ARTIFICIAL INTELLIGENCE BASED ELECTRIC FAULT DETECTION IN PMSM Jayarama Pradeep 1, R.Devanathan 2 and Kannan Prashanth 3 1 Research Scholar, Sathyabama University. 2 Professor, Hindustan Institute of Technology. 3 Student, St. Joseph s College of Engineering. Abstract In Electrical drive system, occurrence of any fault may degrade the entire system s performance. Online fault detection plays a vital role in drive system to detect and rectifying the fault in drive systems used in safety critical applications. In this paper, suitable Electrical fault detection for Permanent Magnet Synchronous Motor using Artificial Neural Network is proposed. The neural network based method is highly efficient computing method and is suited to detect faults that develop gradually in a system. An appropriate selection of the feature extractor will provide the neural network with adequate significant details in the pattern set, so that the highest degree of accuracy in the neural network performance can be obtained. The discrete wavelet transform permits a systematic decomposition of a signal into its sub-band levels as a preprocessing of the system. Since different faults have different effects for stator currents, the wavelet transform can extract the features efficiently. The proposed approach deals with the fault detection system incorporating a neural network which is trained using Levenberg-Marquart algorithm and a Discrete Wavelet Transform based feature extraction block for Permanent magnet synchronous motor drive system. Keywords: Fault detections, Feature Extraction, Discrete wavelet transform, Permanent Magnet Synchronous Motor, Electrical fault, Levenberg-Marquart algorithm INTRODUCTION A fault is defined as an unpermitted deviation of at least one characteristic property of a variable from an acceptable behavior. Faults in a system may lead to degraded performance, malfunctions, or failures. Different from fault, the consequences of a failure are usually more serious, such as partial or complete system breakdown [1]. Faults in engineering systems are difficult to avoid. In complex systems, any fault possesses the potential to impact the entire system s behavior. In a manufacturing process, a simple fault may result in off specification products, higher operation costs, shutdown of production lines, and environmental damage, etc. In a continuously operated system, ignoring 1
a small fault can lead to disastrous consequences. In order to avoid such situation, there has been an increase in the number of fault detection techniques that are generic in nature. There are basically two generic ways to approach the analytical fault detection problem: The model based approach and the data-based approach. In the model-based approach, the engineer has access to a model of the system whose behavior is being monitored. The model could be analytical, or knowledge-based. Most applications of this approach have dealt with linear systems, since they can be easily described and studied. In the data-based approach one bypasses the step of obtaining a mathematical model and deals directly with the data. This is more appealing when the process being monitored is not known to be linear or when it is too complicated to be extracted from the data. It is conceivable that a neural net can be used as a monitoring device, in order to detect major changes in the operation of the system. Specifically, one approach may be that the neural net is trained on a well-behaving system, and then operated with no more training in parallel to the actual system. The neural net output will then be compared to that of the physical system, and any anomalies in the output of our system will be detected. FAULT DETECTION Generally Fault detection in motors is done by incorporating a relay circuit in the main device circuitry. A relay is an electrically operated switch. Many relays use an electromagnet to operate a switching mechanism mechanically, but other operating principles are also used. Relays are used where it is necessary to control a circuit by a lowpower signal (with complete electrical isolation between control and controlled circuits), or where several circuits must be controlled by one signal. Permanent magnet synchronous motors use three phase solid state relays in practical applications. However effective fault diagnosis requires highly efficient relay mechanism using various elements. The total cost of the relay circuitry must not exceed the cost of the machine it intends to protect. This means that lower the cost of a machine, lower would be its relay circuitry cost. Another aspect of a good fault detection system is the time takes to detect a fault. An ideal fault detection system should detect the fault and its location immediately so that the fault can be isolated for further corrective action. This plays a vital role, as faults can have an avalanche effect on the entire system thereby creating more faults and damage. Relays are also affected by temperature changes and aging process. Most electrical systems widely succumb to two types of faults, namely short circuit and open circuit faults. In