Acoustic based fault diagnosis of three-phase induction motor

Applied Acoustics 137 (2018) 82–89

Contents lists available at ScienceDirect

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

Acoustic based fault diagnosis of three-phase induction motor

T

Adam Glowacz AGH University of Science and Technology, Faculty of Electrical Engineering, Automatics, Computer Science and Biomedical Engineering, Department of Automatic Control and Robotics, Al. A. Mickiewicza 30, 30-059 Kraków, Poland

A R T I C LE I N FO

A B S T R A C T

Keywords: Acoustic signal Induction motor Fault Diagnosis Machine Nearest neighbour

The article describes acoustic based fault diagnosis techniques of a three-phase induction motor. Four real states of the three-phase induction motor were analysed: healthy three-phase induction motor, three-phase induction motor with broken rotor bar, three-phase induction motor with 2 broken rotor bars, three-phase induction motor with faulty ring of squirrel-cage. Two feature extraction methods of acoustic signals of the induction motor SMOFS-32-MULTIEXPANDED-2-GROUPS (Shortened Method of Frequencies Selection Multiexpanded 2 Groups) and SMOFS-32-MULTIEXPANDED-1-GROUP were described. The Nearest Neighbour classifier, backpropagation neural network and modified classifier based on words coding were used for recognition of acoustic signals. Results of recognition were very good for the real data and developed fault diagnosis techniques based on acoustic signals. The described fault diagnosis approach can find applications in the industry.

1. Introduction Electrical motors play an important role in industry and home appliances. They are used in many industrial plants such as: refineries, mines, factories, and ironworks. High reliability and cost reductions are essential for mentioned industrial plants. Each year the number of electrical motors is increased. Induction motor is a widely used electrical motor. Induction motors are inexpensive and have high reliability. They are also easy to maintain. Any electrical motor failure causes loss of production. It may also cause permanent failure of electrical motor. Operators of electrical motor can prevent unexpected failure if they use early fault diagnosis system [1]. It is a reason to develop such fault diagnosis system. Diagnosis of electrical motors is discussed in many scientific articles [2,3]. Mechanical faults (mechanical unbalance, bearing failures, shaft misalignment, air-gap deformation) and electrical faults (rotor and stator faults) of induction motors were analysed [4–8]. Vibration, electrical and thermal analyses were used for fault diagnosis of electrical machines [9–19]. Acoustic signals were also analysed [20–28]. Techniques based on thermal and acoustic signals are called non-invasive fault diagnosis techniques (we do not need to connect anything). Techniques based on electrical signals are usually invasive fault diagnosis techniques. Each of technique has advantages and disadvantages. Advantages of analysis of vibration signals are: low cost and instant measurement of vibrations [29–39]. Vibration signals can be used for mechanical and electrical faults of the motor. The problem for vibration analysis is set of accelerometer/data logger. It is required

to put the device in the same way. Moreover the location of the fault cannot be exactly determined. In the literature thermal imaging was used for fault diagnosis of induction motors [40–42]. Advantages of analysis of thermal images are: non-invasive measurement and proper detection of the location of the fault. However this technique has some disadvantages: high cost as well as time to heat up motor, slower processing of the data (images). It is also required to put the thermal camera in the same way and use proper methods of image processing [40–43]. In the literature fault diagnosis techniques based on the analysis of electric currents were described [2–6]. Advantages of analysis of electrical current are: low cost, high recognition efficiency, signal is not mixed. However this technique has some disadvantages: invasive technique, limited faulty states - electrical faults. Descriptions of acoustic based fault diagnosis techniques are also available in the literature [44–50]. Advantages of acoustic based fault diagnosis are: non-invasive technique, low cost and instant measurement of acoustic signals. Acoustic signals can be used for mechanical and electrical faults of the motor. The problem for acoustic analysis is set of microphone. It is required to put the device in the same way. Moreover the location of the fault cannot be exactly determined (similar to vibration signals). Acoustic signals are mixed by other signals (reflected waves). They are mixed more than vibration and electrical current signals. In this paper acoustic based fault diagnosis technique of a threephase induction motor was described. Four states of the three-phase induction motor were analysed (Fig. 1): healthy three-phase induction

E-mail address: [email protected]. https://doi.org/10.1016/j.apacoust.2018.03.010 Received 1 February 2018; Received in revised form 27 February 2018; Accepted 9 March 2018 0003-682X/ © 2018 Elsevier Ltd. All rights reserved.

Applied Acoustics 137 (2018) 82–89

A. Glowacz

Fig. 1. Experimental setup of analysis of acoustic signals of analysed three phase induction motors.

Fig. 2. (a) Rotor of the healthy three-phase induction motor. (b) Rotor of the three-phase induction motor with faulty ring of squirrel-cage.

Fig. 4. Proposed acoustic based fault diagnosis techniques of the three-phase induction motor using SMOFS-32-MULTIEXPANDED-2-GROUPS and SMOFS-32-MULTIEXPANDED-1-GROUP.

classification (Fig. 4). The first step of signal processing is recording of acoustic signal of the three-phase induction motor. For this purpose low-cost capacity microphone with computer or digital voice recorder can be used. Low-cost capacity microphone with computer can be used for online condition monitoring. It has frequency range 20–20,000 Hz. Dynamic microphone has frequency range 100–5000 Hz. Lower frequency (< 100 Hz) are essential for condition monitoring. The author analysed three-phase induction motor using digital voice recorder (format: WAVE, sampling frequency 44,100 Hz, 16 bits, mono channel, omnidirectional). The second step of signal processing is split of soundtrack into smaller audio files. Next the acoustic data were split into 1-second data files. After that signal was normalized in the range [−1, 1]. Normalized signals were processed by the Hamming window, the FFT method (16384 frequency components) and the SMOFS-32-MULTIEXPANDED-2-GROUPS, SMOFS32- MULTIEXPANDED-1-GROUP. Feature extraction methods computed training and test feature vectors (Fig. 4). The last step of signal processing was classification of feature vectors. Classification of feature vectors was based on the Nearest Neighbour classifier, backpropagation neural network and modified classifier based on words coding.

