OLTC Fault detection Based on Acoustic Emission and Supported by Machine Learning

Abstract

Power transformers are an essential part of the power grid. They have a relatively low rate of failure, but removing the consequences is costly when it occurs. One of the elements of power transformers that are often the reason for shutting down the unit is the on-load tap changer (OLTC). Many methods have been developed to assess the technical condition of OLTCs. However, they require the transformer to be taken out of service for the duration of the diagnostics, or they do not enable precise diagnostics. Acoustic emission (AE) signals are widely used in industrial diagnostics. The generated signals are difficult to interpret for complex systems, so artificial intelligence tools are becoming more widely used to simplify the diagnostic process. This article presents the results of research on the possibility of creating an online OLTC diagnostics method based on AE signals. An extensive measurement database containing many frequently occurring OLTC defects was created for this research. A method of feature extraction from AE signals based on wavelet decomposition was developed. Several machine learning models were created to select the most effective one for classifying OLTC defects. The presented method achieved 96% efficiency in OLTC defect classification.

1. Introduction

The main goal of the present research is to assess the possibility of detecting defects occurring in on-load tap changers (OLTCs). The analysis results of time courses of acoustic emission (AE) generated during switching processes using wavelet decomposition are presented in this paper. The creation of a measurement database containing a wide range of tested OLTC defects is also described.

OLTCs are one of the elements of transformer equipment. Their purpose is to allow the voltage levels on the power grid to change by varying the number of active coils in the transformer. The OLTC mechanism works under load conditions so that energy can be supplied to consumers continuously [1].

Power transformers are an essential part of the power grid and have a relatively low rate of failure. However, the cost of remedying the consequences of losses is high. This is why diagnostic tests are so necessary [2]. Increasing emphasis is being placed on condition assessment without shutting down transformers, and such an approach allows us to avoid the costs associated with power outages [3]. One of the elements most frequently causing transformer failures is OLTCs [4]. The reasons for their loss can be divided into three groups [5]:

• Oil faults;
• Mechanical faults;
• Electrical faults.

The research presented in this article focuses on the detection of mechanical damage. During switching, the dynamic forces acting on the OLTC contacts gradually deform the contacts. In uneven contact wear, the OLTC switches the individual phases non-simultaneously. The spring magazine stores mechanical energy from the electric drive and then releases it, ensuring quick switching. It consists of two springs mounted axially. If one of them is damaged, storing the appropriate amount of potential energy will not be possible, leading to the extension of the switching process.

There are many diagnostic methods to assess the technical condition of OLTCs. One of the most recognized and widely used methods is dynamic resistance measurement (DRM). This method is presented in [6]. In the research presented in this manuscript, this method is used to confirm the occurrence of a defect. The most significant disadvantage of DRM is the inability to perform diagnostics on a loaded OLTC.

A direction often chosen when developing OLTC diagnostic methods is the analysis of vibrations generated during switching. Vibration waveforms are challenging to interpret, and expert knowledge is required from the diagnostician to assess the technical condition. For this reason, research is being conducted to use machine learning (ML) tools to simplify the diagnostic process. To train the ML model, descriptors must be extracted from the signal. Many works have presented proposals for this process. In [7], the authors used FFT analysis for feature extraction. Another method of obtaining information from vibration waveforms using the wavelet transform is presented in [8]. An approach using features extracted from the time–frequency domain by a minimum-entropy deconvolution filter is shown in [9]. Feature extraction based on calculating statistical values in the time domain is shown in [10].

Another approach to obtaining data enabling the assessment of the technical condition of the OLTC is recording the AE signals generated during the switching operation. This method uses piezoelectric transducers with a larger bandwidth compared to vibration measurements. Partial discharges and electric arcs generate acoustic signals in higher frequency bands. Thus, combining OLTC diagnostics from a mechanical and electrical perspective is possible.

AE signals, similarly to vibroacoustic signals, are challenging to interpret. Thus, ML models are proposed to classify OLTC defects. Several data feature extraction methods have been presented in the literature. The usage of Hilbert transform was shown in [11]. An attempt to analyze OLTC AE signals in the time–frequency domain is presented in [12]. A feature extraction method joining spectral density, time and frequency measures, and wavelet coefficient energies is shown in [13].

