Detection and Classification of Partial Discharge by Hybrid Neural Network Algorithms in Air-insulated MV Switchgear

1. Introduction

Correct detection of partial discharge problems in MV switchgear is the foundation of its safe usage. Early recognition of the type of fault inside the switchgear can significantly reduce financial losses and, at an early stage of operation, save lives. Incorrect exploitation, environmental conditions (temperature, humidity, etc.), natural ageing of the insulation, as well as human mistake, have a huge impact on the possibility of unwanted PD occurring [1,2].

The dynamic development of Artificial Intelligence (AI) has significantly improved the effectiveness of the safe operation of electrical power equipment. The use of convolutional and deep neural networks, together with additional development of the architecture (the addition of different modules, depending on the type of measurement data), gives a satisfactory level of effectiveness in classifying faults caused by PD [3]. The continuous technological progress allows for the use of increasingly practical and cheaper methods of detecting disturbances from the point of view of switchgear manufacturers. In the industry, the classical electrical method using capacitive TEV sensors, High-Frequency Current Transformers (HFCT) or Ultra-High-Frequency (UHF) sensors still finds wide application. After a thorough review of the literature, a Transient Earth Voltage (TEV) sensor was used for signal measurement due to its reliability, simplicity of construction, ease of installation, and low cost [4,5,6,7]. These sensors are mounted on the outer surface of the metal housing of the switchgear. When PD occurs, the electromagnetic waves propagate away from the discharge point as a transient voltage. The transient voltage inside the metal shell cannot be detected directly from the external surface by the skin effect, However, the EM wave can propagate to any space where there is an electrical discontinuity in the metal shell, which can generate a series of transient voltage peaks at the surface. These peaks are called TEVs and were first proposed by Dr John Reeves in 1974 [8,9,10].

In the process of the research, the last fault, which is a series of many different signals superimposed on each other simultaneously - mainly signals from the partial discharges of faults one and two - was simulated artificially. Detection of single PDs from multi-source is still a challenge for individual differentiation and effective evaluation by different types of classifiers [11,12,13].

During the research process, it was decided to investigate the effectiveness of GoogleNet and SqueezeNet in classification based on the researchers' publications [14,15,16]. GoogleNet was checked, for example, during the recording of images of the corona around power transmission lines [17]. The authors decided to check the effectiveness of the aforementioned networks by creating spectrograms before the learning process [18,19,20]. The article continues with a study of the use of a hybrid combination of networks Stacked Autoencoder (SAE) with Feedforward Neural Network (FNN) [21,22,23,24], Convolutional Neural Network (CNN) combined with Autoencoder (AE) and Long Short-Term Memory network [25,26,27,28,29,30,31].

The first section of the article describes how the previous measurement and classification processes were carried out, presents how the MV switchgear defects were simulated, and then describes the test methods used during the research.

The next section briefly characterises the deep neural networks used for classification. The hybrid combinations of CNN, LSTM, SAE and GoogleNet and SqueezeNet used are described, and the classification results are then presented.

At the end, the last part of the article summarises the results of the research and indicates the next possible research steps.

2. Measurements and Methods

2.1. Previous Research

During previous research [32], the authors carried out experiments to simulate, "artificially", the various possible faults in MV switchgear. TEV sensors were placed on a metal plate. The measured insulation defects originated from electrodes placed behind the metal housing. First, a de-noising process was carried out, then normalisation and finally extraction of the relevant signal features. The complexity of the method is shown in Figure 1.

2.2. Current Research and Measuring System with TEV Sensors

The authors carried out the research at the High Voltage Hall at the Faculty of Electrical Engineering, Warsaw University of Technology. The following measurement equipment was used to carry out the research: a 220/100,000V, 10 kVA test transformer, a NorthStar VD-100 voltage divider and a brushless voltage regulator. The electrical schematic of the measurement system is shown in Figure 2.

2.2.1. Simulation of Possible Switchgear Failures

The authors decided to use a three-phase, single-pole air-insulated MV switchgear as the research object. The measurements were recorded by a Tektronix MSO56B oscilloscope. An internal 100 Hz high-pass filter was used for more precise recording.

It was determined to simulate four types of faults and one additional algorithm validation (five class):

• The first one (Class 1) - a copper wire was attached to one phase of the switchgear, which was at high potential;
• The second (Class 2) - a copper wire was attached to a grounded metal switchgear enclosure;
• Third (Class 3) - surface partial discharges were artificially induced inside the switchgear;
• The fourth (Class 4) - contained a confusion of multiple defects, including: leaving a wrench in the switchgear along with fault one, surface discharge with fault from high potential and surface discharge with fault from low potential;
• The last one (Class 5) - an additional class to validate the algorithm, which was actually the measured noise level.

Figure 3. Simulated faults from partial discharges: (a) Partial discharge from high potential; (b) Partial discharge from low potential; (c) Surface partial discharge; (d) Mounting error - spanner left in switchgear.

Figure 3. Simulated faults from partial discharges: (a) Partial discharge from high potential; (b) Partial discharge from low potential; (c) Surface partial discharge; (d) Mounting error - spanner left in switchgear.

2.2.2. Placement of TEV Sensors

Firstly, the effect of the placement of the HVPD TEV sensors (15 – 70 MHz frequency response) on the quality of the measured signal was checked during the research. Three TEV sensors were placed at three different locations on the switchgear:

• Close to the power supply;
• On the front door of the switchgear;
• At the point furthest from the PD source.

No differences were observed during the signal measurement. Placing the sensors at different distances from the discharge source did not affect the quality of the measured signal.

