Content uploaded by Adam Glowacz
Author content
All content in this area was uploaded by Adam Glowacz on Jan 24, 2020
Content may be subject to copyright.
Content uploaded by Adam Glowacz
Author content
All content in this area was uploaded by Adam Glowacz on Jan 24, 2020
Content may be subject to copyright.
A. Glowacz. Acoustic fault analysis of three commutator motors, Mechanical Systems and 1 Signal Processing, vol. 133, 2019. https://doi.org/10.1016/j.ymssp.2019.07.007 2
3
Acoustic fault analysis of three commutator motors
4
5
Adam Glowacz
1,
*
6
1
AGH University of Science and Technology, Faculty of Electrical Engineering, Automatics, Computer
7
Science and Biomedical Engineering, Department of Automatic Control and Robotics, al. A. Mickiewicza 30,
8
30-059 Kraków, Poland; adglow@agh.edu.pl;
9
10
* Correspondence: adglow@agh.edu.pl
11
12
Abstract: Electrical motors are used in industry and in home appliances. They are important for
13
people. Electrical motors convert electric energy into mechanical work. They generate acoustic
14
signals. In this research the author uses signal processing methods and acoustic signals of two
15
commutator motors of electric coffee grinders and one motor of an electric impact drill. A technique
16
of fault detection of mechanical faults of three commutator motors is presented. Following acoustic
17
signals of the first electric coffee grinder are measured and analysed: healthy, with one missing
18
screw, with a rear faulty sliding bearing and faulty shaft, with a burned out motor (motor off).
19
Following acoustic signals of the second electric coffee grinder are measured and analysed: healthy,
20
with a slightly damaged rear sliding bearing, with a moderately damaged rear sliding bearing,
21
motor off. Following acoustic signals are measured for the electric impact drill: healthy, slightly
22
damaged front bearing, moderately damaged front bearing, motor off. An analysis of acoustic signal
23
was carried out using the developed MSAF-RATIO-24-MULTIEXPANDED-FILTER-8 (Method of
24
Selection of Amplitudes of Frequency Ratio of 24 % Multiexpanded Filter 8 Hz) and k-means
25
clustering. The obtained results are very good. A total efficiency of recognition is in the range of
26
95–96 %.
27
28
Keywords: mechanical fault, bearing, drill, grinder, k-means clustering, motor, acoustic, sound,
29
damaged, diagnosis
30
31
1. Introduction
32
A growing number of electrical motors is a fact. A reason of that is increasing world
33
consumption of electricity. In the years 2002-2017 world consumption of electricity increased 50 %
34
[1]. For example, India consumed 406 TWh in 2002. India consumed 1156 TWh in 2017 [1]. The
35
consumption growth is also associated with the development of industry. Car, mining, fuel, military
36
industry need various devices consisting of motors.
37
Faults of electrical motors can be different: installation errors, manufacturing defects, improper
38
design, damaged insulation, faulty shafts, bearings, gears. Accordingly, it is necessary to develop
39
more reliable diagnostic methods for fault detection. The fault detection of rotating machinery is
40
significant for two reasons. The first reason is to evaluate condition of mechanical parts of the motor,
41
for example: bearings, gears, stator, rotor, shaft, pulley etc. The second reason is to repair the motor,
42
2 of 23
if the fault occurs. In this way, fault detection is useful to avoid unexpected failures. Industrial plants
43
work without unexpected shutdowns and it is important that motors operate efficiently. Developing
44
of new acoustic based fault diagnosis methods is motivation of research. It is also important for
45
industry. The author was also motivated by the literature (please see Section 3).
46
This paper presents the technique of fault detection of mechanical faults of three commutator
47
motors. Two electric coffee grinders and one motor of an electric impact drill are considered. New
48
method of feature extraction MSAF-RATIO-24-MULTIEXPANDED-FILTER-8 is developed.
49
Comparing to other feature extraction methods (please see Section 3), the proposed method was
50
tested for larger data sets – 12 classes. It has also good higher recognition rate. Analysis of acoustic
51
signals was good for 4 classes .
52
In this paper obtained results are computed for three motors. The proposed method allow us to
53
compare many different signals. Application of common frequency bandwidths is good for proper
54
recognition. Bearings and other mechanical faults can be detected using the developed technique. It
55
can find application for power tools, electrical devices, motors and engines.
56
Structure of manuscript is following: 1) Introduction, 2) Considered mechanical faults,
57
3) Theoretical background, 4) Proposed methodology and experimental setup, 5) Results, 6)
58
Discussion, 7) Summary and Conclusions.
59
60
2. Considered mechanical faults
61
Coffee grinders (CG) and electric impact drills (EID) use commutator motors (CM).
62
Commutator motor is often used at home and in the industry. Analysed commutator motors
63
generate different sounds depending on their conditions. Damage of the motor can happen
64
anywhere, for example, the electric drill fell on the floor. Damage can be also caused by operation of
65
the motor.
66
In this work, following acoustic signals of the CG1 (Coffee Grinder 1 – SCG 1050WH, analysed
67
motor HC5420) are measured and analysed: healthy CG1 (Figure 1), CG1 with one missing screw
68
(Figure 2), CG1 with a rear faulty sliding bearing and faulty shaft (Figure 3), CG1 with a burned out
69
motor (motor off, Figure 4).
70
71
Figure 1. Healthy CG1 (Coffee Grinder 1)
72
73
3 of 23
Figure 2. CG1 with one missing screw (missing screw is indicated by pink diamond, slightly
74
unscrewed screw is indicated by green diamond.)
