Partial Discharge Pattern Recognition of Transformers Based on MobileNets Convolutional Neural Network

<span>Partial discharge (PD) pattern recognition is useful to diagnose insulation condition. PD measurement data is commonly represented in phase-resolved partial discharge (PRPD) format. PRPD is useful as it provides a visible pattern for different PD source and various features can be extracted for PD pattern recognition. Shorter PRPD duration will enable more training data but the information in each data is less and vice versa. This works aims to investigate the effects of using very short duration PRPD data on the accuracy of PD pattern recognition. The results conclude that machine learning models such as Artificial Neural Network (ANN) and Support Vector Machine (SVM) are robust enough such that reduction of PRPD duration from 15-seconds to 1-second causes less than 5 % drop in the classification accuracy. However, this is only true for noise free condition. <span>When the same PD data is overlapped with random noise, the classification accuracy suffers a significant reduction up to 19%. Therefore, longer PRPD duration is recommended to withstand the effects of noise contamination.</span></span>

Different types of partial discharge (PD) cause different damage to gas-insulated switchgear (GIS), so it is very important to correctly identify the type of PD for evaluating the GIS insulation condition. The traditional PD pattern recognition algorithm has the limitations of low recognition accuracy and slow recognition speed in engineering applications. This work proposed a Convolutional Block Attention Module Residual Network (CBAM-ResNet) for the GIS PD pattern recognition. In particular, the Squeeze-and-Excitation mechanism was used to improve the recognition accuracy. The GIS phase resolved partial discharge (PRPD) image is used as the input of the pattern recognition model. The results show that the recognition accuracy of the proposed method is as high as 89%, and has a recognition accuracy of 92.72%, 90.19%, 89.36%, and 93.62% for the four PD types, respectively. Compared with normal CNN and SVM recognition algorithms, the PD pattern recognition accuracy of the proposed method has been greatly improved. And the CBAM-ResNet for PD pattern recognition has a good application prospect in engineering practice.

In view of the fact that the statistical feature quantity of traditional partial discharge (PD) pattern recognition relies on expert experience and lacks certain generalization, this paper develops PD pattern recognition based on the convolutional neural network (cnn) and long-term short-term memory network (lstm). Firstly, we constructed the cnn-lstm PD pattern recognition model, which combines the advantages of cnn in mining local spatial information of the PD spectrum and the advantages of lstm in mining the PD spectrum time series feature information. Then, the transformer PD UHF (Ultra High Frequency) experiment was carried out. The performance of the constructed cnn-lstm pattern recognition network was tested by using different types of typical PD spectrums. Experimental results show that: (1) for the floating potential defects, the recognition rates of cnn-lstm and cnn are both 100%; (2) cnn-lstm has better recognition ability than cnn for metal protrusion defects, oil paper void defects, and surface discharge defects; and (3) cnn-lstm has better overall recognition accuracy than cnn and lstm.

This paper discusses the results of an on-going project on the application of artificial intelligence techniques for recognition of partial discharge (PD) patterns from hydrogenerators. The main purpose of the project is to classify data according to the origin of a PD source and create an early warning system for stator winding insulation. More than nine hundreds PD patterns recorded during last two years on more than thirty generators were analysed. The patterns in the form of the maximum and mean pulse height and pulse count discharge distributions were described by statistical parameters such as the skewness, kurtosis, etc., and stored in a data base as a fingerprint. Classification of unknown discharge patterns was carried out by statistical pattern recognition techniques. The results indicate that satisfactory classifications took place.

Widely used method for PD (partial discharge) measurement and analysis of existing underground power transmission cables is to store the detected signal pulses as PRPD (Phase Resolved Partial Discharge) data and to analyze the PD defect factors by the pattern type. Although the PRPD pattern data of various facility defects are acquired through the underground power transmission cable diagnosis, due to the PRPD data storage structure, the pulse generation information is destroyed over time, and reconversion into 3D PD pattern (PRPS: Phase Resolved Pulse Sequence) is impossible. Also recently, as a method for verifying the reliability of the power utility prevention diagnosis system, a simulated partial discharge is generated for each facility defect, and the diagnostic system verification that is capable of simultaneously measuring pulses and PD patterns using various sensors is performed, and it is necessary to develop a PD pulse generator based on simulated PRPS data in order to generate the pulses based on PRPS pattern using the existing PRPD pattern, and input the pulses into the system to perform the defect judgment performance test. To solve this problem, this study developed and verified a PD pattern generator that can generate pulses based on PRPS data, and an algorithm that converts existing PRPD pattern data currently owned by Korea Electric Power Corporation (KEPCO) to PRPS pattern data.

