Applied Acoustics 183 (2021) 108325

Contents lists available at ScienceDirect

Applied Acoustics
journal homepage: www.elsevier.com/locate/apacoust

A sound based method for fault detection with statistical feature

extraction in UAV motors
Ayhan Altinors a,⇑, Ferhat Yol a, Orhan Yaman b
a
Department of Electronics and Automation, Vocational School of Technical Sciences, Firat University, Elazig, Turkey
b
Department of Digital Forensics Engineering, Firat University, Elazig, Turkey

a r t i c l e i n f o a b s t r a c t

Article history: The motors of the Unmanned Aerial Vehicle are critical parts, especially when used in applications such
Received 6 May 2021 as military and defense systems. The fact that the brushless DC (BLDC) motors used in UAVs operate at
Received in revised form 29 June 2021 high speed causes malfunctions. In this study, propeller, eccentric and bearing failures, which are fre-
Accepted 20 July 2021
quently seen in UAV motors, were created. Then the fault diagnosis was made by applying the recom-
Available online 28 July 2021
mended method on the sound data received from the motors. Signal pre-processing, feature
extraction, and machine learning methods were applied to the obtained sound dataset. Decision tree
Keywords:
(DT), Support Vector Machines (SVM), and k Nearest Neighbor (KNN) algorithms are used for machine
UAV motors
Statistical feature extraction
learning. The results have been obtained using three different UAV motors of 1400 KV, 2200 KV, and
Machine learning 2700 KV. For the 2200 KV motor, the accuracy of 99.16%, 99.75%, and 99.75% was calculated in DT,
Sound-based fault detection SVM, and KNN algorithms, respectively. The high accuracy of the proposed method indicates that the
study will contribute to the studies in the relevant field. Another advantage is that the method is fast
and able to work in real-time on embedded systems.
Ó 2021 Elsevier Ltd. All rights reserved.

1. Introduction control. It should be monitored whether there is a malfunction in

the motors, which are the most important parts of the UAV. Our
1.1. Background main motivation is to propose an automatic real-time UAV motor
fault detection method. Current and vibration signals have been
Electric motors are widely used both in daily use and in indus- used in the previous fault detection methods (see Section 1.3).
try. Especially in the defense industry, high-speed electric motors More faults can be detected than current and voltage signals using
are used [1]. Different DC motors are preferred according to the sound signal, and also obtaining sound signals is easier and
characteristics of UAVs. Failure of electric motors creates problems cheaper. Many deep learning-based methods have been proposed
such as the inability of UAVs to fulfill their duties. Thus, both the in the literature to detect UAV motor fault detection. However,
operation fails and causes major accidents. large resources such as servers are needed for the implementation
of deep learning methods. In the proposed method, statistical fea-
ture extraction and machine learning have been used together.
2. Research motivation Therefore, the computational complexity of our model is linear
and it is a lightweight learning method. Moreover, our model can
Unmanned aerial vehicles are increasingly used for civilian and be implemented on the physical-world application easily and it
military purposes in many applications ranging from observation, can be implemented on an embedding system.
search and rescue, target detection, combat, surveillance, environ-
mental monitoring, disaster area research, and scientific mapping. 2.1. Literature review
Failure of UAVs when used mostly for military and defense tech-
nology can have irreparable consequences. For this reason, the UAV motors are widely used in both civil and military fields.
basic components that make up the UAV must be under constant Different faults may occur in UAV motors depending on operating
conditions [2,3]. It is very important to detect and diagnose these
⇑ Corresponding author. faults earlier. There are many studies in the literature about UAV
E-mail addresses: [email protected] (A. Altinors), [email protected] motor failures [4]. Looking at these studies, it is seen that only
(F. Yol), [email protected] (O. Yaman). bearing, magnet, and eccentric failures could be detected [5,6].

https://doi.org/10.1016/j.apacoust.2021.108325
0003-682X/Ó 2021 Elsevier Ltd. All rights reserved.
A. Altinors, F. Yol and O. Yaman Applied Acoustics 183 (2021) 108325

Iannace et al. [7] proposed an Artificial Neural Network-based erature. Generally, propeller and balance faults are detected. In
method for detecting faults in UAV propellers. The dataset was this study, propeller, eccentric and bearing failures can be
collected by using a sound level meter on the UAV. The quadrotor detected. Since many types of faults can be detected simultane-
UAV was fixed on a tripod and operated. For balanced, unbalanced ously with sound-based methods, sound signals are used in this
with one strip, and unbalanced with two strips, the microphone study. The main purpose of this study is to detect more than one
collected sound signals at four positions. By preprocessing the type of fault. At the same time, it is to develop a lightweight
sound signals, 99.03% accuracy was calculated with an Artificial method that can work on UAVs in real-time with high accuracy.
Neural Network. Magdaleno et al. [8] used sound signals to detect
balance in UAV propellers. DWT decomposition was performed on 2.2. Our method
the sound signals and the balance in the propeller was determined
by spectral analysis. Balanced, one unbalanced, two unbalanced, When the articles in the literature are scanned, it is seen that
three unbalanced and four unbalanced classes have been detected. vibration, current, and voltage signals are mostly used for fault
Fu et al. [9] developed a hybrid method based on CNN-LSTM for detection. But in fact, propeller and balance malfunctions are usu-
six-rotor UAVs and performed fault detection. Accuracy was ally detected with these signals. But, in this study, by using sound
calculated for four datasets by applying DNN, CNN, LSTM, and signals, propeller, eccentric and bearing malfunctions in UAV
CNN-LSTM methods. The highest result was obtained with the motors are diagnosed. The use of vibration, current, and voltage
CNN-LSTM algorithm with 94.12%. Liu et al. [10] developed a signals together increases the complexity of the method, and it is
sound-based diagnostic method for Quadrotors. The sound dataset also difficult to detect the mentioned faults. In the proposed
was collected by creating different scenarios on Quadrotors. They method, many faults have been detected with high accuracy by
created spectrograms from audio signals. Propeller failure was using sound signals. Statistical equations have been used to extract
detected by using Spectrograms data with CNN. The magnitude properties from sound signals. DT, SVM, and KNN algorithms were
of the propeller failure was also calculated. Zhang et al. [11] used for the classification of the obtained features. The reason why
proposed a Time-Domain Frequency Estimation-based application statistical equations and machine learning methods are preferred
for detecting UAV propeller failures. There are many studies in the in the proposed method is that the method can work quickly. Mal-
literature for the detection of propeller failures in UAV motors functions that occur during the flight of UAVs should be diagnosed
[12,13]. In the literature, collecting sound signals with accuracy early. For this reason, a real-time method has been developed that
is very important while developing UAV audio-based diagnostic is compatible with embedded systems that can be used on UAV.
methods. Because of the acoustics of the sound change indoors
and outdoors [14,15]. When collecting sound signals indoors,
2.3. Contributions and novelties
attention should be paid to the acoustics of the sound.
In the literature, besides sound-based condition monitoring and
In this study, an acoustic-based method is proposed for detect-
fault detection methods, vibration, current and voltage-based
ing propeller, eccentric and bearing failure, which are the most
methods are also available. Bondyra et al. [16] performed condition
important components of UAV motors. The main contributions of
monitoring and fault diagnosis for UAV propellers using signal pro-
the proposed method are summarized as follows.
cessing. Signals were collected using an IMU sensor. Condition
monitoring and fault detection were performed with the SVM clas-

