USE OF ARTIFICIAL INTELLIGENCE TECHNIQUES AND COMBINED ...
XVIII Congresso Brasileiro de Automática / 12 a 16-setembro-2010, Bonito-MS
USE OF ARTIFICIAL INTELLIGENCE TECHNIQUES AND COMBINED APPROACHES IN INDUCTION MOTOR DIAGNOSIS 1 EDISON R. C. DA SILVA , HUBERT RAZIK2, LANE M. R. BACCARINI3, MAURÍCIO B. DE R. CORRÊA1, AND CURSINO B. JACOBINA1 1
Laboratório de Eletrônica Industrial e Acionamento de Máquinas (LEIAM)/DEE/UFCG Caixa Postal 10105; 58109-970 Campina Grande, PB, Brasil 2 Université Lyon 1 Bât. OMEGA 43 bd du 11 Novembre 1918, 69622 Villeurbanne cedex, France 3 Departamento de Engenharia Elétrica Universidade Federal de São João Del Rei – UFSJ, Brasil
E-mails: edison,mbrcorrea,[email protected],[email protected],
[email protected]
Abstract: The induction motor has become a key component in many industrial plants and in a large number of applications it is supplied by a voltage-source inverter. In spite its advantages, various stresses natures like thermal, electrical, mechanical or environmental could affect the life span of this induction motor drive. In recent years, monitoring and fault detection of electrical machines have moved from traditional techniques to artificial intelligence (AI) techniques. This paper describes the various steps and highlights functions that can be accomplished by using neural networks, fuzzy logic, neural-fuzzy, genetic algorithms, artificial immune system, vector support machine, particle swarm optimization, and gaussian bootstrap process techniques. Keywords: Diagnosis technique, Artificial Intelligence methods, Induction motor fault diagnosis Resumo: O motor de indução tornou-se um equipamento chave em muitas instalações industriais. Além disso, em uma planta industrial, muitos motores são acionados por conversores fonte de tensão. Apesar de suas vantagens, vários estresses térmicos, elétricos, mecânicos ou ambientais, podem afetar seu tempo de vida. Recentemente, o monitoramento e detecção de faltas migraram das técnicas tradicionais para as técnicas de inteligência artificial. Este trabalho descreve várias etapas e evidencia funções que ocorrem com o uso das técnicas de redes neurais, lógica fuzzy, combinações neuro-fuzzy, algoritmos genéticos, sistemas artificiais, máquina de vetores de suporte, otimização através de enxame de partículas e processo de bootstrap gaussiano. Palavrs-chave: Técnicas de diagnóstico, Métodos de inteligência artificial, diagnóstico de faltas em motor de indução.
1
fault identification, and fault severity evaluation and have been focused on for some decades. Different techniques have been developed to accomplish the required tasks for the converter or motor diagnosis, based on the key fault types normally verified in the industry applications. Although monitoring and detecting the converter faults is an important issue (Thorsen and Dalva, 1995), this paper will deal with the monitoring and fault detection of electrical machines. It is important to note that line current signature has been widely used to deal with faults occurring in the stator and the rotor of asynchronous machines and that the frequency components feature can be associated with different rotor faults. One can find in the motor theory that broken bar faults, as well as eccentricity, rotor asymmetry or shaft speed oscillation, show sideband frequencies. In recent years, the monitoring and fault detection of electrical machines have moved from traditional techniques to artificial intelligence techniques [(Fillipetti et al, 1998) (Awadallah and Morcos, 2003) (Siddique et al., 2003) (Qiang et al., 2003)]. When an AI technique is used, fault detection and evaluation can be accomplished without an expert. Among all these approaches used for the diagnostics, some are based on fuzzy logic, neural networks or on the mixed neuro-fuzzy logic.
Introduction
The induction motor has become a key component in many industrial plants. This is due to its reputation of robustness and its low cost of manufacture. In many industrial applications the asynchronous machine is supplied by a voltage-source inverter. In spite its advantages, various stresses natures like thermal, electrical, mechanical or environmental could affect the life span of this induction motor drive. This may cause faults occurring at the power converter stage or at the machine. Among all defects, a three-phase induction motor drive could generate three kinds of problems: rotor (broken rotor bar or end rings, eccentricity), bearing faults, stator (inter-turn or inter-phases short-circuits or disconnection of one phase) (Fuchs, 2003). In the converter could occur: short-circuit or open-circuit in one or more switches, intermittent misfiring (Thorsen and Dalva, 1995). The cost of stopping the drive system for an unplanned maintenance schedule, due to faults, can be high. Knowing that industrial constraints are strong, the reliability and the safe operating system have to be considered. Because of this, many diagnostic procedures have been proposed. Main steps of a diagnostic procedure can be classified as signature extraction,
2024
XVIII Congresso Brasileiro de Automática / 12 a 16-setembro-2010, Bonito-MS
This paper describes the various steps and highlights functions that can be accomplished by using AI techniques. Next, the paper presents some examples of using AI in the diagnostics of an induction machine.
• training based on both time and frequency domain signals obtained via simulation and/or experimental results; • real time, online unsupervised diagnosis; • dynamic updating of the structure with no need to retrain the whole network; • filtering out transients, disturbances, and noise; • fault prediction in incipient stages due to operation anomalies; • operating conditions clustering based on fault types. Example In the following an example of pattern recognition is described: ANN is applied to vibration signals in order to detect mechanical faults. Unbalance, shaft misalignment, and mechanical looseness have been compared with the “healthy” operation for two different ANN techniques: MLP global and PLP (Baccarini 2005). A multi-objective method has been used to improve the generalization capacity of the global MLP network. For investigation, a random choice of training and validation groups was done but took into account the sets dimension: 67% for training patterns and 33% for validation. Deterministic frequencies (fr, 2fr, 3fr, 4fr), and measurement with sensors in six different positions for four cases (healthy, unbalance, misalignment, and mechanical looseness) have been considered, in a total of 978 training and 312 validation patterns.
2 AI-BASED TECHNIQUES There are many types of AI-based techniques. Some of these use expert systems, artificial neural networks (ANNs), fuzzy logic, genetic algorithms (GAs), support vector machines (SVM). Besides giving improved performance these techniques are easy to extend, modify, and combine (with particle swarm, and/or wavelet, for instance). They can also be made adaptive by the incorporation of new data or information (Ivonne et al 2005). Next, considerations on the techniques and examples as well, will be given. 2.1 Expert Systems (ES) The expert system is basically a computer program embodying knowledge about a narrow domain for the solution of problems related to that domain. An ES mainly consist of a knowledge base and an inference mechanism. The knowledge base contains domain knowledge, which may be expressed as any combination of "IF-THEN" rules, factual statements, objects, procedures and cases, while the inference mechanism manipulates the stored knowledge for producing solutions. The system can determine a fault situation doing the signals extraction and fault identification from the combined derived information from behavior of various harmonic components and the machine operating conditions (Rajogopalan et al., 1991). A demerit of ordinary rule-based ES is that they can not handle new situation not covered explicitly in their knowledge bases. These ES can not give any conclusions in these situations.
A. Global MLP Network The Global MLP network, Fig.1, has two binary outputs for four situations: 00 – healthy; 01 misalignment; 10 unbalance; 11 mechanical looseness. The layers activation functions are sigmoid and weights have also been updated via back propagation. It was observed that the best result was obtained for the sensor in the vertical position, in which the vibration levels are more significant in case of mechanical faults. The total success rate was 86.16% for training and 82.37% for validation. A well trained network must adequately respond not only to the pattern used for training but also to all other submitted to them. This is known as network generalization capacity. At the training stage the generating function of data is based on possible realizations of the training sets for same task. This variety of solutions is named variance, which must be minimized to guarantee a good network generalization. On the other hand the number of possibilities increases with the model dimension. A reduction of dimension, by reducing the number of parameters, solves this problem. However, it can originate bias, which reduces the generalization capacity of the network. It is said that the bias occurs when even for different realizations of the training process in the reduced dimension space the solution is practically the same. Bias of solutions must be minimized to preserve the generalization capacity. Therefore, a point of equilibrium must be achieved.
2.2 Artificial Neural Network (ANN) An ANN is a computational model of the brain. It assumes that computation is distributed over several simple units called neurons, which are interconnected and operate in parallel, thus known as parallel distributed processing systems. Implicit knowledge is built into a neural network by training it. ANN can be trained by typical input patterns and corresponding expected output patterns. The error between the actual and expected output is used to strengthen the weights of the connections between the neurons. Awadallah and Morcos (2003) remind that ANNs have been densely applied in the area of motor condition monitoring and fault diagnosis performing one or more of the following tasks: • pattern recognition, parameter estimation, and nonlinear mapping applied to condition monitoring;
2025
XVIII Congresso Brasileiro de Automática / 12 a 16-setembro-2010, Bonito-MS
Fig.1: Scheme for the global MLP network.
