Stator winding short-circuit fault diagnosis in induction ... - IEEE Xplore

2 downloads 0 Views 3MB Size Report
Abstract—In this paper, an approach to detect stator winding short-circuit faults in squirrel-cage induction motors based on. Random Forest and Park's Vector is ...
Stator Winding Short-Circuit Fault Diagnosis in Induction Motors using Random Forest ∗ Tiago

dos Santos, † Fernando J. T. E. Ferreira, Senior Member, IEEE, ‡ Jo˜ao Moura Pires, and § Carlos Dam´asio

∗ ‡ §



Department of Computer Science, NOVA-LINCS, Nova University of Lisbon (FCT/UNL), Lisbon, Portugal Dep. of Electrical and Computer Engineering, Institute of Systems and Robotics, University of Coimbra, Coimbra, Portugal ∗ Altran Portugal, Portugal E-mails: ∗ [email protected], † [email protected], ‡ [email protected], § [email protected]

Abstract—In this paper, an approach to detect stator winding short-circuit faults in squirrel-cage induction motors based on Random Forest and Park’s Vector is proposed. This is accomplished by scoring the unbalance in the current and voltage waveforms as well as in Park’s Vector, both for current and voltage. To score the unbalance in the d-q space, a Principal Component Analysis is applied to Park’s Vector and with the first two principal components the eccentricity is calculated, while the first principal component is used to determine the phase in shortcircuit. The proposed strategy has been experimentally tested on a special 400-V, 50-Hz, 4-pole, 2.2-kW induction motor with reconfigurable stator windings in which it was possible to emulate different types of inter-turn short-circuits. The results are quite promising, even only using 1-kHz sampling frequency to acquire the current and voltage waveforms in the three phases, and the use of the Fast Fourier Transform is avoided. The developed solution may be used for tele-monitoring of the motor condition and to implement advanced predictive maintenance strategies.

Keywords: fault diagnosis; induction motor; inter-turn short-circuit; stator winding; PCA; predictive maintenance; machine learning; classification; random forest.. I. I NTRODUCTION The three-phase squirrel-cage induction motor (SCIM) is the most used kind of electric motor due to its relatively low cost, good efficiency and high availability - representing about 85-90% of the electric motors installed in the industry [1]. However, this kind of motor is not free from unexpected faults [2]. Stator winding short circuits can form 26% of the faults occurring in electrical machines and in small machines can go up to 90% [2]. Although the probability of failures in SCIMs is very low [3], in most industrial facilities these motors are critical and, therefore, condition monitoring and fault detection mechanism are of uttermost importance. It is possible to detect stator winding short-circuits through physical based (model based) approaches or data based approaches - or even a combination of both [4]. A physical approach may be very useful in terms of parameter analytics with the insight of the underlying physical model. A data based approach is an alternative to physical methods where the model structure and parameters are identified through experimental data. A data based approach provides several advantages over physical based approaches, such as non-ideal assumptions about the motor, learning of the sensors behavior [5] and can be used to classify the state of the motor into different types

