A novel multi-agent approach to identify faults in line connected three-phase induction motors

Applied Soft Computing 45 (2016) 1–10

Contents lists available at ScienceDirect

Applied Soft Computing journal homepage: www.elsevier.com/locate/asoc

A novel multi-agent approach to identify faults in line connected three-phase induction motors Rodrigo H. Cunha Palácios a,b,∗ , Ivan N. da Silva a , Alessandro Goedtel b , Wagner F. Godoy a,b a University of São Paulo, São Carlos School of Engineering, Department of Electrical Engineering, Av. Trabalhador São Carlense, 400, Centro, 13.566-590 São Carlos, SP, Brazil b Federal Technological University of Paraná, Department of Electrical Engineering, Av. Alberto Carazzai, 1640, Centro, 86.300-000 Cornélio Procópio, PR, Brazil

a r t i c l e

i n f o

Article history: Received 18 March 2015 Received in revised form 25 February 2016 Accepted 8 April 2016 Available online 22 April 2016 Keywords: Three-phase induction motor (TIM) Pattern recognition Motor faults Faults diagnosis Multi-agent system (MAS)

a b s t r a c t Three-phase induction motors (TIMs) are the key elements of electromechanical energy conversion in a variety of productive sectors. Identifying a defect in a running motor, before a failure occurs, can provide greater security in the decision-making processes for machine maintenance, reduced costs and increased machine operation availability. This paper proposes a new approach for identifying faults and improving performance in three-phase induction motors by means of a multi-agent system (MAS) with distinct behavior classifiers. The faults observed are related to faulty bearings, breakages in squirrel-cage rotor bars, and short-circuits between the coils of the stator winding. By analyzing the amplitudes of the current signals in the time domain, experimental results are obtained through the different methods of pattern classification under various sinusoidal power and mechanical load conditions for TIMs. The use of an MAS to classify induction motor faults allows the agents to work in conjunction in order to perform a specific set of tasks and achieve the goals. This technique proved its effectiveness in the evaluated situations with 1 and 2 hp motors, providing an alternative tool to traditional methods to identify bearing faults, broken rotor bars and stator short-circuit faults in TIMs. © 2016 Published by Elsevier B.V.

1. Introduction Three-phase induction motors (TIMs) consume over 60% of the electrical energy in the industrial sector [1] and are the primary means of energy transformation in mechanical drives [2–4]. Like any other electrical machine, these motors require proper maintenance, as breakdowns can impact on productivity and cause substantial losses in industrial processes. While profitability depends on various factors, equipment maintenance is one of the most important. According to [5], all these factors justify increased efforts to develop new techniques to detect possible motor faults with enough time for proper and planned maintenance procedures. Malfunctioning is divided into two major groups: electrical faults and mechanical faults. The most common failures are related to electrical problems in the stator winding, rotor winding, broken

∗ Corresponding author at: Federal Technological University of Paraná, Department of Electrical Engineering, Av. Alberto Carazzai, 1640, Centro, 86.300-000 Cornélio Procópio, PR, Brazil. E-mail address: [email protected] (R.H.C. Palácios). http://dx.doi.org/10.1016/j.asoc.2016.04.018 1568-4946/© 2016 Published by Elsevier B.V.

rotor bars, broken rotor rings, and connections, among others [6]. Mechanical failures are derived from problems such as wear and tear on the bearings and couplings, eccentricity, and misalignment. According to the research conducted by [7], over 80% of unwanted downtime in electrical motors is related to rotor, stator, and bearing problems. To contribute to the study and development of electric machine fault identification, this paper proposes the implementation of a multi-agent computer system, based on computational intelligence, which identifies faults in the stator, rotor, and bearings of TIMs, where the input parameters are the amplitudes of threephase currents of power supply to the machine, considering the occasional asymmetrical unbalancing of the voltages and covering a wide range of mechanical loads applied to the machine shaft. The early identification of faults is designed to bring greater security to the decision-making process, with the possibility of reducing maintenance costs and increasing the operating availability of the motor. The use of the proposed architecture for MAS provides the opportunity to explore its parallelism, extensibility, redundancy and robustness, also combining the usage of different pattern identification methods. In the case of the need to monitor a network

