Portal de Periódicos da CAPES

Novel colour image encoding system combined with ANN for discharges pattern recognition on polluted insulator model

2020; Institution of Engineering and Technology; Volume: 14; Issue: 6 Linguagem: Inglês

10.1049/iet-smt.2019.0297

1751-8830

Imene Ferrah, A. K. Chaou, D. Maadjoudj, M. Teguar,

Water Quality Monitoring and Analysis

IET Science, Measurement & TechnologyVolume 14, Issue 6 p. 718-725 Research ArticleFree Access Novel colour image encoding system combined with ANN for discharges pattern recognition on polluted insulator model Imene Ferrah, Corresponding Author Imene Ferrah imene.ferrah@g.enp.edu.dz Laboratoire de Recherche en Electrotechnique, Ecole Nationale Polytechnique, 10 Avenue Hassen Badi B.P 182 El-Harrach, 16200 Algiers, AlgeriaSearch for more papers by this authorAhmed Khaled Chaou, Ahmed Khaled Chaou Laboratoire de Recherche en Electrotechnique, Ecole Nationale Polytechnique, 10 Avenue Hassen Badi B.P 182 El-Harrach, 16200 Algiers, AlgeriaSearch for more papers by this authorDjamal Maadjoudj, Djamal Maadjoudj orcid.org/0000-0001-9263-1990 Laboratoire de Recherche en Electrotechnique, Ecole Nationale Polytechnique, 10 Avenue Hassen Badi B.P 182 El-Harrach, 16200 Algiers, AlgeriaSearch for more papers by this authorMadjid Teguar, Madjid Teguar Laboratoire de Recherche en Electrotechnique, Ecole Nationale Polytechnique, 10 Avenue Hassen Badi B.P 182 El-Harrach, 16200 Algiers, AlgeriaSearch for more papers by this author Imene Ferrah, Corresponding Author Imene Ferrah imene.ferrah@g.enp.edu.dz Laboratoire de Recherche en Electrotechnique, Ecole Nationale Polytechnique, 10 Avenue Hassen Badi B.P 182 El-Harrach, 16200 Algiers, AlgeriaSearch for more papers by this authorAhmed Khaled Chaou, Ahmed Khaled Chaou Laboratoire de Recherche en Electrotechnique, Ecole Nationale Polytechnique, 10 Avenue Hassen Badi B.P 182 El-Harrach, 16200 Algiers, AlgeriaSearch for more papers by this authorDjamal Maadjoudj, Djamal Maadjoudj orcid.org/0000-0001-9263-1990 Laboratoire de Recherche en Electrotechnique, Ecole Nationale Polytechnique, 10 Avenue Hassen Badi B.P 182 El-Harrach, 16200 Algiers, AlgeriaSearch for more papers by this authorMadjid Teguar, Madjid Teguar Laboratoire de Recherche en Electrotechnique, Ecole Nationale Polytechnique, 10 Avenue Hassen Badi B.P 182 El-Harrach, 16200 Algiers, AlgeriaSearch for more papers by this author First published: 01 August 2020 https://doi.org/10.1049/iet-smt.2019.0297Citations: 1AboutSectionsPDF ToolsRequest permissionExport citationAdd to favoritesTrack citation ShareShare Give accessShare full text accessShare full-text accessPlease review our Terms and Conditions of Use and check box below to share full-text version of article.I have read and accept the Wiley Online Library Terms and Conditions of UseShareable LinkUse the link below to share a full-text version of this article with your friends and colleagues. Learn more.Copy URL Share a linkShare onFacebookTwitterLinkedInRedditWechat Abstract This study introduces a novel methodology for electrical discharges recognition elaborating an algorithm based on the RGB colour image model and artificial neural network (ANN) classifier. The developed RGB-ANN algorithm aims to detect and monitor the propagation of electrical discharges until flashover, through analysis of six colours appearing in the discharges images, extracted from the flashover videos recorded on a plan glass insulator model under uniform pollution. First, more than 300 colours images are collected and divided into sets to form a large database. Using an RGB encoding system, each pixel is represented by (R, G, B) coordinates and each image is encoded by 3D matrix. For the discharge image, the coordinates of each pixel are compared to all database ones. The colour of the database having the same coordinates of the discharge image pixel is attributed to this latter. Based on the ratio of the pixels number of a given colour to the total pixels number of the discharge image, six indicators are quantified and grouped to form a feature vector. This latter is used as input of the ANN, in order to classify the evolution of discharges into five classes. As the main result, >98% of images have been well classified. 