Abstract
The intention of fault detection is to detect the fault at the beginning stage and shut off the machine immediately to avoid motor failure due to the large fault current. In this work, an online fault diagnosis of stator interturn fault of a three-phase induction motor based on the concept of symmetrical components is presented. A mathematical model of an induction motor with turn fault is developed to interpret machine performance under fault. A Simulink model of a three-phase induction motor with stator interturn fault is created for extraction of sequence components of current and voltage. The negative sequence current can provide a decisive and rapid monitoring technique to detect stator interturn short circuit fault of the induction motor. The per unit change in negative sequence current with positive sequence current is the main fault indicator which is imported to neural network architecture. The output of the feedforward backpropagation neural network classifies the short circuit fault level of stator winding.
1. Introduction
Induction motors overpower the field of electromechanical energy conversion. Their reliability, low cost, and high performance make them the most popular alternating current motors. These motors have the flexibility of being applied in various fields, from household appliances to high power industrial motors. In recent years, the problems of failure in large induction motors have become more significant. For the fault diagnosis problem, it is important to detect if the system has a fault and find its origin [1]. If motor failure is unattended at an early stage, it may become detrimental and damage the motor. This will cause industry production cessation.
In [2], many fault situations are mentioned. One among them is a case in which a disintegrated rotor bar had erupted from the slot and created damage in the stator winding. Faults in the induction motor can be either mechanical or electrical. Major mechanical faults are bearing fault [3–5] and broken rotor bar [6–10]. Electrical fault is influenced by power quality supplied by AC grid, frequency variations, voltage disturbances, and load variation. Another fault is short circuits stator winding [3, 11–14]. About more than one-third of the total fault occurring in the induction motor is that of stator winding fault. Short circuit in the stator winding takes very short time to evolve and totally damage the motor. Normally, the interturn short circuit fault progresses to an intercoil fault, phase winding fault, and single line ground fault, resulting in the breakdown of the motor. Detection of winding fault in the starting stage increases the practicability of repairing the machine by rewinding it or, in large motors, displacing short-circuited coils.
Traditional ways of fault monitoring were in regard to the leakage flux sensing [15], partial discharge [16], harmonics in stator current and voltage [17], etc. Ensuing studies, however, demonstrated that many of these traditional techniques are liable to oddity due to supply voltage distortions [18], built in machine asymmetries [19], coincidental effects of the stator and rotor faults, etc. Motor current signature analysis (MCSA) is an important technique adopted for condition monitoring. Induction motor faults like bearing problems, broken rotor bar, eccentricity abnormalities, and stator winding faults cause variation in amplitude and frequency of motor current signature [3–9, 11–14].
A breakthrough in signal conditioning techniques and advances in computer software has taken machine fault detection to newer heights. Most of the demonstrated work in stator winding fault detection is from a frequency analysis domain. Signal transforming methods like fast Fourier transform (FFT), S-transform, short-time Fourier transform (STFT), wavelet transform, and Hilbert transforms have been adopted in combination with various classification techniques such as expert systems, artificial neural network, fuzzy logic, and support vector machine [20–26] for the motor degradation.
Eminent interest has been shown in [27–29] for artificial neural network in fault detection of the induction motor. The necessary condition to contrivance a successful ANN classifier is the selection of appropriate inputs about each case of fault. In [27, 28], stator interturn fault location detection by ANN is described considering frequency domain parameters as selected input.
In [29], ANN is applied for detecting the severity of the intercoil fault with the selected parameter from time domain. The work in [30–33] focuses on negative sequence current which is caused due to unbalanced windings.
Most of the research work in induction motor stator winding fault detection is from the analysis of the frequency domain. If we choose line currents or line voltages as parameters considered for fault detection, time domain analysis is also equally efficient. This will avoid usage of the spectral analyzer and sophisticated signal conditioning techniques, which makes the fault detection system unit much simpler. The aim is to identify a universal diagnostic technique to detect the winding fault and its severity level without motor constructional data and with knowledge of fault parameters from the time domain analysis.
In this work, we attempt to find a detection technique for stator winding fault based on per unit value of sequence components of current from time domain and classify the severity of fault by using artificial neural network. Here, network has been trained with a full range of input vector obtained from a Simulink model. NN input vector contains experimental values up to a feasible range and Simulink values to complete the input set. This provides a well-trained network. Simulation values are in close agreement with experimental values. Fault detection in the initial stage increases the viability of repairing the machine, and a nip in the bud avoids electrical spark and explosion.
Next session details mathematical modeling of the induction motor with stator winding fault. Using these mathematical equations, the model is created in Simulink which is described in Section 3. The classification method used for fault detection is emphasized in Section 4.
