Fault Diagnosis System for a Multilevel Inverter Using a Neural Network

Transcription

1 Fault Diagnosis System for a Multileel Inerter Using a eural etwor Surin Khomfoi Leon M. Tolbert Electrical and Computer Engineering Electrical and Computer Engineering The Uniersity of Tennessee The Uniersity of Tennessee 44 Ferris Hall 44 Ferris Hall Knoxille,T 37996, USA Knoxille, T 37996, USA surin@ut.edu tolbert@ut.edu Abstract In this paper, a fault diagnosis system in a multileelinerter using a neural networ is deeloped. It is difficult to diagnose a multileelinerter drie (MLID) system using a mathematical model because MLID systems consist of many switching deices and their system complexity has a nonlinear factor. Therefore, a neural networ classification is applied to the fault diagnosis of a MLID system. Fie multilayer perceptron (MLP) networs are used to identify the type and location of occurring faults. The neural networ design process is clearly described. The classification performance of the proposed networ between normal and abnormal condition is about 9 %, and the classification performance among fault features is about 85 %. Thus, by utilizing the proposed neural networ fault diagnosis system, a better understanding about fault behaiors, diagnostics, and detections of a multileel inerter drie system can be accomplished. The results of this analysis are identified in percentage tabular form of faults and switch locations. Index Terms Multileel inerter, neural networ, fault diagnosis I. ITRODUCTIO In recent years, industry has begun to demand higher power ratings, and MLID systems hae become a solution for high power applications. A multileel inerter not only achiees high power ratings, but also enables the use of renewable energy sources. Renewable energy sources such as photooltaic, wind, and fuel cells can be easily interfaced to the multileel inerter system for a high power application []. A schematic of a single phase multileel inerter system is illustrated in Fig.. Because multileel inerter systems are utilized in high power applications, the reliability of the power electronics equipment is ery important. For example, industrial applications such as industrial manufacturing are dependent upon induction motors and their inerter systems for process control. Generally, the conentional protection systems are passie deices such as fuses, oerload relays, and circuit breaers to protect the inerter systems and the induction motors. The protection deices will disconnect the power sources from the multileel inerter system wheneer a fault occurs, stopping the operated process. Downtime of manufacturing equipment can add up to be thousands or hundreds of thousands of dollars per hour, therefore fault detection and diagnosis is ital to a company s bottom line. Many engineers and researchers hae focused on incipient fault detection and preentie maintenance to preent inerter and motor faults from happening. The most important factor, howeer, is how the system could operate continuously while there is an abnormal condition. Va n S A S A SA SA Fig.. Singlephase multileelinerter system. In order to maintain continuous operation for a multileel inerter system, nowledge of fault behaiors, fault prediction, and fault diagnosis are necessary. Faults should be detected as soon as possible after they occur, because if a motor drie runs continuously under abnormal conditions, the drie or motor may quicly fail. The arious fault modes of a conentional PWM oltage source inerter (VSI) system for an induction motor are inestigated in []. Then, the integration of a fault diagnosis system into VSI dries is described in [3]. This integration system introduced remedial control strategies soon after failure occurrences; therefore, system reliability and fault tolerant capability are improed. A noninasie technique for diagnosing VSI drie failures based on the identification of unique signature patterns corresponding to the motor supply current Par s Vector is proposed in [4]. A study of a machine fault diagnosis system by using FFT and neural networs is clearly explained in [5]. Also, a fault diagnosis system for rotary machines based on fuzzy neural networs is deeloped in [6]. Furthermore, a new topology with faulttolerant ability that improes the reliability of multileel conerters is proposed in [7]. The nowledge and information of fault behaiors in the system is important to improe system design, protection, and fault tolerant control. Thus far, limited research has focused on MLID fault diagnosis. Therefore, a MLID fault diagnosis system is proposed in this paper. An example of a MILD open circuit fault at switch S A is represented in Fig.. S A fault will cause unbalanced oltage and current output, while the induction motor is operating. The unbalanced oltage and current may result in ital damage to the induction motor if the induction motor is run lie this for a long time. The unbalanced condition from fault S B S B S B S B HBridge V dc HBridge V dc