2

R.H.C. Palácios et al. / Applied Soft Computing 45 (2016) 1–10

with many motors which requires a higher computational power. It is possible to distribute the processing of the motor fault recognition system on different hardware and even the construction of a system with different classifier types that could better adapt to a specific fault type, although the system may require more information processing power compared to traditional systems found in the literature. This paper is divided into six sections. Section 1 presents the introduction and a review of related research in the field. Section 2 briefly discusses previous works. Section 3 presents the methodology of the study. Experimental results appear in Section 4, while Section 5 presents findings in comparison with previous works. Section 6 presents the conclusion and final discussions. 2. A brief discussion of previous works Induction motors play an important role in industry, a fact which underscores the importance of accurate diagnosis and classification of faults in these motors in the early stages of their evolution. Accordingly, in this work, motor faults are diagnosed by MAS with intelligent pattern classification methods, considering a pre-processing method based on the time domain, by the discretization of the stator current signals for rotor, bearing and stator fault classification. When assessing incipient problems in TIMs, bearing-related issues have been the subject of considerable research, due to the high percentage these components occupy in the classification of motor faults. Bearing faults, according to [7], can account for over 40% of the problems that typically occur in electric motors. Several methods are used to detect these types of faults, such as analyzing mechanical vibration, spectral frequency of the stator current and axial flux. Several techniques are used to identify these faults, as demonstrated in the research presented by [8], where a methodology for monitoring bearing defect conditions by using voltage and stator current signals was developed. This system diagnoses various types of bearing defects in a 0.5 hp motor, considering different loading conditions. The strategy presented by [9], considers a model based on the entropy permutation of the vibration signal that is calculated to detect problems in the bearing during motor operation. In case of bearing failures, the vibration signal is decomposed into a set of intrinsic mode functions by means of the decomposition set in an empirical way. The authors used a variation of the support vector machine (SVM) method to classify and determine the severity of the studied defect. Rotor faults in TIMs account for approximately 10% of the problems presented [7]. Such faults can generate thermal stress, electromagnetic forces, electromagnetic noise and vibrations and are caused by centrifugal forces, environmental stresses (abrasion) and mechanical stress [10]. Several related research projects are ongoing, such as observed in the work presented by [11], where the authors proposed a new approach to identify broken rotor bars in induction motors under different load conditions based on wavelet coefficients of the stator current in a specific frequency range. In case of faults, by increasing the number of broken bars and load levels, the amplitude of some particular components of the stator current sideband will also increase. The quantities – stator current, rotor speed and torque – are used to show the relationship between these parameters and the broken rotor bar severity. An induction motor with 1, 2 or 3 broken rotor bars is used under adversities of load variation on the motor shaft. Finally, this work also evaluates the severity of rotor faults based on data originated from stator current and rotor speed. In the work of [12] an experimental analysis of an induction motor driven at different speeds and load levels was performed. The main objective was to study experimentally the ability of the method by using the motor current signature analysis to diagnose the occurrence of broken rotor bars. In [13], a method

for identifying broken rotor bar defects was also presented, based on the transformed stator current Wavelet in a specific frequency range. The method allows diagnosis of the occurrence and determination of the number of broken bars at different loads. Also, due to the properties of Wavelet transform during transient conditions, it is possible to detect the defect during motor start-up. Experimental results show data concerning motors with 1, 2, 3 and 4 broken bars with the motor operating at rated load. The motor operating with four broken bars was also tested with no load, 33%, 66%, 100% and 133% of load torque. In addition, the authors considered the finite element method to model broken bar defects. Stator coils are subject to various anomalies which can cause different responses in the equipment. Changes in the stator winding can manifest themselves in several ways, including overheating of the motor, electrical overload, and faults in the coil insulation. The stator is subjected to critical situations in the operation of a TIM, such as thermal, electrical, mechanical and environmental, all of which can severely affect the stator condition and lead to breakdowns. As faults in the stator can cause problems with the operation of the motor, several studies to identify these faults are proposed in the literature, such as the work of [14] where the authors presented the analysis of the signature instantaneous active and reactive power in the frequency domain for diagnosing stator faults in motors driven direct on line and also considering an inverter-fed situation. The experimental and simulation results demonstrate the effectiveness of the proposed approach. The severity of the conditions is tested and the behaviors of frequency ranges are observed for the detection of stator faults. In the work presented by [15] the authors carried out a comparative study considering the evaluation of SVM/SMO, multilayer perceptron (MLP) and fuzzy ARTMAP network classifiers to diagnose the severity of stator short-circuit faults. Recently, the work of [16] presented a method considering FFT at the pre-processing stage and an MLP network to classify the severity of stator short-circuit faults in a permanent magnet synchronous motor. Breakdowns related to the stator are responsible for approximately 37% of undesired stoppages in TIMs [7]. The most common methods used in the detection and diagnosis of incipient faults in TIMs are those based on models, specialized systems and intelligent systems that simulate biological models. The diagnosis of such faults can be accomplished through spectral analysis, analysis in the time domain, finite elements, and by means of intelligent systems, as mentioned in the work of [10]. Intelligent systems in this context are presented in various topologies and procedures such as artificial neural network (ANN), fuzzy logic, and hybrid systems. On the other hand, there are few works on distributed artificial intelligence related to recognition of faults in TIMs. According to [17], distributed artificial intelligence is classified by distributed problem solving (DPS), multi-agent systems (MAS) and parallel artificial intelligence (PAI). For over a decade, the proposed use of MAS to address engineering challenges has been reported in several instances in the literature. The use of MAS appears in works on the diagnosis of energy systems related to protection [18], systems monitoring [19], restoration of the power system [20], control [21,22] and automation [23]. For example, the work of [24] employed an intelligent agent methodology for the reconfiguration of power systems. The research conducted by [25] discussed reconfiguration for the restoration of power systems using an MAS. The authors also mentioned that the majority of approaches for restoration found in the literature are centralized solutions which lead to the single point of failure. Additionally, the research conducted by [26] proposed an MAS for fault detection and system reconfiguration for power distribution. Another related study was performed by [27], who designed a MAS integrated with fuzzy systems in order to control reactive power in a distribution system. In the area of protecting power systems, the work developed by [28] proposed an MAS to

R.H.C. Palácios et al. / Applied Soft Computing 45 (2016) 1–10 Table 1 Characteristics of three-phase induction motors used in the experiments. Parameters

Motor 1

Motor 2

Motor 3

Manufacturer Power Poles Slots Yield

WEG 1 hp 4 36 Standard

WEG 1 hp 4 36 High efficiency

WEG 2 hp 4 36 High efficiency

Table 2 Voltage and load variation limits for data acquisition: 1 and 2 hp motors. %Va 100 90 100 90