A multi-objective system approach (Teixeira et al, 2002) allows for equilibrium between the weight norm and the training error of MLP network, and guarantees the generalization capacity of the model. Once a large topology is defined, the algorithm generates a set of solutions with a variety of norms and minimized error for each one, and selects the best response in relation to the validation set. Calculation of network efficiency confirmed the vertical position of the sensor as the most significant, and reached a success rate of 98%.
Fig. 2 The PLN scheme
A fuzzy system is based on the three classical steps which are: the fuzzification step, the inference engine and the defuzzication step. The inputs of the fuzzy expert system are the controlled variables. All inputs can be bounded and normalized. Each input variable is described by membership functions (Small, Medium and High) which can be triangular or take other function shapes. Classically, the inference engine is based on the Max-Min method. The output membership functions are described by Dirac functions (False, True). The evaluation of the centre (center) of gravity is thus easier to compute than using triangular or other functions shapes. The inference engine is based on AND functions and OR function which is described as follows: - For the AND function applied to the two inputs A and B, the output is evaluated as min(μA, μB); - For the OR function applied to the two inputs A and B, the output is evaluated as max(μA, μB). Awadallah and Morcos (2003) presented an extensive list of references, indicating some of the fuzzy and adaptive-fuzzy systems applications to motor fault diagnosis: •evaluating performance using linguistic variables; •predicting abnormal operation and locating faulty element; •utilizing human expertise reflected to fuzzy if—then rules; •system modeling, nonlinear mapping, and optimizing diagnostic •fault classification and prognosis. Example The aim of the example is the diagnosis of signatures of rotor broken bars when the induction machine is fed by an unbalanced line voltage. These signatures are given by the complex spectrum modulus of the line current. The fuzzy logic approach allows us to conclude to the load level operating system as to inform the operator of the rotor fault severity. The sense of velocity can be determined easily using the fuzzy logic. The rules are as follows:
B. Parallel Layers Network (PLN). A Parallel Layers Perceptron was proposed in Caminhas et al (2003) and replaces the original input of the ANFIS net by parallel perceptrons, Fig. 2. Its main objective is to overcome the limitation of working with multiple inputs imposed by the classical ANFIS network due to the resulting exponential increase of operations when each input is combined. Again, the vertical position of the sensor is confirmed as the most significant and the success rate percentage was 94%. 2.3. Fuzzy Logic System (FLS) The FLS are based on a set of rules. One advantage of FLS is that the rules allow the input to be fuzzy, i.e. more like the natural way that human express knowledge. In contrast to ANNs, they give a very clear physical description of how the function approximation is performed (since the rules show clearly the function approximation mechanism). Reasoning procedures, the compositional rule of inference, enable conclusion to be drawn by extrapolation or interpolation from the qualitative information stored in the knowledge base. The fuzzy approach model is a complex problem employing an IF-THEN type of expert rule and linguistic variables to capture directly the qualitative aspects of the human reasoning process involved. However, the problem is shifted to the membership function and rule tuning.
IF (If6 is ) AND (If3 is <small>) THEN Velocity is OR
2026
XVIII Congresso Brasileiro de Automática / 12 a 16-setembro-2010, Bonito-MS
IF (If6 is <small>) AND (If3 is ) THEN Velocity is .
implementing and optimizing fuzzy reasoning via ANN structures (Yoshikazu and Yoshiteru 2003).
Moreover, it is possible to monitor the connection of the motor to the line. A problem occurs when all the currents in the forward and backward have abnormal amplitude. The set of rules can be as follows:
2.5. Support Vector Machines The data-based machine learning is an important aspect of modern intelligent technology, while Statistics Learning Theory (SLT) is a new tool that studies the machine learning methods in the case of a small number of samples. Support Vector Machine (SVM) is derived from the SLT, and has attractive features, such as good generalization ability large dimension robustness, objective function convexity, and well established theory. In fact, SVM based classifier is claimed to have an efficiency does not depend on the number of features of classified entities, to have better generalization properties, cost much less time than NN based classifiers and better accuracy (greater than 97%) than Linear Discriminant analysis, K-Nearest Neighbor, Probabilistic Neural Network, Gaussian Mixture Model pattern recognition techniques (Namburu et al, 2007). Increase in SVM classification accuracy and speed continue a challenge since choices in SVM implementations, such as kernel function and penalty parameters of the support vector, may drastically affect its accuracy. An improvement can be obtained by tuning these parameters with other optimization techniques, like GA (Namburu et al, 2007), artificial immune system (Aydin et al, 2007), or on particle swarm optimization (Yuan and Chu, 2007). The SVMs are essentially binary classifiers (positive and negative classes). However, SVM-based multi-class classifier can be constructed using “one against one” technique, which consists in creating k SVMs, k corresponding to the number of classes. In the generation of each machine, a class is fixed as positive while the other are considered as negative. However, the use of “one against one” technique needs a synthesizing scheme to decide the final results according to the results of sub classifiers. In Fang and Ma (2006) four synthesizing schemes were compared (majority voting; binary tree decision; neural network and hybrid matrix), while in Mayoraz and Alpaydin (1999) ten of them were needed. Depending on what is required in motor diagnosis, there are two possibilities. The first, called simple diagnosis (1-Class), discovers only if the fault has occurred. The second one (complex diagnosis, 2Class) is able to find, for instance, how many bars have been damaged (Kurek and Osowski, 2008). The SVM is gaining application in rotating machinery anomaly detection, due to its superb performance with small samples [(Ortiz and Syrmos 2006), (Rojas and Nandi, 2006)]. Accurate online support vector regression (AOSVR) technique was, for the first time, implemented for machine condition monitoring with applications to motor shaft misalignment (Olufemi, 2007). AOSVR is an algorithm that combines the advantages of SVR with the capability of efficiently updating
IF (If6 is ) AND (If3 is ) THEN Current line is OR IF (If6 is <small>) AND (If3 is <small>) THEN Current line is .
By the same way, we can determine the load level thanks to these three rules: IF (s is <small>) THEN Load level is <Small> OR IF (s is <medium>) THEN Load level is <Medium> OR IF (s is ) THEN Load level is .
The detection of the rotor fault is based on the ratio of the sum of the two sideband currents on fundamental line current. This approach is advocated from one decade and gives a sufficient precision about the fault severity. As we study the forward and the backward sequence of the supply line, we have to calculate this ratio for the two sequences: Irf and Irb. These results will be the additional inputs for the expert system. The set of rules for the expertise of three phase induction motor, in case of broken bar defect, are given by: IF (Irf is <small>) AND (If6 is ) AND (If3 is <small>) THEN Operating Condition is OR IF (Irb is <small>) AND (If6 is <small>) AND (If3 is ) THEN Operating Condition is OR IF (Irf is <medium>) AND (If6 is ) AND (If3 is <small>) THEN Operating Condition is OR IF (Irb is <medium>) AND (If6 is <small>) AND (If3 is) THEN Operating Condition is < Fault in progress > OR IF (Irf is ) AND (If6 is ) AND (If3 is <small>) THEN Operating Condition is OR IF (Irb is < high >) AND (If6 is <small>) AND (If3 is ) THEN Operating Condition is < Broken Bar >
Thanks to these sets of rules, we are able to evaluate several diagnosis indexes. However, this technique requires the knowledge of the behavior for the determination of membership functions. 2.4. Neural-Fuzzy Consider the case of a motor stator faults. NeuralFuzzy fault detection is obtained which learns the stator faults and the condition under which they occur through an inexperienced and noninvasive procedure. The Neural-Fuzzy system is an ANN structured upon fuzzy logic principles, which enables this system to provide qualitative description about the machine condition and the fault detection process. The knowledge is provided by the fuzzy parameters of membership functions and fuzzy rules. The idea behind the fusion of these two technologies is to use the learning ability of ANN to implement and automate the fuzzy system, which use the high-level human-like reasoning capability. Many methods have been proposed for
2027
XVIII Congresso Brasileiro de Automática / 12 a 16-setembro-2010, Bonito-MS
trained support vectors whenever a sample is added to the training set. Example The fault diagnosis of induction machine is a multi-class classification problem. For classification of mechanical faults, as in the ANN example, have been compared to the “healthy” operation (Baccarini, 2005), a software developed by (Poyhonen et al 2002) was used because this work deals with four classes, the technique one-against-all was used in order to group the classes. SVM, together with one-against-all technique, has shown an excellent performance reaching a success rate of 98% for the already mentioned most significant position (vertical). For all cases, the Radial Bases Function (RBF) with variance between 0.01 and 0.1 was used as a kernel.
