2)
𝜋 𝑥 2−𝜔 2

This transform only applies if the linearity, stability, and causality criteria are satisfied, thus

providing a way of validating EIS measurements. However, EIS cannot be measured for

frequencies (𝑥) from zero to infinity, making the KK-transform itself impossible to

implement. To circumvent this limitation, the EIS spectra were fitted to a Voigt circuit

shown in Figure 3-3. A good fit, with random residuals and low mean-square-error (MSE),

indicates a valid EIS spectrum, since the Voigt circuit is known to be KK-transformable

[17], [35].

Figure 3-3: Voigt Circuit

Equation 3-3 defines the Voigt circuit, where the impedance 𝑍 = 𝑍 ′ + 𝑗𝑍′′, the sum of real

and imaginary parts given a frequency 𝜔𝑘 :

𝑅 𝜔 𝜏𝑅
𝑍(𝜔𝑘 ) = ∑𝑛𝑖=1 1+(𝜔𝑖 𝜏 )2 − 𝑗 ∑𝑛𝑖=1 1+(𝜔
𝑘 𝑖 𝑖
𝜏 )2
(3-3)
𝑘 𝑖 𝑘 𝑖

- 85 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

Here, 𝜏𝑖 = 1⁄𝜔𝑖 are the time constants and 𝜔𝑖 are the experimental frequencies. Equation

3-3 was extended as shown in (3-4):

1 𝑅
𝑍(𝜔𝑘 ) = 𝑗𝜔𝑘 𝐿 − 𝑗 𝜔 + 𝑅0 + ∑𝑛𝑖=1 1+𝑗𝜔𝑖 (3-4)
𝑘𝐶 𝑘 𝜏𝑖

A capacitor (with capacitance C) and an inductor (with inductance L) were added in series.

The resistances 𝑅𝑖 are the only unknowns, since the capacitances 𝐶𝑖 = 𝜏𝑖 ⁄𝑅𝑖 . The time

constants 𝜏𝑖 may be computed for each frequency 𝜔𝑖 , but fewer time constants may be

necessary to avoid overfitting. Overfitting can be detected by computing the ratio, μ, of the

sum of positive and negative 𝑅𝑖 parameters as described by Schonleber et. al [36] and

shown in (3-5).

∑𝑅 |𝑅𝑖 |
𝑖 <0
𝜇 =1−∑ (3-5)
𝑅𝑖 ≥0 |𝑅𝑖 |

This ratio 𝜇 measures oscillations in the Voigt fit which result from noise or measurement

errors. Oscillations in the Voigt fit are only mathematically possible with some of the 𝑅𝑖 ’s

(the only unknowns in Voigt fit) are negative. Therefore, 𝜇 relates the amount of positive

𝑅𝑖 elements to the amount of negative 𝑅𝑖 ’s. If μ has a value of one, the amount of negative

𝑅𝑖 ’s is small and no overfitting occurs.

Figure 3-4a shows an example of EIS data used in this study (5-minute rest after 5C

discharge) and the corresponding Voigt circuit fit. The Nyquist plot (Figure 3-4a) as well

as the gain and phase plots (Figure 3-4b) show close agreement between the EIS

measurements and Voigt fit. To further evaluate the goodness of fit, the residuals (Figure

3-4c) and the MSE (Figure 3-4d) are considered. The residuals appear to resemble white

noise, as shown in Figure 3-4c for 5 minutes of rest, suggesting that the Voigt fit is valid

- 86 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

for EIS at short rest times. The MSE stabilizes after 10 minutes indicating that short rest

time EIS results are Kramer-Kronigs transformable. The MSE can be used to determine a

threshold value below which EIS results are considered valid. This is discussed further in

Section 3.5.2 with additional datasets.

Figure 3-4: Real and imaginary EIS data with Voigt fit at 5-minute rest (a), EIS gain and
phase with Voigt fit at 5-minute rest (b), Voigt fit residuals at 5-minute rest (c), Voigt fit
MSE for various rest times and datasets (d) all after 5C discharge.

The validated EIS data still includes a small amount of noise. When analyzing a single

dataset, this noise is small enough to not be of concern. However, when comparing

consecutive EIS measurements, as is done in this study for different rest times, the noise of

- 87 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

different EIS datasets compounds. Figure 3-5 shows the difference of two (raw) EIS gain

plots exhibiting compounded noise in the mid frequency range. The Voigt circuit can help

to reduce this noise. Instead of using the EIS measurements directly, the output of the Voigt

model can be used. To do this, the Voigt-circuit is first fit to the complex impedance (real

and imaginary) pairs of the EIS data. Then the measurement frequency of each complex

impedance pair is used as an input to the Voigt circuit model, creating a new, modeled set

of real and imaginary pairs. The modeled complex impedance pairs are similar to the

measured ones, as shown, for example, in Figure 3-5b in the gain plot. However, when

considering again the difference of two consecutive datasets, a smoothing effect can be

observed. Figure 3-5 also shows the difference of two modeled (filtered) EIS gain plots.

The Voigt-model filtered plot is significantly smoother than the raw EIS plot in Figure 3-5.

In Section 3.4 a method is introduced that uses the difference between consecutive EIS

scans and relies on the smoothing effect of the Voigt filter.

- 88 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

Figure 3-5: Voigt filter smoothing result.

The full procedure combining drift-compensation and Voigt-circuit filtering is shown in

Figure 3-6. During the Voigt-circuit fitting procedure, μ, shown in (3-5), is used to prevent

overfitting. If the MSE of the Voigt-circuit fit is found to be sufficiently low, the output of

the Voigt-circuit fit is used to generate new EIS datapoints at the measurement frequencies.

The combination of all steps shown in Figure 3-6 reduces the impact of the relaxation effect

on EIS measurements. The residuals and MSE resulting from the Voigt circuit fit are used

as criteria to determine if EIS results are valid. In the next section, experiments are

described to test this proposed methodology.

- 89 -
 Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

Figure 3-6: Short rest time EIS measurement methodology.

3.3 EIS CHARACTERIZATION AND EXPERIMENTS

This section describes characterization and experiments designed to further investigate the

impact of the relaxation effect on EIS and to analyze the usefulness and limitations of using

Voigt-circuit filtering together with drift compensation as introduced in the previous

section. To induce relaxation behavior of different magnitudes, a protocol was developed

consisting of discharge pulses between 1C and 5C with each having a discharge duration

of 15s followed by a 3-hour rest. During the rest period, EIS was measured periodically.

The protocol is shown in Figure 3-7 and consists of the following steps:

1. The battery is fully charged (CC-CV mode at 0.3C to 4.2V and 0.02C cutoff

current).

- 90 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

2. The battery is discharged at 0.2C to a specific SoC target determined by coulomb

counting and by subtracting the charge lost during the pulse such that the SoC target will

be reached after the discharge pulse;

3. A 5C discharge pulse is applied for 15 seconds;

4. EIS is measured every 5 minutes until 30 minutes past the initial pulse discharge,

then after every 15 minutes until 1 hour, and every hour until 3 hours is past;

5. Steps 1 to 4 are repeated for 4C, 3C, 2C, and 1C discharge pulses.

Figure 3-7: Experimental protocol showing where EIS measurements are taken with respect
to discharge pulses of different C-rates.

The schematic of the experimental setup that was used for the experiments in this study is

shown in Figure 3-8 for one channel. It consisted of the following devices:

• A BioLogic SP150 potentiostat for EIS measurement with a single multiplexed

measurement channel.

- 91 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

• Agilent loads and power supplies (N3306A, N6733A respectively) with four

channels each to supply and draw current.

• A National Instruments data acquisition (NI-DAQ) module with NI 9219 to

measure cell voltage and NI 9201 to measure shunt voltage to obtain the cell current, with

four measurement channels per module.

• A Testequity 1007C thermal chamber to control cell temperature.

• Custom software (written in Python) to control and automate testing with all the

devices.

The batteries used in this study were fresh Samsung INR21700-50E, 4.9Ah, cylindrical

lithium ion cells also shown in Figure 3-8. Prior to any tests, the new battery cells were

conditioned with 10 charge/discharge cycles.

Figure 3-8: Schematic of experimental setup and battery cell.

EIS was performed between 800 Hz and 0.1 Hz, with a voltage amplitude of 10 mV, 6

points per decade, and two sine wave periods per frequency. The drift compensation option

- 92 -
 Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

of the potentiostat was applied in all cases to reduce the effect of drift as described in

Section 3.2.

3.4 RELAXATION EFFECT ANALYSIS METHOD

Discharge pulses of different C-rates cause different relaxation behavior which causes

differences in EIS results. If this difference is large, the C-rate must be considered for the

interpretation of the EIS results. Otherwise, the results may be mis-interpreted. Visually

evaluating the differences in Nyquist plots is not very useful, so a method is proposed in

this section which allows for a quantitative comparison of the impact of different C-rates

on EIS results. In this method, the percent difference between the magnitude of the

impedance following the lowest (1C) and highest (5C) discharge pulses was calculated for

various rest times as shown in (3-6).

𝑎𝑏𝑠(|𝑍(𝜔)|1𝐶 −|𝑍(𝜔)|5𝐶 )
Δ𝑍(𝜔) = 100 × |𝑍(𝜔)|5𝐶
(3-6)
Here, |𝑍(𝜔)|1𝐶 and |𝑍(𝜔)|5𝐶 are the magnitudes of the impedance data for one frequency

𝜔 after 1C and 5C discharge pulses, respectively. Figure 3-9 shows Δ𝑍(𝜔) for different

frequencies and rest times. Between 10 and 0.1 Hz, Δ𝑍(𝜔) shows a flat region for all rest

times. To further simplify comparison of the rest time effect across different operating

conditions and C-rates, Δ𝑍(5) is used as a representative value from this flat region. This

value at 5Hz will be referred to as simply Δ𝑍 in the remainder of this paper. As evident in

Figure 3-9, Δ𝑍 becomes smaller and smaller as the battery rests. This indicates that the

differences in relaxation rates induced by 1C and 5C discharge pulses is only significant

- 93 -
 Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

for a short time after which the relaxation rates merge. This is discussed further and for

various operating conditions in the next section.

Figure 3-9: ΔZ(ω) for different frequencies and rest times.

3.5 RESULTS & DISCUSSION

In this section the relaxation effect is analyzed using the methods described in the previous

sections with EIS data collected under different battery operating conditions. This section

is organized as follows: 1. The impact of the discharge pulses at different C-rates on the

EIS results is shown under constant operating conditions; 2. The effectiveness of the short-

term measurement and filtering methodology is evaluated for different SoCs and

temperatures using the MSE of the Voight fit; and 3. The combined impact of SoC,

temperature and C-rate of the discharge pulses on the relaxation behavior is presented.

3.5.1 RELAXATION EFFECT AND C-RATE

Figure 3-10 shows how Nyquist plots change over time after pulses of different C-rates

(between 1C and 5C). The plots for each of the different C-rates are clearly separated after

10 minutes of rest time due to the different relaxation behaviors caused by the various

- 94 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

discharge depths. After 30 minutes of rest, however, all five datasets merge together to

become nearly indistinguishable. After 3 hours (180 minutes), the datasets overlap

completely, but have shifted further due to the continuing relaxation effect, however, in a

manner independent of the foregoing discharge pulse depth. Figure 3-10 also shows that

the lower the C-rate, i.e. the lower the initial disturbance, the smaller the difference between

the disturbed and the rested EIS results.

Figure 3-10: Evolution of Nyquist plots measured at different rest times after different C-
rates.

As was shown in Figure 3-9 (Section 3.4) for measurements at 90% SoC, and 25°𝐶, Δ𝑍(𝜔),

which evaluates the difference in EIS results after 1C and 5C discharge pulses, is large

initially for rest times of 10 minutes and 15 minutes. After 30 minutes of rest time, Δ𝑍(𝜔)

is similar to its values after 3 hours. These results suggest that rest times may be shortened

significantly, from the 3 hours conventionally observed to 30 minutes, since the effect of

C-rate becomes small. With this rest time of 30 minutes, EIS results can be interpreted

without considering (or measuring) the C-rate which induced the relaxation effect. If C-

rate is measured it can be used to inform EIS interpretation, or EIS measurements can be
- 95 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

limited to follow low C-rates (1C or less) only. In either case, rest times may be shortened

to less than 30 minutes if validity of the EIS measurement can be ensured. This is discussed

in the next section.

3.5.2 RELAXATION EFFECT AND VOIGT-FIT MSE

The previous section showed that the differences in Nyquist plots after 1C and 5C discharge

pulses becomes small after 30 minutes of rest. In this section, the MSE of the Voigt-fit (see

Section 3.2) is used to further investigate the short-term relaxation effect for different

operating conditions. Figure 3-11a shows the Voigt-fit MSE for different rest times and

different SoCs all measured at 25 °C. For rest times below 25 minutes, the spread of MSE

is relatively large. After 25 minutes, although there are some outliers, the MSE stays within

smaller clusters. Figure 3-11b shows the Voigt-fit MSE for different C-rates and different

temperatures, all measured at 90% SoC. Here, the clusters are similar to Figure 3-11a,

however, the MSE at 10°C is consistently higher and shows a large variance. This suggests

that the Voigt-fit is of low quality, and EIS results may not be valid even after 3 hours of

rest at low temperature. This is due to low temperatures causing the battery to relax at a

much slower rate, which was also shown by Kindermann et al. [28]. The implication for

the short-term relaxation affect is that much longer rest times are required to obtain valid

EIS results at lower temperatures. For all other conditions observed, the MSE appears to

stabilize quickly. To better understand this, the MSEs for the valid conditions were

combined and the mean and standard deviation (SD) for each rest time was obtained, as

shown in Figure 3-11c. After 1-minute rest, the mean is high, and a large SD can be seen.

- 96 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

Between 5 to 20 minutes of rest, the means are similar and the SD error bars overlap, but

the means are much lower than after 1 minute of rest. Finally, between 25 minutes and 3

hours of rest, the means are even lower and in a similar range with overlapping error bars.

Therefore, for best results, a 25 min rest should be observed. However, rest times as low as

5 minutes will produce EIS results with only slightly elevated Voigt-fit MSEs, which may

still be valid. Rest times of less than 5 minutes will not produce valid EIS results.

Figure 3-11: Voigt-fit MSE for different C-rates and SoCs (a), MSE for different C-rates
and temperatures (b), mean and SD of combined conditions (c).
- 97 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

3.5.3 RELAXATION EFFECT AND Δ𝑍

The previous section showed how the Voigt-fit MSE changes with different SoCs,

temperatures and C-rates. To better visualize the effects of different C-rates in combination

with SoC and temperature, Δ𝑍 (Section 3.4) is plotted over rest time in Figure 3-12a and

Figure 3-12b, respectively, with 90% SoC, 25°𝐶 at 30 minutes of rest used as a reference.

The Nyquist plots for each condition are also shown for each case. In Figure 3-12a, the

impact of rest time and SoC on ΔZ is shown. ΔZ increases as the SoC decreases, which

means longer rest times are required for lower SoCs to sufficiently reduce the effect of C-

rate on the EIS results. At 70% SoC Δ𝑍 reaches a value below the reference after 60 minutes

of rest. At 50% SoC Δ𝑍 is still above the threshold even after 3 hours of rest. This can also

be seen in the Nyquist plots from EIS measurements with 30-minute rest times after 1C and

5C discharge as shown in Figure 3-12b. At 90% SoC, the Nyquist plots for the two C-rates

overlap, however, at 50% SoC, a significant gap exists between the two plots. This

suggests, only high SoCs should be used when measuring EIS with short rest times after

large discharge rates precede the measurement.

- 98 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

Figure 3-12: ΔZ for 90%, 70% and 50% SoC compared to reference at 25°C and 90% SoC
(a). Nyquist plots for 90% and 50% SoC after 30 minutes of rest (b).

Figure 3-13a shows ΔZ for different temperatures. Increasing the temperature from 25°C

to 35°C has no significant impact on the relaxation behavior. However, decreasing the

temperature to 10°C increases the time it takes to reduce ΔZ, suggesting that, again, lower

temperatures are not suitable for EIS measurements with short rest times. At 10°C, ΔZ

reaches below the reference after approximately 2.5 hours of rest, however, as was shown

in the previous section, EIS results at 10°C may still not be valid even after 3 hours of rest.

At higher temperatures, ΔZ starts out much lower than at 25°C at 5 minutes rest and then

rapidly decreases to match the values at 25°C for longer rest times. The Nyquist plots

corresponding to the different temperatures are shown in Figure 3-13b, showing

overlapping plots for higher temperatures, but a small gap for the 10°C case. The results

discussed above show that the 30-minutes rest time is sufficient for room temperature and

above, as long as the SoC remains high. Lower SoCs and lower temperatures still show a

large difference in EIS results even after 3 hours of rest, so care must be taken when

- 99 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

interpreting EIS under such conditions. The findings on rest times for different operating

conditions for relaxation induced by a 5C discharge pulse are summarized in Table 3-1.

Figure 3-13: ΔZ at 25°C, 35°C and 10°C compared to reference at 25°C and 90% SoC (a).
Nyquist plots for the same temperatures (b).

Table 3-1: Summary of Minimum Rest Time for Different Conditions

Temperature (°𝐶) SoC (%) C-rate Rest Time (min)

90 5 30
≥ 25 70 5 60
50 5 > 180
10 90 5 > 180
The results shown in this study are applicable to batteries with NMC cathodes. The

relationship between the relaxation effect and battery chemistry is not well researched in

literature. However, the long term (48 hours) relaxation rate of NMC was compared to

Lithium Manganese (LiMn) and Lithium-Iron-Phosphate (LFP) chemistries by

Kindermann et al. [28] showing that NMC and LiMn relaxation rates are within 2% of each

other. This suggests that the NMC results shown in this work may be applicable to other

- 100 -
 Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

chemistries as well. LFP showed a faster relaxation rate compared to NMC and LFP which

suggests rest times are less important for LFP.

The focus of this paper is on the relaxation effect of single cells; however, similar relaxation

behavior can likely be observed for battery modules with cells in parallel, and therefore

equal voltages. Several new methods for BMS-compatible EIS allow for cell voltage

measurement as described in Section 3.1.1 [20], [21]. Combined with such methods, the

findings in this paper are intended to provide BMS engineers with suitable operating points

at which to perform the EIS measurements under different conditions. Furthermore, BMS

generated EIS may require filtering and validation for which the Voigt filter approach

introduced here is suitable.

3.6 CONCLUSION

EIS measurements obtained after C-rate pulses and short rest times must be compensated

and filtered due to drift caused by the rapid change of the battery voltage. In this study, EIS

was measured for batteries of different SoC and at different temperatures, with short rest

times after different discharge pulses by utilizing drift compensation methods and

validating EIS results using the Kramer-Kronigs. Furthermore, a method was proposed to

use the Voigt-circuit as a filter to further improve EIS measurements at short rest times. It

was found that using this method valid EIS results could be obtained for EIS measurements

as low as 5 minutes as long as temperatures remain at room temperature or above.

Furthermore, it was found that the effect of different C-rate pulses disappears after 30

minutes of rest for batteries at 90% SoC and above 25°C. However, batteries at SoCs below

- 101 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

90% as well as batteries at temperatures of 10°C require longer rest times to eliminate the

effect of large C-rates.

ACKNOWLEDGEMENT

We thank Cadex Electronics for their facilities, support and expertise.

REFERENCES
[1] M. A. Hannan, M. S. H. Lipu, A. Hussain, and A. Mohamed, “A review of lithium-
ion battery state of charge estimation and management system in electric vehicle
applications: Challenges and recommendations,” Renewable and Sustainable
Energy Reviews, vol. 78. pp. 834–854, 2017.
[2] R. Xiong, J. Cao, Q. Yu, H. He, and F. Sun, “Critical Review on the Battery State of
Charge Estimation Methods for Electric Vehicles,” IEEE Access, vol. 6, pp. 1832–
1843, 2018.
[3] M. U. Cuma and T. Koroglu, “A comprehensive review on estimation strategies used
in hybrid and battery electric vehicles,” Renew. Sustain. Energy Rev., vol. 42, pp.
517–531, 2015.
[4] L. Ungurean, G. Cârstoiu, M. V Micea, and V. Groza, “Battery state of health
estimation : a structured review of models , methods and commercial devices,” no.
July 2016, pp. 151–181, 2017.
[5] R. Ahmed, M. El Sayed, I. Arasaratnam, J. Tjong, and S. Habibi, “Reduced-Order
Electrochemical Model Parameters Identification and State of Charge Estimation for
Healthy and Aged Li-Ion Batteries—Part II: Aged Battery Model and State of
Charge Estimation,” IEEE J. Emerg. Sel. Top. Power Electron., vol. 2, no. 3, pp.
678–690, 2014.
[6] X. Lai et al., “Co-estimation of state of charge and state of power for lithium-ion
batteries based on fractional variable-order model,” J. Clean. Prod., vol. 255, no. 11,
pp. 10319–10329, 2020.
[7] S. Habibi, “The Smooth Variable Structure Structure Filter,” Proc. IEEE, vol. 95,
no. 5, 2007.
[8] R. Ahmed et al., “Model-Based Parameter Identification of Healthy and Aged Li-
ion Batteries for Electric Vehicle Applications,” SAE Int. J. Altern. Powertrains, vol.
4, no. 2, pp. 233–247, 2015.
[9] S. A. Gadsden, M. Al-Shabi, and S. R. Habibi, “Estimation Strategies for the
Condition Monitoring of a Battery System in a Hybrid Electric Vehicle,” ISRN
Signal Process., vol. 2011, no. 1, 2011.
[10] S. A. Gadsden, M. Al-Shabi, I. Arasaratnam, and S. R. Habibi, “Combined cubature
Kalman and smooth variable structure filtering: A robust nonlinear estimation
strategy,” Signal Processing, vol. 96, no. PART B, pp. 290–299, 2014.
[11] S. A. Gadsden, S. R. Habibi, and T. Kirubarajan, “A novel interacting multiple
model method for nonlinear target tracking,” 13th Conf. Inf. Fusion, Fusion 2010,
2010.
[12] S. A. Gadsden, M. El Sayed, and S. R. Habibi, “Derivation of an optimal boundary
layer width for the smooth variable structure filter,” Proc. Am. Control Conf., pp.
- 102 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

4922–4927, 2011.
[13] M. Ismail, R. Dlyma, A. Elrakaybi, R. Ahmed, and S. Habibi, “Battery state of
charge estimation using an Artificial Neural Network,” 2017 IEEE Transp. Electrif.
Conf. Expo, ITEC 2017, pp. 342–349, 2017.
[14] C. Vidal, P. Malysz, P. Kollmeyer, and A. Emadi, “Machine Learning Applied to
Electrified Vehicle Battery State of Charge and State of Health Estimation: State-of-
the-Art,” IEEE Access, vol. 8, pp. 52796–52814, 2020.
[15] Y. Li et al., “A quick on-line state of health estimation method for Li-ion battery
with incremental capacity curves processed by Gaussian filter,” J. Power Sources,
vol. 373, no. November 2017, pp. 40–53, 2018.
[16] M. Berecibar, I. Gandiaga, I. Villarreal, N. Omar, J. Van Mierlo, and P. Van Den
Bossche, “Critical review of state of health estimation methods of Li-ion batteries
for real applications,” Renew. Sustain. Energy Rev., vol. 56, pp. 572–587, 2016.
[17] A. Lasia, Electrochemical Impedance Spectroscopy EIS, and Corrosion. New York:
Springer, 2014.
[18] J. Meng et al., “An Overview and Comparison of Online Implementable SOC
Estimation Methods for Lithium-Ion Battery,” IEEE Trans. Ind. Appl., vol. 54, no.
2, pp. 1583–1591, 2018.
[19] N. Meddings et al., “Application of electrochemical impedance spectroscopy to
commercial Li-ion cells: A review,” Journal of Power Sources, vol. 480. Elsevier
B.V., 31-Dec-2020.
[20] E. Din, C. Schaef, K. Moffat, and J. T. Stauth, “A scalable active battery
management system with embedded real-time electrochemical impedance
spectroscopy,” IEEE Trans. Power Electron., vol. 32, no. 7, pp. 5688–5698, 2017.
[21] X. Wei, X. Wang, and H. Dai, “Practical on-board measurement of lithium ion
battery impedance based on distributed voltage and current sampling,” Energies, vol.
11, no. 1, 2018.
[22] N. Lohmann, P. Haussmann, P. Wesskamp, J. Melbert, and T. Musch, “Employing
Real Automotive Driving Data for Electrochemical Impedance Spectroscopy on
Lithium-Ion Cells,” SAE Int. J. Altern. Powertrains, vol. 4, no. 2, pp. 308–317, 2015.
[23] P. Kollmeyer, A. Hackl, and A. Emadi, “Li-ion battery model performance for
automotive drive cycles with current pulse and EIS parameterization,” 2017 IEEE
Transp. Electrif. Conf. Expo, ITEC 2017, pp. 486–492, 2017.
[24] C. Pastor-Fernández, K. Uddin, G. H. Chouchelamane, W. D. Widanage, and J.
Marco, “A Comparison between Electrochemical Impedance Spectroscopy and
Incremental Capacity-Differential Voltage as Li-ion Diagnostic Techniques to
Identify and Quantify the Effects of Degradation Modes within Battery Management
Systems,” J. Power Sources, vol. 360, pp. 301–318, 2017.
[25] U. Westerhoff, T. Kroker, K. Kurbach, and M. Kurrat, “Electrochemical impedance
spectroscopy based estimation of the state of charge of lithium-ion batteries,” J.
Energy Storage, vol. 8, pp. 244–256, 2016.
[26] U. Krewer, F. Röder, E. Harinath, R. D. Braatz, B. Bedürftig, and R. Findeisen,
“Review—Dynamic Models of Li-Ion Batteries for Diagnosis and Operation: A
Review and Perspective,” J. Electrochem. Soc., vol. 165, no. 16, pp. A3656–A3673,
2018.
[27] L. Deverakonda and T. Hu, “Effects of rest time on discharge response and
equivalent circuit model for a lead-acid battery,” J. Power Sources, vol. 282, pp. 19–
27, 2015.
[28] F. M. Kindermann, A. Noel, S. V. Erhard, and A. Jossen, “Long-term equalization
effects in Li-ion batteries due to local state of charge inhomogeneities and their
- 103 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

impact on impedance measurements,” Electrochim. Acta, vol. 185, pp. 107–116,

2015.
[29] A. Barai, G. Chouchelamane, Y. Guo, A. McGordon, and P. Jennings, “A study on
the impact of lithium-ion cell relaxation on electrochemical impedance
spectroscopy,” J. Power Sources, vol. 280, no. Special Issue, pp. 74–80, 2015.
[30] W. Waag, S. Käbitz, and D. U. Sauer, “Experimental investigation of the lithium-
ion battery impedance characteristic at various conditions and aging states and its
influence on the application,” Appl. Energy, vol. 102, pp. 885–897, 2013.
[31] H. Wang, M. Tahan, and T. Hu, “Effects of rest time on equivalent circuit model for
a li-ion battery,” Proc. Am. Control Conf., vol. 2016-July, pp. 3101–3106, 2016.
[32] V. Zinth et al., “Inhomogeneity and relaxation phenomena in the graphite anode of
a lithium-ion battery probed by in situ neutron diffraction,” J. Power Sources, vol.
361, pp. 54–60, Sep. 2017.
[33] M. Messing, T. Shoa, and S. Habibi, “Lithium-Ion Battery Relaxation Effects,” in
ITEC 2019, 2019.
[34] Zahner Messsysteme, “EIS - Electrochemical Impedance Sepctroscopy,” 2019.
[35] B. A. Boukamp, “A Linear Kronig-Kramers Transform Test for Immittance Data
Validation,” J. Electrochem. Soc., vol. 142, no. 6, pp. 1885–1901, 1995.
[36] M. Schönleber, D. Klotz, and E. Ivers-Tiffée, “A Method for Improving the
Robustness of linear Kramers-Kronig Validity Tests,” Electrochim. Acta, vol. 131,
pp. 20–27, 2014.

