Electric motor faults diagnosis using artificial neural networks
FAULT DIAGNOSIS
Electric motor faults diagnosis using artificial neural networks Lingxin Li, C K Mechefske and Weidong Li
This paper presents electric motor fault diagnosis using two kinds of Artificial Neural Networks (ANN): feedforward networks and self organising maps (SOM). Major faults such as bearing faults, stator winding fault, unbalanced rotor and broken rotor bars are considered. The ANNs were trained and tested using measurement data from stator currents and mechanical vibration signals. The effects of different network structures and the training set sizes on the performance of the ANNs are discussed. This study shows that the feedforward ANN with a very simple internal structure can give satisfactory results, while SOMs can classify the type of motor faults during steady state working conditions. The experiment results also show that the feedforward ANN is the more promising scheme in this case where fault data from electric motors is available.
1. Introduction With an increasing trend towards the need for mechanisation, automation and high capacity mining equipment, it has become necessary that the reliability of equipment be improved. It has been found that maintenance accounts for about 30-60% of the total operating cost budget for a typical mining company and represents the largest portion of the mine’s controllable operating costs[1]. In addition, because of the nature of mining equipment, the failures of mining equipment may significantly reduce and even stop production. This, coupled with the fact that the mining industry today is faced with an ever more complex and demanding marketplace, has led to an increasing interest in methods to reduce maintenance costs. Syncrude Canada Ltd. is the world’s largest producer of crude oil from oil sands and the largest single source producer in Canada, and currently supplies 13 percent of Canada’s petroleum requirements. At Syncrude, a truck and shovel system is used to mine and deliver the raw oil sand to a central separator[2]. The shovel is primarily an excavating and loading device and its reliability and efficiency makes a notable impact on the mining productivity and the plant-wide profits. Electric motors are critical components in the P&H 4100 type cable shovel. The ever changing mining environment and dynamical loading conditions always strain and wear motors and cause faults. Electric motor manufacturers and users have relied for a long time on protective relays – such as circuit breakers or fuses to monitor faults and disconnect motors. However, following this traditional scheme led to facing some situations where the machine was badly damaged, mainly due to the machine tripping after the fault was well developed[3]. This scheme can not provide warning of impending faults before they occur. Incipient fault detection for electric motors would not only be highly cost-effective in The authors are in the Department of Mechanical and Materials Engineering, Queens’ University, Kingston, Ontario, Canada, K7L 3N6. E-mail: [email protected]; [email protected]; [email protected]
Insight Vol 46 No 10 October 2004
minimising maintenance downtime, but it would also allow preventive maintenance to be scheduled. The work reported in this paper is part of a larger research project: Electric Motor Fault Diagnosis Systems for P&H4100 Shovels at Syncrude Canada Ltd. The primary focus of the work is to employ appropriate technology to monitor and detect normal and faulty electric motors and to provide a warning and diagnose the faults at an early stage. There already exist several monitoring methods for electric motors. Stator current monitoring is one of the most widely employed monitoring techniques[4]. This monitoring is believed to be most sensitive to motor component faults and can provide unique fault patterns. In addition to stator current analysis, vibration monitoring can also provide effective indications especially for motor bearing and broken rotor bar faults. The use of artificial neural networks (ANN) for processing stator current and motor vibration signals is particularly attractive due to the remarkable information processing characteristics such as nonlinearity, robustness, ability to learn and ability to handle imprecise and fuzzy information[5]. The use of artificial neural networks for electric motor fault detection was first proposed by Chow et al[6-8]. They proposed a typical back-propagation (BP) neural network structure for incipient motor fault diagnosis. The incipient faults focused on turn-to-turn insulation faults and bearing wear. Kolla et al[9] investigated motor faults including overloads, single phasing, unbalanced supply voltage, locked rotor, ground fault, over-voltage and under-voltage using detection methods based on the feedforward neural network. Besides BP neural networks, other unsupervised learning neural networks were also used in motor fault diagnosis. Penman and Jin applied the Kohonen’s self-organising map (SOM) to classify multiple faults in a 3 kW, 4-pole, 3-phase, cage induction motor[10]. Two faulty conditions were considered in their paper: mechanical looseness in the mountings of the motor and unbalanced line voltage supply. Czeslaw et al investigated the diagnosis problems of induction motors in the case of rotor, stator and rolling element bearing faults using multilayer perceptron networks and self-organising Kohonen networks[11], and they applied different networks for different fault detection. By reviewing the past work, it is noted that most research has focused on electric motor fault detection using either BP neural networks or self-organising maps separately for specific faults. Some limited work was conducted by combining BP neural networks and self-organising map for motor fault diagnosis, however, this work focused on applying different networks for different faults in isolation. Therefore, in this work, artificial neural networks, including feedforward backpropagation networks and self-organising maps are applied broadly to motor fault diagnosis. Vibration and stator current indices are used as inputs to artificial neural networks to build a comprehensive automatic motor fault diagnosis system. After being trained, neural networks can automatically store knowledge about the faults or malfunctions in the motors being monitored. By learning from this historical data, the characteristics
1
of associative memory are used to build an ongoing diagnostic database. Finally, the effects of different network structures and training set sizes on the performance of different neural network configurations are discussed.