without any change in the children generation (Selection). Two children are next obtained thanks to an artificial reproduction using the genes of two parents (Crossover). The parents with better fitness are selected to take part in reproduction. In Mutation, the genes of one parent are randomly altered to give a child. Finally, all individuals of low Fitness are replaced by new ones, randomly computed. Example In [Razik et al (2008)(2009)] the GA is used to find the global maximum as well as to solve an optimization problem in the spectral lines identification process in case of a rotor broken bar of an induction motor. The N individuals are the supply frequency and the slip frequency and these are the parameters to be found. In order to find the eight main components in the current spectrum in Fig. 3, eight Gaussian functions were used as a window, which only depends on the supply frequency (f s) and the slip frequency (sfs) The fitness is calculated as being the integral of the product of the current spectrum by the spectral window, that is,
2.6. Genetic Algorithms Genetic Algorithm (GA) is based on biology and, in particular, by those biological processes that allow populations of organisms to adapt to their surrounding environment: genetic inheritance and survival of the fittest, that is, natural selection as well as evolutionary process. GA is a stochastic optimization method and needs less prior information about the problems to be solved than the conventional optimization schemes, which often require the derivative of objective functions. It also has the unique features of parallel search and global optimization and it is adapted for the simultaneous evaluation of a large number of points in the search space. GA can be used to determine the coefficients of a regulator (Cvetkoski et al, 1998) or to identify induction machine parameters. It can also be used in the diagnosis of induction motor rotor and stator faults, such as rotor broken bars [(Razik et al (2008) (2009)], open rotor and stator phase (Cristaldi et al 2004), rotor unbalance and misalignment, and bearing loose fault (Wei et al, 2007b). Although GA based approaches have interesting features when used alone, as compared to NeuroFuzzy (NF) based approaches, for instance (Cristaldi et al 2004), GA combined with other AI based approaches will have tremendous scope in future. This is due to the fact that the combination of GA with other motor fault diagnosis schemes has demonstrated enhanced performance in global and near-global minimum search. The GA uses three fundamental operations which are: Selection, Crossover and Mutation. First, an initial population is defined by N individuals created through a random generator. Each one is constituted by a parameters vector, of which the elements are called genes. These are the parameters for the optimization process, to which constraints can be added. The quality of an individual, to be adequate to the problem, is characterized by its fitness F, evaluated, for each individual, via an objective function. Then the parent individuals with the highest fitness are copied
Fitness = ∫
+100
−100
Spectrum( f ).ω ( f )df
where the spectral window is 8 1 ( f − fi )2 ω ( f ) = ∏ exp(− 2 σ2 i −1
(1)
(2)
The approach was tested for a rotor broken bar fault detection composed of four stages. First the acquisition of line currents was done, followed by the calculation of the Concordia’s vector. GA was used in the search lines of the supply frequency and the slip frequency which are inside of the Concordia’s vector spectrum. Fig. 4 depicts the results for line currents spectrum with one full broken bar and 100% of full load level. Results are obtained after few interactions. A comparison with the healthy line currents spectrum allows the operator to distinguish the faulty case and also to be aware of a growing rotor fault. The fault severity detection is based on a fuzzy approach. 2.7. Artificial Immune System The immune system (IS) is an efficient self-defense method that guards the human body from foreign antigens or pathogens. Artificial immune system (AIS) is an emerging soft computing method inspired by natural immune system. Because the AIS abilities of learning, memory and self adaptive control, this method is used in pattern recognition, classification, optimization and anomaly detection problems. Clonal selection is an artificial immune algorithm that is used for optimization problems. It has not crossover operator, so this method is different from genetic algorithm. Affinity proportional reproduction and affinity maturation are two distinctive properties of clonal selection. Because of these properties, clonal selection converges faster than genetic algorithm and doesn’t catch local minimum.
2028
XVIII Congresso Brasileiro de Automática / 12 a 16-setembro-2010, Bonito-MS
one shares its own experience to others particles, they move globally into the search space along a search trajectory. The speed for particles is given by: vik +1 = ωvik + c1 rand (⋅)( pi − xik ) + c 2 rand (⋅)( p g − xik ) where w is the inertia weight, pg is the position of the best particle among all particles, so it is the global solution. pi is the best previous of particle and vi denotes the velocity of the ith particle. rand are random numbers with uniform distribution in the range [0; 1], c1 is the cognition parameter and c2 is the social parameter. This equation updates the particles speeds based on particle momentum, the attraction force towards the global and the best local. In order to improve the performance of the PSO algorithm different modifications have been introduced. A constriction factor was proposed by (Clerc 1999) thus modifying the speed equation. Another improvement was proposed in (Shi and Eberhart, 1999) with the introduction of a time varying inertia weight (TVIW) as a function of the maximum and minimum values of inertia weight, and of the maximum number of search iteration. The algorithm prioritizes the global search. At the end of the search process, inertia decreases linearly to its minimal value prioritizing the local search. Also, in (Liu et al, 2005) was proposed an adaptive algorithm so that the inertia weight is different for each particle. Consequently, particles near the optimal solution refine the results and particles far from the optimal solution continue to explore the search space. PSO method has been applied to the following motor faults: broken bar, part of the end-ring broken (Razik et al, 2009), stator inter-turn and independent winding short circuit (Ethny et al, 2006). PSO method is an efficient method to solve the optimization problem as to extract information quickly from a frequency. Next it will be shown its ability to estimate the line frequency and the fault line frequencies with the induction motor operating under one full broken bar (Razik, 2009b). Example Based on a spectrum of current and calculation of the window function and of Fitness with the same equations used in GAS method, the transient of the PSO was examined thanks to several variables. In the estimation of the fundamental frequency the majority of the particles stayed close to the fundamental frequency of 50 Hz after about twenty iterations. To estimate the slip s, the population moved along the two sidebands to be found. The slip being close to 8.7%. Both the particle best fitness and the global best performance have been reached after 20 iterations. Bad performance particles disappeared with the increasing value of the number of iterations.
Fig. 3. The stator line currents spectrum in case of one broken bar.
Fig. 4. Stator line currents spectrum in case of one broken bar and the induction motor operating at full load, obtained by using GA.
In combination with other diagnosis methods, AIS can effectively improve the accurate rate of fault diagnosis and diagnosis system robustness, bringing into play each of their advantage, so that the accurate rate is improved. Clonal selection has been used to select optimal parameters of SVM, extracted from three phase motor current and constructed based on Park’s vector approach. This has been applied to the study of broken rotor bar and stator short circuit faults (Aydin et al., 2007). Also, for rotating machinery, AIS method has been combined with neural network method for machine fault diagnosis using genetic algorithm to combine diagnosis methods to make each kind of diagnosis method display its advantage in optimal space (de Castro and Timmis, 2002), (Wei et al 2007a). 2.8. PSO Algorithm Particle swarm optimization (PSO) is a semi-global optimization algorithm, first introduced by Kennedy and Eberhart (1995). It simulates social model like those of birds, insects and fish swarm. Its main concept is simulating the movement of these organisms searching for food. PSO is a simple optimization technique without heavy computation and has shown success in solving many optimization problems (Ciuprina et al., 2002), (Razik et al (2009). The technique does not require the computation of derivatives and hessians nor does not need training with heuristic data. Candidates to find the best solution are particles. They share their experience and collaborate to each other suggesting its own solution for the problem. As each
2.9 Bootstrap Gaussian Process Bootstrap Gaussian Process (BGP) has been proposed from the merge of Gaussian process classifiers (GPCs) and bootstrap methods, as an alternative to other classifiers, like the kernel classifier support 2029
XVIII Congresso Brasileiro de Automática / 12 a 16-setembro-2010, Bonito-MS
other classifiers, like the kernel classifier support vector machine (SVM), which has excellent performance towards this purpose, but it has difficulties to optimize relevant hyper-parameters. GPCs are Bayesian probabilistic kernel classifiers and provide a well established Bayesian framework to determine the optimal or near optimal kernel hyper-parameters. They are largely unexplored for anomaly detection applications and, also, a promising statistical tool for both binary and multi-category classification. Moreover, GPCs proved to outperform SVM (Kim et al., 2006). It can be employed to solve a wide range of problems, such as hypothesis tests, model selection and probability distribution estimations. Bootstrap is most useful where little is known about the statistics of the data or too few samples are available to use asymptotic results (Zoubir and Iskander, 2007). In BGP, bootstrap methods are incorporated to improve GPCs’ performance for small machinery anomaly samples by re-sampling at random. The fact that GPCs are strong classifiers suggests that small numbers of bootstrap samples might be sufficient to enhance classification performance. Experiment results for rotating machinery misalignment anomalies detection (Xue et al., 2008) in which wavelet packet is utilized to perform vibration analysis, show that bootstrap GPCs are highly effective and outperform GPCs and SVM with cross validation for anomaly detection. Thus the proposed approach is promising for rotating machinery anomaly detection.