Fig. 3. (a) Rotor of the three-phase induction motor with broken rotor bar. (b) Rotor of the three-phase induction motor with 2 broken rotor bars.

motor (Fig. 2a), three-phase induction motor with faulty ring of squirrel-cage (Fig. 2b), three-phase induction motor with broken rotor bar (Fig. 3a), three-phase induction motor with 2 broken rotor bars (Fig. 3b). The proposed fault diagnosis techniques consist of signal processing methods: preprocessing, feature extraction, classification. Two feature extraction methods of acoustic signals - SMOFS-32-MULTIEXPANDED-2-GROUPS (Shortened Method of Frequencies Selection Multiexpanded 2 Groups) and SMOFS-32-MULTIEXPANDED-1-GROUP were developed. The classification step was performed using the Nearest Neighbour classifier, backpropagation neural network and modified classifier based on words coding. 2. Acoustic based fault diagnosis technique

2.1. Shortened Method of Frequencies Selection Multiexpanded 2 Groups and 1 Group

The proposed acoustic based fault diagnosis techniques consist of signal processing methods, such as: preprocessing, feature extraction,

The Shortened Method of Frequencies Selection Multiexpanded 2 83

Applied Acoustics 137 (2018) 82–89

A. Glowacz

Groups (SMOFS-32-MULTIEXPANDED-2-GROUPS) and the Shortened Method of Frequencies Selection Multiexpanded 1 Group (SMOFS32-MULTIEXPANDED-1-GROUP) used differences of spectra of acoustic signals of the three-phase induction motor. Acoustic signals and their spectra depend on motor size, construction, operational parameters, nominal power, nominal electric current, nominal voltage, rotor speed and the state of the motor (healthy three-phase induction motor, threephase induction motor with broken rotor bar, three-phase induction motor with 2 broken rotor bars, three-phase induction motor with faulty ring of squirrel-cage). The proposed methods SMOFS-32-MULTIEXPANDED-2-GROUPS and SMOFS-32-MULTIEXPANDED-1-GROUP had following steps of processing: (6) (1) Compute the frequency spectrum of an acoustic signal for each state of the three-phase induction motor. The frequency spectrum of an acoustic signal of the healthy three-phase induction motor was expressed as vector ahtim = [ahtim1, ahtim2, … , ahtim16384]. The frequency spectrum of an acoustic signal of the three-phase induction motor with broken rotor bar was denoted as vector atimbrb = [atimbrb1, atimbrb2, … , atimbrb16384]. The frequency spectrum of an acoustic signal of the three-phase induction motor with 2 broken rotor bars was defined as vector atimbrbs = [atimbrbs1, atimbrbs2, … , atimbrbs16384]. The frequency spectrum of an acoustic signal of the three-phase induction motor with faulty ring of squirrel-cage was expressed as vector atimfr = [atimfr1, atimfr2, … , atimfr16384]. (2) Compute differences of frequency spectra of acoustic signals of three-phase induction motors: ahtim − atimbrb, ahtim − atimbrbs, ahtim − atimfr, atimbrb − atimbrbs, atimbrb − atimfr, atimbrbs − atimfr. (3) Compute absolute values of obtained differences: |ahtim − atimbrb|, |ahtim − atimbrbs|, |ahtim − atimfr|, |atimbrb − atimbrbs|, |atimbrb − atimfr|, |atimbrbs − atimfr|. (4) Compute selected frequency components using formula (1). The computed frequency components are greater than an iterated threshold Thn. Moreover the number of iterations depends on acoustic signals and a parameter NFCn. (1)

||FSASX |−|FSASY || > Thn,

(7)

where Thn – threshold of selection after n iterations, ||FSASX| − |FSASY|| – difference between frequency spectrum of acoustic signal of state X of the motor and frequency spectrum of acoustic signal of state Y of the motor, FSASX - frequency spectrum of acoustic signal of state X of the motor, FSASY - frequency spectrum of acoustic signal of state Y of the motor. (5) Compute the threshold Thn for each iteration (the number of iterations n). The computed threshold Thn is denoted as (2): NFC

Thn =

∑NFCnn= 1 ||FSASX |−|FSASY ||

NFCn ⩽ 32,

NFCn

,

(2)

(3) (8) (9)

where NFCn is the number of frequency components after n iterations (initially NFC0 = 16384, the FFT method computes 16,384 frequency components for window length 32,768). If the value of NFCn is greater than 32, the SMOFS-32- carries out formula (2) (loop calculations). The SMOFS-32 stops its calculations, if the value of NFCn is less or equal to 32. The SMOFS-32 computed feature vector with 1-32 features (frequency components). The final value of the parameter NFCn depends on the analysed acoustic signals. Let's see CASE1 (example of using SMOFS-32 without MULTIEXPANDED extension). 3 acoustic signals of states X, Y and Z are considered. The SMOFS-32 selects frequency components 300, 320, 340, 360, 380 Hz for difference of acoustic signals of states X