2. Electric motor faults The reliability of electric motors has been studied for many years. The failures of electric motors are catalogued into mechanical, insulation and magnetic faults. The surveys show that bearing failures cause nearly half of all failures and stator winding failures about 15 to 35%, depending on the application. The combined section of rotor and shaft failures is under 10 percent[12]. The results are shown in Figure 1. From the surveys it is clear that bearing faults, stator winding faults, unbalanced rotor and broken rotor bar are the major fault conditions in electric motors. Thus, this paper focuses on the above most common motor faults which are discussed briefly below. 2.1 Bearing faults Rolling element bearings consist of inner raceways, outer raceways and rolling elements rotating between them. Bearing faults can take place due to fatigue even under normal balanced operation with good shaft alignment, and can also be caused by improper lubrication, installation errors and contamination. One of the results of bearing failures is the increased level of vibration and noise. Vibration frequency domain studies show that when defects exist in a bearing the defects will generate characteristic feature frequencies in the vibration signals. Many publications have discussed the use of these feature frequencies to identify defects in a bearing assembly[13], and these frequencies include ball pass outer raceway frequency (FBPO), ball pass inner raceway frequency (FBPI) and ball spin frequency (FB). Good bearing condition can have FBPO, FBPI and its harmonics, however the amplitude is small and even. When rolling elements, inner raceway and outer raceway defects appear, FB, FBPI, FBPO and their harmonics are exited correspondingly. All of these frequencies can be calculated from bearing geometric dimensions. These frequencies are expressed as follows: N D cosq FBPO = B FS (1- b ) ..........................(1) 2 DC
FBPI =
NB D cosq F (1 + b ) .........................(2) 2 S DC
D cosq 2 D ) ) ......................(3) FB = C FS (1- ( b DC 2DB where, Db is the ball diameter; Dc is the bearing pitch diameter and q is the contact angle of the bearing. Schoen et al[14] have shown that stator current monitoring can detect rolling element bearing faults in induction motors. Line current spectral components are predicted at frequencies of:
Fbng = Fe ± mFV ..................................(4) where Fv is one of the characteristic vibration frequencies; Fe is the supply frequency; m=1,2,3… Although the magnitudes of these harmonic components are small compared to other spectral � constituents, they fall at different locations from those of the supply and machine inherent slot harmonics. This phenomenon makes it feasible to distinguish between healthy and faulty operation. 2.2 Stator winding faults Stator winding faults constitute almost 15-35% of electric motor faults and have a destructive effect on the stator coils. These faults are usually short circuits between a phase winding and the
2
Figure 1. Failure rate of electric motor
ground or between two phases caused by insulation breakdown. This breakdown leads to unbalance in the stator and changes in the current and vibration spectra. Kliman et al[15] observed an increase in the motor phase current leading to shifts in both positive and negative sequence currents. 2.3 Broken rotor bars and unbalanced rotor For the broken rotor bar fault vibration sidebands are expected, which can be seen around the fundamental rotor frequency. The sideband frequencies are given by:
f b = (1 + 2ks) f .................................(5) where k=1,2,3…; s is the motor slip; f is the rotor frequency. The amplitudes and location of the sidebands depend on the physical position of the broken rotor bars, speed and load. When the motor is healthy, no sidebands are visible. With the motor operating under no load, sidebands are not detectable in the vibration spectra whether a rotor bar is broken or not. The reason for this is because of the slip is too small[16]. With increasing load, sidebands appear in the expected locations. Motor current analysis has shown that current spectra are very similar to the vibration spectra for broken rotor bars. When there is no load, the sidebands are not visible as in the vibration spectra. The sidebands will be observed with increasing load. As for an unbalanced rotor, high peaks will show up at the motor running speed and its harmonics in vibration analysis.
3. Experimentation 3.1 Experimental set-up The vibration and current data from electric motor were generated using a Gearbox Fault Simulator(GFS). The construction of the monitoring system is shown in Figure 2. The test motor is a three-phase induction motor with output power of 3 HP (2.24 kW). The vibration signals were measured through vibration sensors at two locations on the motor. One was placed in the vertical direction and the other was set to the axial direction of the motor. One Hall-effect current sensor was used to acquire the stator current. The Gearbox Fault Simulator can create varying and controlled mechanical conditions through motor speed and load control. The signals from all sensors were transmitted to a National Instruments PCI 4474 A/D PC card and sampled at a rate of 1,000 samples/second. Each data set collected contains a 20 second sample time. The experiments for this study included a normal electric motor and electric motors with four main faults (bearing faults, stator winding faults, unbalanced rotor and broken rotor bar). The tests were conducted under 5 speed and 4 load conditions. The speeds were increased from 15 Hz up to 55 Hz with steps of
Insight Vol 46 No 10 October 2004
10 Hz by adjusting the motor drive panel. At each speed four loads were applied to the gearbox output shaft at the percentage of 0, 33%, 66% and 100% of the maximum 70 lb-in (7.9 N-m). 3.2 Signal feature selection and preprocessing Based on the analysis for electric motor faults in section 2 above, the following indices were chosen for the feature inputs for the neural networks. 3.2.1 Vibration signal in the time and frequency domain In order to quantify the vibration signal in the time domain, three statistical parameters are chosen. These parameters are the maximum amplitude Amax, the mean Amean, and the Kurtosis Kf and are defined as: N -1 Amax = max(d(n)) d(n) is time domain data .......(6) n=0 Amean
1 N -1 = Â d(n) ...............................................(7) N n=0
1 N-1 Â (d(n)) 4 N n=0 ..........................................(8) Kf = 1 N -1 ( Â (d(n))2 )2 N n=0
In the frequency domain, three frequencies (FBPO, FBPI, FB) can be applied to determine defects of bearings. Therefore, the amplitudes of the spectrum for frequencies FBPO, FBPI, and FB respectively XFBPO, XFBPI and XFB are utilised. For the broken rotor bar fault, the amplitude of the sideband frequency fb is chosen. For unbalanced rotor, the amplitude of the motor running speed is applied. 3.2.2 Current signal The current signal is sensitive to the broken rotor bar and bearing faults. The amplitude of current sidebands and bearing frequency Fbng are used. All inputs were normalised before being applied to train the networks. Normalisation of data within a uniform range (for example, 0–1) is essential to prevent larger numbers from overwhelming smaller ones and to prevent premature saturation of hidden nodes, which impedes the learning process. This is especially true when actual input data take large values. It is recommended that the data be normalised between slightly offset values such as 0.1 and 0.9 rather than between 0 and 1 to avoid the sigmoid function leading to slow or no learning. The normalisation equation is shown below: c inew = 0.8
c iold - min(c old i ) ......................(9) old + 0.1 max(c old i ) - min( c i )
4. Artificial neural networks Artificial neural networks (ANNs) refer to computing systems designed to simulate the way a biological nervous system operates. Neural networks possess a high level of adaptability that can not be obtained from other completely analytical or numerical procedures. They also provide a data-based heuristic approach to condition monitoring[17]. The choice of network type depends on the nature of the problem to be solved. At present, the two most widely applied ANNs are the multilayer feedforward neural networks trained by backpropagation algorithms (FFBPN) and the Kohonen self-organising map (SOM). In this paper, FFBPN and SOM with different configurations have been evaluated to determine the possible structure of an optimum electric motor fault diagnosis algorithm.