In the study, high-speed UAV motors are used. Sound signals
sifier on the collected signals. According to Benini et al. [17] pro-
received from DC motors are used for fault detection. On the
poses an acceleration measurement-based method to detect
basis of sound signals, more faulties can be detected.
VTOL-UAV propeller failures. Statistical methods were used for fea-

Statistical equations are used for feature extraction from sound
ture extraction. Tri-axial accelerations signals were collected with
signals. In the study, 1200  44100 sound signals were received.
1 kHz sampling. At the same time, location data with GPS is
By using statistical equations, 1200  6 features are extracted
recorded on the SD card. Propeller failures were detected while
from 1200  44100 sound signals. Thus, a fast method that
performing Hovering, Climbing, Landing, and Complete flight
can work in real time has been developed. A high automatic
movements on the VTOL-UAV. 95% accuracy was calculated while
fault classification model has been proposed in this work using
detecting faults. Cheng et al. [18] developed a vibration-based
a lightweight machine learning technique.
method for diagnosing quadcopter faults. In the model developed,
they extracted features using RMS, Sample Entropy, and Standard
Deviation algorithms from the three-axis vibration data obtained 2.4. Study outline
from the accelerometer. Later, in the Self Organization Map
(SOM) model they developed using Tri-axial accelerometer data, In the second part of this study, information about UAV motors
they achieved 96.98% and 99.24% accuracy. Sadhu et al. [19] devel- is given. The motors used in the study are explained, and their
oped a neural network-based method for fault detection and iden- properties are given. In the proposed method, feature extraction
tification. By using Bi-LTSM and CNN, flight data was classified and from sound signals received from motors is explained. In addition,
a fault diagnosis was made. In the experiments performed, fault machine learning methods and parameters used are given. In the
detection was possible with 99.00% accuracy in simulation data fourth chapter, the experimental setup and classification results
and 85.00% accuracy in real-time experimental data. Lu et al. [20] are presented. In the fifth chapter, the results of the proposed
proposed a deep learning-based method for fault detection in method are analyzed. In the conclusion part, the results of the
UAV motors. In the method they developed, the acceleration of method are explained.
the increase in motor temperatures was monitored and fault pre-
diction was made. Bearing failures were determined from current 3. Material
and vibration data on an experimental setup. Current studies on
detecting UAV motor faults are listed in Table 1. Brushless motors make the winding energy exchange with
As can be seen in Table 1, many sensors such as sound, IMU, some electronic circuits instead of brush-collector. There are per-
temperature, current are used to detect malfunctions in UAV manent magnets on the outer wall of conventional DC motors,
motors. Many methods such as machine learning, signal process- and the windings in the rotor are energized in different positions
ing, and deep learning have been applied in the studies in the lit- with the help of a brush, creating opposite forces to each other,
2
A. Altinors, F. Yol and O. Yaman Applied Acoustics 183 (2021) 108325

Table 1
Literature review.

Studies Fault Types Dataset Method Accuracy

Wang et al. [21], 2020 General flight failures Control Parameters Quantified LSTM 95.1%
FDMAE 98.6%
Sadhu et al. [19], 2020 General flight failures GPS, Communication and Sensors CNN’s and LSTMs 90%
Titouna et al. [22], 2020 General flight failures Control Parameters Artificial Neural Networks –
Liu et al. [10], 2020 Propeller Failure Sound Data Spectrogram & CNN 90%
Ghalamchi et al. [23], 2020 Propeller Failure Vibration data (IMU) Discrete Fourier transformation (DFT) –
Kante et al. [24], 2020 Motors Failure ESC PWM Signals – –
Ray et al. [25], 2020 Motor Winding Failure Current Data Discrete Wavelet Transform (DWT) –
Cheng et al. [18], 2019 Motors Vibration Failure Mems Accelerometer & Gyro Machine Learning 96.8%
Keipour et al. [26], 2019 General flight failures Control Parameters Gaussian distribution 86.36%
Rangel et al. [8], 2019 Propeller Balancing Failure Sound Data Discrete Wavelet Transform (DWT) –
Benini et al [17], 2019 Propeller Balancing Failure IMU Data Discrete Fourier transformation (DFT) 95%
Iannace et al. [7], 2019 Propeller Failure Sound Data Artificial neural networks 96%
Wang et al. [27], 2019 Motors and Propeller Failure Vibration data FFT, Hilbert Huang Transform (HHT) –
Lu et al. [20], 2018 Motors Temperature Failure Temperature Data Deep Reinforcement learning –
Pourpanah et al. [28], 2018 Motors and Propeller Failure Vibration and Current Data CART 95.8%
KNN 96.4%
NB 96.5%
SVM 96.7%
QFAM-GA 97.8%
Aboutalebi et al. [29], 2018 Sensors Failure IMU Data Artificial neural networks –
Bondara et al. [16], 2017 Propeller Failure IMU Data FFT, Wavelet Packet Decomposition(WPT)Support –
Vector Machine(SVM)
Saied et al. [13], 2017 Motors Failure Speed Data SVM –

Table 2
Technical specifications of Wester A2212 model UAV motors.