and Y. The SMOFS-32 selects frequency components 310, 330, 350, 370, 390 Hz for difference of acoustic signals of states X and Z. The SMOFS-32 selects frequency components 315, 320, 325, 330, 335, 340, 345, 350 Hz for difference of acoustic signals of states Y and Z. There are no common frequency components for acoustic signals of states X, Y, Z. Component frequencies 320, 340, 330, 350 Hz are common for two differences of frequency spectra of acoustic signals. What will happen for more analysed differences of acoustic signals of states of the motor? To solve this task the author used a parameter TCFC-MTS (Threshold of common frequency components many training sets). This parameter adds MULTIEXPANDED extension. Set the parameter TCFC-MTS = (number of required common frequency components of training sets)/(number of differences between frequency spectra). The number of common frequency components depends on the value of TCFC-MTS. Let's analyse CASE2. There are 3 training sets. Each of them has 4 acoustic signals: (X1, X2, X3, X4), (Y1, Y2, Y3, Y4), (Z1, Z2, Z3, Z4), where X1 acoustic signal of state X1, X2 - acoustic signal of state X2, X3 acoustic signal of state X3, X4 - acoustic signal of state X4. The SMOFS-32-MULTIEXPANDED selects frequency components for each difference in one training set: (|X1 − X2|), (|X1 − X3|), (|X1 − X4|), (|X2 − X3|), (|X2 − X4|), (|X3 − X4|), (|Y1 − Y2|), (|Y1 − Y3|), (|Y1 − Y4|), (|Y2 − Y3|), (|Y2 − Y4|), (|Y3 − Y4|), (|Z1 − Z2|), (|Z1 − Z3|), (|Z1 − Z4|), (|Z2 − Z3|), (|Z2 − Z4|), (|Z3 − Z4|) - 18 differences between frequency spectra of acoustic signals. If the parameter TCFC-MTS = 3/18 = 0.166, then the SMOFS-32-MULTIEXPANDED selects frequency components found 3 times for 18 differences. For example, frequency component 300 Hz were found 4 times (max. 18 times). Frequency component 350 Hz were found 8 times. Frequency component 400 Hz were found 12 times. The SMOFS-32-MULTIEXPANDED selects 300, 350, 400 Hz (if TCFC-MTS = 3/18 = 0.166). If the parameter TCFCMTS = 11/18 = 0.611, then the SMOFS-32-MULTIEXPANDED selects frequency component 400 Hz. If the parameter TCFCMTS = 13/18 = 0.722, then the SMOFS-32-MULTIEXPANDED selects 0 frequency components. It can be noticed that the parameter TCFC-MTS should be set experimentally. Set the number of groups (Group extension). The SMOFS-32-MULTIEXPANDED-2-GROUPS used 2 groups. The SMOFS-32-MULTIEXPANDED-1-GROUP used 1 group. For proper recognition group of essential frequency components required all analysed classes. For example, the frequency component 300 Hz (found 4 times) separates states (|X1 − X2|), (|X1 − X3|), (|X1 − X4|). The frequency component 350 Hz (found 8 times) separates states (|X2 − X3|), (|X2 − X4|), (|X3 − X4|). The frequency component 400 Hz (found 12 times) separates states (|X2 − X3|), (|X2 − X4|), (|X3 − X4|). It can be noticed that the frequency component 400 Hz (found 12 times) is better than the frequency component 350 Hz (found 8 times). The essential frequency components are 300 Hz and 400 Hz, because they separates 4 states together. The essential frequency components 300 Hz and 400 Hz form 1 group of essential frequency components. Select 1-2 groups of essential frequency components. Form a feature vector consisted of essential frequency components.

The proposed method SMOFS-32-MULTIEXPANDED-2-GROUPS was shown in Fig. 5. The author used 5 training sets (20 one-second samples, 30 differences of frequency spectra of acoustic signals of the three-phase induction motor). The obtained absolute values of differences were following: |ahtim − atimbrb|, |ahtim − atimbrbs|, |ahtim − atimfr|, |atimbrb − atimbrbs|, |atimbrb − atimfr|, |atimbrbs − atimfr| were presented in Figs. 6–11 (rotor speed of the analysed motor = 1400 rpm, training set 5). The SMOFS-32-MULTIEXPANDED-2-GROUPS found 6 essential 84

Applied Acoustics 137 (2018) 82–89

A. Glowacz

Fig. 8. The difference between spectra of frequency of acoustic signal of the healthy three-phase induction motor and the three-phase induction motor with faulty ring of squirrel-cage (|ahtim − atimfr|) - after using SMOFS-32-MULTIEXPANDED-2-GROUPS.

Fig. 5. Block diagram of the SMOFS-32-MULTIEXPANDED-2-GROUPS.

Fig. 9. The difference between spectra of frequency of acoustic signal of the three-phase induction motor with broken rotor bar and the three-phase induction motor with 2 broken rotor bars (|atimbrb - atimbrbs) - after using SMOFS-32-MULTIEXPANDED-2GROUPS.

Fig. 6. The difference between spectra of frequency of acoustic signal of the healthy three-phase induction motor and the three-phase induction motor with broken rotor bar (|ahtim − atimbrb|) - after using SMOFS-32-MULTIEXPANDED-2-GROUPS.

Fig. 10. The difference between spectra of frequency of acoustic signal of the three-phase induction motor with broken rotor bar and the three-phase induction motor with faulty ring of squirrel-cage (|atimbrb − atimfr|) - after using SMOFS-32-MULTIEXPANDED-2GROUPS.

MTS = 0.333. The author used 20 one-second training samples for 5 analysed training sets. The computed essential frequency components formed feature vectors. The obtained feature vectors were used in the classification step. The Nearest Neighbour classifier, backpropagation neural network and modified classifier based on words coding were used for recognition of the computed feature vectors. However other classifiers can be also used for fault diagnosis: Support Vector Machine [51], Naive Bayes classifier [51,52], Support Vector Machine [51], classification tree

Fig. 7. The difference between spectra of frequency of acoustic signal of the healthy three-phase induction motor and the three-phase induction motor with 2 broken rotor bars (|ahtim − atimbrbs|) - after using SMOFS-32-MULTIEXPANDED-2-GROUPS.

frequency components - 301, 923, 2222, 300, 2492, 648 Hz, for TCFCMTS = 0.333. The SMOFS-32-MULTIEXPANDED-1-GROUP found 3 essential frequency components - 301, 923, 2222 Hz, for TCFC85

Applied Acoustics 137 (2018) 82–89

A. Glowacz

Fig. 12. Structure of the considered back-propagation neural network.

network can be class of acoustic signal for example: healthy three-phase induction motor. More information about backpropagation neural network is available in the following articles [62–65]. The author used backpropagation neural network as a classifier. The author used back-propagation neural network presented in the Fig. 12. The network had 3 or 6 inputs (essential frequency components), 10 neurons of hidden layer and 4 neurons of output layer. The output values of backpropagation neural network were following: 0001 - class A (healthy three-phase induction motor), 0010 class B (three-phase induction motor with broken rotor bar), 0100 class C (three-phase induction motor with 2 broken rotor bars) and 1000 - class D (three-phase induction motor with faulty ring of squirrelcage).