- 104 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

Chapter 4: Estimating Battery State of Health using

Electrochemical Impedance Spectroscopy and the
Relaxation Effect

Marvin Messing1,2, Tina Shoa2, Saeid Habibi1

1
Department of Mechanical Engineering, McMaster University, Hamilton, ON, Canada,
2
Cadex Electronics, Richmond, BC, Canada
This paper is published in Journal of Energy Storage, vol. 43, Article 103210, 2021, doi:
10.1016/j.est.2021.103210.

ABSTRACT

Among the most important tasks of a Battery Management System (BMS) are State of

Charge (SoC) and State of Health (SoH) estimation. Many SoH estimation techniques are

available, each with their advantages and drawbacks. These include methods based on a

technique known as Electrochemical Impedance Spectroscopy (EIS). This technique

provides detailed information about the battery’s state of health but requires long rest times

to prevent the battery relaxation effect from impacting the EIS measurement. In this paper

EIS is shown to be able to track the short-term relaxation effect for batteries of different

SoH. A SoH estimation method is proposed which combines fractional order impedance

modeling and short-term relaxation effects with EIS characterization for rapid SoH

determination. This empirical method is demonstrated to have an average SoH estimation

error of less than 1%. As new methods arise to simplify EIS hardware requirements for real

time applications, the proposed method offers a new way of utilizing EIS for SoH

estimation.

- 105 -
 Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

4.1 INTRODUCTION

In this paper an empirical model is proposed for battery State of Health (SoH)

estimation. The model is developed based on the battery relaxation effect and

Electrochemical Impedance Spectroscopy (EIS) measurements fitted with a fractional

order impedance model. The rate of change of a combined model resistance is shown to be

a repeatable quantity and is utilized here for SoH estimation. This section summarizes the

motivation and technical challenges, provides a review of recent and relevant literature,

presents the main paper contributions, and provides an outline of the rest of the paper.

4.1.1 MOTIVATION AND TECHNICAL CHALLENGES

Lithium-ion Batteries (LiB) are used in many applications including Electric

Vehicles (EV). Any multi-cell LiB pack requires a Battery Management System (BMS) to

optimize the performance of the battery pack and to ensure each cell remains within safe

operating limits [1]. To do these tasks, the BMS must estimate the State of Charge (SoC)

and SoH of the cells, since these states cannot be measured. The SoC indicates the charge

left in the battery before it must be charged. The SoH indicates the capacity of the cell, i.e.

how much energy the cell can provide after it is fully charged, compared to when it was

new. The accuracy of SoC algorithms depends on the SoH of the battery and many different

methods have been investigated to improve the accuracy of SoH [2–7]. A group of SoH

estimation methods are based on diffusion processes taking place inside the battery

electrodes during rest time, known as the relaxation effect. As a result of the relaxation

effect, the battery voltage changes over time and the rate of change of the voltage can be

related to the SoH of the battery. Another method commonly used in laboratory settings to
- 106 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

investigate SoH is EIS, a non-destructive and time efficient technique that provides detailed

insight into the condition of a battery cell. This is achieved by applying small sinusoidal

signals to the battery over a wide frequency range and measuring the battery response. The

resulting response signal is then analyzed in the frequency domain by considering

magnitude, phase as well as real and imaginary components of the response at each

frequency. Although a BMS would greatly benefit from this technique, the required

hardware is complex and costly [8]. Nonetheless, there is increasing research effort to

develop solutions to make EIS available for BMS applications [9–12] for use in EV

applications. For example, Carkhuff et al. [10] designed multiplexer hardware to deliver an

EIS signal to each cell in a battery pack. More recently, Locorotondo et al. [13] designed a

wireless charging circuit with built-in EIS measurement capability. In this paper, EIS is

used to track the relaxation effect to develop an empirical SoH estimation model. In the

next section, literature relevant to SoH estimation and the relaxation effect is reviewed.

4.1.2 RELAXATION EFFECT AND SOH ESTIMATION

When batteries are allowed to rest, slow diffusion processes take place which slowly

change the voltage and impedance of the battery [14]. This relaxation effect was used for

SoH estimation by several studies. Qian et al. [15] proposed an SoH estimation strategy

based on voltage relaxation for 10 minutes of rest after full charge and full discharge. They

used a second order ideal Equivalent Circuit Model (ECM) which was fit to the voltage

curve and the model parameters were related to SoH. Fang et al. [16] proposed a similar

model using a third order ECM and 2 hours of rest. Baghdadi et al. [17] correlated the

- 107 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

voltage after 30 minutes of rest after full charge to SoH. They used a linear model to

estimate the SoH with model parameters adjusted for different operating conditions. Ran

et al. [18] proposed a data driven model to cluster batteries by SoH using a pulse profile

which included relaxation phases. They find a clustering accuracy of 88%. EIS was used

to characterize the relaxation effect in several studies [14], [19–23]. Several studies also

use EIS for SoH estimation directly. Shabbir et al. [24] built a SoH prediction model based

on multiple features of the Nyquist plot with an estimation error of around 10%. Guha et

al. [25] reconstructed EIS from constant discharge pulses and modeled the change of the

internal resistance from the EIS results with a third order polynomial. Kim et al. [26]

measured EIS with a constant current bias and used linear regression to model the change

in SoH with EIS impedance at 250 Hz. Gismero et al. [27] used EIS to track the calendar

age of batteries and used third order ECM parameters to model SoH with a 2.5% error.

Locorotondo et al. [28] used a Pseudo-Random Binary Sequence (PRBS) signal (instead of

multi-sine) to generate EIS results. They a limited number of cells between 100% and 50%

SoH and proposed an impedance clustering method for SoH detection. They state a required

rest time of 1 hour before measurements can be obtained.

The studies discussed above either use the relaxation effect in time domain or EIS

measurements to estimate SoH. Each study imposes requirements such as long

measurement times (hours), and having the battery to be fully charged or discharged, or

applying a fixed bias or discharge current during measurement. In addition, the SoH

estimation errors seem to be greater than 2% in most cases. In this study, the combination

- 108 -
 Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

of the relationship between the battery relaxation effect and its SoH, complemented by EIS

measurements are used to develop a new SoH estimation strategy.

4.1.3 CONTRIBUTIONS

This paper presents the following original contributions and conclusions: (1) An

empirical SoH estimation method is proposed by combining EIS measurements, fractional

order battery modeling, and the short-term relaxation effect; (2) The proposed SoH

estimation method is experimentally applied in an aging study and is shown to have an

average estimation error of less than 1% SoH.

4.1.4 PAPER OUTLINE

The paper is organized as follows. Section 4.2 describes the experiments conducted

to characterize the relaxation effect with EIS and to obtain aged batteries. Section 4.3

introduces the battery relaxation effect and presents a conceptual view of the underlying

physical phenomenon and its impact on battery impedance. Section 4.4 introduces the

fractional order battery model and describes the modeling approach. Section 4.5 introduces

the SoH estimation method followed by a discussion of its performance in Section 4.6. The

concluding remarks are provided in Section 4.7.

4.2 EIS CHARACTERIZATION AND AGING EXPERIMENTS

The work presented in this paper includes extensive experiments conducted on

Samsung INR2170-50E lithium nickel manganese cobalt (LiNiMnCoO2 or NMC) batteries

of different SoH. The experiments were conducted to provide development and validation

datasets for the proposed SoH estimation method. The experiments consisted of discharge
- 109 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

pulses of different currents used to induce relaxation behavior with different relaxation

rates. Each pulse was followed by a 3 hours rest period, during which EIS was measured at

regular intervals starting after 1 minute of rest. Discharge currents between 1C and 5C were

tested, with each pulse held for 15 seconds. EIS was measured with a Bio-logic SP150

potentiostat for the frequency range of 0.1 to 800 Hz, with a 10 mV amplitude with two

sine waves per period, 6 points per decade, and with the drift compensation setting enabled

[29]. The EIS measurement duration with these settings is 64 seconds for each full scan.

This time interval is considered as part of any rest time intervals. In this study, all results

were obtained at 90% SoC. This SoC target was obtained by first fully charging the battery

according to the manufacturer’s recommendation. Next, batteries were discharged at 0.2C

and the charge counted using coulomb-counting and by subtracting the charge lost during

the pulse. This way, the battery SoC is equal to 90% at the beginning of the rest period.

The temperature was held at 25°C using a Testequity 1007C thermal chamber. A schematic

of the test setup is shown in Figure 4-1.

Figure 4-1: Schematic of experimental setup and battery cell.

- 110 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

Table 4-1: SoH and Test Conditions for 14 Battery Cells at 90% SoC and 25°C.
SoH (%) C-rates Repeats
99.96 1,2,3,4,5 3
98.26 5 1
97.61 5 1
96.00 1,2,3,4,5 1
92.00 5 1
91.27 5 3
90.69 1,2,3,4,5 1
86.54 5 1
86.22 5 1
85.05 1,2,3,4,5 1
83.44 5 1
83.03 5 3
81.69 1,2,3,4,5 1
78.41 5 1

The protocol discussed above was applied to 14 battery cells of different SoH as

shown in Table 4-1. As indicated in the “C-rates” column, selected cells between 100%

and 80% SoH were subjected to discharge pulses of 5 different C-rates to test the impact of

varying relaxation severity on EIS measurements and to investigate how this changes with

various battery SoH. All cells were tested with 5C discharge pulses to form the dataset for

the development and testing of the proposed SoH estimation method. Some cells at

different SoHs were tested three times at the same condition to ensure the repeatability of

the results. Three repeats of the 5C, 15 seconds discharge pulse are shown in Figure 4-2.

- 111 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

Figure 4-2: Three repeats of a 5C, 15s discharge pulse at 90% SoC, 100% SoH.

The cells were aged with simple charge/discharge cycle aging at a constant

temperature of 25°C. The cells were charged at C/2 with the standard constant-current,

constant-voltage (CC-CV) method to 4.2V with a target current of C/20. The cells were

discharged at 1C to 2.5V as recommended by the manufacturer.

The SoH of the batteries was determined at 25°C by first fully charging the battery

(at C/2 with the CC-CV method to 4.2V and C/20 cut-off current), followed by a 1 hour

rest, followed by a discharge at C/5 to 2.5V. This procedure yields a measured capacity 𝐶𝑚

value via coulomb counting. The capacity value of 4.9 mAh is stated as the rated capacity

𝐶𝑟 of the battery according to the manufacturer datasheet and the SoH is calculated using

Equation 4-1.

𝐶𝑚
𝑆𝑜𝐻 (%) = × 100 (4-1)
𝐶𝑟

- 112 -
 Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

4.3 RELAXATION EFFECT AND IMPEDANCE

In this section, the battery relaxation effect and its possible origin are discussed

based on EIS measurements and conceptual explanation of physical processes inside the

battery. The relaxation effect can most easily be observed by monitoring the Open Circuit

Voltage (OCV) of the battery after applying a charge or discharge pulse. The OCV slowly

increases over several hours before reaching a stable value. This happens because of ion

concentration gradients inside the electrodes causing charge imbalances which are resolved

by slow diffusion processes in the absence of an external driving force. Kindermann et al.

[29] presented a mind-model showing the different kinds of diffusion processes, which are

still on-going even after the voltage has stabilized. They show that the relaxation effect not

only impacts voltage, but battery impedance as well. Using EIS measurements, they

detected impedance changes for as long as 40 hours of rest time.

In this paper, the short term relaxation effect for new and aged battery cells is

considered. Figure 4-3 shows the Nyquist plots for EIS measurements obtained from

batteries at 100%, 90% and 80% SoH. At each SoH, the Nyquist plots from EIS

measurements obtained at rest times of 5, 10 and 15 minutes after a discharge pulse are also

shown. Each set of three EIS measurements track the relaxation effect in a distinctive way,

with decreasing distance between the mid to low frequency range of the Nyquist plots for

the same total time interval of 15 minutes. The rate at which the distance between Nyquist

plots changes decreases with SoH. This behavior forms the basis for the SoH estimation

method proposed in this paper.

- 113 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

Figure 4-3: Evolution of Nyquist plot with rest time and SoH.

The change of impedance with rest time can be explained by the change in charge

concentration inside the electrodes over time. Figure 4-4 shows the Nyquist plots of the

EIS results as they change with rest time and battery age. The same plot is shown three

times with a specific dataset highlighted corresponding to the rest time, one for new and

one for aged batteries in each case. Also shown in Figure 4-4 are diagrams showing

conceptual views of the charge concentration inside the anode (similar behaviour also

occurs on the cathode). During relaxation after a discharge pulse the ions diffuse through

- 114 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

the Solid Electrolyte Interface (SEI) into the electrode. However, at first, most of the charge

is located close to the surface of the electrode [30]. As the rest time increases, ions move

further into the electrode, decreasing the readily available surface charge, and increasing

charge transfer impedance, 𝑍𝐶𝑇 , as a result. In the case of aged cells, 𝑍𝐶𝑇 starts out larger

due to an increased SEI and reduced electrode material. The rate of change of 𝑍𝐶𝑇 during

relaxation is slowed also due to the increased SEI impedance and because it takes longer

for ions to arrive at active electrode material. This behaviour for new and aged battery cells

is conceptually shown in the diagrams of Figure 4-4 as it relates to EIS measurements.

- 115 -
 Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

Figure 4-4: Short-term relaxation effect impedance evolution tracked with EIS data (plots highlighted
according to rest time) and conceptually pictured with battery half-cell diagrams.

4.4 BATTERY MODELING

EIS results are best quantified by fitting battery models to the data. ECMs are

frequently used because of their simplicity and because the circuit elements used somewhat

represent the physical behavior of the battery [30]. The simplest model is a single resistor,

which models the internal, ohmic resistance of the battery. However, since the chemical
- 116 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

reactions inside batteries are also time dependent, additional elements are used to model

processes of different time constants, such as charge transfer of both electrodes, impedance

of the SEI and diffusion inside the electrodes. These processes can be modeled using

multiple branches with a resistor in parallel with a capacitor (R||C) [31]. For EIS modeling,

these R||C branches produce ideal semi-circles, however, the EIS results for batteries are

non-ideal, and semi-circles only provide approximate fits.

As shown in Figure 4-5a, the SEI impedance usually manifests as a small semi-

circle in the high to mid frequency range of the EIS measurement. The SEI is a biproduct

of undesirable chemical reactions between the negative electrode and the electrolyte. At the

same time, after its formation, the SEI layer prevents further reactions between electrode

and electrolyte. However, the SEI contributes to the internal resistance of the battery and

further increases this resistance, as the SEI layer grows with battery age. A larger semi-

circle in the mid frequency range is caused by the chemical reactions necessary for lithium

ions to be transferred between electrodes and electrolyte (charge transfer). Finally, lithium

ions which enter the electrodes slowly diffuse deeper into the electrode materials. This

diffusion appears as a line at a 45° angle in the low frequency range of the EIS measurement

[32]. Higher numbers of R||C branches can achieve improved fits, however, at the cost of

adding many model parameters, and possibly overfitting to the data. Near perfect fits can

be obtained using a rendition of capacitors known as constant phase elements (CPE), and

CPEs in parallel with resistors (𝑅||𝑄, 𝛼), known as ZArc elements. CPE and ZArc elements

produce much better fits since they are non-ideal circuit elements, at the cost of only one

additional parameter (𝛼) compared to ideal capacitors. Furthermore, the fitting process can
- 117 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

be controlled more easily since there is low ambiguity in the fitting results. In contrast,

when using ideal circuit elements, two fits with different parameters may be accepted, due

to the large inherent error between the fit and the data. This means that ideal models cannot

distinguish between small changes in the EIS results, making them unsuitable to model the

subtle changes in charge transfer and diffusion impedance occurring as a result of the

relaxation effect.

Figure 4-5: Nyquist plots of low and high SoH EIS measurements with model fits (a), fractional order
impedance model (b).

- 118 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

Figure 4-5a shows two EIS datasets, one for a battery at 100% SoH, and the other

for a battery at 90% SoH. The EIS results at lower SoH show significant increases in SEI

layer and charge transfer impedances. The equivalent circuit structure used to model EIS

results in this work is also shown in Figure 4-5b. The circuit consists of an inductance (𝐿)

to model inductive effects from cell windings and cables (used to connect the potentiostat

to the battery), a resistance (𝑅0 ) modeling ohmic resistance, a ZArc element (𝑅1 ||𝑄1 , 𝛼1 ) to

model the SEI layer impedance, a second ZArc element (𝑅2 ||𝑄2 , 𝛼2 ) to model charge

transfer impedance, and finally a CPE (𝑄3 , 𝛼3 ) to model diffusion. The impedance 𝑍𝑐 of

this circuit is defined by Equation 4-2 for input frequency 𝜔. The time constants are defined
1⁄
by τi = (R i Q i ) αi [26].

𝑅1 𝑅2
𝑍𝑐 (𝜔) = 𝑗𝜔𝐿 + 𝑅0 + +
1 + 𝑅1 𝑄1 (𝑗𝜔𝜏1 ) 1 1 + 𝑅2 𝑄2 (𝑗𝜔𝜏2 )𝛼2
𝛼
(4-2)
1
+
𝑄3 (𝑗𝜔)𝛼3
Further to the above, the circuit in Figure 4-5b provides good fits for both healthy

and aged batteries as also shown in Figure 4-5a. Model fitting was done using Particle

Swarm Optimization (PSO) with the objective function 𝑆 shown in Equation 4-3:

𝜔𝑓

𝑆 = ∑ {[𝑍 ′ (𝜔) − 𝑍𝑐′ (𝜔)]2 + [𝑍 ′′ (𝜔) − 𝑍𝑐′′ (𝜔)]2 } (4-3)

𝜔=𝜔𝑖

This objective function computes the square of the difference between the real value

of the measured impedance 𝑍′ and the real value of the calculated impedance 𝑍𝑐′ for

frequencies 𝜔 between the initial frequency 𝜔𝑖 and the final frequency 𝜔𝑓 . The same is

done for the imaginary parts of the measured and calculated impedance (𝑍′′ and 𝑍𝑐′′
- 119 -
 Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

respectively). The resulting square differences are summed to yield a single objective

function for the PSO algorithm to optimize.

4.5 SOH ESTIMATION METHOD

This section first introduces the empirical SoH model followed by a description of

the fitting process used to obtain consistent fractional order model fits to the EIS data.

4.5.1 EMPIRICAL SOH MODEL

In this section, a new SoH estimation method is proposed based on the rate of

change of an equivalent resistance with relaxation time. This equivalent resistance 𝑅𝑒𝑞 , as

shown in Equation 4-4, is a summation of the branch resistances of the impedance model

of Equation 4-2 introduced in the previous section.

𝑅𝑒𝑞 = 𝑅1 + 𝑅2 (4-4)

- 120 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

Figure 4-6: Evolution of 𝑹𝒆𝒒 with rest time at 100% SoH (a), log-linear trend of 𝑹𝒆𝒒 for rapid relaxation
region (b).

Figure 4-6a shows how 𝑅𝑒𝑞 changes for models fitted to EIS data measured at

different rest times after a 5C, 15 second discharge pulse. The initial rate of change of R eq

with rest times at 5, 10 and 15 minutes can be correlated to SoH. To find the rate of change,

a semi-log-linear fit defined by Equation 4-5 and shown in Figure 4-6b was used.

𝑅𝑒𝑞 (𝑡) = 𝑚 × 𝑙𝑜𝑔(𝑡) + 𝑦𝑟 (4-5)

- 121 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

Here, Req is given by the rate of change 𝑚, y-intercept 𝑦𝑟 and relaxation time t.

The performance of this model is discussed in Section 4.6. The correlation of rate

of change 𝑚 and SoH follows a two-phase exponential decay model shown in Equation 4-

6.

𝑠𝑜ℎ(𝑚) = 𝑎1 exp(𝑚 × 𝑏1 ) + 𝑎2 exp(𝑚 × 𝑏2 ) (4-6)

The parameters 𝑎1 and 𝑏1 describe an initial fast decay from 100% SoH to 95% SoH,

followed by a slower decay from 95% SoH to 80% SoH described by 𝑎2 and 𝑏2 . The

parameter values and performance of this model are discussed in detail in Section 4.6.

4.5.2 MODEL FITTING PROCEDURE

The following model fitting procedure was found to be necessary to guide the PSO

algorithm to find repeatable solutions. First, the impedance model (Equation 4-2) is fit to

the EIS dataset measured after the longest rest time (15 minutes). Further to the model

presented in Equation 4-2, Table 4-2 shows an example of model parameters found for the

100% SoH case. Also shown are the upper and lower bounds used with the PSO algorithm.

For a new EIS dataset at unknown battery state, the upper or lower bounds can be used as

initial values for the PSO algorithm provided the data comes from the same battery model.

Table 4-2: Model Parameter Bounds and Example Values at 90% SoC, 100% SoH, 25°C and 15
minutes Rest

L (𝝁F) R0 (m𝛀) R1 (m𝛀) 𝝉𝟏 (ms) R2 (m𝛀) 𝝉𝟐 (ms) Q3 𝜶𝒊

Upper Bound 0.7 28 40 480 6 30 5 1
Lower Bound 0.4 20 0 0 2 0 1 0
Example 0.58 22.2 0.41 2 4.8 0.5 4.04 1, 0.58, 0.52

- 122 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

After the 15-minute rest EIS fit is obtained, these upper and lower bounds are

updated such that 𝐿, 𝑅0 , and 𝛼3 are fixed to the value of the 15-minute rest fit. Furthermore,

𝑅1 and 𝑅3 from the 15-minute rest EIS fit are set to be the upper bounds of the next fit.

With these new limits in place, 𝑅1 , 𝜏1, 𝑅2 , 𝜏2 and 𝑄3 are fit to the 10-minute rest EIS.

Finally, the limits are updated once more to fix the upper bounds of R 1 and R2 to their

values at 10 minutes rest, and a fit for 5 minutes rest EIS is obtained. The parameters 𝐿, 𝑅0 ,

and 𝛼3 are not affected by the relaxation effect [33], so they can be fixed once an initial fit

is found. This procedure is shown in Figure 4-7 together with the steps of the empirical

SoH model to show a summary of the complete SoH estimation method.

- 123 -
 Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

Figure 4-7: Procedure to obtain SoH estimate from EIS data at different rest times.

4.6 RESULTS AND DISCUSSION

In this section the performance of the proposed SoH estimation method is discussed.

In Section 4.5 the general method was described for one SoH example dataset. With this

method, the rate of change m of the combined branch resistance 𝑅𝑒𝑞 for EIS data recorded
- 124 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

between 5 and 15 minutes of rest after a 5C, 15s discharge pulse can be correlated to SoH.

The rate of change m is found using a semi-log linear fit to 𝑅𝑒𝑞 over rest time defined by

Equation 4-5. Figure 4-8a shows how 𝑅𝑒𝑞 evolves with rest time for a subset of six different

SoHs between 100% and 80%. The semi-log-linear fits are shown in Figure 4-8b for each

of the six datasets.

- 125 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

Figure 4-8: Short term evolution of Req with rest time (a) and semi-log-linear fits to the initial rate of
change of Req (b).