Partial Discharge (PD) characteristics of typical defects in oil-paper insulation under AC, DC and AC-DC combined voltages have been studied extensively. However, there is still limited material available for the pattern recognition of partial discharge under actual stress withstood by valve side windings in converter transformers. In this paper, the PD of needle-plane defect is measured under actual stress in converter transformers. The stress is DC voltage superimposed harmonics, which has been analyzed in PSCAD at first. Then DC voltage superimposed both single harmonic and multiple harmonics, including simulated actual stress and approximate square waveform, have been applied on electrode. PD parameters and patterns are acquired by detector using pulse current method. The test results indicate that the inception voltage and pulse repetition rate of PD have no direct relationship with harmonic order due to the accumulation of space charges injected by DC voltage. PD amplitude shows a growing trend and phase distributions of PD expand with the increase of harmonic order. PD patterns are also closely related to the waveform of applied voltage. PD amplitude rises with the increase of voltage change rate, and the majority of PD pulses occur in the phase range which has a higher voltage change rate. Experimental results show that harmonic order exerts a great effect on PD characteristics of needle-plane defect under DC superimposed harmonic waveforms. The findings of the current study are useful in further research on the pattern recognition of PD and diagnose in converter transformers.

This paper presents an investigation on the off-line Partial Discharge (PD) and localization for 1 unit of in-service 5 MVA 33/6.9 kV oil immersed transformer. First, Dissolved Gas Analysis (DGA) was performed to confirm the existence of PD activities. Next, offline PD measurements were carried out on both LV and HV windings. Once the transformer has been energized, PD localization was performed at different sections of the transformer tank. It is found that HV windings are having possible multiples PD activities based on offline Phase-Resolved Partial Discharge (PRPD) pattern and PD localization.

In this paper, application of self organizing map (SOM) to recognize partial discharge (PD) patterns of molded type transformers (MTTs) is proposed. The PD patterns are measured by using a commercial PD detector. A set of features, used as operators, for each PD pattern are extracted through statistical tools. The proposed SOM classifier has the advantage of high robustness to the ambiguous patterns, which is useful in recognizing the PD patterns of electrical transformers. To verify the effectiveness of the proposed method, the classifier was tested on 120 sets of field-test PD patterns of MTTs. The test results show the proposed approach may achieve a quite satisfactory recognition of PD patterns.

Partial discharge (PD) diagnosis is confirmed to be one of the most effective tools for assessing the health condition of power equipment. Classification and recognition of the measured PD data provide the insulation defects information which facilitate the condition diagnosis of electrical apparatus. This paper presents a new group of image features for partial discharge classification, where the gray images are formed to represent different PD defects. The PD gray images are decomposed into various vectors by two-dimensional linear discriminant analysis (2DLDA), where 9 representative parameters are extracted from each image vector. Finally, fuzzy k-nearest neighbor classifier (FkNNC), multi-class support vector machine (MC-SVM) and back-propagation neural network (BPNN) are respectively employed for PD classification. 419 diversified samples of PD data acquired from typically artificial defect models of oil/pressboard insulation, where the defect size, applied voltage and insulation aging are taken into account, are employed for algorithm validation. The recognition results of 419 PD samples show that the defects are well identified by the proposed 2DLDA features with high accuracy. In addition, the significant increments of average recognition accuracies are obtained by different classifiers compared with the phase-resolved partial discharge (PRPD) features in previous works. The obtained results indicate that the proposed 2DLDA features are potentially effective and reliable in recognizing different PD sources and may be considered as an improved PD recognition tool when compared with the intensively used PRPD features.

This study is focused on solution to the problem that may occur due to noise signals when using the image obtained through the conventional partial discharge preprocessing methods as inputs to the CNN-based partial discharge pattern classifier. To solve such problem, a new data preprocessing method is proposed by considering a projection technique. In the data information obtained through the conventional partial discharge preprocessing methods, data information of useless noise signal leads to the problem of performance degradation when constructing highly qualified pattern classifier based on the data information of partial discharge signal. The proposed preprocessing method called as ‘Projection’ in this study is designed to solve this problem and improve a classification performance of CNN-based partial discharge pattern classifier. First of all, through GIS simulation, one-dimensional partial discharge data is obtained as five cases such as corona discharge, floating discharge, insulator discharge, free particle discharge, and steady state (noise) in an environment with a noise signal by using a UHF sensor. After that, by applying the two conventional partial discharge preprocessing methods such as PRPS(Phase Resolved Pulse Sequence), PRPD(Phase Resolved Partial Discharge) and the proposed partial discharge preprocessing method, one-dimensional partial discharge data is transformed into 3 data types such as image set of PRPS, image set of PRPD, and image set of Projection. Finally each image set is used as inputs to the designed CNN-based partial discharge pattern classifier. Through the comparative analysis of the feature maps of each layer in CNN as well as the performance accuracy of partial discharge pattern classification for each image set, the superiority of the proposed preprocessing method is demonstrated.