References Awadalla M.A, and Morcos, M. M. (2003). Application of AI Tools in Fault Diagnosis of Electrical Machines and Drives-An Overview. IEEE Trans. on Energy Conversion, Vol. 18 (2), pp. 245-252. Aydin, I; Kakaköse, M; Akin, E. (2007). Artificial Immune Based Support Vector Machine Algorithm for Fault Diagnosis of Induction Motors. ACEMP, pp. 217-221. Baccarini L.M.R (2005). Detecção e Diagnóstico de Falhas em Motores de Indução. Universidade Federal de Minas Gerais. Caminhas, W.M; Vieira, D.A.G. and Vasconcelos, J. A. (2003). Parallel layer perceptron. Neurocomputing, pp. 771 – 778. Clerc, M. (1999). The Swarm and the Queen: Towards a Deterministic and Adaptive Particle Swarm Optimization. IEEE Congress on Evolutionary Computation, pp. 1951–1957. Cristaldi L.; Lazzaroni, M.; Monti, A.; Ponci,F. and Zocchi, F. (2004). A Genetic Algorithm for fault identification in electrical drives: a comparison with Neuro-Fuzzy computation. Instrumentation and Measurement Technology Conference IMTC, Vol. 2, pp. 1454-1459. Ciuprina, G.; Loan, D. and Munteanu, I (2002). Use of intelligent-particle swarm optimization in electromagnetic. IEEE Trans. Magnetics, Vol. 38, pp. 1037-1040. Cvetkoski, G.; Petkovsk, L. and Cundev, M. (1998). Mathematical model of a permanent magnet axial field synchronous motor for a genetic algorithm optimization. ICEM, Istanbul, pp. 1172–1177. De Castro, L.N. and Timmis, J. (2002). Artificial Immune Systems: A New Computational Intelligence Approach. Springer Verlag, New York. Ethny, S.; Acarnley, P.P.; Zahawi, B. and Giaouris, D. (2006). Induction Machine Fault Identification using Particle Swarm Algorithms. PEDES, pp. 14. Fang, R. and Ma, H.Z. (2006). Application of MCSA and SVM to Induction Machine Rotor Fault Diagnosis. Proceedings of the 6th World Congress on Intelligent Control and Automation, Vol. 2, pp. 5543-5547. Fuchs, F.W. (2003). Some diagnosis methods for voltage source inverters in variable speed drives with induction machines - A Survey. IEEE Ind. Electron. Conf., pp. 1378-1385. Filippetti, F; Franceschini, G.; Tassoni, C. and Vas, P. (1998) AI Techniques in Induction Machines Diagnosis Including the Speed Ripple Effect. IEEE Transactions on Industry Applications, Vol. 34 (1), 98-108. Ivonne, Y. B. ;Sun D and He, Y.K. (2005). Fault Diagnosis Using Neural-Fuzzy Technique Based on the Simulation Results of Stator Faults for a Three-Phase Induction Motor Drive System. ICEMS, pp. 1966-1971.
3. CONCLUSION In this paper, an overview on Artificial Intelligence (AI) methods-based motor fault diagnosis systems has been given. Several techniques using neural networks, fuzzy logic, neural-fuzzy, genetic algorithms, artificial immune system, vector support machine, particle swarm optimization, and Gaussian bootstrap process were summarized. Their applications and possibilities of combination were discussed as well. AI techniques are a very strong tool for electrical motors diagnosis studies. Although some investigators indicate that they are not yet supposed to compete with conventional methods, tremendous efforts have been made to develop new methods, as it is the case of bootstrap gaussian process. One observation is that AI methods become a strong tool when used in combination with other ones. ACKNOWLEDGEMENT Authors would like to thank CNPq (Conselho Nacional de Pesquisa e Desenvolvimento), FAPESQ (Fundação de Amparo à Pesquisa – Paraíba), CAPES, and COFECUB, for the financial support.
2030
XVIII Congresso Brasileiro de Automática / 12 a 16-setembro-2010, Bonito-MS
Kennedy, J and Eberhart, R. (1995). Particle Swarm Optimization. IEEE International Conference on Neural Networks, pp. 1942-1948. Kim, H.C; Kim, D. and Ghahramani,.Z. And Bang, S.Y. (2006). Appearance-based gender classification with Gaussian processes. Pattern Recognition Letters, Vol. 27, pp. 618–626 Kurek, J. and Osowski, S. (2008). Support Vector Machine for Diagnosis of the Bars of Cage Inductance Motor. ICESS’08, pp. 1022-1025. Liu, B.; Wang, L.; Jin, Y-H ; Tang, F and Huang, D.X. (2005). Improved particle swarm optimization combined with chaos. Elsevier Chaos, Solitons and Fractals, Vol. 25(5), pp. 1261–1271. Mayoraz, E and Alpaydin, E. (1999). Support Vector Machines for Multi-Class Classification. Int. Workshop on Artificial Neural Networks, Vol. 2 (4), pp. 833-842. Namburu, S.M; Chigusa, S.; Prokhorov, D; Qiao, L; Choi, K and Pattipati, K. (2007). Application of an Effective Data-Driven Approach to Real-time Fault Diagnosis in Automotive Engines. IEEE AC, pp. 1-9. Olufemi, A.O., Myong K.J; Adedeji, B.B. and Hines, J.W. (2007). Online Support Vector Regression Approach for the Monitoring of Motor Shaft Misalignment and Feedwater Flow Rate IEEE Trans. On Syst, Man, and Cybernetics. Applications and Reviews, Vol. 37(5), pp. 962-970. Ortiz, E. and Syrmos, V. (2006). Support vector machines and wavelet packet analysis for fault detection and identification. Int. Joint Conf. on Neural Networks, Vancouver, pp. 3449–3456. Poyhonen, S.: Negrea, M.; Arkkio, A; Yotyniemi, H. and Koivo, H. (2002). Support Vector Classification for Fault Diagnosis of in Electrical Machine. ICSP'02, pp. 1719-1722. Qiang, S., Gao, X.Z. and Zhuang, X.(2003). State-ofthe-art in Soft Computing-based Motor Fault Diagnosis. CAC 2003 pp. 381-1386. Rajogopalan, V; Debebe K. and Sankar T.S. (1991). Expert system for fault diagnosis of VSI fed AC drives. IEEE-IAS Annual. Meeting, Dearborn, MI, pp. 368–373. Razik, H; Corrêa, M.B.R. and Silva, E.R.C. (2008). An application of Genetic Algorithm and Fuzzy Logic for the induction motor diagnosis. IEEE IECON, pp. 3067-3072. Razik, H; Corrêa, M.B.R. and Silva, E.R.C.(2009a). A Novel Monitoring of Load Level and Broken Bar Fault Severity Applied to Squirrel-Cage Induction Motors Using a Genetic Algorithm. IEEE Trans. Ind.ustrial Electronics, Vol. 56 (11), pp. 4615-4626. Razik, H; Corrêa, M.B.R. and Silva, E.R.C.(2009b). The use of particle swarm optimization for the tracking of Induction motor faulty lines. Proc. of IEEE PowerEng’09, pp. 680-684. Rojas, A. and Nandi, A.K. (2006). Practical scheme for fast detection and classification of rollingelement bearing faults using support vector ma-
chines. Mechanical Systems and Signal Processing, Vol. 20 pp. 1523–1536. Shi, Y. and Eberhart, R.C. (1999). Empirical Study of Particle Swarm Optimization. Congress on Evolutionary Computation, pp. 1945–1950. Siddique, A., Yadava, G. S. and Singh, B. (2003). Applications of Artificial Intelligence Techniques for Induction Machine Stator Fault Diagnostics: Review. SDEMPED´03, pp. 29-34. Teixeira, R.A; Braga, A.P; Takahashi, R.H.C. and Saldanha, R.R. (2002). Decisor implementation in neural model selection by multiobjective optimization. SBRN, pp. 234–239. Thorsen, O.V. and Dalva, M. (1995). A Survey of the Reliability with an Analysis of Faults on Variable Frequency Drives in the Industry. IEEE Trans. Industry Applications, Vol. 31(5), pp. 1.1861.196. Yoshikazu, F. and Ueki, Y. (2003). Fault Analysis System Using Neural Networks and Artificial Intelligence. IEEE ANNPS´03, pp. 20-25. Yuan, S.F. and Chu, F.L. (2007). Fault diagnostics based on particle swarm optimization and support vector machines. Mechanical Systems and Signal Processing, Vol. 21, pp. 1787–1798. Wei D.; Zhan-sheng L. and Dong-hua W. (2007a). Combination Diagnosis Based on Genetic Algorithm for Rotating Machinery. Third International Conference on Natural Computation, pp. 307309. Wei, D; Zhan-sheng, L. and Xiaowei, W. (2007b). Application of Image Recognition Based on Artificial Immune in Rotating Machinery Fault Diagnosis. International Conference on Bioinformation and Bio-medical Engineering, ICBBE’07, pp. 1047-1052. Xue, W., Daowei Bi, Liang D., and Sheng W. (2008). Bootstrap Gaussian Process Classifiers for Rotating Machinery Anomaly Detection. IEEE International Joint Conference on Neural Networks, pp. 1129-1134. Zoubir, A.M and Iskander, D.R. (2007). Bootstrap methods and applications. IEEE Signal Processing Magazine, Vol. 24 (4), pp. 10–19.