Table 4-3 shows the slopes and intercepts for each case as well as the fitting

statistics. The R squared and adjusted R squared values are all between 0.994 and 1,

suggesting the linear fit is suitable to describe the datasets. Figure 4-8b and Table 4-3 show

that the slopes of the log-linear fits increase as SoH decreases, but not in a linear fashion.

Instead the data follows an exponential decay with two time constants.

Table 4-3: Semi-log Fit (Equation 5) Results and Statistics

SoH (%) 100 96 92 90 85 80

Y intercepts 3.68 2.95 8.24 5.90 10.76 20.63
Slope, m 0.54 0.55 0.78 0.96 2.24 7.884
𝑅2 0.9967 0.9999 0.9992 0.9986 0.9998 0.9971
2
𝑅𝑎𝑑𝑗 0.9934 0.9998 0.9984 0.9972 0.9996 0.9942
RMSE 0.017 0.003 0.012 0.020 0.018 0.236

Figure 4-9 shows the calculated slopes m vs. measured, true SoH for the entire

experimental dataset. Also shown in this figure is the double exponential model given by

Equation 6 fitted to the data, as well as the 95% prediction band (dotted lines), indicating

where 95% of unknown data will fall. The first time constant of this model (Equation 4-6)

describes the initial steep drop between 100% and 90% SoH. The second time constant

describes the much slower, nearly linear decay between 90% and 78% SoH.

- 126 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

Figure 4-9: Req short term relaxation evolution slopes and SoH.

The parameters of the model (Equation 4-6) are shown in Table 4-4, with statistics
2
showing a high 𝑅𝑎𝑑𝑗 equal to 0.98 and a RMSE of less than 1% SoH. The repeatability of

the SoH estimation method was assessed for one dataset at above 90% and another below

90%. Each dataset consists of three repeats of the experimental EIS measurement.

Table 4-4: Double Exponential Model (Equation 6) Fit Results and Statistics

𝒂𝟏 𝒃𝟏 𝒂𝟐 𝒃𝟐 𝑹𝟐𝒂𝒅𝒋 RMSE (% SoH)

22.16 -1.67 88.48 -0.0156 0.98 0.82

Table 4-5: Repeatability of the SoH Estimation Method

Data Set
Variable True SoH (%) mean SD
1 2 3
91.27 1.00 1.09 1.19 1.09 0.095
Rate of change, 𝑚
83.03 3.58 3.67 4.30 3.85 0.394
91.27 91.20 90.58 90.05 90.61 0.573
Estimated SoH (%)
83.03 83.61 83.44 82.41 83.15 0.651

Figure 4-10a shows the estimated SoH with mean and SD error bars for the repeat

sets, as well as the estimated SoHs for the entire dataset. The corresponding numbers are
- 127 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

shown in Table 4-5, where the first two rows show data for m and the last two rows data

for the estimated SoH. Above 90% SoH, the SD of m is small and the corresponding SoH

estimate has a small SD as well. Below 90% SoH, the SD of m is much larger, however,

the SD of the SoH estimate is still small. This is because of the different time constants in

the exponential model, causing the estimated SoH to be sensitive to changes in m for the

high SoH, fast decay portion of the model, but much less sensitive for the low SoH, slow

decay portion of the model. In general, the estimation error is consistently lower for the

low SoH portion of the model as shown in Figure 4-10b. This model has an RMSE SoH

estimation error of 0.82% with an overall maximum error of 1.74%. Below 90% SoH, the

maximum error is much lower, at 0.62%.

- 128 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

Figure 4-10: Estimated SoH vs. True SoH and slope repeatability (a), SoH Estimation Error and
RMSE (b).

The proposed method is demonstrated at 90% SoC, 25°C, with a relaxation effect

induced by a 15s, 5C discharge pulse and a measurement period of up to 15 minutes.

However, the method can be adjusted by accounting for SoC and temperature in the model,

as long as the relaxation effect does not change significantly. 𝑅𝑒𝑞 slopes obtained for a
- 129 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

100% SoH battery for different SoCs and temperatures are shown in Figure 4-11. SoC

values of 70% and 50% fall within one SD (assumed to be 0.095 from Table 4-5) of the

90% SoC used for the proposed method. Similarly, increasing the temperature to 35°𝐶

produces a slope value also within one SD of the 100% SoH (and 25°𝐶) datapoint shown

in Table 4-3. This suggests that the relaxation rate of change behaves similar across a wide

range of operating conditions. However, lowering the temperature to 10°𝐶 results in a

significantly different slope value. Therefore, to be able to test SoH at lower temperatures,

a representative number of batteries must be re-tested at different SoH to understand the

impact of low temperature on Equation 4-6.

Figure 4-11: 𝑹𝒆𝒒 slopes for different SoCs and Temperatures for battery at 100% SoH.

Discharge pulses similar to 5C can also be used with some model adjustment,

however, 5C was found to provide the most consisted results. Reducing the C-rate may

increase the SoH estimation error. In this study, EIS was measured at 5, 10 and 15 minutes,
- 130 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

however, due to the semi-log linear nature of the rate of changed of the combined

resistance, any three points between 5 and 15 minutes can be used to obtain the relaxation

rate. Each EIS measurement for the chosen frequency range has a duration of less than 2

minutes, therefore, the measurement time could be reduced to less than 10 minutes overall.

Further reduction in measurement time may be achieved by utilizing multi-sinewave EIS

methods [34], which allow the sine waves at different frequencies to be stacked into a single

signal, instead of running them consecutively as done in this study. The EIS measurement

time is then limited only by the lowest frequency. Thus, the upper end of the measurement

time of the proposed SoH estimation method should be further investigated to see if it can

be reduced, possibly well below 10 minutes.

Compared to voltage-based methods, introduced in Section 4.1.2, which utilize the

relaxation effect, the proposed method shows gains in accuracy and testing time. Table 4-6

summarizes the method comparison. The method requiring 10 minutes of rest time [14] is

limited to full charge or full discharge conditions and accuracy was not stated, possibly due

to the low number of cells tested. The method requiring 30 minutes of rest time after full

charge [17] also did not state SoH accuracy. A promising method was proposed by [16]

with a rest time of 20 minutes and a wide SoC range (requiring building of lookup tables

at different SoCs). They showed an accuracy of 1.8% SoH, although with only 6 cells

tested.

- 131 -
 Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

Table 4-6: Method Comparision

Number of Aging Procedure
Rest Time SoH
Reference SoC Cells (Temperature,
(min) Accuracy
Tested Charge)
100 or
[75] 10 Not Stated 3 75°C, 1C
0
[76] 20 40-80 1.8% 6 25°C, 1C
[77] 30 100 Not Stated 9 55°C, 1C
This
15 50-90 1% 16 25°C, 0.5C
Study

The proposed SoH estimation method has been experimentally applied here to an

NMC cylindrical battery cells aged with cycle aging at room temperature, but is also

applicable to other chemistries. Kindermann et al. [14] showed that the relaxation rates

measured with EIS of NMC and Lithium Manganese Oxide (LiMn) cells are within 1% of

each other, suggesting that the proposed SoH estimation model would be equally effective.

4.7 CONCLUSION

In this paper a SoH estimation method is proposed that entails a combined

consideration of the short-term relaxation effect, EIS measurements and fractional order

impedance modeling. It was shown that EIS measurements obtained between 5 and 15

minutes of rest can produce repeatable results to be used with an empirical model. The

proposed empirical model was shown to provide an average SoH estimation error of less

than 1%. This method shows promise for SoH estimation with EIS, for potential

implementation with a BMS.

- 132 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

ACKNOWLEDGEMENTS

This work was supported by the Collaborative Research and Training Experience

(CREATE) Grant through the Natural Science and Engineering Research Council of

Canada under Grant 482038-2016. We thank Cadex Electronics for their facilities, support

and expertise.

REFERENCES
[1] M.A. Hannan, M.S.H. Lipu, A. Hussain, A. Mohamed, A review of lithium-ion
battery state of charge estimation and management system in electric vehicle
applications: Challenges and recommendations, Renew. Sustain. Energy Rev. 78
(2017) 834–854. doi:10.1016/j.rser.2017.05.001.
[2] R. Xiong, J. Cao, Q. Yu, H. He, F. Sun, Critical Review on the Battery State of
Charge Estimation Methods for Electric Vehicles, IEEE Access. 6 (2018) 1832–
1843. doi:10.1109/ACCESS.2017.2780258.
[3] M. Ismail, R. Dlyma, A. Elrakaybi, R. Ahmed, S. Habibi, Battery state of charge
estimation using an Artificial Neural Network, 2017 IEEE Transp. Electrif. Conf.
Expo, ITEC 2017. (2017) 342–349. doi:10.1109/ITEC.2017.7993295.
[4] S.A. Gadsden, M. Al-Shabi, S.R. Habibi, Estimation Strategies for the Condition
Monitoring of a Battery System in a Hybrid Electric Vehicle, ISRN Signal Process.
2011 (2011). doi:10.5402/2011/120351.
[5] M.S. Farag, R. Ahmed, S.A. Gadsden, S.R. Habibi, J. Tjong, A comparative study
of Li-ion battery models and nonlinear dual estimation strategies, 2012 IEEE Transp.
Electrif. Conf. Expo, ITEC 2012. (2012) 1–8. doi:10.1109/ITEC.2012.6243485.
[6] R. Ahmed, M. El Sayed, I. Arasaratnam, J. Tjong, S. Habibi, Reduced-Order
Electrochemical Model Parameters Identification and State of Charge Estimation for
Healthy and Aged Li-Ion Batteries—Part II: Aged Battery Model and State of
Charge Estimation, IEEE J. Emerg. Sel. Top. Power Electron. 2 (2014) 678–690.
doi:10.1109/JESTPE.2014.2331062.
[7] Y. Che, Z. Deng, X. Lin, L. Hu, Predictive Battery Health Management With
Transfer Learning and Online Model Correction, IEEE Trans. Veh. Technol. 70
(2021) 1269–1277.
[8] N. Meddings, M. Heinrich, F. Overney, J.S. Lee, V. Ruiz, E. Napolitano, S. Seitz,
G. Hinds, R. Raccichini, M. Gaberšček, J. Park, Application of electrochemical
impedance spectroscopy to commercial Li-ion cells: A review, J. Power Sources.
480 (2020). doi:10.1016/j.jpowsour.2020.228742.
[9] E. Din, C. Schaef, K. Moffat, J.T. Stauth, A scalable active battery management
system with embedded real-time electrochemical impedance spectroscopy, IEEE
Trans. Power Electron. 32 (2017) 5688–5698. doi:10.1109/TPEL.2016.2607519.
[10] B.G. Carkhuff, P.A. Demirev, R. Srinivasan, Impedance-Based Battery
Management System for Safety Monitoring of Lithium-Ion Batteries, IEEE Trans.
Ind. Electron. 65 (2018) 6497–6504. doi:10.1109/TIE.2017.2786199.

- 133 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

[11] X. Wei, X. Wang, H. Dai, Practical on-board measurement of lithium ion battery
impedance based on distributed voltage and current sampling, Energies. 11 (2018).
doi:10.3390/en11010064.
[12] N. Lohmann, P. Haussmann, P. Wesskamp, J. Melbert, T. Musch, Employing Real
Automotive Driving Data for Electrochemical Impedance Spectroscopy on Lithium-
Ion Cells, SAE Int. J. Altern. Powertrains. 4 (2015) 308–317. doi:10.4271/2015-01-
1187.
[13] E. Locorotondo, F. Corti, L. Pugi, L. Berzi, A. Reatti, G. Lutzemberger, Design of a
Wireless Charging System for Online Battery Spectroscopy, Energies. (2021).
[14] F.M. Kindermann, A. Noel, S. V. Erhard, A. Jossen, Long-term equalization effects
in Li-ion batteries due to local state of charge inhomogeneities and their impact on
impedance measurements, Electrochim. Acta. 185 (2015) 107–116.
doi:10.1016/j.electacta.2015.10.108.
[15] K. Qian, B. Huang, A. Ran, Y.B. He, B. Li, F. Kang, State-of-health (SOH)
evaluation on lithium-ion battery by simulating the voltage relaxation curves,
Electrochim. Acta. 303 (2019) 183–191. doi:10.1016/j.electacta.2019.02.055.
[16] Q. Fang, X. Wei, T. Lu, H. Dai, J. Zhu, A state of health estimation method for
lithium-ion batteries based on voltage relaxation model, Energies. 12 (2019).
doi:10.3390/en12071349.
[17] I. Baghdadi, O. Briat, P. Gyan, J.M. Vinassa, State of health assessment for lithium
batteries based on voltage–time relaxation measure, Electrochim. Acta. 194 (2016)
461–472. doi:10.1016/j.electacta.2016.02.109.
[18] A. Ran, Z. Zhou, S. Chen, P. Nie, K. Qian, Z. Li, B. Li, H. Sun, F. Kang, X. Zhang,
G. Wei, Data-Driven Fast Clustering of Second-Life Lithium-Ion Battery:
Mechanism and Algorithm, Adv. Theory Simulations. 3 (2020) 1–9.
doi:10.1002/adts.202000109.
[19] A. Barai, W.D. Widanage, J. Marco, A. Mcgordon, P. Jennings, A study of the open
circuit voltage characterization technique and hysteresis assessment of lithium-ion
cells, J. Power Sources. 295 (2015) 99–107. doi:10.1016/j.jpowsour.2015.06.140.
[20] W. Waag, S. Käbitz, D.U. Sauer, Experimental investigation of the lithium-ion
battery impedance characteristic at various conditions and aging states and its
influence on the application, Appl. Energy. 102 (2013) 885–897.
doi:10.1016/j.apenergy.2012.09.030.
[21] H. Wang, M. Tahan, T. Hu, Effects of rest time on equivalent circuit model for a li-
ion battery, Proc. Am. Control Conf. 2016-July (2016) 3101–3106.
doi:10.1109/ACC.2016.7525394.
[22] V. Zinth, C. von Lüders, J. Wilhelm, S. V. Erhard, M. Hofmann, S. Seidlmayer, J.
Rebelo-Kornmeier, W. Gan, A. Jossen, R. Gilles, Inhomogeneity and relaxation
phenomena in the graphite anode of a lithium-ion battery probed by in situ neutron
diffraction, J. Power Sources. 361 (2017) 54–60.
doi:10.1016/J.JPOWSOUR.2017.06.060.
[23] M. Messing, T. Shoa, S. Habibi, Lithium-Ion Battery Relaxation Effects, in: ITEC
2019, 2019.
[24] H. Shabbir, W. Dunford, T. Shoa, State of health estimation of Li-ion batteries using
electrochemical impedance spectroscopy, 2017 IEEE Transp. Electrif. Conf. Expo,
ITEC 2017. (2017) 108–112. doi:10.1109/ITEC.2017.7993255.
[25] A. Guha, A. Patra, Online Estimation of the Electrochemical Impedance Spectrum
and Remaining Useful Life of Lithium-Ion Batteries, IEEE Trans. Instrum. Meas.
67 (2018) 1836–1849. doi:10.1109/TIM.2018.2809138.

- 134 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

[26] J. Kim, L. Krüger, J. Kowal, On-line state-of-health estimation of Lithium-ion

battery cells using frequency excitation, J. Energy Storage. 32 (2020) 101841.
doi:10.1016/j.est.2020.101841.
[27] A. Gismero, D.I. Stroe, E. Schaltz, Calender Aging Lifetime Model for NMC-based
Lithium-ion Batteries Based on EIS Measurements, in: 2019 Fourteenth Int. Conf.
Ecol. Veh. Renew. Energies, 2019.
[28] E. Locorotondo, V. Cultrera, L. Pugi, L. Berzi, M. Pierini, G. Lutzemberger,
Development of a battery real-time state of health diagnosis based on fast impedance
measurements, J. Energy Storage. 38 (2021) 102566.
doi:10.1016/j.est.2021.102566.
[29] Zahner Messsysteme, EIS - Electrochemical Impedance Sepctroscopy, (2019).
[30] N.A. Chaturvedi, R. Klein, J. Christensen, J. Ahmed, A. Kojic, Modeling,
estimation, and control challenges for lithium-ion batteries, IEEE Control Syst. Mag.
(2010) 1997–2002. doi:10.1109/acc.2010.5531623.
[31] A. Rodríguez, G.L. Plett, Controls-oriented models of lithium-ion cells having blend
electrodes. Part 1: Equivalent circuits, J. Energy Storage. 11 (2017) 162–177.
doi:10.1016/j.est.2017.02.004.
[32] T.R. Jow, S.A. Delp, J.L. Allen, J.-P. Jones, M.C. Smart, Factors Limiting Li +
Charge Transfer Kinetics in Li-Ion Batteries, J. Electrochem. Soc. 165 (2018) A361–
A367. doi:10.1149/2.1221802jes.
[33] A. Maheshwari, M. Heck, M. Santarelli, Cycle aging studies of lithium nickel
manganese cobalt oxide-based batteries using electrochemical impedance
spectroscopy, Electrochim. Acta. 273 (2018) 335–348.
doi:10.1016/j.electacta.2018.04.045.
[34] H. Zappen, F. Ringbeck, D. Sauer, Application of Time-Resolved Multi-Sine
Impedance Spectroscopy for Lithium-Ion Battery Characterization, Batteries. 4
(2018) 64. doi:10.3390/batteries4040064.

- 135 -
 Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

Chapter 5: Battery SoC Estimation from EIS using Neural

Nets

Marvin Messing1,2, Tina Shoa2, Ryan Ahmed1, Saeid Habibi1

1
Department of Mechanical Engineering, McMaster University, Hamilton, ON, Canada,
2
Cadex Electronics, Richmond, BC, Canada
This paper is published in 2020 IEEE Transportation Electrification Conference and
Expo (ITEC), 2020, pp. 588-593, doi: 10.1109/ITEC48692.2020.9161523. This paper is
republished here with permission3.

ABSTRACT

In this paper, a battery state of charge (SoC) estimation strategy with deep neural networks

(DNN) and Electrochemical Impedance Spectroscopy (EIS) is proposed. EIS data was

obtained for a range of conditions and was used as inputs to a DNN. Additionally, a battery

model was fit to the data, and the model parameters were used as inputs to a second DNN.

The Root Mean Square Error (RMSE) of both networks was found to be less than 5% for

SoC above 30%. The dataset used in this study included batteries of different States of

Health (SoH) as well as EIS measured at various rest times after different discharge pulses.

5.1 INTRODUCTION

Battery state estimation accuracy for lithium-ion batteries and their use in electric vehicles

(EV) remains an area of interest to engineers and researchers. The most important battery

states are State of Charge (SoC) [1], State of Health (SoH) [2], and State of available Power

3
In reference to IEEE copyrighted material which is used with permission in this thesis, the IEEE does not
endorse any of McMaster's products or services. Internal or personal use of this material is permitted. If
interested in reprinting/republishing IEEE copyrighted material for advertising or promotional purposes or
for creating new collective works for resale or redistribution, please go to
http://www.ieee.org/publications_standards/publications/rights/rights_link.html to learn how
to obtain a License from RightsLink.
- 136 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

(SoaP) [3]. Further increase in accuracy for the estimation of these states can result in a

decrease in cost, size, and weight of the battery pack, as well as increase in longevity and

safety [4], [5]. However, the information extracted from the battery is usually restricted to

simple voltage, current and temperature measurements, and Kalman filters are used

together with battery models to obtain optimal state estimation in the face of uncertainties

[6], [7],[8]. A more detailed picture of the states of the battery can be obtained using

Electrochemical Impedance Spectroscopy (EIS), a technique which has been widely used

by researchers in electrochemistry and battery engineering to characterize battery

behaviour [9], [10]. EIS applies a voltage or current signal to the battery at different

frequencies and the corresponding response is measured and converted to frequency

domain via the Fourier transform. The real and imaginary components of the impedance

are commonly plotted for each frequency resulting in a Nyquist plot [11]. EIS is not yet

available as part of the battery management system (BMS) in EVs, but efforts are being

made to bring down costs and complexity of this method [12]–[14]. Once EIS becomes

available in the BMS it can be used to augment SoC estimation algorithms. Only a few

studies have attempted to obtain accurate state estimation from EIS data. In [15], an

equivalent circuit model was fit to EIS data and the model parameters were related to SoC

showing an error of 2 to 5%. In [16], a large EIS dataset was used with Gaussian process

regression to estimate SoH. The SoH estimated by the Gaussian algorithm was shown to

be within ±1 standard deviation after 150 cycles. The zero phase frequency was related to

SoH and used to train a simple neural network in [17], however, no information about

estimation accuracy was provided. Deep Neural Networks (DNN) have become a useful
- 137 -
 Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

tool to solve difficult classification and regression problems [18] in areas ranging from

image processing to natural language processing. The relationship of the EIS data and the

battery states is very complex making DNNs a suitable choice for state estimation from

EIS. The aim of this study is to train a DNN with EIS data from lithium-ion batteries of

different SoH to estimate SoC. Two networks are trained, one using raw EIS data, and the

other using parameters from an impedance model.

5.2 EXPERIMENTAL

5.2.1 CHARACTERIZATION TESTS AND DATA

Lithium-ion batteries (Samsung INR21700-50E, 4.8Ah) were tested with EIS using a Bio-

logic SP150 Potentiostat. Battery SoC and SoH were determined with a custom battery

tester to less than 0.5% accuracy, and aging was done by cycling (constant current –

constant voltage charge at C/2, discharge at 1C) using a Cadex C7400ER 4-channel battery

analyzer. EIS and capacity tests were performed under controlled temperature conditions

using a Testequity 1007C thermal chamber. The dataset used in this study consist of EIS

scans for 25 frequencies between 800 Hz and 100 mHz each with a voltage amplitude of

10 mV. To account for relaxation effects [19], EIS was measured periodically after 15

second discharge pulses of 1C, 3C and 5C currents. The measurements were taken every 5

minutes over 30 minutes, then after every 15 minutes for 1 hour, and then every hour over

3 hours. In addition, the batteries were tested at 25°C and at 100%, 90%, and 80% SoH.

Finally, SoC was varied in 5% increments from 30% to 95%. This multi-dimensional

dataset of over 600 data points formed the basis for the findings of this study. A small

- 138 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

subset of Nyquist plots for the different conditions are shown in Figure 5-1a from which

the high complexity of the relationship between SoC, SoH and EIS results can be observed.

Figure 5-1b shows how different rest times after a 5C discharge impact the EIS

measurement.

a)

80% SoH

90% SoH

100% SoH

b)

Figure 5-1: Nyquist plots for 100%, 70% and 30% SoC at 25°C and different SoH (a).
Nyquist plots for 95% SoC, 90% SoH at 25°C after 15 second, 5C discharge pulse measured
different rest times (b).

- 139 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

5.2.2 BATTERY MODEL

The impedance model shown in Figure 5-2b was fit to the EIS data using particle swarm

optimization and the sum of squares cost function with equal weighting [11]. The model is

based on [9] who showed a good fit to EIS data and provided clear reasoning for the

selection of model components. The complex impedance 𝑍 of this fractional model changes
1⁄
with frequency 𝜔 according to (5-1), where 𝜏𝑘 = (𝑅𝑘 𝑄𝑘 ) 𝛼𝑘 .

𝑅1 𝑅2 1
𝑍(𝜔) = 𝑖𝜔𝐿 + 𝑅0 + 1+𝑅 𝛼 + 1+𝑅 𝛼 +𝑄 𝛼3 (5-1)
1 𝑄1 (𝑖𝜔𝜏1 ) 1 1 𝑄1 (𝑖𝜔𝜏1 ) 2 3 (𝑖𝜔)

The model parameters across the different conditions are used as inputs to one of the neural

networks. EIS data at 100% SoH can be modeled using a simpler model, with an inductor

in series with a resistor in series with a constant phase element (CPE) and in series with

only one parallel Zarc element. However, the neural network requires the same amount of

inputs regardless of SoH, therefore, two Zarc elements were used as shown in Figure 5-2

in all cases.

- 140 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

a)

b)
Figure 5-2: Impedance model used to model relaxation effect (a), and model fit to relaxation
data at 25°C and 90% SoC (b).

5.2.3 NETWORK ARCHITECTURES

The architectures for the two neural networks used in this study are shown in Figure 5-3.

Both networks consist of a single output, which is the SoC of a battery. The input

parameters for the first network (Figure 5-3a) are the raw EIS values in frequency domain,

which are a complex pair for each measurement frequency. The complex pairs were split

into real and imaginary parts, yielding 25 numbers for the real parts and 25 numbers for the

imaginary parts of the impedance, since 25 frequencies were used in the EIS measurement.

Therefore, the neural network in Figure 5-3a has a total of 50 inputs. A network structure

consisting of three hidden layers with 180 nodes each was found to result in the lowest

regression error. For the second neural network (Figure 5-3b), the EIS data was used to
- 141 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

parameterize the battery model presented in Section 5.2.2. This resulted in 10 numbers for

each EIS scan, representing the model parameters 𝐿, 𝑅0 , 𝑅1 , 𝑄1 , 𝛼1 , 𝑅2 , 𝑄2 , 𝛼2 , 𝑄3 and 𝛼3 .

These 10 model parameters were used as the inputs to the neural network shown in Figure

5-3b. In this case, a single hidden layer with 70 nodes resulted in the lowest regression

error.

a)

b)

Figure 5-3: Network structures for raw EIS input (a), and EIS model parameters (b).

The models were designed and trained using Keras (based on TensorFlow) and Scikit-

Learn Python libraries. The input parameters were scaled by removing the mean and scaling

to unit variance, using the Scikit-Learn StandardScaler function. The output parameters

- 142 -
 Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

were scaled to values between 0 and 1, using the Scikit-Learn MinMaxScaler function.