Fig. 11. The difference between spectra of frequency of acoustic signal of the three-phase induction motor with 2 broken rotor bars and the three-phase induction motor with faulty ring of squirrel-cage (|atimbrbs − atimfr|) - after using SMOFS-32-MULTIEXPANDED-2GROUPS.

[53,54], fuzzy classifier [55,56]. The author used limited number of classifiers - 3, because the results of recognition were similar. 2.2. Nearest Neighbour classification method In the world literature there are a lot of applications of the Nearest Neighbour (NN) classification method [57–61]. The NN is often used for signal processing and data classification. This classifier can classify linearly separable and non-linearly separable feature vectors (data) properly. Advantages of the method are: high recognition rate, simple to implement, can classify feature vector using small number of training examples, various distance functions, good for multi-class problems. Disadvantages are following: time complexity O(N2), storage of feature vectors, selection of distance function. The Nearest Neighbour classifier can be considered as supervised method. It uses labeled training feature vectors (data). Next it compares labeled training feature vectors with test feature vectors using selected distance function. The author used the Manhattan distance for analysis, however other distance functions such as: Euclidean, Jacquard, Minkowski could be used. The results of classification using Euclidean, Minkowski distance functions were similar. The author decided to use the Manhattan distance. It was defined as follows (4):

2.4. Modified classifier based on words coding Modified classifier based on words transformed numerical values into a string of characters (word). Modified classifier based on words used two steps: pattern creation and testing. The vector of features x = [x1, x2, … , xn] was given, where x1, … , xn were features (features may be frequency components computed from an acoustic signal using the FFT, MSAF, SMOFS, SMOFS-EXPANDED, SMOFS-32-MULTIEXPANDED-2-GROUPS). The classes of patterns are denoted as w1, w2, … , wj, where the index j was the class number. In the process of pattern creation and testing, we received training word vectors va, vb, … , vj, where vj = [v1, v2, … , vn], v1, v2, … , vn were words for j-th class of patterns (training samples). Test word vectors fa, fb, … , fj were defined as fj = [f1, f2, … , fn], where f1, f2, … , fn were words for the j-th class (test samples). Transformation to words was carried out according to the formula (5):

1

d (ahtim−atimbrb) =

∑

|(ahtimi−atimbrbi )|,

i=1

⎧ x i ∈ [k ,2k ) ⇒ x i → vi1 ⎪ x ∈ [2k ,3k ) ⇒ x → v i i i2 ⎨… ⎪ x i ∈ [kg ,kg + k ) ⇒ x i → vig ⎩

(4)

where the test feature vector ahtim = [ahtim224, ahtim686, ahtim1652, ahtim223, ahtim1853, ahtim482] and the training feature vector atimbrb = [atimbrb224, atimbrb686, atimbrb1652, atimbrb223, atimbrb1853, atimbrb482]. The computed distances between test feature vectors and training feature vectors were used for decisions of classifications. More information about the NN classification method can be found in the following literature [57–61].

(5)

where: k ∈ W , g was the number of words, vig was g-th word, xi was i-th coordinate of feature vector xj, i = 0, … , n. In testing process, the word vector f was assigned to class, whose training word vector vj was the closest to the test word vector f. For this purpose, the lexicographic comparison was used. Strings of characters were compared each other (the coordinate of the training word vector vj was compared to the coordinate of the test word vector f). The result of the comparison was true or false. To obtain the recognition result, the following formula was introduced:

2.3. Backpropagation neural network Backpropagation neural network is also supervised learning method. It uses 2 processes - pattern creation process (learning process) and testing process. The neural network consists of many artificial neurons. They are used to receive and process data. The backpropagation method is used in artificial neural network for learning process (using training feature vectors). It computes the error of each neuron (in each layer of neurons). Each neuron has a weight value. Initially the weight value is random. Next in the learning process the weight value is modified. It depends on the computed error. Next testing process is performed using test feature vectors. Next the data (test feature vectors) are multiplied by weights of neurons. After that the output of neural network is computed. The output of neural

Uj =

U1 ·100%, U2

(6)

where: U1 - number of well-compared words belonging to the j-th class, U2 - number of all comparisons (equal to the number of coordinates of the word vector U2 = n), Uj - percentage of well-recognized words belonging to the j-th class . To obtain the final result, the formula was introduced:

max(Uj ) ⇒ f → wj 86

j = 1,2,…,M,

(7)

Applied Acoustics 137 (2018) 82–89

A. Glowacz

Table 1 The results of recognition of acoustic signals of the three-phase induction motor using the SMOFS-32-MULTIEXPANDED-2-GROUPS and the Nearest Neighbour classifier. Type of acoustic signal signal

ERASTPIM [%]

Healthy motor Motor with broken rotor bar Motor with 2 broken rotor bars Motor with faulty ring of squirrel-cage

95.83 100 100 100 TERASTPIM [%] 98.95

Table 2 The results of recognition of acoustic signals of the three-phase induction motor using the SMOFS-32-MULTIEXPANDED-1-GROUP and the Nearest Neighbour classifier. Fig. 13. Analysed three phase induction motors.