2. Mathematical Modeling of Stator Winding Turn Fault
A three-phase induction motor with turn fault in one-phase stator winding is considered, where β is the fraction of shorted turns. Winding in this phase has two parts—shorted turns and unfaulted turns. Machine equations in abc variables for a symmetrical motor with turn fault in one winding can be expressed as [34–36]. Here, we assumed that the leakage inductance of the shorted turns is , where is the per-phase leakage inductance and the fault impedance is resistive .where
The resistance matrices of equation (1) are as follows:
Adding the first two rows of equation (1),where
Inductance matrices are modified as
The voltage and flux linkage equations for the shorted turns (βs2) are
The expression of electromagnetic torque can be in machine abc variables is
The inductance matrices are given by (9)–(11)
3. Sequence Component Analysis and Parameter Extraction
The symmetrical components are solid tool for analyzing and solving the problems of any unbalanced system. The symmetrical components are found reliable indicators of stator turn faults. By principle, symmetrical (without fault) motors supplied with symmetrical three-phase voltage sources will not produce negative sequence currents. When a turn fault occurs, symmetry will disturb and generate negative and zero sequence currents. In reference to the symmetrical components practice, three sets of symmetrical balanced phases are derived from any set of unbalanced parameters. They are recognized as positive, negative, and zero sequence components. Using Fortescue’s transformation given by equation (12), symmetrical components (IP, IN, I0) are calculated from unbalanced phase currents (Ia, Ib, Ic).where .
Fundamentally, the three-phase induction motor is a symmetrical system in healthy conditions and produces only positive sequence currents. It generates positive, negative, and zero sequence when symmetry is disturbed during a fault situation.
MATLAB software is used for creating a simulation model of a three-phase motor with turn fault in one of the phase winding. Due to the difficulty in creating the fault and measuring the phase currents experimentally for high values of percentage shorting, we are forced to create the Simulink model. Simulink of the induction motor with stator winding shorting is constructed based on the fundamental equations mentioned in Section 2. The Simulink model of the motor with interturn short is shown in Figure 1.
![](https://static.hindawi.com/articles/jece/volume-2019/4825787/figures/4825787.fig.001.jpg)
The model is simulated for different levels of shorting in one-phase winding and values of phase currents are saved in the MATLAB workspace. From these values, negative sequence current, positive sequence current, and zero sequence currents are calculated. Phase currents and sequence currents for different fault levels are tabulated in Table 1.
Here, simulation is carried out by introducing short circuiting in one phase at a time. The severity level of fault has been increased from zero percentage gradually. A negative sequence component of current is found to be increasing gradually with an increase in the fault level.
The Simulink data are verified by conducting experiments on three-phase induction motors rated 5 hp and 1 hp. Three-phase induction motors of Kirloskar electric company having specification 5 hp, 1430 rpm, 415 V, and 10 A and a digital oscilloscope-DS 1150 are used in the experiment. The experimental setup of stator interturn fault is shown in Figure 2. Interturn short circuit is created by taking out tapings from one of the phase windings. Table 2 gives the comparison of phase current values of winding with shorted turns from experimental setup and that from the Simulink model of the same rated motor.
![](https://static.hindawi.com/articles/jece/volume-2019/4825787/figures/4825787.fig.002.jpg)
Experimental values and simulated values are comparable since error percentage of both values is very low, thus proving the authenticity of the Simulink model.
For 5 hp motor, current values obtained practically are compared with those obtained from the mathematical model, described under Section 2. The details of winding parameters used for calculation are Stator resistance, Rs = 0.2777 Ω Rotor resistance, Rr = 0.183 Ω Stator inductance, Ls = 0.0553 H Rotor inductance, Lr = 0.056 H Mutual inductance, Lm = 0.0538 H
Error percentage of simulation values with mathematical model values listed in Table 3 is very low. Comparison output shows compliance or proves the validity of the Simulink model.
Simulated values are compared with practical values as well as analytical values from a mathematical model. Mean Error percentage of phase current values obtained from experiment and simulation are 0.00249. Mean error percentage of simulated values with analytical values of the mathematical model is 0.0051. The comparison output shows compliance and proves the validity of Simulink model.
The per unit change in a negative sequence current with positive sequence current is considered as the main input parameter for classification of severity of the fault level in phase windings. Case 1: with zero percentage shorting (healthy) Inegative = 0 Ipositive = 10.0248 And so δ = 1 Case 2: turn fault of 1.4 percentage Inegative = 0.0254 Ipositive = 10.0493 δ = (10.0493 − 0.0254)/10.0493 = 0.997472461
The value of δ for various levels of short circuiting is tabulated in Table 4. Selection of effective parameter is very important in fault detection along with the selection of classifier.
Generalization capacity of fault indicator parameter (δ) has been checked for five motors. The specifications of motors (I to V) considered for analysis are M I- 1.1 KW, 400 V, 50 Hz, 1447 rpm, 2.7 A M II- 5.5 KW, 400 V, 50 Hz, 1457 rpm, 11.6 A M III- 55 KW, 400 V, 50 Hz, 1480 rpm, 102 A M IV-110 KW, 400 V, 50 Hz, 1487 rpm, 194 A M V-250 KW, 400 V, 50 Hz, 1488 rpm, 445 A
The value of δ varies from 1 to 0.95 for short circuit levels 0% to 15%. In this case, the input vector to NN or fault indicator values (δ) is identical for all motors for a specific fault level. Figure 3 shows the variation of δ values at different interturn fault levels for motors under analysis.