Model Wester A2212 1400 KV Wester A2212 2200 KV Wester A2212 2700 KV
Operating voltage 2S-3S Li-po (7.4–11.1 V) 2S-3S Li-po (7.4–11.1 V) 2S-3S Li-po (7.4–11.1 V)
Current Capacity 12A(60S) 18.5A(60S) 18.5A(60S)
No load Current 10 V @ 0.5A 10 V @ 0.5A 10 V @ 1.8A
Maximum Efficiency 80% 82% 92%
Motor Dimensions 27.5x30mm 27.8x27mm 27x26mm
Weight 47gr 48gr 60gr
Shaft Diameter 3.17 mm 3.17 mm 3.17 mm

converting electrical energy into mechanical energy. But in brush- study, Wester A2212 model 1400 KV, 2200 KV, and 2700 KV UAV
less motors, the rotor and stator are in the opposite position. Since motors were used. The technical characteristics of the motors
1980 s, it has been preferred for applications that do not require shown in Fig. 1 are given in Table 2.
much power, especially with the development of magnets and Since the brushless motors used in UAVs are the types of motors
semiconductors. UAV motors are usually wound in 2 or 3 phases, operating at high speeds, malfunctions may occur during long-
but there are also brushless motors with more phases as well. term operations. In UAV motors, rotor-shaft, bearing, and stator
There are 4 and 8 poles UAV motors with at least two poles. In this failures usually occur [30,31].

Fig. 1. Motors used in the study (a) Wester A2212 1400 KV (b) Wester A2212 2200 KV (c) Wester A2212 2700 KV.

3
A. Altinors, F. Yol and O. Yaman Applied Acoustics 183 (2021) 108325

3.1. Rotor-shaft faults reason, a large laboratory was used and a microphone was placed
about 1 m away. A dataset was created in a noise-free environment
UAV motors are motor types that contain permanent magnets for three different motor types. A total of 300 s of sound was col-
and whose operation is relatively not very complicated. These per- lected for each failure condition. Since the recording frequency is
manent magnets, which are placed in the rotor at a certain angle, 44 kHz, 300  44100 data were collected. This process was
follow the rotating field on the stator and rotate the rotor. These repeated for four classes, and 1200  44100 data were collected
permanent magnets on the rotor are usually attached to the rotor for each motor type. Sample sound signals obtained from the
using strong adhesives. Especially if the motor is impacted or oper- microphone in the study are shown in Fig. 3.
ated with an unbalanced propeller, vibrations on the motor may
cause permanent magnets to be out of position. A very small slip 4. The proposed method
on the magnet can cause the motor to work with vibration and
noise and decrease the motor efficiency and cause the motor to In this study, sound data were collected in four different condi-
overheat [32]. At the same time, a physical impact that the motor tions: Healthy motor, propeller failure, eccentric failure, and bear-
may receive from the outside (generally the blows that may occur ing failure. The block diagram of the proposed method is shown in
during crushing) may cause the motor shaft to bend; however, the Fig. 4.
motor balance may deteriorate. In this case, the motor will run in a
loud and vibrating manner again. 4.1. Pre-Processing
Bearing faults
UAV motors used in UAVs are generally designed to allow good There is a pre-processing step before extracting features from
cooling with open windings. Although this design method has a the sound data collected for three different motors. For pre-
positive effect on the cooling of the motor windings, it has negative processing, the 1200  44100 sound data is normalized using Eq.
effects on the bearings of the motors, especially in dusty environ- (1).
ments. Dust particles in the air can cause the bearings to lose their
properties in a short time and to operate noisily. At the same time, Z  Z min
Z0 ¼ ð1Þ
faulties on the Rotor and Shaft, and also the use of improperly bal- Z max  Z min
anced propellers cause vibrations on the motor. However, motor In Eq. (1), the Z value denotes the property. The values of Z_max
bearings deteriorate in a short time and may cause noisy operation and Z_min represent the maximum and minimum values in the
[32]. array. After the normalization step, the sound data were arranged
Stator faults between 0 and 1. It has become suitable for feature extraction from
Stator faults are usually caused by the breakdown of the stator sound data.
winding insulation. There are several main reasons for these mal-
functions. The most common of these situations is that the motor 4.2. Feature extraction
gets hotter than normal due to the overload. In such overload sit-
uations, because excessive current will flow through the motor Statistical equations are used for feature extraction from sound
windings, there is more heating of the windings. If the high heat data. The Average, Standard Deviation, Variance, Correlation, Kur-
generated in the windings is not thrown out of the motor in a short tosis, and Skewness values of 1200  44100 sound data were cal-
time, the insulation on the windings will deteriorate and cause the culated. The statistical equations used in this study are given in
motor to burn and become inoperable [32]. Eqs. (2)–(7).
Propeller faults
PN
In the event that unmanned aerial vehicles hit any obstacle dur- i¼1 X i
f1 ¼ ð2Þ
ing their flight, breakage or deterioration of the wings may occur. N
This prevents the UAV from flying correctly [33]. In this study, fault vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
types are created by using three different UAV motors. These fail- uN
uP
ures are Propeller, Eccentric and Bearing failures. The experimental u ðX i  f 1Þ2
ti¼1
setup developed in the study is shown in Fig. 2. f2 ¼ ð3Þ
N
This process is repeated for three different speed UAV motors.
Thus, four classes of sound data are obtained for three motors. A PN 2
Samsung mobile phone and an external microphone connected it i¼1 ðX i  f 1Þ
f3 ¼ ð4Þ
were used while collecting sound data from the running motors. N
Thus, it was aimed to collect better quality audio dataset. The fol-
X
N
i  Xi  f 1
lowing are the specifications of the external microphone used: f4 ¼ ð5Þ
i¼1
rx

Polarity: Omnidirectional

Frequency range: 65 Hz–44KHz pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi PN 3
!3=2
NðN  1Þ x¼1 ðX i  f 1Þ
1