3. Analysis of acoustic signals of three-phase induction motors Following acoustic signals of 4 states of three-phase induction motor were measured, processed and analysed (Fig. 13): healthy three-phase induction motor, three-phase induction motor with broken rotor bar, three-phase induction motor with 2 broken rotor bars, three-phase induction motor with faulty ring of squirrel-cage. Operational parameters of induction motors were: SM = 1400 rpm, PNOM = 0.55 kW, fNOM = 50 Hz, UNOM = 220 V, ECNOM = 2.52/1.47 A (Δ/Y), where SM rotor speed of the motor, PNOM - nominal power of the motor, fNOM nominal frequency of the motor, UNOM - nominal voltage of the motor, ECNOM - nominal electric current of the motor. 20 one-second samples of acoustic signals were analysed for pattern creation. 188 one-second samples were analysed for testing process. Training and test samples of acoustic signals of three-phase induction motors were analysed using proposed approach. Efficiency of recognition of acoustic signal of three-phase induction motor was defined as (8):

ERASTPIM = (NPTSAS )/(NATSAS ) 100%

ERASTPIM [%]

Healthy motor Motor with broken rotor bar Motor with 2 broken rotor bars Motor with faulty ring of squirrel-cage

100 100 93 100 TERASTPIM [%] 98.26

Table 3 The results of recognition of acoustic signals of the three-phase induction motor using the SMOFS-32-MULTIEXPANDED-2-GROUPS and the backpropagation neural network. Type of acoustic signal signal

ERASTPIM [%]

Healthy motor Motor with broken rotor bar Motor with 2 broken rotor bars Motor with faulty ring of squirrel-cage

100 100 100 100 TERASTPIM [%] 100

Table 4 The results of recognition of acoustic signals of the three-phase induction motor using the SMOFS-32-MULTIEXPANDED-1-GROUP and the backpropagation neural network.

(8)

where: ERASTPIM – efficiency of recognition of acoustic signal of threephase induction motor for selected class, NPTSAS – number of test samples of acoustic signals of three-phase induction motor for selected class tested properly, NATSAS – number of all test samples of acoustic signals of three-phase induction motor for selected class. Total efficiency of recognition of acoustic signal of three-phase induction motor was defined by following formula (9):

TERASTPIM = (ERASTPIM1 + ERASTPIM 2 + ERASTPIM 3 + ERASTPIM 4 )/4

Type of acoustic signal signal

Type of acoustic signal signal

ERASTPIM [%]

Healthy motor Motor with broken rotor bar Motor with 2 broken rotor bars Motor with faulty ring of squirrel-cage

100 98.6 95.83 100 TERASTPIM [%] 98.87

(9)

where TERASTPIM - total efficiency of recognition of acoustic signal of three-phase induction motor, ERASTPIM1 - efficiency of recognition of acoustic signal of the healthy three-phase induction motor, ERASTPIM2 efficiency of recognition of acoustic signal of the three-phase induction motor with broken rotor bar, ERASTPIM3 - efficiency of recognition of acoustic signal of the three-phase induction motor with 2 broken rotor bars, ERASTPIM4 - efficiency of recognition of acoustic signal of the threephase induction motor with faulty ring of squirrel-cage. Four states of the three-phase induction motor were analysed. The results were presented in Tables 1–5. In Table 1, the author presented the results of recognition of acoustic signals of the three-phase induction motor using the SMOFS-32-MULTIEXPANDED-2-GROUPS (6 analysed frequency components) and the Nearest Neighbour classifier. In Table 2, the author presented the results of recognition of acoustic signals of the three-phase induction motor using the SMOFS32-MULTIEXPANDED-1-GROUP (3 analysed frequency components) and the Nearest Neighbour classifier. In Table 3, the author presented the results of recognition of acoustic signals of the three-phase induction motor using the SMOFS32-MULTIEXPANDED-2-GROUPS (6 analysed frequency components)

Table 5 The results of recognition of acoustic signals of the three-phase induction motor using the SMOFS-32-MULTIEXPANDED-1-GROUP and the modified classifier based on words coding. Type of acoustic signal signal

ERASTPIM [%]

Healthy motor Motor with broken rotor bar Motor with 2 broken rotor bars Motor with faulty ring of squirrel-cage

100 84.72 95.83 72.22 TERASTPIM [%] 88.19

and the backpropagation neural network. In Table 4, the author presented the results of recognition of acoustic signals of the three-phase induction motor using the SMOFS32-MULTIEXPANDED-1-GROUP (3 analysed frequency components) and the backpropagation neural network. In Table 5, the author presented the results of recognition of 87

Applied Acoustics 137 (2018) 82–89

A. Glowacz

acoustic signals of the three-phase induction motor using the SMOFS32-MULTIEXPANDED-1-GROUP (3 analysed frequency components) and the modified classifier based on words coding (k = 0.003). The obtained results of analysed acoustic signals and methods were very good (TERASTPIM was in the range of 88.19–100%). The SMOFS32-MULTIEXPANDED-2-GROUPS selected 6 frequency components (for analysed three-phase induction motors). The SMOFS-32-MULTIEXPANDED-1-GROUP selected 3 frequency components. The best recognition results were obtained for SMOFS-32-MULTIEXPANDED-2-GROUPS (6 analysed frequency components) and the backpropagation neural network TERASTPIM = 100% (Table 3). The proposed feature extraction methods found proper features for analysed three-phase induction motors. Next obtained feature vectors were classified by the Nearest Neighbour classifier, backpropagation neural network and the modified classifier based on words coding.