2 S A can be soled if the fault location is correctly identified. Switching patterns and the modulation index of other actie switches in the MLID can be adjusted to maintain output oltage and current in a balanced condition. Although the MLID can continuously operate in a balanced condition, the MLID will not be able to operate at its rated power. Therefore, the MLID can operate in balanced condition at reduced power while the fault occurs until the operator nows and repairs the inactie switch. In this research, we will attempt to diagnose the fault location in a MLID from its output oltage waeform. MLID open circuit faults at each switch are considered. Although the MLID system usually consists of three phases of H bridge inerters and also can hae short circuit faults, the fault diagnosis system will be the same topology as a single phase and open circuit case. Moreoer, one leel of multileel inerter is focused in this research; howeer, other inerter leels can be extended by using this proposed topology with more training data. The proposed networ utilizes output oltage signals of the MLID to train the neural networs. The acquired data is transformed by using Fast Fourier Transform technique to rate a signal alue as an important characteristic. The signal feature extraction is discussed, and the process of neural networ design is clearly described in the following section. II. GEERAL OTIO OF FAULT DIAGOSIS SYSTEM A. Structure of fault diagnosis system The structure for a fault diagnosis system is illustrated in Fig. 3. The system is composed of four major states: feature extraction, neural networ classification, fault diagnosis, and switching pattern calculation with gate signal output. The feature extraction, neural classification, and fault diagnosis are the focus of this research. The feature extraction performs the oltage input signal transformation, with rated signal alues as important features, and the output of the transformed signal is transferred to the neural networ classification. The networs are trained with both normal and abnormal data for the MLID; thus, the output of this networ is nearly and as binary code. The binary code is sent to the fault diagnosis to decode the fault type and its location. Then, the switching pattern is calculated. B. Feature extraction system The Simpower Matlab toolbox in Simulin is used to simulate data of fault features with.8 modulation index (m a ) out of. and as illustrated in Fig. 4. Also the output oltages are measured from the MLID prototype in noload condition as shown in Fig. 5. As can be seen, the signals are difficult to rate as an important characteristic for classifying a fault hypothesis. Therefore, a signal transformation technique is needed. An appropriate selection of the feature extractor is to proide the neural networ with adequate significant details in the pattern set so that the highest degree of accuracy in the neural networ performance can be obtained. One possible technique for implementation with a digital V a n S A S A S B S B HBridge V dc Fig.. HBridge Switch S A open circuit fault at second leel of singlephase multileelinerter. V dc V dc Input Multileel Inerter System Gate output signals Switching pattern and time calculation system Fault diagnosis system eural networ fault classification Output Output Voltage Signals Feature Extraction System Fig. 3. Structure of fault diagnosis system. Induction Motor signal processing microchip is Fast Fourier Transform [8]. Beginning with the discrete Fourier Transform in (), then the FFT using the decimation in time decomposition algorithm is illustrated in (). F = n= where F F f n W n W = e = G = G W W π j H H for for = for = =,...,,...,,..., G is for eennumbered elements of f n, whereas H is for oddnumber elements of f n. G and H can be computed as shown in (3) and (4). n fnw n= () () G = (3)

3 sine IV A B C sine ref ormal Fault A Fault A Fault B Fault B ormal Fault A Fault A Fault B Fault B sine IV sw signal A sine IV sw signal B sw signal C In In In sdcs A sdcs B sdcs C step load A B C Tm m Asynchronous Machine SI Units m is_abc is_qd wm Te Machines Measurement Demux Fault Diagnosis Simulation Scope3 VC VB VA VAB VBC VCA Scope Input data Fig. 4. Simulation of fault features at S A, S A, S B, and S B of Hbridge with modulation index =.8 out of.. Scope Fault Diagnosis Experiment Input data Fig. 5. Experiment of fault features at S A, S A, S B, and S B of Hbridge with modulation index =.8 out of.. n fn W n= H = (4) Together, the computational saings of the FFT becomes log compared to for the DFT [8]. Other popular signal transformation techniques such as Hartley and Waelet are explained in [8]. otmal Fault A Fault A Fault B Fault B FFT Simulation Harmonic order Fig. 6. Signal transformation of simulation output oltages by using FFT with modulation index =.8 out of.. ormal Fault A Fault A Fault B Fault B FFT Experiment Harmonic order Fig. 7. Signal transformation of experimental output oltages by using FFT with modulation index =.8 out of.. C. Simulation and experiment comparison The transformed signals of both simulation and experiment are represented in Fig. 6 and Fig. 7, respectiely. Obiously, the results are nearly identical fault features. The FFT technique has a good identical feature to classify normal and abnormal features; therefore, FFT is used to transform oltage output signals in this research in order to rate signal alue for important features so that the features for a fault hypothesis can be classified. In the experiment, a threephase wyeconnected cascaded multileel inerter using V, 7 A MOSFET as the switching deices was used to carry out the output oltage signals. A FPGA chip with 8 µsec time step is utilized to generate gate drie signals and interfaces with the gate drie board. The switching angles are calculated by using Simulin based on sinusoidal PWM. A olt battery is supplied to each HBridge inerter in both simulation and experiment.