a permanent magnet synchronous machine these faults can occur either on the inverter side or in the machine windings itself. The effect of the fault on the circuit is the same irrespective on which side the fault has occurred. Short circuit fault is defined as a low-resistance connection established by accident or intention between two points in an electric circuit. The current tends to flow through the area of low resistance, bypassing the rest of the circuit [2]. Open circuit fault is defined as any interruption in the circuit, such as an open switch, a break in the wiring, or a component such as a resistor that has changed its resistance to an extremely high value will cause current to cease. The EMF will still be present, but voltages and currents around the circuit will have changed or ceased altogether [3]. 2
NEURAL NETWORK USED IN FAULT DETECTION The objective of the paper is to provide a neural network based fault detecting system that is highly efficient and cost effective. The neural network solution for the machine is more accurate and faster than other types as its behavior is much like the human brain. The various outputs of the machine are provided to the neural network after applying a wellresearched feature extraction technique. The neural network is initially trained with a selected range of data. This could include the performance of the machine during ideal and faulted conditions. The neural network must be properly trained as it plays an important role in detecting faults. The neural network acts on the input it receives based on the activation function it is provided. The weights and bias of the neural network are changed for each instance of input it receives [4].When a faulted condition occurs in the permanent magnet synchronous machine either in the inverter side or in the machine windings, the neural network can recognize the faulted condition immediately. The detection of faulted condition by the neural network works similar to the human biological neural network. The input for the neural network is taken from the motor output instead of the inverter as shown in Fig.1. By this the short circuit and open circuit faults that occur at both the inverter and machine windings can be detected easily. Fig 1. Basic Block Diagram The direct and quadrature axis currents drawn from the machine are subjected to discrete wavelet transform. The Fourier transform is a useful tool to analyze the frequency components of the signal. However, if we take the Fourier transform over the whole time axis, we cannot tell at what instant a particular frequency rises. Short-time Fourier transform (STFT) uses a sliding window to find spectrogram, which gives the information of both time and frequency. But still another problem exists: The length of window limits the resolution in frequency. Wavelet transform seems to be a solution to the problem above. Wavelet transforms are based on small wavelets with limited duration. The translated-version wavelets locate where we are concerned, whereas the scaled-version wavelets allow us to analyze the signal in different scale [5]. 3
PERMANENT MAGNET SYNCHRONOUS MOTOR PMSMs are attractive for industrial applications, their high power density, which is defined as the amount of output power for a unit weight (power (watts) / weight), is a distinct advantage over other types of electric machines. The stator winding fault is the most likely electrical fault in PMSM. Stator winding faults are usually caused by insulation breakdown between coils in the same phase or different phases [6]. While the winding short usually emerges locally as an incipient fault, it may propagate rapidly and result in the failure of the entire phase. This is due to the increased ohmic heating associated with the large current in the shorted portion of the winding. The excessive heating will lead to significant temperature increase and faster deterioration of the insulation system. The PMSM control is equivalent to that of the dc motor by a decoupling control known as field oriented control or vector control. The vector control separates the torque component of current and flux channels in the motor through its stator excitation. The stator currents must be transformed to the rotor reference frame with the rotor speed ω, using Park s transformation. The q and d axis currents are constants in the rotor reference frames since α is a constant for a given load torque. As these constants, they are similar to the armature and field currents in the separately excited dc machine. The q axis current is distinctly equivalent to the armature current of the dc machine; the d axis current is field current, but not in its entirety. It is only a partial field current; the other part is contributed by the equivalent current source representing the permanent magnet field. For this reason the q axis current is called the torque producing component of the stator current and the d axis current is called the flux producing component of the stator current. ARTIFICIAL NEURAL NETWORKS An Artificial Neural Network (ANN) is an information processing paradigm that is