%Vb

%Vc

1 hp (Nm)

2 hp (Nm)

100 100 110 100

100 100 100 110

0.5–6.0 0.5–6.0 0.5–6.0 0.5–6.0

0.5–9.0 0.5–9.0 0.5–9.0 0.5–9.0

help protect distributed electricity generation systems, the main features of which are detection and troubleshooting. The research conducted by [19] proposed an intelligent MAS to monitor the operating conditions of power transformers. The proposal for an intelligent multi-agent architecture presented in this paper is designed to explore the capability of autonomy and encapsulation inherent in the architecture, which leads agents to offer distinct behavior patterns, based on classifier standards, which can be applied collaboratively to present motor fault diagnosis. 3. Methodological aspects applied to determining induction motor faults with MAS As presented by [29,30], the traditional classification for multiple faults in TIMs is essentially composed of the existence of the combination module, which requires all the outputs of the classifiers in order to generate the final output. Accordingly, a multi-classifying system can become an inflexible system, as the decision making becomes a centralized process or even a model created from training a dataset with multiple faults, which can generate too many rules on a centralized basis. The purpose of this work is to remove the layer of decision and allow entities to make the classification interactively, concurrently and independent of the platform, whereby the classifiers can decide the best output from the system, as shown in Fig. 1. The structure features a communication system composed of autonomous entities, in which an input is presented to every entity. These entities produce their respective outputs and can communicate by agreeing over the output, which reinforces the behavior of an MAS. Each classifier is an intelligent agent which behaves in order to recognize patterns. MAS are a common means of exploring the potential ability of agents and combine many agents in an individual system. Each agent has incomplete information and is unable to solve an entire problem on its own, but together, the agents form a system with enough information and ability to solve the problem. Fig. 1 shows the flowchart of the proposed MAS. The experimental data of the three-phase stator currents are obtained directly from the motor by using the laboratory test bench shown in Fig. 2. The specifications of the motors used in this study are provided in Table 1. The tests, where the motors produce signals due to faults in the bearings, rotor, and stator, are also characterized by applying a range of load torque and voltage unbalance conditions, as shown in Table 2. During the acquisition, the load torque was varied from 0.5 Nm to 6 Nm with intervals of 0.5 Nm. For each load torque condition, the voltage is unbalanced between phases up to ±10% with intervals of 2%, the voltage unbalance occurring

3

exclusively in the phase “a” with phases “b” and “c” balanced, an overvoltage in phase “a”, undervoltage in phase “b” and balanced phase “c” was also considered, as well as experiments with all phases properly balanced. The test data were acquired using a National Instruments USB-6221 signal conditioning board as well as Matlab software, operating at 25,000 samples per second. The stator current signal is suitable for the acquisition system, which uses a Lem Hall sensor to condition the signals in the range of ±10 V. More information about the test bench can be found in recent papers such as [2,3]. In the data processing module the data set with signals from the currents of each phase (Ia,b,c ) are selected from motors operating in steady-state. For each test, half a wave period from each phase is selected from the currents in the time domain. The data are then organized and discretized in ten points per half cycle of the current waveform for each phase before being normalized by the peak value of the signal [2,31], thus creating vectors with 30 points (absolute values) for use as input in the MAS. Each point represents a value of the normalized amplitude of an instant of time for each phase current and the behavior of these 30 values refers to the standard sample. After the treatment phase of the signals, when the intelligent multi-agent classifier module is activated, the standard classifier model of each specialist agent has already been determined. When the system receives the sample to be identified, the main class processes the data and activates the operation of the multi-agent platform. All the agents register availability in the system through the directory facilitator (DF) agent, available in Jade platform. The coordinating agent then processes the initial input information and verifies if the TIM is operating under normal conditions, sending a message to the non-defective motor classifier agent. The functionality that checks whether the motor signal sample is under standard operation conditions with no fault comprises the usage of an ANN method of MLP type trained with data collected from a 1 hp motor covering stator, rotor and bearing faults with a negative classification standard (0) and data gathered from a healthy motor with positive classification standard (1). The learning rate value is defined as 0.3, the momentum term receives 0.2, and the maximum number of epochs for training the network is 500. The ANN architecture is composed of fifteen neurons which are used in the single hidden layer, empirically defined. For the hidden layer, the activation function is hyperbolic tangent, and as for the output layer linear function is used. After the system is satisfied with the accuracy of the data received, it terminates processing of the sample and returns the result. Otherwise, the coordinator agent calls for sample evaluation from the agents responsible for the identification of bearing, stator or rotor faults. Based on statistical methods in accuracy and errors, provided in the validation of previously trained classifier models, agents communicate with each other to determine the acceptable threshold to diagnose the fault. The initial activation of the proposed MAS causes specific agents to register with the DF agent and at the same time train their respective predetermined models. While the agents are active together with the platform, it is possible to validate any samples that are in the organizational pattern of the system. This pattern is related to a vector line with 31 positions, where positions 1 through 10 represent the sinusoidal current of phase “a”, positions 11–20 represent the current signal of phase “b”, while positions 21–30 show the current of phase “c” of a specific sample that represents the motor test. In position 31, the sample pattern is determined. According to [32], for an initial evaluation based on the results shown in Fig. 3, classification tests of samples acquired through the testing bench were performed using 5 classification methods to obtain information to define the best behavior of each agent responsible for fault identification in the MAS proposed in this paper. All