1 Introduction High voltage (HV) insulators are very useful for electrical transmission and distribution grids. They are intended to perform and operate effectively under the most severe climatic conditions. Consequently, monitoring the performance of these insulators under pollution is of the upmost importance to maintain safe and continuous operation of power on the network [1-6]. If insulators are not correctly monitored especially under severe pollution conditions, the flashover can occur through the following steps: accumulation of contamination layer, wetting of the insulator, increasing of the leakage current, the formation of dry band arcs and finally the extension of such arcs to cover the leakage path [1, 7-9]. To avoid the flashover occurrence and limit the progression of electrical discharges, it is necessary to develop reliable tools to evaluate in real-time the performance of outdoor insulators [10, 11]. Insulator performance monitoring goes through the prediction and forecasting of the contamination severity [3, 7]. This pollution assessment is carried out usually basing on the investigation of the leakage current (LC), which can be represented in time or in the frequency domain [5, 12]. It is noted that for the frequency analysis, the wavelet transform has shown clear superiority compared to the classical Fourier method [13]. Based on the interpretation of the signals of LC, discharges or applied voltage, several studies have been devoted to monitor and/or diagnosis polluted insulators. Chaou et al. [6] also used the wavelets to monitor and diagnose the surface state of the HV insulator model under uniform and non-uniform pollution. For this, continuous wavelet transform (CWT) and discrete wavelet transform (DWT) have been used to detect electrical discharges on the applied voltage signal and the LC one, respectively. The wavelet technique has been shown as an effective tool for the diagnostic and monitoring of polluted insulators performance. Recurrent plot (RP) technique has been developed to analyse LC through the contaminated outdoor insulator [5] and rime-iced composite insulator [11]. In both investigations, the authors decomposed the LC into different frequency components using DWT. It is shown that the topological structure of the high-frequency components is prominent to identify non-linear properties of discharge activities. The obtained results indicate that RP gives visual recurrent patterns of discharge activities for monitoring outdoor insulators performance; it can therefore be considered as an effective technique for studying discharge evolution during the flashover process. However, RP provides only a qualitative overview of the state of the insulation. To overcome this inconvenience, Chaou et al. [2] employed the recurrence quantification analysis (RQA) through eight indicators to quantify and investigate LC waveforms under various pollution conductivities. The mean values of those indicators are used as inputs of three classification methods namely, K-nearest neighbours (KNNs), Naïve Bayes (NBs) and support vector machines (SVMs), to classify the contamination severity into five classes. The obtained results have shown a good correlation between the RQA indicators and the pollution severity level. Dhahbi and Beroual [3] investigated the variation of the LC characteristics in time and frequency domains, for insulator model under different pollution configurations. Such characteristics consist in the peak value, the phase shift, the total harmonic distortion and the harmonic contents. The obtained results portray that the magnitude of the harmonic components can be well correlated with the LC distortions. However, the LC peak value and the phase shift cannot be reliable indications of the discharge activity. Otherwise, the time–frequency analysis can be used as a tool for patterns recognition and classification of the LC. As a non-electrical technique, Yi et al. [4] developed an algorithm exploiting the acoustic emissions (AEs) generated by partial discharge activity for monitoring the polluted insulators. This algorithm combines the empirical mode decomposition (EMD) and fast Fourier transform (FFT) to extract the frequency characteristics of the AE signals. It has been successfully applied to distinguish between corona discharge, partial discharge and arc discharge. The prediction and the classification of pollution level or the flashover risk one using artificial neural network (ANN) remains the subject of many types of research. To monitor LC of post insulators, Mi et al. [14] installed an optical fibre sensor system to measure LC amplitude, LC RMS, LC impulses number and relative humidity. These parameters have been used as inputs of a radial basis function neural network (RBFNN) to predict the contamination level (as output). The predicted results are to be in good agreement with the experimental ones confirming the effectiveness of the RBFNN. Al Khafaf and El-Hag [15] proposed a feed-forward neural network (FFNN) algorithm in order to predict and monitor the change in LC peak on the polluted insulator surface. A good correlation between the change in the average peak of LC and the contamination level on the insulator surface has been obtained. Qaddoumi et al. [16] proposed a back-propagation ANN (BPANN)-based near-field microwave technique to detect and classify different types of faults (air voids, conductive inclusions etc.) in 33 kV no-ceramic insulators. Eight signals of different frequencies have been used to train the network. As results, a recognition rate of 95% for silicon rubber insulators and 97% for fibreglass core ones has been found. Quizhpi-Cuesta et