Signal/Noise: 74 dB SPL f5 ¼ N
PN ð6Þ

Output impedance: 1000 Ohms or less N2 1
x¼1 ðX i  f 1Þ
2
N

Connector: 3.5 mm stereo jack " ! ! #
PN 4

Battery type: LR44 N1 1
x¼1 ðX i  f 1Þ
f6 ¼ ðN þ 1Þ N
PN 3 þ6 ð7Þ
ðN  2ÞðN  3Þ 1
x¼1 ðX i  f 1Þ
2
N
The sound was collected with the ‘‘Voice Recorder” program on
the phone. While creating sound data, AAC (m4a) format, 320 kbps, In this study, statistical feature extraction was used for a
and 44 kHz sampling frequency were preferred [34]. A quiet labo- real-time method [35]. Average, Standard Deviation, Variance,
ratory of about 50 square meters was used to collect the sound sig- Correlation, Kurtosis, and Skewness values parameters were used
nals from the motors. The microphone is fixed at a distance of for statistical feature extraction. Thus, 1200  6 features were
about 1 m from the UAV motor. Being close to the UAV motor calculated for the 1200  44100 dataset. As a result of feature
may adversely affect the audio dataset because it is indoor. For this extraction, 1200  6 features were obtained for three motor types.
4
A. Altinors, F. Yol and O. Yaman Applied Acoustics 183 (2021) 108325

Fig. 2. Images of the experimental setup developed in this study (a) Healthy motor, (b) Propeller failure (c) Eccentric failure (d) Bearing failure.

Fig. 3. Sound samples collected for fault detection (a) Healthy (b) Propeller failure (c) Eccentric failure (d) Bearing failure.

In the study, it was aimed to extract a small number of features. 4.3. Classification methods
The reason for this is to develop a fast method that can work on
embedded systems. The classification was made after feature The properties obtained in the study were used together with
extraction. the classification algorithms. The result of the classification algo-

5
A. Altinors, F. Yol and O. Yaman Applied Acoustics 183 (2021) 108325

Fig. 4. Block diagram of the proposed method.

with this method, it simply performs the decision-making process

Table 3 by dividing the data into sections. The data to be classified are clas-
Parameters of proposed classification methods. sified and distributed sequentially into sub-sections until they
Fine DT Maximum number of splits 100 reach a leaf, depending on the rule mechanism to be used while
Split criterion Gini’s diversity index creating the decision tree. In order to understand the generaliza-
Surrogate decision splits Off tion ability of the created decision tree structure, test data are
Quadratic SVM Kernel function Quadratic tested at the root of the tree and remain bound by a single decision
Box constraint level 2 rule until the data reach the bottom. It is possible to separate these
Kernel scale mode Auto
groups by drawing a boundary between labeled groups in a plane
Manual kernel scale 1
Multiclass method One-vs-one for classification in the SVM algorithm. The place where the deci-
sion boundary to separate the planes will be drawn should be
Medium KNN Number of neighbors 10
Distance metric Euclidean the farest place from the group members. Support Vector Machines
Distance weight Equal set these limits. KNN algorithm is a method that is frequently used
Standardize data True in machine learning applications and other classification applica-
tions due to its simple and strong structure. This method realizes
the classification based on the proximity to the training examples
rithm gives the success of the method. DT, SVM, and KNN algo- in the feature space. The KNN algorithm performs class assign-
rithms are used in the proposed method. ments according to the k value. Although the implementation of
DT is a frequently used method in classification studies because the method is simple and powerful, it is one of the disadvantages
the rules used in creating tree structures are simple and under- of this method that it requires a large memory space, and the pro-
standable. In the classification problem that needs to be solved cessing load increases as a result of the increase in the volume of

Fig. 5. The accuracy results calculated for Fine DT, Medium DT, Coarse DT, Linear Discriminant, Gaussian Naïve Bayes, Linear SVM, Quadratic SVM, Cubic SVM, Fine KNN,
Medium KNN, Cosine KNN, and Ensemble Boosted Trees.

6
A. Altinors, F. Yol and O. Yaman Applied Acoustics 183 (2021) 108325

data. In the study, Fine_DT was preferred as the DT algorithm, 5. Experimental results
Quadratic_SVM as the SVM algorithm, and Medium_KNN as the
KNN algorithm. Classification parameters of Fine_DT, Quadra- The proposed method in the study was developed in the
tic_SVM and Medium_KNN algorithms are given in Table 3. MATLAB 2020a program. Results were obtained from a computer
Fine DT, Medium KNN, and Quadratic SVM algorithms were with i7-9700 CPU 3.00 GHz processor, 32 GB RAM and 6 GB Video
used to classify the calculated features in the proposed method. card. A four motor quadrotor was used in the experimental setup.
The reason why these algorithms are preferred is that they are Propeller, eccentric, and bearing failures were created for this
higher than the results of other algorithms. 12 classifiers have been quadrotor, and sound data was created. Dataset has sound data
used for classification. These classifiers are Fine DT, Medium DT, of four classes: Healthy, propeller, eccentric, and bearing. This pro-
Coarse DT, Linear Discriminant, Gaussian Naïve Bayes, Linear cess has been applied for Wester A2212 model 1400 KV, 2200 KV
SVM, Quadratic SVM, Cubic SVM, Fine KNN, Medium KNN, Cosine and 2700 KV UAV motors. Thus, four classes of sound data are gen-
KNN, and Ensemble Boosted Trees. The accuracy results calculated erated for three types of UAV motors.
for 12 classifiers are shown in Fig. 5. The proposed method was applied to the sound data, and the
As can be seen in Fig. 5, Fine DT, Medium KNN and Quadratic ROC curves given in Fig. 6 were calculated.
SVM were preferred among these classifications with high The best ROC curve for 1400 KV motor, as can be seen in Fig. 6,
accuracy. is AUC = 0.97. This ROC curve is calculated by the SVM algorithm.

Fig. 6. ROC curves obtained in classification of faults in 1400 KV, 2200 KV and 2700 KV UAV motors.

7
A. Altinors, F. Yol and O. Yaman Applied Acoustics 183 (2021) 108325

Fig. 7. Confusion matrix calculated for classification of faults in 1400 KV, 2200 KV and 2700 KV UAV motors.