of faulty states and different fault severity degree - without the need of complex physical models [4]. Techniques based on the traditional Motor Current Signature Analysis (MCSA) approach or other more recent methods based on modern signal analysis methodologies have been proposed [4], [6], [7], being many of them based on the comparison of the spectrum between the diagnosed machine and the same machine in a healthy state. Other proposed techniques are the Extended Park’s Vector (EPVA) [8] and the Park-Hilbert method [9], also relying on the spectral analysis. Data based approaches have also been proposed, such as [10]–[12]. In [10], a power logger and a vibration analyzer are used to extract features and, by using a Support Vector Machine (SVM), stator winding short circuits and bearing faults are detected. In [11], the authors use the alpha-beta stator current as input to a Hebbian-based neural network to extract the Principal Components Analysis (PCA). By rotating the data with the PCA’s matrix output, it is decided where the fault lies. In [12], the authors use wavelet transform to analyze motor line currents and use a feedforward Artificial Neural Network (ANN) to fault characterization based on fault features extracted using continuous wavelet transform. We propose a data based approach where a Random Forest (RF) is used. The input for the RF is the result of a feature extraction process that takes place in the raw data, as well as in the d-q space. In the d-q space, by applying PCA, the components are used to score the unbalance and to associate the direction of the first principal component to the phase in fault. The Random Forest algorithm has the property of being inherently parallelizable, which is a major advantage when considering the application of this model to an industrial scale. II. DATA G ENERATION S ETUP In order to apply the proposed approach, the 50-Hz line currents and line-to-line voltages waveforms of the three phases (R, S and T), have been acquired at a 1-kHz sampling rate (which is relatively low), using a commercial datalogger (InMonitor) with Wi-Fi communication capability, as it can be seen in Fig. 1. Five cycles were acquired (100 points, 100 ms) and an averaged 1.6 cycles were computed (32 points, 32 ms) and sent via Wi-Fi to an online database, in which the data is finally recorded and used by high-level computing.

c 978-1-5090-4281-4/17/$31.00 2017 IEEE.

TABLE I: Data size in faulty state by load level and shortcircuit (SC) mode. SC Current Phase A Phase B Phase C

Fig. 1: Experimental setup diagram to implement the proposed fault diagnosis strategy.

Fig. 2: Motor stator winding configuration with an external resistor (RSC ) to emulate short-circuits.

The average waveform is a way to mitigate some noise and/or interference transients during the data acquisition process. This process has been repeated 1-in-1 or 2-in-2 minutes, depending on the restriction of the Wi-Fi communication. The acquisition of the waveforms of the three currents and three voltage was made simultaneously (parallel acquisition, without delays between the analogue inputs). Since the starting point of each acquired cycle (with 6 waveforms) is not always the same and there is naturally a phase shift between the different phases, all the waveforms have been properly rotated in time to the same phase, i.e., have the zero, maximum and minimum points at the same time instant. Of the 32 points, only 20 have been actually used, since they represent a full period. For the experiments, a special 400-V, 50-Hz, 4-pole, 2.2-kW delta-connected SCIM (Fig. 3), with reconfigurable stator windings with external access to the terminals of all the coils. The stator core has 36 slots, and the stator winding has two sets of coils, each one with 6 coils per phase. Each slot contains 2 coils sides (double-layer winding). One coil in each phase has also external access to

(A) L NL L NL L NL

2.5 151 147 151 146 130 155

2.0 18 19 0 0 0 0

50 turns SC 1.5 1.25 24 19 22 26 0 0 0 0 0 0 0 0

0.25 19 24 0 0 0 0

25 turns SC 2.5 25 23 19 18 19 18

intermediate/tap points. Each coil has a maximum current of 2.5 A. When both coil sets are connected in parallel, the motor can provide full power and the maximum phase current is 5 A. For the sake of simplicity, only one set of coils has been used, limiting the phase current to 2.5 A and the motor output power to 1.1 kW. To emulate the stator winding short-circuits, an external resistance has been connected to the external terminals, as illustrated in Fig. 2 . To avoid damaging the motor winding, the maximum allowed short-circuit current was the coil rated current (2.5 A). Data have been collected for different motor states, namely, healthy with and without load, 2.5-A shortcircuit on all the paired coil terminals in each phase, with and without load, and 1.5-, 1.0-, 0.25-A short-circuits in one coil of a given phase. Table I shows the number of cycles generated given the phase in fault, the number of coils in fault, the fault current and the motor condition. When one coil is short-circuited (total of 50 turns) it represents a shortcircuit of 16.7% of the total number of turns per phase (total of 300 turns; 6 coils per phase connected in series; Fig. 2), whereas when half a coil is short circuited (total of 25 turns) it represents a short-circuit of 8.3% of the total number of turns per phase. III. R ANDOM F OREST Ensemble Learning is a learning process where a set of learning algorithms are used to be joined in a final model. The motivation to use ensemble techniques is that the ensemble classifier is likely to have a lower error rate or the variance of the ensemble classifier will be lower when compared with certain unstable classification models, such as decision trees, that have high variance [13]. One technique that can help the task of minimizing generalization error is bagging, which can reduce the variance of classifier models. Bagging, which refers to Bootstrap Aggregating, consists on taking repeatedly samples with replacement from the training dataset and, for each bootstrap sample, a model is trained. The bagging ensemble prediction is defined to be the class with the most votes among models predictions (for classification models) or the average of the predictions among ensemble models (for estimation models). One important observation is that the process of aggregation has the effect of reduce the error due to model variance, since the ensemble is formed by different model instances that were trained with different datasets (sampled from the training dataset) [13].