Without these scaling methods, the neural networks were unable to estimate the SoC. All

neurons have Sigmoid activation functions and optimization was done using the standard

stochastic gradient descent (SGD), with the mean squared error (MSE) cost function.

Learning rates were varied between 0.001 and 0.1. The dataset was split into training (80%)

and validation (20%) sets, and half of the training set was used as a batch size. In addition,

L2 regularization was employed, with a factor of 0.0001 for raw EIS input, and 0.00016

for model parameter input to prevent overfitting. Higher and lower regularization factors

resulted in increased RMSEs or over-fitting. Network training was performed to a

maximum of 10 million epochs. The hyper parameters are summarized in Table 5-1.

Table 5-1 Network Hyper Parameters

Hyper Parameter Value
Activation Function Sigmoid
Optimizer SGD
Cost Function MSE
Learning Rate 0.001-0.1
Training/Validation Split 20%
Batch Size ½ of Training Set
Regularization L2
Epochs 10M

5.3 RESULTS AND DISCUSSION

The performance comparison of the two networks is visualized in Figure 5-4 for each

measured SoC. All estimates from the neural networks are shown (50 points per SoC), as

well as their mean and standard deviation for each SoC. For both networks, some SoC

values are estimated more consistently than others. With raw EIS input (Figure 5-4a), the

standard deviation is large in the mid SoC range (50-70%), which is where the EIS data
- 143 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

mostly overlaps, making it harder for the neural network to distinguish SoCs. However, for

model parameter inputs, the standard deviation is higher for lower SoCs (50-30%). In Table

5-2, performance parameters are summarized for both networks. The RMSE values for the

raw EIS input network are similar between training and validation sets suggesting that

overfitting is minimal. However, the training and validation RMSE values for the model

parameter input network are further apart, suggesting that some overfitting is present. The

maximum standard deviation per SoC is slightly higher when using raw EIS inputs,

however, it is also more robust.

- 144 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

a)

b)

Figure 5-4: RMSE evolution during training epochs for DNN using raw EIS data (a), and
DNN using EIS model parameters (b).

Table 5-2: Network Performance

Input
Parameter
Raw EIS Model Param.
Overall RMSE 3.80 3.87
Training RMSE 3.81 3.78
Validation RMSE 3.74 4.22
Max. Std. Dev. 4.39 3.83

Figure 5-5 shows the evolution of the RMSE during training for both types of networks.

With raw EIS inputs, the neural network reaches a low RMSE within 1 million epochs.
- 145 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

During the rest of the training time the performance gains are much smaller, but the training

and validation RMSE approach the same value. With model parameters as inputs, on the

other hand, the network training is slower in general, with over 6 million epochs required

before the RMSE stabilizes. The difference between training and validation RMSEs,

however, remains significant. The training process for this network is not smooth but shows

steps at epoch numbers 2 million and 6 million, which may be a result of the regularization

algorithm.

- 146 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

a)

b)

Figure 5-5: RMSE evolution during training epochs for DNN using raw EIS data (a), and
DNN using EIS model parameters (b).

To find suitable numbers for hidden layers and nodes per layer for each network, the

number of layers and nodes were changed in different directions. Figure 5-6 shows how

the RMSE changes for both training and validation values, for the two types of inputs (raw

EIS and model parameters) and for different numbers of layers (Figure 5-6a) and different

- 147 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

number of nodes per layer (Figure 5-6b). In each case, the criteria for best performance was

the lowest RMSE and the shortest distance between training and validation, favouring the

latter over the former. The configurations for number of layers and nodes per layer selected

based on this criterion are circled in Figure 5-6. For raw EIS inputs, 3 hidden layers (Figure

5-6a) with 180 nodes each (Figure 5-6b) were selected, since this configuration results in

both the lowest RMSE, as well as the shortest distance between training and validation. For

battery model inputs, the lowest RMSE occurs at 3 layers for the training set (Figure 5-6a),

however, the validation set RMSE is much higher. This indicates that the network over-fits

to the training data. The next lowest RMSE occurs with 1 layer, and here, the distance

between training and validation is smaller. However, the RMSE for the training set is still

lower than the RMSE for the validation set, even at 1 layer, so overfitting was not

eliminated. Similarly, the smallest distance between training and validation is found for 70

nodes per layer (Figure 5-6b).

- 148 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

Input: Model Param. Raw EIS

a)

b)

Figure 5-6: RMSE for validation and training for different number of network layers (a),
and different number of nodes per layer (b).

The results of this study show that when using simple, feed-forward neural networks, there

are no performance gains from using battery model parameters over raw EIS data. The

smaller number of input nodes required for model parameters is desirable, however, the

modeling process itself can add noise to the data, and as shown in this study, the neural

network does not perform significantly better than with all 50 raw EIS values as inputs.

Furthermore, all neural network configurations used in this study for model parameter

inputs tended to over-fit to the training data. In contrast, using raw EIS as inputs to the

- 149 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

neural networks resulted in RMSE values which were similar for both training and

validation sets, as well as lower RMSE values for most SoC values, as shown in Figure 5-7.

Figure 5-7 further shows that for SoCs above 30%, all RMSE values are below 5%. At 30%

SoC, however, a clear increase in RMSE is seen for both types of inputs.

Figure 5-7: RMSE for each SoC.

If EIS measurement capability is available on board EVs or, for smaller battery packs used

in portable applications, as part of a charger, SoC may be measured by feeding raw EIS

data into a neural network. The purpose of this study was to provide a starting point for this

method and show how simple neural networks perform in this task. Various values for

number of layers, nodes per layer, regularization and learning rate were investigated,

however, an even more systematic approach considering a wider range of values for each

parameter may improve performance. In this study, the training-validation split was held

constant, however, a k-fold validation approach may provide more information of the

robustness of the neural network performance. Finally, the dataset used in this study
- 150 -
 Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

included EIS scans taken with rest times ranging from as short as 5 minutes to 3 hours. The

difference in EIS results especially at short rest times (Figure 5-1b) creates an additional

challenge for the neural network regression. As part of a future study, scans at different rest

times should be isolated to investigate the impact of rest times on neural network

performance. Shorter rest times are desirable for real time applications; however, any

discharge or charge has a significant impact on the EIS measurement during initial rest.

Beyond simple, feed-forward neural networks, more sophisticated network structures are

available such as recurrent neural networks (RNN) and convolutional neural networks

(CNN). However, using EIS data with these types of neural networks is less straight

forward since special formatting of the input data may be required. CNNs are frequently

used to find features in images and RNNs have a short-term memory component, neither

of which are directly applicable to EIS data. However, strategies to take advantage of RNNs

and CNNs will be investigated in future studies.

5.4 CONCLUSION

Neural networks were used in this work to estimate the SoC of Samsung INR21700-50E

lithium-ion batteries from EIS scans at different SoH, rest time, and prior discharge pulse

depths. The EIS data was used to parameterize a fractional order battery model. Two neural

networks were designed, one to use for inputs, the values of the real and imaginary EIS

data components, and another to use the battery model parameters as its inputs. Different

network structures were evaluated for each case. The ones with lowest RMSE and lowest

overfitting were presented. It was found that both cases can provide SoC estimates with

- 151 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

RMSEs of less than 5% for the SoC range of 35% to 95%. However, using battery model

parameters as inputs to the neural network still showed some degree of overfitting. Further

work is suggested to include more advanced neural network architectures as well as

investigate the impact of rest time on the estimation accuracy.

ACKNOWLEDGMENT

This research was conducted under a CREATE grant from the Natural Sciences and

Engineering Research Council of Canada.

REFERENCES
[1] R. Xiong, J. Cao, Q. Yu, H. He, and F. Sun, “Critical Review on the Battery State of
Charge Estimation Methods for Electric Vehicles,” IEEE Access, vol. 6, pp. 1832–
1843, 2018.
[2] L. Ungurean, G. Cârstoiu, M. V Micea, and V. Groza, “Battery state of health
estimation : a structured review of models , methods and commercial devices,” no.
July 2016, pp. 151–181, 2017.
[3] A. Farmann and D. U. Sauer, “Comparative study of reduced order equivalent circuit
models for on-board state-of-available-power prediction of lithium-ion batteries in
electric vehicles,” Appl. Energy, vol. 225, no. May, pp. 1102–1122, 2018.
[4] M. U. Cuma and T. Koroglu, “A comprehensive review on estimation strategies used
in hybrid and battery electric vehicles,” Renew. Sustain. Energy Rev., vol. 42, pp.
517–531, 2015.
[5] M. A. Hannan, M. S. H. Lipu, A. Hussain, and A. Mohamed, “A review of lithium-
ion battery state of charge estimation and management system in electric vehicle
applications: Challenges and recommendations,” Renew. Sustain. Energy Rev., vol.
78, no. August 2016, pp. 834–854, 2017.
[6] Y. Bar-Shalom, X.-R. Li, and T. Kirubarajan, Estimation with Applications to
Tracking and Navigation, vol. 9. 2001.
[7] H. H. Afshari, S. A. Gadsden, and S. Habibi, “Gaussian fi lters for parameter and
state estimation : A general review of theory and recent trends,” Signal Processing,
vol. 135, no. March 2016, pp. 218–238, 2017.
[8] R. Ahmed, M. El Sayed, I. Arasaratnam, Jimi Tjong, and S. Habibi, “Reduced-Order
Electrochemical Model Parameters Identification and SOC Estimation for Healthy
and Aged Li-Ion Batteries Part I: Parameterization Model Development for Healthy
Batteries,” IEEE J. Emerg. Sel. Top. Power Electron., vol. 2, no. 3, pp. 659–677,
2014.
[9] A. Maheshwari, M. Heck, and M. Santarelli, “Cycle aging studies of lithium nickel
manganese cobalt oxide-based batteries using electrochemical impedance
spectroscopy,” Electrochim. Acta, vol. 273, pp. 335–348, 2018.

- 152 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

[10] X. Zhu et al., “Electrochemical impedance study of commercial

LiNi0.80Co0.15Al0.05O2electrodes as a function of state of charge and aging,”
Electrochim. Acta, vol. 287, pp. 10–20, 2018.
[11] A. Lasia, Electrochemical Impedance Spectroscopy EIS, and Corrosion. 2011.
[12] E. Din, C. Schaef, K. Moffat, and J. T. Stauth, “A scalable active battery
management system with embedded real-time electrochemical impedance
spectroscopy,” IEEE Trans. Power Electron., vol. 32, no. 7, pp. 5688–5698, 2017.
[13] X. Wei, X. Wang, and H. Dai, “Practical on-board measurement of lithium ion
battery impedance based on distributed voltage and current sampling,” Energies, vol.
11, no. 1, 2018.
[14] H. Piret, P. Granjon, N. Guillet, and V. Cattin, “Tracking of electrochemical
impedance of batteries,” J. Power Sources, vol. 312, pp. 60–69, 2016.
[15] U. Westerhoff, T. Kroker, K. Kurbach, and M. Kurrat, “Electrochemical impedance
spectroscopy based estimation of the state of charge of lithium-ion batteries,” J.
Energy Storage, vol. 8, pp. 244–256, 2016.
[16] Y. Zhang, Q. Tang, Y. Zhang, J. Wang, U. Stimming, and A. A. Lee, “Identifying
degradation patterns of lithium ion batteries from impedance spectroscopy using
machine learning,” Nat. Commun., vol. 11, no. 1, 2020.
[17] Z. Xia and J. A. Abu Qahouq, “Adaptive and fast state of health estimation method
for lithium-ion batteries using online complex impedance and artificial neural
network,” Conf. Proc. - IEEE Appl. Power Electron. Conf. Expo. - APEC, vol. 2019-
March, pp. 3361–3365, 2019.
[18] M. Nielson, Neural Networks and Deep Learning. Determination Press, 2015.
[19] M. Messing, T. Shoa, and S. Habibi, “Lithium-Ion Battery Relaxation Effects,” 2019
IEEE Transp. Electrif. Conf. Expo, pp. 1–6, Jun. 2019.

- 153 -
 Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

Chapter 6: Low Temperature, Current Dependant Battery

State Estimation using Interacting Multiple Model
Strategy

Marvin Messing1,2, Sara Rahimifard1, Tina Shoa2, Saeid Habibi1

1
Department of Mechanical Engineering, McMaster University, Hamilton, ON, Canada,
2
Cadex Electronics, Richmond, BC, Canada
This paper is published in IEEE Access, 2021, doi: 10.1109/ACCESS.2021.3095938.
This paper is republished here with permission.

ABSTRACT

Lithium-ion battery State of Charge (SoC) estimation for Electric Vehicle (EV)

applications must be robust and as accurate as possible to maximize battery utilization and

ensure safe operation over a wide range of operating conditions. SoC estimation commonly

utilizes filters such as the Extended Kalman Filter (EKF) which rely on battery models,

usually in the form of Equivalent Circuit Models (ECM). At low temperatures the battery

response to current draw becomes increasingly non-linear, resulting in amplified SoC

estimation errors. In this study, current dependent SoC estimation at low temperature is

proposed using an Interacting Multiple Model (IMM) filter with three ECMs covering a

range of C-rates. The IMM is combined with the Smooth Variable Structure Filter (SVSF)

to obtain robust SoC estimates within a SoC estimation error of 2%.

6.1 INTRODUCTION

In this paper an improved low temperature State of Charge (SoC) estimation strategy is

proposed. Reduced SoC estimation errors are achieved by using different battery models

optimized for different drive cycle discharge current magnitudes (or C-rates). Three

- 154 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

Equivalent Circuit Models (ECM) are parameterized to cover low, medium, and high C-

rate drive cycles. The Interacting Multiple Model (IMM) filter is combined with the ECMs

and the Smooth Variable Structure (SVSF) filter to estimate SoC at low temperature. The

performance of the proposed estimation strategy is evaluated for lithium-ion cells at 100%

and 90% State of Health (SoH). Furthermore, the IMM-SVSF strategy is compared to the

conventional EKF in place of the SVSF (i.e. IMM-EKF). The remainder of this section

discusses the motivation and technical challenges, presents a review of relevant literature,

and provides the main contributions as well as an outline of the paper.

6.1.1 MOTIVATION AND TECHNICAL CHALLENGES

Electric Vehicles (EV) depend on Battery Management Systems (BMS) to ensure optimal

operation of the battery pack. One of the most important tasks of a BMS is the estimation

of several battery states, including SoC, State of Power (SoP), and SoH. A large selection

of methods exists for battery state estimation across a wide range of operating conditions

[1]. For low temperature operation, special care must be taken to ensure sufficient state

estimation accuracy. The behavior of the battery changes considerably at low temperatures

[2] and to account for that, model parameters are adjusted for different temperatures. Huo

et al. [3] developed a temperature dependent 2nd order ECM and used it to improve SoC

estimation using an EKF. Guo et al. [4] combined a temperature adjusted 2nd order ECM

with a dual extended Kalman filter and showed that their method greatly improves SoC

estimation compared to an EKF without temperature compensation. Xiong et al. [5] used a

temperature dependent battery model and a membership theory based estimation algorithm

- 155 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

to estimate SoC within 5% down to -40°𝐶. Furthermore, they show that even with

temperature compensation, the average estimation errors increase as temperature decreases.

Yang et al. [6] used a temperature compensated model with the EKF with an estimation

error within 3% for temperatures as low as 0°𝐶. Zhu et al. [7] proposed a temperature, SoC

and current compensated electro-thermal battery model and showed reduced voltage errors

down to -20°𝐶. Shen et al. [8] used a Square Root Cubature Kalman Filter (SRCKF) with

a temperature compensated, 2nd order ECM and showed improved SoC estimation accuracy

between -20°𝐶 and 60°𝐶 for medium C-rate drive cycles. At low temperature, the battery

response becomes increasingly non-linear with C-rate, creating additional modeling errors.

Kollmeyer et al. [9] compared different models and different parameterization methods for

various drive cycles and temperatures and showed the increased impact of C-rate at low

temperatures on battery model parameters. The battery models can be improved to include

current-dependent mode parameters; however, this creates additional challenges with filter-

based state estimation methods. Temperature dependent model parameters (much like SoC

dependent parameters) are compatible with state estimation filters such as the EKF. This is

because temperature and SoC are low frequency signals, changing at a slow rate and mostly

in one direction. However, drive cycle currents for EV applications are high frequency

signals, able to rapidly change magnitude and direction from one second to the next. This

means that current dependent battery models could exhibit parameter updates from one

filter sample to the next, creating a system which is no-longer observable. Observability is

a requirement for filters such as the EKF to produce robust estimates [10], [11]. Therefore,

current dependent battery models are not feasible for use with EKF and similar filters. In
- 156 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

this paper, the IMM is used to circumvent this limitation while maintaining observability

and robust state estimates. Relevant literature on the usage of the IMM for battery state

estimation is reviewed next.

6.1.2 INTERACTING MULTIPLE MODEL FILTER

When the uncertainty of a single model used with an EKF or similar filter is high, the

estimation error will also be high [12]. In such cases, estimation can be improved by

combining information from several models into one filter. This can be useful if the system

model is uncertain, but will be close to within a range of different models or their blended

combination. A static multi-model algorithm can then pick the most suitable one to use for

the given inputs. A dynamic multi-model algorithm can be used if the system model

changes during the estimation process. However, in the dynamic case, an optimal estimator

is not computationally feasible [10]. Instead, a sub-optimal algorithm such as the IMM

derived by Blom et al. [13] can be used. The IMM is used in many practical applications

such as air traffic tracking, autonomous vehicle tracking and missile tracking, and new

applications are still being explored [14]–[16]. For battery applications, the IMM has been

used to allow for multiple battery models to be considered simultaneously as part of state

estimation algorithms. Smiley et al. [17] used the IMM with EKF to select from 21 physics-

based, reduced order battery models each representing different battery aging states. In [18]

the same group introduced a post processing method to improve the predictions obtained

from the IMM by analyzing the probability mass function. Su et al. [19] considered three

different aging models, a polynomial, an exponential, and a Verhulst model, and used the

- 157 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

IMM to combine the models for an improved prediction of the Remaining Useful Life

(RUL) of the battery. In this paper, the IMM is used to improve SoC estimation at low

temperatures by switching between three different models, which are parameterized for

different C-rate regions.

6.1.3 SMOOTH VARIABLE STRUCTURE FILTER

The SVSF approach is a predictor-corrector method first presented in [20]. The SVSF is

based on sliding mode control (SMC) which uses a discontinuous gain and a smoothing

boundary layer. The stability and robustness of the SVSF method has been illustrated

against uncertainties and noise in relation to the filter model. The SVSF was later improved

with several advancements, including the covariance formulation, time-varying smoothing

boundary layer (SVSF-VBL) and combinations with different filters such as KF, EKF,

UKF, Particle filter (PF) and more [21]–[23]. The SVSF-VBL approach provides more

accuracy in the presence of noise and model uncertainties [24]. However, observability of

the system needs to be guaranteed in use of these algorithms. Gadsden et al. [25] compared

the SVSF to EKF and other filters for a target tracking problem and showed superior

performance of the SVSF especially in the presence of modeling errors. In a similar study,

Gadsden et al. [26] showed improved state and parameter estimation performance using the

SVSF for a hybrid EV application. In another study, Gadsden et al. [27] used the SVSF

with an IMM and compared the IMM-SVSF to the IMM-EKF for an air traffic control

problem. They found a significant improvement of the IMM-SVSF over the IMM-EKF.

- 158 -
 Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

This paper compares SVSF and EKF in terms of accuracy and robustness and their

performance as part of the IMM for low temperature battery SoC estimation.

6.1.4 CONTRIBUTIONS

In this work the IMM is used with the SVSF to reduce SoC estimation errors at low

temperatures by considering different battery models for different C-rates. This paper

includes the following contributions: 1) multiple models are used to reduce SoC estimation

errors at low temperatures using the IMM with C-rate specific models, and 2) the IMM-

SVSF is demonstrated to have higher SoC estimation accuracy compared to the IMM-EKF.

6.1.5 PAPER OUTLINE

The remainder of this paper is organized as follows. Section 6.2 contains background theory

on the battery model, EKF, SVSF, IMM, as well as vehicle modeling to obtain drive cycles.

In Section 6.3 the datasets used to parameterise and validate the proposed estimation method

are introduced. Section 6.4 presents the results and discusses the performance of the IMM-

SVSF compared to IMM-EKF and single model IMM and SVSF filters at low temperatures.

Concluding remarks are provided in Section 6.5.

6.2 THEORY

This section first defines the equivalent circuit model used in this study, followed by

descriptions of the EKF and SVSF SoC estimation algorithms. Finally, the IMM algorithm

is introduced.

- 159 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

6.2.1 BATTERY MODELING & PARAMETERIZATION

The battery model used in this study as part of the SoC estimation methods is shown in

Figure 6-1. This 3rd order ECM is a common choice for SoC estimation algorithms. A

resistor 𝑅0 is used to model the internal resistance of the battery and three RC-pairs are

used to model the time dependent, transient response of the battery caused by charge

transfer and diffusion reactions [28]. The resistances 𝑅𝑗 and capacitances 𝐶𝑗 are for the jth

element of the circuit with elements from 1 to 3. The ECM can be used together with the

OCV-SoC relationships of the battery to obtain terminal voltage using (6-1), (6-2), (6-3),

and (6-4) given an SoC and a cell current demand 𝑢𝑘 .

𝑧̂𝑘+1 = 𝑂𝐶𝑉(𝑥𝑘,𝑘+1 ) − 𝑅0 𝑢𝑘 − ∑𝑛𝑗=1 𝑣𝑗,𝑘+1 (6-1)

𝛥𝑡 𝛥𝑡
𝑣𝑗,𝑘+1 = 𝑣𝑗,𝑘 (1 − 𝜏 ) + 𝐶 𝑢𝑘 (6-2)
𝑗 𝑗

𝜏𝑗 = 𝑅𝑗 𝐶𝑗 (6-3)
Δ𝑡
𝑥𝑘+1 = 𝑥𝑘 + 𝐶 𝑢𝑘 (6-4)
𝑟

Figure 6-1: 3rd order battery equivalent circuit model.

Here, 𝑛 is the order of the ECM (in this case 𝑛 = 3), Δ𝑡 is the time between steps 𝑘 and

𝑘 + 1 (sample rate), 𝑣𝑗,𝑘 is the voltage drop across RC branch 𝑗 at time 𝑘 and 𝜏𝑗 is the time

constant associated with branch 𝑗. Finally, 𝑥𝑘 is the SoC of the battery at time 𝑘 and (4) is

the coulomb counting equation used to calculate the gain or loss of SoC given 𝑢𝑘 and Δ𝑡.

Finally, 𝐶𝑟 (in Amp-seconds) is the rated capacity of the cell. The OCV-SoC relationship,

OCV(𝑥𝑘 ) can be used in the form of a lookup table obtained from experimental data. Figure
- 160 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

6-2 shows two OCV-SoC datasets used in this study, one at 25°C used for baseline ECMs,

and a second at 0°C used for low temperature ECMs. The lower temperature OCV-SoC

curve shows slightly lower OCV throughout the operating range of the cell.

Figure 6-2: OCV-SoC Curve for different temperatures.

The ECM is used with different model parameter sets to match the behavior of batteries at

different temperatures and SoH. The parameters are found using the Simulink Parameter

Estimator with the Non-linear Least Squares method and the Trust-Region-Reflective

algorithm. The sum of squared errors was used as a cost function.

6.2.2 EXTENDED KALMAN FILTER

In this section, the EKF equations are presented. The EKF is widely used for battery state

estimation and, therefore, considered in this work as a baseline for comparison to the

proposed implementation of the IMM-SVSF SoC estimation strategy. The EKF estimates

- 161 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

the state vector 𝑥̂ using non-linear system and measurement equations as shown in (6-5)

and (6-6), respectively.

𝑥̂𝑘+1 = 𝑓(𝑥̂𝑘 , 𝑢𝑘 , 𝑤𝑘 ) (6-5)

𝑧̂𝑘 = ℎ(𝑥̂𝑘 , 𝑢𝑘 , 𝑣𝑘 ) (6-6)
where 𝑢𝑘 is the input to the system, 𝑤𝑘 is the system noise, 𝑣𝑘 is the measurement noise,

and 𝑧̂𝑘 is the measurement prediction. The battery model described in Section 6.2.1. can

be rearranged into matrix form to be used with the EKF as shown in (6-7).

𝛥𝑡
𝛥𝑡
1−𝐶 0 0 0 𝐶1
−1
𝑣1,𝑘 1 𝛥𝑡
𝑣2,𝑘 0 1−
𝛥𝑡
0 0 𝐶2
−1
𝑥̂𝑘,𝑘 = [𝑣 ] , 𝐹 =
3,𝑘
𝐶2 , 𝐺= 𝛥𝑡 ,𝐻 = −1 (6-7)
𝛥𝑡 𝑑𝑂𝐶𝑉(𝑥̂𝑘,𝑘+1)
𝑥𝑘 0 0 1−𝐶 0 𝐶3
2 𝛥𝑡 [ 𝑑𝑥 ]
[ 0 1]
0 0 [− 𝐶𝑟 ]

The matrices 𝐹 and 𝐺 make up the parts of (6-2) as well as (6-4) to produce the state vector

𝑥̂𝑘,𝑘 at filter step 𝑘. The state vector contains the three branch voltages as well as the SoC

estimate 𝑥𝑘 . The matrix 𝐻 is the Jacobian derived from (6-1).