4 (b) (d) (a) Fig. 8. Experiment of fault features at (a) normal, (b) S A fault, (c) S A fault, (d) S B fault, and (e) S B fault of Hbridge with modulation index =.8 out of.. Fault occurrence is created by physically remoing the switch in the desired position. A Yoogawa DL 54c is used to measure output oltage signals shown in Fig. 8 as ASCII files. The measured signals are set to =3; sampling frequency is Hz. Voltage spectrum is calculated and transferred to the neural networ fault classification. (c) (e) III. METHODOLOGY OF EURAL ETWORK FAULT CLASSIFICATIO All fault features, as preiously discussed, can be classified based upon their effects upon the output oltages. The transformation of output oltage signals is achieed by using FFT as shown by simulation and experimental results in Figs. 6 and 7. As mentioned before, the systematic mathematical technique may be complicated to implement in the practical real time control system; therefore, a feedforward neural networ technique permitting input/output mapping with a nonlinear relationship between nodes will be utilized [9]. eural networs proide the ability to recognize anomalous situations because of their intrinsic capacity to classify and generalize. Especially, the sensitiity and response time of the original procedure presented for the online analysis of fault set repetition enable online fault location techniques to be deeloped []. The stages of neural networ fault classification are explained as follows: A. eural networ architecture design The architecture of the proposed fault diagnosis neural networ,, is illustrated in Fig. 9. The fie multilayer feedforward networs, or MLP, are used in this research because the input data contain continuous features. A networ has one hidden layer with 4 input nodes corresponding to harmonic order and magnitude, hidden nodes, and output node. The sigmoid actiation function is used: tansig for hidden nodes and logsig for an output node. A logsig actiation function is used for an output node because the target output is between and. The implementation of the proposed neural networ classification system, consisting of fie s is shown in Fig.. It should be noted that the number of nodes for the input and output layers depends on the specific application. The selection of number and dimension in the hidden layer is based on neural networ accuracy in preliminary tests. Indeed, optimizing of networ architecture selection is a significant topic in a study of artificial intelligence aspects [9, ]. B. Input/output data Each networ is trained with one set of normal data and four sets of abnormal data, thus the size of the input matrix is 5 input data rows with 4 columns, [5 4]. The size of the output target is [5 ]. The target output corresponding with classification data is represented in Table I. A selection of training data and test data, from the original signal, is discussed in the next section. C. eural networ training The Leenberg Marquardt training paradigm, trainlm is utilized in this research because trainlm not only performs ery fast training time but also has inherent regularization properties. Regularization is a technique which adds constraints so that the results are more consistent. The % misclassification and % input data error rate are chosen to calculate a sum of square error goal, SSE; therefore, a.334 SSE goal is used to train the networ by calculating from (5). The training process will be finished when the SSE goal is met. SSE = L ( y ij d ij ) i= j = where y ij is an output alue of neural networ, d ij is an output of training data, is the number of training data, L is the number of units in the output layer.. Training and testing data set selection The training data set should also coer the operating region; thus, the test data sets are generated from simulation with arious operation points (different modulation indices). The testing sets are measured from experiment. Training and testing sets hae Hz sampling frequency. Both data sets are transformed by FFT from to 39 harmonic orders. Zero harmonic order means the dc component of the signals. (5)