Fig. 3: Motor used to emulate the stator winding short-circuits.

Another important observation comes from the expected prediction error value in this ensemble method, that can be defined as: 1 × σ 2 (Yˆ ) (1) m where Z is the average of models predictions, m is the number of models trained with resampled, and σ 2 (Yˆ ) represents the variance of the set of models predictions Y . Given that, we can observe that the higher the m is, the lower the expected prediction error gets. RF is an ensemble learning technique that makes use of the bagging task, and can be used both for classification and regression. RF are based on a combination of decision tree learners, where each decision tree is trained on a independently drawn random sample (bootstrap sample) of the training dataset. As expected from (1), the expected prediction error of a RF converges to a limit as the number of trees becomes higher. Nevertheless, the RF generalization error will depend on the correlation between the trees [14]. Therefore, during the training process, each tree in the forest uses a random selection of features to split at each node in order to minimize the correlation between trees, improve accuracy of the RF and maintain the strength of the tree learner [14]. E((Z − y)2 ) =

IV. F EATURE E XTRACTION The proposed feature extraction process takes place in the raw data and in the d-q space. This feature extraction process takes into account the variations of the supply current which may happen due to external factors. To do so, an average of the previous 6 records unbalance attributes is calculated and joined to the record that is being processed. The presented feature set were select after a preliminary assessment where the predictive power of different features set were evaluated, being the presented the one with best predictive power among all the considered. The features that will be extracted are: (a) first Principal Component (PC) coordinates of the current in d-q space, (b)

eccentricity of the current in d-q space using the two PCs, (c) averages of the previous 6 record’s first PC coordinates of the current in d-q space, (d) averages of the previous 6 records’s eccentricity of the two PC of the current in d-q space, (e) score of the three phase currents unbalance, (f) average of the previous 6 records’s score of the three phase currents unbalance, (g) root mean square value for each phase current, (h) averages of the RMS value of the previous 6 records for each phase current, (i) first PC coordinates of the voltage in d-q space and (j) eccentricity of the two PC of the voltage in d-q space. Summed up, the total input number for the RF are 18 features. For some records, there are no 6 previous records. For that subset, the values of the features (c)(d)(f) are a missing value and they are not considered neither for training or for testing purposes. A. Raw Data The features obtained from the raw data are (e)(f)(g)(h)(i), forming a total of 9 features. Before extracting any feature, each record is normalized by its maximum peak value. This allows for the unbalance study without penalizing the cases where the motor is on load condition (which would occur in this scoring process as well as in the PCA transform, described in subsection IV-C). To score the three-phase current unbalance, the current RMS value for each phase (IA , IB , IC ) is calculated and then the current unbalance was calculated according to (2).  unbI =

max (|IA − avg|, |IB − avg|, |IC − avg|) avg IA + IB + IC avg = 3

 (2)

B. Park’s Vector For each record, the Park’s Transform is applied to the currents and voltages, resulting into the Park’s Vector, which, over an entire period, results into a circle (symmetrical condition) or an ellipse containing some ripple (asymmetrical condition).

