3966 IEEE TRANSACTIONS ON TRANSPORTATION ELECTRIFICATION, VOL. 10, NO.

2, JUNE 2024

Two-Level Battery Health Diagnosis Using

Encoder–Decoder Framework and Gaussian Mixture
Ensemble Learning Based on Relaxation Voltage
Haoxiang Xiang , Yujie Wang , Senior Member, IEEE, Xingchen Zhang , Graduate Student Member, IEEE,
and Zonghai Chen , Senior Member, IEEE

Abstract— Accurate diagnosis of the state-of-health (SOH) of multiple intertwined physical and chemical mechanisms, such
the lithium-ion battery is crucial for its safe and reliable opera- as lithium-ion migration, structural changes in electrode mate-
tion. In this article, a two-level battery health diagnosis model is rials, and electrochemical reactions, among others [2]. These
proposed using relaxation voltage. First, the health features of the
relaxation voltage sequence are extracted using the autoencoder mechanisms add complexity and difficulty to the diagnosis of
based on the encoder–decoder framework, and the Gaussian the SOH of lithium-ion batteries.
mixture model (GMM) is used to cluster distinct aging levels of Currently, two predominant approaches exist for diagnosing
the battery, thereby enabling a preliminary diagnosis of battery the SOH of lithium-ion batteries: model-based and data-driven.
SOH. Then, a novel Gaussian mixture ensemble learning (GMEL) Model-based approaches depend on a priori knowledge and
method is presented that leverages prior knowledge for accu-
rate diagnosis of battery SOH. Furthermore, the performance utilize dynamic models to estimate the SOH of batteries. These
of the ensemble model is enhanced by the sequential model- models include complex electrochemical models (EMs) such
based algorithm configuration (SMAC) algorithm to optimize as P2D [3] and SP [4] models, as well as simpler equivalent
the hyperparameters, resulting in mean-absolute-error (MAE) circuit models (ECMs) such as integer-order [5], [6], [7] and
and root-mean-square-error (RMSE) of 0.736% and 1.013%,
fractional-order ECM [8], [9], [10]. Although EM models can
respectively. In addition, a data reconstruction model is developed
using the encoder–decoder framework to address the challenge provide a detailed description of the electrochemical processes
of obtaining complete relaxation voltage sequences in the real within the battery, their design involves numerous nonlinear
world. Utilizing only 4 min of incomplete relaxation voltage data, partial differential equations and various parameters, mak-
the presented model achieves the MAE of 1.265% and RMSE of ing them impractical for real-world applications. Conversely,
1.681% in diagnosing the battery SOH. Finally, the superiority
ECM is computationally efficient, but it demands a significant
of the proposed method is verified by several sets of experiments.
amount of prior knowledge, which is often difficult to obtain
Index Terms— Battery health diagnosis, data reconstruction in practice. Data-driven approaches for battery SOH evaluation
model, encoder–decoder framework, Gaussian mixture ensemble
learning (GMEL), relaxation voltage.
are gaining popularity due to advancements in artificial intel-
ligence and machine learning. These methods learn models
I. I NTRODUCTION automatically from measurable data such as voltage, current,
and temperature without prior knowledge of battery aging.
L ITHIUM-ION batteries are pivotal energy storage
devices, providing extensive applications in electric vehi-
cles, portable electronic devices, and other domains. However,
The data-driven approach involves two primary steps: feature
extraction and model construction. Feature extraction is to
during their operation, lithium-ion batteries undergo a gradual obtain the characteristic parameters reflecting the degradation
aging process. This ultimately leads to a deterioration in of the battery, while model construction involves fitting a
performance [1]. Without timely diagnosis and maintenance model to the nonlinear relationship between the extracted
of their health status, these issues can worsen, resulting in feature data and the corresponding battery SOH labels.
the failure of lithium-ion batteries to operate correctly and Practical health features must be obtained from available
potentially causing safety hazards. Consequently, diagnosing data, and the charging process of batteries is relatively stable,
the state-of-health (SOH) of lithium-ion batteries is indispens- so charging data are often used to extract feature quantities
able. The aging of lithium-ion batteries is characterized by that portray the SOH of batteries [11], [12]. In [11], the
peak area of the partial incremental capacity curve and the
Manuscript received 17 June 2023; revised 9 August 2023; accepted
17 September 2023. Date of publication 20 September 2023; date of current duration of the constant current (CC) and constant voltage
version 18 June 2024. This work was supported in part by the National Natural (CV) charging periods are regarded as health indicators of the
Science Foundation of China under Grant 62373340 and Grant 61803359 and battery. In [12], the D-NSGA-II algorithm is used to select
in part by the Natural Science Foundation of Anhui Province under Grant
2208085UD12. (Corresponding author: Yujie Wang.) the voltage range in the charging voltage profile so that the
The authors are with the Department of Automation, University of Science optimal range for different environments can be dynamically
and Technology of China, Hefei 230027, China (e-mail: xianghx@mail. captured to estimate the SOH. In addition to the charging
ustc.edu.cn; [email protected]; [email protected];
[email protected]). process, researchers have recently placed greater emphasis
Digital Object Identifier 10.1109/TTE.2023.3317449 on the relaxation process of the battery after the current
2332-7782 © 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: University of Science & Technology of China. Downloaded on July 05,2024 at 11:26:17 UTC from IEEE Xplore. Restrictions apply.
XIANG et al.: TWO-LEVEL BATTERY HEALTH DIAGNOSIS USING ENCODER–DECODER FRAMEWORK 3967

