Universidade Federal de Minas Gerais Departamento de Engenharia Elétrica Evaluation of Electrical Insulation in Three-Phase Induction Motors and Classification of Failures Using Neural Networks Armando Souza Guedes . Sidelmo Magalhães Silva . Braz de Jesus Cardoso Filho Cláudio Alvares Conceição Published in: Electric Power Systems Research, June 2016
General Rights Copyright and moral rights for publications made accessible in the public portals are retained by the authors and/or other copyright owners and it is a condition of accessing publications that users recognize and abide by the legal requirements associated with these rights. Users may: • download and print one copy of the publication from the public portal for the purpose of private study or research. • not further distribute the material or use it for any profit-making activity or commercial gain • freely distribute the URL identifying the publication in the public portal If you believe that this document breaches copyright please contact us providing details, and we will remove access to the work immediately and investigate your claim.
Evaluation of Electrical Insulation in Three-Phase Induction Motors and Classification of Failures Using Neural Networks Armando Souza Guedesa,⇤⇤, Sidelmo Magalh˜aes Silvaa,⇤, Braz de Jesus Cardoso Filhoa,⇤, Cl´audio Alvares Concei¸c˜aob,⇤⇤ a Graduate
Program in Electrical Engineering - Federal University of Minas Gerais - Av. Antˆ onio Carlos 6627, 31270-901, Belo Horizonte, MG, Brazil b Departament of Industrial Electrical Maintenance - Petrobras/REGAP - Rodovia BR 381 km 427, 32689-898, Betim, MG, Brazil
Abstract This paper presents a study on the evaluation of stator electrical insulation of Three-Phase Induction Motors (IM) and the classification of the failure mechanism using an approach based on Computational Intelligence Tools (CIT). A brief review showing the main parameters for the evaluation of insulation condition and testing of IMs is presented, as well as the promising use of CITs for fault diagnosis of industrial equipment, including motors. This paper proposes a new methodology for evaluation and classification of insulation conditions with the aid of K-means clustering and of a classifier based on ANNs (Artificial Neural Networks). Keywords: induction motor, insulation resistance, artificial neural networks, clustering data. 2010 MSC: 00-02, 99-00 1. Introduction
5
10
15
20
25
The use of electric drives based on induction motors increased sharply in industry, especially due to the increase of the reliability and flexibility of operation of these devices 30 compared to other electric machines [1]. The robustness of its construction and the low maintainability give the IM a main role in most critical processes of an industrial plant, replacing drives previously performed by gas or steam turbines [1]. 35 Despite the high Mean Time Between Failure (MTBF) of induction machines, factors related with the coupling between the motor and the load, environmental, operational and power quality problems can lead to their untimely unavailability. The predictive detection of failures 40 in IMs have been analyzed in several studies and various techniques have been developed in order to allow maintenance of these machines without production interruption. In this context, insulation failures, responsible for about 40% of the stops of IMs, are an important issue to be in- 45 vestigated from the predictive point of view allowing the recognition and monitoring of this kind of failure [2]. A failure in the motor insulation leads to its immediate unavailability, causing production stops, undesirable financial losses and, in some cases, high maintenance costs.To 50 identify low insulation resistance in the stator of Low and ⇤ UFMG,
Belo Horizonte-MG, Brazil S.A., Brazil Email addresses:
[email protected] (Armando Souza Guedes),
[email protected] (Sidelmo Magalh˜ aes Silva),
[email protected] (Braz de Jesus Cardoso Filho),
[email protected] (Cl´ audio Alvares Concei¸c˜ ao) ⇤⇤ Petrobras
Preprint submitted to Electric Power Systems Research
55
