ETH Library

Tiny On-Device Structural Health

Monitoring for Wind Turbines
Using MEMS Pressure Sensors

Conference Paper

Author(s):
von Däniken, Elias ; Mikhaylov, Denis ; Moallemi, Amirhossein; Polonelli, Tommaso ; Magno, Michele

Publication date:
2024

Permanent link:
https://doi.org/10.3929/ethz-b-000690066

Rights / license:
In Copyright - Non-Commercial Use Permitted

Originally published in:

https://doi.org/10.1109/SAS60918.2024.10636460

This page was generated automatically upon download from the ETH Zurich Research Collection.
For more information, please consult the Terms of use.
 Tiny On-Device Structural Health Monitoring for
Wind Turbines Using MEMS Pressure Sensors
Elias von Däniken Denis Mikhaylov Amirhossein Moallemi Tommaso Polonelli Michele Magno
D-ITET D-ITET DEI D-ITET D-ITET
ETH Zürich ETH Zürich University of Bologna ETH Zürich ETH Zürich
Zürich, Switzerland Zürich, Switzerland Bologna, Italy Zürich, Switzerland Zürich, Switzerland
0009-0002-6100-8818 0000-0003-2427-8958 0000-0002-7973-5493 0000-0003-0405-3612 0000-0003-0368-8923