interruption, as the relaxation voltage curve contains rich aging it may be more practical to achieve a fast and approximate
information [13], [14]. In [13], multiple statistical properties estimate of the battery aging level in addition to an accurate
of the relaxation voltage sequences are extracted as battery estimate of the battery SOH. Therefore, models that can adapt
health features. In [14], relaxation voltage sequences with to different application scenarios hold more value.
a wider state of charge range are input to a dual Gaussian To address the above-mentioned challenges, a novel two-
process regression (GPR) to estimate battery SOH and open level battery health diagnosis model is proposed in this article.
circuit voltage (OCV). Although progress has been made The major contributions are as follows.
in extracting battery health features, two key issues remain. 1) Features strongly correlated with battery SOH are auto-
First, the process is often manual, time-consuming, and lacks matically extracted from the relaxation voltage sequence
universality and reliability. Second, with the increasing diver- using the autoencoder based on the encoder–decoder
sity of battery charging methods, it is necessary to develop framework. A Gaussian mixture model (GMM) is then
a relaxation voltage-based approach that is not dependent utilized to cluster the feature data, probabilistically
on specific charging modes. However, obtaining a complete dividing it based on the battery’s aging degree. This
relaxation voltage sequence is difficult in practical applications approach provides a preliminary diagnosis of the bat-
due to its long duration, which poses challenges for feature tery’s health status without SOH labels.
extraction. 2) A novel Gaussian mixture ensemble learning (GMEL)
Data-driven models have proven to be effective in learn- method is proposed. This overcomes the limitations of
ing the nonlinear mapping of extracted features to battery traditional ensemble methods by dividing subdatasets
SOH. Traditional machine learning methods such as extreme according to clustering results and assigning weights
learning machine (ELM) [15], GPR [16], and artificial neural based on posterior probability. The sequential model-
network (ANN) [17] have demonstrated success in evaluat- based algorithm configuration (SMAC) algorithm is
ing battery SOH. Recently, deep-learning methods employing utilized to optimize the hyperparameters of the ensemble
complex neural networks have gained popularity. For instance. model, further enhancing its performance. The two-
Ye and Yu [18] proposed a deep domain adversarial net- level battery health diagnosis model constructed through
work to tackle the issue of distribution discrepancy between the above work provides both preliminary and accurate
different domains caused by diverse working conditions. SOH diagnoses and is adaptable to various application
Wang et al. [19] applied a deep belief network based scenarios.
on particle swarm optimization to estimate battery SOH 3) A data reconstruction model based on the encoder–
while considering various discharge depths. Most data-driven decoder framework is presented to address the challenge
approaches are designed with specific assumptions and appli- of obtaining the complete relaxation voltage sequence
cation constraints, which can result in poor generalization in practice. This enables the diagnosis model to be
and robustness, as the model’s performance becomes heav- applied to short-time relaxation scenarios with greater
ily dependent on the quality of training samples. Instead, practicality.
the ensemble learning that combines multiple weak learners The rest of this work is arranged as follows. Section II
through a specific set of combining strategies can generate introduces the battery aging dataset utilized in this study. Then,
a more robust and accurate estimation model [20]. In [21], Section III presents feature extraction and data reconstruction
a hierarchical ensemble learning structure is proposed to methods based on the encoder–decoder framework. Section IV
reduce the prediction error of a single ELM model, and this outlines the proposed two-level battery diagnosis model and
ensemble approach can significantly improve the accuracy the hyperparameter optimization method. Several experiments
and reliability of estimation. In [22], a convolutional neural are conducted in Section V. Finally, Section IV summarizes
network model that combines transfer learning and ensemble the whole article.
learning concepts for battery capacity estimation is employed.
While the battery SOH diagnosis method based on ensemble
learning has achieved more stable and accurate results than II. BATTERY AGING DATASET
before, there are still pressing issues to address. First, the The aging cycle data of 38 commercial 18650 NCA lithium-
training dataset used by base learners often suffers from the ion batteries sourced from the publicly available dataset [13]
problem of idiosyncratic similarity. This can result in a lack are used in this study. The dataset provides complete infor-
of diversity among the set of base learners, which may lead mation on the voltage and current in the battery cycle aging
to a significant bias in the prediction results of the ensemble experiments. These batteries have a nominal capacity of 3.5 Ah
model [20]. Second, accurately quantifying the contribution and a cutoff voltage range of 2.65–4.2 V. They are categorized
of each base learner to the final prediction results has proven based on the temperature and charging current C rate of the
challenging, and using crude weight assignments may not aging experiments, as shown in Table I. Their discharging
accurately reflect the differences in their contributions. This current C rates are all 1C.
highlights the need for a new method that can construct a The batteries undergo CC charging until the voltage reaches
diverse set of base learners and accurately determine their con- the upper threshold. Then, the charging switches to CV. The
tributions. Moreover, most data-driven battery SOH diagnosis batteries’ current is disconnected when it drops to 0.05C, and
methods require a complex pipeline process to obtain specific a 30-min relaxation process follows. The batteries are then
estimation results [23]. In practical applications, sometimes discharged with a CC of 1C until the voltage drops to the

Authorized licensed use limited to: University of Science & Technology of China. Downloaded on July 05,2024 at 11:26:17 UTC from IEEE Xplore. Restrictions apply.
3968 IEEE TRANSACTIONS ON TRANSPORTATION ELECTRIFICATION, VOL. 10, NO. 2, JUNE 2024

TABLE I
BATTERY DATASET

Fig. 2. Relaxtion voltage varies with batteries aging (cX denotes the
X th battery in the corresponding cycle condition).

Fig. 1. Discharging capacity varies with batteries aging.