![](https://static.hindawi.com/articles/jece/volume-2019/4825787/figures/4825787.fig.003.jpg)
4. Neural Network for Classification
Artificial neural networks are indulgent to noise and respond quickly, so they can be employed in real-time fault detection [27–29]. Since it is not possible to create a look-up table storing data for all conditions, a feedforward neural network is used for classifying the fault. Anticipating maximum accuracy from the trained neural network, input vector is created using possible experimental values and Simulink values for high percentage shorting. The various processes involved in the work of detection of severity level of shorting in stator winding are described in the block diagram (Figure 4).
![](https://static.hindawi.com/articles/jece/volume-2019/4825787/figures/4825787.fig.004.jpg)
The design and development of neural networks comprises preparation of input data set for the neural network, selection of a network structure, training of the network, testing, and evaluation of the classifier.
Backpropagation (BP), which is the most popular supervised learning method, is adopted for this process. This learning algorithm increases the efficiency of the network by minimizing the error, and so the gradient of the error curve slopes down.
The input data to NN are array of δ values. Target value is fixed for each value of input, δ. Both input data and target values for different classification levels are displayed in Table 5. Data set for training is selected in such a way that it contains the actual practical values till maximum measurable value during short circuiting and full range of values from simulation.
Target value [10000] represents healthy condition of winding. Target value is fixed as 1xxxx where xxxx represents xx.xx% level of shorting. Target for 00.23% of severity is 10023, for 01.40% of severity is 10140 and for 23.17% severity is 12317.
The performance of the algorithm is sensitive to the setting of the learning rate. A very small learning rate will lead to longer converging time, and a very high learning rate can lead to oscillating and unstable algorithm.
Backpropagation training with an adaptive learning rate is implemented with gradient descent function. Hyperbolic tangent sigmoid transfer function is used, which calculates a layer’s output from its net input. Mean squared normalized error performance function measures the network performance according to the mean of squared errors, when incorporated into the training process enhances the efficiency of the synaptic weight’s adjustment. A very low MSE reflects that the desired output and the ANNs outputs are close to each other, and thereby the network is well trained.
5. Result and Performance Validation
The proposed networks were subjected to training with input signals as described in Section 4. While analyzing the performance and regression plot of networks, it has been found that neural networks reacted well with training and validation samples. Validation performance of the network gave 100% accuracy (with 50 samples). So, accuracy percentage is calculated with error between target value and actual output vector. The level of accuracy obtained is 99.05%. The performance plot of the neural network is given in Figure 5 Stopping criterion is established with means squared error of 1.02e − 005.
![](https://static.hindawi.com/articles/jece/volume-2019/4825787/figures/4825787.fig.005.jpg)
NN with 2 hidden layers and number neurons in hidden layers as 16 and 1, respectively, shows the highest accuracy of 99.6%.
Comparison is made between the performance of neural networks which are fed with three-phase currents as input and δ as input. Neural network with δ as input vector shows steady increase in percentage accuracy. The graph in Figure 6 shows the percentage accuracy of the neural network when it is trained with phase currents as input vector and when trained with proposed fault indicator δ as input vector where, δ = (IP − IN)/IP.
![](https://static.hindawi.com/articles/jece/volume-2019/4825787/figures/4825787.fig.006.jpg)
6. Conclusion
This research focus is to make progress as well as to simplify the field of condition monitoring and fault detection in an induction motor. Negative sequence current monitoring is one of the simplest but solid and substantial techniques for stator short circuit detection. It is relevant to specify that selection of fault indicator is very critical in the classification process. In this work, per unit change in a negative sequence current with positive sequence current is considered as the fault indicator and so found to be more generalized technique for interturn stator winding fault detection. While phase current is considered as fault indicator, input vector of NN is different for different motors. When taken δ as fault indicator, input vector of NN is the same for different motors because of its per unit nature.
The fault parameter under consideration is derived from the time domain, which avoids the use of sophisticated signal conditioning techniques used in frequency domain for fault detection.
The work presents an application of neural network to classify the stator interturn fault. The network is trained with a full range of input vector using experimental values (for a small level of fault) as well as Simulink values (for high level of fault). Thus, NN is well trained with a full range of data sets. Simulink modeling helps create an infinite database, which is not possible by means of experiment. The performance of the NN is found to be accurate and fast. Fault detection in the beginning stage increases the feasibility of repairing the machine and avoids the risk of fire and explosion. Future extension of NN is possible to account detection of other electrical and mechanical faults possible in an induction motor.
Data Availability
The Simulink model data used for parameter calculation and the neural network training and testing data used to support the findings of this study are included within the article. Consolidated data, which are used for checking the generalization of five different motors, are included within the article.
Conflicts of Interest
The authors declare that they have no conflicts of interest.
Acknowledgments
This work was carried out in state-of-the-art Power System Protection Laboratory, Department of Electrical and Electronics, KCG College of Technology, Chennai, India. The lab is established under FIST (Fund for Improvement of S&T Infrastructure) scheme, supported by the Department of Science and Technology (DST), Government of India.