Classification of common discharges in outdoor insulation using acoustic signals and artificial neural network

: Condition monitoring of outdoor insulation systems is crucial to the integrity of distribution and transmission overhead lines and substations. The objective of this study is to use a commercial acoustic sensor along with artificial neural network (ANN), to classify different typical types of discharges in outdoor insulation systems. First, ANN was used to distinguish between five common electrical discharges that were generated under controlled conditions. Next, this approach was extended to include outdoor ceramic insulators. Three types of defects were tested under laboratory conditions, i.e. a crack in the ceramic disc, surface pollution discharge, and corona near the insulator surface. Both a single disc, and three discs connected in an insulator string were tested with respect to these defects. For both controlled samples and full insulators, a recognition rate of more than 85% was achieved.


Introduction
Overhead lines supported by outdoor insulators have been used by utilities at both the transmission and distribution voltage levels. Outdoor insulators play a critical role as they insulate the high voltage power line from the grounded tower. However, due to pollution accumulation, outdoor insulators may suffer from flashover leading to power interruption. These flashovers may lead to sustained faults which will result in power system locks-out as a means of preventing damage to the overhead assets. This type of disruption can be very costly; as the insulators need to be either cleaned of the contaminant or be replaced because of physical damage leading to the system shut down for hours or days. Therefore, monitoring of outdoor insulators is crucial to the integrity of the power grid.
Both ceramic and non-ceramic insulators have been used as outdoor insulators. Non-ceramic insulators show several merits over ceramic insulators, like lightweight and their excellent pollution performance [1]. Hence, most of the newly installed insulators in North America are non-ceramic insulators. However, ceramic insulators are older than non-ceramic insulators and nearly 150 million porcelain suspension insulators are estimated to be currently deployed in the North American grid [2]. A significant number of these ceramic insulators suffer from different types of defects, approaching their end of life and hence need to be identified to avoid complete line failure.
In ceramic insulators, partial discharge (PD) can be initiated because of electric field enhancement due to surface pollution, internal crack, and/or hardware accessories. Moreover, when a water film builds on the insulator surface, this will lead to leakage current (LC) development. Since the LC density is not homogeneous along the insulator surface, in some areas enough heat develops to evaporate the water layer, forming dry bands; resulting in arcing across these dry areas called dry-band arcing (DBA).
When PD and DBA occur, several physical phenomena occur like acoustic emission (AE), electromagnetic (EM) radiation, light emission, or LC signals. Hence, different transducers can be utilised to monitor PD and DBA. Most of the published research focuses either on the measurement of PD, or DBA; but usually, there is a lack of comprehensive approach, which can be used to detect both types of discharge activities.
For PD monitoring, several techniques have been investigated. High-frequency current transformer has been used to detect PD activities resulted from artificially polluted outdoor insulators [3]. It has been found that the signature of the captured PD signal depends significantly on the pollution severity level. Moreover, PD for both ceramic and non-ceramic insulators was measured using radio frequency (RF) antenna in laboratory conditions [4]. It has been reported that there were clear differences in the PD spectra associated with non-ceramic and ceramic samples [4]. Furthermore, an RF antenna was successfully used to detect PD in ceramic insulators and identify damaged ones without specifying the damage type [5].
RF antenna along with machine learning (ML) were used to detect and identify different defects in ceramic insulators [6]. Recognition rate of more than 95% was reported. The main short coming of the proposed system is the relatively high cost of the RF data acquisition system. A fibre optic-based sensor to detect PD has been employed on 500 kV string insulators [7]. The sensor used has the advantage of immunity to EM radiation, but the overall system is complex to implement. DBA, however, has been mainly monitored directly using current measurements. LC can be either measured as a voltage drop across a shunt resistor or through magnetic coupling using a current transformer. It has been found that the level of lowfrequency harmonics in DBA, is highly correlated to both the degree of insulator surface damage [8] and the likelihood of flashover occurrence [9].
Since PD is a high-frequency phenomenon (MHz -GHz range), and DBA is a low-frequency phenomenon (100's of Hz), most of the existing techniques cannot be used to simultaneously measure both PD and DBA.
Alternatively, ultrasound AE sensors have been used to measure both PD and DBA [10,11]. Measuring both PD and DBA simultaneously is a cost-effective option for power utility companies. However, the major disadvantage of this approach is the difficulty to identify the source of the measured signal, i.e. PD or DBA. ML has been implemented to identify different types of defects on controlled polymer samples [12,13]. Moreover, ML along with image processing has been used to identify physical damage in ceramic insulators [14,15]. However, image-based techniques can only detect visible damages and cannot detect internal defects.
From the aforementioned discussions, it can be stated that there is a lack of a comprehensive and cost-effective condition monitoring technique that can be used to identify different types of defects in ceramic insulators. The objective of this paper is employing AE techniques to assess the condition of both controlled samples and line insulators. The research investigates the possibility to distinguish between contamination and physical damages of line insulators by measuring the AE resulted from electrical discharges. Fig. 1 shows the schematic diagram for the overall setup used to generate and measure different types of PDs. Five common types of PDs are considered for the multi-class classification problem as shown in Fig. 2. The test was extended to study the possibility to detect different defects in practical ceramic insulators. These defects include cracks (Fig. 3a), hardware corona (Fig. 3b), and wet surface discharge due to surface contamination. A 150 kV/20 kVA transformer with PD level less than 2 pC is used to generate the required high voltage to initiate PD in the test samples. Commercially available acoustic sensor (MK-720) with 40 kHz as the centre frequency with a frequency range of 16-80 kHz is used to record the PD AEs. A classical PD detector is employed to measure simultaneously PD along with the acoustic sensor. The   The experimental setup used to investigate the effect of the captured acoustic signal sensitivity as a result of a change in distance and angle of measurement on the classification of PD sources is shown in Fig. 4. The distance D represents the distance between the acoustic sensor and PD source, H is the vertical distance from the floor level to the PD source and h is the vertical distance from floor level to the acoustic sensor.