Abstract—Wind turbines are extensively used as a renewable fore, the primary approach is to time-schedule data collection
source of energy. These turbines are exposed to turbulence with pre-determined intervals to be transmitted to the cloud
and highly variable weather conditions such as wind, rain, and for storage and processing [5]. While being effective so far,
ice, which can cause significant damage and degraded power
generation performance if not detected in time. The emergence this approach imposes limitations in terms of damage detection
of the Industry 4.0 paradigm enables a reduction in maintenance latency and data traffic volumes. Low-power wireless links are
costs and increased efficiency. Recent research has demonstrated often slow (<1 Mbps), requiring several minutes to forward
the possibility of monitoring the aerodynamic performance of the the data for processing. Moreover, in the case of a wind farm,
wind turbine directly from the blades with today’s state-of-the- the wireless channel can be easily saturated when several tens
art wireless monitoring systems. However, these need external
support to process and analyze the collected data. This paper of devices are transmitting simultaneously [3]. Furthermore,
proposes an effective tiny machine learning approach to monitor current Structural Health Monitoring (SHM) solutions for
the blade structural integrity directly onboard the sensor node, wind turbine blades focus mainly on vibration-based analyses,
requiring only 3 kB of RAM and 35 kB of Flash running on such as modal analysis, assuming a correlation between struc-
a low-power ARM Cortex-M4F at 48MHz. We experimentally tural anomalies and spectral components. However, surface
evaluate the proposed model with a dataset acquired in a wind
tunnel with 40 MEMS barometers placed around the airfoil in 6 damage or corrosion can be invisible for this vibration-based
different damage conditions. The XGBoost classification model is approach despite decreasing energy generation efficiency. For
purely based and trained on the blade aerodynamics, featuring these reasons, alternatives to SoA approaches still need to
a classification accuracy above 83% over 6 labeled classes of be investigated based on other physical sources, such as the
increasing damage. With a total execution time of 579 ms while it airflow behavior around the blade airfoil, to extract additional
only consumes 9 mJ, it enables real-time analysis for a sustainable
battery-based monitoring system. operative health indicators.
Index Terms—SHM, wind turbines, tinyML, edge computing,
Industry 4.0 This paper focuses on the twofold research challenge of
enabling on-device and near zero-latency SHM of wind tur-
bine blades using a novel approach based on aerodynamic
I. I NTRODUCTION
measurements of the pressure distribution around the blade
Today, wind energy is considered one of the most important surface. More specifically, the contributions of this work are as
technologies to significantly reduce CO2 emissions and help follows: (i) A blade damage detection tiny Machine Learning
mitigate global warming based on the recently agreed EU (tinyML) model based on a recently released open source
targets [1]. For the widespread adoption of wind turbines as a dataset [6] featuring 6 damage classes. The XGBoost model
power source, it is essential to optimize maintenance costs and can detect a crack with varying severity, from 5 mm to 20 mm,
operation efficiency while reducing negative environmental and a classification accuracy above 83 %. Unlike previous SoA
impacts [2]. Current industry trends in this field call for real- publications based on vibration analysis, this work focuses
time and always-on monitoring systems not only on the elec- only on aerodynamic features acquired through an array of 40
tric generator and the tower but also including a deeper under- barometers placed around the blade surface. (ii) A computa-
standing of the blade aerodynamics and working load during tionally optimized feature extraction algorithm to preprocess
normal operations [3]. Specifically for offshore wind farms, acquired barometric data into a lower dimensional representa-
where high installation and periodic maintenance costs require tion. Starting from an input matrix of [40×512], the algorithm
a high volume of generated electricity to be commercially reduces it to [20 × 512] and subsequently to a list of extracted
viable, wind turbines operate in turbulent airflow, encountering features, such as mean, variance, skew, and kurtosis, with
strong and sometimes sudden wind speed changes, which can dimension [5 × 20]. (iii) A robust data handling approach
cause significant load fluctuations, reducing the programmed supporting missing or faulty sensors, which is a common
lifespan and decreasing maintenance intervals [4]. situation for sensor nodes deployed in harsh environments
Current State of the Art (SoA) on continuous blade moni- such as offshore wind turbines. The proposed methodology
toring is based on wireless and low-power sensor nodes with is autonomously able to detect faulty barometers and discard
limited computational capabilities and energy onboard. There- them. Thanks to the redundant information from an array
of 40 barometers, the model can work with up to 50 % of around the airfoil have only been measured indirectly via audio
invalid input data. (iv) An optimized implementation of the emission or anomalous vibrations [4], [13].
crack detection model to run completely onboard the sensor Moreover, all the methods above are inherently limited
node used to acquire the dataset [6]. The model classifier by steps (ii) and (iii) since they outsource the data analysis
runs in 3.4 ms, while feature extraction takes an extra 575 ms and damage classification to remote servers. However, this