This paper mainly investigates the possibility of OLTC mechanical defect detection using ML tools. The proposed diagnostic method uses wavelet decomposition to extract details from EA signals, and then numerical values are determined for each detail. The present research stands out from other studies by using an original database of waveforms of EA signals recorded for OLTCs with defects. A wide range of contact wear was modeled, which may contribute to the creation of a diagnostic system that allows for the assessment of not only whether the OLTC is functional or damaged but also the extent to which it is worn. Such a system will allow for the estimation of the life of the device and the planning of renovations more effectively. A measurement database for over 3500 OLTC switches was created. This allowed for effective training with large amounts of data and reliable validation of the effectiveness of defect classification. The literature proposes using various types of transducers, most often accelerometers. A universal broadband transducer was used during the measurements, allowing for simultaneous OLTC diagnostics for mechanical and electrical defects. The number of defects examined during this research was much higher than in the previously presented studies. The possibility of determining the degree of symmetrical and asymmetrical contact wear was examined. Additionally, a case of damage to the spring energy storage was modeled. A wide range of ML algorithms were used to classify defects and were compared in terms of the effectiveness of OLTC diagnostics. In order to achieve the best classification performance, Bayesian optimization was used.

2. Materials and Methods

Databases are used during the process of generating ML models. The more samples they contain, the more effective classifiers can be created. Currently, there is no open database containing OLTC measurements enabling the classification of defects with the level of detail desired by the authors. Therefore, it was decided to perform a series of measurements to provide research materials.

2.1. Experimental Setup

The conducted research used an actual OLTC model. The OLTC used consists of a selector shortened to six taps and a VEL-110-type power switch. The switch is shown in Figure 1. The device was placed in a tank filled with transformer oil, corresponding to the actual operating environment of the tested OLTC. The system has an electric drive, providing the mechanical energy necessary for switching.

To avoid the need to switch the OLTC manually, a PLC controller was used. It performed a series of switches, which significantly accelerated the measurements. The next step taken to minimize the need to interfere in the data acquisition process was to use the drive current measurement as a trigger for the measurement card.

During the measurements, the following defects that may occur in OLTCs were examined:

• Class 1: Fully functional OLTC;
• Class 2: Symmetrical contact wear by 1 mm;
• Class 3: Symmetrical contact wear by 2 mm;
• Class 4: Symmetrical contact wear by 3 mm;
• Class 5: Symmetrical contact wear by 4 mm;
• Class 6: Asymmetrical contact wear;
• Class 7: Broken spring;
• Class 8: Damaged contact surface.

Symmetrical contact wear occurs when the contacts for the three phases wear evenly (Classes 2–5). This manifests in the extension of the switching time, which causes a longer current flow through the additional resistance of the auxiliary circuits. This may lead to overheating of the resistors and, in the long run, to a complete loss of the OLTC’s switching capacity. Several contact sets were used to model contact-wear defects, either symmetrical or asymmetrical. Each group was characterized by a varying degree of wear. Wear was simulated by mechanically machining the contacts. The view of the mounted, modified contacts is presented in Figure 2.

Asymmetrical wear of contacts (Class 6) results in the switching of individual phases, not simultaneously. The tests included asymmetrical switching of one phase. This defect was modeled using contacts with various degrees of wear.

The kinetic energy stored in the switch consists of two springs, one inside the other. A common cause of OLTC failure is damage to one of these springs. This defect was modeled by removing one of the springs (Class 7).

Class 8 covers contact damage by modifying their geometry without changing the effective contact thickness. This models extensive mechanical distortion, which often occurs in OLTCs. For this case, contacts with an increased contact area were used. The effective thickness of these elements was not changed.

2.2. Measurement Line

The measurement system used to record the AE signals generated by the OLTC during switching consisted of the following components:

• Piezoelectric transducer;
• Preamplifier;
• Amplifier;
• Data acquisition card;
• Laptop.

WD17AH was chosen as the acoustic wave transducer. This type of transducer was selected because it has a frequency response that matches the expected frequencies. Its additional advantage is high sensitivity. The data of the WD17AH piezoelectric transducer are shown in

Table 1. Technical data of the WD17AH transducer.