Application of ANN
The MK-720 acoustic sensor is provided with internal signal conditioning circuits. First, the output of the acoustic sensor run through a high-pass filter to remove the noise, following which, the outer envelope of the filtered signal was detected prior to the final step, which was the application of a fast Fourier transform (FFT). The signal conditioning stages are identified in Fig. 5.
A typical measured acoustic signal and its FFT from a sharp electrode are shown in Fig. 6.
The intensity of the AE originated from a PD source varies periodically with the alternating voltage. The acoustic sensor uses the amplitude modulation technique where a carrier wave with a particular frequency is changed according to the intensity of the audio signal. Here, the sensor has inbuilt technology to use a carrier waveform frequency same as the supply voltage frequency, which means that the amplitude spikes of the processed acoustic signal dominate at the fundamental frequency of the supply voltage and its integral multiples. Hence, the acoustic signal components at 60, 120 and 180 Hz were used as the input feature vector for the ANN.
The structure of the two implemented ANNs is presented in Fig. 7. The first ANN (Fig. 7a) was used to distinguish between five different controlled PD sources. The input feature vector was the 60, 120 and 180 Hz frequency components of the acquired PD signal envelope. A total of 450 test data have been collected from the five different PD sources. The second ANN (Fig. 7b) was implemented to classify three different defects in ceramic insulators and used three input feature vector (same as the first ANN). 150 input data sets from the three defects of line insulators were collected. The data for both neural networks were randomly mixed and 70% of the total data collected were used for training, 15% for validation, and 15% for testing of developed ANN classifier. This random mixing of data was repeated for 5 times and the average recognition rate was reported in this study.

Classification of controlled PD sources
The average classification accuracy obtained from the ANN classifier for the five different controlled discharges is depicted in Table 1.
It is apparent from Table 1 that a relatively high overall classification rate was achieved with an average recognition rate of 92% for five different trials. The relatively high classification rate is due to the nature of the measured PD acoustic signal envelopes. Examples of two envelopes for classes I and V are shown in Fig. 8. Each envelope has a distinct signature. The same was observed for other PD signal envelopes. For example, it has been observed that the amplitude and repetition rate of the envelope spikes are relatively high for the discharge from sharp electrodes compared to those from a smooth electrode. Differences in the acoustic signal envelope signatures are due to the different types of streamers formed on the dielectric surface and/or around the high voltage electrode.
Moreover, the 3D plots for the 60, 120, and 180 Hz for the five different discharges envelopes are shown in Fig. 9. It is evident from Fig. 9 that the five different discharge types are separable with some overlap between the classes which explains the high recognition rate.