A single cycle of the EKF algorithm consists of prediction, correction and update steps

defined as follows:

• Prediction: Equation 6-8 is the state equation of the EKF producing the a-priori state
estimate 𝑥̂𝑘,𝑘+1 , where the a-priori is indicated by incrementing the second subscript
𝑘. Equation 6-9 is the covariance corresponding the a-priori state estimate. Equation
6-11 represents the error 𝜈 between the measured voltage 𝑧𝑘 and the voltage estimate
𝑧̂𝑘 .

𝑥̂𝑘,𝑘+1 = 𝐹𝑥𝑘,𝑘 + 𝐺𝑢𝑘 (6-8)

′
𝑃𝑘,𝑘+1 = 𝐹𝑃𝑘,𝑘 𝐹 + 𝑄 (6-9)
𝑧̂𝑘|𝑘+1 = 𝐻𝑥̂𝑘|𝑘+1 (6-10)
𝜈𝑘,𝑘+1 = 𝑧𝑘+1 − 𝑧̂𝑘,𝑘+1 (6-11)

- 162 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

• Correction: In (6-12), the innovation covariance 𝑆 is calculated given the

measurement noise covariance 𝑅 and used in (6-13) to calculate the EKF gain 𝑊:

𝑆 = 𝑅 + 𝐻𝑃𝑘,𝑘+1 𝐻′ (6-12)
′ −1
𝑊 = 𝑃𝑘,𝑘+1 𝐻 𝑆 (6-13)
• Update: Finally, the gain is used in (6-14) and (6-15) to calculate the a-posteriori
state vector 𝑥̂𝑘+1,𝑘+1 and state covariance 𝑃𝑘+1,𝑘+1 respectively. Here, 𝐼 is the
identity matrix.

𝑥̂𝑘+1,𝑘+1 = 𝑥𝑘,𝑘+1 + 𝑊𝜈𝑘+1 (6-14)

𝑃𝑘+1,𝑘+1 = (𝐼 − 𝑊𝐻)𝑃𝑘,𝑘+1 (6-15)
𝜈𝑘+1,𝑘+1 = 𝑧𝑘+1 − 𝑧̂𝑘+1,𝑘+1 (6-16)
6.2.3 SMOOTH VARIABLE STRUCTURE FILTER

This section provides the SVSF algorithm which is applicable to any observable and

differentiable system. The SVSF employs a smoothing boundary layer 𝝍 and a

discontinuous gain. The SVSF gain forces the states to converge to a neighborhood of the

true value. This paper employs the SVSF with a time-varying smoothing boundary layer

(SVSF-VBL) to enhance estimation accuracy [24]. The width of the boundary layer

depends on the uncertainty of the filter model, as well as the system and measurement noise.

The SVSF-VBL algorithm uses a time-varying boundary layer with saturated limits to

guarantee stability and estimation convergence [21]. The SVSF-VBL has the same

prediction and update steps as the EKF but a different correction step:

• Correction: In (6-17) the combination of measurement errors 𝑬𝒌+𝟏 is calculated

given the SVSF convergence parameter 𝜸. Equation (6-18) calculates the SVSF
smoothing boundary layer width 𝝍. The SVSF gain is shown in (6-19), given a
tuning parameter 𝝍𝒍𝒊𝒎 .

- 163 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

𝐸 = |𝜈𝑘,𝑘+1 | + 𝛾|𝜈𝑘 | (6-17)

′
𝜓 = (𝐸 −1 𝐻′ 𝑆𝑘+1 )−1 (6-18)
−1 −1
𝐻 𝐸 𝑠𝑎𝑡(𝜈𝑘 𝜓 −1 ) 𝜈𝑘,𝑘+1 , 𝜓 ≥ 𝜓𝑙𝑖𝑚
𝑊={ (6-19)
𝐻 −1 𝐸𝜓 −1 , 𝜓 < 𝜓𝑙𝑖𝑚

6.2.4 INTERACTING MULTIPLE MODEL FILTERS

The IMM allows several estimation filters to run in parallel, each using a different model.

The estimates from each of the filters are then blended together using mixing probabilities

obtained based on the filters’ error covariance matrices. The main function of the IMM is

to compute mixing probabilities, which determine how much weight is assigned to each

filter estimate. The IMM must be initialized with a mode probability vector 𝝁𝒌 and a mode

transition probability matrix 𝒑. For the 𝒓 = 𝟑 filters used in this work, the initialization is

shown in (6-20).

𝜇𝑘,1 𝑝11 ⋯ 𝑝13

𝜇 = [𝜇𝑘,2 ] , 𝑝 = [ ⋮ ⋱ ⋮ ] (6-20)
𝜇𝑘,3 𝑝31 ⋯ 𝑝33

The mode probabilities indicate the probability of each filter (or mode) using the correct

model at each time step 𝒌. The elements 𝒊, 𝒋 of the mixing probability matrix 𝝁𝒌,𝒌 are

calculated using (6-21), where 𝒄̅𝒋 is calculated using (6-22).

𝑝𝑖𝑗
𝜇𝑘,𝑘,𝑖𝑗 =
𝑐̅𝑗
𝜇𝑘,𝑖 𝑖, 𝑗 = 1, … , 𝑟 (6-21)
𝑐̅𝑗 = ∑𝑟𝑖=1 𝑝𝑖𝑗 𝜇𝑘,𝑗 𝑗 = 1, … , 𝑟 (6-22)
The mixing probability matrix is then used in (6-23) and (6-24) to calculate new state

estimates 𝑥̃𝑗 and new state covariances 𝑃𝑗 for each of the 𝑟 filters, using the last filter output

𝑥̂𝑘+1,𝑘+1 and 𝑃𝑘+1,𝑘+1 . Here, 𝐴𝑥,𝑖 = (𝑥̂ 𝑘+1,𝑘+1,𝑖 − 𝑥̃𝑗 ). The new 𝑥̃𝑗 and 𝑃𝑗 are then used for

blending of the estimates and as new starting parameters for the next filter steps.

𝑥̃𝑗 = ∑𝑟𝑖=1 𝜇𝑘,𝑘,𝑖𝑗 𝑥̂𝑘+1,𝑘+1,𝑖 𝑗 = 1, … , 𝑟 (6-23)

- 164 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

𝑃𝑗 = ∑𝑟𝑖=1 𝜇𝑘𝑘,𝑖𝑗 (𝑃𝑘+1,𝑘+1,𝑖 + 𝐴𝑥,𝑖 𝐴𝑥,𝑖 ′) 𝑗 = 1, … , 𝑟 (6-24)

The remaining steps of the IMM are used to calculate the mixing probabilities for the next

iteration, given the voltage estimate 𝒛̂𝒌,𝒋 and associated innovation covariance 𝑺𝒋 from the

filters. This is done by first computing the likelihood 𝚲𝒋 using the gaussian probability

density function (PDF) 𝑵(𝒙, 𝝁, 𝝈) as shown in (6-25).

𝛬𝑗 = 𝑁(𝑧𝑘 , 𝑧̂𝑘,𝑗 , 𝑆𝑗 ) (6-25)

The likelihood is then used together with 𝒄̅𝒋 from (6-22) and 𝒄 from (6-27) to update the

mode probabilities 𝝁𝒌+𝟏,𝒋 in (6-26).

𝑐̅𝑗
𝜇𝑘+1,𝑗 =
𝑐
𝛬𝑗 , 𝑗 = 1, … , 𝑟 (6-26)
𝑐 = ∑𝑟𝑗=1 𝛬𝑗 𝑐̅𝑗 (6-27)
̂𝒌+𝟏,𝒌+𝟏 and associated covariance 𝑷𝒌+𝟏,𝒌+𝟏 can be
Finally, the IMM state vector 𝒙

̂𝒌+𝟏,𝒌+𝟏,𝒋 − 𝒙
computed using (28) and (29), where 𝑩𝒙,𝒋 = [𝒙 ̂𝒌+𝟏,𝒌+𝟏 ].

𝑥̂𝑘+1,𝑘+1 = ∑𝑟𝑗=1 𝜇𝑘+1,𝑗 𝑥̂𝑘+1,𝑘+1,𝑗 (6-28)

𝑃𝑘+1,𝑘+1 = ∑𝑟𝑗=1 𝜇𝑘+1,𝑗 (𝑃𝑘,𝑗 + 𝐵𝑥,𝑗 𝐵′ 𝑥,𝑗 ) (6-29)
Figure 6-3 shows an overview of the IMM algorithm. The algorithm starts at the “IMM

mixing probabilities” block to compute the mixing probabilities given initial values from

(20). The mixing probabilities are used in the next block (“IMM mixing”) together with the

initial state vectors and state covariances of each of the filters (𝑟 = 3 ) used in this diagram

to compute updated state vectors and covariances as shown in (6-23). The updated states

are then used to initialize regular filters, such as the EKF or the SVSF. At this point, the

IMM estimate can also be computed in the “IMM Estimate” block with (6-28) and (6-29).

The output voltage estimates and measurement covariances of each of the filters are sent to

the “IMM likelihood” block to compute likelihoods using the measured voltage and (6-25).
- 165 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

The state estimates and state covariances are sent to the “IMM mixing” block to be used in

the next iteration instead of the initial values used for the first iteration. Finally, the mode

probabilities are updated in the “IMM update” block with (6-26) and used in the “IMM

mixing probabilities” block for the start of the next iteration instead of the initial mode

probabilities used in the first iteration. In Figure 6-3 only three filters are shown, but

additional filters can be used.

Figure 6-3: IMM Algorithm Diagram.

- 166 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

6.2.5 VEHICLE MODELING

The IMM-EKF or IMM-SVSF SoC estimation filters allow multiple battery models to run

in parallel to achieve an overall improved SoC estimate. In this study, the IMM filters are

used to improve SoC estimation at low temperatures where the three different battery

models are optimized for different C-rates. First, the battery models must be parameterized

with datasets that represent different C-rate ranges as well as the dynamic voltage response

specific to a given application. This study is intended for EV applications; therefore,

automotive drive cycles are used to obtain parameterization datasets. Automotive drive

cycles as provided by the US Environmental Protection Agency (EPA) consist of

representative sections of vehicle speed vs. time data for different driving conditions. This

drive cycle data can be used to simulate battery load demand for a given vehicle by

calculating the power required to overcome rolling resistance, aerodynamic drag, and in

some situations, gravity to achieve the speed of the drive cycle [78]. In (6-30), the power

required to accelerate the mass of the vehicle 𝑷𝒂𝒄𝒄 is calculated using the desired speed 𝑽

from the drive cycle, the mass of the car 𝒎𝒄𝒂𝒓 and acceleration given by 𝒂 = 𝒅𝑽/𝒅𝒕.

𝑃𝑎𝑐𝑐 = 𝑚𝑐𝑎𝑟 𝑎 𝑉 (6-30)

Given the air density 𝜌𝑎𝑖𝑟 , the frontal area of the vehicle 𝐴𝑓𝑟𝑜𝑛𝑡 and the drag coefficient of

the vehicle 𝐶𝐷 , (6-31) can be used to calculate the power required to overcome drag, 𝑃𝑑𝑟𝑎𝑔 .

1
𝑃𝑑𝑟𝑎𝑔 = 𝜌𝑎𝑖𝑟 𝐴𝑓𝑟𝑜𝑛𝑡 𝐶𝐷 𝑉 3 (6-31)
2

Under the assumption that the vehicle is driven on a flat road (i.e. without inclines of

declines), (6-32) calculates the power 𝑃𝑟𝑜𝑙𝑙,𝑓𝑙𝑎𝑡 to overcome rolling resistance, given

- 167 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

acceleration due to gravity 𝑔=9.81𝑚\𝑠 2 . The constants 𝐶𝑟𝑎 , 𝐶𝑟𝑏 , and 𝐶ℎ describe the

conditions of the road surface.

𝑃𝑟𝑜𝑙𝑙,𝑓𝑙𝑎𝑡 = (𝐶𝑟𝑎 + 𝐶𝑟𝑏 𝑉) 𝐶ℎ 𝑚𝑐𝑎𝑟 𝑔 (6-32)

To achieve the given drive cycle speeds, the battery must at least provide enough power to

overcome 𝑃𝑡𝑜𝑡 = 𝑃𝑎𝑐𝑐 + 𝑃𝑑𝑟𝑎𝑔 + 𝑃𝑟𝑜𝑙𝑙,𝑓𝑙𝑎𝑡 . However, additional power is lost due to non-

ideal battery efficiency 𝜂𝑏𝑎𝑡𝑡 and losses associated with the drivetrain efficiency 𝜂𝑑𝑟 .

Furthermore, parasitic power 𝑃𝑎𝑢𝑥 is required to run the auxiliary systems of the car, such

as lights, air-conditioning, radio, etc. The total power required by the battery pack is then

given by (6-33).

𝑃𝑏𝑎𝑡𝑡 = [(𝑃𝑡𝑜𝑡⁄𝜂𝑑𝑟 ) + 𝑃𝑎𝑢𝑥 ]⁄𝜂𝑏𝑎𝑡𝑡 (6-33)

Since current is usually easier to control in laboratory equipment than power, the power

can be converted into battery pack current. As this is a simple model, a fixed nominal

voltage 𝐸𝑐𝑒𝑙𝑙,𝑛𝑜𝑚 is assumed. The total voltage of the pack can be calculated given the

number of cells in series 𝑁𝑠𝑒𝑟𝑖𝑒𝑠 . Dividing 𝑃𝑏𝑎𝑡𝑡 by the total pack voltage and the number

of cells in parallel 𝑁𝑝𝑎𝑟𝑎𝑙𝑙𝑒𝑙 yields the battery cell current 𝐼𝑐𝑒𝑙𝑙 as shown in (6-34).

𝐼𝑐𝑒𝑙𝑙 = [𝑃𝑏𝑎𝑡𝑡 ⁄(𝑁𝑠𝑒𝑟𝑖𝑒𝑠 𝐸𝑐𝑒𝑙𝑙,𝑛𝑜𝑚 )]/𝑁𝑝𝑎𝑟𝑎𝑙𝑙𝑒𝑙 (6-34)

The battery cells used in this study are similar to those used by the Tesla Model 3 Long

range EV. The simple vehicle model described above can be used to approximate the

behavior of a Tesla Model 3 by using the parameters as well as assumed efficiencies shown

in Table 6-1. The efficiency of the battery pack and the drive train depend on many factors

and conditions, but for this simple vehicle model, constant, conservative numbers are

assumed [30]. The original Tesla Model 3 battery pack configuration was scaled by

- 168 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

reducing the number of modules in parallel to achieve slightly higher currents to cover a

C-rate range between 0 and 4C across different drive cycles.

Table 6-1: Simple Vehicle Model Parameters [29], [34], [35]

Name Symbol Value

Mass of the vehicle. 𝑚𝑐𝑎𝑟 1730 kg
Rolling resistance calculation factor. 𝐶𝑟𝑎 0.0041
Rolling resistance calculation factor. 𝐶𝑟𝑏 0.000018
Rolling resistance calculation factor. 𝐶ℎ 1.5
Density of air. 𝜌𝑎𝑖𝑟 1.2
Vehicle frontal area. 𝐴𝑓𝑟𝑜𝑛𝑡 3.25 𝑚 2
Vehicle drag coefficient. 𝐶𝑑 0.23
Vehicle auxiliary power draw. 𝑃𝑎𝑢𝑥 350 W
Drivetrain efficiency. 𝜂𝑑𝑟 0.765
Battery efficiency. 𝜂𝑏𝑎𝑡𝑡 0.95
Number of modules in parallel. 𝑁𝑝𝑎𝑟𝑎𝑙𝑙𝑒𝑙,𝑠𝑐𝑎𝑙𝑒𝑑 18
Number of series cells per module. 𝑁𝑠𝑒𝑟𝑖𝑒𝑠 96
Nominal battery cell voltage. 𝐸𝑐𝑒𝑙𝑙,𝑛𝑜𝑚 3.6 V

Figure 6-4 shows three drive cycle speed profiles and the corresponding current profiles

obtained with the simple vehicle model described above. Figure 6-4a shows the US Urban

Dynamometer Driving Schedule (UDDS) from which two low C-rate sections were

selected to parameterize the low C-rate model. Similarly, Figure 6-4b shows the aggressive

driving schedule called US06 from which two medium C-rate and two high C-rate sections

were chosen to parameterize the medium C-rate and high C-rate models, respectively. This

was done to achieve significant differences in model parameters for the three models, which

enables the IMM to identify the correct model more clearly. In each case, only discharge

currents were considered to simplify the model used for this study. The model can be

extended to include components to handle charge currents following methods described in

[31]. Figure 6-4c shows a drive cycle mix used for validation consisting of the unified

- 169 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

driving schedule called LA92 and the UDDS. In the next section, the drive cycle samples

and the validation cycle are used in experiments to obtain voltage profiles from battery

cells. These voltage profiles are later used to parameterize the three C-rate specific models

as well as to benchmark the proposed IMM based estimation strategy.

- 170 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

Figure 6-4: Current profile samples from EPA drive cycles for low C-rates (a) medium and
high C-rates (b). Validation drive cycle current profile (c).

- 171 -
 Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

6.3 LOW TEMPERATURE DRIVE CYCLE TESTING

This section describes the experiments and experimental setup used to obtain the

parameterization and validation drive cycle datasets from Samsung INR2170-50e lithium-

ion battery cells. The resulting datasets are also shown in this section.

6.3.1 EXPERIMENTS AND EXPERIMENTAL SETUP

The data used to parameterize the battery models and to validate the algorithms was

obtained using a custom designed battery tester. The main components of the tester are

shown in Figure 6-5a. The setup includes an Agilent load and power supply to discharge

and charge the battery cells as well as to implement the drive cycle load demand profiles.

An NI9219 Data Acquisition (DAQ) unit was used to measure the cell voltage via a custom

Kelvin connection. The DAQ also measures battery cell temperature as well as cell current

via the voltage drop across a shunt resistor. The battery cells were housed in a thermal

chamber for controlled environmental conditions and to simulate low temperature

conditions. The entire system was controlled and automated by custom software written in

the Python programming language. Figure 6-5b shows a picture of the setup in the lab as

well as a screen capture of the custom control software. A custom designed 4-point

measurement battery holder is also shown in magnified form since the cells cannot be seen

through the thermal chamber window otherwise.

- 172 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

Figure 6-5: Experimental setup diagram (a), picture and custom control software screen
capture (b).

New Samsung INR21700-50e 4.9Ah cylindrical lithium-ion battery cells were first broken

in with 10 charge/discharge cycles at 25°C with a standard constant-current constant-

- 173 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

voltage (CCCV) charge protocol at 0.5C to 4.2V and 0.02C charge cut-off current as

recommended by the manufacturer. The discharge current was set to 0.2C. The same

settings are used to verify the battery SoH. OCV-SoC tests were performed at 25°C and

0°C using C/15 discharge and charge currents. Prior to each OCV-SoC test, the batteries

were fully charged at 25°C. For the drive cycle tests, the batteries were first fully charged

at 25°C. Then, the environmental temperature was set to 0°C followed by a rest period of

1hour after the temperature setpoint was reached. Finally, the drive cycle current profile

was implemented to draw the required current from the battery. After the drive cycle profile

was completed, the temperature was raised back to 25°C and the battery was charged again.

In this study, results at 100% and 90% SoH are presented. The aged dataset was obtained

by aging battery cells with continued charge/discharge operations involving 1C discharge

and 0.5C charge currents and, the CCCV charge method to 0.02C cut-off current. The

ambient temperature was set to 25°C.

6.3.2 DRIVE CYCLE DATASETS

With the experimental procedures described in the previous section, the battery cell voltage

response at 0°𝐶 corresponding to different drive cycle current profiles are obtained. Four

different drive cycles are used as described in Section 6.2.5. Three of the current profiles

are used to produce high, medium, and low C-rate voltage cycles to parameterize each

battery model. These current profiles consist of drive cycle samples representative of

different C-rates (see Section 6.2.5) and are repeated several times in a row to cover a wider

SoC window. The fourth current cycle is used as a validation cycle to evaluate the

- 174 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

performance of the algorithms considered in this study. The validation cycle covers an SoC

range of approximately 5%. To include the full range of SoC, the procedures described in

this study can be repeated to construct lookup tables for the parameters of each model

across the entire range of SoC (for example, at every 5% SoC). However, for the purpose

of demonstrating the IMM-SVSF estimation approach with C-rate specific models, only

the 5% SoC window is presented.

Figure 6-6 shows each of the 4 cycles and their current and voltage responses obtained at

0°C for a battery at 100% SoH. In Figure 6-6a, high current loads up to 4C (20A) produce

the voltage profile to parameterize the high C-rate model. Figure 6-6b shows current loads

up to 2C (10A) and the corresponding voltage profile to parameterize the medium C-rate

model. The low C-rate model was parameterized with the voltage profile in Figure 6-6c

resulting from current loads of less than 1C (5A). The validation cycle current profile,

Figure 6-6d, includes loads between less than 1C and up to 4C. Similar voltage profiles

were obtained at 25°C to establish baseline results as well as for a battery at 90% SoH at

0°𝐶 for further evaluation of the proposed SoC estimation strategy. The model

parameterization results and the performance of the EKF, SVSF, and IMM algorithms are

presented and discussed in the next section.

- 175 -
 Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

Figure 6-6: Drive Cycles at 0°C and 100% SoH for parameterization of the high C-rate
model (a), the medium C-rate model (b), the low C-rate model (c), and for validation (d).

6.4 RESULTS & DISCUSSION

In this section, the models optimized for different C-rate ranges are discussed and the

behaviour of the IMM is investigated using mode probabilities. The performance of the

IMM-SVSF and IMM-EKF are then compared.

6.4.1 C-RATE SPECIFIC EQUIVALENT CIRCUIT MODELS

To evaluate the proposed SoC estimation strategy, the ECM introduced in Section 6.2.1.

was parameterized for three different C-rate ranges at 100% and 90% SoH and 0°C. Table

2 shows an overview of the model parameters obtained for each condition. Models L1, M1,
- 176 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

and H1 are used in Section 6.4.3. with the IMM for low temperature SoC estimation at

100% SoH, and L2, M2, and H2 for 90% SoH. Models B1 and B2 are used in Section 6.4.2.

to establish a baseline performance of the EKF and SVSF single filters at 25°C and 0°C,

respectively, for a 100% SoH battery. Model B2 has the same parameters as model M1 as

both are at the same conditions.

For each model, the drive cycle inputs were kept within different C-rate ranges (as shown

in Section 6.3.2) such that models are obtained for low C-rate (0-1C), medium C-rate (1-

2C) and high C-rate (2-4C). The performance of the 100% SoH, 0°C models (L1, M1 and

H1) compared to the parameterization profiles from Section 6.3.2. is shown in Figure 6-7.

For each C-rate range, a close fit is obtained using the 3rd order ECM introduced in Section

6.2.1 The Root Mean Square Error (RMSE) of the model fits is also included in Table 2.

The high C-rate models (H1 and H2) show the highest RMSEs, which indicates that the 3 rd

order ECM has some difficulty modeling battery response at high C-rates and low

temperatures. The remaining models (L1, L2, M1, M2, B1, B2) all fall into a range of

RMSEs between 0.003 and 0.03. The three time constants of the ECM fall within three

ranges, 0.19-0.36, 3.7-10, and 18-42, with model B1, 𝜏2 being the only exception. Model

B1 is the only model at 25°C which suggests that at higher temperatures, the ECM does not

need all three time constants to achieve a good RMSE of 0.0071 which falls well within

the range of the RMSEs of the other models.

- 177 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

Table 6-2: Battery Model Parameters

Model Name L1 L2 M1 M2 H1 H2 B1 B2

Type Low C-rate Medium C-rate High C-rate Baseline

C-rate Range 0-1 1-2 2-4 1-2
SoH [%] 100 90 100 90 100 90 100 100
Temperature
0 0 0 0 0 0 25 0
[°𝐶]
𝑅0 [𝑚Ω] 43.1 25.6 39.1 30.4 34.8 18.8 22.8 39.1
𝑅1 [𝑚Ω] 12.4 34.7 9.5 21.8 8.8 24.5 7.6 9.5
𝑅2 [𝑚Ω] 3.8 16.3 2.5 5 2 12.6 11.5 2.5
𝑅3 [𝑚Ω] 49.7 8 37.4 28.4 26.9 9.2 14.2 37.4
𝜏1 [𝑠] 0.25 0.23 0.3 0.25 0.36 0.19 0.24 0.3

𝜏2 [𝑠] 3.7 5.74 4.1 10 6.6 9.3 3824.3 4.1

𝜏3 [𝑠] 42 18 31.54 21.1 24.5 18 18.4 31.5

Fit RMSE
10.1 3.5 12.1 24.5 31.5 63.3 7.1 12.1
[× 10−3 ]

- 178 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

Figure 6-7: Performance for L1 (a), M1 (b), and H1 (c) c-rate specific models at 100% SoH
and 0°C.

Figure 6-8 shows how the series resistance 𝑅0 and the branch resistances of the models L1,

M1 and H1 vary. All of the resistances decrease as C-rate increases. At this low temperature

condition, the model parameters are significantly different for different C-rate regions,

therefore, the IMM will be able to extract different information from each model to achieve

an improved SoC estimate. In the next section, the models are used with IMM-EKF and

IMM-SVSF filters for SoC estimation. The medium C-rate models (M1 and M2) are used

for the EKF and SVSF single filters.

- 179 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

Figure 6-8: Change in model resistances with C-rate at 0°C.