lower cut-off voltage, followed by another 30-min relaxation

process. The above operations are iteratively repeated with a
relaxation process sampling time of 120 s.
In this work, SOH is defined using the mainstream capacity-
based definition, which is the ratio of the current maximum
available capacity to the nominal capacity of the cell, with the
Fig. 3. Structure of encoder–decoder framework.
following expression:
Ccurrent
SOH = × 100%. (1) inconsistencies to build a battery SOH diagnostic model with
Cnominal
strong generalization and high accuracy.
The SOH degradation curves of all NCA batteries are
shown in Fig. 1. It is observed that the NCA batteries exhibit III. F EATURE E XTRACTION AND DATA R ECONSTRUCTION
better capacity retention under slightly elevated temperatures.
In this section, an encoder–decoder framework is initially
Furthermore, distinct aging trajectories are observed among
introduced. Subsequently, the investigation on health feature
batteries under identical aging conditions, primarily attributed
extraction and data reconstruction is conducted based on this
to the inevitable variations in battery manufacturing processes
framework, aiming to achieve the extraction of representative
and raw materials. In summary, battery aging is influenced
features that indicate the SOH of batteries from short-term
by multiple factors, and the chemical reactions and physical
relaxation voltage sequences in real-world scenarios.
changes within the battery are highly intricate. Therefore,
it is necessary to develop a battery SOH diagnostic model
applicable to different aging conditions. A. Encoder–Decoder Framework
The relaxation voltage is the voltage when the battery is The encoder–decoder is a neural network framework desig-
disconnected from the current during the relaxation process ned to handle sequence-to-sequence conversion tasks [24].
after a full charge. This voltage is independent of a fixed In this framework, the encoder transforms the input sequence
charging pattern and constitutes an ideal electrical signal into a fixed-length vector while the decoder utilizes the vector
for battery health assessment. In this article, the relaxation to create the output sequence. Fig. 3 shows the encoder–
voltage sequence of the first 14 sampling points is studied, decoder framework employing fully connected layers.
corresponding to a 28-min relaxation process. Two batteries Assuming the input sequences X = {x1 , x2 , . . . , xk }, where
are randomly selected under different cycle conditions, and xi ∈ R n , the encoder transforms it into an encoded sequences
their relaxation voltages are plotted against aging cycles, Z = {z 1 , z 2 , . . . , z k }, where z i ∈ R l . The decoder, in turn,
as shown in Fig. 2. The results demonstrate a clear correlation decodes the encoded sequences to new sequences X ′ =
between relaxation voltage and battery aging, as evidenced {x1′ , x2′ , . . . , xk′ }. The data flow of the encoding and decoding
by the faster initial voltage drop with increasing battery process follows the equation:
age and the slower rate at which the voltage returns to
z i = f (wz · xi + bz ) (2)
OCV equilibrium. In Sections IV and V, we take the NCA
battery dataset with different aging conditions and monomer xi′ = g(wx ′ · z i + bx ′ ) (3)

Authorized licensed use limited to: University of Science & Technology of China. Downloaded on July 05,2024 at 11:26:17 UTC from IEEE Xplore. Restrictions apply.
XIANG et al.: TWO-LEVEL BATTERY HEALTH DIAGNOSIS USING ENCODER–DECODER FRAMEWORK 3969

Fig. 4. Results of autoencoder with two features. (a) Loss curve. (b) Decoded Fig. 5. Results of encoder–decoder data reconstruction model. (a) Loss curve.
curves. (b) Decoded curves.

where f (·) and g(·) represent the activation functions, while C. Relaxation Voltage Sequence Reconstruction
wz , wx ′ , bz , and bx ′ denote the weights and biases, respec- In practical scenarios, attaining the complete relaxation
tively. Throughout the training process, the encoder–decoder voltage sequence is challenging due to the prolonged duration
network endeavors to minimize the reconstruction error, which required. Moreover, a shorter relaxation voltage sequence
measures the deviation between the true output sequences and would render the previous feature extraction model based on
the predicted output sequences as the autoencoder invalid. However, this issue can be resolved by
n employing data reconstruction based on the encoder–decoder

1X 2
L X r′ , X ′ = x ′ − xi′ . (4) framework. Specifically, the reconstruction of the complete
2 i=1 r,i
sequence by a shorter relaxation voltage sequence is consistent
with the architecture shown in Fig. 3, where the input and
B. Feature Extraction and Correlation Analysis output are both sequences. The neural network uses consists of
The encoder–decoder network functions as an autoencoder an encoder and decoder, each comprising two fully connected
when every pair of input and output sequences are identical layers. The data reconstruction model utilized for practical
(xi = xi′ ) [25]. An important characteristic of the autoencoder applications is obtained through the training of the neural
is that its hidden layer dimension is usually smaller than the network. Mathematically, the input sequences are denoted
dimension of the input data. This forces the autoencoder to as X = {x1 , x2 , . . . , xk }, where xi ∈ R n , and the output
learn to a compressed representation of the input data, thus sequences are represented as X ′ = {x1′ , x2′ , . . . , xk′ }, where
allowing the network to capture the main features of the input xi′ ∈ R m . It is important to note that in data reconstruction,
data. This constraint can help the autocoder learn more useful the dimensionality of the input sequence (n) is smaller than
representations for efficient data dimensionality reduction. The that of the output sequence (m).
model’s encoded low-dimensional sequence z i captures the The sequence data of the first two sampling points (4 min
principal characteristics of the original data, allowing for effec- of relaxation) are used as input, and the complete relaxation
tive dimensionality reduction, while the decoder restores the voltage sequences are used as output to verify the validity
encoding to the original signal. Upon completing autoencoder of the reconstructed model. We assign 75% of the relaxation
training, the encoder can reduce computation and storage costs voltage sequence data to the training set and the remaining
while being less sensitive to noise than the original high- data to the test set. The loss curve of the training set is shown
dimensional data. in Fig. 5(a), where the loss function decreases rapidly and
In this study, a fully connected neural network as shown in converges smoothly, indicating a highly effective training pro-
Fig. 3, is employed for the analysis. As an example, the dimen- cess. The performance of the model is evaluated by analyzing
sionality of the features is reduced to two. The loss curve and the reconstructed voltage curves shown in Fig. 5(b) on the
decoder reconstructed voltage are shown in Fig. 4. The loss test set. It is noteworthy that the network achieves impressive
of the neural network decreased smoothly without significant results by accurately reconstructing the complete voltage curve
oscillations or fluctuations, indicating stable network training. using shorter relaxation voltage sequences in the early stages,
The decoded voltage curves accurately followed the original approximating the real original curve closely. These results
curves, demonstrating the autoencoder’s ability to effectively provide robust evidence to support the effectiveness of the
compress and reconstruct critical features of the original input. proposed data reconstruction model.
The Spearman correlation coefficient is employed to ana-
lyze the relationship between the extracted features and IV. BATTERY H EALTH D IAGNOSIS M ODEL D EVELOPMENT
SOH [26]. The final correlation coefficients of features 1 and 2 In this section, a two-level battery health diagnosis model
are calculated as 0.8544 and −0.9061, respectively. The result based on the extracted health features is proposed, as shown
reveals a strong positive correlation between feature 1 and in Fig. 6. The preliminary diagnosis of battery health is
SOH, while feature 2 exhibited a strong negative correlation performed using the GMM clustering algorithm [27], followed
with SOH. These findings suggest that the autoencoder can by the GMEL method for accurate diagnosis. The SMAC
automatically extract features that are closely related to the algorithm is utilized to optimize the hyperparameters of the
battery’s health status. model [28].