al. [17] interested in classifying discharges into partial discharges or partial breakdown in cap and pin insulators. Discharges have been detected and measured using a digital finite impulse response (FIR) filter. The classification of such discharges is achieved, employing a three-layer back-propagation neural network. Note that the discharge patterns have been generated using a statistical analysis called fingerprints. The proposed procedure allowed reducing the time as well as the cost required to analyse partial discharges. Among the numerous investigations devoted to the polluted insulators monitoring field, just a few number of studies dealing with image processing techniques. For instance, Maraaba et al. [18] elaborated multi-layer FNN (MFNN) to predict the contamination level. To achieve this goal, ten features have been extracted from HV insulator images and then used as inputs of the network. A good correlation between the predicted contamination level and visual images has been found. Chaou et al. [1] are the sole who developed (black and white) image processing algorithm to recognise arcing discharges pattern on the insulator model through the extraction of four indicators from the filtered images. These indicators have been used as inputs of three classification algorithms, namely Knns, NBs, SVMs, to distinguish between the presence or the absence of arc discharges on the insulating surface. SVM classifier has shown the best performance comparing to Knn and NB. The novelty of the present investigation is to recognise and classify electrical discharges pattern propagating on polluted insulating surfaces basing on the colour of the discharge. For this purpose, we have developed a colour image model for discharges recognition and a BPANN for their classification. First, the images of different discharge steps, including luminosities, sparks, brushes and arcs, have been extracted from the flashover videos acquisition. The six colours contained in the extracted images are blue, purple, red, orange, yellow and white. Images are then encoded using the RGB (red, green and blue) encoding system. In this latter, each image pixel is encoded by triple coordinates (R, G, B), and then each numerical image is encoded by 3D matrix. The colour of each pixel is identified using as reference a large and sufficient database. The next step resides in the counting of the pixels number for each colour. Afterwards, we calculate the feature vector elements in which each element corresponds to the ratio of the pixels number of one colour to the total pixels number in an image. Constituted by six indicators (one indicator per colour), the feature vector is used as input of the elaborated artificial neural network. This latter allows classifying the discharges images into five classes. 2 Experimental setup Experiments have been carried out in the High Voltage Laboratory of ENP (Ecole Nationale Polytechnique) on a plan insulator model under uniform pollution. The experimental setup consists in a high-voltage test transformer (300 kV/50 kVA, 50 Hz), supplied by a regulator voltage (autotransformer) (220/500 V, 50 kVA, 50 Hz), a capacitive divider (with 1000:1 ratio) and a glass insulator plan model (Fig. 1). Two rectangular electrodes (50 × 3 cm2 × 2 µm), made up of aluminium, have been used. The distance between them is 29.2 cm. It corresponds to the leakage path of the 1512L cap and pin outdoor insulator used mainly by the Algerian Company of Gas and Electric Power (SONELGAZ). The insulator model is placed on wooden support at 100 cm height from the ground (Fig. 2). Installed at 100 cm height from the insulator model, a video camera (Full HD_20 Megapixels) is used to record the discharges evolution until flashover. The data have been processed through a PC. Fig 1Open in figure viewerPowerPoint Experimental set-up H.V.T: high voltage transformer, R.T: regulating transformer, I.T: isolating transformer, V.R: voltage regulator, I.M: insulator model, P.C: personal computer Fig 2Open in figure viewerPowerPoint Disposition of insulator model and video camera Before each test, the insulating surface is first washed with tap water and dried. It is cleaned, then after, with isopropyl alcohol to eliminate any pollution traces. The insulator model is submitted to uniform pollution. Two different types of pollution have been adopted. The first one consists of a saline solution (distilled water and NaCl). This later has been sprayed five times at a distance of 50 cm from each side of the insulator model as shown in Fig. 3a. Five conductivities have been selected (0.003, 0.03, 0.3, 3 and 10 mS/cm). The sand, collected from Naama city (southern Algeria) at 20 m high, has been used as second pollution agent. Four amounts, namely 15, 30, 45 and 60 g, have been considered. The corresponding NSDD values are 0.01, 0.02, 0.03 and 0.04 g/cm2, respectively. NSDD values are calculated by IEC 60815 [19] as follows: (1) where A is the polluted