Insight Vol 46 No 10 October 2004
Figure 2. Construction of monitoring system
4.1 Feedforward neural network The multilayer feedforward backpropagation network is based on a supervised procedure. The network constructs a model based on examples of data with known outputs. The network builds the model solely from the examples presented, which are together assumed to implicitly contain the information necessary to establish the relation being sought[18,19]. The determination of the appropriate number of hidden layers and the number of hidden nodes in each layer is one of the most critical tasks in FFBPN design. Unlike the input and output layers, one starts with no prior knowledge as to the number and size of training hidden layers[16]. Too few hidden nodes would be incapable of differentiating between complex patterns. While if the network has too many hidden nodes it will follow the noise in the data due to overparameterisation leading to poor generalisation. In addition, with an increasing number of hidden nodes, the training becomes excessively time-consuming. Since a three-layer neural network is universal in the sense that essentially any function can be implemented to any desired degree of accuracy with sufficient hidden neurons[20], this work focuses on a three-layer neural network. A general three-layer feedforward backpropagation neural network is shown in Figure 3. The network consists of the input layer, the output layer and the hidden layer in between. The hidden layer is used to process and connect the information from the input layer to the output layer only in a forward direction. The hidden layer performs feature extraction on the input data[21]. Each neuron in the hidden layer sums up its input signals after weighting them with the strengths of the respective connections, wnm, and computes an output Zm as a function of the sum. The following equation is used: N
Zm = f ( Âw nm c n ) ...............................(10) 21
n
21 where wnm is the connection weight between the nodes of the input layer and the hidden layer; cn is the input feature data. The outputs of the neuron in the output layer yj are similar:
M
y j = f ( Âw mj Zm ) m
32
(j=1,…,J) .........(11)
In this work, three-layer feedforward networks are modelled with the MatLab Neural Network Toolbox 6.5.1. The networks are trained using a Levenberg-Marquardt algorithm. The activation functions at the hidden layer and the output layer are the hyperbolic tangent function. The initial values of the weights and offsets are randomly assigned. Each model uses 300 epochs for training. During the training phase, the weights are repeatedly adjusted until the calculated outputs have converged sufficiently close to the target output or an iteration limit has been met. The FFBPN
3
Where wik(t +1) is the new value of the kth component of the weight factor vector of neuron i; wikt is the old value of the kth component of the weight factor vector of neuron i; and hci is a scalar gaussian kernel function. 2
hci = a (t)exp(-
ri - rc ) ..........................(13) 2s(t) 2
where a(t) is the training rate factor, s(t) is a factor that implies the size of the effective neighbourhood, and the r’s are the coordinates of the neurons. In this work, SOM networks are also modelled with the MatLab Neural Network Toolbox 6.5.1
5. Results and discussion Figure 3. Feedforward backpropagation neural network
learning rate h and FFBPN momentum coefficient m, were selected by experimentation with different values in the range (0.001,0.005). According to signal feature selection, the inputs for FFBPN are X=[Fs, Load, Amax, Amean, Kf, XFS, XFBPO, XFBPI, XFB, current sideband1, current sideband2, vibration sideband1, vibration sideband2]. There are five outputs corresponding to the four faults described above and a no fault condition. The output goes to 1 if the specific condition exists, otherwise it is zero. Figure 4 illustrates the inputs and outputs of the FFBPN.
�
Figure 4. FFBPN to identity electric motor faults
4.2 Self-organising map The self-organising map (SOM) algorithm was originally proposed by Kohonen on the grounds of biological plausibility. SOM networks are designed primarily for unsupervised learning. Whereas in supervised learning the training data set contains cases featuring input variables together with the associated outputs (and the network must infer a mapping from the inputs to the outputs), in unsupervised learning the training data set contains only input variables. SOM networks attempt to discover some underlying structure of the data. The SOM network can learn to recognise clusters of data, and can also relate similar classes to one another. The user can build up an understanding of the data, which is used to refine the network. As classes of data are recognised, they can be labelled, so that the network becomes capable of classification tasks. A SOM network has only two layers: the input layer, and an output layer of radial units (also known as the topological map layer). The units in the topological map layer are laid out in space typically in two dimensions as used in this work. Training a SOM is an iterative process in which a best matching unit (BMU) must first be found for each data vector. Each data vector must therefore be compared with each neuron on the map in order to find the BMU. A neuron on the map that most closely resembles the current vector is selected as its BMU and the weight factor of the neuron and its neighbouring neurons are adjusted according to the following formula[12]: wik (t + 1) = w ik (t) + hci (t)[c k (t) - w ik (t)] ..............(12)
4
All of the neural networks tested in this work were trained with the same training data. The training data set had 800 samples and the test data set had 200 samples used to check the network response for untrained data which were different from the training data. 5.1 FFBPN In order to evaluate the impact of the neural network architecture on the performance of motor fault diagnosis, different neural networks were selected. The number of hidden neurons varied in the range (15,40), and the two learning rates used were 0.001 and 0.005. Table 1 shows the performance of FFBPN with different network structures. Table 2 shows the performance results where the 13 ¥ 25 ¥ 5 network with the learning rate of 0.005 being implemented, where BE is bearing fault; SW: stator winding fault; UR: unbalanced rotor and BB: broken rotor bars The results demonstrate that with the proper signal feature selection and training procedure, the electric motor fault diagnosis scheme using FFBPN can diagnose motor faults with the desired accuracy. Table 1 also indicated that the accuracy rate for motor fault diagnosis depended on the neural network structure, ranging from 50 to 99.5 percent. It was found that the increase of neurons did not necessarily improve the performance and the 13 ¥ 25 ¥ 5 neural network yielded the best results. In addition, the accuracy rate increased significantly from 55.5%, 79%, 95% up to 99% as the training set size increased as shown in Figure 5. The 13 ¥ 25 ¥ 5 neural network is used in this application. Table 1. Performance of FFBPN with different network structure during test FFBPN structure
Learning rate
Accuracy (%)
FFBPN structure
Learning rate
Accuracy (%)
13×15×5
0.001
97.5
13×15×5
0.005
98.5
13×20×5
0.001
98
13×20×5
0.005
94.5
13×25×5
0.001
99
13×25×5
0.005
99.5
13×30×5
0.001
96
13×30×5
0.005
97
13×35×5
0.001
97.5
13×35×5
0.005
96
13×40×5
0.001
99
13×40×5
0.005
50
Table 2. Performance results of 13¥25¥5 network Conditions
# of samples
Diagnosis results Normal
BE
SW
BB
UR
Accuracy (%)
Normal
40
40
0
0
0
0
BE
40
0
40
0
0
0
100 100
SW
40
0
1
39
0
0
97.5
BB
40
0
0
0
40
0
100
UR
40
0
0
0
0
40
100
Total
200
40
41
39
40
40
99.5
Insight Vol 46 No 10 October 2004
network structures was compared. More time was consumed when the FFBPN was trained as compared to the time for SOM. Acknowledgements Financial support for this work was provided by the Natural Sciences and Engineering Research Council Of Canada (NSERC) and Synrude Canada Ltd. References Figure 5. Effects of the training set sizes on the accuracy rate
5.2 SOM Two-dimensional SOM networks with various numbers of neurons: 10 ¥ 10, 15 ¥ 15 and 20 ¥ 20 were modelled. In all cases the rectangle network topology was used and the inputs of the SOM were the same as the ones used for the FFBPN. The training epoch is 1000. In Figure 6 the responses from SOM network 10 ¥ 10 are demonstrated. The shapes mean normal, bearing fault, stator winding fault, unbalanced rotor and broken rotor bars respectively. From Figure 6, it is clear that the SOM has not clustered the different faults into distinct regions for normal and fault motors. Speed and load affected the amplitudes and location of the sidebands of the normal and faulty motors. Figure 7 compares the influence of speed on sideband amplitudes at 100% load condition. Figure 7(a) and 7(b) show that the normal and faulty vibration sideband amplitudes increased with speed, and the vibration sideband amplitudes for a broken rotor bar were higher than normal sideband amplitudes at the same speed. On the other hand, the current sideband amplitudes remain almost constant for all speeds stated as shown in Figure 7(c). The comparisons of sideband amplitudes at other load conditions show similar trends against the speed. Therefore, SOM classification was tested at different speed and load levels. Figure 8 shows a Kohonen feature map at a speed of 55 Hz and 100% load . It can be seen that in this case the network separated the characteristic regions including normal, bearing fault, stator winding fault, unbalanced rotor and broken rotor bars into different locations on the feature map. The other Kohonen feature maps showed similar results at different speeds and loads. From the above experiment results it is shown that neural networks are effective tools for electric motor fault diagnosis. The feedforward ANN is the more promising scheme in this case where fault data of electric motors are available, while SOM can give good classification during steady state of working conditions. However, it is difficult for SOM to cluster the different faults at the working conditions of variable speeds and loads.