Medium Voltage (LV, MV) IMs, preventive tools are being currently applied only during scheduled shutdowns of the machine [3]. During a schedule shutdown, Resistance to Ground (RTG), Polarization Index (PI) and the Absorption Index (AI) are currently measured, without any analysis or inference of the degradation state of the insulation. For this reason test results are restricted to condition “approved” or “not approved”, without any information on the degradation status of the insulation. Recently, many industries are adopting predictive tools for monitoring and evaluation of the insulation status of motors with the use of predictive analyzers [4]. The utilization of these devices along with a maintenance plan are considered success factors for identifying causes of insulation degradation in the planned maintenance of motors. In the context of detection and evaluation of failures in industrial equipment, the use of CITs is becoming more popular. Techniques using ANNs and data clustering has been successfully implemented in applications such as robotics and industrial systems [5, 6], partial discharge analysis [7] and current electrical signature analysis in motors [8, 9, 10, 11, 12], diagnosis of dissolved gases in transformer oil [13], vibration analysis in steam turbine [14], bearing fault detection [15], etc. The capability of generalization and learning of the ANNs has motivated its use in various areas of knowledge (engineering, medicine, economics, biology, etc.). Applications of classification, pattern recognition, control, prediction, diagnosis and fault detection with the use of ANNs are being performed with great success [16, 17]. This work presented a bibliographical review of the main parameters for evaluation of insulation and some May 2, 2016
60
65
70
techniques for diagnostics of failure in industrial equipment and IMs. The purpose was to study mechanisms of degradation and failure of stator insulation system, presented a new methodology for evaluation and classification of insulation conditions with the aid of K-means clustering 105 and of a classifier based on ANNs. In this context the classifier based in ANN will have a fundamental function. The diagnosis informed by ANN enables the technician and the electrician, performers of test, to take a rapid decision about operation condition or the best procedure for maintenance of the motor, discar-110 ding the work of the specialist. The classifier also assist in the root causes analysis of cases of premature failure of the insulation. 2. Techniques for Evaluation of Stator Insulation
75
80
85
90
95
100
References indicate that 60 to 70% of failures in IMs are due to mechanical problems, especially bearing failure caused by inadequate lubrication [2]. The second largest cause of faults, 30 to 40%, in IMs are of electrical origin, with remark to electrical insulation failures in the stator covering 80 to 90%. In medium voltage machines electrical insulation failures represent a percentage of 60 to 70% of the overall causes of electrical failures [2]. In case of mechanical failure, predictive maintenance programs based on vibration analysis are well disseminated, bringing significant reductions of maintenance costs as well as reducing undesirable shutdowns [18]. For electrical insulation failures in the stator of IMs, the most widespread preventive and predictive techniques are: 1. RTG measurement, PI [3, 19] and AI calculating [20]; 2. Partial Discharge Analysis (PDA)[7, 21, 22, 23]; 3. Dissipation Factor (DF) or Power Factor (PF) insulation measurement and Power Factor Tip-Up (PFTU) 115 or Dissipation Factor Tip-Up (DFTU) test [19, 21, 24, 25, 26, 27, 28]; 4. Capacitance to Ground measurement (CTG) and Capacitance Tip-Up (CTU) test [21, 24, 29]; 5. AC Hipot [30], DC Hipot [31] and Surge test [32].
< 1000
500
1000
2500
500
1000
2501
5000
1000
2500
5001
1200
2500
5000
5000
10000
> 12001
- absorption current IA : current portion responsible for the polarization of the dielectric material. It has exponential decay and typically becomes constant after 10 minutes; - capacitive current IC : current portion occurring due to the capacitive nature of the isolation system. It has fast exponential decay and approaches zero after a few minutes.
Figure 1: Decomposition of current IG in the RTG test [3].
The minimum RTG values recommended by the IEEE Std-43 are seen in Table 2. The RTG value is measured 1 minute after the start of the test. Its value is corrected to 40o C to allow the comparison between RTG values along the motor lifecycle.
RT GM IN (M ⌦)
Motor test Motors manufactured before 1970 and others not listed below. Motors manufactured after 1970 with windings form coils. For motors with random windings in the stator, motors with windings form coils rated below 1kV and dc armatures.
Table 1: Guidelines for direct voltages to be applied during insulation resistance test.
Insulation resistance test direct voltage (V )
- conductive current IG : current portion constant during the test, responsible for the dielectric losses in the insulating material;
Table 2: Minimum RTG values recommended at 40o C.
2.1. RTG Test, PI and AI Calculation The RTG test is performed with direct voltage for 10 minutes, with a magnitude that depends on the winding rated voltage as shown in Table 1 [3].
Winding rated voltage (V )
The total current IT that circulates in the insulation during the test can be decomposed into three components, as shown in Figure 1 [3].
1
kV 1+1 100 5
Nominal voltage motor or winding.