5 TABLE I TARGET AD CLASSIFICATIO DATA umber of Classification Target ormal [ ;;;; ] 5x Fault at S A [ ;;;; ] 5x 3 Fault at S A [ ;;;; ] 5x 4 Fault at S B [ ;;;; ] 5x 5 Fault at S B [ ;;;; ] 5x ormal Fault A Input data Fault A Output Binary [ ] Input 3 Output 4 Fault B 4 Fault B Fig.. Implementation of proposed neural networ classification system.. Scaling data Fig. 9. Proposed neural networ architecture. The input training data are scaled by using the mean center and unit ariance method (Zscore scaling). The scaling data will aoid premature saturation of sigmoidal units and also allow the use of a specific output neuron [9]. The scaling parameters: the mean alue (X M ) and the standard deiation alue (X S ) are saed with the same data file as weights and biases. The testing data set will be scaled with the same scaling parameters as the training data set when the networ is examined. D. eural networ testing The networs are examined with the test data sets, as mentioned aboe, when the proposed networs hae trained to the desired error goal. Testing the networ inoles presenting the test set to the networ and calculating the error. If the error goal is met, the training is complete. IV. FAULT CLASSIFICATIO RESULTS The performances of the proposed networs are tested in two categories. First, the networs are tested with the selected test set at the same operating point (m a =.8) as preiously mentioned. The tested results along with the testing data set are illustrated in Table II. As can be seen, the error between the actual and target output data after testing the networ is less than the SSE goal (.334), thus the training process is complete. TABLE II COFUSIO TABLE FOR TESTIG DATA SET (m a =.8) Classification Target ormal [ ] Fault A [ ] Fault A [ ] Fault B [ ] Fault B [ ] Actual Data ormal Fault A Fault A Fault B Fault B 9.993x 99.93% 8.854x 4.88%.95x 3.9% 9.88x 4.93% 6.63x 4.66% 6.45x 4.64% 9.896x 98.96%.89x 4.3% 6.99x 5.693% 7.386x 4.74% 9.873x 5.98%.7x 3.3% 9.99x 99.99%.696x 3.69%.58x 3.5% 9.94x 4.994% 8.853x 4.88%.64x 3.6% 9.989x 99.89%.8x 3.% TABLE III COFUSIO TABLE FOR DIFFERET OPERATIO DATA SET (m a=) Actual Data Classification ormal Fault A Fault A Fault B Fault B ormal 95.4% 4.5%.98%.94%.358% Fault A.8% 75.94%.4% 5.8%.8% Fault A.9% 3.3% 84.9%.6%.4% Fault B.8%.36%.69% 78.89%.76% Fault B.5% 6.74%.5%.77% 86.83%.7x 3.%.4x 3.4%.46x 3.5% 3.476x 5.35% 9.997x 99.97%

6 Clearly, each networ has classification performance of more than 98%; therefore, the classification performance of the networ is quite satisfactory in the first category. Second, the new testing data are simulated at different operating points by changing the modulation index (m a ) from.8 to.. The new testing data set, at a unity modulation index, is transferred to the neural networs. In the MLID application, the modulation index is dependent on the operation point of the induction motor. Therefore, it would be better if the networs hae good classification performance for a wide range of operation. The second category testing results are illustrated in Table III. Obiously, the classification performance between normal and abnormal conditions is about 9 %, which is ery good performance, whereas the classification performance among fault features is about 85 % in Fault A and B and about 75% in Fault A and B (The proposed networ misclassified only out of 4 test sets). It should be noted that the future classification performance is important for the application. The classification performance among fault occurrences is acceptable. The neural networ can be trained with a larger, more complete fault data set to get more accurate classification results. As shown in Table III, the classification performance decreases when the operation point of MLID is changed. This suggests that an additional training set, or more training data, may be needed to train the fault classification neural networ especially in Fault A and Fault B training set. Also, a better data transformation technique may be required; howeer, the timeconsuming process