6. Conclusion We can draw the following conclusions based on the above experimental results. n Artificial neural networks are effective for electric motor fault diagnosis. A feedforward NN with very simple internal structure can give satisfactory results, and an accuracy rate up to 99.5 percent; SOM can classify the type of motor faults during steady state working conditions. n The performance of different neural networks were compared. The accuracy rate depended on the neural network architecture. Furthermore, the increase of neurons did not necessarily improve the performance. n The performance of the neural network technique depended on the training set size. The larger the training set size, the better the diagnosis performance. n The time used for the training process of different neural
Insight Vol 46 No 10 October 2004
1. R A Hall, P F Knights and L K Daneshmend, ‘Pareto analysis and condition-based maintenance of underground mining equipment’, Institution of Mining and Metallurgy, pp109, January-April 2000. 2. http://www.syncrude.com/research/04_06.html 3. Mohamed A Awadallah and Medhat M Morco, ‘Application of AI tools in fault diagnosis of electrical machines and drives – an overview’, IEEE Transaction on Energy Conversation, Vol 18, No 2, pp245, June 2003. 4. T B Breen, G B Kliman and S C Patel, ‘New developments in non-invasive on-line motor diagnostics’, IEEE Proceedings. Petroleum and Chemical Industry Conference, pp231-236, September 1996. 5. I A Basheer and M Hajmeer, ‘Artificial neural networks: fundamentals, computing, design, and application’ Journal of Microbiological Methods 43, pp3–31, 2000. 6. M-Y Chow and S O Yee, ‘Using neural networks to detect incipient faults in Induction motors’, Journal of Neural Network Computing 2, pp26-32, 1991. 7. M-Y Chow and S O Yee, ‘Methodology for on-line incipient fault detection in single-phase squirrel-cage induction motors using artificial neural networks’, IEEE Transactions on Energy Conversion 6, pp536-545, 1991. 8. M-Y Chow, R N Sharpe and J C Hung, ‘On the application and design of artificial neural networks for motor fault detection’, IEEE Transactions on Industrial Electronics 40, pp181-196, 1993. 9. Sri Kolla and Logan Varatharasa, ‘Identifying three-phase induction motor faults using artificial neural networks’, ISA Transactions 39, pp433-439, 2000. 10. J Penman and C M Yin, ‘Feasibility of using unsupervised learning artificial neural networks for the condition monitoring of electrical machines’, IEE Proceedings. B 141, pp317–322, 1994. 11. Czeslaw T Kowalski, Teresa Orlowska-Kowalska, ‘Neural networks application for induction motor faults diagnosis’, Mathematics and Computers in Simulation 63, pp 435–448, 2003. 12. Voitto Kokko, PhD thesis, ‘Condition monitoring of squirrelcage motors by axial magnetic flux measurements’, University of Oulu, 2003. 13. Bo Li, Mo-Yuen Chow, Yodyium Tipsuwan and James C Hung, ‘Neural-network-based motor rolling bearing fault diagnosis’, IEEE Transactions on Industrial Electronics, Vol 47, No 5, pp1060-1068, 2000. 14. R R Schoen, T G Habetler, F Kamran and R G Bartheld, ‘Motor bearing damage detection using stator current monitoring’, IEEE Transactions on Industrial Application, Vol 31, No 6, pp1274-1279, November.-December, 1995. 15. G B Klimen, W J Premerlanri, R A Koegl and D Hoeweler, ‘A new approach to on-line turn fault detection in AC motors’, IEEE IAS Annu. meeting, pp687-693, 1996. 16. B Liang, S D Iwnicki and A D Ball, ‘Asymmetrical stator and rotor faulty detection using vibration, phase current and transient speed analysis’, Mechanical Systems and Signal Processing 17(4), pp857–869, 2003.
5
17. J Tranter, ‘The fundamentals of, and the applicaton of computers to condition monitoring and predictive maintenance’, Proceeding of the 1st International Machinery Monitoring and Diagnosis Conference, Las Vegas, pp394-401, 1989. 18. D E Rumelhart, G E Hinton and R J Williams, ‘Learning representations by backpropagating errors’, Nature, 323, pp533-536, 1986. 19. R P Lippman, ‘An introduction to computing with neural nets’, IEEE ASSP Magazine, pp4-22, April 1987. 20. T I Liu and J M Mengel, ‘Intelligent monitoring of ball bearing conditions’, Mechanical Systems and Signal Processing 6(5), pp419-431, 1992. 21. A K Nandi, ‘Vibration based fault detection-features, classifiers and novelty detection’, Condition Monitoring and Diagnostic Engineering Management, COMADEM International, 2002.
Figure 6. Two-dimensional Kohonen feature map 10¥10
Figure 8. Two-dimensional 10¥10 Kohonen feature map (speed 55 Hz and 100% load)
Figure 7. Influence of speed on the left sideband amplitudes at 100% load (a - radial vibration, b - axial vibration, c – current)
6
Editor’s Note: In Figures 6 and 8, the shapes mean normal, bearing fault, stator winding fault, unbalanced rotor and broken rotor bars respectively.