inspired by the way biological nervous systems, such as the brain, process information. The key element of this paradigm is the novel structure of the information processing system. It is composed of a large number of highly interconnected processing elements (neurons) working in unison to solve specific problems. ANNs, like people, learn by example. An ANN is configured for a specific application, such as pattern recognition or data classification, through a learning process. Learning in biological systems involves adjustments to the synaptic connections that exist between the neurons. This is true of ANNs as well [7]. Neural networks, with their remarkable ability to derive meaning from complicated or imprecise data, can be used to extract patterns and detect trends that are too complex to be noticed by either humans or other computer techniques. A trained neural network can be thought of as an "expert" in the category of information it has been given to analyze [7]. This expert can then be used to provide projections given new situations of interest and answer "what if" questions. Artificial neural networks consist of many nodes. Each node has a node function, associated with it which along with a set of local parameters determines the output of the node, given an 4
input. Modifying the local parameters may alter the node function. Artificial Neural Networks thus is an information-processing system. In this information-processing system, the elements called neurons, process the information. The signals are transmitted by means of connection links. The links possess an associated weight, which is multiplied long with the incoming signal (net input) for any typical neural net. The output signal is obtained by applying activations to the net input. The arrangement of neurons into layers and the pattern of connection within and inbetween layer are generally called as the architecture of the net as shown in Fig.2. Fig.2 Neural Network Architecture The neurons within a layer are found to be fully interconnected or not interconnected. The number of layers in the net can be defined to be the number of layers of weighted interconnected links between the particular slabs of neurons. If two layers of interconnected weights are present, then it is found to have hidden layers. There are various types of network architectures: Feed forward, feedback, fully interconnected net, competitive net, etc. TRAINING THE NEURAL NETWORK What has attracted interest in neural networks is the possibility of learning. Given a specific task to solve, and a class of function, learning means using a set of observations to find which solves the task in some optimal sense. This entails defining a cost function such that, for the optimal solution, - i.e., no solution has a cost less than the cost of the optimal solution The cost function is an important concept in learning, as it is a measure of how far away a particular solution is from an optimal solution to the problem to be solved. Learning algorithms search through the solution space to find a function that has the smallest possible cost. For applications where the solution is dependent on some data, the cost must necessarily be a function of the observations; otherwise we would not be modeling anything related to the data. It is frequently defined as a statistic to which only approximations can be made. As a simple example, consider the problem of finding the model, which minimizes, for data pairs drawn from some distribution. In practical situations we would only have samples from and thus, for the above example, we would only 5
minimize. Thus, the cost is minimized over a sample of the data rather than the entire data set. When some form of online machine learning must be used, where the cost is partially minimized as each new example is seen. While online machine learning is often used when is fixed, it is most useful in the case where the distribution changes slowly over time. In neural network methods, some form of online machine learning is frequently used for finite datasets. LEVENBERG MARQUART ALGORITHM The Levenberg Marquardt algorithm was independently developed by Kenneth. Levenberg and Donald Marquardt, provides a numerical solution to the problem of minimizing a nonlinear function. It is fast and has stable convergence. In the artificial neuralnetworks field, this algorithm is suitable for training small- and medium-sized problems[9] [10]. The Levenberg-Marquardt method is an efficient and popular damped least square technique. This method is a combination between the Gauss and the steepest gradient descent methods, where the amount of damping used in each iteration is central in establishing the behavior of the system. Further, the damping is determined by four parameters, whose optimum values vary from model to model. An inappropriate selection of the damping parameters could trigger a decrease in the rapidness of convergence, convergence to local minimum, or system instability. Therefore, finding proper values for these parameters is fundamental in the use of this method and implies a great deal of extra time. This