Authorized licensed use limited to: University of Science & Technology of China. Downloaded on July 05,2024 at 11:26:17 UTC from IEEE Xplore. Restrictions apply.
3970 IEEE TRANSACTIONS ON TRANSPORTATION ELECTRIFICATION, VOL. 10, NO. 2, JUNE 2024

Fig. 6. Framework of the proposed two-level battery health diagnosis model.

A. Preliminary Diagnosis Based on GMM Clustering

To further investigate the intrinsic relationship between bat-
tery health features and the SOH, this study employs the GMM
clustering method on the feature data. The GMM clustering
algorithm considers the entire dataset as a combination of
multiple Gaussian distributions, i.e.,
K
X
πk N x | µk , 6 k


p(x) = (5)
k=1
where x denotes the feature data, N (x | µk , 6 k ) indicates
the Gaussian distribution density function with µk as the
mean and 6 k as the covariance. K stands for the number of
Gaussian distributions, which is also the number of clusters.
πk represents
P K the weight of the kth Gaussian distribution
satisfying k=1 πk = 1, 0 ≤ πk ≤ 1.
After establishing a GMM that characterizes the data dis-
tribution, the posterior probability distribution of samples
Fig. 7. Results of GMM clustering. (a) Selection of cluster number based on
belonging to distinct clusters can be computed by applying BIC. (b) Two-dimensional visualization of clustering. (c) Three-dimensional
Bayes’ theorem as visualization of clustering. (d) Match between clustering result and battery
SOH.
πk N x | µ k , 6 k


p(k) p(x | k)
γk (x) ≡ p(k | x) = P =P
l πl N x | µl , 6 l


l p(l) p(x | l)
(6)
the number of data samples. BIC has two terms: one measures
where γk (x) is the probability distribution of belonging to model fit to data and the other penalizes model complexity.
cluster k under a given feature sample x. By utilizing this By computing the BIC values under various numbers of
distribution, the output can more accurately reflect the degree clusters, we can determine the optimal number of clusters for
of belonging of each sample to different clusters. It is reason- the GMM. As shown in Fig. 7(a), the BIC demonstrates a
able to assume that the samples belong to the clusters with significant decrease as the number of clusters increases from
the highest probability. one to three, after which the rate of decline slows down
Determining the optimal number of clusters, denoted by and gradually converges. This observation suggests that the
K , in GMM generally requires manual selection, which can balance between the degree of fit and model complexity is
be a challenging task. However, the Bayesian information achieved when the number of clusters is three. Moreover, with
criterion (BIC) has proven to be a valuable tool to aid in three clusters, the battery SOH can be interpreted as three
this process [29]. The BIC is a model selection criterion that different decay degrees, which more closely reflects the actual
balances the degree of model fit with its complexity and is condition of the battery. Thus, we select three clusters after
given by considering model fit, complexity, and interpretability. A 2-D
visualization of the clustering results is shown in Fig. 7(b),
BIC = −2 · log(L) + k · log(n) (7)
where the sample points of different clusters are color-coded,
where L denotes the likelihood function of the data, k repre- and the purple asterisk represents the centroid of each cluster.
sents the number of parameters in the model, and n represents The distance between the centers of the three clusters is large,

Authorized licensed use limited to: University of Science & Technology of China. Downloaded on July 05,2024 at 11:26:17 UTC from IEEE Xplore. Restrictions apply.
XIANG et al.: TWO-LEVEL BATTERY HEALTH DIAGNOSIS USING ENCODER–DECODER FRAMEWORK 3971