In order to improve the recognition accuracy of partial discharge (PD) by making full use of the time-frequency characteristics of PD signals and employing deep learning theory, a kind of PD pattern recognition method based on variational mode decomposition (VMD)-Choi-Williams distribution (CWD) spectrum and optimized convolutional neural network (CNN) with cross-layer feature fusion is proposed in this paper. Firstly, a PD signal is decomposed into several components by VMD algorithm, and the CWD analysis of the obtained components is carried out to obtain the VMD-CWD time-frequency spectrum. Secondly, the cross-layer feature fusion and optimization CNN (CFFO-CNN) is constructed by introducing cross-layer connection and optimization algorithm. Thirdly, the VMD-CWD is regarded as the input vector to train CFFO-CNN to learn and extract the intrinsic features of the spectrum. Finally, the trained network is used to recognize the PD types of the testing samples. The proposed method is compared with traditional recognition methods such as BP neural network (BPNN) and support vector machine (SVM), as well as some commonly used deep learning algorithms. The experimental results indicate that the recognition performance of the proposed method is significantly better than that of existing recognition methods with accuracy up to 99.5%. It is proved that CFFO-CNN has superior feature extraction ability, which can extract the internal features of the VMD-CWD spectrum independently with higher recognition accuracy and wider application prospect.

This paper proposes a practical partial discharge (PD) severity analysis based on three factors as repetition rate, amplitude and gap time between positive and negative pattern on a phase resolved partial discharge (PRPD) pattern. The PDs are measured by using conventional and unconventional method. The PD amplitude represents PD size. Repetition rate is an average number of PD pulses per second. Gap time is a time distance between two clusters on positive and negative sides in PRPD pattern. Seven 6.6 kV motors were investigated by analysing three parameters, the results show that seven motors encountered high, moderate and low PD severity according to amplitude, repetition rate and gap time. In the first motor, the PRPD pattern shows slot discharge with a high repetition rate and a narrow gap time. This indicates the high risk to motor failure within a short period. The slot discharge was found at stator conductor through ground wall insulation to stator core. In the second motor, the PRPD pattern shows the slot discharge with high repetition rate with the gap time about 5.88 ms. Subsequently, a conventional method according to IEC60270 was performed. The problem was confirmed by visual inspection as stator slot damage by loosen wedge. Finally, the repetition rate, amplitude and gap time analysis were performed with another 5 motors. The results show the different PD severity that were further investigated and confirmed by visual inspection. This PD severity analysis could help to easily analyse, identify problem, predict PD severity and possibility of failure.

Detection and classification of partial discharge (PD) signal in XLPE cable is very important to find out the root cause of insulation failure. The purpose of this work is to improve the accuracy and effectiveness in classifying various types of PD signals including corona, surface, and internal discharges in a noisy environment using machine learning (ML) and deep learning (DL) techniques. For ML models, statistical features such as skewness, kurtosis, standard deviation, and variance are extracted from the PD signals and used for training. In contrast, the DL model focuses on continuous wavelet transforms (CWT) based scalogram images which highlight the resolution of the signal’s energy and are used for feature extraction in the classification study. Instead of using traditional phase-resolved partial discharge (PRPD) plots, this work introduces a novel approach to converting PD signals to scalogram images using CWT. These scalogram images provide a visual representation of the frequency components of different PD types and how they change over time. The wavelet-based scalogram images are used as input features for the proposed convolutional neural network (CNN) and state-of-the-art (SOTA) DL models, along with shallow ML classifiers for model training, validation, and testing. Along with this k-fold cross-validation and hyperparameter tuning are also employed to enhance the overall performance of the model. The classification results demonstrate that the proposed CNN model achieves up to 97.33% recognition accuracy and less computational complexity as compared to various ML and DL models.

This paper presents a feature extraction algorithm combining S transform (ST) and two-directional two-dimensional principal component analysis ((2D) <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> PCA) for partial discharge (PD) pattern recognition. S transform (ST) is firstly employed to obtain a time-frequency representation of the recorded UHF signals. Then, (2D) <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> PCA is applied to compress the ST amplitude(STA) matrices to extract various feature vectors with different (d <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</sub> , d <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sub> ) combinations, i.e. (5, 5), (5, 10), (10, 5) and (10, 10). The extracted features are examined by both PSO-SVM classifier and BPNN. Experimental results show that the classification accuracies by PSO-SVM are all higher than that by BPNN under four circumstances of (d <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</sub> , d <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sub> ) combinations. The success rates of the PSO-SVM with the four feature vectors are above 94% in all cases. It can be found that the proposed feature extraction and classification algorithm can be effectively applied to PD pattern recognition.

Partial Discharge (PD) detection is one of the important methods used to evaluate the insulation condition of in-service high voltage (HV) electrical equipment. On-site PD detection methods are preferred compared to other methods as non-intrusive sensors are easier to install and have no interruptions on the operation. An electrical PD testing utilizing High Frequency Current Transformer (HFCT) installed at the ground terminal of the in-service instrument transformer was used to detect partial discharge activities. However, the limitation of this method is the measurement often lacks of accuracy due to interference in the measured PD waveforms. Therefore, such PD signals are hard to be distinguished from the stochastic noise pulses. Hence, the characteristics of PD signals and interference signals are studied in the Phase Resolved Partial Discharge (PRPD) pattern in order to identify the extent of PD occurrence in the signals. The discharge magnitude could increase significantly because of the noise signals superposed with PD signals. Therefore a method that can remove high magnitude noise will be proposed by using statistical signal processing technique.