lack of efficiency is considered a disadvantage in comparison to other techniques FEATURE EXTRACTION The computation unit cannot directly visualize as a human does. The signals are difficult to sample as an important characteristic and have high correlation coefficient for classifying a fault hypothesis. Therefore, a signal transformation technique is required. Sampling number is deficient it is difficult for diagnosis, and a large sampling number is a burden for transferring and calculation. So feature extraction of the signal is a critical initial step in any monitoring and fault diagnosis system. Its accuracy directly affects the final monitoring results. Thus, the feature extraction should preserve the critical information for decision making. An appropriate selection of the feature extractor is to provide the neural network with adequate significant details in the pattern set so that the highest degree of accuracy in the neural network performance can be obtained. DISCRETE WAVELET TRANSFORM Often times, the information that cannot be readily seen in the time-domain can be seen in the frequency domain. When current signals show non-stationary or transient conditions, the conventional fourier transform technique is not suitable. The analysis of non-stationary signals can be performed 6
using time-frequency techniques (short-time Fourier transform) or time-scale techniques (wavelet transforms) [10]. The discrete wavelet transform (DWT) permits a systematic decomposition of a signal into its sub-band levels as a preprocessing of the system. Since different faults have different effects for stator currents, the wavelet transform can extract the features, which provides a good basis for the next feature extraction. Although the discretized continuous wavelet transform enables the computation of the continuous wavelet transform by computers, it is not a true discrete transform. As a matter of fact, the wavelet series is simply a sampled version of the CWT, and the information it provides is highly redundant as far as the reconstruction of the signal is concerned. This redundancy, on the other hand, requires a significant amount of computation time and resources. The DWT, on the other hand, provides sufficient information both for analysis and synthesis of the original signal, with a significant reduction in the computation time. The DWT is considerably easier to implement when compared to the CWT. DWT employs two sets of functions, called scaling functions and wavelet functions, which are associated with low pass and high-pass filters, respectively. The decomposition of the signal into different frequency bands is simply obtained by successive high-pass and low-pass filtering of the time domain signal. The original signal x[n] is first passed through a halfband high-pass filter g[n] and a low-pass filter h[n]. After the filtering, half of the samples can be eliminated according to the Nyquist s rule, since the signal now has a highest frequency of π /2 radians instead of π. The signal can therefore be sub-sampled by 2, simply by discarding every other sample. This decomposition halves the time resolution since only half the number of samples now characterizes the entire signal. However, this operation doubles the frequency resolution, since the frequency band of the signal now spans only half the previous frequency band, effectively reducing the uncertainty in the frequency by half. The above procedure, which is also known as the sub-band coding, can be repeated for further decomposition. At every level, the filtering and sub-sampling will result in half the number of samples (and hence half the time resolution) and half the frequency band spanned (and hence doubles the frequency) interpreting the DWT coefficients can sometimes be rather difficult because the way DWT coefficients are presented is rather peculiar. In brief, DWT coefficients of each level are concatenated, starting with the last level. SIMULINK MODELLING OF THE SYSTEM PMSM Drive System is simulated as shown in Fig.3 using Matlab/Simulink package for fault diagnosis using neural network Discrete wavelet transform algorithm. The machine is controlled by the phase switching provided by the MOSFET inverter which is controlled by the triggering system. The permanent magnet synchronous motor is given a constant torque input and is assumed to be running at a stable speed. 7 Fig 3. PMSM Drive System
Fig.4. Simulink model of fault detection System Fig.5 Feature Extraction system Fig.4 shows the Simulink model of the fault detection system for various fault condition as mentioned in section VI. The information from the motor is given to the Matrix Concatenate block as the Neural Network cannot work on data provided to it in discrete form. The neural network output is given to a quantizer. The neural network is provided with three layers and gives the output as 1 if a fault is detected and remains 0 in the absence of a fault. Fig 5. shows the Simulink model of feature extraction system. The direct axis and the quadrature axis stator currents are extracted for the neural network. The stator currents possess efficient information in its characteristics for the neural network to detect the presence of fault in either the converter or the motor circuit. The extracted currents are passed to a zero-order hold and then into a buffer to send the data to the Discrete Wavelet Transform block. PMSM FAULT DETECTION 8