4

R.H.C. Palácios et al. / Applied Soft Computing 45 (2016) 1–10

Fig. 1. Flowchart of the overall functionality of the proposed system.

methods were tested for each type of observed failure through the cross-validation method, with 10-folds for training and validation. Based on the results presented in Table 3 the agent responsible for identifying bearing faults is based on the SVM/sequential minimal optimization (SMO) [33,34] and k-NN [35] methods. The classifier behavior for the agent responsible for identifying rotor faults is based on the ANN/MLP [36] and k-NN methods. The

behavior classifier for the agent that identifies stator faults is based on the k-NN and ANN/MLP methods. For all situations, the configuration of the k-NN classification method for executing the tests uses only three neighbors, while the search algorithm is based on the calculation of the Euclidean distance method. The basic settings of the SVM/SMO method defined in this paper are: a rounding error rate value of 1012, the core

R.H.C. Palácios et al. / Applied Soft Computing 45 (2016) 1–10

5

model. These samples, composed from signals of defective stator, rotor, and bearings, as well as from the motors without faults, are organized where each specialized agent recognizes its respective standard classification. For stator faults, short-circuits were created in the stator of Motor 1, considering severity levels of 1%, 3%, 7%, and 10%, and 1%, 3%, 5%, 10%, 15%, and 20% for Motor 2. As for Motor 3, the following stator short-circuit severity levels were used; 1%, 3%, 5%, and 10%. Fig. 4 shows an illustration of a 1 hp motor, 4 poles, manufactured by WEG, which was rewound with winding derivation taps. When these keys are driven, a short-circuit occurs between the turns of the same coil, according to the selected percentage. As for the faulty rotor bars, it is possible to emulate the occurrence of one, two, and four consecutive bars, as well as two consecutive broken bars in opposite sides of the rotor. Fig. 5 presents an image with the insertion of holes on the rotor. These rotor faults are performed by drilling, using a drill bit with a diameter sufficiently larger than the bar. This procedure results in a set of rotors with artificially created defects. For testing motors with bearing faults, a summary of reproduction of these faults, in which the data were acquired for testing in this work, can be seen in Fig. 6, based on the distributed and localized conditions, which emulate the operating conditions in the industrial environment. Faults were introduced in bearings related to excessive wear (distributed), grooves, inner and outer tracks, electric shock and defect in the spheres (located). Excessive wear defects were reproduced in order to evaluate the similarity of the bearing degradation process through excessive use, lack of lubrication and excessive load on the shaft. After cleaning the bearing, the lubrication was replaced by an abrasive slurry used

Fig. 2. Experimental test bench.

function of the SVM/SMO is the Polynomial Kernel [37], and an error tolerance parameter of 0.001. The ANN/MLP configurations are: learning rate value defined as 0.3, the momentum term receives 0.2 and the maximum number of epochs for training the network is 500. Fifteen neurons are used in the single hidden layer, defined empirically. For the hidden layer, the hyperbolic tangent activation function is used, and one linear function is used in the output layer. Each classification method is added to the behavior of the previously trained agents, and the respective models are adjusted to the 960 sample signals gathered from the 1 hp motors (Motors 1 and 2). A third motor (Motor 3) of 2 hp was used exclusively in the validation stage aiming to verify the robustness of the proposed

Fig. 3. Results of tests performed with standard methods for classifying faults in the stator, rotor and bearings in three-phase, 1 hp induction motors.

Table 3 Results of the validation of sample lots with stator faults in a 1 hp motor, and stator agent performance based on the k-NN and ANN/MLP method. Lots with stator faults

1

2

3

4

5

6

7

8

9

10

Samples % Accuracy of stator agent (k-NN) % Accuracy of stator agent (ANN/MLP) % Accuracy of bearing agent (k-NN) % Accuracy of bearing agent (SVM/SMO) % Accuracy of rotor agent (ANN/MLP) % Accuracy of rotor agent (k-NN) % Accuracy of normal conditions agent (k-NN) % Accuracy of normal conditions agent (ANN/MLP)

10 100 100 80 100 100 100 100 100

10 100 100 90 100 80 100 80 80

10 100 90 80 100 100 80 90 80

10 100 80 90 100 100 100 100 80

10 100 90 90 100 90 80 90 80

10 100 100 90 100 100 90 90 60

10 100 100 70 100 80 90 80 70

10 100 100 90 100 100 80 90 70

10 100 100 80 100 100 90 100 90

10 100 100 90 100 100 100 90 80

6

R.H.C. Palácios et al. / Applied Soft Computing 45 (2016) 1–10

Fig. 4. Creation of stator short-circuit faults: (1) Winding process. (2) Creation of derivations to emulate short-circuit faults. (3) Conclusion of rewind process. (4) Motor prepared for the emulation of 5% of short-circuit in the coil of phase “a”.