on a 48 MHz ARM Cortex-M4F with 80 kB of RAM. To approach limits system performance due to the bottleneck
process the data onboard and execute the tinyML classification generated by the data transfer and increases the system latency
model, only 9 mJ of energy is consumed. Our model features a response from minutes to even hours [14]. Transferring large
classifier with 8.6 k parameters stored in Flash memory, while numbers of samples via low-power wireless link is a slow
feature extraction takes only 3 kB of RAM excluding data process, often involving a bit rate below 1 Mbps [3].
buffers. The full damage detection pipeline runs in 579 ms, Therefore, this paper focuses on the open challenge of
while the acquisition of an input window takes 5 s; allowing detecting structural damages via aerodynamic measurements
real-time and continuous monitoring. acquired through an array of 40 barometers placed on the blade
Therefore, this paper not only proposes a unique approach surface. At the same time, the proposed anomaly detection
for structural health monitoring on wind turbines but also model runs completely onboard the sensor node, featuring a
decreases the overall system energy consumption by a factor latency below 1 s.
of 13× compared to SoA without onboard processing solu- III. DATASET
tions [3], [5]. This is because data transmission from the sensor
node for remote processing would require 117 mJ and 1 s, This work utilizes the open source dataset in [6]. It was
even before the energy consumption of the remote processing collected using the Aerosense system [3], [5] composed of
system is taken into account. a single sensor node which is described further below. The
test depicted in Figure 1 is conducted to mimic a progressive
II. R ELATED W ORKS structural degradation of a wind turbine blade by applying
damage in the form of an artificial crack. It consists of six
SHM and predictive maintenance are an active topic in the damage classes. The first two cases are with an undamaged
recent advancement of Industry 4.0 [7]. Specifically, SHM cantilever beam; one of them is with an added mass of 246 g
is particularly attractive for wind turbines, where an active mounted 116 cm from the base. Then, the other 4 classes
research strand is pursuing different approaches [3], especially are with an increasing crack size, namely 5 mm, 10 mm,
with non-destructive techniques [2]. Generally, all SHM sys- 15 mm and 20 mm. For each class, the structure is excited
tems apply the following steps: (i) sensor data acquisition; at two different frequencies by the cantilever, namely 1.0 Hz
(ii) data transmission; (iii) signal processing; (iv) data inter- or 1.9 Hz, the wind speed is set to 10 m s−1 or 20 m s−1 the
pretation. [3], [7] For data acquisition, one can differentiate Angle of Attack (AoA) is set to either 0◦ or 8◦ , although for
between active and passive excitation. In active systems, an data quality reasons only the 0◦ AoA data is used in this work.
actuator excites the blade while an accelerometer measures the This results in 4 combinations for each damage class. For each
response [8]. This solution cannot be implemented outside a of these, 3 recordings exist, each approximately 150 s long.
research environment and thus is not discussed further. Mean- In detail, the dataset includes 228 experiments of ≈2 min
while, global passive measurement approaches are mostly each including 40 barometric traces sampled at 100 Hz, 5
vibration-based, detecting the modal frequency variations of acceleration signals placed at different blade positions sampled
the structure under study. In this case, arrays of accelerometers at 2 kHz, a front-facing video recording, and a stream-wise
are often employed. Conversely, local measurements focus motion capture.
on individual parts of the blade, e.g., using strain gauges to The dataset [6] was collected in an open wind tunnel,
detect structural deformations in predetermined locations [9]. featuring a test section of 40 cm × 40 cm. The measurement
Further, acoustic phenomena have been investigated to detect system was mounted onto a cross-section of a 3D-printed
crack formation, delamination, corrosion, and debonding [10], NACA633418 airfoil with a chord length of 16 cm. The width
[11]. In addition to the sensor-based approaches within or of the airfoil is 45 cm, and it is mounted onto a heaving
on the blade’s surface described above, other work has ap- cantilever beam made out of aluminum with the dimensions
plied techniques to monitor turbine blades externally using of 1 cm × 4 cm, as depicted in Figure 1. At the tip of the
recent breakthroughs in computer vision and machine learning. cantilever, a small motor is mounted with an eccentric mass of
Drones equipped with cameras have been used to look for 80 g and a radius of 8 cm to induce a harmonic excitation. To
damage [12]. More recent devices can be deployed directly simulate an increasing crack size, the cantilever beam cross-
on the blade surface while in operation [3] to measure its section was reduced with a common metal saw. This reduction
aerodynamic behavior with arrays of barometers and differen- is located 10 cm away from the support of the cantilever (see
tial pressure sensors [5] without requiring any modification of Figure 1). To adjust the AoA, two airfoils were printed with
the blade itself. As there is a direct correlation between wind different ducts, one for an AoA of 0◦ and one for 8◦ . The
turbine efficiency and the blade aerodynamic performance, experimental setup has a Reynolds number Re of 1.28e5 at
i.e., the generated lift, monitoring it in real-time would be 10 m s−1 .
beneficial to predict detrimental events, such as structural The Aerosense system [3], [5] is a custom sensor node
and surface damage. However, so far air flow discrepancies system which features 40 Micro-Electro-Mechanical Systems
 0.18
0.20
0.22
0.24