75
76
Figure 3 CG1 with a rear faulty sliding bearing and faulty shaft (indicated by red diamond)
77
78
Figure 4 CG1 with a burned out motor (burned out motor and burned electrical wiring are indicated
79
by green diamond)
80
81
Following acoustic signals of the CG2 (Coffee Grinder 2 – ME-1498, analysed motor FY5420) are
82
measured and analysed: healthy CG2 (Figure 5), CG2 with a slightly damaged rear sliding bearing
83
(Figure 6), CG2 with a moderately damaged rear sliding bearing (Figure 7), motor off (Figure 8).
84
85
4 of 23
Figure 5. Healthy CG2
86
87
Figure 6. CG2 with a slightly damaged rear sliding bearing (indicated by red diamond)
88
89
Figure 7 CG2 with a moderately damaged rear sliding bearing (indicated by red diamond)
90
91
Figure 8 Motor off
92
5 of 23
93
Following acoustic signals of the EID (Electric Impact Drill – 50G515) are measured and analysed:
94
healthy (Figure 9), slightly damaged front bearing (Figure 10), moderately damaged front bearing
95
(Figure 11), motor off (Figure 12).
96
97
Figure 9. Healthy front bearing of the EID
98
99
Figure 10. Slightly damaged front bearing of the EID (indicated by red diamond)
100
101
Figure 11. Moderately damaged front bearing of the EID (indicated by red diamond)
102
103
6 of 23
Figure 12. Healthy EID (motor off)
104
105
3. Theoretical background
106
107
Current based fault diagnosis is used for electrical motors [2-6]. It can not be used, if we do not
108
have electrical current signal. An ammeter with a measuring card are used for measurement of
109
electrical current. The methods based on the current analysis is often used for faults of rolling
110
element bearings [2, 3]. A fault diagnosis approach is developed using current signals and MCSA
111
(Motor Current Signature Analysis) [4]. The MSCA is able to detect the different types of gear faults.
112
The MCSA is also used to detect faults of induction motors [4, 5]. The current based methods are
113
effective, because the electrical current signal is less noisy than vibration and acoustic signal. The
114
recognition rate of electrical signals is high. Electrical current signal is not difficult to process. In the
115
literature [3] electrical and mechanical signals are analysed for wind turbine of the induction
116
generator. The fault diagnosis technique for both vibration and current signals is presented.
117
Moreover, the development of the correlation between the current components and torque
118
disturbances is shown [3]. The article [6] describes the approach of fault location of broken rotor bars
119
of induction motor. The proposed approach is based on stator current and Hilbert Transform [6].
120
The fault diagnosis from vibration signals is used for rotating machinery [2, 5, 7, 8]. It is the
121
most common technique for the fault diagnosis of rotating machinery. Many vibration based
122
techniques are presented in the literature [2]. Vibration analysis is often used for bearing,
123
misalignment, unbalance and crack of gear tooth diagnosis [2, 5]. It can detect faulty state
124
(mechanical or electrical). However, location of the fault can be difficult. The vibration based fault
125
diagnosis is also affected by a mounting position of measuring device (accelerometer). It requires
126
prior knowledge about this mounting position. Reciprocating compressor system of model WH64
127
and rolling bearings are analysed using vibration signals [7]. A method of feature extraction ISAX
128
(Improved Symbolic Aggregate ApproXimation) is developed. The experiment results of
129
recognition for ISAX method are in the range of 80.40–95.60 % [7]. Next paper [8] presents a bearing
130
fault diagnosis method FF-FC-MIC (Feature-to-Feature and Feature-to-Category- Maximum
131
Information Coefficient). It can reach diagnosis accuracy in the range of 97.50–98.75 % for CWRU
132
dataset (Bearing Data Center of Case Western Reserve University) and 91.75–99.07 % for CUT-2
133
dataset [8].
134
Alternatively, the acoustic based fault diagnosis can be used [9–19]. The microphone can be
135
installed instead of the accelerometer. The microphone can capture more information (frequency
136
bandwidth) than the accelerometer. Ultrasounds and acoustic signals (20–20,000 Hz) are often used
137
in industry. The acoustic signals also allow us to detect faults of rotating machinery [9–10].
138
However, acoustic signal can be distracted by ambient noise. It is very important, because the
139
7 of 23
acoustic signal is more difficult to recognise. The paper [9], describes the acoustic based fault
140
detection of broken rotor bars and defects in bearings of induction motors. Bearings fault diagnosis
141
using sound signal is also presented [10]. The article [11] describes the method of fault detection of
142
the rotating machine using the acoustic and vibration signal. The correlation information between
143
acoustic and vibration signal of the rotating machine is described [11]. The article [12] presents a
144
system using a cellphone with a camera and a microphone. It uses video clip and IFR
145
(Instantaneous Frequency of Rotation). It also uses acoustic signal and IFCF (Instantaneous Fault
146
Characteristic Frequency) to diagnose bearing faults: outer raceway and inner raceway. The success
147
rate is 100 % for all analysed states and distance 0.2–0.3 m [12]. Compound fault prediction of rolling
148
bearing is preseneted in the paper [13]. Acoustic and acceleration sensor to capture compound
149
signal. Achieved an accuracy of recognition is equal to 94 % for bearing faults. The proposed method
150
is based on CNN (Convolution Neural Network), CMF (Combined Mode Function) and EEMD
151
(Ensemble Empirical Mode Decomposition) [13]. The paper [14] presents technique for prediction of
152
the torque. It uses the acoustic signals of induction machine and dyadic wavelet transform. The
153
success rate of proposed approach is 90–95 % [14]. In the paper [15], a fault diagnosis method of
154
gearboxes is presented. The proposed approach is based on acoustic signals and a two layer
155
scattering transform. An average accuracy of 97 % is achieved for four gearbox faults [15]. Next
156
paper [16] describes a review of the developments on mill load soft measuring techniques. The
157
measuring techniques are based on acoustic and vibration signals [16]. The study [17] presents fault
158
detection of induction machine using smartphone and recorded audible noise. The approach is
159
compared with vibration analysis. The fault detection using neural network is equal to 96.4 %.