indicating that the characteristics of the different clusters are posterior probability of sample points belonging to each clus-
more distinct and have a higher degree of differentiation. ter, representing the confidence level of the aging degree of
Fig. 7(c) shows a 3-D visualization of the clusters. All three that sample point. Consequently, the prediction results of each
clusters are represented by a 2-D Gaussian distribution whose base learner are weighted and integrated, which can more
mean is the cluster center. The staggered nature of the three accurately reflect the contribution of each base learner to the
distributions indicates that the three clusters are well-defined. final prediction results and improve the prediction accuracy
The clustering results plotted alongside the changes in of the ensemble model. Furthermore, the GMEL is a versatile
battery SOH are shown in Fig. 7(d), revealing that the GMM framework that enhances the generalization ability of ensemble
clustering algorithm can effectively capture the trend of battery models by augmenting the dissimilarity among base learners,
SOH changes. The three clusters identified in the figure thereby not being confined to specific base learners and having
roughly correspond to severe, moderate, and slight battery a broad range of applications. In this study, the multilayer
aging, indicating that the GMM algorithm can provide a perceptron (MLP) is selected as the base learner due to its
preliminary diagnosis of the battery health condition. More- robust nonlinear modeling capability.
over, the GMM clustering method can compute the posterior
probability of each data point belonging to each cluster C. Model Hyperparameter Optimization Based on
using (6), in addition to assigning each data point to a cluster. SMAC Algorithm
The probabilistic clustering approach is shown in Fig. 7(b),
Hyperparameters impact the model’s performance in
which shows the probability distribution plots of three data
machine learning and their optimal combination is crucial.
points. The height of each bar represents the probability
Bayesian optimization-based methods use the GPR model
that the data point belongs to its corresponding cluster. This
to estimate the objective function’s probability distribution
approach quantifies the similarity between data points and
and maximize improvement. However, discrete variables pose
each cluster, making it particularly useful when data point
a challenge as they are transformed into high-dimensional
boundaries are blurred or overlap exists. To summarize, the
continuous spaces, reducing the algorithm’s efficiency [28].
preliminary diagnosis model constructed by GMM clustering
The SMAC algorithm uses a random forest model to estimate
is well-suited for numerous applications that do not necessitate
the objective function’s probability distribution, overcom-
precise estimation of battery SOH. Furthermore, the model
ing the limitation of Gaussian process solutions for discrete
can provide valuable reference and guidance for subsequent
variables [28]. Random forest is an ensemble learning method
accurate estimation of battery SOH.
consisting of independently trained decision trees. The final
B. Accurate Diagnosis Based on GMEL regression is obtained by averaging the predictions of these
trees [30]. Random forest’s high diversity improves explo-
We propose a GMEL method for achieving a more precise ration in high-dimensional spaces, while subtree results reduce
diagnosis of battery SOH based on the preliminary diagnosis computational costs and improve efficiency. The pseudo-code
model presented earlier. The algorithm framework is shown in of the SMAC algorithm is shown in Algorithm 1. Specifically,
Fig. 6. Initially, GMM clustering is employed to partition the the algorithm constructs a random forest model based on the
training set into subdatasets corresponding to distinct aging initial sample points, computes the mean and variance of the
patterns. Then, local base learners for each subdatasets are proxy model output based on the results of each subtree, and
constructed. Finally, in the test set, we input the samples to uses expected improvement (EI) as its acquisition function to
each base learner to obtain prediction results and to the GMM generate new sampling points. The algorithm then refines the
to derive the posterior probability distribution of each cluster. original random forest model with the new information and
The results of each base learner are then weighted with the iterates until the stopping condition is met.
posterior probability distribution to generate the final battery
SOH diagnosis result. GMEL effectively leverages local prior
D. Workflow of the Two-Level Battery Health
knowledge to enhance the accuracy and robustness of battery
Diagnosis Model
health diagnosis.
Bagging is a well-known ensemble learning method [20], The workflow of the two-level battery health diagnosis
but it has two flaws. First, bootstrap sampling is a form model can be seen in Fig. 6. Specifically, when the model is
of random sampling with put-back that results in a lack of successfully trained on the training set, the operational process
significant differentiation among the base learners. Second, in real scenarios is as follows.
the Bagging method simply averages the results of multiple 1) After the battery is fully charged and left at rest, if the
base learners, which leads to the inability to fully utilize resting time is short, the data reconstruction model can
the diversity among the base learners and impairs the gen- reconstruct the short relaxation voltage sequence into a
eralization ability of the model. The GMEL proposed in complete sequence.
this study effectively addresses the limitations of Bagging. 2) The complete relaxation voltage sequence is fed into the
First, we employ GMM clustering to partition the dataset autoencoder for health features extraction.
into subdatasets with distinct aging patterns to train each 3) The extracted health features are input into the GMM to
base learner, ensuring diversity among the base learners and obtain the probability distribution of the battery’s current
significantly improving the generalization performance of the health status as severe, moderate, or slight. This is the
ensemble model. Second, GMM clustering can generate the first-level battery health diagnosis.

Authorized licensed use limited to: University of Science & Technology of China. Downloaded on July 05,2024 at 11:26:17 UTC from IEEE Xplore. Restrictions apply.
3972 IEEE TRANSACTIONS ON TRANSPORTATION ELECTRIFICATION, VOL. 10, NO. 2, JUNE 2024

Algorithm 1 SMAC Algorithm TABLE II

B EST H YPERPARAMETER OF MLP A FTER SAMC O PTIMIZATION
1 Randomly initialize n 0 combinations of
hyperparameters and get n 0 points
{(x1 , f (x1 )), (x2 , f (x2 )), . . . , (xn 0 , f (xn 0 ))} (xi is a
combination of hyperparameters, f is the model to be
optimized);
2 Initialize the maximum number of iterations N ;
3 Train the random forest model with n 0 points;
4 for n from 1 to N do
5 Generate candidate results by large-scale random
sampling, import the candidate points into the
random forest to get the mean and standard and root-mean-square-error (RMSE), which are defined as
deviation, and calculate the acquisition function N
1 X
an (x); MAE = [i
SOHi − SOH (8)
6 Choose the point that maximizes an (x) as xn+1 ; N i=1
v
7 The new point xn+1 is brought in model to obtain u N
u1 X 2
f (xn+1 ); RMSE = t [i
SOHi − SOH (9)
8 Update n = n + 1, and retrain the random forest N i=1
with the current n points;
9 end where SOHi is the real SOH and SOH [i is the estimated SOH.
10 return The point x ∗ that maximizes f (x). The performance of the different models for SOH estimation
on the test set is shown in Fig. 8. The true SOH is represented
by the blue line, while the estimated SOH is shown by scatter.
The density of the scatter plot is expressed by different colors,
4) Based on the results of the first-level battery health with red indicating a lower density and blue indicating a higher
diagnosis, the extracted health features and probabil- density. Therefore, the more concentrated the scatters on the
ity distribution of the current health status of the blue line (i.e., the density of scatters on the blue line is higher
battery are input into the ensemble model based on than the density of scatters that deviate from the blue line),
the GMEL architecture, which is optimized by the the greater the accuracy of the SOH estimation is indicated.
SMAC algorithm, to obtain the accurate SOH of the Table III provides the estimated metrics for the quantification
battery. This constitutes the second-level battery health of each model. From Fig. 8(a)–(d), it is evident that the GMEL
diagnosis. model achieves the highest accuracy of SOH estimation, with
its estimates closely concentrated on the true SOH blue line.
The GMEL model exhibits an MAE of 0.800% and RMSE
V. E XPERIMENT R ESULTS AND D ISCUSSION
of 1.107%, as shown in Table III. In contrast, the models
This section aims to validate the effectiveness of the pro- with poor estimation performance are averaging, bagging,
posed battery health diagnosis model by comparing it with and the MLP model in that order. The MLP model shows
other methods. Specifically, we compare the diagnostic per- the worst estimation performance, with its estimated values
formance of various models, both before and after SMAC differing greatly from the true SOH, indicating that it does
optimization, and assess different feature extraction methods not perform as well as the ensemble models. In conclusion,
and hyperparameter optimization algorithms to highlight the the ensemble method plays a pivotal role in the performance
superiority of the proposed approach. Finally, we demonstrate of ensemble models, and the proposed GMEL model achieves
the feasibility of the relaxation voltage sequence reconstruction superior performance compared to other ensemble methods by
method. enhancing diversity.
Fig. 8(e)–(h) shows a significant enhancement in the per-
formance of all models following the optimization of the
A. Comparison of Diagnosis Models base learner’s hyperparameters using SMAC algorithm, lead-
Several battery SOH diagnosis models are compared, ing to highly accurate estimation of SOH. This observation
including single MLP, averaging ensemble, bagging ensemble, demonstrates the criticality of hyperparameter optimization
and GMEL. The base learners of the ensemble models are all in improving model performance, as it has the potential to
single hidden layer MLPs, and the ensemble is composed of enhance the generalization performance of the ensemble model
three base learners. NCA battery data are randomly divided by optimizing the performance of the base learner. Notably,
into a 75% training set and a 25% test set. Hyperparameter even after optimization, the GMEL model proposed in this
optimization is conducted using a fivefold cross-validation work remains the best-performing model, achieving MAE and
approach on the training set with the SAMC method. The RMSE of 0.736% and 1.013%, respectively. These results pro-
resulting hyperparameter optimization results are shown in vide robust evidence for the efficacy of the SMAC algorithm
Table II. The model’s performance on the test set is evaluated in optimizing the model. Optimizing the hyperparameters of
using two metrics, namely, the mean-absolute-error (MAE) the base learner can significantly improve the generalization