CP
0.26
0.28
0.30
0 1 2 3 4 5 0 5 10 15 20 25
Time [s] # Samples
102 10

1
101

CP
100 0.5 1.5 2.1 5

10 1

0 10 20 30 40 50
Frequency [Hz]
Fig. 1. Dataset acquisition setup in [6]. The acquisition system comprises 5
accelerometers and 40 barometers installed on a cross-section of a 3D-printed Fig. 2. Typical barometric signal of a heaving airfoil in a wind tunnel in the
NACA633418 airfoil supported by a 160.5 cm aluminum bar. A crack of healthy state. The upper part shows the time domain and the distribution of
increasing severity is manually applied. the time series. From this distribution, simple temporal features are extracted.
The frequency domain is shown in the lower section, with the bandwidth of
the band-pass filter highlighted in red and magnified.

(MEMS) barometers of the type LPS27HHW from STMicro- Preprocessing

electronics. They are mounted on a flexible strip along a 2D [Input] [40x512] [20x512]
Chopper Reducer
airfoil section and sample with 100 Hz. The main processing
unit is a TI CC2652P running at 48 MHz. It has an ARM Window Size Sensors Set
Cortex-M4F core and has 80 kB of RAM. In addition to the
Feature Extraction
352 kB on-chip flash, the system includes a Flash memory
[20x1]
of 512 MB. To transmit the collected data to a local server, Summation FIR Filter Hann Window

it features a Bluetooth Low Energy (BLE) downlink with a

bandwidth of 1.2 Mbps, requiring up to 178 nJ per bit [5]. Skew Mean
[20x4]
Figure 2 shows a typical signal of a barometer. The window + Kurtosis Variance
is from a healthy experiment, with a wind speed of 20 m s−1
and an excitation of 1.9 Hz. The upper part of the Figure 2
Classifier
shows the signal in the time domain, where the noise generated [20x5] [Class]
by the wind turbulence is visible. On the right side, the Score Sigmoid

distribution of the same signal is shown. Figure 2 also depicts

the signal in the frequency domain. There, one can see that
mainly the bias (DC) part and the modal frequency of the Fig. 3. Functional diagram of the model. The model is split into 3 functional
blocks indicated by different colors. The preprocessing handles the input
blade are present. data, which is generated at 100 Hz by 40 sensors. Between the blocks, the
data size described of the form [channels x samples]. The feature extraction
block extracts 1 frequency- and 4 time-domain features for each channel,
IV. B LADE DAMAGE D ETECTION M ODEL respectively. The classifier outputs the most likely damage class.

Since this work aims to detect wind turbine structural A. Preprocessing

damage using only the aerodynamic properties of the blade,
only the 40 LPS27HHW barometers are used as input data. As shown in Figure 3, the preprocessing step consists of
The model architecture consists of multiple stages, as shown two stages. First, the 40-channel time-series data sampled at
in Figure 3. In the first step (Preprocessing), the data is 100 Hz is split into windows of N samples, and a subset
divided into fixed-length time windows, and the redundant of sensors is selected for further processing. The window
sensors are removed to reduce the input dimensionality. In size is the optimum outcome of a trade-off between memory
the feature extraction step, the main components of the signal limitations and classification accuracy. To keep the memory
are extracted in the time and frequency domain. The resulting footprint small, a short window size is beneficial because
values are then fed to the computationally optimized classifier multiple parts of the pipeline scale accordingly. For example,
based on an XGBoost gradient-boosted decision tree structure the memory needed to store the weights for the Hann-Window.
to minimize the memory and computation requirements. The On the other hand, the measurement of the energy in the
model assigns a class proportional to the crack severity to eigenfrequency of the blade is less prone to noise with longer
each window, described in Section III. Each time window is periods. An investigation showed that a window size of 512
classified separately and independently of other windows. samples (5 s) gave the highest classification accuracy while
also being easily implementable on the CC2652P with existing 5mm Crack
102
Digital Signal Processing (DSP) libraries, enabling optimal 10mm Crack
15mm Crack
execution speed on the MCU. 20mm Crack
101
The other parameter, the sub-set of sensors, comes from