6.4.2 FILTER INITIALIZATION AND BASELINE

The initialization parameters of the standard EKF are the initial state vector 𝑥̂𝑘,𝑘 , the initial

state covariance 𝑃𝑘,𝑘 , the process covariance 𝑄 and the measurement covariance 𝑅.

Equation (35) shows the initial values used for 𝑥̂𝑘,𝑘 , 𝑃𝑘,𝑘 , and 𝑄, where 𝐼4 is the 4-by-4

identity matrix and 𝑠𝑜𝑐𝑖𝑛𝑖𝑡 changes depending on the drive cycle used or when evaluating

filter convergence speed. 𝑄 was chosen sufficiently low to indicate some confidence in the

process. If 𝑄 is too small and the input currents and measured voltages are very accurate,

the filter will favor the process over the model. In addition to choosing a suitably high 𝑄, a

bias was added to the current and voltage signals to simulate realistic conditions. Similarly,

the measurement covariance 𝑅 is chosen to reflect some uncertainty. 𝑥̂𝑘,𝑘 and 𝑃𝑘,𝑘 are

updated as part of the filter and their values mostly impact convergence speed. The same

values for the three parameters in (6-35) are used for all filters in this study.

10−3
10−3
𝑥̂𝑘,𝑘 = [ −3 ] , 𝑃𝑘,𝑘 = 𝐼4 × 4 × 10−6 , 𝑄 = 𝐼4 × 2 × 10−2 (6-35)
10
𝑠𝑜𝑐𝑖𝑛𝑖𝑡

- 180 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

The SVSF uses the same initial parameters as the EKF plus the two additional parameters

ψ_lim, the constant smoothing boundary width, and γ, the convergence rate. The

parameters are summarized in Table 6-3. The measurement covariance is the only value

that is adjusted between the 100% and 90% SoH datasets as shown.

Table 6-3: Filter Initialization Parameters

Name Symbol Value

Measurement Covariance (100%/90% SoH) R 0.2/0.8
SVSF Convergence Rate 𝛾 0.1
SVSF Constant Smoothing Boundary Width 𝜓𝑙𝑖𝑚 70
Current Bias 𝐼𝑏𝑖𝑎𝑠 80 mA
Voltage Bias 𝑉𝑏𝑖𝑎𝑠 1 mV

With these initial settings as well as a suitable battery model, the battery SoC can be

estimated using EKF and SVSF filters. To establish a baseline performance, the model B1

(see Table 6-2) was then used with the EKF and SVSF and the validation drive cycle

voltage profile also at 25°C. Similarly, the mode B2 was used with the validation drive

cycle at 0°C. Figure 6-9a shows the SoC estimation performance at 25°C. Although the

medium C-rate model B1 is very specific to the medium C-rate range, good SoC tracking

is achieved. In contrast, the SoC estimation performance shown in Figure 6-9b, shows large

deviation of both EKF and SVSF from the measured SoC. The objective of this paper is to

improve the SoC tracking of EKF and SVSF at 0°C by using the IMM algorithm. The IMM

requires two additional initialization parameters, 𝜇𝑘 and 𝑝, with initial values for this study

shown in (6-36). The initial mode probability 𝜇𝑘 is chosen such that each filter has equal

probability of using the correct model at the start, since the correct filter is unknown. The

- 181 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

mode transition probabilities p impact how quickly the IMM switches between modes. The

chosen values result in smooth switching behaviour of the IMM.

0.33 0.9998 0.0001 0.0001

𝜇𝑘 = [0.33] , 𝑝 = [0.0001 0.9998 0.0001] (6-36)
0.33 0.0001 0.0001 0.9998

Figure 6-9: Baseline filter performance at 100% SoH, 25°C (a) and 0°C (b).

In the next section, the performance of the IMM-EKF and IMM-SVSF are presented and

compared.

6.4.3 PERFORMANCE COMPARISON

In this section, the low-temperature performance of the IMM-SVSF and IMM-EKF

estimation strategies is demonstrated for 100% SoH and 90% SoH batteries. Figure 6-10

shows SoC estimation performance, SoC error and C-rate (for reference) for the validation
- 182 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

drive cycle profile. Figure 6-10a shows the measured SoC for the validation drive cycle at

100% SoH as well as the estimated SoC from the EKF, SVSF, IMM-EKF, and IMM-SVSF.

Figure 6-10b shows similar results for the 90% SoH case. The SoC estimation errors for

100% and 90% SoH are shown in Figure 6-10c and Figure 6-10d, respectively. The C-rate

load profile of the validation drive cycle is shown in Figure 6-10e and Figure 6-10f for

reference. The validation drive cycle starts with C-rates of less than 1.5C until about 300

seconds into the cycle, then C-rates increase to around 2C on average. Towards the end the

average C-rate is raised again with peaks of up to 4C. Initially, in Figure 6-10a and Figure

6-10b, all filters perform in a similar way up until around 200 seconds into the drive cycle.

In this first part, the C-rates are mostly between 0.5 and 1.5C which can still be captured

well by the medium C-rate models (M1 and M2) used by the single filters. Between 200

seconds and 300 seconds, the drive cycle C-rates are mostly below 0.5C and the single

filters show an increased estimation error. Beyond 300 seconds, the C-rates become

increasingly larger and the single model filters (EKF and SVSF) show further increased

estimation errors. The EKF-IMM estimation error also increases, however, at a much

slower rate. The SVSF-IMM shows the lowest estimation error during the second half of

the drive cycle and overall remains within ±2%. Table 6-4 summarizes the performance

of each filter using the RMSE of the SoC estimate for the 100% and 90% SoH validation

cycles at 0°C. The IMM based low temperature SoC estimation strategy can achieve

significant improvements over single filters. Furthermore, the SVSF on its own performs

similar to the EKF, however, when combined with the IMM strategy, the IMM-SVSF

produces the lowest SoC estimation RMSE.

- 183 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

Table 6-4: Filter RMSE Comparison for validation cycles at 0°C

Filter RMSE 100% SoH RMSE 90% SoH

EKF 3.29 2.26
SVSF 3.03 2.36
IMM-EKF 1.01 0.85
IMM-SVSF 0.70 0.71

Figure 6-10: IMM SoC Estimation and error at 0°C for 100% SoH (a, c) and 90% SoH
(b,d), C-rates for reference (e, f).

A performance metric for the filters is how well they respond to increased uncertainty due

to, for example, increased input error. Figure 6-11 shows the impact of increasing the input

current bias on the RMSE of the SoC estimation during the validation cycle. Figure 6-11a

shows how the RMSE changes for each filter with the 100% SoH dataset, and Figure 6-11b
- 184 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

for the 90% SoH dataset. The RMSE increases in all cases with increased current bias,

however, the IMM-SVSF shows a slightly slower rate of RMSE increase compared to the

other filters.

Figure 6-11: Impact of current bias on RMSE for 100% SoH (a) and 90% SoH datasets.

In the next section, the performance of the IMM-SVSF and IMM-EKF is investigated

further using the voltage modeled by each filter as well as the mode probabilities calculated

by the IMM algorithm.

- 185 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

6.4.4 IMM VOLTAGE MODELING & MODE PROBABILITIES

The results shown in the previous section demonstrate the effectiveness of the IMM-SVSF

combination for low temperature SoC estimation. The reason for the lower estimation

errors of the IMM based filter is that the IMM is able to consider trade-offs between

different models. In this study, the models L1, M1, and H1 are used for 100% SoH data at

0°C. Each model is parameterized to suite different C-rate ranges, so the IMM can adjust

the SoC estimate based on the input C-rate by selecting the most suitable combination of

models. In this section, the behavior of the IMM is investigated further by first considering

the modeled battery cell voltages produced by the three different models. Figure 6-12

shows the voltage response of the three models used in the IMM-SVSF compared to the

measured voltage of the battery for the same drive cycle input. Also shown is the voltage

error for each model. When the drive cycle C-rates are low, the voltage error of the high C-

rate model H1 increases as shown, for example, between 220 and 320 seconds in the

magnified portion of Figure 6-12b. However, when voltage drops are large, such as shown

in Figure 6-12a.3, the model H1 is the most accurate. Similarly, the medium C-rate model

M1 is the most accurate for medium voltage drops and the low C-rate model L1 for low

voltage drops and rest periods as shown in Figure 6-12a.1 and Figure 6-12a.2 respectively.

This shows that the models perform as intended when they are used as part of the IMM.

- 186 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

Figure 6-12: Voltage and voltage error of the three models (L1, M1, H1) used in the IMM-
SVSF at 100% SoH and 0°C.

Another way to visualize the behaviour of the IMM is by considering the mode

probabilities. The mode probabilities indicate which model is most likely to be the correct

one, but effectively, mode probabilities also represent how much each model impacts the

final IMM estimate. Figure 6-13a and Figure 6-13b show the mode probabilities for the

IMM-EKF for the 100% and 90% SoH cases, respectively. Similarly, Figure 6-13c and d
- 187 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

show the mode probabilities for the IMM-SVSF. For each filter, the mode probabilities of

the three models are shown, indicating the contribution of each model to the final state

estimate. Figure 6-13e and Figure 6-13f show again the C-rate load profile of the validation

drive cycle for reference. In the previous section, the IMM-EKF was shown to have a

reduced estimation error, however, the IMM-SVSF estimation error was better and lower

still. In general, the IMM filters first favour the low C-rate models (L1 and L2), then the

medium C-rate models (M1 and M2), and finally the high C-rate models (H1 and H2). This

roughly coincides with the gradual increase in average C-rate of the drive cycles, however,

the IMM requires some time to switch between models. Subtle differences are apparent

between the IMM-EKF and IMM-SVSF and for the two SoH conditions. In Figure 6-13c,

the IMM-SVSF briefly favours the low C-rate model L1 again during a low C-rate region

(less than 0.5C) around 300 seconds. For the same 100% SoH dataset, the medium C-rate

model M1 remains dominant in the IMM-EKF as shown in Figure 6-13a. As shown in the

previous section (Figure 6-10a), the IMM-SVSF computes an improved SoC estimate after

300 seconds compared to the IMM-EKF. For the 90% SoH dataset, the IMM-SVSF in

Figure 6-13d shows a smooth transition between models and a clear dominant model for

each region. The IMM-EKF, on the other hand, shows similar overall trends, but with lower

mode probabilities in each region. Again, the SoC estimation performance of the IMM-

SVSF is also improved (Figure 6-10b). The IMM-SVSF is able to switch between models

more quickly and more confidently, resulting in improved SoC estimates.

- 188 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

Figure 6-13: Mode probabilities for IMM-EKF at 100% SoH (a), 90% SoH, (b) and IMM-
SVSF 100% SoH (c), 90% SoH (d). C-rates shown for reference in (e) and (f).

The performance of the filters is demonstrated based on single cell data. However, the

IMM-SVSF method can be scaled to full battery packs consisting of series and parallel

connected cells. Sub-modules of parallel cells are often treated as single cells with higher

capacities to reduce the number of sensors used in the battery pack [32]. This is valid under

the assumption that the cells are similar enough in their characteristics. Improvements can

be made to ECMs to include resistances of interconnects [33] and those models can then

- 189 -
 Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

be used with the IMM-SVSF. Single cells or parallel sub-modules in series equipped with

individual voltage sensors can each be tracked with dedicated IMM-SVSF algorithms.

6.5 CONCLUSION

Low temperatures increase the C-rate dependence of the response of lithium-ion batteries.

As a result, SoC estimation accuracy can decreases due to voltage modeling errors at

different C-rates. In this paper, an IMM-SVSF SoC estimation method is proposed that

utilizes three models, each for a different range of C-rates. The proposed strategy is able to

significantly increases SoC estimation accuracy at low temperatures. The IMM-SVSF is

further compared to IMM-EKF and still shows improved accuracy. The IMM behavior is

discussed in terms of voltage modeling and mode probabilities.

ACKNOWLEDGMENT

We thank Cadex Electronics for their facilities, support and expertise.

REFERENCES
[1] Y. Wang et al., “A comprehensive review of battery modeling and state estimation
approaches for advanced battery management systems,” Renew. Sustain. Energy
Rev., vol. 131, no. March, p. 110015, 2020.
[2] J. Jaguemont, L. Boulon, and Y. Dubé, “A comprehensive review of lithium-ion
batteries used in hybrid and electric vehicles at cold temperatures,” Appl. Energy,
vol. 164, pp. 99–114, 2016.
[3] Y. T. Huo, W. Hu, Z. Li, and Z. Rao, “Research on parameter identification and
state of charge estimation of improved equivalent circuit model of Li-ion battery
based on temperature effects for battery thermal management,” Int. J. Energy Res.,
vol. 44, no. 14, pp. 11583–11596, 2020.
[4] F. Guo, G. Hu, P. Zhou, J. Hu, and Y. Sai, “State of charge estimation in electric
vehicles at various ambient temperatures,” Int. J. Energy Res., vol. 44, no. 9, pp.
7357–7370, 2020.
[5] R. Xiong, L. Li, Q. Yu, Q. Jin, and R. Yang, “A set membership theory based
parameter and state of charge co-estimation method for all-climate batteries,” J.
Clean. Prod., vol. 249, no. 5, p. 119380, 2020.
[6] S. Yang, C. Deng, Y. Zhang, and Y. He, “State of charge estimation for lithium-ion
battery with a temperature-compensated model,” Energies, vol. 10, no. 10, 2017.
- 190 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

[7] J. Zhu et al., “An improved electro-thermal battery model complemented by

current dependent parameters for vehicular low temperature application,” Appl.
Energy, 2019.
[8] J. Shen et al., “State of charge estimation framework for lithium-ion batteries based
on square root cubature Kalman filter under wide operation temperature range,”
Int. J. Energy Res., vol. 45, no. 4, pp. 5586–5601, 2021.
[9] P. Kollmeyer, A. Hackl, and A. Emadi, “Li-ion battery model performance for
automotive drive cycles with current pulse and EIS parameterization,” 2017 IEEE
Transp. Electrif. Conf. Expo, ITEC 2017, pp. 486–492, 2017.
[10] Y. Bar-Shalom, X.-R. Li, and T. Kirubarajan, Estimation with Applications to
Tracking and Navigation, vol. 9. 2001.
[11] E. R. Dougherty, “Optimal Filtering,” Random Process. Image Signal Process., pp.
307–482, 2009.
[12] Y. Zheng, M. Ouyang, X. Han, L. Lu, and J. Li, “Investigating the error sources of
the online state of charge estimation methods for lithium-ion batteries in electric
vehicles,” J. Power Sources, vol. 377, pp. 161–188, Feb. 2018.
[13] H. A. P. Blom and Y. Bar-Shalom, “The Interacting Multiple Model Algorithm for
Systems with Markovian Switching Coefficients,” IEEE Trans. Automat. Contr.,
vol. 33, no. 8, pp. 780–783, 1988.
[14] E. Mazor, A. Averbuch, Y. Bar-Shalom, and J. Dayan, “Interacting Multiple
Model Methods in Target Tracking: A Survey,” Ieee Trans. Aerosp. Electron.
Syst., vol. 34, no. 1, 1998.
[15] K. Jo, K. Chu, and M. Sunwoo, “Interacting Multiple Model Filter-Based Sensor
Fusion of GPS With In-Vehicle Sensors for Real-Time Vehicle Positioning,” IEEE
Trans. Intell. Transp. Syst., vol. 13, no. 2, p. 973, 2012.
[16] Z. He, X. Bu, H. Yang, and Y. Song, “Interacting multiple model cubature Kalman
filter for geomagnetic/ infrared projectile attitude measurement,” Meas. J. Int.
Meas. Confed., vol. 174, no. January, p. 109077, 2021.
[17] A. Smiley and G. L. Plett, “An adaptive physics-based reduced-order model of an
aged lithium-ion cell, selected using an interacting multiple-model Kalman filter,”
J. Energy Storage, vol. 19, no. June, pp. 120–134, 2018.
[18] A. J. Smiley, W. K. Harrison, and G. L. Plett, “Postprocessing the outputs of an
interacting multiple-model Kalman filter using a Markovian trellis to estimate
parameter values of aged Li-ion cells,” J. Energy Storage, vol. 27, no. October
2019, p. 101043, 2020.
[19] X. Su, S. Wang, M. Pecht, L. Zhao, and Z. Ye, “Interacting multiple model particle
filter for prognostics of lithium-ion batteries,” 2017.
[20] S. Habibi, “The Smooth Variable Structure Structure Filter,” Proc. IEEE, vol. 95,
no. 5, 2007.
[21] S. A. Gadsden and S. R. Habibi, “A new robust filtering strategy for linear
systems,” J. Dyn. Syst. Meas. Control. Trans. ASME, vol. 135, no. 1, 2013.
[22] S. A. Gadsden, S. Habibi, and T. Kirubarajan, “Kalman and smooth variable
structure filters for robust estimation,” IEEE Trans. Aerosp. Electron. Syst., vol.
50, no. 2, pp. 1038–1050, 2014.
[23] S. A. Gadsden and S. R. Habibi, “A new form of the smooth variable structure
filter with a covariance derivation,” Proc. IEEE Conf. Decis. Control, pp. 7389–
7394, 2010.
[24] S. A. Gadsden, M. El Sayed, and S. R. Habibi, “Derivation of an optimal boundary
layer width for the smooth variable structure filter,” Proc. Am. Control Conf., pp.
4922–4927, 2011.
- 191 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

[25] S. A. Gadsden, D. Dunne, S. R. Habibi, and T. Kirubarajan, “Comparison of

extended and unscented Kalman, particle, and smooth variable structure filters on a
bearing-only target tracking problem,” Signal Data Process. Small Targets 2009,
vol. 7445, no. September 2009, p. 74450B, 2009.
[26] S. A. Gadsden, M. Al-Shabi, and S. R. Habibi, “Estimation Strategies for the
Condition Monitoring of a Battery System in a Hybrid Electric Vehicle,” ISRN
Signal Process., vol. 2011, no. 1, 2011.
[27] S. A. Gadsden, S. R. Habibi, and T. Kirubarajan, “A novel interacting multiple
model method for nonlinear target tracking,” 13th Conf. Inf. Fusion, Fusion 2010,
2010.
[28] T. R. Jow, S. A. Delp, J. L. Allen, J.-P. Jones, and M. C. Smart, “Factors Limiting
Li + Charge Transfer Kinetics in Li-Ion Batteries,” J. Electrochem. Soc., vol. 165,
no. 2, pp. A361–A367, 2018.
[29] P. Ahmadi and E. Kjeang, “Realistic simulation of fuel economy and life cycle
metrics for hydrogen fuel cell vehicles,” Int. J. Energy Res., vol. 41, no. 5, pp.
714–727, 2017.
[30] J. Wu, J. Liang, J. Ruan, N. Zhang, and P. D. Walker, “Efficiency comparison of
electric vehicles powertrains with dual motor and single motor input,” Mech.
Mach. Theory, vol. 128, pp. 569–585, 2018.
[31] G. L. Plett, “Extended Kalman filtering for battery management systems of LiPB-
based HEV battery packs Part 2. Modeling and identification,” J. Power Sources,
vol. 134, pp. 262–276, 2004.
[32] M. Lelie et al., “Battery management system hardware concepts: An overview,”
Appl. Sci., vol. 8, no. 4, 2018.
[33] L. Chang, B. Duan, P. Li, K. Zhang, C. Zhang, and L. Xiao, “Influence of
interconnect resistances on parallel- connected LiFePO 4 cells performance,” 2019
3rd Conf. Veh. Control Intell., no. 61527809, pp. 5–8, 2018.
[34] F. Lambert, “Tesla Model 3: Exclusive first look at Tesla’s new battery pack
architecture,” electrek.co, 2017. [Online]. Available:
https://electrek.co/2017/08/24/tesla-model-3-exclusive-battery-pack-architecture/.
[35] Tesla, “Model 3,” Tesla Press Information, 2018. [Online]. Available:
https://web.archive.org/web/20180315094117/https://www.tesla.com/presskit.

- 192 -
 Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

Chapter 7: EIS from Accelerated and Realistic Battery

Aging

Marvin Messing1,2, Tina Shoa2, Saeid Habibi1

1
Department of Mechanical Engineering, McMaster University, Hamilton, ON, Canada,
2
Cadex Electronics, Richmond, BC, Canada
This paper is published in 2021 IEEE Transportation Electrification Conference and
Expo (ITEC), 2021, pp. 720-725, doi: 10.1109/ITEC51675.2021.9490091. This paper is
republished here with permission4.

ABSTRACT

In this paper, the behaviour of lithium-ion batteries aged with an accelerated

(charge/discharge) and a realistic (drive cycle) protocol is compared using Electrochemical

Impedance Spectroscopy (EIS) characterization. Two equivalent circuit models are used to

further analyse degradation trends. The results show increased impedances for drive cycle

aged batteries above 90% State of Health (SoH) compared to charge/discharge aged cells

at the same SoH. Below 90% SoH, the opposite is found, with charge/discharged aged cells

showing higher impedance.

7.1 INTRODUCTION

Lithium-ion batteries (LiB) used in Electric Vehicles (EV) require precise battery

management systems (BMS) for safety monitoring, fault detection and state estimation.

The state of charge (SoC) [1], state of health (SoH) [2], and state of available power (SoaP)

4
In reference to IEEE copyrighted material which is used with permission in this thesis, the IEEE does not
endorse any of McMaster's products or services. Internal or personal use of this material is permitted. If
interested in reprinting/republishing IEEE copyrighted material for advertising or promotional purposes or
for creating new collective works for resale or redistribution, please go to
http://www.ieee.org/publications_standards/publications/rights/rights_link.html to learn how
to obtain a License from RightsLink.
- 193 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

[3] must be estimated by the BMS to ensure optimal operation and energy usage. The

accuracy of the state estimation not only impacts longevity and safety of the battery, but

also contributes to the cost and size of the battery pack [4], [5]. One of the main challenges

with SoH estimation is the long time it takes to obtain representative battery data to use for

algorithm development. Many studies that characterize batteries at different SoH employ

accelerated aging techniques, such as increasing operating temperature or current rates [6]–

[9]. However, the use of acceleration factors results in data which is not necessarily

representative of batteries that were aged as part of an EV battery pack. The impact of

different aging methods on the battery behaviour becomes especially apparent if the

batteries are analyzed with Electrochemical Impedance Spectroscopy (EIS). EIS is a non-

destructive material characterization technique that applies low amplitude, sinusoidal

current or voltage signals at different frequencies to a battery [10]. The EIS excitation

signals cause a battery current or voltage response which can be transformed using the Fast

Fourier Transform (FFT) and analyzed in the frequency domain. The EIS results are most

commonly visualized in the form of Nyquist plots, which show the complex impedance

pairs (real and imaginary) for each input signal frequency. The Nyquist plots exhibit several

features corresponding to the internal impedance response of the various layers of the

battery. The EIS data can further be analysed by using battery models such as Equivalent

Circuit Models (ECM).

Several studies show how EIS changes under different operating and aging conditions.

Schuster et al. [11] used EIS to track the evolution of ohmic resistance and charge transfer

resistance under different charging/discharging rates, temperatures, and voltage limits and
- 194 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

showed that high charging rates, high temperature, and high voltage limits cause the

resistances to increase significantly. However, a range of conditions are shown to result in

similar resistance evolution. In another study, Schuster et al. [12] showed how ohmic and

charge transfer resistances change with calendric, cyclic, and mild cyclic/high temperature

aging protocols. They highlight the importance of using realistic aging protocols to obtain

aged battery data when developing algorithms to utilize these resistances. Both studies use

simple resistor-capacitor (RC) based ECMs. Zhu et al. [13] showed how different charging

protocols change the evolution of EIS and corresponding fractional order model fits.

Olofsson et al. [13] showed how EIS results change for drive cycle tests performed at room

temperature and elevated temperature.

The objective of this study is to analyse the differences between EIS results from batteries

aged using accelerated and realistic cycling protocols by modeling EIS data with different

ECMs. The aging protocols are designed to use similar temperatures and charging methods,

while still resulting in different aging rates due to differences in cycling profiles. The

results show that the realistic test causes increased impedance measurements in the

frequency regions associated with SEI layer growth and charge transfer resistance between

100% and 90% SoH, even though it takes three times the amount of time to reach the same

SoH as the accelerated test. However, this trend reverses for data below 90% SoH, showing

faster SEI and charge transfer resistance increase with SoH for the accelerated test at 87%

and 85% SoH.

- 195 -
 Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

The paper is organized as follows. In Section 7.2, a simple EV model based on the Tesla

Model 3 is presented. This model is used to generate drive cycle current profiles for realistic

battery aging. The aging experiment design is introduced next, followed by a description

of the battery models used in this study. Section 7.3 contains the findings of the aging study

and EIS characterization for the accelerated and realistic aging methods. Finally,

concluding remarks are provided and future work is discussed in Section 7.4.

7.2 METHODS AND THEORY

7.2.1 ELECTRIC VEHICLE MODELING

In this study, standard drive cycles as provided by the US Environment Protection Agency

(EPA) are used to apply current profiles to battery cells such that the cells age in a way

similar to cells used in EVs. However, the EPA drive cycles merely provide speed vs. time

information for different scenarios. To convert this data into battery cell current, a vehicle

model is required. The first step is to calculate the power required to achieve the desired

speeds by overcoming opposing forces. Equation (7-1) shows how to calculate 𝑃𝑎𝑐𝑐 , the

power required to accelerate the mass of the vehicle, given the desired speed 𝑉, the mass

of the car 𝑚𝑐𝑎𝑟 and acceleration calculated using (7-2).