Frequency Band (kHz)	Peak Sensitivity, Ref V/(m/s)	Peak Sensitivity, Ref V/µbar
100–900	56 dB	−61 dB

. The transducer was mounted to the OLTC tank using a magnetic holder.

The signal from the transducer was then amplified using a preamplifier with a gain of 20 dB. The 2/4/6 PREAMPLIFIER manufactured by Physical Acoustics Corporation was used. The next element of the measurement path was the Wide-Bandwidth AE Amplifier, also manufactured by Physical Acoustics Corporation.

An Acquitek CH3160 acquisition card was used to acquire the AE signals. The signals were sampled with a frequency of 350 kHz. The KEW 8146 Leakage Clap Sensor was also connected to the card, allowing for motor load current monitoring. The current signal was used as a trigger for the measuring card. When the current reached the threshold value, the measurement was triggered. AE signals were recorded using AcquiFlex v2.0 software dedicated to the card used. In Figure 3, the AE measurement line is shown.

To enable verification of the occurrence of OLTC defects, the DRM method was used. This method is widely recognized for assessing the technical condition of OLTCs. An MT3 device was used to perform DRM. The device manufacturer provides OLTC.exe v2.10.510 software that allows for the control of MT3 and record waveforms. MT3 with a laptop is shown in Figure 4.

Over 400 switches were made for each tested defect, during which measurements were taken. This allowed us to train ML tools with a significant amount of data and to check their effectiveness reliably. The general diagram of the measurement line is presented in Figure 5.

2.3. Signal Processing and Classification

Sample recording started when the drive started and ended when the switchover was completed. All signals were normalized by dividing them by the maximum value, so the amplitude of each signal was in the range <−1;1>. The recorded runs consisted of approximately 2 million samples. The fragment corresponding to the switch itself was much shorter. Therefore, to speed up further work, it was decided to cut out significant fragments of the signals. The algorithm searched for the maximum amplitude and cut out a chunk of 50,000 samples. An example of the original waveform and the detected fragment where the switching occurred are presented in Figure 6.

The present research used discrete wavelet transform (DWT) to extract descriptors. DWT decomposes a waveform into several details. This is done by sliding a wavelet through the signal and multiplying it in each location. The results are the coefficients. The process is repeated for a larger-scale wavelet. An important issue when using DWT is the selection of the mother wavelet. During preliminary research, it was observed that the sym8 wavelet allows for achieving the best results in classifying OLTC defects. A similar mother wavelet was proposed in [14]. In the present study, the signal was decomposed into five details.

The following characteristic values were determined for each detail:

• Mean;
• Root mean squared;
• Form factor;
• Crest factor;
• Standard deviation;
• Skewness.

To achieve good results in ML training, the data must be normalized. A standard scaler was used for the data features. That is, the data features were shifted so that their mean was zero and scaled to achieve unit variance. The determined values were used as data for training ML models. The dataset was split in half into training and test datasets. The following ML models were selected:

• K-nearest neighbors (KNN);
• Decision tree (DT);
• Random forest (RF);
• Support-vector machine (SVM);
• Gradient boosting (GB);
• Adaptive boosting (AdaBoost).

KNN is a supervised learning classifier. Proximity is used to predict the class to which a given data point belongs. The basic assumption is that points belonging to the same classes are close to one another [15]. On the other hand, DT looks for the best series of tests to classify the data. This algorithm creates decision rules based on if-else statements. Each test is defined in a way that allows the data to be divided into classes as precisely as possible. Decision trees are very susceptible to overfitting [16]. This problem can be overcome by using RF. The RF classifier consists of a series of DTs. Each tree is trained on a different dataset. The classification result is the average output value of all trees [17].

SVM is a classifier whose training aims to determine a hyperplane separating examples belonging to two classes with a maximum margin. This algorithm is used to classify between two classes. When the issue of multi-class classification arises, the one-versus-rest approach is used. This approach involves training one classifier for each category. Each model predicts whether a data point belongs to the corresponding class [18]. Another technique used for classification in this research is GB, which generates a set of simpler ML models. Most often, these are DTs. In each iteration, a set of DTs are updated. They are fitted based on a negative gradient of the loss function [19]. A slightly different approach is used in AdaBoost classifiers. During model fitting, the significance of previously misclassified observations is boosted, and they are more likely to be used again to train the model to be more effective. As a weak learner for AdaBoost, DT was used [20].