The PI and AI values are defined as: PI = 2
RT G10minutes RT G1minute
(1)
AI = 120
RT G60seconds RT G30seconds
(2) 155
The minimum recommended PI values according to IEEE Std-43 and AI according to NFPA (National Fire Protection Association) are seen Table 3.
2.3. DF or PF Measurement and PFTU or DFTU Test
Table 3: Minimum recommended PI and AI values.
AI (NFPA)
PI (IEEE) Thermal class rating A Thermal class rating B, F e H
> 1.5 > 2.0
160
> 1.5 165
125
Some manufacturers recommend the evaluation of PI and AI within ranges that may indicate the insulation degradation condition. Table 4 shows values of PI, AI and the estimated insulation state [29, 33]. Table 4: PI and AI values recommended by manufacturers.
PI
AI
PI< 1.0
-
bad
1.0 PI< 1.25
questionable
2.0 PI< 4.0 PI
4.0
1.25 AI< 1.4 1.4 AI< 1.6 AI
1.6
170
Insulating state
1.0 100M ⌦. The NEMA MG-1 [36] recommends only that these tests be performed on new or repaired motors [37]. The Surge test is recommended according IEEE Std522 for motors above 100kW . The test consist in the application of DC voltage pulses on two identical stator windings. The waveform obtained from the test reveals the insulation state to ground and between turns. Despite being recommended for new and repaired machines, monitoring test responses can diagnose the insulation deterioration over the motor lifecycle [32]. Figure 4 illustrates typical responses expected in the Surge test: - Figure 4-a: shows the waveform for good winding (overlapping signals); - Figure 4-b: shows the waveform for winding with short circuit between turns (delayed signals); - Figure 4-c: shows the waveform to the disconnected winding; - Figure 4-d: shows the waveform for completely winding short circuited to ground.
285
290
295
Figure 4: Typical waveforms expected in the Surge test. 300
250
255
260
265
3. ANN Applications on Motors Diagnostics Currently Artificial Neural Networks (ANNs) have been applied in di↵erent fields of knowledge assisting in classi-305 fication applications, pattern recognition, control, prediction, failures detection, diagnosis, etc [16]. In industrial systems ANNs find applications in control process and fault diagnosis such as in machines and robots [5, 6]. In industrial equipments ANNs find applications in vi-310 bration analysis aiding in the classification of failure of bearings, axial collision or friction, unbalance, oil whip and damaged shaft coupling [14, 38, 39, 40]. ANNs applications in failures diagnosis in IMs are concentrated mainly in Electric Current Signature Analysis (ECSA), enabling the detection and evaluation of the fol-315 lowing defects [8, 9, 10, 11]: 1. 2. 3. 4.
270
275
280
broken rotor bars detection; unbalance and rotor eccentricity; misalignment and coupling fatigue; short circuit between the stator coils.
320
Reference [8] presents the application of a Multilayer Perceptron (MLP) Artificial Neural Network for diagnosing faults in broken rotor bars, voltage imbalance and short circuit between stator coils. The network is trained and validated using 6 di↵erent types of learning algorithms: 325 1. 2. 3. 4.
Levenberg Marquardt (LM); Scaled Conjugate Gradient (SCG); Gradient Descent (GD); 330 Gradient Descent with Adaptive and Momentum Learning Rate Backpropagation (GDX); 5. Conjugate Gradient Backpropagation (CGB); 6. BFGS Quasi-Newton (BFGS). 335
5
The learning algorithm that achieved greater accuracy (95%), with fewer interactions for convergence of the Mean Square Error (MSE) and the lowest time, was the LM. The remaining algorithms achieved accuracy between 75% and 90% and greater computational cost. Reference [9] shows an application of a hybrid network with the use of inference mechanism Fuzzy Max-Min (FMM) and the neural network Classification and Regression Tree (CART), denominated by the authors as FMMCART network. In this paper the authors use the FMM for extracting learning rules as input CART network. From the construction of a hierarchical tree, FMMCART algorithm determines the best separation of classes. The network was applied to diagnose faults of eccentricity in the motor, with input data as the stator current spectrum. The results show accuracy greater than 99%, above the traditional networks MLP, but with values close to SVM (Support Vector Machines) [9]. Reference [10] shows the application of ANN with activation functions type Radial Basis Function (RBF) for identifying faults related the motor condition, load and power supply. The ANN is used as the base algorithm for fault detection replacing the traditional algorithms of motor protection relays. The purpose of this study was to detect and classify faults as: 1. 2. 3. 4. 5.
motor overload; undervoltage and overvoltage in the motor; voltage imbalance in power supply; phase fault in the motor; healthy motor.