CP
the measurement redundancy of the acquisition system, with
100
the barometers placed a few millimeters apart. Maximum
accuracy was reached when using sensors covering all the
airfoil sections. For example, one cannot completely remove 10 1
1.5 1.6 1.7 1.8 1.9 2.0 2.1
the sensors around the trailing edge. On the other hand, only Frequency [Hz]
using alternate sensors (with either odd or even index) as
inputs to the model halves the input dimension, and thus the Fig. 4. Frequency domain of the data zoomed in on the modal frequency of
the blade. A representative selection of the different damage classes is shown.
model complexity, while maintaining similar accuracy. The
reduction from [40×512] to [20×512] is shown in Figure 3. At
the same time, the remaining sensors can be used as a backup C. Data Augmentation
in case of hardware failure, ensuring correct system operation To increase the number of data points available for model
even if a large subset of barometers malfunction over time. For training and evaluation, two data augmentation strategies were
example, in [6], many experiments have been conducted with used. First, although data from only 20 of the 40 pressure
one or more faulty sensors. In Figure 3, the reducer block sensors is supplied to the model, the redundant sensor data
automatically discards channels providing meaningless data, was not discarded. Second, a striding mechanism with a stride
for example, if it is outside a realistic threshold or too variable. overlap of 50% was used in the windowing step for training
Assuming the sensor adjacent to a defective sensor is not also and validation data. No striding mechanism is used in the test
faulty, its data can be substituted for the neighboring defective set (see Section IV-D).
sensor without significantly affecting performance.
D. Classifier
B. Feature Extraction For the classifier, only architectures that are executable
on a resource-constrained microcontroller were considered.
As shown in Section III, the pressure signals are noisy, and Decision trees were chosen for two reasons. Firstly, they
the modal frequency component is often not very prominent. are one of the smallest model families in terms of memory
However, the clustering of the different damage classes can consumption. Secondly, decision trees have no library depen-
be seen in the frequency domain, as shown in Figure 4. The dencies, which are otherwise often an insurmountable problem
features were selected to be computationally and memory- when converting models from high-level languages such as
efficient to run in real-time on a low-power MicroController Python to plain ANSI C code. Specifically, gradient-boosted
Unit (MCU) while still allowing the model classes to be dis- trees (XGBoost) proved to be powerful and self-contained
tinguished. A total of 5 features were selected and computed classification architecture [15], especially since Python li-
using two separate signal paths, shown in Figure 3. The first braries exist to automatically generate C-code functions based
path computes four simple statistical features - the mean, on the pre-trained XGBoost model.
variance, skew, and kurtosis. The second path, composed of To train the crack classifier, the data set is split into a train
the Summation, FIR, and Hann Window blocks, measures the and a test set. Since the dataset has three recordings for each
energy of the modal frequency. First, a Hann-Window removes set of parameters, two were chosen for the training set and
artifacts in the frequency domain. Then, a band-pass filter is one for the test set. This results in a 66/33 train-test split.
applied featuring 200 taps, tuned to the modal frequency, start- In addition to the data augmentation strategies described in
ing from 1.5 Hz to 2.1 Hz. This is also indicated in Figure 2 Section IV-C, a stratified 10-fold cross-validation strategy was
by the red section in the frequency domain. After the filter, used during training to minimize model overfitting.
the vector is squared and summed, giving the signal energy by The training resulted in a classifier consisting of 60 es-
Parseval’s theorem. Since the signal contains only the modal timators per damage class. XGBoost internally creates an
frequency, this calculation gives a good approximation of the independent estimator per class, giving 360 estimators with
energy of the modal frequency of the signal, which can be a maximum depth of 4. In total, the estimators have 4126
used as a model feature. splits and 4488 leaves, resulting in 8614 parameters. Since
Notably, other than the modal frequency, each channel’s the parameters are saved as 32-bit floats, the model needs
DC component (mean) was an essential feature for damage approximately 35 kB of storage. The classifier achieves 84.1%
classification. From an aerodynamic perspective, it describes accuracy over 6 classes of damage, with an increasing crack
the flow distribution around the blade section and its aerody- size from 0 mm to 20 mm, on the test set executing in the
namic performance in generating suctions as expected. Thus, a training environment, also listed in Table I.
variance in the pressure distribution around the airfoil indicates
the blade’s structural behavior that can deviate from the orig- E. On-Device Implementation
inally intended response. Therefore, even if computationally To implement the model onto the Aerosense system in
lightweight, the channel mean is a fundamental indicator of real-time while minimizing energy consumption, the blade
the blade’s condition. monitoring pipeline of Figure 3 is implemented in C and
optimized explicitly for the DSP library of the ARM Cortex- TABLE I
M4F. C LASSIFICATION ACCURACY AND WEIGHTED PRECISION OF THE
PROPOSED PIPELINE DEPLOYED ON A P ERSONAL C OMPUTER (PC) AND
The feature extraction uses the ARM CMSIS library, specifi- THE ARM C ORTEX -M4F.
cally the DSP variant for the Cortex-M series. The first path of
the feature extraction contains the filter, as shown in Figure 3. PC (x86 Architecture) Aerosense (Cortex-M4F)
Accuracy 84.1% 83.1%
Next, the second path of feature extraction consists of the
Precision 84.4% 83.6%
statistical features. Also, the ARM CMSIS has built-in support
for the mean (x) and the variance (s2 ). Both are necessary
to compute the skew (g) and the kurtosis (w). Equation (1) Healthy 13.6% 0.0% 2.1% 0.0% 0.9% 0.1% 0.14
shows the formula for the intermediate values (zi ), which are
calculated once for every input window. H.with AM 0.5% 13.5% 0.2% 0.4% 0.0% 2.2% 0.12