The best ROC curve for 2200 KV motors obtained from SVM and 1 XNC
TP C
KNN algorithms is AUC = 1.00. The best ROC curve for 2700 KV UAR ¼ ð10Þ
NC c¼1 TP c þ FNc
motor is AUC = 1.00. This ROC curve is calculated with the SVM
algorithm as in a 1400 KV motor. It is calculated as AUC = 0.99 sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
for 2700 KV motor by DT and KNN algorithms. The proposed YNC NC TP c
method was run with 1000 iterations, and the confusion matrix GM ¼ c¼1 TP þ FN
ð11Þ
c c
obtained was shown in Fig. 7.
As can be seen in Fig. 7, the best classification for 1400 KV
motor is calculated with the SVM algorithm. When viewed on a 2  ðUAP  UARÞ
F1 ¼ ð12Þ
class basis, the highest value for Healthy (Class 1) and Propeller UAP þ UAR
(Class 2) classes was obtained with the SVM algorithm. The highest
value for the eccentric (Class 3) class was obtained with KNN. The
TP c
highest value for the bearing (Class 4) class was calculated with DT IoU ¼ ; c ¼ f1; 2;    NCg ð13Þ
TP c þ FPc þ FNc
and SVM. The highest results for the 2200 KV motor were calcu-
lated with SVM and KNN algorithm. For 2700 KV motor, SVM for Where TPC, TNT, FPC, and FNc are the number of true positives,
Healthy (Class 1), DT for Propeller (Class 2), SVM for Eccentric true negatives, false positives, and false negatives, respectively. N c
(Class 3), SVM and KNN for Bearing (Class 4), the highest classifica- is the number of classes. The results obtained for 1400 KV,
tion was calculated. For 1400 KV, 2200 KV, and 2700 KV UAV 2200 KV, and 2700 KV Brushless DC motors are given in Table 4.
motors, DT, SVM, and KNN algorithms were run 1000 iterations As seen in Table 4, the highest accuracy value for a 1400 KV
and Accuracy (CACC), Unweighted Average Precision (UAP), motor was calculated as 91.25% with the SVM algorithm. In DT
Unweighted Average Recall (UAR), Geometric mean (GM), F- and KNN algorithms, accuracy values were calculated as 89.16%
measure (F1) and IoU parameters were calculated. These parame- and 86.50%, respectively. For the 2200 KV motor, the highest accu-
ters are given in Eqs. (8)–(13), respectively. racy value was calculated as 99.75% by SVM and KNN algorithms.
With the DT algorithm, the highest accuracy value is 99.16%. For
TP c þ TNc
CACC ¼ ; c ¼ f1; 2;    NCg ð8Þ the 2700 KV motor, the highest accuracy value was calculated as
TP c þ TN c þ FPc þ FNc
90.00% with the SVM algorithm. In DT and KNN algorithms, accu-
racy values were calculated as 88.91% and 89.00%, respectively.
1 XNC
TP C Accuracy, precision, recall, geometric mean, F1-measure, and IoU
UAP ¼ ð9Þ
NC c¼1 TPc þ FPc results obtained from the proposed method were examined, and
it was seen that the highest results were calculated for 2200 KV.
8
A. Altinors, F. Yol and O. Yaman Applied Acoustics 183 (2021) 108325

Table 4
Accuracy, precision, recall, geometric mean, F1-measure, and IoU (%) results of the used 3 classifiers.

Case Classification Parameter Accuracy Precision Recall Geo-mean F1-measure IoU

1400 KV DT Max 89.16 89.32 89.16 88.70 89.24 100
Min 85.08 85.23 85.08 84.04 85.15 97.79
Mean 86.89 86.98 86.89 56.16 86.93 99.05
Std 0.59 0.60 0.59 0.67 0.59 0.33
SVM Max 91.25 91.27 91.25 90.90 91.26 99.61
Min 89.50 89.46 89.50 89.00 89.48 98.06
Mean 90.45 90.44 90.45 90.04 90.45 99.32
Std 0.29 0.29 0.29 0.31 0.29 0.28
KNN Max 86.50 86.78 86.50 85.21 86.63 99.59
Min 83.50 83.58 83.50 81.60 83.54 96.38
Mean 85.11 85.30 85.11 83.62 85.20 98.05
Std 0.45 0.47 0.45 0.54 0.46 0.49
2200 KV DT Max 99.16 99.16 99.16 99.16 99.16 98.01
Min 98.25 98.25 98.25 98.23 98.25 94.17
Mean 98.78 98.79 98.78 98.78 98.78 96.22
Std 0.15 0.15 0.15 0.15 0.15 0.60
SVM Max 99.75 99.75 99.75 99.74 99.75 100
Min 99.33 99.33 99.33 99.33 99.33 98.34
Mean 99.60 99.61 99.60 99.60 99.60 99.25
Std 0.06 0.06 0.06 0.06 0.06 0.33
KNN Max 99.75 99.75 99.75 99.74 99.75 99.66
Min 99.58 99.58 99.58 99.58 99.58 99.00
Mean 99.66 99.66 99.66 99.66 99.66 99.33
Std 0.01 0.01 0.01 0.01 0.01 0.02
2700 KV DT Max 88.91 89.04 88.91 88.42 88.97 100
Min 84.33 84.49 84.33 83.27 84.41 98.64
Mean 86.86 87.01 86.86 86.19 86.93 99.51
Std 0.64 0.64 0.64 0.72 0.64 0.29
SVM Max 90.0 90.43 90.0 89.32 90.21 100
Min 87.66 88.18 87.66 86.66 87.92 99.32
Mean 88.93 89.44 88.93 88.11 89.19 99.92
Std 0.29 0.30 0.29 0.33 0.29 0.15
KNN Max 89.00 89.55 89.00 88.40 89.27 100
Min 87.25 87.76 87.25 86.48 87.50 99.66
Mean 88.19 88.74 88.19 87.53 88.47 99.66
Std 0.28 0.29 0.28 0.31 0.28 0.02

Table 5
Time of feature extraction, training, and test operations for DT, SVM, and KNN algorithms.