Authorized licensed use limited to: University of Science & Technology of China. Downloaded on July 05,2024 at 11:26:17 UTC from IEEE Xplore. Restrictions apply.
XIANG et al.: TWO-LEVEL BATTERY HEALTH DIAGNOSIS USING ENCODER–DECODER FRAMEWORK 3973

Fig. 8. Test results of battery SOH diagnosis models. (a) MLP. (b) Averaging. (c) Bagging. (d) GMEL. (e) SMAC-MLP. (f) SMAC-Averaging.
(g) SMAC-Bagging. (h) SMAC-GMEL.

TABLE III
T EST R ESULTS OF BATTERY SOH D IAGNOSIS M ODELS

performance of the ensemble model, resulting in superior

predictive outcomes.

B. Comparison of Feature Extraction Algorithms

To assess the feature extraction ability of the autoencoder,
a comparison with the commonly used dimensionality reduc-
tion algorithm, PCA, is conducted. Specifically, the features in
the GMEL model and the SMAC-GMEL model are replaced
with those obtained by PCA dimensionality reduction, and the
estimation errors on the test set are compared. The results
shown in Fig. 9 indicate that both models based on the
autoencoder exhibit lower estimation errors on the test set
than the PCA-based model. The findings demonstrate that the
autoencoder is capable of extracting features that describe the
SOH of batteries with greater accuracy when compared to Fig. 9. Test results of battery SOH diagnosis models based on different
feature extraction algorithms.
PCA. This superiority may be attributed to the fact that PCA is
only able to learn the linear relationships within the original
data, while autoencoder, utilizing neural network structures,
is able to learn nonlinear relationships. Therefore, autoencoder hyperparameters of the base learner using fivefold cross-
is a more powerful nonlinear dimensionality reduction method validation in the training set, and their optimization times are
than PCA. compared. Subsequently, the estimation performance of the
GMEL ensemble model for SOH based on different optimiza-
tion algorithms is compared in the testing set. The results are
C. Comparison of Hyperparameter Optimization Algorithms shown in Table IV. This indicates that SMAC outperforms
Different hyperparameter optimization algorithms, includ- the other algorithms in terms of time efficiency and accuracy,
ing grid search, random search, and Bayesian optimization, achieving an MAE of 0.736% and RMSE of 1.013% in only
are compared with the SMAC algorithm used in this study. 82 s. These results are comparable to those of the grid search.
All hyperparameter optimization algorithms optimize the However, the grid search’s computation time is prohibitively

Authorized licensed use limited to: University of Science & Technology of China. Downloaded on July 05,2024 at 11:26:17 UTC from IEEE Xplore. Restrictions apply.
3974 IEEE TRANSACTIONS ON TRANSPORTATION ELECTRIFICATION, VOL. 10, NO. 2, JUNE 2024