for feature extraction should be considered in order to implement it in an online fault diagnosis system. Online fault diagnosis with a wide range operation due to desired modulation index is important in MLID. Although the classification performance decreases when the operating point is changed, the classification performance of the proposed networ is adequate to classify the fault s occurrence with its location. V. COCLUSIOS A study of a fault diagnosis system in a multileel inerter using neural networs has been proposed. The feature extraction system has been discussed. An FFT technique is utilized to transform output oltage signals in order to rate the signal alue as an important characteristic for classifying a fault hypothesis. FFT has an adantage in that less time is needed to calculate this signal transformation process, and it is possible to implement with a digital signal processing microchip. The feature extraction system is an important process because bad transformation signals will lead to poor classification performance. A study suggests that the feature extraction is a challenging research topic. The proposed networs perform ery well with the selected testing data set. The classification performance is ery high, more than 98%. Obiously, the classification performance between normal and abnormal condition is quite satisfactory in both testing categories; whereas, the classification performance among fault features is acceptable in the second category. The classification performance decreases when the operating point of the MLID is different from the training set. The results tell us that a new training set, or more training data, may be needed to accomplish a wide range of operation. Although the classification performance decreases when the operating point is changed, the oerall classification performance of the proposed fault diagnosis system is acceptable. The proposed networs hae the ability to classify normal and abnormal conditions, including the fault location. Therefore, by utilizing the proposed neural networ fault diagnosis system in this research, a better understanding of fault behaiors, diagnostics, and detections for multileel inerter drie systems can be achieed. REFERECES [] L. M. Tolbert, F. Z. Peng, Multileel Conerters as a Utility Interface for Renewable Energy Systems, in Proceedings of IEEE Power Engineering Society Summer Meeting, pp [] D. Kastha, B. K. Bose, Inestigation of Fault Modes of Voltagefed Inerter System for Induction Motor Drie, IEEE Trans. Industry Applications, ol. 3, no. 4, Jul. 994, pp [3] D. Kastha, B. K. Bose, OnLine Search Based Pulsating Torque Compensation of a Fault Mode SinglePhase Variable Frequency Induction Motor Drie, IEEE Trans. Industry Applications, ol. 3, no. 4, Jul./Aug. 995, pp. 88. [4] A. M. S. Mendes, A. J. Marques Cardoso, E. S. Saraia, Voltage Source Inerter Fault Diagnosis in Variable Speed AC Dries by Par s Vector Approach, in Proceedings of the 998 IEE 7 th International Conference on Power Electronics and Variable Speed Dries, pp [5] S. Hayashi, T. Asaura, S. Zhang, Study of Machine Fault Diagnosis Using eural etwors, in Proceedings of the eural etwors, IJC ', Vol., pp [6] S. Zhang, T. Asaura, X. Xu, B. Xu, Fault Diagnosis System for Rotary Machines Based on Fuzzy eural etwors, in Proceedings of the 3 IEEE/ASME Adanced Intelligent Mechatronics, AIM 3, pp [7] A. Chen, L. Hu, L. Chen, Y. Deng, and X. He, A Multileel Conerter Topology With FaultTolerant Ability, IEEE Trans. on Power Electronics, ol., no., March. 5, pp [8] J. A. Momoh, W. E. Olier Jr, J. L. Dolc, Comparison of Feature Extractors on DC Power System Faults for Improing A Fault Diagnosis Accuracy, in Proceedings of the 995 IEEE Intelligent Systems for the st Century, ol. 4, pp [9] L. H. Tsoualas, R. E. Uhrig, Fuzzy and eural Approaches in Engineering, John Wiley & Sons, Inc [] A. Bernieri, M. D Apuzzo, L. Sansone, M. Saastano, A eural etwor Approach for Identification and Fault Diagnosis on Dynamic Systems, IEEE Trans. Instrumentation and Measurement, ol. 43, no. 6, Dec. 994, pp