2031
USE OF ARTIFICIAL INTELLIGENCE TECHNIQUES AND COMBINED APPROACHES IN INDUCTION MOTOR DIAGNOSIS 1 EDISON R. C. DA SILVA , HUBERT RAZIK2, LANE M. R. BACCARINI3, MAURÍCIO B. DE R. CORRÊA1, AND CURSINO B. JACOBINA1 1
Laboratório de Eletrônica Industrial e Acionamento de Máquinas (LEIAM)/DEE/UFCG Caixa Postal 10105; 58109-970 Campina Grande, PB, Brasil 2 Université Lyon 1 Bât. OMEGA 43 bd du 11 Novembre 1918, 69622 Villeurbanne cedex, France 3 Departamento de Engenharia Elétrica Universidade Federal de São João Del Rei – UFSJ, Brasil
E-mails: edison,mbrcorrea,[email protected],[email protected],
[email protected]
Abstract: The induction motor has become a key component in many industrial plants and in a large number of applications it is supplied by a voltage-source inverter. In spite its advantages, various stresses natures like thermal, electrical, mechanical or environmental could affect the life span of this induction motor drive. In recent years, monitoring and fault detection of electrical machines have moved from traditional techniques to artificial intelligence (AI) techniques. This paper describes the various steps and highlights functions that can be accomplished by using neural networks, fuzzy logic, neural-fuzzy, genetic algorithms, artificial immune system, vector support machine, particle swarm optimization, and gaussian bootstrap process techniques. Keywords: Diagnosis technique, Artificial Intelligence methods, Induction motor fault diagnosis Resumo: O motor de indução tornou-se um equipamento chave em muitas instalações industriais. Além disso, em uma planta industrial, muitos motores são acionados por conversores fonte de tensão. Apesar de suas vantagens, vários estresses térmicos, elétricos, mecânicos ou ambientais, podem afetar seu tempo de vida. Recentemente, o monitoramento e detecção de faltas migraram das técnicas tradicionais para as técnicas de inteligência artificial. Este trabalho descreve várias etapas e evidencia funções que ocorrem com o uso das técnicas de redes neurais, lógica fuzzy, combinações neuro-fuzzy, algoritmos genéticos, sistemas artificiais, máquina de vetores de suporte, otimização através de enxame de partículas e processo de bootstrap gaussiano. Palavrs-chave: Técnicas de diagnóstico, Métodos de inteligência artificial, diagnóstico de faltas em motor de indução.
1
fault identification, and fault severity evaluation and have been focused on for some decades. Different techniques have been developed to accomplish the required tasks for the converter or motor diagnosis, based on the key fault types normally verified in the industry applications. Although monitoring and detecting the converter faults is an important issue (Thorsen and Dalva, 1995), this paper will deal with the monitoring and fault detection of electrical machines. It is important to note that line current signature has been widely used to deal with faults occurring in the stator and the rotor of asynchronous machines and that the frequency components feature can be associated with different rotor faults. One can find in the motor theory that broken bar faults, as well as eccentricity, rotor asymmetry or shaft speed oscillation, show sideband frequencies. In recent years, the monitoring and fault detection of electrical machines have moved from traditional techniques to artificial intelligence techniques [(Fillipetti et al, 1998) (Awadallah and Morcos, 2003) (Siddique et al., 2003) (Qiang et al., 2003)]. When an AI technique is used, fault detection and evaluation can be accomplished without an expert. Among all these approaches used for the diagnostics, some are based on fuzzy logic, neural networks or on the mixed neuro-fuzzy logic.
Introduction
The induction motor has become a key component in many industrial plants. This is due to its reputation of robustness and its low cost of manufacture. In many industrial applications the asynchronous machine is supplied by a voltage-source inverter. In spite its advantages, various stresses natures like thermal, electrical, mechanical or environmental could affect the life span of this induction motor drive. This may cause faults occurring at the power converter stage or at the machine. Among all defects, a three-phase induction motor drive could generate three kinds of problems: rotor (broken rotor bar or end rings, eccentricity), bearing faults, stator (inter-turn or inter-phases short-circuits or disconnection of one phase) (Fuchs, 2003). In the converter could occur: short-circuit or open-circuit in one or more switches, intermittent misfiring (Thorsen and Dalva, 1995). The cost of stopping the drive system for an unplanned maintenance schedule, due to faults, can be high. Knowing that industrial constraints are strong, the reliability and the safe operating system have to be considered. Because of this, many diagnostic procedures have been proposed. Main steps of a diagnostic procedure can be classified as signature extraction,
2024
XVIII Congresso Brasileiro de Automática / 12 a 16-setembro-2010, Bonito-MS
This paper describes the various steps and highlights functions that can be accomplished by using AI techniques. Next, the paper presents some examples of using AI in the diagnostics of an induction machine.
• training based on both time and frequency domain signals obtained via simulation and/or experimental results; • real time, online unsupervised diagnosis; • dynamic updating of the structure with no need to retrain the whole network; • filtering out transients, disturbances, and noise; • fault prediction in incipient stages due to operation anomalies; • operating conditions clustering based on fault types. Example In the following an example of pattern recognition is described: ANN is applied to vibration signals in order to detect mechanical faults. Unbalance, shaft misalignment, and mechanical looseness have been compared with the “healthy” operation for two different ANN techniques: MLP global and PLP (Baccarini 2005). A multi-objective method has been used to improve the generalization capacity of the global MLP network. For investigation, a random choice of training and validation groups was done but took into account the sets dimension: 67% for training patterns and 33% for validation. Deterministic frequencies (fr, 2fr, 3fr, 4fr), and measurement with sensors in six different positions for four cases (healthy, unbalance, misalignment, and mechanical looseness) have been considered, in a total of 978 training and 312 validation patterns.
2 AI-BASED TECHNIQUES There are many types of AI-based techniques. Some of these use expert systems, artificial neural networks (ANNs), fuzzy logic, genetic algorithms (GAs), support vector machines (SVM). Besides giving improved performance these techniques are easy to extend, modify, and combine (with particle swarm, and/or wavelet, for instance). They can also be made adaptive by the incorporation of new data or information (Ivonne et al 2005). Next, considerations on the techniques and examples as well, will be given. 2.1 Expert Systems (ES) The expert system is basically a computer program embodying knowledge about a narrow domain for the solution of problems related to that domain. An ES mainly consist of a knowledge base and an inference mechanism. The knowledge base contains domain knowledge, which may be expressed as any combination of "IF-THEN" rules, factual statements, objects, procedures and cases, while the inference mechanism manipulates the stored knowledge for producing solutions. The system can determine a fault situation doing the signals extraction and fault identification from the combined derived information from behavior of various harmonic components and the machine operating conditions (Rajogopalan et al., 1991). A demerit of ordinary rule-based ES is that they can not handle new situation not covered explicitly in their knowledge bases. These ES can not give any conclusions in these situations.
A. Global MLP Network The Global MLP network, Fig.1, has two binary outputs for four situations: 00 – healthy; 01 misalignment; 10 unbalance; 11 mechanical looseness. The layers activation functions are sigmoid and weights have also been updated via back propagation. It was observed that the best result was obtained for the sensor in the vertical position, in which the vibration levels are more significant in case of mechanical faults. The total success rate was 86.16% for training and 82.37% for validation. A well trained network must adequately respond not only to the pattern used for training but also to all other submitted to them. This is known as network generalization capacity. At the training stage the generating function of data is based on possible realizations of the training sets for same task. This variety of solutions is named variance, which must be minimized to guarantee a good network generalization. On the other hand the number of possibilities increases with the model dimension. A reduction of dimension, by reducing the number of parameters, solves this problem. However, it can originate bias, which reduces the generalization capacity of the network. It is said that the bias occurs when even for different realizations of the training process in the reduced dimension space the solution is practically the same. Bias of solutions must be minimized to preserve the generalization capacity. Therefore, a point of equilibrium must be achieved.
2.2 Artificial Neural Network (ANN) An ANN is a computational model of the brain. It assumes that computation is distributed over several simple units called neurons, which are interconnected and operate in parallel, thus known as parallel distributed processing systems. Implicit knowledge is built into a neural network by training it. ANN can be trained by typical input patterns and corresponding expected output patterns. The error between the actual and expected output is used to strengthen the weights of the connections between the neurons. Awadallah and Morcos (2003) remind that ANNs have been densely applied in the area of motor condition monitoring and fault diagnosis performing one or more of the following tasks: • pattern recognition, parameter estimation, and nonlinear mapping applied to condition monitoring;
2025
XVIII Congresso Brasileiro de Automática / 12 a 16-setembro-2010, Bonito-MS
Fig.1: Scheme for the global MLP network.