Insight Vol 46 No 10 October 2004
Electric motor faults diagnosis using artificial neural networks Lingxin Li, C K Mechefske and Weidong Li
This paper presents electric motor fault diagnosis using two kinds of Artificial Neural Networks (ANN): feedforward networks and self organising maps (SOM). Major faults such as bearing faults, stator winding fault, unbalanced rotor and broken rotor bars are considered. The ANNs were trained and tested using measurement data from stator currents and mechanical vibration signals. The effects of different network structures and the training set sizes on the performance of the ANNs are discussed. This study shows that the feedforward ANN with a very simple internal structure can give satisfactory results, while SOMs can classify the type of motor faults during steady state working conditions. The experiment results also show that the feedforward ANN is the more promising scheme in this case where fault data from electric motors is available.
1. Introduction With an increasing trend towards the need for mechanisation, automation and high capacity mining equipment, it has become necessary that the reliability of equipment be improved. It has been found that maintenance accounts for about 30-60% of the total operating cost budget for a typical mining company and represents the largest portion of the mine’s controllable operating costs[1]. In addition, because of the nature of mining equipment, the failures of mining equipment may significantly reduce and even stop production. This, coupled with the fact that the mining industry today is faced with an ever more complex and demanding marketplace, has led to an increasing interest in methods to reduce maintenance costs. Syncrude Canada Ltd. is the world’s largest producer of crude oil from oil sands and the largest single source producer in Canada, and currently supplies 13 percent of Canada’s petroleum requirements. At Syncrude, a truck and shovel system is used to mine and deliver the raw oil sand to a central separator[2]. The shovel is primarily an excavating and loading device and its reliability and efficiency makes a notable impact on the mining productivity and the plant-wide profits. Electric motors are critical components in the P&H 4100 type cable shovel. The ever changing mining environment and dynamical loading conditions always strain and wear motors and cause faults. Electric motor manufacturers and users have relied for a long time on protective relays – such as circuit breakers or fuses to monitor faults and disconnect motors. However, following this traditional scheme led to facing some situations where the machine was badly damaged, mainly due to the machine tripping after the fault was well developed[3]. This scheme can not provide warning of impending faults before they occur. Incipient fault detection for electric motors would not only be highly cost-effective in The authors are in the Department of Mechanical and Materials Engineering, Queens’ University, Kingston, Ontario, Canada, K7L 3N6. E-mail: [email protected]; [email protected]; [email protected]
Insight Vol 46 No 10 October 2004
minimising maintenance downtime, but it would also allow preventive maintenance to be scheduled. The work reported in this paper is part of a larger research project: Electric Motor Fault Diagnosis Systems for P&H4100 Shovels at Syncrude Canada Ltd. The primary focus of the work is to employ appropriate technology to monitor and detect normal and faulty electric motors and to provide a warning and diagnose the faults at an early stage. There already exist several monitoring methods for electric motors. Stator current monitoring is one of the most widely employed monitoring techniques[4]. This monitoring is believed to be most sensitive to motor component faults and can provide unique fault patterns. In addition to stator current analysis, vibration monitoring can also provide effective indications especially for motor bearing and broken rotor bar faults. The use of artificial neural networks (ANN) for processing stator current and motor vibration signals is particularly attractive due to the remarkable information processing characteristics such as nonlinearity, robustness, ability to learn and ability to handle imprecise and fuzzy information[5]. The use of artificial neural networks for electric motor fault detection was first proposed by Chow et al[6-8]. They proposed a typical back-propagation (BP) neural network structure for incipient motor fault diagnosis. The incipient faults focused on turn-to-turn insulation faults and bearing wear. Kolla et al[9] investigated motor faults including overloads, single phasing, unbalanced supply voltage, locked rotor, ground fault, over-voltage and under-voltage using detection methods based on the feedforward neural network. Besides BP neural networks, other unsupervised learning neural networks were also used in motor fault diagnosis. Penman and Jin applied the Kohonen’s self-organising map (SOM) to classify multiple faults in a 3 kW, 4-pole, 3-phase, cage induction motor[10]. Two faulty conditions were considered in their paper: mechanical looseness in the mountings of the motor and unbalanced line voltage supply. Czeslaw et al investigated the diagnosis problems of induction motors in the case of rotor, stator and rolling element bearing faults using multilayer perceptron networks and self-organising Kohonen networks[11], and they applied different networks for different fault detection. By reviewing the past work, it is noted that most research has focused on electric motor fault detection using either BP neural networks or self-organising maps separately for specific faults. Some limited work was conducted by combining BP neural networks and self-organising map for motor fault diagnosis, however, this work focused on applying different networks for different faults in isolation. Therefore, in this work, artificial neural networks, including feedforward backpropagation networks and self-organising maps are applied broadly to motor fault diagnosis. Vibration and stator current indices are used as inputs to artificial neural networks to build a comprehensive automatic motor fault diagnosis system. After being trained, neural networks can automatically store knowledge about the faults or malfunctions in the motors being monitored. By learning from this historical data, the characteristics
1
of associative memory are used to build an ongoing diagnostic database. Finally, the effects of different network structures and training set sizes on the performance of different neural network configurations are discussed.