Fig. 5. Creation of broken rotor bars faults: (1) Drilling to emulate one broken rotor bar. (2) Demonstration of three rotors with one, two and four consecutive broken bars.

for grinding parts and the motor run again for 30 min in order to degrade the bearing. After this process, the bearings were cleaned and lubricated again for proper data acquisition. The insertion grooves in the inner and outer races were held with the help of a mini grinding machine. Faulty tests considering the electric discharge in the bearings were carried out by using a controlled discharge generated by an arc welding machine, performed on a clean, dry bearing, with its inner ring grounded and the electrode promptly applied on the outer race. One example of how the system operates is the condition of the agent responsible for identifying stator faults. In this case, the training data is prepared using samples from all the faults in this study, as well as samples from motors operating under normal conditions. For this agent, all samples are determined with a classification value of 0, except for the samples with stator faults, which are given a value of 1. Using this logic, the training data

prepared for the classifiers inherent to the behavior of the agents belonging to the system are also organized. The Java Agent Development Framework (JADE) environment [38] is adopted in this project in order to develop the proposed MAS designed to address the implementation of applications based on agents, using FIPA specifications for the interoperability between MAS. 4. Experimental results Section 3 presented the methodological aspects of preprocessing in the proposed MAS experiments to identify stator, rotor and bearing faults in TIMs. A block diagram of the various steps involved from the pre-processing to the classification of faults through the MAS is also provided. For the presentation of the experimental results, the data is obtained from a laboratory test bench

R.H.C. Palácios et al. / Applied Soft Computing 45 (2016) 1–10

7

Fig. 6. Creation of bearing faults. (1) Outer race grooves. (2) Inner race grooves. (3) Bearing prepared to faults emulation. (4) Electric shock and defect in the spheres (outer race). (5) Simulation of wear and breakage on the balls. (6) Use of abrasive slurry to simulate excessive use and lack of lubrication.

and is divided into training data and validation for the consolidation of the results. For each TIM defect proposed in this study, an intelligent agent comprising a specific classifier is determined to solve the task. Each agent comprises a model which has been previously trained with experimental data. All data used in training the agents is related to 1 hp TIMs, as detailed in Table 1. To validate the effectiveness of the MAS, as shown in Table 3 ten lots are selected with ten samples from each class which are: stator, rotor, bearing and non-defective motors. Each lot has a set of ten samples of current signals. Based on the results presented in Fig. 3, the classification strategy which should be employed for each specific type of fault implicit for the agents is defined. Not all samples presented in the validation are presented in the training of the intelligent system classifiers used in the agents, but are taken at random from the general set of data for each type of defect. Table 3 shows the results of the validation samples with stator, rotor and bearing faults in the 1 hp TIM, and also presents the classification performance based on the k-NN and ANN/MLP methods for the stator agent, k-NN and SVM/SMO methods for the bearing agent, ANN/MLP and k-NN methods for the rotor agent, and k-NN and ANN/MLP methods for the normal conditions agent. For the purpose of verifying samples from motors with stator faults, the probability of the proposed MAS is 100% accurate for stator faults in every lot of samples, considering the k-NN classifier. As for the ANN/MLP method, the test results for the same batches of samples with the behavior specified in the stator agent are over 80% accurate. The k-NN method used in the behavior of the stator agent is shown to be more effective on average for the tested samples. The validation of bearing defect samples appears in the same table, where the bearing agent was tested with behavior classifiers based on the k-NN and SVM/SMO methods. As for the k-NN classifier the results demonstrate over 70% accuracy for the validation of ten lots with bearing faults (balls, outer race, inner race and wear) in a 1 hp motor, while the other agents obtain a 0% possibility for all lots tested. Based on the best classifiers of the previous tests, presented in Fig. 3, the behavior classifier based on the SVM/SMO method is assigned to the bearing agent in order to verify the dynamic of the results. Still considering Table 3 for these same ten lots tested using a k-NN classifier, the bearing agent improves its accuracy.

The results for the sample tests with rotor faults in TIMs are also given in Table 3. The rotor agent lots tested with the ANN/MLP classifier method are over 80% accurate. The accuracy of the other agents is 0%, which better represents the certainty of the decision of these agents. As for the k-NN method the rotor agent is also over 80% accurate, which is comparable to the ANN/MLP method. In addition to the tests based on ten lots of ten samples, data is prepared for motors under normal operating conditions. The final two lines shown in Table 3 focus on the performance of the k-NN and ANN/MLP classifier methods tested for the agent without faults, and the k-NN method shows better than 80% accuracy for all mentioned validation samples. In the case of the ANN/MLP method, for these same lots 60–80% accuracy was achieved, while the accuracy of the other agents is 0%. In order to increase the number of tests and verify the validity of the proposed MAS, larger additional lots are prepared for signals from the 1 hp motors under normal operating conditions, with 240 samples for stator, bearings and rotor faults, plus non-defective motors, respectively, as detailed in Table 4. The samples tested are not presented in the training of the methods used for the agents and the results are compared to those obtained in tests with smaller lots of samples. After checking samples concerning motor with stator faults, it is noticed that the agent responsible for determining stator faults is 98.3% accurate, while the other agents do not considered any samples from this set as belonging to their domains. As for the samples related with rotor faults, the agent responsible for determining such failures correctly determining the classification of 239 of 240 samples, whereas 2.91% of these samples were correctly Table 4 Results of validation of lots of 240 samples from the 1 hp motor with stator, bearing and rotor faults and without faults. Lots with 240 faults samples

Stator

Rotor

Bearing

NC

Samples % Accuracy of stator agent (k-NN) % Accuracy of rotor agent (ANN/MLP) % Accuracy of bearing agent (SVM/SMO) % Accuracy of N.C. agent (k-NN)

240 98.3 0 0 0

240 0 99.58 0 2.91

240 0 0 100 0

240 12.5 0 0 88.3

N.C. = normal conditions.