The RBF network inputs are current and instantaneous voltages, that after training has achieved 100% accuracy in the tested scenarios. Reference [11] shows comparative study using di↵erent CITs for the detection of failures in IMs. The methods discussed in this work are: 1. 2. 3. 4. 5.
Naive Bayes; k-Nearest Neighbor; SVM (Sequential Minimal Optimization); ANN (MLP); Repeated Incremental Pruning to Produce Error Reduction; 6. C4.5 Decision Tree. Bearing failures, short circuit between stator coils, broken rotor bars, power supply and mechanical loading conditions were evaluated using ECSA. ANN and SVM presented the highest accuracies with values greater than 99% in all scenarios considered. For the diagnosis of defects and insulation failures of electrical equipment research is focused at PDAs of high voltage motors, estimated insulating lifetime in the laboratory and gases dissolved in oil transformers. Reference [7] presents MLP application of the occurrence of partial discharges for classification and location:
340
1. 2. 3. 4. 5. 6.
due to the corona e↵ect; internal discharges in isolation; end-winding discharges; surface discharges; slot discharges; not PD.
390
395
345
350
355
360
365
The learning algorithm used in MLP was the BackPropagation type GD. The authors report an accuracy of 94.5% in the tests, up from 79% of other methodologies that do not use computational intelligence. Reference [13] shows the application of an ANN for predicting failure in oil insulated power transformers from the analysis of the concentration of dissolved gases. The implemented ANN is based on MLPs with Back-Propagation type LM learning algorithm. The results obtained after network training achieved an accuracy of 93.75% for the diagnosis and 91.66% for the forecast of failure time, much superior than the traditional methods used. Reference [41] presents the ANN application for life prediction of insulating material Relanex. In this work, a MLP network is used with a learning algorithm BackPropagation type LM. The structure of the network is composed by the following inputs: -
Ba : activation energy of polarization; BV : activation energy of conduction; Uk : breakdown voltage; t: evaluation time.
Table 5: Summary of characteristics of the database motors.
380
385
Power (hp)
Voltage class
Number of motors
Refinery
50 250
LV MV
18 39
50 250
LV MV
18 40
To perform the grouping of the IRP the K-means clustering method was used. The K-means algorithm has great application in non-supervised machine learning for organizing, building and understanding models[42, 43]. The principle of the K-means is build clusters (Ck ) of similar data minimizing intra-cluster variance [44]. 8 9 K,m K n 4 and AI>1.5, without a perceived abrupt change in resistance during the test [2, 20].
i, j = 1 : n , i 6= j
410
415
420
425
430
The type of measure distance employed in the K-means algorithm depends on the application. In temporal series, where the format is more important, at a distance of correlation better results are shown when compared to the metrics that are based on the magnitude, as for example the Euclidean or Manhattan distance [42]. Each IRP of the database has 120 points of RTG measurement every 5 seconds (10 minutes of test according to the IEEE Std-43). A priori, was defined a number of clusters equal to 4, because it was expected to find 3 classes of defective motors and 1 class of motors without defect. Applying the K-means algorithm in curves IRPs with the correlation distance as the measure of similarity (equation12), were obtained 4 distinct clusters. Some reports in database have inspection photographs of failed motors. This information helped the specialist to classify the 4 groups returned by K-means algorithm in: -
motor motor motor motor
insulation insulation insulation insulation
Figure 5: Motor with insulation stator in good condition.
in good condition; featuring moisture degradation; featuring oil/dirt degradation; featuring thermal degradation.
Table 6 shows the quantitative IRPs per group made by K-means and classified by the specialist . Table 6: Quantitative per class made by specialist resulting from K-means clustering.
435
Clustering
Number of motors
Moisture
34
Oil/dirt
40
Good
132
Thermal
10
Figure 6: Typical IRP of insulation in good condition at 400 C.