DT SVM KNN
Mean (ms) Std (ms) Mean (ms) Std (ms) Mean (ms) Std (ms)
Feature Extraction 4.14 0.03 4.14 0.03 4.14 0.03
Training (10-Fold CV) 169.5 68.2 1353.8 147.7 181.1 67.3
Test Model 26.0 24.4 36.0 28.7 28.5 25.4
Total Training 173.6 – 1357.94 – 185.24 –
Total Test 30.14 – 40.14 – 32.64 –

Running times of feature extraction, training, and test processes for algorithms. As a result, DT and KNN algorithms are more suitable
DT, SVM, and KNN algorithms are calculated in Table 5. for real-time operation in embedded systems.
As seen in Table 5, the feature extraction step is equal in all In the proposed method, 10 Fold Cross-validation was used and
algorithms. After the feature extraction process, training was car- the results were calculated. In Fig. 8, the results for DT and BT algo-
ried out with 10-Fold CV. It is the DT algorithm that completes rithms were calculated by fold-wise accuracies.
the fastest training process. Then it performs the KNN algorithm
training step with 181.1 ms. SVM algorithm is very slow compared 6. Discussion
to DT and KNN algorithm. The fastest algorithm for the test step is
DT algorithm with 26 ms. The test step was calculated as 36 ms Class by class results are calculated for three different motor
with the SVM algorithm and 28.5 ms with the KNN algorithm. types used in the study. The obtained graphics are shown in Fig. 9.
The total test time of 1  44100 sound data received in 1 s was The highest class by the class result for the 1400 KV UAV motor
30.14 ms with DT algorithm, 40.14 ms with SVM algorithm, and has been obtained for Propeller (Class 2) and Eccentric (Class 3)
32.64 ms with KNN algorithm. When the total training and total classes. For Class 2, the highest accuracy value, with 99.66% was
test times of the proposed method are examined, it is clearly seen calculated by the SVM algorithm. For Class 3, the highest accuracy
that DT is the fastest algorithm. It is seen that the KNN algorithm is value with 99.33% was obtained by the KNN algorithm. All classes
also fast. The slowest algorithm is calculated as SVM. When the gave high results for the 2200 KV UAV motor. The lowest accuracy
performances of the algorithms are examined in Tables 4 and 5, value for Class 2 was calculated with 97.33% by DT algorithm. All
it is seen that DT and KNN algorithms are better. Although high accuracy values for other classes are higher than 98%. The class’s
success is achieved with the SVM algorithm, it is slower than other highest class result for the 2700 KV UAV motor is calculated for
9
A. Altinors, F. Yol and O. Yaman Applied Acoustics 183 (2021) 108325

Fig. 8. The calculated fold-wise accuracies employing DT, KNN and SVM classifiers of the defined 1400 KV, 2200 KV and 2700 KV.

Healthy (Class 1) and Eccentric (Class 3). When the class by class and KNN algorithms. In other motors, the accuracy is slightly
results obtained for three motor types were examined, it was seen lower. The difference in this accuracy value is related to the run-
that the best results were obtained with a 2200 KV motor. Class2 ning speed of the motors. In future studies, the motors will be
and Class3 for 1400 KV and Class1 and Class3 for 2700 KV motor operated at different speeds, and methods will be developed for
have the highest values. The accuracy values of Class 4 for the troubleshooting. When compared with the studies in the literature,
1400 KV motor and Class 2 for the 2700 KV motor are low. It is it is seen that the proposed method is efficient. The proposed
thought that the reason for this situation may be that it was method is highly accurate and fast. This shows that the proposed
affected by the noise in the environment while creating the dataset method can be applied to the embedded system. The lightweight
or that the artificially created faults were not sufficiently formed. of the method provides fast classification. Thus, it will be able to
The Scatter Plot graphs are given in Fig. 10 verify the class by class work on embedded systems in real-time.
results.
Three types of motors are used in the proposed method. Sound
7. Conclusions
data was created by operating the 1400 VK DC motor at approxi-
mately 40% speed. The 2200 KV motor was operated at about
In this study, the classification of faults in UAV motors com-
60%, and the 2700 VK motor at about 80% speed. As can be seen
monly used in UAVs is discussed. The importance of UAVs in both
in Table 6, the faults in the 2200 KV motor are classified with
military and civilian fields has been the motivation of our work. In
99.16%, 99.75%, and 99.75% accuracy, respectively, by DT, SVM,
order to diagnose the malfunctions in the motors used in UAVs,
10
A. Altinors, F. Yol and O. Yaman Applied Acoustics 183 (2021) 108325

Fig. 9. Classification accuracies (%) obtained for various classes using our proposed method with BLDC acoustic datasets.

Table 6
Comparison of the proposed method with similar studies in the literature.

Studies Fault Types Features Program Method Accuracy Time

(ms)
Sadhu et al. [19], 2020 General flight failures i7-3.4 GHz, 8 GB RAM AirSim drone simulator CNN’s and LSTMs 90% 82
1.2 GHz, 1 GB RAM Raspberry Pi 3 Model B and 312
2 GHz, 8 GB RAM Nvidia Jetson TX2 83
Liu et al. [10], 2020 Propeller Failure – MATLAB Spectrogram & CNN 90% –
Park et al. [36], 2020 Motors Failure i7- 9700 K- 3.60 GHz MATLAB Linear discriminant analysis 30.35% –
Principal component analysis 86.43% –
Multi-principal component 90.35% –
analysis
Fisher discriminant analysis 83.08% –
Partial least squares regression 98.04% –
Canonical variate analysis 96.92% –
Cheng et al. [18], 2019 Motors Vibration – – Machine Learning 96.8% –
Failure
Keipour et al. [26], 2019 General flight failures 2 GHz, 8 GB RAM Nvidia Jetson TX2 Gaussian distribution 86.36% 2020
Wang et al. [21], 2020 General flight failures – Jetson Nano ARM Airbone LSTM 95.1% 2612
ECP PCA-LSTM 98.6% 2588
Our Method Study i7-3.00 GHz, 32 GB MATLAB DT 89.16% 30.14
1400 KV Propeller failure RAM SVM 91.25% 40.14
Eccentric failure KNN 86.5% 32.64
Our Method Bearing failure DT 99.16% 30.14
2200 KV SVM 99.75% 40.14
KNN 99.75% 32.64
Our Method 2700 KV DT 88.91% 30.14
SVM 90.0% 40.14
KNN 89.0% 32.64

artificial faults have been created over the sample motors. The these data, and a fast algorithm has been developed. The results
sound dataset was created by creating malfunctions according to are obtained by using the obtained properties together with DT,
different motor types. Feature extraction has been made from SVM, and KNN algorithms. MATLAB2020a software was used to

11
A. Altinors, F. Yol and O. Yaman Applied Acoustics 183 (2021) 108325

Fig. 10. Scatter plots of the selected features (a) 1400 KV (b) 2200 KV (c) 2700 KV. The comparison of the proposed method with similar studies in the literature is shown in
Table 6.