[5] Glowacz A, Glowacz W, Glowacz Z, Kozik J, Gutten M, Korenciak D, et al. Fault diagnosis of three phase induction motor using current signal, MSAF-Ratio15 and selected classifiers. Arch Metall Mater 2017;62(4):2413–9. http://dx.doi.org/10. 1515/amm-2017-0355. [6] Glowacz A, Glowacz W, Glowacz Z. Recognition of armature current of DC generator depending on rotor speed using FFT, MSAF-1 and LDA. Eksploatacja i Niezawodnosc-Maintenance and Reliability 2015;17(1):64–9. http://dx.doi.org/10. 17531/ein.2015.1.9. [7] Li Y, Chai F, Song ZX, Li ZY. Analysis of vibrations in interior permanent magnet synchronous motors considering air-gap deformation. Energies 2017;10(9). http:// dx.doi.org/10.3390/en10091259. [8] Irfan M, Saad N, Ibrahim R, Asirvadam VS. An on-line condition monitoring system for induction motors via instantaneous power analysis. J Mech Sci Technol 2015;29(4):1483–92. http://dx.doi.org/10.1007/s12206-015-0321-9. [9] Yousefi R, Yusof R, Arfa R. Fault diagnosis of three-phase induction motors based on vibration and electrical current signals. Jurnal Teknologi 2016;78:129–33. [10] Bazan GH, Scalassara PR, Endo W, Goedtel A, Godoy WF, Palacios RHC. Stator fault analysis of three-phase induction motors using information measures and artificial neural networks. Electr Power Syst Res 2017;143:347–56. http://dx.doi.org/10. 1016/j.epsr.2016.09.031. [11] Antonino-Daviu JA, Gyftakis KN, Garcia-Hernandez R, Razik H, Cardoso AJM. Comparative influence of adjacent and non-adjacent broken rotor bars on the induction motor diagnosis through MCSA and ZSC methods. In: IECON 2015 - 41st annual conference of the IEEE industrial electronics society, Book series. Yokohama (Japan): IEEE Industrial Electronics Society; 2015. p. 1680–5. [12] Yang T, Pen HB, Wang ZX, Chang CS. Feature knowledge based fault detection of induction motors through the analysis of stator current data. IEEE Trans Instrum Measur 2016;65:549–58. http://dx.doi.org/10.1109/TIM.2015.2498978. [13] Gutten M, Korenciak D, Kucera M, Sebok M, Opielak M, Zukowski P, et al. Maintenance diagnostics of transformers considering the influence of short-circuit currents during operation. Eksploatacja i Niezawodnosc-Maintenance and Reliability 2017;19(3):459–66. http://dx.doi.org/10.17531/ein.2017.3.17. [14] Gutten M, Janura R, Sebok M, Korenciak D, Kucera M. Measurement of short-circuit effects on transformer winding with SFRA method and impact test. Metrol Measur Syst 2016;23(4):521–9. http://dx.doi.org/10.1515/mms-2016-0044. [15] Sebok M, Gutten M, Kucera M. Diagnostics of electric equipments by means of thermovision. Przeglad Elektrotechniczny 2011;87(10):313–7. [16] Liu H, Wang ZX, Zhong J, Xie ZW. Early detection of spontaneous combustion disaster of sulphide ore stockpiles. Tehnicki Vjesnik-Technical Gazette 2015;22(6):1579–87. [17] Singh SK, Kumar S, Dwivedi JP. Compound fault prediction of rolling bearing using multimedia data. Multimedia Tools Appl 2017;76(18):18771–88. http://dx.doi. org/10.1007/s11042-017-4419-1. [18] Zhang GT, Wu JY, Hao LL. Fast calculation model and theoretical analysis of rotor unbalanced magnetic pull for inter-turn short circuit of field windings of non-salient pole generators. Energies 2017;10(5). http://dx.doi.org/10.3390/en10050732. [19] Ouachtouk I, El Hani S, Dahi K. Wireless health monitoring system for rotor eccentricity faults detection in induction machine. Adv Electr Electron Eng 2017;15(3):376–82. http://dx.doi.org/10.15598/aeee.v15i3.2193. [20] Jozwik J. Identification and monitoring of noise sources of cnc machine tools by acoustic holography methods. Adv Sci Technol-Res J 2016;10(30):127–37. http:// dx.doi.org/10.12913/22998624/63386. [21] Tra V, Kim J, Khan SA, Kim JM. Bearing fault diagnosis under variable speed using convolutional neural networks and the stochastic diagonal Levenberg-Marquardt algorithm. Sensors 2017;17(12). http://dx.doi.org/10.3390/s17122834. [22] Sangeetha P, Hemamalini S. Dyadic wavelet transform-based acoustic signal analysis for torque prediction of a three-phase induction motor. IET Signal Process 2017;11(5):604–12. http://dx.doi.org/10.1049/iet-spr.2016.0165. [23] Glowacz A, Glowacz W, Glowacz Z, Kozik J. Early fault diagnosis of bearing and stator faults of the single-phase induction motor using acoustic signals. Measurement 2018;113:1–9. http://dx.doi.org/10.1016/j.measurement.2017.08. 036. [24] Glowacz A. Recognition of acoustic signals of loaded synchronous motor using FFT, MSAF-5 and LSVM. Arch Acoust 2015;40(2):197–203. http://dx.doi.org/10.1515/ aoa-2015-0022. [25] Mika D, Jozwik J. Normative measurements of noise at CNC machines work stations. Adv Sci Technol Res J 2016;10(30):138–43. http://dx.doi.org/10.12913/ 22998624/63387. [26] Deptula A, Kunderman D, Osinski P, Radziwanowska U, Wlostowski R. Acoustic diagnostics applications in the study of technical condition of combustion engine. Arch Acoust 2016;41(2):345–50. http://dx.doi.org/10.1515/aoa-2016-0036. [27] Glowacz A. Fault diagnostics of acoustic signals of loaded synchronous motor using SMOFS-25-EXPANDED and selected classifiers. Tehnicki vjesnik–Technical Gazette 2016;23(5):1365–72. http://dx.doi.org/10.17559/TV-20150328135652. [28] Glowacz A. DC motor fault analysis with the use of acoustic signals, coiflet wavelet transform, and K-nearest neighbor classifier. Arch Acoust 2015;40(3):321–7. http://dx.doi.org/10.1515/aoa-2015-0035. [29] Carletti E, Miccoli G, Pedrielli F, Parise G. Vibroacoustic measurements and simulations applied to external gear pumps an integrated simplified approach. Arch Acoust 2016;41(2):285–96. http://dx.doi.org/10.1515/aoa-2016-0028. [30] Delgado-Arredondo PA, Garcia-Perez A, Morinigo-Sotelo D, Osornio-Rios RA, Avina-Cervantes JG, Rostro-Gonzalez H, et al. Comparative study of time-frequency decomposition techniques for fault detection in induction motors using vibration analysis during startup transient. Shock Vib 2015. http://dx.doi.org/10.1155/ 2015/708034. [31] Krolczyk GM, Krolczyk JB, Legutko S, Hunjet A. Effect of the disc processing