The instantaneous values of the direct (d) and quadrature (q) vectors, id and iq , resulting from the application of the Park Transform to the instantaneous values of the three line currents, ia , ib and ic , are given by:  q q q id = 2 ia − 1 ib − 1 ic 3 6 6 q q (3) iq = 1 ib − 1 ic 2 2 This is an intermediate step so that the PCA can be applied. In the Park’s Vector there are several characteristics that allow the determination of several short-circuit faults. From the resultant ellipse’s format, it is possible to know if there is a short circuit in phase A (Fig 4a on the left), phase B (Fig 4a in the middle) or phase C (Fig 4a on the right). This strategy avoids the use of the FFT applied to the Park’s Vector (Extended Park’s Vector Approach – EPVA), in which a component of 2 times the supply frequency reflects clearly the existence of a winding unbalance resulting from an interturn short circuit [8], but requires a higher computing effort, which could be a drawback if thousands of motors are to be permanently monitored in a given industrial facility. C. PCA Transform The features obtained from the d-q space are (a)(b)(c)(d)(i)(j). PCA is a widely used method (a linear orthogonal transformation) which yields directions of maximum variance. For this reason, this method is normally employed for dimension reduction, given that the directions that have a greater data variance are those that describe better the data. As a result, PCA is able to provide the main directions of the given data on the space-vector and, when applied to the elliptical form of the Park’s Vector, the two first components correspond to the major and minor axis (as shown in Fig. 4b). From here, an unbalance score is calculated as an eccentricity: score = 1 − e2 /e1

(4)

where e2 is the eigen-value of the second principal component and the e1 is the eigen-value of the first principal component. PCA is, in principle, easier to implement than the FFT/EPVA, requiring less computational effort, but it has more limitations in terms of detection of different motor fault modes. Anyway, as we shall show, for detecting stator winding shortcircuits it is an effective approach. V. M ODEL S ELECTION An assessment on the predictive power of several classification algorithm have been done. The caret [15] package (version 6.0-73) was used and from the 168 classification algorithms provided by the package, a total of 135 were used in this assessment. The models were train in a context of a binary problem (classify examples as healthy or shortcircuited) with bootstrap (25 repeats, Accuracy was used to choose the best model instance) and a model parameter tune length of 3. From this set of 135 models, only those that scored

more or equal than 0.95 regarding Accuracy metric where consider. From those, after a further study of the techniques, the final subset of models to consider were reduced to 10 (C5.0, RF, Stochastic Gradient Boosting, Neural Networks, Boosted Logistic Regression, Least Squares SVM with RBF, SVM RBF, SVM with Class weights, k-NN). After this preliminary model evaluation, another assessment for binary classification regarding the influence of metaparameters in the models (class-balance, resample technique used, chosen metric to chose the best model during the resample process, type of preprocessment applied to data) was made, were the modeling process consisted on randomly dividing the dataset in train set (50%), validation set (20%) and test set (30%). As a final result of this assessment, there were two models instances that stood out by having a total accuracy: RF and Stochastic Gradient Boosting. At this point, the RF was chosen due to its inherently parallelizable nature. The Stochastic Gradient Boosting is an ensemble method that makes use of the task Boosting, which is inherently not parallelizable. The observation that the chosen model is an inherently parallelizable one is important since it is desirable to apply the presented model in an industrial scale, since it can process a large set of information faster than a non-parallelizable model. VI. M ODEL T RAINING AND E VALUATION The number of records available for the learning process after the feature extraction process and outlier removal (Tukey’s method, regarding d-q unbalance feature) were 1159, where 357 are healthy cases and 802 are faulty cases. The dataset was divided into train (50%), validation (20%) and test (30%) sets. In order to detect short-circuit faults with identification of the phase in fault, this paper presents two approaches: one approach is to model one binary-class classification problem to classify between Healthy state and Faulty state plus model a multi-class classification problem to identify the phase in fault (presented in subsection VI-A); the other approach is to model a multi-class classification problem to classify between Healthy, Short-Circuit Phase A, Short-Circuit Phase B and Short-Circuit Phase C (presented in subsection VI-B). Both approaches will be evaluated on how accurate they are at distinguish an healthy example from a short-circuited example, how sensitive they are to the available short-circuit conditions, as well as how accurate they are at identifying the phase in short-circuit. To evaluate the models, the metrics used in this work are Accuracy, Recall and Specificity given by (5), (6), and (7), respectively, where T P is the True Positives count, T N is the True Negatives count, F N is the False Negatives count, F P is the False Positive count and C is the count of all instances involved (T P + T N + F P + F N ). Accuracy =

TP + TN C

(5)

(a) Park’s Vector for a 2.5-A coil-terminal short-circuit in phases A, B and C (respectively).