In a PMSM, open and short circuit faults can occur either in the inverter side or in the machine windings. Fault phenomenon is provided with a respective neural network setup. The resultant effect is however similar when the PMSM does not have any fault the output of the neural network is 0 which is shown in Fig 6. When the neural network detects any fault in either the inverter or the motor system the output becomes 1as shown in Fig 7. Fig.6 Output of Fault Indicator under normal condition. Fig.7 Output of Fault Indicator under one inverter switch condition at t= 0.1s. Fig.8 Motor phase voltage under normal condition. switch Fig.9 Motor phase voltage under one inverter open circuit at t=0.1s. 9
Fig. 10 Stator Currents under normal condition. Fig. 11 Stator Currents under fault condition. Fig. 12 Quadrature and Direct axes under normal condition Fig. 13 Quadrature and Direct axes under fault condition Fig.14 Speed response under normal operation. 10
Fig. 15 Speed response under fault condition Fig. 16 Torque response under normal operation Fig.17 Torque response under fault condition Fig. 8 indicates the normal response of phase voltage fed to the motor. When a fault is developed in the windings of the motor or in an inverter side, for e.g., an open circuit of any one switch in the inverter part is simulated at t=0.1s and immediately the neural identifies the fault and fault type indicator show level 1 as shown in Fig.7. The phase voltage of motor is shown in Fig.9 under fault condition. Fig.10 & 11 indicates the three phase stator currents of the motor under normal and fault condition. Fig. 12 & 13 represents the quadrature and direct axes current under normal and abnormal condition. It can be observed that the quadrature current is decreases and direct axis current increases around 20A which is well above the rated value under fault condition. Fig 13 & 14 indicates the speed response of the motor under normal and inverter switch open circuit fault condition where the motor normally runs at 700rpm and when fault occurs at t=0.1s the speed reaches zero after oscillation. The torque response is shown Fig.15 & 16 under normal and fault condition. CONCLUSION 11
Electric machines are important components for industries and special applications. The continuous healthy operation of machines is critical for the reliability of the entire system. PMSM has been provided with a fault detecting system by incorporating a neural network which is trained using Levenberg-Marquart algorithm and a DWT based feature extraction block. The various parameters of the permanent magnet synchronous motor were considered and the direct axis stator current and quadrature axis stator currents were found to contain more information instances in its characteristics. Hence they were given as inputs for the neural network after sampling and transformation. The present project deals with providing a fast fault detection system that can detect any kind of fault occurring in the inverter or motor circuitry. However the idea can be extended to diagnose the fault and provide an effective isolation technique REFERENCES 1. LI LIU Robust Fault Detection And Diagnosis For Permanent Magnet Synchronous Motors Dissertation submitted to the Department of Mechanical Engineering, The Florida State University, College Of Engineering June 9th, 2006, pp 2-12. 2. Introduction to short circuit analysis, www.pdhonline.org. 3. Open circuit fault definition, http://www.learnabout-electronics.org/resistors_18.php, 4. Hua Su and Kil To Chong, Induction Machine Condition Monitoring Using Neural Network Modeling, IEEE Transactions on Industrial Electronics, vol. 54, no. 1, February 2007, pp36-42. 5. Robi Polikar, The Engineer's Ultimate Guide To Wavelet Analysis The Wavelet Tutorial, Part I, May 2, 2001, pp 10-12. 6. D.R. Hush, C.T. Abdallah, G.L. Heileman, and D. Docampo, Neural Networks In Fault Detection: A Case Study, University of New Mexico,Albuquerque, NM 87131, USA., pp2-4. 7. Fiona Nielsen, Neural Networks algorithms and applications, Ph.D thesis, Brock Business College, page 2 8. P.Baby, P.Bhuvaneswari and K.Sasirekha, A Study on Learning Techniques in Artificial Neural Network, JESCSET 13 & vol. 2,no.4, pp 1001-1006. 9. Bogdan M. Wilamowski, and Hao Yu, Improved Computation for Levenberg Marquardt IEEE Transactions on Neural Networks, vol. 21, no. 6, June 2010,v pp 930-937. 10. Evelyn Araneda, A Variation of the Levenberg Marquardt Method An attempt to improve efficiency, M.Sc Thesis, Massachusetts Institute of Technology, May 2004. 11. Ru-Shan Wu and Ling Chen, Wave Propagation and Imaging Using Gabor-Daubechies Beamlets, Proc. 5th International Conference on Theoretical and Computational Acoustics, May 21-25, 2001, Beijing, China. Jayarama Pradeep Research Scholar, Sathyabama University. 12
13