Every classification model must be tuned to the currently examined dataset. Each model has different parameters, and their proper selection allows for a significant increase in classification efficiency. In the present research, the optimization of classifier parameters was performed using the Bayesian optimization method. This is a popular strategy for black-box function optimization. Initially, the optimizer checks the performance of the classifiers for random sets of parameters. Then, based on the results obtained, it approximates the objective function. The subsequent matched sets of parameters are determined at the maximum of the objective function. The algorithm continues for a given number of iterations.

3. Results

The first step in analyzing the results was to determine the actual existence of the modeled defects. The DRM was used for this purpose. Two sample current waveforms are presented in Figure 7. The presented results show the results for a fully functional OLTC (Figure 7a) and with asymmetry (Figure 7b). DRM results were analyzed for each defect. A generally accepted way to interpret DRM results is to compare the times read from the waveforms. Figure 7 shows three characteristic times:

• T₁—from the moment of opening of the first primary contact to the closing of the second resistive contact;
• T₂—from the moment of closing the second resistive contact to the moment of opening the first resistive contact;
• T_C—total switching time.

The read characteristic times for all classes without asymmetry are presented in

Table 2. DRM characteristic times for classes without asymmetry.

Class	1	2	3	4	5	7	8
	Characteristic Times (ms)
T1	16	19	24	29	35	43	17
T2	23	21	20	18	14	37	22
TC	50	54	59	62	67	81	49

. The results presented are the average times read for ten randomly selected measurements from each class. Between Classes 1 and 5, a gradual increase in T₁ and a decrease in T₂ can be noticed. This is a characteristic symptom of contact wear. As expected, the damaged spring (Class 7) resulted in a significant extension of T₁, T₂, and T_C. For Class 8, there are no significant differences compared to Class 1. This is due to the lack of change in the contact thickness affecting the individual times.

Checking the correctness of modeling the defect described in Class 6 is slightly different than in the case of the other classes. If OLTC switching asymmetry is checked, the characteristic times for individual phases should be compared. The times are not compared to those of other classes. The appearance of differences in the lengths of individual connection stages means the asymmetrical operation of the OLTC. The read times for Class 6 are presented in

Table 3. DRM characteristic times and time differences for asymmetry (Class 6).

	Characteristic Times (ms)			Time Difference (ms)
Phase	A	B	C	A–B	B–C	C–A
T1	10	6	14	4	8	4
T2	20	35	15	15	20	5
TC	41	39	62	2	23	21

.

The interpretation of DRM measurements allowed for the detection of defects. It was found that the defect modeling methods used allowed for achieving the desired results. However, the DRM diagnostics did not allow for the diagnosis of mechanical deformation of the contact (Class 8).

Sample AE waveforms for each class are presented in Figure 8. Detailed interpretation of the results of the AE method is problematic because it is impossible to assign individual acoustic events to specific switching stages. The most visible defect is the asymmetrical operation of the OLTC. The area corresponding to contact switching (for Class 6: 45–90 ms) is characterized by more acoustic events with a lower amplitude. This is caused by non-simultaneous switching of the contacts of the individual phases. For defects without asymmetry, diagnosis is possible only by overlapping the recorded waveforms with the reference waveforms, but despite this procedure, the diagnostic results may be ambiguous. Therefore, we propose using ML to simplify OLTC diagnostics.

The method of feature extraction from AE waveforms proposed in this study is based on wavelet decomposition. Figure 9 shows the results of the decomposition of an example waveform. The statistical values calculated, 30 for each detail, constituted a database used to train the ML models.