Fig. 11. Real-time fault detection method for UAV motors considered for future studies.

test the proposed method. For 1400 KV motor, 89.16%, 91.25%, and For the 2700 VK motor, 88.91%, 90.0%, and 89.0% accuracy were
86.5% accuracy was calculated from DT, SVM, and KNN algorithms. obtained, respectively. In addition, the classification test times of
For the 2200 KV motor, the accuracy of 99.16%, 99.75%, and 99.75% these algorithms are calculated as 30.14 ms, 40.14 ms, and
was calculated from DT, SVM, and KNN algorithms, respectively. 32.64 ms, respectively.

12
A. Altinors, F. Yol and O. Yaman Applied Acoustics 183 (2021) 108325

As a result of this study, a real-time method has been developed Meet Power, Electron Comput ROPEC 2018 2019:2–6. 10.1109/
ROPEC.2018.8661459.
with the proposed statistical-based feature extraction and machine
[9] Fu J, Sun C, Yu Z, Liu L. A hybrid CNN-LSTM model based actuator fault
learning methods. The proposed method has been tested in a lab- diagnosis for six-rotor UAVs. Proc 31st Chinese Control Decis Conf CCDC 2019
oratory environment. Datasets were collected for three different 2019:410–4. 10.1109/CCDC.2019.8832706.
speed motor types using an external microphone connected to [10] Liu W, Chen Z, Zheng M. An Audio-Based Fault Diagnosis Method for
Quadrotors Using Convolutional Neural Network and Transfer Learning. Proc
the phone. When the accuracy and the working time of the algo- Am Control Conf 2020;2020-July:1367–72. 10.23919/ACC45564.2020.
rithm are examined, it is seen that it can work in real-time. 9148044.
[11] Zhang X, Luo H, Li K, Kaynak O. Time-domain frequency estimation with
application to fault diagnosis of the UAVs blade damage. IEEE Trans Ind
8. Future works Electron 2021;0046:1. https://doi.org/10.1109/tie.2021.3084177.
[12] Lee Jun-yong, Lee Won-tak, Ko Sang-ho, Oh Hwa-suk. Fault classification and
Unmanned Aerial Vehicles are used in both military and civilian diagnosis of UAV motor based on estimated nonlinear parameter of steady-
state model. Int J Mech Eng Robot Res 2020:22–31. https://doi.org/10.18178/
fields, especially in defense technology. The use of UAVs for imag-
ijmerr10.18178/ijmerr.10.1.22-31.
ing and defense purposes further increases the interest in these [13] Saied Majd, Lussier Benjamin, Fantoni Isabelle, Shraim Hassan, Francis Clovis.
vehicles. High-speed brushless motors are used in UAVs. Failures Fault diagnosis and fault-tolerant control of an octorotor UAV using motors
speeds measurements. IFAC-PapersOnLine 2017;50(1):5263–8. https://doi.
in these motors cause big problems. In the next study, the pro-
org/10.1016/j.ifacol.2017.08.468.
posed method will be applied to the embedded system and trans- [14] Iannace Gino, Ciaburro Giuseppe, Trematerra Amelia. Acoustical unmanned
formed into a real-time diagnostic application. In this study, the aerial vehicle detection in indoor scenarios using logistic regression model.
dataset was created by operating the motors sequentially while Build Acoust 2021;28(1):77–96. https://doi.org/10.1177/1351010X20917856.
[15] Ciaburro Giuseppe, Iannace Gino, Trematerra Amelia. Research for the
receiving the sound data. In the next study, while the sound is presence of unmanned aerial vehicle inside closed environments with
received from the faulty motor, other motors will also run in the acoustic measurements. Buildings 2020;10(5):96. https://doi.org/10.3390/
background. This will eliminate background sounds from other buildings10050096.
[16] Bondyra A, Gasior P, Gardecki S, Kasinski A. Fault diagnosis and condition
motors while working in real-time. Different methods will be monitoring of UAV rotor using signal processing. Signal Process - Algorithms,
developed in order not to be affected by environmental conditions. Archit Arrange Appl Conf Proceedings, SPA 2017;2017-Septe:233–8. 10.23919/
The real-time fault detection method we are considering as our SPA.2017.8166870.
[17] Benini A, Ferracuti F, Monteriu A, Radensleben S. Fault detection of a vtol uav
further studies for UAV motors is shown in Fig. 11. using acceleration measurements. 2019 18th Eur Control Conf ECC 2019
As can be seen in Fig. 11, it will be tested on UAVs with different 2019:3990–5. 10.23919/ECC.2019.8796198.
rotor numbers in future studies. The lightweight, high-accuracy [18] Cheng DL, Lai WH. Application of self-organizing map on flight data analysis
for quadcopter health diagnosis system. Int Arch Photogramm Remote Sens
and fast method developed in this study can be loaded into the
Spat Inf Sci – ISPRS Arch 2019;42:241–6. https://doi.org/10.5194/isprs-
embedded card. Thus, the real-time condition of the UAV can be archives-XLII-2-W13-241-2019.
monitored with the microphone and embedded card fixed on the [19] Sadhu V, Zonouz S, Pompili D. On-board deep-learning-based unmanned aerial
vehicle fault cause detection and identification. Proc – IEEE Int Conf Robot
UAV. In case of any abnormality/fault, our intended UAV condition
Autom 2020:5255–61. https://doi.org/10.1109/ICRA40945.2020.9197071.
monitoring system can send alerts to users. [20] Lu Huimin, Li Yujie, Mu Shenglin, Wang Dong, Kim Hyoungseop, Serikawa
Seiichi. Motor anomaly detection for unmanned aerial vehicles using
Declaration of Competing Interest reinforcement learning. IEEE Int Things J 2018;5(4):2315–22.