Some examples of IRPs for the groupings defined in 450 Table 6 are shown bellow. Figure 5 shows a photograph of the stator of a motor exemplifying the insulation in a good condition. This is a LV motor that was removed for maintenance due to a mechanical problem. Its stator was inspected and tested 7
Figure 7 shows the stator of LV motor removed for maintenance due to moisture in the insulation. In this figure, rust marks are observed in the core due to oxidation caused by presence of water.
Figure 7: Motor with insulation stator in moisture contamination process.
455
Figure 8: IRP of motor with insulation stator in moisture contamination process at 400 C.
Table 7 presents the date and results of tests. The IRP this motor is seen in Figure 8. In this case, shutdown and maintenance of machine occurred before insulation failure. Table 7: Tests of insulation stator in good state and moisture contamination process.
Test date
460
465
470
RT G(M ⌦) CT G(pF )
PI
AI
DF
Dec/2012
4750
7100
3.7
1.54
0.00128
Jul/2013
581
50800
1.57
1.33
0.00156
Figure 8 illustrates the e↵ect of moisture on the IRP of an insulation of a motor. The e↵ect of moisture increases the CTG and decreases the RTG, leading to a fast exponential decay of current IC (a few seconds). Current IA is also a↵ected by moisture, because water molecules require less bias energy in relation to an insulating material, thus the time constant of the decay of current IA will be smaller. In this case, PI and AI indexes have similar values, but smaller than recommended. Figure 9 shows the stator of LV motor removed for maintenance due to thermal degradation in the insulation. In this figure one can observe dryness and breakage of the coils lashings caused by the action thermal and aging. Table 8 presents the date and results of tests. The IRP of this motor can be seen in Figure 10. Also in this case, the shutdown and maintenance of the machine occurred before the insulation failure.
Figure 9: Motor with insulation stator in thermal degradation process.
Table 8: Tests of insulation stator in good state and thermal degradation process.
Test date
475
480
8
PI
AI
DF
Fev/2013
RT G(M ⌦) CT G(pF ) 1175
11200
4.39
1.75
0.00202
Mai/2014
371
6980
8.57
7.87
0.01024
Figure 10 shows the IRP of a motor insulation that was subjected to thermal degradation due to overloading or aging. The e↵ect of temperature causes a decrease of the CTG, decrease of current IC , close values of PI and AI, although larger than recommended. In Figure 10 small variations of resistance shown over the test can be explained by the occurrence of failure in the insulation layers and partial discharges.
495
500
senting contamination by external agent such as oil, grease, dirty, etc. The e↵ect of the contamination increases CTG and decreases RTG, in a less accentuated manner, compared to the e↵ect caused by moisture. Currents IC and IA will have their exponential decay occurring faster, but with abrupt variations in the insulation resistance during the test. The resistance variations seen in Figure 12 may be associated with a more superficial contamination, generating partial discharges on the insulation surface. In a similar manner as the failure mechanism by thermal degradation, in the contamination by external agent, partial discharges also decrease the RTG and cause the insulation failure.
Figure 10: IRP of motor with insulation stator in thermal degradation process at 400 C.
485
Figure 11 shows the stator of MV motor removed for maintenance due to contamination by oil and dirty in the insulation. In this figure it is observed a large quantity of lubricating oil with dirty in the coils. Table 9 presents some data and the resulting RTG of this test. The IRP of this motor is seen in Figure 12. Shutdown and maintenance of the machine occurred before of the insulation failure.
Figure 12: IRP of motor with insulation stator in contamination process by oil/dirt at 400 C.
505
510
Table 10 summarizes the changes in parameters RTG, CTG, PI, PI and DF in relation to the type of insulation degradation mechanism and its typical IRP. An arrow pointing upward indicates an increase in the parameter by the specific mechanism. An arrow pointing downward indicates that the considered mechanism decreases the indicated parameter. Table 10: Summary of trend parameters for a type of insulation degradation mechanism.
Mechanism RTG Figure 11: Motor with insulation stator in contamination process by oil/dirt.
Table 9: Tests of insulation stator in good state and contamination process by oil/dirt.
Test date
490
RT G(M ⌦) CT G(pF )
PI
AI
DF
Nov/2012
5700
22300
4.71
1.64
0.00072
Dec/2013
878
52800
1.81
1.33
0.00086
Figure 12 shows the IRP of a motor with insulation pre-515 9
CTG
PI
AI
DF
Moisture
#
""