insulator model surface (=29.2 × 50 cm2); WS is the weight of sand amount (g). Fig 3Open in figure viewerPowerPoint Pollution agent (a) Saline solution, (b) Sand and distilled water Each sand layer has been moistened with distilled water (of 2 μS/cm conductivity) five times at the same distance 50 cm from each insulator model side as illustrated in Fig. 3b. 3 Discharge flashover process In this section, we are interested in investigating the evolution of discharges from their initiation to final flashover. Figs. 4-6 represent an example of the extracted images from video recorded when using moistened sand pollution of 0.02 g/cm2 NSDD. For this, the applied voltage is increased (with ∼2 kV/s) from zero to flashover voltage. Fig 4Open in figure viewerPowerPoint Luminosities and brushes (a) 53 kV, (b) 55 kV, (c) 60 kV Fig 5Open in figure viewerPowerPoint Arcs progression (a) 63 kV, (b) 65 kV,(c) 66 kV,(d) 67 kV Fig 6Open in figure viewerPowerPoint Flashover arc (a) 67.5 kV, (b) 68 kV The first luminosities begin to appear at both electrodes sides of the insulator model at 25 kV. Such luminosities become remarkable at ∼50 kV and evolve into scattered bright points at 53 kV (Fig. 4a). Increasing further the applied voltage at 55 kV, these points start to merge to form weak sparks (Fig. 4b). These latter take the form of discharges brushes at 60 kV (Fig. 4c). Until now, the purple colour is dominant. The arc structure consisting of small red/orange/yellow partial arcs starts to appear from 63 kV (Fig. 5a). It is worth noting that arcs developed so far are not localised. From 65 kV, we have observed a new phase characterising the discharges evolution process. During this phase, some arcs began to disappear. On the other hand, two localised arcs emerge at both electrodes sides. The earth arc is longer than the HV one. The remaining arcs are denser, lighter and brighter (Fig. 5b). From 66 kV, the two localised arcs increase in length and the number of the other arcs slightly decreases (Fig. 5c). From 67 kV, the localised arcs, which become thicker, denser and longer, form the main arcs. Their colour becomes to be red/orange. The number of other arcs still decreasing (Fig. 5d). Until now, the insulating surface, not short-circuited by the partial arcs, shows black colour. From 67.5 kV, the flashover occurs following the contact of the two localised arcs as shown in Figs. 6a and b corresponding respectively to the first and the final steps of the flashover arc. The formed (resultant) arc possesses a white colour. Its thickness reduces from the first to the final flashover steps. The distribution of colours that appears when moving laterally from the arc to the top and or bottom of the model side, changes from light purple to black passing through dark purple. Such a situation does not remain for a long time since the yellow, orange and red colours (Fig. 6b) replace the light rapidly to dark purple (Fig. 6a). Figs. 4-6 show that the discharges progress into eight steps as summarised in Table 1. Such steps have been inspired by other investigations in the field [1, 7]. Table 1. Typical discharge evolution until flashover Step Typical discharge phenomena 1 no obvious arc discharge at 14.2 s (Fig. 4a) 2 weak spark at 15.5 s (Fig. 4b) 3 discharge in the shape of brushes at 15.9 s (Fig. 4c) 4 short local arc discharge at 16.3 s (Fig. 5a) 5 dense small arc discharge at 17 s (Fig. 5b) 6 bright main arc discharge 17.2 s (Fig. 5c) 7 intensive main arc discharge at 17.8 s (Fig. 5d) 8 first and final arc flashover stages at 18 s (Figs. 6a and b) The same discharges flashover process previously described has been observed for the all NSDD (moistened sand) and conductivities (saline solution) values. The relationship between the values of flashover voltage and NSDD ones are illustrated in Table 2. Table 2. Flashover voltage versus NSDD NSDD, g/cm2 0.01 0.02 0.03 0.04 flashover voltage, kV 82 68 59 56 We observe that the flashover voltage decreases abruptly when NSSD increases from 0.01 to 0.03 g/cm2, and slowly elsewhere. More details about the behaviour of the same insulator model under uniform and non-uniform desert pollution (humidified sand) are presented in previous work [20] recently published by our research team. The sand has been collected from Naama region (South-West Algeria) on a dune of about 20 m of height. A large board of experimental investigations has been carried out by analysing the flashover process and the leakage current versus NSDD and layer widths (in non-uniform pollution configuration). Furthermore, a wavelet transform technique has been used to study the frequency-time characteristics and to calculate the standard deviation (STD) of different harmonic components of leakage current. The extracted components have been used as indicators of the pollution level. The same ascertainment is made when applying the saline solution on the experimental model in terms of flashover voltage values, since we have obtained 95, 95, 73, 48 and 37 kV for 0.003, 0.03, 0.3, 3 and 10 mS/cm, respectively. 