8

R.H.C. Palácios et al. / Applied Soft Computing 45 (2016) 1–10

Table 5 Results of validation of lots with 2 hp motor samples with stator defect short-circuits of 1%, 3%, 5%, and 10%. Lots of 2 hp motor with stator faults

1%

3%

5%

10%

Samples % Accuracy of stator agent (k-NN) % Accuracy of rotor agent (ANN/MLP) % Accuracy of bearing agent (SVM/SMO) % Accuracy of normal conditions agent (k-NN)

120 45 10 27 0

120 66 9 15 0

120 89 3 3 0

120 100 0 0 0

Table 6 Results of validation of sample lots with bearing faults in the 2 hp motor. Lots of 2 hp motor with bearing faults

Balls

Outer race

Samples % Accuracy of stator agent (k-NN) % Accuracy of rotor agent (ANN/MLP) % Accuracy of bearing agent (SVM/SMO) % Accuracy of normal conditions agent (k-NN)

120 5 36.7 75.8 1.6

120 7.5 30 77 0

classified by the normal conditions agent. Samples concerning bearing faults were classified with 100% of accuracy by the agent responsible for determining this type of defect, while the other agents do not recognized this fault. Finally, the samples originated from healthy motor were checked by these four agents. Normal conditions agent achieved an accuracy of 83.3% for the classification of these samples whereas the stator agent presented 12.5% of error. Finally, based on the necessity to generalize the system, sets of samples comprising data concerning stator and bearings faults for the 2 hp motors are prepared and detailed in Tables 5 and 6. The data presented in the tests are not used in the training of the classifier methods used in the system agents and all adjustment samples of classifiers models are derived from 1 hp motor. Table 5 presents the results concerning the evaluations of 480 samples of stator faults, separated according to severity percentages. Higher severity levels showed better faults detection accuracy. As for the samples considering short-circuit level of 1% the agent responsible for determining stator faults obtained an accuracy of 45%, impacting in a false positive on the final classification. The agents responsible to identify rotor and bearing faults obtained respectively 10% and 27% of accuracy. As for the validation tests considering exclusively samples concerning 10% of short-circuit fault, the agent responsible for determining stator failures reach 100% of accuracy, whereas the other agents did not identify any sample submitted to the MAS. Although there is determination of percentage classification of incorrect agents, the process of mutual decision makes the final diagnosis as expected. Complementing the MAS robustness tests proposed in this paper, bearing faults in the 2 hp motor was evaluated. Table 6 shows results of validation tests performed using samples of bearings with defects in balls and outer race grooves. The rotor agent is 36.7%

and 30% accurate, respectively, for the classifications of the samples with balls and outer race defects, which is considered a false positive. However, the bearing agent is 75.8% and 77% accurate, allowing the MAS to define the conditions for the bearing faults. Although in case of false positives occur in some agents, the final diagnosis is correctly performed by the MAS, once the decision rules are favorable for the determination of the winning agent. Based on these results, the generalization of the identification of faults shows satisfactory accuracy rates for higher powered motors, without including the respective samples in the training of classifier standards inherent in the performance of the agents integrated within the system. 5. Findings in comparison with previous works Table 7 compares the results, including scope and methodology of this research, with some recent studies in the literature. Considering studies related to fault classification in induction motors, this paper presents a method of pre-processing based on the discretization of stator current signal in the time domain, while the researched literature shows the use of techniques based on motor current signature analysis or vibration in the frequency domain in order to obtain the best accuracy indices [39–41] or approaches exclusively based on the time and frequency domains [42,43]. Another important aspect is the fact that this work considered the use of different classification strategies associated with MAS agents for fault diagnosis, while other works available in the literature basically employ the use of a single technique. In this study voltage unbalance of ±10% was applied under a wide level of load torque ranging from 8–150% and 6–112% respectively for the 1 and 2 hp motors. According to the studied literature, the implication of voltage unbalance in the task of pattern recognition is considered in the works of [43,40]. A large variation in load conditions appears in the works of [39,40]. In addition, this work considers the use of three different sizes of induction motors (1 and 2 hp), which allows observation of classification accuracy under different dynamic situations whereas other studies available in the literature employ the use of a single motor. Furthermore, this work presented the concept of fault multi classification where bearing, broken rotor bar and stator short-circuit faults were evaluated, while other studies presented results considering 1 or maximum 2 fault conditions. With regards to the accuracy of the results, the work of [40] reached 100% accuracy; however, this work did not cover an evaluation of bearing faults. Also in the work of [43], an accuracy of close to 100% was obtained disregarding the evaluation of stator faults and also considering reduced conditions of load variation. In this work, the average accuracy was 90% overall, considering the classification of three types of failures, their severity, and mainly the adversities employed in all tested conditions.

Table 7 Summary of recent published paper comparing the results of this research. References

[42]

[39]

[43]

[40]

[41]

This study

Motors Preprocessing Data source Class method Unbalance Load variation Stator fault Rotor fault Bearing fault Accuracy Multiclassification

2 hp T/F-FFT Vibration ANN No No No No Yes ≤90% No

2 hp F-FFT Current SVM No Yes No Yes No 64–90% No

0.75 hp T/F-FourierBessel Current ARTMAP Yes No No Yes Yes ≈100% Yes

1 hp F-FFT Current FUZZY Yes Yes Yes Yes No ≈100% No

A/C motor – NS F-FFT Vibration/Current ANN/ANFIS No No No No Yes 90–92% Yes

1 and 2 hp T-Discretize Current MAS Yes Yes Yes Yes Yes ≈90% Yes

T = time domain; F = frequency domain; NS = not specified.