160
Microphones of smartphones are able to capture good quality sound in the near distance from the
161
rotating machinery. The acoustic signals of the machine are interfered by ambient noise in the far
162
distance from the rotating machinery. The acoustic signals have lower faulty amplitude ratio than
163
vibration signal. The accelerometer is fastened to the frame of the motor. Whereas, the acoustic
164
signals are measured at a certain distance from the motor [17]. In the article [18], the authors propose
165
fault diagnosis technique of gear based on acoustic signals. The CNN is used for two datasets. The
166
accuracy of proposed approach is in the range of 96.5–98.5 % [18]. The fault diagnosis method using
167
acoustic signals of rolling bearing is presented [18]. The method uses PCA–GCC–SVM (Principal
168
Component Analysis – General Cross-Correlation – Support Vector Machine). The computed
169
recognition ratio is equal to 91.7 % [19].
170
4. Proposed methodology and experimental setup
171
The approach based on acoustic signals of commutator motors uses PC or notebook. The Tracer
172
KTM 43948 microphone is used, to capture an acoustic signal. The microphone has following
173
parameters: frequency response 30–16000 Hz, sensitivity 58 dB +/-3 dB. The microphone is located
174
0.3 m from the CG and EID. Friction of mechanical parts of the CG, EID and rotation of the
175
commutator motor (CM) generate acoustic signals. The captured acoustic signal has sampling rate
176
equal to 44.1 kHz and single channel. Next the acoustic data of the CG/EID are split into 1-s samples.
177
The obtained samples are normalized in range of <-1, 1>. The amplitude normalization makes it
178
possible to compare two signals, if the microphones are not at the same distance. The next step of
179
processing is feature extraction using the MSAF-RATIO-24-MULTIEXPANDED-FILTER-8. The
180
computed feature vectors are classified using k-means clustering. The proposed approach needs
181
historical data for training. The author used two soundtracks. The first of them is used for training.
182
The second is used for testing. The approach based on acoustic signals of the CM is depicted in
183
Figure 13. Notebook, microphone and electric coffee grinder are shown in Figure 14a. An
184
experimental setup consists of the mentioned devices (Figure 14b).
185
8 of 23
186
Figure 13. Approach based on acoustic signals of the CM
187
188
(a)
9 of 23
(b)
Figure 14. a) Notebook, microphone and electric coffee grinder 1, b) Diagram of the proposed
189
experimental setup
190
4.1. MSAF-RATIO-24-MULTIEXPANDED-FILTER-8
191
Feature vectors are obtained using the MSAF-RATIO-24-MULTIEXPANDED-FILTER-8
192
(Method of Selection of Amplitudes of Frequency Ratio of 24 % Multiexpanded Filter 8 Hz). This
193
method uses FFT (Fast Fourier transform) spectra of acoustic signals. Next it processes FFT spectra
194
in order to compute the best feature vectors. The developed method has 8 steps:
195
1. Compute FFT spectra of all analysed acoustic signals of the commutator motor (CM). Each FFT
196
spectrum consists of 16384-elements (frequency components). It is a vector. The frequency
197
component corresponds to 1.3458251953125 Hz (16,384*1.3458251953125 = 22050). Let's consider
198
acoustic signals of the coffeee grinder 1. The analysed classes of the CG1 are represented by the
199
vectors of FFT spectrum: healthy CG1 − hcg=[hcg
1
, hcg
2
, ..., hcg
16,384
], CG1 with one missing screw
200
− mscg=[mscg
1
, mscg
2
, ..., mscg
16,384
], CG1 with a rear faulty sliding bearing and faulty shaft −
201
fbcg=[fbcg
1
, fbcg
2
, ..., fbcg
16,384
].
202
2. For training phase, compute following differences: hcg - fbcg, hcg - mscg, fbcg - mscg.
203
3. Compute: |hcg - fbcg|, |hcg - mscg|, |fbcg - mscg|.
204
4. Select the best differences using a parameter Ratio. The parameter Ratio is defined as:
205
206
Ratio=(100%)VFC
i
/MaxVFC (1)
207
where MaxVFC
−
maximum absolute value of difference (for example max|hcg - fbcg| can be
208
max|hcg
21
- fbcg
21
|= 0.05), VFC
i
−
absolute value of difference between frequency components
209
with index i (for example i=32, then |hcg
32
- fbcg
32
| = 0.01), Ratio
−
threshold (set experimentally,
210
for example 0.01/0.05 = 20 %, then |hcg
32
- fbcg
32
| is selected). If value of difference of frequency
211
components (with index i) is larger than Ratio then frequency component (with index i) is
212
10 of 23
analysed in the next step. The found frequency components (with index i) are called Common
213
Frequency Components (CFCs).
214
5. Select CFCs. If CFCs can not be found, then use a parameter TCFCs (Threshold of CFCs). The
215
parameter TCFCs is expressed by equation (2):
216
217
sdifferenceanalysedofNumber CFCsrequiredofNumber
TCFCs =
(2)
218
Let's consider EXAMPLE 1 of using the parameter TCFCs. Five training sets are analysed. Each
219
training sets has 3 samples of acoustic signals (hcg, mscg, fbcg). There are 15 differences (3
220
difference for each training set). Frequency components 400 and 500 Hz are found 5 times for
221
|hcg - fbcg|. Frequency components 410 and 510 Hz are found 4 times for |hcg - mscg|.