4 RGB encoding system-based image processing method 4.1 RGB encoding system A digital image is composed of a limited number of elements, called pixels (px). Each pixel contains only one colour. The RGB representation is the most often used in image processing [21, 22]. In RGB, each colour (pixel) is encoded by triple Cartesian coordinates of primary colours, namely red, green and blue [23]. Inspired from other investigations [24, 25], RGB has been represented, in our work, using a colour cube, where the black is at the origin point (0,0,0) and the white is at the farthest corner from the origin (255,255,255) as depicted in Fig. 7. As each pixel is characterised by triple coordinates (R, G, B) belonging to [0, 255], each image should be represented by a 3D matrix also called matrix in three plans (i.e. the first plan corresponds to R, the second to G and the third to B). Fig 7Open in figure viewerPowerPoint RGB colour cube In the RGB encoding system, any colour reconstitution is carried out combining, with specific percentages, the three primary colours (red, green and blue). This combination process is called 'additive synthesis'. For instance, at the same percentage equal to 100% (respectively 0%), the sum of the three primary colours gives the white (respectively the black). 4.2 Database creation and RGB validation Recognition of discharges image has been carried out through the identification of its colours pixels. The different colours recorded during the flashover process are blue, purple, red, orange, yellow and white. Except, the white colour, the other possess various intensity degrees from the lightest to the darkest. More than 300 images containing colours possessing different intensity degrees of the previous colours have been collected from various web sites (especially, allrgb.com). All collected images of the same colour (regardless of their intensity degrees) are pasted together to form only one image of only one colour. Fig. 8 presents an example of a collage of some images in blue colour. Fig 8Open in figure viewerPowerPoint Collection and collage of images of blue colour The previous procedure allows obtaining one image by colour. Each image (colour) is encoded by the 3D matrix in the database. Table 3 gives the pixels number used in each colour of the collected database. The white colour does not need to be represented by an image in the database since it does not possess intensity degrees. Table 3. Pixels number for each colour Colour Blue Purple Red Orange Yellow pixel number 79,661 179,073 83,917 135,579 22,029 Note that our RGB encoding system has been successfully tested on real predefined images. An illustrative example is shown in Fig. 9 in which the black area indicates the pixels detected for a given colour. Fig 9Open in figure viewerPowerPoint Colours detection 4.3 Image processing method Each colour (characterising by different intensity degrees) of the previous database has been introduced in the elaborated algorithm to be represented by the 3D matrix. In fact, the collected database allows our algorithm to identify, from the images discharges, the colour of each pixel by comparing its coordinates (R, G, B) to all database ones. Basing on the equality decision, the colour of the database having the same coordinates of the discharge image pixel is attributed to this latter. It is worth noting that our database is sufficient to be used because it widely covers all combinations of (R, G, B) coordinates of the different extracted images of electrical discharges from flashover video. 5 ANN architecture An ANN is a system whose design is inspired by the operation of biological neurons. ANNs generally have a number of highly interconnected neurons organised into layers [26]. These latter are constituted by nodes where computation happens by combining inputs from the data with a set of coefficients, or weights. These input-weight products are summed. The sum passes through the activation function. Starting by introducing the data in an initial input layer, the output of each layer is simultaneously the input of the next one, and so on. Among several structures and architectures of ANNs, the multi-layer perceptron neural network (MLPNN) has been used in the present investigation, because it is one of the most reported and the most commonly used considering its simplicity and classification performance [27, 28]. This network has been trained using the back-propagation algorithm. The calculation has been carried out with feed-forward algorithm. The parameters of the MLPNN model are the weights and biases. The training is used to adjust the weights and biases. The MLPNN is trained with random initial parameters. The training process is repeated several times, in order to determine the best model having the highest classification precision. Two hidden layers have been used. 12 neurons have been considered