(b) Principal Components of the Park’s Vector with their eigen-value.

Fig. 4: Park’s Vector representation in the d-q space.

TP Recall = TP + FN Specif icity =

TN TN + FP

(6)

TABLE II: Confusion Matrix for multiclass classification in approach VI-A. (a) Confusion Matrix for multiclass RF (phase in fault identification) in Validation Set. Phase A

Actual Phase B

Phase C

72 0 0

0 34 1

0 0 53

(7)

A. Two-Classifier Approach This approach consists on training two models: a model for short-circuit detection and a second model, which is only used when the first model detects a short-circuit, that identifies the phase in short-circuit. A RF was trained to classify the data in two classes - healthy class and faulty class. For this process, the convention was that a positive case is a faulty case and a negative case is an healthy case. During the training process, 4-fold crossvalidation resample technique was used and to choose the best model during the resample stage the F1 Score metric was used. To train this RF, the selected number of trees to grow was 500 and the number of features randomly sampled as candidates at each split was 2. The selected training parameters were obtained through a lot of experimentation, where the presented configuration was the one with highest performance. When evaluating on the training set, the binary RF scores an Accuracy metric of 1, which means that the binary RF predicts every example correctly. When evaluating on the validation set, as well as on the test set, the binary RF scores an Accuracy metric of 1, meaning that every example is predicted correctly. Subsequently, a multi-class RF was trained to classify the data into Phase A, Phase B and Phase C - indicating in which phase was the short-circuit occurring, being the model trained only with examples in short-circuit state. During the train process, 4-fold cross validation resample technique was used and to choose the best model during the resample stage the Kappa metric was used. To train this RF, the selected number of trees to grow was 500 and the number of features randomly sampled as candidates at each split was 2.

Predicted

Phase A Phase B Phase C

(b) Confusion Matrix for multiclass RF (phase in fault identification) in Test Set.

Predicted

Phase A

Actual Phase B

Phase C

80 0 0

0 69 2

0 0 90

Phase A Phase B Phase C

When evaluated on the training set to predict the phase in fault of the examples, the trained model shows an Accuracy metric of 1, meaning that all the examples are predicted correctly. When evaluated on the validation set o predict the phase in fault of the examples, the trained model shows an Accuracy of 0.99%, presenting only one misclassification case (as shown in Table II(a)). When evaluated on the test set set o predict the phase in fault of the examples, the trained model shows an Accuracy of 0.99%, presenting only two misclassification cases (as shown in Table II(b)). B. Single-Classifier Approach This approach consists on training a model to classify between four classes: healthy, short-circuit in Phase A, shortcircuit in Phase B and short-circuit in Phase C. To this end, a RF was trained using 4-fold cross validation resample technique (using Kappa to choose the best model),

TABLE III: Confusion Matrix for multiclass classification in approach VI-B.

TABLE IV: Accuracy on short-circuit detection by shortcircuited turns and short-circuit current.