0.10
5mm 2.5% 0.0% 12.2% 0.8% 1.1% 0.0%
i=0 i=0
xi − x 1 X 1 X 0.08
zi = (1) g= zi3 (2) k= zi4 (3)
s n n 10mm 0.9% 0.5% 1.6% 13.4% 0.0% 0.2%
n n 0.06

15mm 0.5% 0.1% 0.2% 0.2% 15.7% 0.0% 0.04

Splitting the feature computation into basic successive steps
0.02
drastically reduces the computation time since each time series 20mm 0.8% 1.0% 0.0% 0.0% 0.0% 14.9%
is only accessed in memory once, and the total number of 0.00
Healthy H.with AM 5mm 10mm 15mm 20mm
instructions is reduced by a factor of 3 compared to performing
the calculations separately. Fig. 5. Confusion Matrix of the model’s classification executed on the
Lastly, the porting of the decision tree is performed via the Aerosense system. H. with AM stands for Healthy with Additional Mass.
m2cgen Python module to convert the pre-trained XGBoost
model to C code in 32-bit floating-point precision to be used
for inference on the sensor node. distributed class accuracy among the dataset experiments as
shown in the diagonal of Figure 5.
V. E XPERIMENTAL R ESULTS Figure 5 shows the confusion matrix of the model on the
A. Measurement Setup Aerosense system. The squares on the diagonal are the true
The execution time and the power consumption are mea- positives per class, which is always better than 12 %, where
sured to characterize the model on the sensor node. To provide 16.6 % is a perfect score. Further, the colors indicate that
realistic measurements, the blade health classifier described in the few errors are spread over all classes, with only 3 minor
this paper is tested on the Aerosense MCU, measuring the concentrations of errors - two between the healthy state and the
firmware performances via CPU cycles. In practice, this is 5 mm crack and one between the healthy state with additional
achieved with the help of the Data Watchpoint and Trace Unit mass and the 20 mm crack. Most other errors are between sim-
(DWT). The DWT provides low-level access to the Cortex- ilar damage classes. A classifier only differentiating between
M4F, monitoring different events in the execution flow, such crack sizes, smaller or bigger than 5 mm, will achieve more
as stalls and cycles. All latency measurements were executed than 80 % accuracy.
1000 times to decrease the impact of interrupts and external Table II shows the energy and latency of the model executed
events. on the Aerosense system. The Preprocessing block in Figure 3
Further, the power consumption was characterized using a is excluded from the table because the latency and energy
Nordic Power Profiler Kit II from Nordic Semiconductor. The contributions of the Chopper and Reducer are insignificant.
power consumption measurements were conducted ten times, Therefore, Table II is split into the remaining parts of the
and the average value is given in Table II. pipeline. Because of the FIR filter, the energy feature (Sum-
mation, FIR Filter, and Hann Window in Figure 3) is 25×
B. Model Accuracy and Latency more expensive time-wise than the other preprocessing stages
Table I compares the classification accuracy and weighed (Skew, Mean, Kurtosis, and Variance in Figure 3). Further,
precision of the on-device implementation on the Aerosense the full model needs about 579 ms to execute. Since the model
system to the original XGBoost implementation in Python. can be stored completely in the MCU, the energy consumption
Since the Python version runs on common x86 Hardware, scales linearly with the execution time, as reported in Table II.
we expect the models to behave similarly, as the main differ- The latency of the processing of the 512-sample window at
ences are only the instruction set and the differing hardware 579 ms is one order of magnitude faster than real-time – at
platforms. Although Python was forced to use 32-bit floats, 100 Hz each window takes approximately 5 s to collect.