TABLE IV diagnosis and supervised accurate diagnosis to meet various

P ERFORMANCE C OMPARISON OF H YPERPARAMETER application requirements. The main results of this article can
O PTIMIZATION A LGORITHMS IN T EST S ET
be ordered as follows.
1) The autoencoder is utilized to extract health features
from the relaxation voltage sequence. Subsequently,
GMM clustering is applied to automatically partition
the data samples into distinct clusters based on the
health features, which correspond to different degrees of
aging. This model achieves an unsupervised preliminary
diagnosis of battery health without SOH labels.
2) For accurate diagnosis of SOH, a novel ensemble learn-
ing method named GMEL is presented. The ensemble
model constructed by GMEL has the lowest MAE and
RMSE of 0.800% and 1.107%, respectively, in the
test set compared to other commonly used models.
In addition, the SMAC algorithm is used to optimize the
model, resulting in lower MAE and RMSE of 0.736%
and 1.013%, respectively.
3) A data reconstruction model based on the encoder–
decoder framework is raised that can reconstruct 28 min
Fig. 10. Test results of battery SOH diagnosis models based on GMEL model of complete data based on four minutes of short-time
using reconstructed voltage sequence data. (a) GMEL. (b) SMAC-GMEL.
relaxation voltage data. In this scenario, the SMAC-
GMEL model attains accurate SOH diagnosis with MAE
long in practice, taking 4304 s to complete the search. In con- and RMSE of 1.265% and 1.681%, respectively. The
trast, Bayesian optimization is relatively inefficient in this aforementioned results demonstrate the practicality and
comparison, requiring 133 s to complete the search and not effectiveness of the proposed diagnostic model.
performing as well as the random search, which has a slightly
shorter search time. This disparity may stem from Bayesian R EFERENCES
optimization’s complexity and time consumption for discrete [1] H. Rauf, M. Khalid, and N. Arshad, “Machine learning in state of health
and remaining useful life estimation: Theoretical and technological
parameter variables, which do not fully exploit the algorithm’s development in battery degradation modelling,” Renew. Sustain. Energy
potential within a shorter time frame. The above discussion Rev., vol. 156, Mar. 2022, Art. no. 111903.
indicates SMAC as a highly efficient hyperparameter opti- [2] R. Xiong, Y. Pan, W. Shen, H. Li, and F. Sun, “Lithium-ion battery aging
mechanisms and diagnosis method for automotive applications: Recent
mization algorithm that can identify the optimal solution in a advances and perspectives,” Renew. Sustain. Energy Rev., vol. 131,
shorter time compared to other algorithms, making it suitable Oct. 2020, Art. no. 110048.
for diverse machine learning applications. [3] D. Wang, Q. Zhang, H. Huang, B. Yang, H. Dong, and J. Zhang, “An
electrochemical–thermal model of lithium-ion battery and state of health
estimation,” J. Energy Storage, vol. 47, Mar. 2022, Art. no. 103528.
D. Validation of Voltage Sequence Reconstruction [4] D. Fang et al., “Performance simulation method and state of health esti-
mation for lithium-ion batteries based on aging-effect coupling model,”
Validation of the encoder–decoder model’s accuracy in Green Energy Intell. Transp., vol. 2, no. 3, Jun. 2023, Art. no. 100082.
reconstructing the complete relaxation voltage sequence is [5] T. Ouyang, P. Xu, J. Chen, J. Lu, and N. Chen, “An online prediction
shown in Fig. 5. Subsequently, the reconstructed voltage of capacity and remaining useful life of lithium-ion batteries based on
simultaneous input and state estimation algorithm,” IEEE Trans. Power
sequence undergoes feature extraction through the autoen- Electron., vol. 36, no. 7, pp. 8102–8113, Jul. 2021.
coder, and the proposed GMEL model is utilized to diagnose [6] T. Ouyang, J. Ye, P. Xu, C. Wang, and E. Xu, “Estimation of state-
the battery SOH. As shown in Fig. 10, the GMEL estimation of-charge and state-of-health for lithium-ion battery based on improved
firefly optimized particle filter,” J. Energy Storage, vol. 68, Sep. 2023,
performance is satisfactory, and then the SMAC optimiza- Art. no. 107733.
tion further improves the estimation accuracy. After SMAC [7] J. Tian et al., “Lithium-ion battery health estimation with real-world data
optimization, the MAE and RMSE of the GMEL model are for electric vehicles,” Energy, vol. 270, May 2023, Art. no. 126855.
[8] Y. Xu, B. Hu, T. Wu, and T. Xiao, “Joint estimation of state of charge and
only 1.265% and 1.681%, respectively, reflecting very low state of health of lithium-ion battery based on fractional order model,”
estimation errors. In summary, the proposed diagnosis model, J. Power Electron., vol. 22, pp. 318–330, Dec. 2022.
in conjunction with a brief relaxation voltage sequence (only [9] P. Xu, X. Hu, B. Liu, T. Ouyang, and N. Chen, “Hierarchical esti-
mation model of state-of-charge and state-of-health for power batteries
4 min), facilitates precise battery SOH diagnosis, rendering considering current rate,” IEEE Trans. Ind. Informat., vol. 18, no. 9,
the method practically applicable. pp. 6150–6159, Sep. 2022.
[10] T. Ouyang, P. Xu, J. Lu, X. Hu, B. Liu, and N. Chen, “Coestimation
VI. C ONCLUSION of state-of-charge and state-of-health for power batteries based on
multithread dynamic optimization method,” IEEE Trans. Ind. Electron.,
Accurate diagnosis of the SOH of lithium-ion batteries is vol. 69, no. 2, pp. 1157–1166, Feb. 2022.
crucial for their normal operation. In this work, a two-level [11] J. Wang et al., “Multiple indicators-based health diagnostics and prog-
nostics for energy storage technologies using fuzzy comprehensive
battery health diagnosis model based on relaxation voltage is evaluation and improved multivariate grey model,” IEEE Trans. Power
proposed. The model can perform unsupervised preliminary Electron., vol. 36, no. 11, pp. 12309–12320, Nov. 2021.

Authorized licensed use limited to: University of Science & Technology of China. Downloaded on July 05,2024 at 11:26:17 UTC from IEEE Xplore. Restrictions apply.
XIANG et al.: TWO-LEVEL BATTERY HEALTH DIAGNOSIS USING ENCODER–DECODER FRAMEWORK 3975