R.H.C. Palácios et al. / Applied Soft Computing 45 (2016) 1–10

As for the studies considering the classification of short-circuit faults in the stator coils, several bearing faults and broken rotor bars in induction motors connected directly on line, the main contribution of this work is to present an alternative approach/concept of fault multi classification by means of an MAS. The adversities applied in the tests, as well as the demonstration of the system robustness, may provide enhanced security and encourage the implementation of a real-time system designed to monitor TIMs directly in the industry. 6. Conclusion In this work data signals were acquired from TIMs through the laboratory test bench. Accordingly, tests were conducted with three types of motors, under varying load and voltage supply unbalances, in order to detect bearing and broken rotor bar faults of varying severity, and different levels of short-circuits in the stator coils. A proper database was created in order to test various methods of classification standards and implement an intelligent MAS to detect the faults being investigated. Pre-processing of the data was based on the discretization method for the signal currents of the motors. This technique involved the construction of a vector with equidistant points of a wave half cycle of the same magnitude. These sets of points characterize the behavior of the motor and allow the classifier methods to create a standard model to identify problems related to the state of the machine. The use of an intelligent MAS allowed the classifying agents to interact or work together in order to perform a given set of tasks and achieve goals. As a result of the need for a hybrid system to enable the diversification of classifier methods in order to produce more reliable results, a system was developed with multi-agent entities responsible for different types of faults and classification performance. This technique proved to be efficient for the situations evaluated, demonstrating the multi-agent approach as an alternative to traditional methods for identifying faults in TIMs. Acknowledgments The authors gratefully acknowledge the contributions of CNPq (process #552269/2011-5), FAPESP (process #2011/17610-0), Araucária Foundation/CAPES (CP 13/2014), the University of São Paulo and the Federal Technological University of Paraná for their financial support toward the development of this research. References [1] M. Hajian, J. Soltani, G. Markadeh, S. Hosseinnia, Adaptive nonlinear direct torque control of sensorless IM drives with efficiency optimization, IEEE Trans. Ind. Electron. 57 (3) (2010) 975–985. [2] R.H.C. Palácios, I.N. da Silva, A. Goedtel, W.F. Godoy, M. Oleskovicz, A robust neural method to estimate torque in three-phase induction motor, J. Control Autom. Electr. Syst. 25 (4) (2014) 493–502. [3] T.H. dos Santos, A. Goedtel, S.A.O. da Silva, M. Suetake, Scalar control of an induction motor using a neural sensorless technique, Electr. Power Syst. Res. 108 (2014) 322–330. [4] M. Suetake, I. da Silva, A. Goedtel, Embedded DSP-based compact fuzzy system and its application for induction-motor V/f speed control, IEEE Trans. Ind. Electron. 58 (3) (2011) 750–760. [5] M. Riera-Guasp, J. Antonino-Daviu, G.-A. Capolino, Advances in electrical machine, power electronic, and drive condition monitoring and fault detection: state of the art, IEEE Trans. Ind. Electron. 62 (3) (2015) 1746–1759. [6] C.-C. Yeh, N.A.O. Demerdash, Induction motor-drive systems with fault tolerant inverter-motor capabilities, IEEE International Electric Machines Drives Conference 2 (2007) 1451–1458. [7] J. Zarei, M.A. Tajeddini, H.R. Karimi, Vibration analysis for bearing fault detection and classification using an intelligent filter, Mechatronics 24 (2) (2014) 151–157. [8] M. Irfan, N. Saad, R. Ibrahim, V. Asirvadam, Condition monitoring of induction motors via instantaneous power analysis, J. Intell. Manuf. (2015) 1–9.