[21] Wang B, Peng X, Jiang M, Liu D. Real-time fault detection for UAV based on
model acceleration motor. IEEE Trans Instrum Meas 2020;69:9505–16.
The authors declare that they have no known competing finan- https://doi.org/10.1109/TIM.2020.3001659.
cial interests or personal relationships that could have appeared [22] Titouna C, Nait-Abdesselam F, Moungla H. An Online Anomaly Detection
to influence the work reported in this paper. Approach for Unmanned Aerial Vehicles. 2020 Int Wirel Commun Mob
Comput IWCMC 2020 2020:469–74. 10.1109/IWCMC48107.2020.9148073.
[23] Ghalamchi Behnam, Jia Zheng, Mueller Mark Wilfried. Real-time vibration-
Acknowledgments based propeller fault diagnosis for multicopters. IEEE/ASME Trans
Mechatronics 2020;25(1):395–405. https://doi.org/10.1109/TMECH.351610.
This work is supported by Firat University Research Fund, Tur- 1109/TMECH.2019.2947250.
[24] Kantue P, Pedro JO. Integrated Fault Detection and Diagnosis of an Unmanned
key. Project Numbers: TBMYO.20.01. Aerial Vehicle using Time Difference of Arrival 2020:336–42. 10.1109/
icstcc50638.2020.9259777.
References [25] Ray DK, Roy T, Chattopadhyay S. Skewness Scanning for Diagnosis of a Small
Inter-Turn Fault in Quadcopter’s Motor based on Motor Current Signature
analysis. IEEE Sens J 2020;XX:1–1. 10.1109/jsen.2020.3038786.
[1] Olson IJ, Atkins EM. Qualitative failure analysis for a small quadrotor
[26] Keipour A, Mousaei M, Scherer S. Automatic real-time anomaly detection for
unmanned aircraft system. AIAA Guid Navig Control Conf 2013:1–11.
autonomous aerial vehicles. ArXiv 2019:5679–85.
https://doi.org/10.2514/6.2013-4761.
[27] Wang X, Fan W, Li X, Wang L. Weak degradation characteristics analysis of
[2] Mitronikas E, Papathanasopoulos D, Athanasiou G, Tsotoulidis S. Hall-effect
UAV motors based on laplacian eigenmaps and variational mode
sensor fault identification in brushless DC motor drives using wavelets. Proc
decomposition. Sensors (Switzerland) 2019;19. 10.3390/s19030524.
2017 IEEE 11th Int Symp Diagnostics Electr Mach Power Electron Drives,
[28] Pourpanah F, Zhang B, Ma R, Hao Q. Anomaly Detection and Condition
SDEMPED 2017 2017;2017-Janua:434–40. 10.1109/DEMPED.2017.8062391.
Monitoring of UAV Motors and Propellers. Proc IEEE Sensors 2018;2018-
[3] Ciaburro G, Iannace G. Improving smart cities safety using sound events
Octob:2018–21. 10.1109/ICSENS.2018.8589572.
detection based on deep neural network algorithms. Informatics 2020;7(3):23.
[29] Aboutalebi Payam, Abbaspour Alireza, Forouzannezhad Parisa, Sargolzaei
https://doi.org/10.3390/informatics7030023.
Arman. A novel sensor fault detection in an unmanned quadrotor based on
[4] Faiz J, Jafari A. Interturn fault diagnosis in brushless direct current motors - A
adaptive neural observer. J Intell Robot Syst Theory Appl 2018;90(3-
review. Proc IEEE Int Conf Ind Technol 2018;2018-Febru:437–44. 10.1109/
4):473–84. https://doi.org/10.1007/s10846-017-0690-7.
ICIT.2018.8352217.
[30] Ferrão IG, Pigatto DF, Fontes JVC, Silva NBF, Espes D, Dezan C, et al. STUART:
[5] Medeiros RLV, Filho ACL, Ramos JGGS, Nascimento TP, Brito AV. A novel
ReSilient archiTecture to dynamically manage Unmanned aeriAl vehicle
approach for speed and failure detection in brushless DC motors based on
networks under atTack. Proc - IEEE Symp Comput Commun 2020;2020-July.
chaos. IEEE Trans Ind Electron 2019;66(11):8751–9. https://doi.org/10.1109/
10.1109/ISCC50000.2020.9219689.
TIE.4110.1109/TIE.2018.2886766.
[31] Johry A, Kapoor M. Unmanned aerial vehicle (UAV): fault tolerant design. Int J
[6] Veras FC, Lima TLV, Souza JS, Ramos JGGS, Lima Filho AC, Brito AV. Eccentricity
Eng Technol Sci Res IJETSR WwwIjetsrCom ISSN 2016;3:2394–3386.
failure detection of brushless DC motors from sound signals based on density
[32] Awadallah MA, Morcos MM, Gopalakrishnan S, Nehl TW. A neuro-fuzzy
of maxima. IEEE Access 2019;7:150318–26. https://doi.org/10.1109/
approach to automatic diagnosis and location of stator inter-turn faults in CSI-
Access.628763910.1109/ACCESS.2019.2946502.
fed PM brushless DC motors. IEEE Trans Energy Convers 2005;20(2):253–9.
[7] Iannace G, Ciaburro G, Trematerra A. Fault diagnosis for UAV blades using
https://doi.org/10.1109/TEC.2005.847976.
artificial neural network. Robotics 2019;8(3):59. https://doi.org/10.3390/
[33] Sadeghzadeh I, Zhang Y. A review on fault-tolerant control for unmanned
robotics8030059.
aerial vehicles (UAVs). AIAA Infotech Aerosp Conf Exhib 2011;2011:1–12.
[8] Rangel-Magdaleno JDJ, Urena-Urena J, Hernandez A, Perez-Rubio C. Detection
https://doi.org/10.2514/6.2011-1472.
of unbalanced blade on UAV by means of audio signal. 2018 IEEE Int Autumn

13
A. Altinors, F. Yol and O. Yaman Applied Acoustics 183 (2021) 108325

[34] Yaman Orhan, Tuncer Turker, Tasar Beyda. DES-Pat: a novel DES pattern-based [36] Park J-H, Jun C-Y, Jeong J-Y, Chang DE. Real-time quadrotor actuator fault
propeller recognition method using underwater acoustical sounds. Appl detection and isolation using multivariate statistical analysis techniques with
Acoust 2021;175:107859. https://doi.org/10.1016/j.apacoust.2020.107859. sensor measurements 2020:33–7. 10.23919/iccas50221.2020.9268391.
[35] Yaman Orhan, Ertam Fatih, Tuncer Turker. Automated Parkinson’s disease
recognition based on statistical pooling method using acoustic features. Med
Hypotheses 2020;135:109483. https://doi.org/10.1016/j.mehy.2019.109483.

14