(a) Confusion Matrix for binary classifier. Healthy

Predicted

Healthy Phase A Phase B Phase C

65 1 2 0

Actual Phase A Phase B 0 66 0 0

1 0 53 0

Phase C 0 0 0 44

(b) Confusion Matrix for multiclass classifier. Healthy

Predicted

Healthy Phase A Phase B Phase C

96 1 6 0

Actual Phase A Phase B 0 107 0 0

2 0 64 1

Phase C 5 1 0 64

with a selected number of trees to grow of 500 and a number of features randomly sample as candidates at each split of 2. The performance of the model on the validation set is presented in Table III(a), while the performance of the model on the test set is presented in Table III(b). When evaluating the capacity of the model distinguish between an healthy state and a faulty state, the model scores a Recall of 0.97 on both validation and test set (meaning there are examples where the model predicts Healthy instead of a short-circuit) and a specificity of 0.94 (meaning there are example that where the model predicts as Faulty state (SC A, SC B, SC C) instead of an Healthy state). When calculating the weighted accuracy for Healthy and Faulty detection for the validation and test set, this model present scores 0.97. When evaluating the Accuracy with which the model detects the phase in short-circuit, the model scores an Accuracy of 0.99 in the validation set and an Accuracy of 0.98 in the test set. These calculations were made by ignoring the Healthy column on the confusion matrix in Table III and calculating the Accuracy (total of correct classifications divided by all considered cases). C. Evaluation of the Two Approaches 1) Short-Circuit Fault Detection: For the task of shortcircuit detection, approach VI-A presented a total accuracy on both Validation and Test set, which represent 50% of the dataset. By training with 50% of the dataset, this approach was able to classify correctly all the remaining examples. For the same task, approach VI-B presents a recall of 0.97 when evaluated on both validation and test set, which indicates there were examples where the model predicts Healthy instead of a Short-Circuit. Not only there are short-circuit misclassified as Healthy, but approach VI-B scores a Specificity of 0.94, meaning there are examples that the model predicts as Faulty instead of Healthy state.

SC Current (A) Approach VI-A Approach VI-B

2.5 1.00 1.00

50 turns SC 2.0 1.5 1.25 1.00 1.00 1.00 1.00 1.00 1.00

0.25 1.00 1.00

25 turns SC 2.5 1.00 0.93

In spite of approach VI-B presenting an inferior performance than approach VI-A, the performance of approach VI-B is still high and quite good. 2) Model sensitivity of the short-circuits: The presented approaches can also be evaluated based on the sensitivity they present for different short-circuit conditions. Table IV present the Accuracy that each approach performed in the aggregate of Validation and Test set. It is possible to observe that approach VI-A was able to identify all the short-circuit example as such, and therefore the Accuracy for all the short-circuit conditions is 1. Still on Table IV, it is possible to observe that in the shortcircuit examples where 50 turns were shorted approach VI-B scores an Accuracy of 1, but in the short-circuits examples where 25 turns were shorted it scores an Accuracy of 0.93. When analyzing the misclassified cases, it was possible to observe that there were three misclassified Phase B short-circuits in no load condition, while there were five misclassified Phase C short-circuits in load condition. This way, approach VI-A presents a better performance than approach VI-B on detecting short-circuits with 25 shorted turns. 3) Identification of the Faulty Phase: Regarding the task of identifying the phase in short-circuit, the presented approaches’ evaluation is presented in Table V by measuring the Recall and Specificity metrics for both approaches. Analyzing the results for Phase A, both approach VI-A and approach VI-B present a Recall of 1 (meaning every shortcircuit occurring in phase A was classified as such). However, while approach VI-A scores a Specificity of 1 (meaning no short-circuit was misclassified as occurring in phase A), approach VI-B presents a Specificity of 0.99 (meaning there were occurrences of short-circuit misclassified as occurring in phase A). Analyzing the results for Phase B, both approaches present a Recall of 0.97 (meaning that there were short-circuit in phase B that were classified as being in phase A or phase C). Regarding Specificity, approach VI-A scores 1 (meaning no short-circuit was misclassified as occurring in phase B), while approach VI-B scores 0.98 (meaning there were occurrences of short-circuit misclassified as occurring in phase B). Analyzing the results for Phase C, approach VI-A presents a Recall of 1 while approach VI-B presents a recall of 0.95. Regarding Specificity, both approaches present a score of 0.99. This way, approach VI-A presents a better performance than approach VI-B in the task of identifying the phase in shortcircuit.

TABLE V: Performance Analysis on short-circuit phase identification.