222
Frequency components 420 and 520 Hz are found 4 times for |fbcg - mscg|. CFCs can not be
223
found. Frequency components 400, 410, 420, 500, 510, 520 Hz seem to be good for recognition.
224
They are selected, if we set the parameter TCFCs = 0.2666 (4/15). If we set the parameter TCFCs =
225
0.3333 (5/15), then frequency components 400 and 500 Hz will be selected.
226
6. Form groups of frequency. It can be noticed that the best frequency components are 400, 410,
227
420, 500, 510, 520 Hz. The MSAF-RATIO-24-MULTIEXPANDED-FILTER-8 method computed 2
228
groups <400 Hz, 410 Hz, 420 Hz> and <500 Hz, 510 Hz, 520 Hz>.
229
7. Form bandwidths of frequency. The MSAF-RATIO-24-MULTIEXPANDED-FILTER-8 method
230
uses 8 Hz bandwidths. Based on EXAMPLE 1, the method selects 6 bandwidths. The middle of
231
the first bandwidth is equal to 400 Hz. Next middles are at 410, 420, 500, 510, 520 Hz. Finally the
232
method selects bandwidths: <396–404 Hz>, <406–414 Hz>, <416–424 Hz>, <496–504 Hz>,
233
<506–514 Hz>, <516–524 Hz>.
234
8. Based on selected bandwidths, form a feature vector.
235
Figure 15. Block diagram of the MSAF-RATIO-24-MULTIEXPANDED-FILTER-8 method. FFT: Fast
236
Fourier Transform; TCFCs: (Threshold of CFCs).
237
11 of 23
238
A block diagram of the MSAF-RATIO-24-MULTIEXPANDED-FILTER-8 method is depicted in
239
Figure 15. Differences |hcg - fbcg|, |hcg - mscg|, |fbcg - mscg| are depicted in Figures 16-18.
240
241
Figure 16. Difference (|hcg - fbcg|)
242
243
Figure 17. Difference (|hcg - mscg|)
244
245
12 of 23
Figure 18. Difference (|fbcg - mscg |)
246
247
Figures 16-18 present peaks of differences of acoustic signals. However, it can be noticed that some
248
highest peaks are not selected. The proposed method selects bandwidths of frequency for all
249
analysed acoustic signals (training samples).
250
Based on acoustic signals of the CG1, the MSAF-RATIO-24-MULTIEXPANDED-FILTER-8 method
251
(TCFCs=0.25) selects two frequency components: 51, 52 Hz. Based on computed frequency
252
components, following frequency bandwidth is selected: <47–56 Hz>. The frequency bandwidth
253
<47–56 Hz> is selected using frequency components 51, 52 Hz. The first frequency bandwidth is <51 -
254
4, 51 + 4 Hz>. The second frequency bandwidth is <52 - 4, 52 + 4 Hz>. A common frequency
255
bandwidth is <47–56 Hz>.
256
The common frequency bandwidths (8 features, <47–56 Hz>) of the CG1 are shown in Figures
257
(19–21).
258
259
13 of 23
Figure 19. Values of features of healthy CG1 (8 features, frequency bandwidth <47–56 Hz>)
260
261
Figure 20. Values of features of CG1 with one missing screw (8 features, frequency bandwidth
262
<47–56 Hz>)
263
264
14 of 23
Figure 21. Values of features of CG1 with a rear faulty sliding bearing and faulty shaft
265
(8 features, frequency bandwidth <47–56 Hz>)
266
267
Based on acoustic signals of the CG2, the MSAF-RATIO-24-MULTIEXPANDED-FILTER-8 method
268
(TCFCs=0.4166) selects five frequency components:: 522, 537, 538, 539, 1610 Hz. Based on the
269
computed frequency components, following frequency bandwidths are selected: <518–526 Hz>,
270
<533–543 Hz>, <1606–1614 Hz>. The common frequency bandwidths (23 features) of the CG2 are
271
shown in Figures 22–24.
272
273
Figure 22. Values of features of healthy CG2 (23 features, frequency bandwidths <518–526 Hz>,
274
<533–543 Hz>, <1606–1614 Hz>)
275
276
15 of 23
Figure 23. Values of features of CG2 with a slightly damaged rear sliding bearing (23 features,
277
frequency bandwidths <518–526 Hz>, <533–543 Hz>, <1606–1614 Hz>
278
279
Figure 24. Values of features of CG2 with a moderately damaged rear sliding bearing (23 features,
280
frequency bandwidths <518–526 Hz>, <533–543 Hz>, <1606–1614 Hz>
281
282
Based on acoustic signals of the EID, the MSAF-RATIO-24-MULTIEXPANDED-FILTER-8 method
283
(TCFCs=0.25) selects four frequency components:: 410, 426, 479, 480 Hz. Based on the computed
284
frequency components, following frequency bandwidths are selected: <406–414 Hz>, <422–430 Hz>,
285
<475–484 Hz>. The common frequency bandwidths (22 features) of the EID are shown in Figures
286
25–27.
287
288
16 of 23
Figure 25. Values of features of the healthy EID (22 features, frequency bandwidths <406–414 Hz>,
289
<422–430 Hz>, <475–484 Hz>)
290
291
292
293
294
Figure 26. Values of features of the slightly damaged front bearing of the EID (22 features, frequency
295
bandwidths <406–414 Hz>, <422–430 Hz>, <475–484 Hz>)
296
297
17 of 23
Figure 27. Values of features of the moderately damaged front bearing of the EID (22 features,
298
frequency bandwidths <406–414 Hz>, <422–430 Hz>, <475–484 Hz>)
299
300
The common frequency bandwidths (8 features) of the CG1 form the first feature vector. The
301
common frequency bandwidths (23 features) of the CG2 form the second feature vector. The
302
common frequency bandwidths (22 features) of the EID form the third feature vector. The k-means
303
clustering is used to classify the computed vectors [20–23]. It can be noticed (Figures 19–27), that
304
features are selected quite well. Many classifier can be used for classification, for example: The
305
Nearest Neigbour classifier [24–26], Naive Bayes, neural networks [27–29], Self-organizing map [30],
306
Support Vector Machine [31, 32] etc. The k-means clustering is used, however mentioned classifiers
307
will be also proper for recognition.