4. Conclusions In this article the author described acoustic based fault diagnosis technique of the three-phase induction motor. Four states of the threephase induction motor were analysed: healthy three-phase induction motor, three-phase induction motor with broken rotor bar, three-phase induction motor with 2 broken rotor bars, three-phase induction motor with faulty ring of squirrel-cage. The proposed feature extraction methods: SMOFS-32-MULTIEXPANDED-2-GROUPS and SMOFS32-MULTIEXPANDED-1-GROUP were analysed. The Nearest Neighbour classifier, backpropagation neural network and the modified classifier based on words coding were used for recognition. The results of recognition of acoustic signals were very good for the real data (TERASTPIM was in the range of 88.19–100%). The presented acoustic based fault diagnostic techniques were not expensive. The laptop and the capacity microphone can be bought for 350$. The described acoustic based fault diagnosis techniques can be used to protect other types of rotating electric motors. Many rotating electric motors can be diagnosed using acoustic signals. It can prevent unexpected failure and improve maintenance of electric motors. Advantages of the proposed acoustic based fault diagnosis technique are: non-invasive technique, low cost and instant measurement of acoustic signals. However analysis was carried out for 4 states of the motor. The proposed methods were also proper for more states. In the future, it is good idea to analyse more faults and motors for better fault diagnosis. There is also idea to mix various techniques of fault diagnosis such as: thermal imaging, vibration analysis and electric current analysis. Acknowledgments The author thanks unknown reviewers for their valuable suggestions. Conflicts of interest The author declares no conflict of interest. References [1] Merizalde Y, Hernandez-Callejo L, Duque-Perez O. State of the art and trends in the monitoring, detection and diagnosis of failures in electric induction motors. Energies 2017;10(7). http://dx.doi.org/10.3390/en10071056. [2] Ciszewski T, Gelman L, Swedrowski L. Current-based higher-order spectral covariance as a bearing diagnostic feature for induction motors. Insight 2016;58(8):431–4. http://dx.doi.org/10.1784/insi.2016.58.8.431. [3] Verucchi C, Bossio J, Bossio G, Acosta G. Misalignment detection in induction motors with flexible coupling by means of estimated torque analysis and MCSA. Mech Syst Sig Process 2016;80:570–81. http://dx.doi.org/10.1016/j.ymssp.2016. 04.035. [4] Sulowicz M, Weinreb K, Mielnik R, Zywczak T, Jaraczewski M. The method of current measurement in the rotor cage bars of prototype induction motor with the use of Rogowski coils. In: 2015 International conference on information and digital technologies (IDT). p. 357–65.

88

Applied Acoustics 137 (2018) 82–89

A. Glowacz

[32]

[33]

[34]

[35]

[36]

[37]

[38]

[39]

[40]

[41]

[42]

[43]

[44]

[45]

[46]

[47]

[48]

technology on the vibration level of the chipper during operations. Tehnicki Vjesnik-Technical Gazette 2014;21(2):447–50. Li ZX, Jiang Y, Hu C, Peng Z. Recent progress on decoupling diagnosis of hybrid failures in gear transmission systems using vibration sensor signal: a review. Measurement 2016;90:4–19. http://dx.doi.org/10.1016/j.measurement.2016.04. 036. Sawczuk W. The application of vibration accelerations in the assessment of average friction coefficient of a railway brake disc. Measur Sci Rev 2017;17(3):125–34. http://dx.doi.org/10.1515/msr-2017-0016. Sawczuk W, Szymanski GM. Diagnostics of the railway friction disc brake based on the analysis of the vibration signals in terms of resonant frequency. Arch Appl Mech 2017;87(5):801–15. http://dx.doi.org/10.1007/s00419-016-1202-0. Adamczak S, Stepien K, Wrzochal M. Comparative study of measurement systems used to evaluate vibrations of rolling bearings. Proc Eng 2017;192:971–5. http:// dx.doi.org/10.1016/j.proeng.2017.06.167. Zhang C, Li ZX, Hu C, Chen S, Wang JG, Zhang XG. An optimized ensemble local mean decomposition method for fault detection of mechanical components. Measur Sci Technol 2017;28(3). http://dx.doi.org/10.1088/1361-6501/aa56d3. Li ZX, Jiang Y, Hu CQ, Peng ZX. Difference equation based empirical mode decomposition with application to separation enhancement of multi-fault vibration signals. J Diff Eq Appl 2017;23(1–2):457–67. http://dx.doi.org/10.1080/ 10236198.2016.1254206. Hreha P, Radvanska A, Knapcikova L, Krolczyk GM, Legutko S, Krolczyk JB, et al. Roughness parameters calculation by means of on-line vibration monitoring emerging from AWJ interaction with material. Metrol Measur Syst 2015;22(2):315–26. http://dx.doi.org/10.1515/mms-2015-0024. Zhang C, Peng ZX, Chen S, Li ZX, Wang JG. A gearbox fault diagnosis method based on frequency-modulated empirical mode decomposition and support vector machine. Proc Inst Mech Engin Part C-J Mech Eng Sci 2018;232(2):369–80. http://dx. doi.org/10.1177/0954406216677102. Singh G, Kumar TCA, Naikan VNA. Induction motor inter turn fault detection using infrared thermographic analysis. Infr Phys Technol 2016;77:277–82. http://dx.doi. org/10.1016/j.infrared.2016.06.010. Garcia-Ramirez AG, Morales-Hernandez LA, Osornio-Rios RA, Garcia-Perez A, Romero-Troncoso RJ. Thermographic technique as a complement for MCSA in induction motor fault detection. In: 2014 International conference on electrical machines (ICEM). p. 1940–5. Singh G, Naikan VNA. Infrared thermography based diagnosis of inter-turn fault and cooling system failure in three phase induction motor. Infr Phys Technol 2017;87:134–8. http://dx.doi.org/10.1016/j.infrared.2017.10.007. Kubiak I. Video signal level (colour intensity) and effectiveness of electromagnetic infiltration. Bull Pol Acad Sci-Tech Sci 2016;64(1):207–18. http://dx.doi.org/10. 1515/bpasts-2016-0023. O'Brien RJ, Fontana JM, Ponso N, Molisani L. A pattern recognition system based on acoustic signals for fault detection on composite materials. Euro J Mech A-Solids 2017;64:1–10. http://dx.doi.org/10.1016/j.euromechsol.2017.01.007. Wang WJ, Cui LL, Chen DY. Multi-scale morphology analysis of acoustic emission signal and quantitative diagnosis for bearing fault. Acta Mech Sinica 2016;32(2):265–72. http://dx.doi.org/10.1007/s10409-015-0529-z. Caesarendra W, Kosasih B, Tieu AK, Zhu HT, Moodie CAS, Zhu Q. Acoustic emission-based condition monitoring methods: review and application for low speed slew bearing. Mech Syst Sig Process 2016;72–73:134–59. http://dx.doi.org/10. 1016/j.ymssp.2015.10.020. Rzeszucinski P, Orman M, Pinto CT, Tkaczyk A, Sulowicz M. A signal processing approach to bearing fault detection with the use of a mobile phone. In: 2015 IEEE 10th international symposium on diagnostics for electric machines, power electronics and drives (SDEMPED). p. 310–5. Germen E, Basaran M, Fidan M. Sound based induction motor fault diagnosis using