minor numerical differences are expected due to the 64-bit The RAM usage of the model is optimized to be mini-
nature of the host platform. Additionally, Table I shows that mal, using only 3 kB of the heap for the filter taps and at
the precision and accuracy for both variants are very similar. maximum 1 kB of the stack for classifier inference. Further,
Notably, the optimized model for the Aerosense still reaches an the pre-trained model occupies 35 kB in FLASH for the
accuracy above 83 % over 6 damage classes, with an evenly 8.6 kparameters.
 TABLE II ACKNOWLEDGMENT
L ATENCY, T IME , AND E NERGY RESULTS OF THE M ODEL RUNNING ON
THE A EROSENSE SENSOR NODE . ST (S IMPLE TIME ) F EATURES INCLUDE The experiments were performed within the French-Swiss
S KEW, M EAN , K URTOSIS , AND VARIANCE . E (E NERGY ) F EATURES project MISTERY funded by the French National Research
INCLUDE S UMMATION , FIR F ILTER , AND H ANN W INDOW. Agency (ANR PRCI grant no. 266157) and the Swiss National
ST Features E Features Classifier Total Science Foundation (grant no. 200021L 21271). Moreover,
Cycles 919 k 26.7 M 164 k 27.7 M Imad Abdallah and Yuriy Marykovskiy supported the realiza-
Time [ms] 19.2 556 3.40 579 tion of this work.
Energy [mJ] 0.300 8.419 0.053 9.068
R EFERENCES
[1] S. D. Ahmed, F. S. Al-Ismail, M. Shafiullah, F. A. Al-Sulaiman, and
I. M. El-Amin, “Grid integration challenges of wind energy: A review,”
VI. D ISCUSSION IEEE Access, vol. 8, pp. 10 857–10 878, 2020.
[2] H. Zhao, G. Chen, H. Hong, and X. Zhu, “Remote structural health
The results show that it is possible to detect wind turbine monitoring for industrial wind turbines using short-range doppler radar,”
blade faults in real time using only the pressure sensors in IEEE Transactions on Instrumentation and measurement, vol. 70, pp. 1–
9, 2021.
wind tunnel conditions. The onboard processing promises [3] T. Polonelli, H. Müller, W. Kong, R. Fischer, L. Benini, and M. Magno,
a significant reduction in the volume of data that must be “Aerosense: a self-sustainable and long-range bluetooth wireless sensor
transmitted for remote processing, which in turn promises node for aerodynamic and aeroacoustic monitoring on wind turbines,”
IEEE Sensors Journal, vol. 23, no. 1, pp. 715–723, 2022.
a significant reduction (by a factor of 13) in the energy [4] M. Rezamand, M. Kordestani, R. Carriveau, D. S.-K. Ting, M. E.
consumption and the latency of fault detection. This is because Orchard, and M. Saif, “Critical wind turbine components prognostics:
data transmission of a single measurement window of 512 A comprehensive review,” IEEE Transactions on Instrumentation and
Measurement, vol. 69, no. 12, pp. 9306–9328, 2020.
samples from the 40 barometers of the sensor node for remote [5] R. Fischer, H. Mueller, T. Polonelli, L. Benini, and M. Magno,
processing would require 117 mJ and 1 s [3], [5], even before “Windnode: A long-lasting and long-range bluetooth wireless sensor
the energy consumption of the remote processing system is node for pressure and acoustic monitoring on wind turbines,” in 2021 4th
IEEE International Conference on Industrial Cyber-Physical Systems
taken into account. On the other hand, processing onboard (ICPS). IEEE, 2021, pp. 393–399.
requires just 9 mJ. Moreover, data transmission could be [6] I. Abdallah, J. Deparday, Y. Marykovskiy, and S. Bar-
avoided entirely if no fault is detected, and be as low as a ber, “Aerosense measurements: Wind tunnel EPFL,” 2023.
[Online]. Available: https://mostwiedzy.pl/en/open-research-data/
single byte if a problem occurs. aerosense-measurements-wind-tunnel-epfl,60901070464152-0
A model developed for real-world use would likely need [7] T. Polonelli, A. Bentivogli, G. Comai, and M. Magno, “Self-sustainable
IoT wireless sensor node for predictive maintenance on electric motors,”