308
309
310
4.2. K-means clustering
311
The k-means clustering is a type of unsupervised classifier. It uses unlabeled feature vectors for
312
training and testing. The method computes k clusters using training feature vectors. In the proposed
313
analysis k is equal to 4 (4 types of acoustic signals). The method has 3 steps for training phase:
314
1. Randomly initialize k cluster centers.
315
2. Move the cluster centers to the average of feature vectors of the specific cluster.
316
3. Repeat the step 2. If the value of cluster center is equal to the value of average of feature
317
vectors, stop computations.
318
319
Next the method uses the computed cluster centers and new unknown test feature vectors for
320
testing. The nearest distance between feature vectors (cluster centers and test feature vector) is
321
computed using distance metric. In this analysis, the Euclidean distance (3) is used. Other types of
322
metrics can be also used. They will provide similar recognition results.
323
324
∑
1
2
|)-(|)-(
n
iii
cuED
=
=c u
(3)
325
18 of 23
where ED(u-c) − the Euclidean distance, u − unknown test feature vector, c − cluster center, n=<1, ..., 8>
326
or n=<1, ..., 23> − index of feature vector.
327
328
The description of the k-means clustering is available in the literature [20–23].
329
330
5. Results
331
Acoustic analyses of the CG1 (coffee grinder 1), CG2 (coffee grinder 2) and EID (Electric Impact
332
Drill) were carried out for three commutator motors. The devices were powered from the mains 230
333
V/50 Hz. Acoustic signals were measured in the room 3 m x 3 m. Four acoustic signals were
334
measured for CG1: healthy, with one missing screw, with a rear faulty sliding bearing and faulty
335
shaft, with burned out motor (motor off). The model of the motor was HC5420 for the CG1. The
336
rated power of the CG1 was equal to 150 W. The rotor speed of the motor was equal to 11,300 rpm.
337
The rated current of the motor was equal to 0.41 A.
338
Four acoustic signals were measured for CG2: healthy, with a slightly damaged rear sliding
339
bearing, with a moderately damaged rear sliding bearing, motor off. The model of the motor was
340
FY5420 for CG2. The rated power of the CG2 was equal to 140 W. The rotor speed of the motor was
341
equal to 28,000-30,000 rpm.
342
Four acoustic signals were measured for the EID: healthy, slightly damaged front bearing,
343
moderately damaged front bearing, motor off. The motor of the EID had rated power equal to 500 W.
344
The rotor speed of the motor was equal to 3,000 rpm.
345
The researcher used 32 training samples of the CG1, 32 training samples of the CG2, and 32
346
training samples of the EID for training phase of acoustic analysis. Next the author used 200 test
347
samples of the CG1, CG2 and EID (600 test samples). The researcher used cross-validation.
348
Cross-validation was based on training and testing sets. It was used for prediction of unknown test
349
sample.
350
Following equation (4) was introduced, to evaluate the efficiency of recognition of the proposed
351
analysis of the CG − ER
CG
.
352
353
100% )( / )(
CGhALLCGhCG
NNER
−
=
(4)
354
where: N
CGh
– number of test feature vectors assigned to healthy CG (healthy CG is a class CGh),
355
N
ALL-CGh
– number of all test feature vectors of the healthy CG (CGh class), ER
CG
– efficiency of
356
recognition for healthy CG (class CGh).
357
Following equation (5) was introduced, to evaluate the total ER
CG
− TER
CG
(for 4 analysed
358
acoustic signals):
359
360
4/)(
CGbCGfCGmCGhCG
ERERERERTER +++=
(5)
361
where TER
CG
− total ER
CG
(for 4 signals of the CG), ER
CGh
− ER
CG
for the CGh class (healthy CG), ER
CGm
362
− ER
CG
for the CGm class (CG with one missing screw), ER
CGf
− ER
CG
for the CGf class (CG with a rear
363
faulty sliding bearing and faulty shaft), ER
CGb
− ER
CG
for the CGb class (CG with a burned out motor).
364
TER
CG1
, TER
CG2
and TER
EID
(TER of the EID) are computed similarly.
365
366
19 of 23
The obtained results of the acoustic analysis of the CG1 was presented in Table 1. The results were
367
computed using the MSAF-RATIO-24-MULTIEXPANDED-FILTER-8
and the k-means clustering
368
(Table 1).
369
370
Table 1. Results of the acoustic analysis of the CG1 using the MSAF-RATIO-24-MULTIEXPANDED-FILTER-8
371
and the k-means clustering
372
Type of acoustic signal ER
CG
[%]
Healthy CG1 100
CG1 with one missing screw 92
CG1 with a rear faulty sliding bearing and faulty shaft 88
CG1 with a burned out motor (motor off) 100
TER
CG1
[%]
TER
CG1
[%] 95
373
The obtained results were following: TER
CG1
= 95 % and ER
CG
in the range of 88
–
100 % (for the
374
CG1).
375
376
The obtained results of the acoustic analysis of the CG2 was presented in Table 2. The results were
377
computed using the MSAF-RATIO-24-MULTIEXPANDED-FILTER-8
and the k-means clustering
378
(Table 2).