2. Electric motor faults The reliability of electric motors has been studied for many years. The failures of electric motors are catalogued into mechanical, insulation and magnetic faults. The surveys show that bearing failures cause nearly half of all failures and stator winding failures about 15 to 35%, depending on the application. The combined section of rotor and shaft failures is under 10 percent[12]. The results are shown in Figure 1. From the surveys it is clear that bearing faults, stator winding faults, unbalanced rotor and broken rotor bar are the major fault conditions in electric motors. Thus, this paper focuses on the above most common motor faults which are discussed briefly below. 2.1 Bearing faults Rolling element bearings consist of inner raceways, outer raceways and rolling elements rotating between them. Bearing faults can take place due to fatigue even under normal balanced operation with good shaft alignment, and can also be caused by improper lubrication, installation errors and contamination. One of the results of bearing failures is the increased level of vibration and noise. Vibration frequency domain studies show that when defects exist in a bearing the defects will generate characteristic feature frequencies in the vibration signals. Many publications have discussed the use of these feature frequencies to identify defects in a bearing assembly[13], and these frequencies include ball pass outer raceway frequency (FBPO), ball pass inner raceway frequency (FBPI) and ball spin frequency (FB). Good bearing condition can have FBPO, FBPI and its harmonics, however the amplitude is small and even. When rolling elements, inner raceway and outer raceway defects appear, FB, FBPI, FBPO and their harmonics are exited correspondingly. All of these frequencies can be calculated from bearing geometric dimensions. These frequencies are expressed as follows: N D cosq FBPO = B FS (1- b ) ..........................(1) 2 DC
FBPI =
NB D cosq F (1 + b ) .........................(2) 2 S DC
D cosq 2 D ) ) ......................(3) FB = C FS (1- ( b DC 2DB where, Db is the ball diameter; Dc is the bearing pitch diameter and q is the contact angle of the bearing. Schoen et al[14] have shown that stator current monitoring can detect rolling element bearing faults in induction motors. Line current spectral components are predicted at frequencies of:
Fbng = Fe ± mFV ..................................(4) where Fv is one of the characteristic vibration frequencies; Fe is the supply frequency; m=1,2,3… Although the magnitudes of these harmonic components are small compared to other spectral � constituents, they fall at different locations from those of the supply and machine inherent slot harmonics. This phenomenon makes it feasible to distinguish between healthy and faulty operation. 2.2 Stator winding faults Stator winding faults constitute almost 15-35% of electric motor faults and have a destructive effect on the stator coils. These faults are usually short circuits between a phase winding and the
2
Figure 1. Failure rate of electric motor
ground or between two phases caused by insulation breakdown. This breakdown leads to unbalance in the stator and changes in the current and vibration spectra. Kliman et al[15] observed an increase in the motor phase current leading to shifts in both positive and negative sequence currents. 2.3 Broken rotor bars and unbalanced rotor For the broken rotor bar fault vibration sidebands are expected, which can be seen around the fundamental rotor frequency. The sideband frequencies are given by:
f b = (1 + 2ks) f .................................(5) where k=1,2,3…; s is the motor slip; f is the rotor frequency. The amplitudes and location of the sidebands depend on the physical position of the broken rotor bars, speed and load. When the motor is healthy, no sidebands are visible. With the motor operating under no load, sidebands are not detectable in the vibration spectra whether a rotor bar is broken or not. The reason for this is because of the slip is too small[16]. With increasing load, sidebands appear in the expected locations. Motor current analysis has shown that current spectra are very similar to the vibration spectra for broken rotor bars. When there is no load, the sidebands are not visible as in the vibration spectra. The sidebands will be observed with increasing load. As for an unbalanced rotor, high peaks will show up at the motor running speed and its harmonics in vibration analysis.
3. Experimentation 3.1 Experimental set-up The vibration and current data from electric motor were generated using a Gearbox Fault Simulator(GFS). The construction of the monitoring system is shown in Figure 2. The test motor is a three-phase induction motor with output power of 3 HP (2.24 kW). The vibration signals were measured through vibration sensors at two locations on the motor. One was placed in the vertical direction and the other was set to the axial direction of the motor. One Hall-effect current sensor was used to acquire the stator current. The Gearbox Fault Simulator can create varying and controlled mechanical conditions through motor speed and load control. The signals from all sensors were transmitted to a National Instruments PCI 4474 A/D PC card and sampled at a rate of 1,000 samples/second. Each data set collected contains a 20 second sample time. The experiments for this study included a normal electric motor and electric motors with four main faults (bearing faults, stator winding faults, unbalanced rotor and broken rotor bar). The tests were conducted under 5 speed and 4 load conditions. The speeds were increased from 15 Hz up to 55 Hz with steps of
Insight Vol 46 No 10 October 2004
10 Hz by adjusting the motor drive panel. At each speed four loads were applied to the gearbox output shaft at the percentage of 0, 33%, 66% and 100% of the maximum 70 lb-in (7.9 N-m). 3.2 Signal feature selection and preprocessing Based on the analysis for electric motor faults in section 2 above, the following indices were chosen for the feature inputs for the neural networks. 3.2.1 Vibration signal in the time and frequency domain In order to quantify the vibration signal in the time domain, three statistical parameters are chosen. These parameters are the maximum amplitude Amax, the mean Amean, and the Kurtosis Kf and are defined as: N -1 Amax = max(d(n)) d(n) is time domain data .......(6) n=0 Amean
1 N -1 = Â d(n) ...............................................(7) N n=0
1 N-1 Â (d(n)) 4 N n=0 ..........................................(8) Kf = 1 N -1 ( Â (d(n))2 )2 N n=0
In the frequency domain, three frequencies (FBPO, FBPI, FB) can be applied to determine defects of bearings. Therefore, the amplitudes of the spectrum for frequencies FBPO, FBPI, and FB respectively XFBPO, XFBPI and XFB are utilised. For the broken rotor bar fault, the amplitude of the sideband frequency fb is chosen. For unbalanced rotor, the amplitude of the motor running speed is applied. 3.2.2 Current signal The current signal is sensitive to the broken rotor bar and bearing faults. The amplitude of current sidebands and bearing frequency Fbng are used. All inputs were normalised before being applied to train the networks. Normalisation of data within a uniform range (for example, 0–1) is essential to prevent larger numbers from overwhelming smaller ones and to prevent premature saturation of hidden nodes, which impedes the learning process. This is especially true when actual input data take large values. It is recommended that the data be normalised between slightly offset values such as 0.1 and 0.9 rather than between 0 and 1 to avoid the sigmoid function leading to slow or no learning. The normalisation equation is shown below: c inew = 0.8
c iold - min(c old i ) ......................(9) old + 0.1 max(c old i ) - min( c i )
4. Artificial neural networks Artificial neural networks (ANNs) refer to computing systems designed to simulate the way a biological nervous system operates. Neural networks possess a high level of adaptability that can not be obtained from other completely analytical or numerical procedures. They also provide a data-based heuristic approach to condition monitoring[17]. The choice of network type depends on the nature of the problem to be solved. At present, the two most widely applied ANNs are the multilayer feedforward neural networks trained by backpropagation algorithms (FFBPN) and the Kohonen self-organising map (SOM). In this paper, FFBPN and SOM with different configurations have been evaluated to determine the possible structure of an optimum electric motor fault diagnosis algorithm.