9

[9] X. Zhang, Y. Liang, J. Zhou, Y. Zang, A novel bearing fault diagnosis model integrated permutation entropy, ensemble empirical mode decomposition and optimized {SVM}, Measurement 69 (2015) 164–179. [10] A. Bellini, F. Filippetti, C. Tassoni, G.A. Capolino, Advances in diagnostic techniques for induction machines, IEEE Trans. Ind. Electron. 55 (12) (2008) 4109–4126. [11] P. Shi, Z. Chen, Y. Vagapov, Z. Zouaoui, A new diagnosis of broken rotor bar fault extent in three phase squirrel cage induction motor, Mech. Syst. Signal Process. 42 (1–2) (2014) 388–403. [12] K. Yahia, A. Cardoso, S. Zouzou, S. Gueddidi, Broken rotor bars diagnosis in an induction motor fed from a frequency converter: experimental research, Int. J. Syst. Assur. Eng. Manage. 3 (1) (2012) 40–46. [13] B.M. Ebrahimi, J. Faiz, S. Lotfi-fard, P. Pillay, Novel indices for broken rotor bars fault diagnosis in induction motors using wavelet transform, Mech. Syst. Signal Process. 30 (2012) 131–145. [14] M. Drif, A. Cardoso, Stator fault diagnostics in squirrel cage three-phase induction motor drives using the instantaneous active and reactive power signature analyses, IEEE Trans. Ind. Inform. 10 (2) (2014) 1348–1360. [15] W.F. Godoy, I.N. da Silva, A. Goedtel, R.H.C. Palácios, Evaluation of stator winding faults severity in inverter-fed induction motors, Appl. Soft Comput. 32 (2015) 420–431. [16] S. Moosavi, A. Djerdir, Y. Ait-Amirat, D. Khaburi, ANN based fault diagnosis of permanent magnet synchronous motor under stator winding shorted turn, Electr. Power Syst. Res. 125 (2015) 67–82. [17] I.S. Torsun, Foundations of Intelligent Knowledge-based Systems, Academic Press, 1995. [18] E. Davidson, S. McArthur, J.R. McDonald, T. Cumming, I. Watt, Applying multi-agent system technology in practice: automated management and analysis of SCADA and digital fault recorder data, in: IEEE Power Engineering Society General Meeting, 2006, pp. 1–10. [19] S. McArthur, S. Strachan, G. Jahn, The design of a multi-agent transformer condition monitoring system, IEEE Trans. Power Syst. 19 (4) (2004) 1845–1852. [20] T. Nagata, H. Watanabe, M. Ohno, H. Sasaki, A multi-agent approach to power system restoration, International Conference on Power System Technology. PowerCon 2000 3 (2000) 1551–1556. [21] D. Koesrindartoto, J. Sun, L. Tesfatsion, An agent-based computational laboratory for testing the economic reliability of wholesale power market designs, in: IEEE Power Engineering Society General Meeting, 2005, pp. 2818–2823. [22] A. Dimeas, N. Hatziargyriou, Operation of a multiagent system for microgrid control, IEEE Trans. Power Syst. 20 (3) (2005) 1447–1455. [23] D. Buse, P. Sun, Q. Wu, J. Fitch, Agent-based substation automation, IEEE Power Energy Mag. 1 (2) (2003) 50–55. [24] J. Solanki, N. Schulz, Using intelligent multi-agent systems for shipboard power systems reconfiguration, in: 13th International Conference on Intelligent Systems Application to Power Systems, 2005, pp. 1–3. [25] J. Solanki, N. Schulz, W. Gao, Reconfiguration for restoration of power systems using a multi-agent system, in: 37th Annual North American Power Symposium, 2005, pp. 390–395. [26] K. Nareshkumar, M. Choudhry, J. Lai, A. Feliachi, Application of multi-agents for fault detection and reconfiguration of power distribution systems, IEEE Power Energy Society General Meeting (2009) 1–8. [27] B. Monchusi, A. Yusuff, J. Munda, A. Jimoh, Fuzzy multi-agent based voltage and reactive power control, in: International Conference on Renewable Energies and Power Quality, 2011, pp. 1–5. [28] Z. Xiangjun, K.K. Li, W. Chan, S. Sheng, Multi-agents based protection for distributed generation systems, in: IEEE International Conference on Electric Utility Deregulation, Restructuring and Power Technologies, vol. 1, 2004, pp. 393–397. [29] R. Battiti, A.M. Colla, Democracy in neural nets: voting schemes for classification, Neural Netw. 7 (4) (1994) 691–707. [30] G. Niu, T. Han, B.-S. Yang, A.C.C. Tan, Multi-agent decision fusion for motor fault diagnosis, Mech. Syst. Signal Process. 21 (3) (2007) 1285–1299. [31] C.F. do Nascimento, A.A. de Oliveira Jr., A. Goedtel, P.J.A. Serni, Harmonic identification using parallel neural networks in single-phase systems, Appl. Soft Comput. 11 (2) (2011) 2178–2185. [32] R.H.C. Palácios, I.N. da Silva, A. Goedtel, W.F. Godoy, A comprehensive evaluation of intelligent classifiers for fault identification in three-phase induction motors, Electr. Power Syst. Res. 127 (2015) 249–258. [33] V.N. Vapnik, The Nature of Statistical Learning Theory, Springer-Verlag New York, Inc., New York, NY, USA, 1995. [34] J. Platt, Fast training of support vector machines using sequential minimal optimization, in: B. Schoelkopf, C. Burges, A. Smola (Eds.), Advances in Kernel Methods – Support Vector Learning, MIT Press, 1998. [35] D. Aha, D. Kibler, Instance-based learning algorithms, Mach. Learn. 6 (1991) 37–66. [36] S. Haykin, Neural Networks: A Comprehensive Foundation, 2nd ed., Prentice Hall PTR, Upper Saddle River, NJ, USA, 1998. [37] E. Alpaydin, Introduction to Machine Learning, 2nd ed., The MIT Press, 2010. [38] F.L. Bellifemine, G. Caire, D. Greenwood, Developing Multi-agent Systems with JADE, Wiley, 2007. ´ F. Kulic, ´ M. Pineda-Sánchez, I. Kamenko, Support vector machine [39] D. Matic, classifier for diagnosis in electrical machines: application to broken bar, Expert Syst. Appl. 39 (10) (2012) 8681–8689.

10

R.H.C. Palácios et al. / Applied Soft Computing 45 (2016) 1–10

[40] M. Seera, C.P. Lim, D. Ishak, H. Singh, Offline and online fault detection and diagnosis of induction motors using a hybrid soft computing model, Appl. Soft Comput. 13 (12) (2013) 4493–4507. [41] H. Ertunc, H. Ocak, C. Aliustaoglu, ANN and ANFIS based multi-staged decision algorithm for the detection and diagnosis of bearing faults, Neural Comput. Appl. 22 (1) (2013) 435–446.

[42] J. Zarei, Induction motors bearing fault detection using pattern recognition techniques, Expert Syst. Appl. 39 (1) (2012) 68–73. [43] V.T. Tran, F. AlThobiani, A. Ball, B.-K. Choi, An application to transient current signal based induction motor fault diagnosis of Fourier–Bessel expansion and simplified fuzzy ARTMAP, Expert Syst. Appl. 40 (13) (2013) 5372–5384.