𝑃𝑎𝑐𝑐 = 𝑚𝑐𝑎𝑟 𝑎 𝑉 (7-1)

𝑎 = 𝑑𝑉⁄𝑑𝑡 (7-2)
In (7-3), the power 𝑃𝑟𝑜𝑙𝑙,𝑓𝑙𝑎𝑡 to overcome rolling resistance assuming a flat surface is

calculated given acceleration due to gravity 𝑔=9.81𝑚\𝑠 2 and a road surface parameter 𝐶𝑟

calculated using (7-4).

𝑃𝑟𝑜𝑙𝑙,𝑓𝑙𝑎𝑡 = 𝐶𝑟 𝑚𝑐𝑎𝑟 𝑔 (7-3)

- 196 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

The constants 𝐶𝑟𝑎 , 𝐶𝑟𝑏 , and 𝐶ℎ describe the conditions of the road surface.

𝐶𝑟 = (𝐶𝑟𝑎 + 𝐶𝑟𝑏 𝑉) 𝐶ℎ (7-4)

The power required to overcome drag, 𝑃𝑑𝑟𝑎𝑔 , is calculated in (7-5) given the air density

𝜌𝑎𝑖𝑟 , the frontal area of the vehicle 𝐴𝑓𝑟𝑜𝑛𝑡 and the drag coefficient of the vehicle 𝐶𝐷 .

1
𝑃𝑑𝑟𝑎𝑔 = 𝜌𝑎𝑖𝑟 𝐴𝑓𝑟𝑜𝑛𝑡 𝐶𝐷 𝑉 3 (7-5)
2

Equations (7-1), (7-3), and (7-5) provide the total power required to achieve the desired

speeds and is summarized in (7-6) as 𝑃𝑡𝑜𝑡 .

𝑃𝑡𝑜𝑡 = 𝑃𝑎𝑐𝑐 + 𝑃𝑟𝑜𝑙𝑙,𝑓𝑙𝑎𝑡 + 𝑃𝑑𝑟𝑎𝑔 (7-6)

Next, the power required from the battery pack is calculated in (7-7) by considering further

losses due to drive train efficiency 𝜂𝑑𝑟 , battery efficiency 𝜂𝑏𝑎𝑡𝑡 and parasitic power 𝑃𝑎𝑢𝑥

required by the auxiliary systems of the car.

𝑃𝑏𝑎𝑡𝑡 = [(𝑃𝑡𝑜𝑡⁄𝜂𝑑𝑟 ) + 𝑃𝑎𝑢𝑥 ]⁄𝜂𝑏𝑎𝑡𝑡 (7-7)

Using (7-8), the power can be converted into battery pack current, assuming a fixed

nominal voltage 𝐸𝑐𝑒𝑙𝑙,𝑛𝑜𝑚 and given the number of cells in parallel 𝑁𝑠𝑒𝑟𝑖𝑒𝑠 , which make up

the total pack voltage.

𝐼𝑝𝑎𝑐𝑘 = 𝑃𝑏𝑎𝑡𝑡 ⁄(𝑁𝑠𝑒𝑟𝑖𝑒𝑠 𝐸𝑐𝑒𝑙𝑙,𝑛𝑜𝑚 ) (7-8)

Finally, the battery cell current is obtained in (7-9) given the number of cells in parallel

𝑁𝑝𝑎𝑟𝑎𝑙𝑙𝑒𝑙 .

𝐼𝑐𝑒𝑙𝑙 = 𝐼𝑝𝑎𝑐𝑘 ⁄𝑁𝑝𝑎𝑟𝑎𝑙𝑙𝑒𝑙 (7-9)

In this study, the simple vehicle model described above is used to approximate the

behaviour of a Tesla Model 3 Long range EV. The parameters specific to the EV as well as

assumed efficiencies are shown in Table 7-1: Vehicle Model Parameters for Tesla Model

- 197 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

3 Long Range. The original Tesla Model 3 battery pack configuration was scaled by

reducing the number of modules in parallel to achieve slightly higher currents. This results

in utilization of more of the battery voltage range and reduced aging times. Nevertheless,

the current profiles remain sufficiently realistic to compare to accelerated charge/discharge

aging.

Table 7-1: Vehicle Model Parameters for Tesla Model 3 Long Range
Name Symbol Value
Mass of the vehicle. 𝑚𝑐𝑎𝑟 1730 kg
Rolling resistance calculation factor. 𝐶𝑟𝑎 0.0041
Rolling resistance calculation factor. 𝐶𝑟𝑏 0.000018
Rolling resistance calculation factor. 𝐶ℎ 1.5
Density of air. 𝜌𝑎𝑖𝑟 1.2
Vehicle frontal area. 𝐴𝑓𝑟𝑜𝑛𝑡 3.25 𝑚 2
Vehicle drag coefficient. 𝐶𝑑 0.23
Vehicle auxiliary power draw. 𝑃𝑎𝑢𝑥 350 W
Drivetrain efficiency. 𝜂𝑑𝑟 0.765
Battery efficiency. 𝜂𝑏𝑎𝑡𝑡 0.95
Number of modules in parallel. 𝑁𝑝𝑎𝑟𝑎𝑙𝑙𝑒𝑙,𝑠𝑐𝑎𝑙𝑒𝑑 18
Number of series cells per module. 𝑁𝑠𝑒𝑟𝑖𝑒𝑠 96
Nominal battery cell voltage. 𝐸𝑐𝑒𝑙𝑙,𝑛𝑜𝑚 3.6 V

7.2.2 BATTERY AGING

An extensive aging study was conducted with battery cells undergoing accelerated

charge/discharge aging and other cells undergoing aging in the form of a combination of

drive cycles to simulate battery operation as encountered in EVs. For each aging condition,

cells were aged to 85% SoH and EIS measurements were obtained at 100%, 95%, 90%,

87%, and 85% SoH. All cells used in this study are Samsung INR21700-50E, 4.8Ah

lithium-ion batteries. A custom battery tester was used to determine battery states to less

than 0.5% error, as well as to simulate drive cycles for realistic aging. The accelerated aging

- 198 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

(constant current – constant voltage charge at C/2, discharge at 1C) was performed on

Cadex C7400ER 4-channel battery analyzers. EIS (10 mV amplitude, 800Hz to 0.1Hz, 3-

hour pre-measurement rest) was performed using a Bio-logic SP150 potentiostat. Testing

was done at a constant temperature of 25°C using a Testequity 1007C thermal chamber.

Figure 7-1 shows the drive cycle aging protocol current, scaled by battery capacity (c-rate).

The protocol consists of a series of “Week Cycles” (trips from home to work and back plus

an evening errand and charge at home), a “Weekend Cycle” (trip through the city, onto a

highway, fast charging (1C) at destination, and back), as well as a capacity check (constant

current – constant voltage charge at C/2, discharge at C/5). The “Week Cycles” are a

combination of urban (UDDS) and urban-aggressive (US06) EPA drive cycles, and the

“Weekend Cycles” are a combination of US06 and highway (HWYFET) EPA drive cycles.

Home charging is done with constant current – constant voltage charge at C/2. The week

cycles are repeated 10 times, amounting to two work weeks of driving. This is followed by

a weekend cycle (i.e. weekend trips occur every 2 weeks). Capacity checks are done after

every 8 weekend cycles. If the capacity check indicates that the target capacity has been

reached, the EIS characterization is performed. This procedure is summarized in Figure

7-2.

- 199 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

UDDS US06 UDDS US06

-2

-4
0 50 100
Week Cycle Weekend Cycle

0
C-rate

-2
Fast Charge

-4
0 200 400 600 800
(a)

0
C-rate

-2

-4
0 200 400 600 800
Time (min)
(b)
Figure 7-1: Aging test load profiles in units of C-rate for drive cycle aging test (a) and
charge/discharge aging test (b).

- 200 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

Figure 7-2: Aging test procedure showing the usage of different drive cycle protocols,
different charge rates, and EIS steps.

7.2.3 BATTERY MODELING

To understand the degradation behaviour of the batteries as they age, two ECMs were used

in this study. The first model is a fractional order impedance model which can follow the

shape of the data very well and fitting can be automated if suitable initial conditions for the

model parameters are provided. The second model is a 2 nd order RC based ECM. Some of

the parameters obtain from the RC model can reveal more realistic trends even though the

model fit is poor due to the ideal nature of the RC circuit elements. Figure 7-3a shows

Nyquist plots for EIS data at 100% SoH and 90% SoH and corresponding model fits for

the fractional order model and the RC model. The fractional impedance model shown in

Figure 7-3b consists of an inductor 𝐿, representing inductance effects due to cell windings

- 201 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

and cables, a series resistance 𝑅0 , for ohmic resistance effects, a resistance 𝑅1 in parallel

with a constant phase element (CPE) defined by 𝑄1 and 𝛼1 , together known as a ZArc

element, to model the solid electrolyte interface impedance, another ZArc element defined

by 𝑅2 , 𝑄2 and 𝛼2 to model charge transfer impedance, and finally a CPE defined by 𝑄3

and 𝛼3 to model diffusion. The complex impedance 𝑍𝑓𝑟 of this ECM changes with

frequency 𝜔 according to (9), where 𝜏𝑘 = (𝑅𝑘 𝑄𝑘 )1⁄𝛼𝑘 .

𝑅𝑘 1
𝑍𝑓𝑟 (𝜔) = 𝑖𝜔𝐿 + 𝑅0 + ∑2𝑘=1 1+𝑅 ) 𝛼𝑘
+𝑄 𝛼 (9)
𝑘 𝑄𝑘 (𝑖𝜔𝜏𝑘 3 (𝑖𝜔) 3

The RC model contains two RC pairs instead of the ZArc and CPE elements as shown in

Figure 7-3c. The RC branches are used to model the charge transfer impedance and

diffusion impedance, but neglecting the SEI layer impedance. This is done because the

exact positioning of an additional RC semi-circle to model SEI layer impedance is

somewhat arbitrary and does not yield useful information for this study. This is discussed

further in Section 7.3.2. The complex impedance 𝑍𝑟𝑐 of the RC model is described by (10),

where 𝜏2 = 𝑅2 𝐶2 and 𝜏3 = 𝑅3 𝐶3 .

𝑅2 𝑅3
𝑍𝑟𝑐 (𝜔) = 𝑖𝜔𝐿 + 𝑅0 + + (10)
1+𝑖𝜔𝜏2 1+𝑖𝜔𝜏3

The fractional order ECM was parameterized using the Particle Swarm Optimization (PSO)

method. The PSO was also used for the RC ECM, however, 𝑅0 and 𝑅2 were fixed such that

𝑅0 coincides with the real impedance value corresponding to the smallest imaginary value.

𝑅2 was fixed to the difference between the lowest point in the charge transfer region and

𝑅0 [12].

- 202 -
 Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

-Z'' (mΩ)
2
100 Fractional Fit
90 RC Fit
0
25 30 35 40
-2 Z' (mΩ)
(a)

(b)

(c)

Figure 7-3: Equivalent circuit battery model fits to Nyquist plots (a), fractional order model
circuit (b), RC-based circuit without SEI layer branch (c).

7.3 RESULTS AND DISCUSSION

7.3.1 VOLTAGE AND TEMPERATURE RESPONSE

Figure 7-4a shows the voltage response of battery cells at 100% SoH resulting from drive

cycle (week cycle followed by weekend cycle) and charge/discharge protocols. The same

is shown in Figure 7-4b but for batteries at 85% SoH. The voltage response of the

charge/discharge aging profile did not change much, only charging shows longer CV mode

periods for low SoH. The voltage data from the drive cycle, however, shows a significantly

altered response. At low SoH, the drive cycles use a much larger portion of the battery

voltage range, with the weekend cycle nearly using the full range. This will also decrease
- 203 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

the average voltage of the cell and could possibly impact degradation. The fresh cell only

uses the top half of the voltage range (3.6 to 4.2 V) and has a higher average voltage which

can accelerate degradation. This degradation effect due to voltage may be slower for the

aged cell which uses the full voltage range (2.5 to 4.2V).

Charge/Discharge Drive Cycle

4.0
Voltage (V)

3.5

3.0

2.5
0 180 360 540 720 900 1080
(a)

4.0
Voltage (V)

3.5

3.0

2.5
0 180 360 540 720 900 1080
Time (min)
(b)

Figure 7-4: Drive cycle aged and charge/discharge aged voltage profiles for fresh cells (a)
and cells at 85% SoH (b).

Figure 7-5a shows a comparison of the temperature profiles between fresh (100% SoH) and

aged (85% SoH) batteries aged with the drive cycle profile. The temperature data for the

aged cell shows a slightly wider temperature range, reaching both lower and higher

- 204 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

temperatures compared to the fresh cell. The lower temperatures reached in the aged case

are likely due to the lower average voltage. The temperature data for fresh and aged cells

aged with the charge/discharge profile is shown in Figure 7-5b. Here, the temperature peaks

for the aged cell are lower for almost all peaks shown. This is likely a result of the extended

CV region, where currents are dropping and less heat is generated. In contrast, the fresh

cell spends more time at the 0.5C charging current, generating more heat. For both drive

cycle aged and charge/discharge aged cells, charging results in the biggest temperature

increase, with the overall maximum of close to 30°C reached during the fast charge as part

of the weekend cycle. In general, however, the temperatures are similar between the two

aging protocols as well as between fresh and aged cells, with differences not exceeding

0.5°C. Therefore, temperature is expected to affect the two aging protocols in the same way

and to not cause any significant differences in aging behaviour between charge/discharge

aged and drive cycle aged battery cells.

- 205 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

Fresh Aged
30

Temperature (°C)
28

26

24
0 180 360 540 720 900 1080
29 (a)

28
Temperature (°C)

27

26

25

24
0 180 360 540 720 900 1080
Time (min)
(b)

Figure 7-5: Comparison of temperature profiles for fresh cells and cells at 85% SoH during
drive cycle aging protocol (a) and charge/discharge protocol (c).

7.3.2 EIS CHARACTERIZATION

From the realistic drive cycle based aging tests and the accelerated charge/discharge aging

tests, EIS datasets were obtained for batteries at 100%, 95%, 90%, 87%, and 85% SoH.

The charge/discharge aging protocol required 3 months of continuous aging to achieve 85%

SoH. The drive cycle aging protocol required 9 months to reach the same SoH, 3 times the

amount of time. Figure 7-6a shows the Nyquist plots from EIS data recorded with batteries

that underwent drive cycle aging conditions, and Figure 7-6b shows the Nyquist plots from

EIS data recorded with batteries that underwent charge/discharge aging. For the drive cycle
- 206 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

aged batteries, three datasets were available down to 90% SoH. Above 90% SoH the

variance between the three sets is negligible but at 90% SoH the variance is larger and the

standard deviation (Std. Dev.) is shown around the 90% SoH mean Nyquist plot in Figure

7-6a. Between 100% and 90% SoH, the drive cycle aged data shows a faster increase in the

semi-circles compared to the charge/discharge aged results. This suggests that impedance

growth has a bigger impact on battery degradation when drive cycles are used above 90%

SoH. The charge/discharge protocol reached 90% SoH much faster than the drive cycle

protocol, however, with significantly smaller impedance growth. Below 90% SoH, the

charge/discharge aged Nyquist plots become much larger than the drive cycle aged ones,

indicating that in this lower SoH phase, charge/discharge cycles result in higher impedance

increase at the same SoH. The variance observed for the drive cycle data at 90% SoH, while

noticeable, is still small compared to the change between SoH targets and even more so,

when compared to charge/discharge aging.

- 207 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

10.0 Drive Cycle Aged

100 (mean)
95 (mean)
7.5
90 (mean)

-Z'' (mΩ)
5.0 87
85
2.5 Std. Dev.
Fractional Fit
0.0
25 30 35 40 45 50
Z' (mΩ)
(a)
10.0 Charge/Discharge Aged
95
90
7.5
87
-Z'' (mΩ)

5.0 85

2.5

0.0
25 30 35 40 45 50
Z' (mΩ)
(b)
Figure 7-6: Nyquist plot evolution at various SoH targets with fractional order model fits
for drive cycle aged EIS data (a) and charge/discharge aged EIS data (b).

Figure 7-7 shows the evolution with SoH of the battery model parameters from the two

battery models introduced in Section 7.2.3. Figure 7-7a shows an increase in 𝑅0 obtained

from the RC model, but no clear trend for 𝑅0 from the fractional model. Between 100%

and 90% SoH, 𝑅0 from charge/discharge aging data is within one standard deviation of 𝑅0

from the drive cycle aging data. Figure 7-7b (fractional model only) shows a large increase

in 𝑅1 for drive cycle aged cells down to 90% SoH and a smaller increase for the

charge/discharge aged cells in the same SoH region. Below 90% SoH, 𝑅1 eventually

decreases for drive cycle aged cells. For the charge/discharge aged cells, 𝑅1 continues to

- 208 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

increase up to 87% SoH, and decreases after that. Assuming the variance of the drive cycle

data below 90% SoH is similar (or larger) to the variance at 90% SoH, 𝑅1 from the two

aging datasets becomes indistinguishable. This behaviour is not obvious from the Nyquist

plots. However, the charge/discharge Nyquist data shows a much more pronounced

inflection point in the SEI layer impedance region. A more defined inflection point leaves

the PSO algorithm with fewer options for the exact placement of the SEI layer impedance

model branch. Figure 7-7c shows trends for 𝑅2 which are similar for both models, with the

RC model only offset by the missing SEI layer resistance. For both models, 𝑅2 increases

faster for drive cycle aged batteries to 90% SoH, and faster for charge/discharge aged cells

bellow 90% SoH. The variance for the drive cycle data is also small in all available cases

for the 𝑅2 parameter. Therefore, 𝑅2 is the most reliable parameter to consider for both

fractional and RC models, when comparing EIS results. The inflection point in the Nyquist

plot defining 𝑅2 is well defined in all cases, resulting in clear model parameter trends.

Considering the Nyquist plots and model parameters, it is clear that a change in impedance

growth rate with SoH occurs around 90% SoH. The voltage ranges utilized by the different

aging protocols (as discussing in Section 7.3.1) may offer one explanation. The higher

average voltage of fresh cells during drive cycle aging can cause accelerated degradation,

and the average voltage decreases as the battery ages. For the charge/discharge aged cells,

the duration spent at the maximum voltage during charge increases as the battery ages,

causing the opposite degradation trend. This also happens in the drive cycle aging case, but

the lower average voltage range during discharge seems to outweigh the longer CV mode

periods. The higher frequency cycles of the drive cycle protocol may also be a factor
- 209 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

contributing to the aging behaviour, but is difficult to isolate among more dominant

stressors such as temperature and voltage.

Drive Cycle Charge/Discharge Std. Dev.

26 RC-Model

24
R0 (mΩ)

22

Fractional-Model
20
8 (a)

6
R1 (mΩ)

2 Fractional-Model

0
30 (b)

20
R2 (mΩ)

10 RC-Model

Fractional-Model
0
100 95 90 85
SoH (%)
(c)

Figure 7-7: Battery model parameter evolution with SoH for drive cycle and
charge/discharge aging for ohmic parameter R_0 (a), SEI layer parameter R_1, and charge
transfer parameter R_2 (c).

- 210 -
 Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

7.4 CONCLUSIONS AND FUTURE WORK

In this work, accelerated (charge/discharge) and realistic (drive cycle based) aging

protocols were employed to age LiBs. EIS results are presented for 100% to 85% SoH

samples from both aging protocols and show increasing impedance with SoH. Two battery

models are used to further analyse the EIS data. Impedance initially increases at a faster

rate with SoH for the drive cycle aging dataset, but is overtaken by the charge/discharge

aging dataset after 90% SoH. The voltage profiles for the different aging sets suggest higher

average voltages at the beginning of drive cycle testing and at the end (85% SoH) of

charge/discharge testing may explain the observed trends. The impact of drive cycle

frequencies on battery degradation should be investigated further.

ACKNOWLEDGMENT

This research was conducted under a CREATE grant from the Natural Sciences and

Engineering Research Council of Canada.

REFERENCES
[1] R. Xiong, J. Cao, Q. Yu, H. He, and F. Sun, “Critical Review on the Battery State
of Charge Estimation Methods for Electric Vehicles,” IEEE Access, vol. 6, pp.
1832–1843, 2018.
[2] L. Ungurean, G. Cârstoiu, M. V Micea, and V. Groza, “Battery state of health
estimation : a structured review of models , methods and commercial devices,” no.
July 2016, pp. 151–181, 2017.
[3] A. Farmann and D. U. Sauer, “Comparative study of reduced order equivalent
circuit models for on-board state-of-available-power prediction of lithium-ion
batteries in electric vehicles,” Appl. Energy, vol. 225, no. May, pp. 1102–1122,
2018.
[4] M. U. Cuma and T. Koroglu, “A comprehensive review on estimation strategies
used in hybrid and battery electric vehicles,” Renew. Sustain. Energy Rev., vol. 42,
pp. 517–531, 2015.
[5] M. A. Hannan, M. S. H. Lipu, A. Hussain, and A. Mohamed, “A review of lithium-
ion battery state of charge estimation and management system in electric vehicle
applications: Challenges and recommendations,” Renew. Sustain. Energy Rev.,
- 211 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

vol. 78, no. August 2016, pp. 834–854, 2017.

[6] A. Maheshwari, M. Heck, and M. Santarelli, “Cycle aging studies of lithium nickel
manganese cobalt oxide-based batteries using electrochemical impedance
spectroscopy,” Electrochim. Acta, 2018.
[7] Y. Gao, J. Jiang, C. Zhang, W. Zhang, Z. Ma, and Y. Jiang, “Lithium-ion battery
aging mechanisms and life model under different charging stresses,” J. Power
Sources, vol. 356, pp. 103–114, 2017.
[8] L. Su et al., “Identifying main factors of capacity fading in lithium ion cells using
orthogonal design of experiments,” Appl. Energy, vol. 163, pp. 201–210, 2016.
[9] X. Han, M. Ouyang, L. Lu, J. Li, Y. Zheng, and Z. Li, “A comparative study of
commercial lithium ion battery cycle life in electrical vehicle : Aging mechanism
identi fi cation,” J. Power Sources, vol. 251, pp. 38–54, 2014.
[10] A. Lasia, Electrochemical Impedance Spectroscopy EIS, and Corrosion. 2011.
[11] S. F. Schuster et al., “Nonlinear aging characteristics of lithium-ion cells under
different operational conditions,” 2015.
[12] S. F. Schuster, M. J. Brand, C. Campestrini, M. Gleissenberger, and A. Jossen,
“Correlation between capacity and impedance of lithium-ion cells during calendar
and cycle life,” 2015.
[13] J. Zhu et al., “Investigation of lithium-ion battery degradation mechanisms by
combining differential voltage analysis and alternating current impedance,” J.
Power Sources, vol. 448, p. 227575, 2020.

- 212 -
 Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

Chapter 8: Summary, Conclusions and Recommendations

In this chapter, the major conclusions of this thesis are summarized and recommendations

are made for possible future projects based on the findings.

8.1 RESEARCH SUMMARY

This thesis presents several advancements in the area of battery characterization through

the development, improvement and implementation of methods including Electrochemical

Impedance Spectroscopy (EIS), Deep Neural Networks (DNN), and adaptive filters such

as the Interacting Multiple Model (IMM) filter. The estimation of battery State of Charge

(SoC) and State of Health (SoH) is a critical component of Battery Management Systems

(BMS), especially for applications in Electric Vehicles (EV). Many methods exist,

however, each with their own advantages and drawbacks. The main objective of this thesis

is to investigate and improve different state estimation techniques by characterizing battery

dynamics under different temperatures, aging conditions and relaxation effects.

Chapters 2, 3, 4 and 5 utilize EIS to estimate battery SoC and SoH. One of the main

drawbacks of EIS is the requirement for long rest times to avoid interference from the

battery relaxation effect. In Chapter 2, the relaxation effect is characterized with EIS and

its impact on battery modeling is investigated. The findings from this work provide insight

into the relaxation behaviour of batteries and the impact of relaxation on battery model

parameters. In addition, impedance models such as Fractional Order Models (FOM) were

found to accurately track the relaxation effect. The combined effects of relaxation, SoC,

and temperature were analysed to show that relaxation effects must be carefully considered

- 213 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

when processing EIS measurements. With the growing interest in improving EIS methods

to be more suitable for BMS applications, the requirement for long rest times becomes

impractical. Yet the results discussed in Chapter 2 show an increased impact of the

relaxation effect on model parameters as rest times are reduced. Therefore, the short-term

relaxation effect must be investigated in more detail to determine its impact on EIS under

different operating conditions. This is the subject of Chapter 3.

Chapter 3 presents methodologies to significantly decrease the rest times required for EIS

under a wide range of battery operating conditions, circumventing a major hurdle to the use

of EIS in real time applications. The impact of the relaxation effect during short rest times,

where it is most severe, is investigated for different SoCs, temperatures, and excitation

amplitudes. A filtering technique is developed that utilizes the Voigt circuit to smooth out

EIS measurements. Using this filter, rest times of as low as 5 minutes are shown to produce

valid EIS results following low battery excitation. If the initial excitation is large, rest times

can be as low as 30 minutes, provided the SoC remains high. The findings of this work

provide operating points for the most time efficient EIS measurements.