379
380
Table 2. Results of the acoustic analysis of the CG2 using the MSAF-RATIO-24-MULTIEXPANDED-FILTER-8
381
and the k-means clustering
382
Type of acoustic signal ER
CG
[%]
Healthy CG2 100
CG2 with a slightly damaged rear sliding bearing 92
CG2 with a moderately damaged rear sliding bearing 92
Motor off 100
TER
CG2
[%]
TER
CG2
[%] 96
383
The obtained results were following: TER
CG2
= 96 % and ER
CG
in the range of 92
–
100 % (for the
384
CG2).
385
386
The obtained results of the acoustic analysis of the EID was presented in Table 3. The results were
387
computed using the MSAF-RATIO-24-MULTIEXPANDED-FILTER-8
and the k-means clustering
388
(Table 3).
389
390
Table 3. Results of the acoustic analysis of the EID using the MSAF-RATIO-24-MULTIEXPANDED-FILTER-8
391
and the k-means clustering
392
Type of acoustic signal ER
EID
[%]
Healthy 100
Slightly damaged front bearing 100
Moderately damaged front bearing 80
Motor off 100
TER
EID
[%]
TER
EID
[%] 95
393
The obtained results were following: TER
EID
= 95 % and ER
EID
in the range of 80
–
100 % (for the
394
EID
).
395
20 of 23
6. Discussion
396
Nowadays, innovative techniques of fault detection using acoustic signals are significant for the
397
industry and safety. Fault detection from acoustic data can be useful, for example places very
398
difficult to access. It can find applications for fault detection of industrial motors etc. It is not
399
expensive when compared to diagnostic technique based on thermal imaging. However, the acoustic
400
signals can be affected by refraction and reflection of other sounds.
401
In this research the author uses acoustic signals of three commutator motors and signal
402
processing methods. Twelve types of acoustic signals are analysed. The
403
MSAF-RATIO-24-MULTIEXPANDED- FILTER-8 method is presented in this paper.
404
Some problems were noticed when performing signal processing. The first of them was
405
selection of training samples. There are different ways, in which the acoustic signals can be captured.
406
Different types of microphones can be used for recording. Using different microphones causes
407
errors. The errors are caused by different parameters of microphones such as: impedance,
408
sensitivity, frequency response, directivity etc. Furthermore, it is necessary to use the same or similar
409
type of microphone for all measurements of acoustic signals. It is the best, if the microphones are the
410
same.
411
A database of acoustic data consisted of training and test samples. The next problem was
412
related with types of acoustic signals of the database. Let's suppose we would like to detect fault of
413
the commutator motor, but we have training samples of car and train. We do not have training
414
samples of the commutator motor. In this example the proposed approach uses k-means clustering
415
and selects the nearest cluster center. If we have training samples of car, train and an induction
416
motor. It will probably recognises the class "induction motor". It depends on analysed acoustic
417
signals. Therefore, it is necessary to have training samples of a similar motor in the training
418
database. This approach can be used for the production line of motors or other devices generating an
419
acoustic signal.
420
The normalization of amplitude was used for better recognition. However distance from the
421
microphone to machine is important. The distance (0.1–1 m) is good for recognition.
422
The proposed MSAF-RATIO-24-MULTIEXPANDED-FILTER-8 method has high recognition results:
423
TER
CG
is in the range of 95–96 %. The parameter TER
EID
is equal to 95 %. Compared to other acoustic
424
fault diagnosis methods, the proposed method has a bit better recognition results (see Section 3,
425
accuracy of other methods in the range of 90-95%).
426
The developed MSAF-RATIO-24-MULTIEXPANDED- FILTER-8 uses frequency bandwidths.
427
The proper frequency bandwidths can be selected using a lot of training samples. In some cases, a
428
spectral leakage may cause errors. Therefore, the computed frequency bandwidths are better for
429
recognition. It allows to select better features.
430
In this paper, the author developed the original method of feature extraction –
431
MSAF-RATIO-24-MULTIEXPANDED-FILTER-8. The Tracer KTM 43948 microphone is used for
432
measurements. Analysis of three types of commutator motors was carried out – Electric Impact Drill
433
– 50G515, ME-1498, analysed motor FY5420, SCG 1050WH, analysed motor HC5420. Twelve
434
acoustic signals of commutator motors were analysed.
435
436
7. Summary and Conclusions
437
438
439
440
21 of 23
Advanced intelligent systems are now being used to enhance the safety of electrical motors. The
441
presented paper develops useful diagnostic technique. The original method of feature extraction
442
MSAF-RATIO-24-MULTIEXPANDED-FILTER-8 is proposed. In this research the author uses
443
acoustic signals of three commutator motors.
444
The acoustic analysis was carried out using the
445
MSAF-RATIO-24-MULTIEXPANDED-FILTER-8 and k-means clustering. The MSAF-RATIO-24-
446
MULTIEXPANDED-FILTER-8 method extracted features from acoustic signal. The
447
MSAF-RATIO-24- MULTIEXPANDED-FILTER-8 method is developed and presented in this paper.
448
The obtained experimental results are very good. The parameter TER
CG
is in the range of 95–96 %.
449
The parameter TER
EID
is equal to 95 %. The experimental results show that:
450
1) The proposed technique is useful for diagnosis of mechanical faults such as bearings faults.
451
2) Other mechanical faults can be also detected by the proposed technique, if there are good
452
training data sets (recorded by the same microphone, at the same distance from microphone,
453
the same environmental noises etc.).
454
3) The proposed technique was analysed for different commutator motors.
455
4) The proposed acoustic based technique can be useful for power tools, electrical devices,
456
motors and engines.
457
5) The results of recognition depends on training data sets.
458
This approach provides means by which acoustic signals can be used to advantage in fault
459
detection. Fault detection from acoustic data can be useful, for example places very difficult to
460
access. It can find applications for fault detection of industrial motors etc. It is not expensive when
461
compared to diagnostic technique based on thermal imaging.