for the first layer and 5 for the second. The input layer contains the independent variables consisting in B, P, R, O, Y and W corresponding, respectively, to the ratios (in %) of the colours blue, purple, red, orange, yellow and white. The output layer consists of five classes related to the flashover discharges steps. The structure of the developed MLPNN is shown in Fig. 10. Fig 10Open in figure viewerPowerPoint MLPNN structure The ANN was implemented under (9.3.0 version) MATLAB environment using neural networks pattern recognition tool 'nprtool'. This latter uses the 'sigmoid' (respectively, the 'softmax') as the activation function for the first (respectively, the second) hidden layer. The 'sigmoid' (respectively, the 'softmax') has been chosen because it is commonly used in non-linear problems [27] (respectively, in input/output vectors [29]). 6 RGB-ANN arcing discharges recognition algorithm This section is dedicated to explain the different steps of the proposed algorithm for arcing discharges recognition. This algorithm aims to detect the different discharges propagation stages on polluted insulator model and classify them into five classes. The execution of this algorithm is carried out in two steps. The first one consists in extracting six indicators for each flashover image in order to construct the corresponding feature vector. Next, the calculated feature vectors are used as inputs to train the proposed MLPNN. The flowchart of the arcing discharges recognition algorithm is presented in Fig. 11. Fig 11Open in figure viewerPowerPoint Flowchart of the arcing discharge recognition algorithm 6.1 Feature extraction The feature extraction consists of the calculation of the characteristic vector of each image. This vector describes discharges properties and represents the input of the classification method. The feature vector contains six indicators, namely B, P, R, O, Y and W corresponding to the ratio (in %) of the pixels number of the blue, the purple, the red, the orange, the yellow and the white, respectively, to the total pixels number of the discharges image. This total number which depends only on the capture emplacement and the resolution of the used video camera is 4 Megapixels. 6.2 Discharge classification To discern between the flashover steps, we have elaborated an ANN classification tool. This latter aims to differentiate between five relevant discharge patterns classes illustrated in Table 4. Indeed, for every image, a class from five classes is carefully attributed. If the image shows only luminosities, it is classified as no arc discharge class, this is the first class. If the image contains short arcing discharges, it is classified in the second class. The third class contains images with dense small or bright main arc discharge. If the image contains any intensive red main arc discharge, it is assigned to the fourth class. The last class concerns the flashover arc images. Table 4. Classification of typical discharges Typical discharge phenomena Class no obvious blue/purple arc discharge 1 weak blue/purple spark purple discharge in the shape of brushes short red/orange/yellow local arc discharge 2 dense red/orange/yellow small arc discharge 3 bright red/orange/yellow main arc discharge intensive red/orange/yellow main arc discharge 4 flashover white arc 5 As illustrated in Table 5, 695 discharges images have been extracted from flashover video (357 when applying saline solution pollution on the insulator model and 338 using sand moistened with distilled water). 591 images (i.e. 85% of the total number) have been dedicated to the training process and the rest (104 images) for the test. Table 5. Image classification database Number of images training 591 test 104 total 695 (=357 + 338) 7 Results This section exposes the classification results of the proposed pattern recognition algorithm, which classifies the discharge type, basing on a proposed image processing method. 7.1 Feature vector Extracted from video recorded when applying moistened sand of 0.02 g/cm2 NSDD on the experimental insulator model, the discharges images carefully analysed with regard to the applied voltage in Section 3 allow obtaining the feature vector containing the six indicators (B, P, W, R, Y and O in %) whose variations are presented in Fig. 10. In this latter, labels or classes (from 1 to 5) of discharge pattern have been added at specific values of the applied voltage. For more precision, the different classes shown in Fig. 12 are summarised according to the applied voltage in Table 6. Fig 12Open in figure viewerPowerPoint Feature vector indicators during the flashover process for 0.02 g/cm2 NSDD Table 6. Classification of evolution typical discharges according to the applied voltage Voltage, kV Typical discharge phenomenon Class <62 absence of arc discharges on the insulating surface 1 63 short local arc discharge 2 [65–66] dense small arc discharge 3 bright main arc discharge 67 intensive main arc disc