Chapter 4 shows how EIS can be used to extract battery SoH information from the

relaxation effect and presents a method for battery SoH estimation to within 1%. The

relaxation effect is found to slow down as batteries age and the rate of the slowdown can

be detected with EIS. Using a FOM and an equivalent resistance model parameter, the rate

of change of the relaxation effect can be related to the battery SoH. A double exponential

model is used to capture two sperate time constants, a fast time constant between 100% and

- 214 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

90% SoH, and a slow time constant below 90% SoH. This second region exhibits a near

linear trend, allowing for increased SoH estimation accuracy at lower SoH. Since batteries

become unsuitable for EV applications below 80% SoH, the estimation of SoH becomes

more important closer to 80%. With accurate SoH estimates, a BMS can act to ensure the

safe operation of the battery as it ages while still maintaining performance and maximum

battery utilization. The SoH method combined with the findings of Chapter 3 can bring EIS

one step closer to being a viable option for battery state estimation and diagnostics in BMS

applications.

In Chapter 5, EIS data is used together with DNNs to estimate battery SoC to within 5%

for a dataset which represents a wide range of battery operating conditions. Two DNNs are

designed and compared. The first DNN accepts EIS data in frequency domain (real and

imaginary pairs) as inputs to estimate SoC as the only output. For the second DNN, the EIS

data was first used to parameterize a FOM. The parameters of the FOM were then used as

inputs to the second DNN to estimate SoC. Both DNNs showed similar accuracies,

however, using model parameters instead of EIS data directly, resulted in overfitting. The

use of DNNs to estimate battery states from EIS is promising, especially since the shape of

EIS results changes in a complex way with battery operating conditions and states. This

kind of complexity presents a challenge for model-based state estimation methods but can

be handled well by DNNs, as long as a sufficiently large dataset is available for training.

Model-based methods typically require only a fraction of parameters compared to DNNs,

which is a concern of real-time, embedded applications. However, methods such a pruning

can be used to reduce DNN model parameters.

- 215 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

Chapter 6 combines the IMM filter with the Smooth Variable Structure (SVSF) filter to

significantly improve battery SoC estimation at low temperatures. An estimation error of

less than 2% is demonstrated. At low temperatures, the battery response to current becomes

increasingly non-linear. This means that a single Equivalent Circuit Model (ECM) cannot

accurately reflect the battery behaviour across the entire range of operating currents. As a

result, state estimation filters such as the Extended Kalman Filter (EKF) become less

accurate as temperature decreases. The IMM allows several EKFs (each with a different

ECM) to run in parallel, and blends the different state estimates into an overall improved

estimate. The IMM uses mode probabilities to weight each EKF estimate based on the

likelihood of the battery model representing the real battery response. Three battery models

were trained each for a different current range and used with three EKFs combined with

the IMM. This IMM-EKF algorithm showed a significant improvement over the single

model EKF for low temperature SoC estimation. In addition, it was shown that the

estimation errors could be further reduced by using the SVSF instead of the EKF. The

adaptive properties of the SVSF result in more robust estimation and estimation errors of

less than 2% were demonstrated for new and aged battery data.

Chapter 7 investigates the impact of different battery aging pathways on EIS measurements.

An extensive aging study was designed to compare accelerated, charge-discharge based

aging to more realistic, drive cycle based aging. The accelerated test showed a faster

impedance increase between 100% and 90% SoH, but was overtaken by the impedance

increase of the realistic test below 90% SoH. The results were analysed with both ECMs

and FOMs. Understanding how EIS results change depending on the way a battery was
- 216 -
 Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

aged has a profound impact on experimental design, battery modeling, and, by extension,

battery state estimation. The results shown in Chapter 7 can be used to inform accelerated

aging experiments to significantly reduce testing time, while still obtaining realistic results.

The research findings of this thesis advance the field of battery engineering by providing

improved techniques for state estimation as well as insights into battery aging and

impedance behaviour. The methods presented can improve battery management strategies,

leading to more efficient EV battery packs. As a result, the cost of EVs can be reduced,

increasing accessibility to a wider market and in turn contribute to the reduction of air

pollution and Greenhouse Gas (GHG) emissions.

8.2 RECOMMENDATION FOR FUTURE RESEARCH

Many areas of future work can be recommended surrounding the topic of EIS and its

applications to EVs. Some of the methods shown in this thesis are intended to improve the

EIS analysis for real time applications, however, the methods were only tested with lab-

grade equipment and not combined with Battery Management System (BMS) hardware.

Implementing EIS as part of a BMS itself is a large area of research.

The proposed SoH estimation method combining EIS and the relaxation effect was

validated using charge/discharge cycle aged batteries only. This study can be extended to

include additional aging pathways and different types of batteries to understand the

robustness of the method.

The IMM-SVSF method presented for increasing SoC estimation accuracy at low

temperatures shows promising performance. However, since multiple filters run in parallel,
- 217 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

some concern exists surrounding computational efficiency. More research is required to

understand the computational impact of the IMM, develop efficiency improvements, and

demonstrate its functionality on real-time systems.

Another area of research that continues to be relevant is the aging behaviour of batteries.

The results shown in this thesis demonstrate the need for careful consideration of the aging

pathways when conducting experiments and analysing results. Additional aging studies

utilizing EIS characterization and a wider range of stressors would greatly benefit the

understanding of the relationship between accelerated and realistic aging tests.

- 218 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

References
[1] Health Effects Institute, State of Global Air 2019. Special Report., Boston, MA,
2019. http://www.stateofglobalair.org/sites/default/files/soga_2019_report.pdf
(accessed April 7, 2019).
[2] Intergovernmental Panel on Climate Change, Climate Change 2021- The Physical
Science Basis, 2021.
[3] Center for Climate and Energy Solutions, Global Emissions, Glob. Emiss. (2021).
https://www.c2es.org/content/international-emissions/.
[4] US EPA, Fast Facts on Transportation Greenhouse Gas Emissions, Green Veh.
Guid. (2019). https://www.epa.gov/greenvehicles/fast-facts-transportation-
greenhouse-gas-emissions.
[5] R. Ahmed, M. El Sayed, I. Arasaratnam, Jimi Tjong, S. Habibi, Reduced-Order
Electrochemical Model Parameters Identification and SOC Estimation for Healthy
and Aged Li-Ion Batteries Part I: Parameterization Model Development for
Healthy Batteries, IEEE J. Emerg. Sel. Top. Power Electron. 2 (2014) 659–677.
doi:10.1109/JESTPE.2014.2331059.
[6] E. Din, C. Schaef, K. Moffat, J.T. Stauth, A scalable active battery management
system with embedded real-time electrochemical impedance spectroscopy, IEEE
Trans. Power Electron. 32 (2017) 5688–5698. doi:10.1109/TPEL.2016.2607519.
[7] A. Farmann, D.U. Sauer, Comparative study of reduced order equivalent circuit
models for on-board state-of-available-power prediction of lithium-ion batteries in
electric vehicles, Appl. Energy. 225 (2018) 1102–1122.
doi:10.1016/j.apenergy.2018.05.066.
[8] A. Opitz, P. Badami, L. Shen, K. Vignarooban, A.M. Kannan, Can Li-Ion batteries
be the panacea for automotive applications?, Renew. Sustain. Energy Rev. 68
(2017) 685–692. doi:10.1016/j.rser.2016.10.019.
[9] Y. Wu, Lithium-Ion Batteries: Fundamentals and Applications, CRC Press, Boca
Raton, 2015.
[10] T.R. Jow, S.A. Delp, J.L. Allen, J.-P. Jones, M.C. Smart, Factors Limiting Li +
Charge Transfer Kinetics in Li-Ion Batteries, J. Electrochem. Soc. 165 (2018)
A361–A367. doi:10.1149/2.1221802jes.
[11] M.M. Kabir, D. Demirocak, Degradation mechanisms in Li-ion batteries: a state-
of- the-art review, Int. J. Energy Res. 41 (2017) 1963–1986. doi:10.1002/er.
[12] A. Mauger, C.M. Julien, Critical review on lithium-ion batteries: are they safe?
Sustainable?, Ionics (Kiel). 23 (2017) 1933–1947. doi:10.1007/s11581-017-2177-
8.
[13] M.B. Dixit, J.-S. Park, P. Kenesei, J. Almer, K.B. Hatzell, Status and prospect of
in situ and operando characterization of solid-state batteries , Energy Environ. Sci.
(2021). doi:10.1039/d1ee00638j.
[14] Z. Gao, H. Sun, L. Fu, F. Ye, Y. Zhang, W. Luo, Y. Huang, Promises, Challenges,
and Recent Progress of Inorganic Solid-State Electrolytes for All-Solid-State
Lithium Batteries, Adv. Mater. 30 (2018) 1705702. doi:10.1002/adma.201705702.
[15] C. Zhao, L. Liu, X. Qi, Y. Lu, F. Wu, J. Zhao, Y. Yu, Y.-S. Hu, L. Chen, Solid-
State Sodium Batteries, Adv. Energy Mater. 8 (2018) 1703012.
doi:10.1002/aenm.201703012.
[16] V.T. DOE, Office, Batteries 2017, (2017).
[17] Y. Gao, J. Jiang, C. Zhang, W. Zhang, Z. Ma, Y. Jiang, Lithium-ion battery aging
mechanisms and life model under different charging stresses, J. Power Sources.
356 (2017) 103–114. doi:10.1016/j.jpowsour.2017.04.084.
- 219 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

[18] L. Su, J. Zhang, C. Wang, Y. Zhang, Z. Li, Y. Song, T. Jin, Z. Ma, Identifying
main factors of capacity fading in lithium ion cells using orthogonal design of
experiments, Appl. Energy. 163 (2016) 201–210.
doi:10.1016/j.apenergy.2015.11.014.
[19] T. Waldmann, M. Wilka, M. Kasper, M. Fleischhammer, M. Wohlfahrt-Mehrens,
Temperature dependent ageing mechanisms in Lithium-ion batteries - A Post-
Mortem study, J. Power Sources. 262 (2014) 129–135.
doi:10.1016/j.jpowsour.2014.03.112.
[20] W. Waag, S. Käbitz, D.U. Sauer, Experimental investigation of the lithium-ion
battery impedance characteristic at various conditions and aging states and its
influence on the application, Appl. Energy. 102 (2013) 885–897.
doi:10.1016/j.apenergy.2012.09.030.
[21] F. Leng, C.M. Tan, M. Pecht, Effect of Temperature on the Aging rate of Li Ion
Battery Operating above Room Temperature, Sci. Rep. 5 (2015) 1–12.
doi:10.1038/srep12967.
[22] J. Schmitt, A. Maheshwari, M. Heck, S. Lux, M. Vetter, Impedance change and
capacity fade of lithium nickel manganese cobalt oxide-based batteries during
calendar aging, J. Power Sources. 353 (2017) 183–194.
doi:10.1016/j.jpowsour.2017.03.090.
[23] M. Ecker, J.B. Gerschler, J. Vogel, S. K??bitz, F. Hust, P. Dechent, D.U. Sauer,
Development of a lifetime prediction model for lithium-ion batteries based on
extended accelerated aging test data, J. Power Sources. 215 (2012) 248–257.
doi:10.1016/j.jpowsour.2012.05.012.
[24] D. Stroe, M. Swierczy, A. Stan, R. Teodorescu, S.J. Andreasen, Experimental
Investigation on the Internal Resistance of Lithium Iron Phosphate Battery Cells
during Calendar Ageing, (2013) 6734–6739.
[25] A. Eddahech, O. Briat, J. Vinassa, Performance comparison of four lithium e ion
battery technologies under calendar aging, Energy. 84 (2015) 542–550.
doi:10.1016/j.energy.2015.03.019.
[26] J.M. Hooper, J. Marco, G.H. Chouchelamane, C. Lyness, Vibration durability
testing of nickel manganese cobalt oxide (NMC) lithium-ion 18,650 battery cells,
Energies. 9 (2016) 1–27. doi:10.3390/en9010052.
[27] L. Somerville, J.M. Hooper, J. Marco, A. McGordon, C. Lyness, M. Walker, P.
Jennings, Impact of vibration on the surface film of lithium-ion cells, Energies. 10
(2017) 1–12. doi:10.3390/en10060741.
[28] J.M. Hooper, J. Marco, G.H. Chouchelamane, J.S. Chevalier, D. Williams, Multi-
axis vibration durability testing of lithium ion 18650 NCA cylindrical cells, J.
Energy Storage. 15 (2018) 103–123. doi:10.1016/j.est.2017.11.006.
[29] F.M. Kindermann, A. Noel, S. V. Erhard, A. Jossen, Long-term equalization
effects in Li-ion batteries due to local state of charge inhomogeneities and their
impact on impedance measurements, Electrochim. Acta. 185 (2015) 107–116.
doi:10.1016/j.electacta.2015.10.108.
[30] A. Barai, G. Chouchelamane, Y. Guo, A. McGordon, P. Jennings, A study on the
impact of lithium-ion cell relaxation on electrochemical impedance spectroscopy,
J. Power Sources. 280 (2015) 74–80. doi:10.1016/j.jpowsour.2015.01.097.
[31] A. Lasia, Electrochemical Impedance Spectroscopy EIS, and Corrosion, Springer,
New York, 2014.
[32] M. Messing, Case Study: EIS Model Fitting and Analysis, HEVPD&D-CREATE.
(2019). http://hevpdd.ca/publications-and-case-studies/eis-model-fitting-and-
analysis/.
- 220 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

[33] M. Abedi Varnosfaderani, D. Strickland, A Comparison of Online Electrochemical

Spectroscopy Impedance Estimation of Batteries, IEEE Access. 6 (2018) 23668–
23677. doi:10.1109/ACCESS.2018.2808412.
[34] N.A. Chaturvedi, R. Klein, J. Christensen, J. Ahmed, A. Kojic, Modeling,
estimation, and control challenges for lithium-ion batteries, IEEE Control Syst.
Mag. (2010) 1997–2002. doi:10.1109/acc.2010.5531623.
[35] R. Xiong, J. Cao, Q. Yu, H. He, F. Sun, Critical Review on the Battery State of
Charge Estimation Methods for Electric Vehicles, IEEE Access. 6 (2018) 1832–
1843. doi:10.1109/ACCESS.2017.2780258.
[36] Z. Li, J. Huang, B.Y. Liaw, J. Zhang, On state-of-charge determination for lithium-
ion batteries, J. Power Sources. 348 (2017) 281–301.
doi:10.1016/j.jpowsour.2017.03.001.
[37] A. Rodríguez, G.L. Plett, Controls-oriented models of lithium-ion cells having
blend electrodes. Part 1: Equivalent circuits, J. Energy Storage. 11 (2017) 162–177.
doi:10.1016/j.est.2017.02.004.
[38] R. Ahmed, Modeling and State of Charge Estimation of Electric Vehicle, Dr. Diss.
(2014). doi:http://dx.doi.org/10.1016/j.idairyj.2012.02.006.
[39] L. Zhang, H. Peng, Z. Ning, Z. Mu, C. Sun, Comparative Research on RC
Equivalent Circuit Models for Lithium-Ion Batteries of Electric Vehicles, Appl.
Sci. 7 (2017) 1002. doi:10.3390/app7101002.
[40] H.H. Afshari, M. Attari, R. Ahmed, A. Delbari, S. Habibi, T. Shoa, Reliable state
of charge and state of health estimation using the smooth variable structure filter,
Control Eng. Pract. 77 (2018) 1–14. doi:10.1016/j.conengprac.2018.04.015.
[41] J.P. Christopherson, Battery Test Manual For Electric Vehicles, 2015.
http://www.inl.gov (accessed May 20, 2019).
[42] M. Schönleber, C. Uhlmann, P. Braun, A. Weber, E. Ivers-Tiffée, A Consistent
Derivation of the Impedance of a Lithium-Ion Battery Electrode and its
Dependency on the State-of-Charge, Electrochim. Acta. 243 (2017) 250–259.
doi:10.1016/j.electacta.2017.05.009.
[43] U. Krewer, F. Röder, E. Harinath, R.D. Braatz, B. Bedürftig, R. Findeisen,
Review—Dynamic Models of Li-Ion Batteries for Diagnosis and Operation: A
Review and Perspective, J. Electrochem. Soc. 165 (2018) A3656–A3673.
doi:10.1149/2.1061814jes.
[44] L. De Sutter, Y. Firouz, J. De Hoog, N. Omar, J. Van Mierlo, Battery aging
assessment and parametric study of lithium-ion batteries by means of a fractional
differential model, Electrochim. Acta. 305 (2019) 24–36.
doi:10.1016/j.electacta.2019.02.104.
[45] X. Zhu, L. Fernández Macía, J. Jaguemont, J. de Hoog, A. Nikolian, N. Omar, A.
Hubin, Electrochemical impedance study of commercial
LiNi0.80Co0.15Al0.05O2electrodes as a function of state of charge and aging,
Electrochim. Acta. 287 (2018) 10–20. doi:10.1016/j.electacta.2018.08.054.
[46] H. Mu, R. Xiong, H. Zheng, Y. Chang, Z. Chen, A novel fractional order model
based state-of-charge estimation method for lithium-ion battery, Appl. Energy. 207
(2017) 384–393. doi:10.1016/j.apenergy.2017.07.003.
[47] R. Ahmed, M. El Sayed, I. Arasaratnam, J. Tjong, S. Habibi, Reduced-Order
Electrochemical Model Parameters Identification and State of Charge Estimation
for Healthy and Aged Li-Ion Batteries—Part II: Aged Battery Model and State of
Charge Estimation, IEEE J. Emerg. Sel. Top. Power Electron. 2 (2014) 678–690.
doi:10.1109/JESTPE.2014.2331062.
[48] R. Yang, R. Xiong, H. He, H. Mu, C. Wang, A novel method on estimating the
- 221 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

degradation and state of charge of lithium-ion batteries used for electrical vehicles,
Appl. Energy. 207 (2017) 336–345. doi:10.1016/j.apenergy.2017.05.183.
[49] A. Farmann, D.U.D.U. Sauer, A comprehensive review of on-board State-of-
Available-Power prediction techniques for lithium-ion batteries in electric vehicles,
J. Power Sources. 329 (2016) 123–137. doi:10.1016/j.jpowsour.2016.08.031.
[50] Y. Zheng, M. Ouyang, X. Han, L. Lu, J. Li, Investigating the error sources of the
online state of charge estimation methods for lithium-ion batteries in electric
vehicles, J. Power Sources. 377 (2018) 161–188.
doi:10.1016/j.jpowsour.2017.11.094.
[51] M.A. Hannan, M.S.H. Lipu, A. Hussain, A. Mohamed, A review of lithium-ion
battery state of charge estimation and management system in electric vehicle
applications: Challenges and recommendations, Renew. Sustain. Energy Rev. 78
(2017) 834–854. doi:10.1016/j.rser.2017.05.001.
[52] Y. Bar-Shalom, X.-R. Li, T. Kirubarajan, Estimation with Applications to Tracking
and Navigation, John Wiley & Sons, Inc., New York, 2001.
doi:10.1002/0471221279.
[53] S. Sepasi, R. Ghorbani, B.Y. Liaw, Improved extended Kalman filter for state of
charge estimation of battery pack, J. Power Sources. (2014).
doi:10.1016/j.jpowsour.2013.12.093.
[54] H. Pan, Z. Lü, W. Lin, J. Li, L. Chen, State of charge estimation of lithium-ion
batteries using a grey extended Kalman filter and a novel open-circuit voltage
model, Energy. (2017). doi:10.1016/j.energy.2017.07.099.
[55] S. Tong, J.H. Lacap, J.W. Park, Battery state of charge estimation using a load-
classifying neural network, J. Energy Storage. 7 (2016) 236–243.
doi:10.1016/j.est.2016.07.002.
[56] E. Chemali, P.J. Kollmeyer, M. Preindl, A. Emadi, State-of-charge estimation of
Li-ion batteries using deep neural networks: A machine learning approach, J.
Power Sources. 400 (2018) 242–255. doi:10.1016/j.jpowsour.2018.06.104.
[57] R. Xiong, L. Li, J. Tian, Towards a smarter battery management system: A critical
review on battery state of health monitoring methods, J. Power Sources. 405
(2018) 18–29. doi:10.1016/j.jpowsour.2018.10.019.
[58] J. Li, K. Adewuyi, N. Lotfi, R.G. Landers, J. Park, A single particle model with
chemical/mechanical degradation physics for lithium ion battery State of Health
(SOH) estimation, Appl. Energy. 212 (2018) 1178–1190.
doi:10.1016/j.apenergy.2018.01.011.
[59] C. Pastor-Fernández, K. Uddin, G.H. Chouchelamane, W.D. Widanage, J. Marco,
A Comparison between Electrochemical Impedance Spectroscopy and Incremental
Capacity-Differential Voltage as Li-ion Diagnostic Techniques to Identify and
Quantify the Effects of Degradation Modes within Battery Management Systems,
J. Power Sources. 360 (2017) 301–318. doi:10.1016/j.jpowsour.2017.03.042.
[60] R. Mingant, J. Bernard, V. Sauvant-Moynot, Novel state-of-health diagnostic
method for Li-ion battery in service, Appl. Energy. 183 (2016) 390–398.
doi:10.1016/j.apenergy.2016.08.118.
[61] J. Tian, R. Xiong, Q. Yu, Fractional-Order Model-Based Incremental Capacity
Analysis for Degradation State Recognition of Lithium-Ion Batteries, IEEE Trans.
Ind. Electron. 66 (2019) 1576–1584. doi:10.1109/TIE.2018.2798606.
[62] Y. Li, M. Abdel-Monem, R. Gopalakrishnan, M. Berecibar, E. Nanini-Maury, N.
Omar, P. van den Bossche, J. Van Mierlo, A quick on-line state of health
estimation method for Li-ion battery with incremental capacity curves processed
by Gaussian filter, J. Power Sources. 373 (2018) 40–53.
- 222 -
Ph.D. Thesis - Marvin Messing McMaster University - Mechanical Engineering

doi:10.1016/j.jpowsour.2017.10.092.
[63] J. Yang, B. Xia, W. Huang, Y. Fu, C. Mi, Online state-of-health estimation for
lithium-ion batteries using constant-voltage charging current analysis, Appl.
Energy. 212 (2018) 1589–1600. doi:10.1016/j.apenergy.2018.01.010.
[64] A. Eddahech, O. Briat, J.M. Vinassa, Determination of lithium-ion battery state-of-
health based on constant-voltage charge phase, J. Power Sources. (2014).
doi:10.1016/j.jpowsour.2014.02.020.
[65] X. Hu, J. Jiang, D. Cao, B. Egardt, Battery health prognosis for electric vehicles
using sample entropy and sparse Bayesian predictive modeling, IEEE Trans. Ind.
Electron. 63 (2016) 2645–2656. doi:10.1109/TIE.2015.2461523.
[66] M. Hu, Y. Li, S. Li, C. Fu, D. Qin, Z. Li, Lithium-ion battery modeling and
parameter identification based on fractional theory, Energy. 165 (2018) 153–163.
doi:10.1016/j.energy.2018.09.101.
[67] A.N. Eddine, B. Huard, J.-D. Gabano, T. Poinot, A. Thomas, S. Martemianov,
Time domain diffusion parameters identification of electrochemical impedance
models using fractional order system, IFAC-PapersOnLine. 51 (2018) 377–382.
doi:10.1016/j.ifacol.2018.09.174.
[68] J. Xu, C.C. Mi, B. Cao, J. Cao, A new method to estimate the state of charge of
lithium-ion batteries based on the battery impedance model, J. Power Sources. 233
(2013) 277–284. doi:10.1016/j.jpowsour.2013.01.094.
[69] P. Mauracher, E. Karden, Dynamic modelling of lead/acid batteries using
impedance spectroscopy for parameter identification, J. Power Sources. 67 (1997)
69–84. doi:10.1016/S0378-7753(97)02498-1.
[70] B. Wang, S.E. Li, H. Peng, Z. Liu, Fractional-order modeling and parameter
identification for lithium-ion batteries, J. Power Sources. 293 (2015) 151–161.
doi:10.1016/j.jpowsour.2015.05.059.
[71] P. Kollmeyer, A. Hackl, A. Emadi, Li-ion battery model performance for
automotive drive cycles with current pulse and EIS parameterization, 2017 IEEE
Transp. Electrif. Conf. Expo, ITEC 2017. (2017) 486–492.
doi:10.1109/ITEC.2017.7993319.
[72] S. Russell, P. Norvig, Artificial Neural Networks, in: M. Hirsch, T. Dunkelberger
(Eds.), Artif. Intell. A Mod. Approach, 3rd ed., Pearson Education, Inc., New
Jersey, 2010: pp. 727–744.
[73] M. Nielson, Neural Networks and Deep Learning, Determination Press, 2015.
[74] Zahner Messsysteme, EIS - Electrochemical Impedance Sepctroscopy, (2019).
http://zahner.de/pdf/EIS.pdf (accessed January 17, 2020).
[75] K. Qian, B. Huang, A. Ran, Y.B. He, B. Li, F. Kang, State-of-health (SOH)
evaluation on lithium-ion battery by simulating the voltage relaxation curves,
Electrochim. Acta. 303 (2019) 183–191. doi:10.1016/j.electacta.2019.02.055.
[76] Q. Fang, X. Wei, T. Lu, H. Dai, J. Zhu, A state of health estimation method for
lithium-ion batteries based on voltage relaxation model, Energies. 12 (2019).
doi:10.3390/en12071349.
[77] I. Baghdadi, O. Briat, P. Gyan, J.M. Vinassa, State of health assessment for lithium
batteries based on voltage–time relaxation measure, Electrochim. Acta. 194 (2016)
461–472. doi:10.1016/j.electacta.2016.02.109.
[78] P. Ahmadi, E. Kjeang, Realistic simulation of fuel economy and life cycle metrics
for hydrogen fuel cell vehicles, Int. J. Energy Res. 41 (2017) 714–727.
doi:10.1002/er.3672.

- 223 -