A multi-objective system approach (Teixeira et al, 2002) allows for equilibrium between the weight norm and the training error of MLP network, and guarantees the generalization capacity of the model. Once a large topology is defined, the algorithm generates a set of solutions with a variety of norms and minimized error for each one, and selects the best response in relation to the validation set. Calculation of network efficiency confirmed the vertical position of the sensor as the most significant, and reached a success rate of 98%.
Fig. 2 The PLN scheme
A fuzzy system is based on the three classical steps which are: the fuzzification step, the inference engine and the defuzzication step. The inputs of the fuzzy expert system are the controlled variables. All inputs can be bounded and normalized. Each input variable is described by membership functions (Small, Medium and High) which can be triangular or take other function shapes. Classically, the inference engine is based on the Max-Min method. The output membership functions are described by Dirac functions (False, True). The evaluation of the centre (center) of gravity is thus easier to compute than using triangular or other functions shapes. The inference engine is based on AND functions and OR function which is described as follows: - For the AND function applied to the two inputs A and B, the output is evaluated as min(μA, μB); - For the OR function applied to the two inputs A and B, the output is evaluated as max(μA, μB). Awadallah and Morcos (2003) presented an extensive list of references, indicating some of the fuzzy and adaptive-fuzzy systems applications to motor fault diagnosis: •evaluating performance using linguistic variables; •predicting abnormal operation and locating faulty element; •utilizing human expertise reflected to fuzzy if—then rules; •system modeling, nonlinear mapping, and optimizing diagnostic •fault classification and prognosis. Example The aim of the example is the diagnosis of signatures of rotor broken bars when the induction machine is fed by an unbalanced line voltage. These signatures are given by the complex spectrum modulus of the line current. The fuzzy logic approach allows us to conclude to the load level operating system as to inform the operator of the rotor fault severity. The sense of velocity can be determined easily using the fuzzy logic. The rules are as follows:
B. Parallel Layers Network (PLN). A Parallel Layers Perceptron was proposed in Caminhas et al (2003) and replaces the original input of the ANFIS net by parallel perceptrons, Fig. 2. Its main objective is to overcome the limitation of working with multiple inputs imposed by the classical ANFIS network due to the resulting exponential increase of operations when each input is combined. Again, the vertical position of the sensor is confirmed as the most significant and the success rate percentage was 94%. 2.3. Fuzzy Logic System (FLS) The FLS are based on a set of rules. One advantage of FLS is that the rules allow the input to be fuzzy, i.e. more like the natural way that human express knowledge. In contrast to ANNs, they give a very clear physical description of how the function approximation is performed (since the rules show clearly the function approximation mechanism). Reasoning procedures, the compositional rule of inference, enable conclusion to be drawn by extrapolation or interpolation from the qualitative information stored in the knowledge base. The fuzzy approach model is a complex problem employing an IF-THEN type of expert rule and linguistic variables to capture directly the qualitative aspects of the human reasoning process involved. However, the problem is shifted to the membership function and rule tuning.
IF (If6 is ) AND (If3 is <small>) THEN Velocity is OR
2026
XVIII Congresso Brasileiro de Automática / 12 a 16-setembro-2010, Bonito-MS
IF (If6 is <small>) AND (If3 is ) THEN Velocity is .
implementing and optimizing fuzzy reasoning via ANN structures (Yoshikazu and Yoshiteru 2003).
Moreover, it is possible to monitor the connection of the motor to the line. A problem occurs when all the currents in the forward and backward have abnormal amplitude. The set of rules can be as follows:
2.5. Support Vector Machines The data-based machine learning is an important aspect of modern intelligent technology, while Statistics Learning Theory (SLT) is a new tool that studies the machine learning methods in the case of a small number of samples. Support Vector Machine (SVM) is derived from the SLT, and has attractive features, such as good generalization ability large dimension robustness, objective function convexity, and well established theory. In fact, SVM based classifier is claimed to have an efficiency does not depend on the number of features of classified entities, to have better generalization properties, cost much less time than NN based classifiers and better accuracy (greater than 97%) than Linear Discriminant analysis, K-Nearest Neighbor, Probabilistic Neural Network, Gaussian Mixture Model pattern recognition techniques (Namburu et al, 2007). Increase in SVM classification accuracy and speed continue a challenge since choices in SVM implementations, such as kernel function and penalty parameters of the support vector, may drastically affect its accuracy. An improvement can be obtained by tuning these parameters with other optimization techniques, like GA (Namburu et al, 2007), artificial immune system (Aydin et al, 2007), or on particle swarm optimization (Yuan and Chu, 2007). The SVMs are essentially binary classifiers (positive and negative classes). However, SVM-based multi-class classifier can be constructed using “one against one” technique, which consists in creating k SVMs, k corresponding to the number of classes. In the generation of each machine, a class is fixed as positive while the other are considered as negative. However, the use of “one against one” technique needs a synthesizing scheme to decide the final results according to the results of sub classifiers. In Fang and Ma (2006) four synthesizing schemes were compared (majority voting; binary tree decision; neural network and hybrid matrix), while in Mayoraz and Alpaydin (1999) ten of them were needed. Depending on what is required in motor diagnosis, there are two possibilities. The first, called simple diagnosis (1-Class), discovers only if the fault has occurred. The second one (complex diagnosis, 2Class) is able to find, for instance, how many bars have been damaged (Kurek and Osowski, 2008). The SVM is gaining application in rotating machinery anomaly detection, due to its superb performance with small samples [(Ortiz and Syrmos 2006), (Rojas and Nandi, 2006)]. Accurate online support vector regression (AOSVR) technique was, for the first time, implemented for machine condition monitoring with applications to motor shaft misalignment (Olufemi, 2007). AOSVR is an algorithm that combines the advantages of SVR with the capability of efficiently updating
IF (If6 is ) AND (If3 is ) THEN Current line is OR IF (If6 is <small>) AND (If3 is <small>) THEN Current line is .
By the same way, we can determine the load level thanks to these three rules: IF (s is <small>) THEN Load level is <Small> OR IF (s is <medium>) THEN Load level is <Medium> OR IF (s is ) THEN Load level is .
The detection of the rotor fault is based on the ratio of the sum of the two sideband currents on fundamental line current. This approach is advocated from one decade and gives a sufficient precision about the fault severity. As we study the forward and the backward sequence of the supply line, we have to calculate this ratio for the two sequences: Irf and Irb. These results will be the additional inputs for the expert system. The set of rules for the expertise of three phase induction motor, in case of broken bar defect, are given by: IF (Irf is <small>) AND (If6 is ) AND (If3 is <small>) THEN Operating Condition is OR IF (Irb is <small>) AND (If6 is <small>) AND (If3 is ) THEN Operating Condition is OR IF (Irf is <medium>) AND (If6 is ) AND (If3 is <small>) THEN Operating Condition is OR IF (Irb is <medium>) AND (If6 is <small>) AND (If3 is) THEN Operating Condition is < Fault in progress > OR IF (Irf is ) AND (If6 is ) AND (If3 is <small>) THEN Operating Condition is OR IF (Irb is < high >) AND (If6 is <small>) AND (If3 is ) THEN Operating Condition is < Broken Bar >
Thanks to these sets of rules, we are able to evaluate several diagnosis indexes. However, this technique requires the knowledge of the behavior for the determination of membership functions. 2.4. Neural-Fuzzy Consider the case of a motor stator faults. NeuralFuzzy fault detection is obtained which learns the stator faults and the condition under which they occur through an inexperienced and noninvasive procedure. The Neural-Fuzzy system is an ANN structured upon fuzzy logic principles, which enables this system to provide qualitative description about the machine condition and the fault detection process. The knowledge is provided by the fuzzy parameters of membership functions and fuzzy rules. The idea behind the fusion of these two technologies is to use the learning ability of ANN to implement and automate the fuzzy system, which use the high-level human-like reasoning capability. Many methods have been proposed for
2027
XVIII Congresso Brasileiro de Automática / 12 a 16-setembro-2010, Bonito-MS
trained support vectors whenever a sample is added to the training set. Example The fault diagnosis of induction machine is a multi-class classification problem. For classification of mechanical faults, as in the ANN example, have been compared to the “healthy” operation (Baccarini, 2005), a software developed by (Poyhonen et al 2002) was used because this work deals with four classes, the technique one-against-all was used in order to group the classes. SVM, together with one-against-all technique, has shown an excellent performance reaching a success rate of 98% for the already mentioned most significant position (vertical). For all cases, the Radial Bases Function (RBF) with variance between 0.01 and 0.1 was used as a kernel.