Phase A Phase B Phase C

Recall VI-A VI-B 1.00 1.00 0.97 0.97 1.00 0.95

Specificity VI-A VI-B 1.00 0.99 1.00 0.98 0.99 0.99

VII. C OMPARISON WITH OTHER M ETHODS The methods with which the proposed method will be compared are the methods refereed in related work, Section I. It was not possible to compare with the method presented in [6] with the presented one due to lack of information about sample frequency and experimental setup. In [8] the EPVA method for short-circuit detection is presented, where for the experimental setup it was used a 15-kW SCIM with a rated current of 29.5 A. The motor have a total of 147 turns per phase, and the experiments consisted on reading the motor line current (no information about the sample frequency is provided) for different scenarios: healthy, 3 shortedturns, 12 shorted turns and 36 shorted turns - representing a percentage of shorted turns per phase of 2%, 8.2% and 24%. In these experiments no more information is provided regarding phases in fault, load condition or under/over-voltage. When comparing [8] with the proposed method, in which the setup has data of the three phases under short-circuit condition and with different faulty currents, being the less severe faults the ones with only 25 shorted turns, corresponding to 8.3% of the total turns shorted in phase. Given this data representativity, the proposed methodology is capable of accurately detect all the faults without the need of spectrum analysis. In [9] the authors present the Park-Hilbert method for shortcircuit detection, using a sample frequency of 10-kHz to read the line current of a 3-kW SCIM. The motor have a total of 200 turns per phase, and they are capable of detect short-circuit faults with 4 turns and 10 turns, representing a percentage of turns per phase shorted of 2% and 5%. When comparing [9] with the proposed method, the proposed method only uses a sample frequency of 1-kHz (which generates less data and therefore is friendlier to the industry) and, in spite of the tests do not have examples of short-circuits with only 2% severity (regarding turns shorter per total turns in phase), we believe that with the proposed methodology it is possible to detect such. Still, the proposed methodology is capable of accurately detect short-circuits with a severity of 8.3% (turns shorter per total turns in phase) without the need of spectrum analysis. In [10], the authors present a SVM to detect short-circuits, using a sample frequency of 1kHz to read the line current of a 2.2-kW induction motor. Nothing is said about the total number of turns per phase, and the experimental tests are made with short-circuits faults on phase B of 3, 5 and 5 turns. The SVM presents an accuracy of 99.33% for detecting shorter turns. When comparing [9] with the proposed method, the proposed approach VI-A presents a better performance for more diverse data, while approach VI-B presents a similar but lower performance at detecting short-circuits.

In [12], the authors present an ANN receiving its input from the wavelet transform. Nothing is said about the sample frequency at which the motor line current is read. The motor used during experimental tests is a 1.5-kW induction motor and the number of turns per phase is 200. The experiment data contains short-circuits with 4 and 10 turns, which represent 2% and 5% of turns shorter per number of turns in phase. The ANN presents an Accuracy of 1 at detecting short-circuits. When comparing [12] with the proposed method, given the lack of information regarding sample frequency, it is not possible to make a fair comparison. Still, the proposed method presents a wider set of conditions on training data and it is also capable of score an accuracy of 1 (approach addressed in Section VI-A) at detecting short-circuits. VIII. C ONCLUSION AND F UTURE W ORK The presented work intend to present a more practical way for industrial condition monitoring, given that the goal is to apply the presented method to thousands of motors. On the basis of the experimental tests, it can be concluded that the approach where two models are used (Section VI-A) has better predictive power than the approach where one model is used (Section VI-B), at the expense of having to train two models and to have the diagnosis process to be processed by two models (which may not be a disadvantage at all). Still, in scenarios where the use of just one model is mandatory, it is shown that the approach in Section VI-B is also capable of achieving high metric values and sensitivity on detecting short-circuits. On the basis of the experimental tests, it can be concluded that both approaches can effectively detect inter-turn shortcircuits using only 120 data points (20 points per phase and variable) and without the use of FFT, potentially leading to a less computing effort. Since the industrial three-phase SCIMs do not use the neutral conductor, only two line current and two line-to-line voltages have to be acquired for this method, leading to a total of 80 points to be sent to the database, reducing the cost of the setup. The certainty level of the performed diagnosis was quite good, as shown in Section VI-C. The presented work can be improved by acquiring data associated with undervoltage, overvoltage and unbalanced voltage conditions. Another aspect that would improve the presented method is the estimation of the stator winding short-circuits severity. ACKNOWLEDGMENT The authors would like to thank Altran Portugal for supporting this work under the project REARM (pRedictivE mAintenance of electRic Motors). R EFERENCES [1] F. J. T. E. Ferreira and S. M. A. Cruz, “Vis˜ao geral sobre selecc¸a˜ o , controlo e manutenc¸a˜ o de motores de induc¸a˜ o trif´asicos,” Manutenc¸a˜ o, vol. 101, pp. 46–53, 2009. [2] R. M. Mccoy and E. L. Owen, “Assessment of the reliability of motors in utility application - UPDATED,” no. 1, pp. 39–46, 1986.