462
It was shown that the proposed approach is useful for diagnosis. The developed approach
463
depends on acoustic data. This is advantage or disadvantage. It is possible to recognise acoustic
464
signals properly for specific conditions, for example distance from the microphone to machine (0.1–2
465
m). The approach can be also used for noisy room with 10 motors. However, faults and training
466
samples of the analysed motor should be prepared each time. It is based on pattern recognition. It is
467
also good idea to add noisy components (noisy sample of acoustic signal) in training sets. Of course
468
there are also disadvantages. Acoustic signal is more noisy than other diagnostic signals such as:
469
electric current and vibration. There is a need of large data set of training samples.
470
The future acoustic analysis can be based on new signal processing methods. Furthermore, the
471
analysis will be extended by new measurements for example acoustic camera. It can be also
472
extended by adding vibration signals or ultrasounds. Cases of different motors, operating conditions
473
and faults should be also analysed.
474
475
476
Funding: This research was funded by the AGH University of Science and Technology, grant No.
477
16.16.120.773
.
478
Acknowledgments: This work has been supported by AGH University of Science and Technology, grant no.
479
11.11.120.714. The author thanks unknown reviewers for the valuable suggestions.
480
Conflicts of Interest: The author declares no conflict of interest.
481
482
References
483
484
1. World Energy Statistics, Enerdata, 2018, https://yearbook.enerdata.net
485
22 of 23
2. Immovilli F., Bellini A., Rubini R., Tassoni C., Diagnosis of Bearing Faults in Induction Machines by
486
Vibration or Current Signals: A Critical Comparison, IEEE Transactions on Industry Applications 2010, 46 (4),
487
1350-1359. DOI: 10.1109/TIA.2010.2049623
488
3. Zappala D., Sarma N., Djurovic S., Crabtree CJ., Mohammad A., Tavner PJ., Electrical & mechanical
489
diagnostic indicators of wind turbine induction generator rotor faults, Renewable energy 2019, 131, 14-24.
490
DOI: 10.1016/j.renene.2018.06.098
491
4. Aouabdi S., Taibi M., Bouras S., Boutasseta N., Using multi-scale entropy and principal component
492
analysis to monitor gears degradation via the motor current signature analysis, Mechanical Systems and
493
Signal Processing 2017, 90, 298-316. DOI: 10.1016/j.ymssp.2016.12.027
494
5. Yao Y., Li YS., Yin Q., A novel method based on self-sensing motor drive system for misalignment
495
detection, Mechanical Systems and Signal Processing 2019, 116, 217-229. DOI: 10.1016/j.ymssp.2018.06.030
496
6. Abd-el-Malek MB., Abdelsalam AK., Hassan OE., Novel approach using Hilbert Transform for multiple
497
broken rotor bars fault location detection for three phase induction motor, ISA Transactions 2018, 80,
498
439-457. DOI: 10.1016/j.isatra.2018.07.020
499
7. Zhang YL., Duan LX., Duan MH., A new feature extraction approach using improved symbolic aggregate
500
approximation for machinery intelligent diagnosis, Measurement 2019, 133, 468-478. DOI:
501
10.1016/j.measurement.2018.10.045
502
8. Tang XH., Wang JC., Lu JG., Liu GK., Chen JD., Improving Bearing Fault Diagnosis Using Maximum
503
Information Coefficient Based Feature Selection, Applied Sciences-Basel 2018, 8 (11), Article Number: 2143.
504
DOI: 10.3390/app8112143
505
9. Delgado-Arredondo PA., Morinigo-Sotelo D., Osornio-Rios RA., Avina-Cervantes JG., Rostro-Gonzalez
506
H., Romero-Troncoso RD., Methodology for fault detection in induction motors via sound and vibration
507
signals, Mechanical Systems and Signal Processing 2017, 83, 568-589. DOI: 10.1016/j.ymssp.2016.06.032
508
10. Kumar H., Sugumaran V., Amarnath M., Fault Diagnosis of Bearings through Sound Signal Using
509
Statistical Features and Bayes Classifier, Journal of Vibration Engineering & Technologies 2016, 4 (2), 87-96.
510
11. Orimoto H., Ikuta A., Statistical faults diagnosis method by Using Higher-order correlation information
511
between sound and vibration of rotational machine, Proceedings of the 23rd International Congress on Sound
512
and Vibration: From Ancient to Modern Acoustics, Book Series: Proceedings of the International Congress on Sound
513
and Vibration 2016.
514
12. Lu SL., Wang XX., Liu F., He QB., Liu YB., Zhao JW., Fault Diagnosis of Motor Bearing by Analyzing a
515
Video Clip, Mathematical Problems in Engineering 2016, Article Number: 8139273. DOI:
516
10.1155/2016/8139273
517
13. Singh SK., Kumar S., Dwivedi JP., Compound fault prediction of rolling bearing using multimedia data,
518
Multimedia Tools and Applications 2017, 76 (18), 18771-18788. DOI: 10.1007/s11042-017-4419-1
519
14. Sangeetha P., Hemamalini S., Dyadic wavelet transform-based acoustic signal analysis for torque
520
prediction of a three-phase induction motor, IET Signal Processing 2017, 11 (5), 604-612. DOI:
521
10.1049/iet-spr.2016.0165
522
15. Heydarzadeh M., Nourani M., Hansen J., Kia SH., Non-invasive Gearbox Fault Diagnosis Using Scattering
523
Transform of Acoustic Emission, 2017 IEEE International Conference on Acoustics, Speech and Signal Processing
524
(ICASSP), Book Series: International Conference on Acoustics Speech and Signal Processing ICASSP 2017,
525
371-375.