without any change in the children generation (Selection). Two children are next obtained thanks to an artificial reproduction using the genes of two parents (Crossover). The parents with better fitness are selected to take part in reproduction. In Mutation, the genes of one parent are randomly altered to give a child. Finally, all individuals of low Fitness are replaced by new ones, randomly computed. Example In [Razik et al (2008)(2009)] the GA is used to find the global maximum as well as to solve an optimization problem in the spectral lines identification process in case of a rotor broken bar of an induction motor. The N individuals are the supply frequency and the slip frequency and these are the parameters to be found. In order to find the eight main components in the current spectrum in Fig. 3, eight Gaussian functions were used as a window, which only depends on the supply frequency (f s) and the slip frequency (sfs) The fitness is calculated as being the integral of the product of the current spectrum by the spectral window, that is,
2.6. Genetic Algorithms Genetic Algorithm (GA) is based on biology and, in particular, by those biological processes that allow populations of organisms to adapt to their surrounding environment: genetic inheritance and survival of the fittest, that is, natural selection as well as evolutionary process. GA is a stochastic optimization method and needs less prior information about the problems to be solved than the conventional optimization schemes, which often require the derivative of objective functions. It also has the unique features of parallel search and global optimization and it is adapted for the simultaneous evaluation of a large number of points in the search space. GA can be used to determine the coefficients of a regulator (Cvetkoski et al, 1998) or to identify induction machine parameters. It can also be used in the diagnosis of induction motor rotor and stator faults, such as rotor broken bars [(Razik et al (2008) (2009)], open rotor and stator phase (Cristaldi et al 2004), rotor unbalance and misalignment, and bearing loose fault (Wei et al, 2007b). Although GA based approaches have interesting features when used alone, as compared to NeuroFuzzy (NF) based approaches, for instance (Cristaldi et al 2004), GA combined with other AI based approaches will have tremendous scope in future. This is due to the fact that the combination of GA with other motor fault diagnosis schemes has demonstrated enhanced performance in global and near-global minimum search. The GA uses three fundamental operations which are: Selection, Crossover and Mutation. First, an initial population is defined by N individuals created through a random generator. Each one is constituted by a parameters vector, of which the elements are called genes. These are the parameters for the optimization process, to which constraints can be added. The quality of an individual, to be adequate to the problem, is characterized by its fitness F, evaluated, for each individual, via an objective function. Then the parent individuals with the highest fitness are copied
Fitness = ∫
+100
−100
Spectrum( f ).ω ( f )df
where the spectral window is 8 1 ( f − fi )2 ω ( f ) = ∏ exp(− 2 σ2 i −1
(1)
(2)
The approach was tested for a rotor broken bar fault detection composed of four stages. First the acquisition of line currents was done, followed by the calculation of the Concordia’s vector. GA was used in the search lines of the supply frequency and the slip frequency which are inside of the Concordia’s vector spectrum. Fig. 4 depicts the results for line currents spectrum with one full broken bar and 100% of full load level. Results are obtained after few interactions. A comparison with the healthy line currents spectrum allows the operator to distinguish the faulty case and also to be aware of a growing rotor fault. The fault severity detection is based on a fuzzy approach. 2.7. Artificial Immune System The immune system (IS) is an efficient self-defense method that guards the human body from foreign antigens or pathogens. Artificial immune system (AIS) is an emerging soft computing method inspired by natural immune system. Because the AIS abilities of learning, memory and self adaptive control, this method is used in pattern recognition, classification, optimization and anomaly detection problems. Clonal selection is an artificial immune algorithm that is used for optimization problems. It has not crossover operator, so this method is different from genetic algorithm. Affinity proportional reproduction and affinity maturation are two distinctive properties of clonal selection. Because of these properties, clonal selection converges faster than genetic algorithm and doesn’t catch local minimum.
2028
XVIII Congresso Brasileiro de Automática / 12 a 16-setembro-2010, Bonito-MS
one shares its own experience to others particles, they move globally into the search space along a search trajectory. The speed for particles is given by: vik +1 = ωvik + c1 rand (⋅)( pi − xik ) + c 2 rand (⋅)( p g − xik ) where w is the inertia weight, pg is the position of the best particle among all particles, so it is the global solution. pi is the best previous of particle and vi denotes the velocity of the ith particle. rand are random numbers with uniform distribution in the range [0; 1], c1 is the cognition parameter and c2 is the social parameter. This equation updates the particles speeds based on particle momentum, the attraction force towards the global and the best local. In order to improve the performance of the PSO algorithm different modifications have been introduced. A constriction factor was proposed by (Clerc 1999) thus modifying the speed equation. Another improvement was proposed in (Shi and Eberhart, 1999) with the introduction of a time varying inertia weight (TVIW) as a function of the maximum and minimum values of inertia weight, and of the maximum number of search iteration. The algorithm prioritizes the global search. At the end of the search process, inertia decreases linearly to its minimal value prioritizing the local search. Also, in (Liu et al, 2005) was proposed an adaptive algorithm so that the inertia weight is different for each particle. Consequently, particles near the optimal solution refine the results and particles far from the optimal solution continue to explore the search space. PSO method has been applied to the following motor faults: broken bar, part of the end-ring broken (Razik et al, 2009), stator inter-turn and independent winding short circuit (Ethny et al, 2006). PSO method is an efficient method to solve the optimization problem as to extract information quickly from a frequency. Next it will be shown its ability to estimate the line frequency and the fault line frequencies with the induction motor operating under one full broken bar (Razik, 2009b). Example Based on a spectrum of current and calculation of the window function and of Fitness with the same equations used in GAS method, the transient of the PSO was examined thanks to several variables. In the estimation of the fundamental frequency the majority of the particles stayed close to the fundamental frequency of 50 Hz after about twenty iterations. To estimate the slip s, the population moved along the two sidebands to be found. The slip being close to 8.7%. Both the particle best fitness and the global best performance have been reached after 20 iterations. Bad performance particles disappeared with the increasing value of the number of iterations.
Fig. 3. The stator line currents spectrum in case of one broken bar.
Fig. 4. Stator line currents spectrum in case of one broken bar and the induction motor operating at full load, obtained by using GA.
In combination with other diagnosis methods, AIS can effectively improve the accurate rate of fault diagnosis and diagnosis system robustness, bringing into play each of their advantage, so that the accurate rate is improved. Clonal selection has been used to select optimal parameters of SVM, extracted from three phase motor current and constructed based on Park’s vector approach. This has been applied to the study of broken rotor bar and stator short circuit faults (Aydin et al., 2007). Also, for rotating machinery, AIS method has been combined with neural network method for machine fault diagnosis using genetic algorithm to combine diagnosis methods to make each kind of diagnosis method display its advantage in optimal space (de Castro and Timmis, 2002), (Wei et al 2007a). 2.8. PSO Algorithm Particle swarm optimization (PSO) is a semi-global optimization algorithm, first introduced by Kennedy and Eberhart (1995). It simulates social model like those of birds, insects and fish swarm. Its main concept is simulating the movement of these organisms searching for food. PSO is a simple optimization technique without heavy computation and has shown success in solving many optimization problems (Ciuprina et al., 2002), (Razik et al (2009). The technique does not require the computation of derivatives and hessians nor does not need training with heuristic data. Candidates to find the best solution are particles. They share their experience and collaborate to each other suggesting its own solution for the problem. As each
2.9 Bootstrap Gaussian Process Bootstrap Gaussian Process (BGP) has been proposed from the merge of Gaussian process classifiers (GPCs) and bootstrap methods, as an alternative to other classifiers, like the kernel classifier support 2029
XVIII Congresso Brasileiro de Automática / 12 a 16-setembro-2010, Bonito-MS
other classifiers, like the kernel classifier support vector machine (SVM), which has excellent performance towards this purpose, but it has difficulties to optimize relevant hyper-parameters. GPCs are Bayesian probabilistic kernel classifiers and provide a well established Bayesian framework to determine the optimal or near optimal kernel hyper-parameters. They are largely unexplored for anomaly detection applications and, also, a promising statistical tool for both binary and multi-category classification. Moreover, GPCs proved to outperform SVM (Kim et al., 2006). It can be employed to solve a wide range of problems, such as hypothesis tests, model selection and probability distribution estimations. Bootstrap is most useful where little is known about the statistics of the data or too few samples are available to use asymptotic results (Zoubir and Iskander, 2007). In BGP, bootstrap methods are incorporated to improve GPCs’ performance for small machinery anomaly samples by re-sampling at random. The fact that GPCs are strong classifiers suggests that small numbers of bootstrap samples might be sufficient to enhance classification performance. Experiment results for rotating machinery misalignment anomalies detection (Xue et al., 2008) in which wavelet packet is utilized to perform vibration analysis, show that bootstrap GPCs are highly effective and outperform GPCs and SVM with cross validation for anomaly detection. Thus the proposed approach is promising for rotating machinery anomaly detection.