[12] J. Hu, L. Weng, Z. Gao, and B. Yang, “State of health estimation and Haoxiang Xiang was born in Guangdong, China,
remaining useful life prediction of electric vehicles based on real-world in February 2000. He received the B.S. degree
driving and charging data,” IEEE Trans. Veh. Technol., vol. 72, no. 1, in automation from the Guangdong University of
pp. 382–394, Jan. 2023. Technology, Guangdong, in 2022. He is currently
[13] J. Zhu et al., “Data-driven capacity estimation of commercial lithium- pursuing the master’s degree in control engineering
ion batteries from voltage relaxation,” Nature Commun., vol. 13, no. 1, with the University of Science and Technology of
p. 2261, Apr. 2022. China, Hefei, China.
[14] B. Jiang, Y. Zhu, J. Zhu, X. Wei, and H. Dai, “An adaptive capacity His research interests include state estimation,
estimation approach for lithium-ion battery using 10-min relaxation control strategies, and health management of energy
voltage within high state of charge range,” Energy, vol. 263, Jan. 2023, storage systems.
Art. no. 125802.
[15] L. Chen et al., “Online estimating state of health of lithium-ion batteries
using hierarchical extreme learning machine,” IEEE Trans. Transport.
Electrific., vol. 8, no. 1, pp. 965–975, Mar. 2022.
[16] Z. Wang, C. Yuan, and X. Li, “Lithium battery state-of-health estimation
via differential thermal voltammetry with Gaussian process regression,”
IEEE Trans. Transport. Electrific., vol. 7, no. 1, pp. 16–25, Mar. 2021. Yujie Wang (Senior Member, IEEE) received the
[17] Z. Wei, H. Ruan, Y. Li, J. Li, C. Zhang, and H. He, “Multistage Ph.D. degree in control science and engineering from
state of health estimation of lithium-ion battery with high tolerance to the University of Science and Technology of China,
heavily partial charging,” IEEE Trans. Power Electron., vol. 37, no. 6, Hefei, China, in 2017.
pp. 7432–7442, Jun. 2022. He has been with the School of Information Sci-
[18] Z. Ye and J. Yu, “State-of-Health estimation for lithium-ion batteries ence and Technology, University of Science and
using domain adversarial transfer learning,” IEEE Trans. Power Elec- Technology of China, since 2017, where he is cur-
tron., vol. 37, no. 3, pp. 3528–3543, Mar. 2022. rently an Associate Professor at the Department of
Automation. His research interests include model-
[19] Y. Wang, K. Li, P. Peng, and Z. Chen, “Health diagnosis for lithium-ion
ing, simulation, and control of renewable energy
battery by combining partial incremental capacity and deep belief net-
systems and electric and fuel cell vehicles.
work during insufficient discharge profile,” IEEE Trans. Ind. Electron.,
Dr. Wang is currently the Secretary of the System Simulation Committee,
vol. 70, no. 11, pp. 11242–11250, Nov. 2023.
Chinese Association of Automation, a member of the Simulation Technology
[20] Z.-H. Zhou, Ensemble Methods: Foundations and Algorithms. Boca
and Applications Committee, China Simulation Federation, and the Director
Raton, FL, USA: CRC Press, 2012.
of the IEEE PES Electric Vehicle Technical Committee.
[21] W. Liu, Y. Xu, and X. Feng, “A hierarchical and flexible data-driven
method for online state-of-health estimation of Li-ion battery,” IEEE
Trans. Veh. Technol., vol. 69, no. 12, pp. 14739–14748, Dec. 2020.
[22] S. Shen, M. Sadoughi, M. Li, Z. Wang, and C. Hu, “Deep convolu-
tional neural networks with ensemble learning and transfer learning for
capacity estimation of lithium-ion batteries,” Appl. Energy, vol. 260, Xingchen Zhang (Graduate Student Member, IEEE)
Feb. 2020, Art. no. 114296. was born in Jiangsu, China, in October 1998.
[23] D. Roman, S. Saxena, V. Robu, M. Pecht, and D. Flynn, “Machine He received the B.S. degree in electrical engineering
learning pipeline for battery state-of-health estimation,” Nature Mach. from Soochow University, Soochow, China, in 2017.
Intell., vol. 3, no. 5, pp. 447–456, Apr. 2021. He is currently pursuing the master’s degree in
[24] C. Bian, H. He, S. Yang, and T. Huang, “State-of-charge sequence control engineering with the University of Science
estimation of lithium-ion battery based on bidirectional long short- and Technology of China, Hefei, China.
term memory encoder–decoder architecture,” J. Power Sources, vol. 449, His research interests include modeling and con-
Feb. 2020, Art. no. 227558. trol for complex systems and fault diagnosis with
[25] Y. Jiang, Y. Chen, F. Yang, and W. Peng, “State of health estimation automotive applications.
of lithium-ion battery with automatic feature extraction and self-
attention learning mechanism,” J. Power Sources, vol. 556, Feb. 2023,
Art. no. 232466.
[26] X. Tan, D. Zhan, P. Lyu, J. Rao, and Y. Fan, “Online state-of-health esti-
mation of lithium-ion battery based on dynamic parameter identification
at multi timescale and support vector regression,” J. Power Sources,
vol. 484, Feb. 2021, Art. no. 229233. Zonghai Chen (Senior Member, IEEE) was born
[27] J. Tian et al., “Feature fusion-based inconsistency evaluation for battery in Anhui, China, in December 1963. He received
pack: Improved Gaussian mixture model,” IEEE Trans. Intell. Transp. the B.Sc. and M.E. degrees from the University of
Syst., vol. 24, no. 1, pp. 446–458, Jan. 2022. Science and Technology of China (USTC), Hefei,
[28] F. Hutter, H. H. Hoos, and K. Leyton-Brown, “Sequential model- China, in 1988 and 1991, respectively.
based optimization for general algorithm configuration,” in Proc. Learn. He has been a Professor with the Department of
Intell. Optim., 5th Int. Conf. (LION). Rome, Italy: Springer, Jan. 2011, Automation, USTC, since 1998. His main research
pp. 507–523. interests include modeling and control of com-
[29] X. Liu, J. Yang, L. Wang, and J. Wu, “Bayesian information criterion plex systems, energy management technologies for
based data-driven state of charge estimation for lithium-ion battery,” electric vehicles and smart microgrids, intelligent
J. Energy Storage, vol. 55, Nov. 2022, Art. no. 105669. robotics, and information processing.
[30] K. S. R. Mawonou, A. Eddahech, D. Dumur, D. Beauvois, and E. Godoy, Prof. Chen was a recipient of special allowances from the State Council
“State-of-health estimators coupled to a random forest approach for of China. He is a member of the Robotics Technical Committee and Model-
lithium-ion battery aging factor ranking,” J. Power Sources, vol. 484, ing, Identification and Signal Processing Technical Committee, International
Feb. 2021, Art. no. 229154. Federation of Automation Control (IFAC).

Authorized licensed use limited to: University of Science & Technology of China. Downloaded on July 05,2024 at 11:26:17 UTC from IEEE Xplore. Restrictions apply.