[3] M. Reliability, W. Group, P. Systems, R. Subcommittee, P. Systems, E. Committee, I. Industry, and A. Society, “Report of Large Motor Reliability Survey of Industrial and Commercial Installations : Part III,” vol. I, no. 1, pp. 153–158, 1987. [4] M. Riera-Guasp, J. A. Antonino-Daviu, and G.-A. Capolino, “Advances in Electrical Machine, Power Electronic, and Drive Condition Monitoring and Fault Detection: State of the Art,” IEEE Transactions on Industrial Electronics, vol. 62, no. 3, pp. 1746–1759, 2015. [5] V. M. Janakiraman, “Machine Learning for Identification and Optimal Control of Advanced Automotive Engines,” Ph.D. dissertation, The University of Michigan, 2013. [6] V. F. Pires and D. Foito, “Detection of Stator Winding Fault in Induction Motors Using a Motor Square Current Signature Analysis (MSCSA),” vol. 5, no. 1. [7] A. Gandhi, S. Member, T. Corrigan, S. Member, L. Parsa, and S. Member, “Recent Advances in Modeling and Online Detection of Stator Interturn Faults in Electrical Motors,” vol. 58, no. 5, pp. 1564–1575, 2011. [8] S. M. A. Cruz and A. J. M. Cardoso, “Stator Winding Fault Diagnosis in Three-Phase Synchronous and Asynchronous Motors , by the Extended Park ’ s Vector Approach,” vol. 37, no. 5, pp. 1227–1233, 2001. [9] S. M.a, G. A.b, G. S.a, and Z. S.E.a, “Detection of inter-turn short-circuit in induction motors using Park-Hilbert method,” International Journal of Systems Assurance Engineering and Management, vol. 5, no. 3, pp. 337–351, 2014. [10] R. A. Patel and B. R. Bhalja, “Condition Monitoring and Fault Diagnosis of Induction Motor Using Support Vector Machine,” Electric Power Components and Systems, vol. 44, no. 6, pp. 683–692, 2016. [11] J. F. Martins, V. F. Pires, and A. J. Pires, “Unsupervised Neural-NetworkBased Algorithm for an On-Line Diagnosis of Three-Phase Induction Motor Stator Fault,” vol. 54, no. 1, pp. 259–264, 2007. [12] A. Jawadekar, S. Paraskar, S. Jadhav, and G. Dhole, “Artificial neural network-based induction motor fault classifier using continuous wavelet transform,” Systems Science & Control Engineering: An Open Access Journal, vol. 2, no. June 2016, pp. 684–690, 2014. [Online]. Available: http://dx.doi.org/10.1080/21642583.2014.956266 [13] D. Larose and C. Larose, Data Mining and Predictive Analytics, 2015. [14] L. Breiman, “RANDOM FORESTS,” pp. 1–33, 2001. [15] M. Kuhn, “The caret Package,” 2016. [Online]. Available: https://topepo.github.io/caret/