Effect of distance on the classification accuracy
To quantify the detection sensitivity of the acoustic sensor, measurements are made with corona and surface discharge at a distance of 150, 200, 250 and 300 cm between the PD source and the acoustic sensor. Fig. 10 depicts the effect of the distance on the magnitude of the 60 Hz component of the corona signal. It is apparent that a decreasing trend is evident and the same trend was noticed for both the 120 and 180 Hz. It is worth mentioning that increasing the distance did not result in overlap between the frequency components of the same class, but caused overlap between different classes that resulted in reduction in the classification accuracy as described below.
To investigate the influence of distance on the classification accuracy, two classifiers were tested, one trained with only the data at 150 cm (ANN 1) and the other was trained with data from all distances (ANN 2). Table 2 summarises the recognition rates obtained from the two classifiers.   Both classifiers showed a decreasing trend in their accuracies, however, ANN 2 showed consistent higher accuracy. Nevertheless, ANN 1 still provides accuracy higher than 80% when tested at distances up to 250 cm. So, to achieve a robust classifier, it is paramount to train the classifier at different distances.

Effect of measurement angle on the classification accuracy
The attenuation in the signal strength as a result of the change in the angle of measurement is studied by placing the acoustic sensor at different angles (5°, 20° and 30°) as defined in Fig. 4. The results reveal that the change in the angle has minimum effect on the 60 Hz component as depicted in Fig. 11. Similar behaviour was observed for the 120 and 180 Hz components.
Subsequently, the influence on the recognition accuracy as a result of changing the angle between the acoustic sensor and the PD source is investigated. An ANN classifier was trained using data measured at a deviation angle of 0° from the line of sight and then was tested using data recorded at 5°, 20°, and 30° deviation angles. The results are shown in Table 3. Unlike the influence of distance on the classification accuracy, it is evident that changing the angle does not impact the classification accuracy. This can be attributed to the fact that changing the angle does not influence the strength of the received signal by the acoustic sensor as evident in Fig. 11.

Classification of defects of ceramic insulators
PD measurement using acoustic sensor along with ANN have been employed to identify three types of defects in ceramic insulators. The defects were corona (class I), crack (class II) and surface discharge (class III). Table 4 shows the average recognition rate obtained from five sets of testing data conducted on the three classes of single and string insulators, respectively.
It is evident from Table 4 that perfect recognition of corona discharge has been achieved. On the other hand, a recognition rate of more than 80% was achieved for the other two classes. This difference in the recognition rate can be explained by showing the 3D plot of the input feature vector for both the single disc and the insulator string, Figs. 12 and 13, respectively. It is evident that the

Analysis with varied defective insulator disc position
Moving the position of the cracked insulator to the middle of the three-disc string insulator resulted in different recognition rates as depicted in Table 5. The average recognition rate of corona (class I) dropped from 100 to 80.66% and the recognition rate of surface discharge (class III) increased from 80.02 to 96.58%. This change in the recognition rate can be explained by showing the 3D plot of the input feature vector as depicted in Fig. 14.
It is apparent that the data cluster for the crack insulator class moved upward and mixed with the corona data cluster and away from the surface discharge data cluster resulted in lowering the recognition rate of the corona class and increased recognition rate for the surface discharge. This observable shift in the 3D pattern is attributable to the fact that the voltage levels across each disc insulator in the string vary according to the position of the disc. Hence moving the cracked insulator upward resulted in different level of discharge and hence acoustic level.

Conclusion
Based on the aforementioned investigations, the following conclusions can be drawn: • The acoustic signal components at 60, 120 and 180 Hz were used as the input feature vector of the ANN to classify controlled PD sources. The trained classifier has been proven effective for the classification of each specific type of controlled PD, with a recognition rate of more than 90%. Furthermore, it has been found that increasing the linear distance from the PD source resulted in decreasing the recognition rate. On the other hand, no impact on the recognition rate has been found from changing the angular distance. • The study was extended to include the classification of defects and contamination conditions in both string and individual ceramic insulators. In this regard, the classifier produced a recognition rate of more than 85%. • Varying the position of a cracked insulator disc within a string has only a very small effect on the classification accuracy.
To further enhance the classification accuracy of the proposed system, larger database is collected, and the authors are intending to utilise deep learning.