[49]

[50]

[51]

[52]

[53]

[54] [55]

[56]

[57]

[58]

[59]

[60]

[61]

[62]

[63]

[64]

[65]

89

Kohonen self-organizing map. Mech Syst Sig Process 2014;46:45–58. http://dx.doi. org/10.1016/j.ymssp.2013.12.002. Delgado-Arredondo PA, Morinigo-Sotelo D, Osornio-Rios RA, Avina-Cervantes JG, Rostro-Gonzalez H, Romero-Troncoso RD. Methodology for fault detection in induction motors via sound and vibration signals. Mech Syst Sig Process 2017;83:568–89. http://dx.doi.org/10.1016/j.ymssp.2016.06.032. Lara R, Jimenez-Romero R, Perez-Hidalgo F, Redel-Macias MD. Influence of constructive parameters and power signals on sound quality and airborne noise radiated by inverter-fed induction motors. Measurement 2015;73:503–14. http://dx. doi.org/10.1016/j.measurement.2015.05.049. Palacios RHC, da Silva IN, Goedtel A, Godoy WF. A comprehensive evaluation of intelligent classifiers for fault identification in three-phase induction motors. Electr Power Syst Res 2015;127:249–58. http://dx.doi.org/10.1016/j.epsr.2015.06.008. Swetapadma A, Yadav A. Protection of parallel transmission lines including intercircuit faults using Naive Bayes classifier. Alex Eng J 2016;55(2):1411–9. http://dx. doi.org/10.1016/j.aej.2016.03.029. Regulski K, Jakubski J, Opalinski A, Brzezinski M, Glowacki M. The prediction of moulding sand moisture content based on the knowledge acquired by data mining techniques. Arch Metal Mater 2016;61(3):1363–8. http://dx.doi.org/10.1515/ amm-2016-0277. Koprowski R. Classification. Book Series: Studies in Computational Intelligence, vol. 682. 2017. p. 83–109. http://dx.doi.org/10.1007/978-3-319-50490-2_5. Valis D, Zak L. Contribution to prediction of soft and hard failure occurrence in combustion engine using oil tribo data. Eng Fail Anal 2017;82:583–98. http://dx. doi.org/10.1016/j.engfailanal.2017.04.018. Valis D, Zak L, Pokora O, Lansky P. Perspective analysis outcomes of selected tribodiagnostic data used as input for condition based maintenance. Reliab Eng Syst Safety 2016;145:231–42. http://dx.doi.org/10.1016/j.ress.2015.07.026. Vitola J, Pozo F, Tibaduiza DA, Anaya M. A sensor data fusion system based on knearest neighbor pattern classification for structural health monitoring applications. Sensors 2017;17(2). http://dx.doi.org/10.3390/s17020417. Tian J, Morillo C, Azarian MH, Pecht M. Motor bearing fault detection using spectral kurtosis-based feature extraction coupled with K-nearest neighbor distance analysis. IEEE Trans Indust Electron 2016;63(3):1793–803. http://dx.doi.org/10. 1109/TIE.2015.2509913. Dou DY, Zhou SS. Comparison of four direct classification methods for intelligent fault diagnosis of rotating machinery. Appl Soft Comput 2016;46:459–68. http:// dx.doi.org/10.1016/j.asoc.2016.05.015. Baraldi P, Cannarile F, Di Maio F, Zio E. Hierarchical k-nearest neighbours classification and binary differential evolution for fault diagnostics of automotive bearings operating under variable conditions. Eng Appl Artif Intell 2016;56:1–13. http://dx.doi.org/10.1016/j.engappai.2016.08.011. Han T, Jiang DX, Zhang XC, Sun YK. Intelligent diagnosis method for rotating machinery using dictionary learning and singular value decomposition. Sensors 2017;17(4). http://dx.doi.org/10.3390/s17040689. Hu ZB, Su J, Jotsov V, Kochan O, Mykyichuk M, Kochan R, et al. Data science applications to improve accuracy of thermocouples. In: 2016 IEEE 8th international conference on intelligent systems (IS); 2016. p. 180–8. Koprowski R, Lanza M, Irregolare C. Corneal power evaluation after myopic corneal refractive surgery using artificial neural networks. BioMed Eng OnLine 2016;15. http://dx.doi.org/10.1186/s12938-016-0243-5. Khan ZF, Al Sayyari AS, Quadri SU. Automated segmentation of lung images using textural echo state neural networks. In: 2017 International conference on informatics, health & technology (ICIHT). Riyadh (Saudi Arabia); 2017. Izworski A, Tadeusiewicz R, Wszolek W. Artificial intelligence methods in diagnostics of the pathological speech signals. In: Knowledge-based intelligent information and engineering systems, Pt 3, Proceedings, Lecture notes in artificial intelligence, vol. 3215; 2004. p. 740–8.