References Awadalla M.A, and Morcos, M. M. (2003). Application of AI Tools in Fault Diagnosis of Electrical Machines and Drives-An Overview. IEEE Trans. on Energy Conversion, Vol. 18 (2), pp. 245-252. Aydin, I; Kakaköse, M; Akin, E. (2007). Artificial Immune Based Support Vector Machine Algorithm for Fault Diagnosis of Induction Motors. ACEMP, pp. 217-221. Baccarini L.M.R (2005). Detecção e Diagnóstico de Falhas em Motores de Indução. Universidade Federal de Minas Gerais. Caminhas, W.M; Vieira, D.A.G. and Vasconcelos, J. A. (2003). Parallel layer perceptron. Neurocomputing, pp. 771 – 778. Clerc, M. (1999). The Swarm and the Queen: Towards a Deterministic and Adaptive Particle Swarm Optimization. IEEE Congress on Evolutionary Computation, pp. 1951–1957. Cristaldi L.; Lazzaroni, M.; Monti, A.; Ponci,F. and Zocchi, F. (2004). A Genetic Algorithm for fault identification in electrical drives: a comparison with Neuro-Fuzzy computation. Instrumentation and Measurement Technology Conference IMTC, Vol. 2, pp. 1454-1459. Ciuprina, G.; Loan, D. and Munteanu, I (2002). Use of intelligent-particle swarm optimization in electromagnetic. IEEE Trans. Magnetics, Vol. 38, pp. 1037-1040. Cvetkoski, G.; Petkovsk, L. and Cundev, M. (1998). Mathematical model of a permanent magnet axial field synchronous motor for a genetic algorithm optimization. ICEM, Istanbul, pp. 1172–1177. De Castro, L.N. and Timmis, J. (2002). Artificial Immune Systems: A New Computational Intelligence Approach. Springer Verlag, New York. Ethny, S.; Acarnley, P.P.; Zahawi, B. and Giaouris, D. (2006). Induction Machine Fault Identification using Particle Swarm Algorithms. PEDES, pp. 14. Fang, R. and Ma, H.Z. (2006). Application of MCSA and SVM to Induction Machine Rotor Fault Diagnosis. Proceedings of the 6th World Congress on Intelligent Control and Automation, Vol. 2, pp. 5543-5547. Fuchs, F.W. (2003). Some diagnosis methods for voltage source inverters in variable speed drives with induction machines - A Survey. IEEE Ind. Electron. Conf., pp. 1378-1385. Filippetti, F; Franceschini, G.; Tassoni, C. and Vas, P. (1998) AI Techniques in Induction Machines Diagnosis Including the Speed Ripple Effect. IEEE Transactions on Industry Applications, Vol. 34 (1), 98-108. Ivonne, Y. B. ;Sun D and He, Y.K. (2005). Fault Diagnosis Using Neural-Fuzzy Technique Based on the Simulation Results of Stator Faults for a Three-Phase Induction Motor Drive System. ICEMS, pp. 1966-1971.
3. CONCLUSION In this paper, an overview on Artificial Intelligence (AI) methods-based motor fault diagnosis systems has been given. Several techniques using neural networks, fuzzy logic, neural-fuzzy, genetic algorithms, artificial immune system, vector support machine, particle swarm optimization, and Gaussian bootstrap process were summarized. Their applications and possibilities of combination were discussed as well. AI techniques are a very strong tool for electrical motors diagnosis studies. Although some investigators indicate that they are not yet supposed to compete with conventional methods, tremendous efforts have been made to develop new methods, as it is the case of bootstrap gaussian process. One observation is that AI methods become a strong tool when used in combination with other ones. ACKNOWLEDGEMENT Authors would like to thank CNPq (Conselho Nacional de Pesquisa e Desenvolvimento), FAPESQ (Fundação de Amparo à Pesquisa – Paraíba), CAPES, and COFECUB, for the financial support.
2030
XVIII Congresso Brasileiro de Automática / 12 a 16-setembro-2010, Bonito-MS
Kennedy, J and Eberhart, R. (1995). Particle Swarm Optimization. IEEE International Conference on Neural Networks, pp. 1942-1948. Kim, H.C; Kim, D. and Ghahramani,.Z. And Bang, S.Y. (2006). Appearance-based gender classification with Gaussian processes. Pattern Recognition Letters, Vol. 27, pp. 618–626 Kurek, J. and Osowski, S. (2008). Support Vector Machine for Diagnosis of the Bars of Cage Inductance Motor. ICESS’08, pp. 1022-1025. Liu, B.; Wang, L.; Jin, Y-H ; Tang, F and Huang, D.X. (2005). Improved particle swarm optimization combined with chaos. Elsevier Chaos, Solitons and Fractals, Vol. 25(5), pp. 1261–1271. Mayoraz, E and Alpaydin, E. (1999). Support Vector Machines for Multi-Class Classification. Int. Workshop on Artificial Neural Networks, Vol. 2 (4), pp. 833-842. Namburu, S.M; Chigusa, S.; Prokhorov, D; Qiao, L; Choi, K and Pattipati, K. (2007). Application of an Effective Data-Driven Approach to Real-time Fault Diagnosis in Automotive Engines. IEEE AC, pp. 1-9. Olufemi, A.O., Myong K.J; Adedeji, B.B. and Hines, J.W. (2007). Online Support Vector Regression Approach for the Monitoring of Motor Shaft Misalignment and Feedwater Flow Rate IEEE Trans. On Syst, Man, and Cybernetics. Applications and Reviews, Vol. 37(5), pp. 962-970. Ortiz, E. and Syrmos, V. (2006). Support vector machines and wavelet packet analysis for fault detection and identification. Int. Joint Conf. on Neural Networks, Vancouver, pp. 3449–3456. Poyhonen, S.: Negrea, M.; Arkkio, A; Yotyniemi, H. and Koivo, H. (2002). Support Vector Classification for Fault Diagnosis of in Electrical Machine. ICSP'02, pp. 1719-1722. Qiang, S., Gao, X.Z. and Zhuang, X.(2003). State-ofthe-art in Soft Computing-based Motor Fault Diagnosis. CAC 2003 pp. 381-1386. Rajogopalan, V; Debebe K. and Sankar T.S. (1991). Expert system for fault diagnosis of VSI fed AC drives. IEEE-IAS Annual. Meeting, Dearborn, MI, pp. 368–373. Razik, H; Corrêa, M.B.R. and Silva, E.R.C. (2008). An application of Genetic Algorithm and Fuzzy Logic for the induction motor diagnosis. IEEE IECON, pp. 3067-3072. Razik, H; Corrêa, M.B.R. and Silva, E.R.C.(2009a). A Novel Monitoring of Load Level and Broken Bar Fault Severity Applied to Squirrel-Cage Induction Motors Using a Genetic Algorithm. IEEE Trans. Ind.ustrial Electronics, Vol. 56 (11), pp. 4615-4626. Razik, H; Corrêa, M.B.R. and Silva, E.R.C.(2009b). The use of particle swarm optimization for the tracking of Induction motor faulty lines. Proc. of IEEE PowerEng’09, pp. 680-684. Rojas, A. and Nandi, A.K. (2006). Practical scheme for fast detection and classification of rollingelement bearing faults using support vector ma-
chines. Mechanical Systems and Signal Processing, Vol. 20 pp. 1523–1536. Shi, Y. and Eberhart, R.C. (1999). Empirical Study of Particle Swarm Optimization. Congress on Evolutionary Computation, pp. 1945–1950. Siddique, A., Yadava, G. S. and Singh, B. (2003). Applications of Artificial Intelligence Techniques for Induction Machine Stator Fault Diagnostics: Review. SDEMPED´03, pp. 29-34. Teixeira, R.A; Braga, A.P; Takahashi, R.H.C. and Saldanha, R.R. (2002). Decisor implementation in neural model selection by multiobjective optimization. SBRN, pp. 234–239. Thorsen, O.V. and Dalva, M. (1995). A Survey of the Reliability with an Analysis of Faults on Variable Frequency Drives in the Industry. IEEE Trans. Industry Applications, Vol. 31(5), pp. 1.1861.196. Yoshikazu, F. and Ueki, Y. (2003). Fault Analysis System Using Neural Networks and Artificial Intelligence. IEEE ANNPS´03, pp. 20-25. Yuan, S.F. and Chu, F.L. (2007). Fault diagnostics based on particle swarm optimization and support vector machines. Mechanical Systems and Signal Processing, Vol. 21, pp. 1787–1798. Wei D.; Zhan-sheng L. and Dong-hua W. (2007a). Combination Diagnosis Based on Genetic Algorithm for Rotating Machinery. Third International Conference on Natural Computation, pp. 307309. Wei, D; Zhan-sheng, L. and Xiaowei, W. (2007b). Application of Image Recognition Based on Artificial Immune in Rotating Machinery Fault Diagnosis. International Conference on Bioinformation and Bio-medical Engineering, ICBBE’07, pp. 1047-1052. Xue, W., Daowei Bi, Liang D., and Sheng W. (2008). Bootstrap Gaussian Process Classifiers for Rotating Machinery Anomaly Detection. IEEE International Joint Conference on Neural Networks, pp. 1129-1134. Zoubir, A.M and Iskander, D.R. (2007). Bootstrap methods and applications. IEEE Signal Processing Magazine, Vol. 24 (4), pp. 10–19.
2031