Insight Vol 46 No 10 October 2004
Figure 2. Construction of monitoring system
4.1 Feedforward neural network The multilayer feedforward backpropagation network is based on a supervised procedure. The network constructs a model based on examples of data with known outputs. The network builds the model solely from the examples presented, which are together assumed to implicitly contain the information necessary to establish the relation being sought[18,19]. The determination of the appropriate number of hidden layers and the number of hidden nodes in each layer is one of the most critical tasks in FFBPN design. Unlike the input and output layers, one starts with no prior knowledge as to the number and size of training hidden layers[16]. Too few hidden nodes would be incapable of differentiating between complex patterns. While if the network has too many hidden nodes it will follow the noise in the data due to overparameterisation leading to poor generalisation. In addition, with an increasing number of hidden nodes, the training becomes excessively time-consuming. Since a three-layer neural network is universal in the sense that essentially any function can be implemented to any desired degree of accuracy with sufficient hidden neurons[20], this work focuses on a three-layer neural network. A general three-layer feedforward backpropagation neural network is shown in Figure 3. The network consists of the input layer, the output layer and the hidden layer in between. The hidden layer is used to process and connect the information from the input layer to the output layer only in a forward direction. The hidden layer performs feature extraction on the input data[21]. Each neuron in the hidden layer sums up its input signals after weighting them with the strengths of the respective connections, wnm, and computes an output Zm as a function of the sum. The following equation is used: N
Zm = f ( Âw nm c n ) ...............................(10) 21
n
21 where wnm is the connection weight between the nodes of the input layer and the hidden layer; cn is the input feature data. The outputs of the neuron in the output layer yj are similar:
M
y j = f ( Âw mj Zm ) m
32
(j=1,…,J) .........(11)
In this work, three-layer feedforward networks are modelled with the MatLab Neural Network Toolbox 6.5.1. The networks are trained using a Levenberg-Marquardt algorithm. The activation functions at the hidden layer and the output layer are the hyperbolic tangent function. The initial values of the weights and offsets are randomly assigned. Each model uses 300 epochs for training. During the training phase, the weights are repeatedly adjusted until the calculated outputs have converged sufficiently close to the target output or an iteration limit has been met. The FFBPN
3
Where wik(t +1) is the new value of the kth component of the weight factor vector of neuron i; wikt is the old value of the kth component of the weight factor vector of neuron i; and hci is a scalar gaussian kernel function. 2
hci = a (t)exp(-
ri - rc ) ..........................(13) 2s(t) 2
where a(t) is the training rate factor, s(t) is a factor that implies the size of the effective neighbourhood, and the r’s are the coordinates of the neurons. In this work, SOM networks are also modelled with the MatLab Neural Network Toolbox 6.5.1
5. Results and discussion Figure 3. Feedforward backpropagation neural network
learning rate h and FFBPN momentum coefficient m, were selected by experimentation with different values in the range (0.001,0.005). According to signal feature selection, the inputs for FFBPN are X=[Fs, Load, Amax, Amean, Kf, XFS, XFBPO, XFBPI, XFB, current sideband1, current sideband2, vibration sideband1, vibration sideband2]. There are five outputs corresponding to the four faults described above and a no fault condition. The output goes to 1 if the specific condition exists, otherwise it is zero. Figure 4 illustrates the inputs and outputs of the FFBPN.
�
Figure 4. FFBPN to identity electric motor faults
4.2 Self-organising map The self-organising map (SOM) algorithm was originally proposed by Kohonen on the grounds of biological plausibility. SOM networks are designed primarily for unsupervised learning. Whereas in supervised learning the training data set contains cases featuring input variables together with the associated outputs (and the network must infer a mapping from the inputs to the outputs), in unsupervised learning the training data set contains only input variables. SOM networks attempt to discover some underlying structure of the data. The SOM network can learn to recognise clusters of data, and can also relate similar classes to one another. The user can build up an understanding of the data, which is used to refine the network. As classes of data are recognised, they can be labelled, so that the network becomes capable of classification tasks. A SOM network has only two layers: the input layer, and an output layer of radial units (also known as the topological map layer). The units in the topological map layer are laid out in space typically in two dimensions as used in this work. Training a SOM is an iterative process in which a best matching unit (BMU) must first be found for each data vector. Each data vector must therefore be compared with each neuron on the map in order to find the BMU. A neuron on the map that most closely resembles the current vector is selected as its BMU and the weight factor of the neuron and its neighbouring neurons are adjusted according to the following formula[12]: wik (t + 1) = w ik (t) + hci (t)[c k (t) - w ik (t)] ..............(12)
4
All of the neural networks tested in this work were trained with the same training data. The training data set had 800 samples and the test data set had 200 samples used to check the network response for untrained data which were different from the training data. 5.1 FFBPN In order to evaluate the impact of the neural network architecture on the performance of motor fault diagnosis, different neural networks were selected. The number of hidden neurons varied in the range (15,40), and the two learning rates used were 0.001 and 0.005. Table 1 shows the performance of FFBPN with different network structures. Table 2 shows the performance results where the 13 ¥ 25 ¥ 5 network with the learning rate of 0.005 being implemented, where BE is bearing fault; SW: stator winding fault; UR: unbalanced rotor and BB: broken rotor bars The results demonstrate that with the proper signal feature selection and training procedure, the electric motor fault diagnosis scheme using FFBPN can diagnose motor faults with the desired accuracy. Table 1 also indicated that the accuracy rate for motor fault diagnosis depended on the neural network structure, ranging from 50 to 99.5 percent. It was found that the increase of neurons did not necessarily improve the performance and the 13 ¥ 25 ¥ 5 neural network yielded the best results. In addition, the accuracy rate increased significantly from 55.5%, 79%, 95% up to 99% as the training set size increased as shown in Figure 5. The 13 ¥ 25 ¥ 5 neural network is used in this application. Table 1. Performance of FFBPN with different network structure during test FFBPN structure
Learning rate
Accuracy (%)
FFBPN structure
Learning rate
Accuracy (%)
13×15×5
0.001
97.5
13×15×5
0.005
98.5
13×20×5
0.001
98
13×20×5
0.005
94.5
13×25×5
0.001
99
13×25×5
0.005
99.5
13×30×5
0.001
96
13×30×5
0.005
97
13×35×5
0.001
97.5
13×35×5
0.005
96
13×40×5
0.001
99
13×40×5
0.005
50
Table 2. Performance results of 13¥25¥5 network Conditions
# of samples
Diagnosis results Normal
BE
SW
BB
UR
Accuracy (%)
Normal
40
40
0
0
0
0
BE
40
0
40
0
0
0
100 100
SW
40
0
1
39
0
0
97.5
BB
40
0
0
0
40
0
100
UR
40
0
0
0
0
40
100
Total
200
40
41
39
40
40
99.5
Insight Vol 46 No 10 October 2004
network structures was compared. More time was consumed when the FFBPN was trained as compared to the time for SOM. Acknowledgements Financial support for this work was provided by the Natural Sciences and Engineering Research Council Of Canada (NSERC) and Synrude Canada Ltd. References Figure 5. Effects of the training set sizes on the accuracy rate
5.2 SOM Two-dimensional SOM networks with various numbers of neurons: 10 ¥ 10, 15 ¥ 15 and 20 ¥ 20 were modelled. In all cases the rectangle network topology was used and the inputs of the SOM were the same as the ones used for the FFBPN. The training epoch is 1000. In Figure 6 the responses from SOM network 10 ¥ 10 are demonstrated. The shapes mean normal, bearing fault, stator winding fault, unbalanced rotor and broken rotor bars respectively. From Figure 6, it is clear that the SOM has not clustered the different faults into distinct regions for normal and fault motors. Speed and load affected the amplitudes and location of the sidebands of the normal and faulty motors. Figure 7 compares the influence of speed on sideband amplitudes at 100% load condition. Figure 7(a) and 7(b) show that the normal and faulty vibration sideband amplitudes increased with speed, and the vibration sideband amplitudes for a broken rotor bar were higher than normal sideband amplitudes at the same speed. On the other hand, the current sideband amplitudes remain almost constant for all speeds stated as shown in Figure 7(c). The comparisons of sideband amplitudes at other load conditions show similar trends against the speed. Therefore, SOM classification was tested at different speed and load levels. Figure 8 shows a Kohonen feature map at a speed of 55 Hz and 100% load . It can be seen that in this case the network separated the characteristic regions including normal, bearing fault, stator winding fault, unbalanced rotor and broken rotor bars into different locations on the feature map. The other Kohonen feature maps showed similar results at different speeds and loads. From the above experiment results it is shown that neural networks are effective tools for electric motor fault diagnosis. The feedforward ANN is the more promising scheme in this case where fault data of electric motors are available, while SOM can give good classification during steady state of working conditions. However, it is difficult for SOM to cluster the different faults at the working conditions of variable speeds and loads.