The method used to extract the data features together with the selected method of optimizing the ML models allowed for achieving satisfactory results. The percentage accuracy results achieved using the individual algorithms are presented in Figure 10. Each model was trained five times. Each time, the training set was randomly selected from the entire set of data features. The presented results were achieved using half of the database as training data. The SVM algorithm performed the best, achieving an efficiency of 96%. The RF and GB algorithms also demonstrated effectiveness in classifying OLTC defects. Their efficacy was slightly worse than that of SVM. DT is susceptible to overfitting, especially in classification cases containing many classes. An example of such a situation is the presented OLTC defect classification. This is the reason for the relatively low effectiveness of DT in the proposed studies. Using many decision trees, the RF algorithm reduces the risk of overtraining. Using a weak classifier more advanced than, for example, tree stumps (one-level DT) made the model susceptible to overtraining. Additionally, the large number of classes in the training set increased the chance of this unfavorable situation. AdaBoost using DT turned out to be the worst classifier. The parameters of the most effective model are presented in

Table 4. SVM model parameters optimized using Bayesian optimization.

Kernel	Gamma	C
RBF	0.087	3.136

.

Figure 11 presents the classification accuracy depending on the size of the training set. The dataset consists of features extracted from over 3500 tap changes. The smallest training set tested contained 1% of the total dataset. The rest of the data that were not used for training were used to test the model. Initially, one may notice a strong correlation between training dataset size and performance. More than 30% of the data were used as training data, and no significant increase in effectiveness was observed. The size of the created measurement database allows for effective training and validation of ML models.

An analysis of the prediction effectiveness for individual classes of OLTC defects was performed for the best classifier. The results are shown in Figure 12. Every bar represents the percentage rate at which the corresponding class was predicted correctly. Each fault was diagnosed with a high efficiency of over 90%. The feature extraction method combined with SVM allows for determining the good technical condition of the OLTC, with 100% compliance. It can be seen that the defect recognized with the greatest efficiency was Class 4, which corresponds to contact wear of 4 mm.

A confusion matrix allows for the determination of which defects are confused with one another. The confusion matrix for SVM, shown in Figure 13, was normalized over columns corresponding to accuracy in successful class prediction. The percentage values on the diagonal represent how many predictions for each class were made correctly. For the SVM model, the class with the least effective prediction was Class 2 (symmetrical contact wear by 1 mm). This was mainly confused with classes describing other degrees of contact wear.

The most common classification error was incorrectly classifying a damaged spring (Class 7) as asymmetrically worn contacts (Class 6).

Generally speaking, each defect was classified with high efficiency. This is a satisfactory result considering the high level of detail used in dividing the causes of OLTC failure. Particular attention should be paid to the increased effectiveness of the proposed method in determining the absence of a defect. The SVM model produced no false positive predictions for Class 1.

4. Discussion

The research results presented in this article confirm the possibility of diagnosing OLTC faults using EA signals. The effectiveness in determining the OLTC defects was similar to the results presented in the literature [21,22,23]. However, our research used a measurement database containing a much larger number of defects occurring during OLTC operation. The possibility of diagnosing varying degrees of symmetrical contact wear, along with spring damage, contact deformation, and asymmetrical contact wear, was examined. The lack of false positive predictions of a properly functioning OLTC allows us to define the proposed tool as useful for maintaining OLTCs. Detailed fault diagnosis is expected to allow for the intelligent planning of transformer shutdowns for maintenance, reducing the risk of failure. The AE diagnostic method is a non-invasive method that will enable the creation of a system that monitors the OLTC status online, returning information about the OLTC status after each switchover. However, before introducing such a system, tests should be conducted on real OLTCs operating with transformers. The impact of noise on the fault classification performance should be investigated and, if necessary, a method should be developed to reduce noise in the signals.

The presented feature extraction method allowed us to achieve satisfactory results for all of the tested ML algorithms. The original measurement database created for this research contained large samples. This allowed for effective training of the ML models and reliable performance verification. It is expected that the proposed method of analyzing AE signals will allow for the effective diagnosis of various types of OLTC. For each type of switch, a database should be created storing AE measurements for devices with defects. It is proposed that such a database would contain similar amounts of data to that created for the purposes of the present research.

Future work will focus on determining additional descriptors to increase the effectiveness of diagnostic ML models. We also plan to test the efficacy of other artificial intelligence tools, such as neural networks. Convolutional neural networks are expected to enable further improvement of the diagnostic method using AE. Work will be carried out to determine the possibility of using the method described in this article to combine mechanical and electrical aspects of OLTC diagnostics.

The research presented here is part of broader research aimed at developing an online OLTC condition monitoring system.