526
16. Tang J., Qiao JF., Liu Z., Zhou XJ., Yu G., Zhao JJ., Mechanism characteristic analysis and soft measuring
527
method review for ball mill load based on mechanical vibration and acoustic signals in the grinding
528
process, Minerals Engineering 2018, 128, 294-311. DOI: 10.1016/j.mineng.2018.09.006
529
17. Vaimann T., Sobra J., Belahcen A., Rassolkin A., Rolak M., Kallaste A., Induction machine fault detection
530
using smartphone recorded audible noise, IET Science Measurement & Technology 2018, 12, 4, pp. 554-560.
531
DOI: 10.1049/iet-smt.2017.0104
532
18. Yao Y., Wang HL., Li SB., Liu ZH., Gui G., Dan YB., Hu JJ., End-To-End Convolutional Neural Network
533
Model for Gear Fault Diagnosis Based on Sound Signals, Applied Sciences-Basel 2018, 8 (9), Article Number:
534
1584, DOI: 10.3390/app8091584
535
19. Li HK., Luo Y., Huang J., Kanemoto T., Guo MY., Tang FL., New acoustic monitoring method using
536
cross-correlation of primary frequency spectrum, Journal of Ambient Intelligence and Humanized Computing
537
2013, 4 (3), 293-301. DOI: 10.1007/s12652-011-0105-8
538
23 of 23
20. Camarena-Martinez D., Valtierra-Rodriguez M., Amezquita-Sanchez JP., Granados-Lieberman D.,
539
Romero-Troncoso RJ., Garcia-Perez A., Shannon entropy and k-means method for automatic diagnosis of
540
broken rotor bars in induction motors using vibration signals, Shock and Vibration 2016, Article Number:
541
4860309. DOI: 10.1155/2016/4860309
542
21. Jiang ZN., Hu MH., Feng K., Wang H., A SVDD and K-Means Based Early Warning Method for
543
Dual-Rotor Equipment under Time-Varying Operating Conditions, Shock and Vibration 2018, Article
544
Number: 5382398, DOI: 10.1155/2018/5382398
545
22. Wang LM., Shao YM., Crack Fault Classification for Planetary Gearbox Based on Feature Selection
546
Technique and K-means Clustering Method, Chinese Journal of Mechanical Engineering 2018, 31 (1), Article
547
Number: 4. DOI: 10.1186/s10033-018-0202-0
548
23. Liu JW., Li Q., Chen WR., Cao TQ., A discrete hidden Markov model fault diagnosis strategy based on
549
K-means clustering dedicated to PEM fuel cell systems of tramways, International Journal of Hydrogen
550
Energy 2018, 43 (27), 12428-12441. DOI: 10.1016/j.ijhydene.2018.04.163
551
24. Gou JP., Ma HX., Ou WH., Zeng SN., Rao YB., Yang HB., A generalized mean distance-based k-nearest
552
neighbor classifier, Expert Systems With Applications 2019, 115, 356-372. DOI: 10.1016/j.eswa.2018.08.021
553
25. Zhang YQ., Cao G., Wang BS., Li XS., A novel ensemble method for k-nearest neighbor, Pattern Recognition
554
2019, 85, 13-25. DOI: 10.1016/j.patcog.2018.08.003
555
26. Bandaragoda TR., Ting KM., Albrecht D., Liu FT., Zhu, Y., Wells JR., Isolation-based anomaly detection
556
using nearest-neighbor ensembles. Computational Intelligence 2018, 34 (4), 968-998. DOI: 10.1111/coin.12156
557
27. Kochan O., Sapojnyk H., Kochan R. Temperature field control method based on neural network. 2013 IEEE
558
7th International Conference on Intelligent Data Acquisition and Advanced Computing Systems (IDAACS 2013),
559
2013. DOI: 10.1109/IDAACS.2013.6662632
560
28. Gajewski, J.; Valis, D. The determination of combustion engine condition and reliability using oil analysis
561
by MLP and RBF neural networks. Tribology International 2017, 115, 557-572. DOI:
562
10.1016/j.triboint.2017.06.032
563
29. Caesarendra W., Wijayaa T., Tjahjowidodob T., Pappachana B. K., Weec A., Izzat Roslan M., Adaptive
564
neuro-fuzzy inference system for deburring stage classification and prediction for indirect quality
565
monitoring, Applied Soft Computing 2018, 72, 565–578. DOI: 10.1016/j.asoc.2018.01.008
566
30. Prieto MD., Millan DZ., Chromatic Monitoring of Gear Mechanical Degradation Based on Acoustic
567
Emission, IEEE Transactions on Industrial Electronics 2017, 64 (11), 8707-8717. DOI: 10.1109/TIE.2017.2701761
568
31. Zhang C., Peng ZX., Chen S., Li ZX., Wang JG., A gearbox fault diagnosis method based on
569
frequency-modulated empirical mode decomposition and support vector machine, Proceedings of the
570
Institution of Mechanical Engineers Part C-Journal of Mechanical Engineering Science 2018, 232 (2), 369-380. DOI:
571
10.1177/0954406216677102
572
32. Pandiyan V., Caesarendra W., Tjahjowidodo T., Tan HH., In-process tool condition monitoring in
573
compliant abrasive belt grinding process using support vector machine and genetic algorithm, Journal of
574
Manufacturing Processes 2018, 31, 199-213. DOI: 10.1016/j.jmapro.2017.11.014
575
33. Lu SL., He QB., Zhang HB., Kong FR., Rotating machine fault diagnosis through enhanced stochastic
576
resonance by full-wave signal construction, Mechanical Systems and Signal Processing 2017, 85, 82-97. DOI:
577
10.1016/j.ymssp.2016.08.003
578
579