to be more complex, as real wind turbines are exposed to in 2022 IEEE Sensors Applications Symposium (SAS). IEEE, 2022, pp.
significantly more variation in environmental and operating 1–6.
conditions and types of faults than those present in the dataset [8] D. Garcı́a and D. Tcherniak, “An experimental study on the data-
driven structural health monitoring of large wind turbine blades
used for the scope of this paper, in particular concerning crack using a single accelerometer and actuator,” Mechanical Systems and
position and size. However, the additional algorithmic com- Signal Processing, vol. 127, pp. 102–119, 2019. [Online]. Available:
plexity of an unsupervised algorithm could be compensated https://www.sciencedirect.com/science/article/pii/S0888327019301530
[9] P. Wu, Y. Wang, X. Zhang, J. Gao, L. Wang, and Y. Liu, “Wind turbine
by implementing further optimizations such as quantization blade breakage monitoring with mogrifier LSTM autoencoder,” IEEE
[16], and upgrading the MCU used on the Aerosense system Transactions on Instrumentation and Measurement, 2023.
to the newest generation of architectures, which provide lower [10] M. Dimitrova, A. Aminzadeh, M. S. Meiabadi, S. Sattarpanah Kar-
ganroudi, H. Taheri, and H. Ibrahim, “A survey on non-destructive
power consumption with higher computational performance. smart inspection of wind turbine blades based on industry 4.0 strategy,”
Applied Mechanics, vol. 3, no. 4, pp. 1299–1326, 2022.
[11] A. Mohammed, B. Hu, Z. Hu, S. Djurović, L. Ran, M. Barnes,
VII. C ONCLUSION and P. A. Mawby, “Distributed thermal monitoring of wind turbine
power electronic modules using FBG sensing technology,” IEEE Sensors
Journal, vol. 20, no. 17, pp. 9886–9894, 2020.
This paper presents a lightweight model for classifying the [12] D. Waleed, S. Mukhopadhyay, U. Tariq, and A. H. El-Hag, “Drone-
crack severity at the base of a wind turbine blade, relying only based ceramic insulators condition monitoring,” IEEE Transactions on
on aerodynamic data collected via an array of barometers. It is Instrumentation and Measurement, vol. 70, pp. 1–12, 2021.
[13] M. Khazaee, P. Derian, and A. Mouraud, “A comprehensive study on
implemented and profiled on a low-power and computationally structural health monitoring (SHM) of wind turbine blades by instru-
constrained sensor node, where it has an execution time of menting tower using machine learning methods,” Renewable Energy,
579 ms and a memory consumption of 3 kB and 35 kB for vol. 199, pp. 1568–1579, 2022.
[14] F. Di Nuzzo, D. Brunelli, T. Polonelli, and L. Benini, “Structural health
RAM and Flash memory, respectively. Trained on a dataset monitoring system with narrowband IoT and MEMS sensors,” IEEE
collected in a wind tunnel, our classifier reaches an accuracy Sensors Journal, vol. 21, no. 14, pp. 16 371–16 380, 2021.
above 84% with 6 classes of structural damage. This model [15] G. Kumaresan, K. Devi, S. Shanthi, B. Muthusenthil, and A. Samy-
durai, “Hybrid fuzzy archimedes-based light gbm-xgboost model for
can enable efficient real-time on-device SHM directly on a distributed task scheduling in mobile edge computing,” Transactions on
sensor node deployed on an operating wind turbine, leading Emerging Telecommunications Technologies, vol. 34, no. 4, p. e4733,
to a decreased detection latency and a wireless bandwidth 2023.
[16] M. Giordano, N. Baumann, M. Crabolu, R. Fischer, G. Bellusci, and
decongestion. Moreover, the power and bandwidth saved pro- M. Magno, “Design and performance evaluation of an ultralow-power
cessing the data onboard support battery-operated devices to smart IoT device with embedded TinyML for asset activity monitoring,”
last longer and to be deployed more extensively on wind farms IEEE Transactions on Instrumentation and Measurement, vol. 71, pp.
1–11, 2022.
at reduced costs.