This time it was decided to use the raw signal waveform for the classification process, without manual feature extraction and de-noising process. The number of signals was 2000: 360 for fault 1, 480 for Class 2, 420 for fault 3, 600 for fault 4 and 140 for last Class 5. The length of each signal was 62494. Three examples (for Class 1, 2 and 3) of signals used during the conducted research are shown in Figure 4. In the given waveforms it can be observed a high background level and the individual pins come from the PD.

2.3. GoogleNet and SqueezeNet + Spectrograms

GoogleNet is a convolutional neural network (CNN) model developed in 2014. It consists of 22 neural network layers, including convolutional layers, dimension reduction layers, batch normalisation layers, ReLU activation layers and softmax output layers. A distinguishing feature of the GoogleNet architecture is the use of a so-called 'Inception module', which allows objects of different sizes and orientations to be detected in a single network laye [33].

SqueezeNet is a deep neural network model characterised by a smaller size than traditional models. This allows it to be used on devices with limited computing power, such as mobile devices or embedded systems. The main feature of SqueezeNet is the use of 1x1 convolutional layers instead of the traditional 3x3 or 5x5. These layers use fewer parameters, leading to a smaller model size, while still providing similar classification performance. Squee-zeNet also includes so-called 'fire modules', which are blocks that combine 1x1 convolution layers and 3x3 or 5x5 [34].

First, a spectrogram was created for each signal. These were then passed through the two networks mentioned above - GoogleNet and SqueezeNet - the process is shown in Figure 5.

2.4. Hybrid Neural Network

2.4.1. SAE+FNN

A hybrid model was used to train the neural networks: Stacked Autoencoder consisting of two hidden layers of 100 and 50 neurons each. The layers were trained in an unsupervised manner to extract features from the training data. Testing was carried out in MatLab - in the Deep Learning Toolbox environment. Network parameters: Epochs: 100; L2WeightRegularisation: 0.001; SparsityRegularisation: 4; Sparsity-Proportion: 0.1. The FNN was trained using the Adam optimisation algorithm with the following parameters: Learning Rate: 0.001; Epochs: 100 Batch size: 32.

Figure 6. Stacked autoencoder structure.

Figure 6. Stacked autoencoder structure.

2.4.2. 2. D-CNN+LSTM

For the hybrid combination of CNN and LSTM, the Keras [35] and Tensorflow [36] libraries were used. Network learning was carried out on an RTX 3070 graphics card, using mainly Conv2D layers with 64 filters, BatchNormalisation, Re-Lu activation, drop-out, as well as an LSTM recursion layer with 200 units and a sequence return option. The network model was compiled with the Adam optimiser taking on the following parameters: Learning Rate: 0.001; Epochs: 50 Batch size: 32.

Figure 7. 2D CNN + LSTM architecture.

Figure 7. 2D CNN + LSTM architecture.

2.4.3. AE+1D-CNN+LSTM

The last hybrid combination investigated, also using the Keras and Tensorflow library, was a combination of SA, CNN and LSTM networks. This time an autoencoder with a hidden layer of size 1984 was used. Learning parameters: Epochs: 150; Batch size: 32; Validation split: 0.2. The created CNN and LSTM network model consisted of Conv1D layers with 64 filters, batch normalisation, MaxPooling, as well as LSTM(256 and 128 respectively) and Drop-out. The optima-lizer Adam was used, with paramters: Learning Rate: 0.0005; Epochs: 1200 Batch size: 32.

The full architecture of the layers of the networks mentioned above is shown in Table 1.

3. Results and Discussion

Based on the collected measurement data during operation of the MV switchgear, a number of experiments were carried out to illustrate the effectiveness of the developed and used algorithms in PD detection and classification in the switchgear. The highest accuracy was obtained for the SqueezeNet and GoogleNet convolutional networks. MatLab's Deep Learning Toolbox environment allowed easy access to both networks. The modelled algorithm first created a spectrogram for each signal to be analysed, then the completed images were subjected to learning. Unfortunately, the time to create the spectrograms was quite long. The results for both cases are shown in Figure 8 and Figure 9.

The classification results for each classification method are summarised in Table 2 and Table 3. Based on Table 2, it can be determined that the classification of fault 1 - PD from high potential - caused the most problems. This may be due to the least amount of measurement data for this class. It is noteworthy that the classification of Class four was successful, despite the combination of signals resulting from multiple disturbances occurring simultaneously in the switchgear or human errors of the assemblers. The results of the last Class confirm that the algorithm performs well with basic PD detection.

The combination of SAE+FNN, CNN+LSTM and AE+CNN+LSTM networks achieved an accuracy of approximately 82%. The best result was achieved by the networks SqueezeNet and GoogleNet – 98.39% and 97.31%.

4. Conclusions

In the presented article, different methods were used to classify partial discharges occurring in MV switchgear in air insulation. The results obtained show that it was possible to significantly speed up the processing of the signal measured by the TEV sensors by using hybrid classification algorithms. Both GoogleNet and SqueezeNet achieved satisfactory classification results – 98% - on the basis of the created spectrograms, despite the addition of an ambiguous fourth fault. The other results obtained show the effectiveness of the adopted classification methods to be around 82%.

The result of the research prompts the creation of a universal capacitive sensor that continuously measures the possibility of partial discharge in MV switchgear. For this purpose, the authors will consider how to accelerate the creation of the spectrogram, as well as how to create an even more universal hybrid neural network.

In addition, the authors plan to investigate the effects of EMC disturbances, such as Burst - which simulates sparking at the contacts of electrical apparatuses in MV switchgear - and Surge, on the correct operation of a universal sensor placed on the external enclosure of an MV switchgear.