6. Conclusion We can draw the following conclusions based on the above experimental results. n Artificial neural networks are effective for electric motor fault diagnosis. A feedforward NN with very simple internal structure can give satisfactory results, and an accuracy rate up to 99.5 percent; SOM can classify the type of motor faults during steady state working conditions. n The performance of different neural networks were compared. The accuracy rate depended on the neural network architecture. Furthermore, the increase of neurons did not necessarily improve the performance. n The performance of the neural network technique depended on the training set size. The larger the training set size, the better the diagnosis performance. n The time used for the training process of different neural
Insight Vol 46 No 10 October 2004
1. R A Hall, P F Knights and L K Daneshmend, ‘Pareto analysis and condition-based maintenance of underground mining equipment’, Institution of Mining and Metallurgy, pp109, January-April 2000. 2. http://www.syncrude.com/research/04_06.html 3. Mohamed A Awadallah and Medhat M Morco, ‘Application of AI tools in fault diagnosis of electrical machines and drives – an overview’, IEEE Transaction on Energy Conversation, Vol 18, No 2, pp245, June 2003. 4. T B Breen, G B Kliman and S C Patel, ‘New developments in non-invasive on-line motor diagnostics’, IEEE Proceedings. Petroleum and Chemical Industry Conference, pp231-236, September 1996. 5. I A Basheer and M Hajmeer, ‘Artificial neural networks: fundamentals, computing, design, and application’ Journal of Microbiological Methods 43, pp3–31, 2000. 6. M-Y Chow and S O Yee, ‘Using neural networks to detect incipient faults in Induction motors’, Journal of Neural Network Computing 2, pp26-32, 1991. 7. M-Y Chow and S O Yee, ‘Methodology for on-line incipient fault detection in single-phase squirrel-cage induction motors using artificial neural networks’, IEEE Transactions on Energy Conversion 6, pp536-545, 1991. 8. M-Y Chow, R N Sharpe and J C Hung, ‘On the application and design of artificial neural networks for motor fault detection’, IEEE Transactions on Industrial Electronics 40, pp181-196, 1993. 9. Sri Kolla and Logan Varatharasa, ‘Identifying three-phase induction motor faults using artificial neural networks’, ISA Transactions 39, pp433-439, 2000. 10. J Penman and C M Yin, ‘Feasibility of using unsupervised learning artificial neural networks for the condition monitoring of electrical machines’, IEE Proceedings. B 141, pp317–322, 1994. 11. Czeslaw T Kowalski, Teresa Orlowska-Kowalska, ‘Neural networks application for induction motor faults diagnosis’, Mathematics and Computers in Simulation 63, pp 435–448, 2003. 12. Voitto Kokko, PhD thesis, ‘Condition monitoring of squirrelcage motors by axial magnetic flux measurements’, University of Oulu, 2003. 13. Bo Li, Mo-Yuen Chow, Yodyium Tipsuwan and James C Hung, ‘Neural-network-based motor rolling bearing fault diagnosis’, IEEE Transactions on Industrial Electronics, Vol 47, No 5, pp1060-1068, 2000. 14. R R Schoen, T G Habetler, F Kamran and R G Bartheld, ‘Motor bearing damage detection using stator current monitoring’, IEEE Transactions on Industrial Application, Vol 31, No 6, pp1274-1279, November.-December, 1995. 15. G B Klimen, W J Premerlanri, R A Koegl and D Hoeweler, ‘A new approach to on-line turn fault detection in AC motors’, IEEE IAS Annu. meeting, pp687-693, 1996. 16. B Liang, S D Iwnicki and A D Ball, ‘Asymmetrical stator and rotor faulty detection using vibration, phase current and transient speed analysis’, Mechanical Systems and Signal Processing 17(4), pp857–869, 2003.
5
17. J Tranter, ‘The fundamentals of, and the applicaton of computers to condition monitoring and predictive maintenance’, Proceeding of the 1st International Machinery Monitoring and Diagnosis Conference, Las Vegas, pp394-401, 1989. 18. D E Rumelhart, G E Hinton and R J Williams, ‘Learning representations by backpropagating errors’, Nature, 323, pp533-536, 1986. 19. R P Lippman, ‘An introduction to computing with neural nets’, IEEE ASSP Magazine, pp4-22, April 1987. 20. T I Liu and J M Mengel, ‘Intelligent monitoring of ball bearing conditions’, Mechanical Systems and Signal Processing 6(5), pp419-431, 1992. 21. A K Nandi, ‘Vibration based fault detection-features, classifiers and novelty detection’, Condition Monitoring and Diagnostic Engineering Management, COMADEM International, 2002.
Figure 6. Two-dimensional Kohonen feature map 10¥10
Figure 8. Two-dimensional 10¥10 Kohonen feature map (speed 55 Hz and 100% load)
Figure 7. Influence of speed on the left sideband amplitudes at 100% load (a - radial vibration, b - axial vibration, c – current)
6
Editor’s Note: In Figures 6 and 8, the shapes mean normal, bearing fault, stator winding fault, unbalanced rotor and broken rotor bars respectively.
Insight Vol 46 No 10 October 2004
