Random Forest Based Optimal Feature Selection for Partial Discharge Pattern Recognition in HV Cables

Partial Discharge (PD) detection in High Voltage (HV) equipment is an essential technique for assessing the insulation condition in the power system apparatus. The presence of PD is one of the critical indications for the defect and deterioration of the insulating material. Effective signal processing techniques are required to acquire, analyse, and classify PD signals into a specific fault type. The PD measurement device is prone to be affected by various noise sources, which leads to misinterpretation of the PD signal. This research focuses on developing a suitable algorithm for denoising PD signals and proposes a novel feature extraction process for classifying PD signals. The signal decomposition-based denoising algorithms in various engineering disciplines have become popular, and such techniques are being applied to PD signals. A novel PD signal denoising using adaptive decomposition, optimized by entropy and statistical parameters is proposed. Furthermore, a novel entropy-driven PD signal classification method is proposed. Initially, the PD signal is decomposed using various adaptive decomposition methods to get Intrinsic Mode Functions (IMFs). Then, IMFs are analysed using statistical or information theory measures such as Kurtosis, Mutual Information (MI) and Dispersion Entropy (DE). The IMFs are classified as noise, noise–dominant and signal IMFs, based on Kurtosis and MI values. The DE value of IMFs is used to select the regularisation parameter of Group-Sparse Total Variation (GSTV). The noise–dominant IMFs are subjected to the GSTV method, and the signal IMFs are added to the output vector to reconstruct the denoised output. The noisy IMFs are discarded for further processing. The proposed methods are applied under different input signal-noise ratios, lengths, and noise sources. The proposed denoising methods with the results and findings are presented and discussed, along with the performance metrics related to the filters. In addition, Tsallis entropy and a novel Generalized Gaussian Distribution – Refined Composite Multiscale Dispersion Entropy (GGD–RCMDE) based features are proposed for the time series signal. The suggested feature extraction methods are applied to the PD dataset to extract the features from various PD signals such as ‘corona’, ‘surface’ and ‘void’. A novel preprocessing method followed by the min-redundancy max-relevance (mRMR) feature selection method is applied to select sensitive features. A multiclass classifier model is trained using the most sensitive features to classify PD faults. The robust experimental analysis of the feature extraction method was carried out and reported. The performance metrics are computed using the results obtained through the experiments. The effectiveness of the proposed denoising and classification methods was presented with a suitable discussion and conclusion.

The feature extraction and pattern recognition techniques are of great importance to assess the insulation condition of gas‐insulated switchgear. In this work, the ultra‐high‐frequency partial discharge (PD) signals generated from four types of typical insulation defects are analysed using S‐transform, and the greyscale image in time‐frequency representation is divided into five regions according to the cutoff frequencies of TEm1 modes. Then, the three low‐order moments of every subregion are extracted and the feature selection is performed based on the J criterion. To confirm the effectiveness of selected moment features after considering the electromagnetic modes, the support vector machine, k‐nearest neighbour and particle swarm‐optimised extreme learning machine (ELM) are utilised to classify the type of PD, and they achieve the recognition accuracies of 92, 88.5 and 95%, respectively. In addition, the results show that the ELM offers good generalisation performance at the fastest learning and testing speeds, thus more suitable for a real‐time PD detection.

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.

Two signal classification methods have been examined to discover their suitability for the task of partial discharge (PD) identification. An experiment has been designed to artificially mimic signals produced by a range of PD sources that are known to occur within high voltage (HV) items of plant. The bushing tap point of a large Auto-transformer has been highlighted as a possible point on which to attach PD sensing equipment and is utilized in this experiment. Artificial PD signals are injected into the HV electrode of the bushing itself and a high frequency current transformer (HFCT) is used to monitor the current between the tap-point and earth. The experimentally produced data was analyzed using two different signal processing algorithms and their classification performance compared. The signals produced by four different artificial PD sources (surface discharge in air, corona discharge in air, floating discharge in oil and internal discharge in oil) have been processed, then classified using two machine learning techniques, namely the support vector machine (SVM) and probabilistic neural network (PNN). The feature extraction algorithms involve performing wavelet packet analysis on the PD signals recorded over a single power cycle. The dimensionality of the data has been reduced by finding the first four moments of the probability density function (Mean, Standard deviation, Skew and Kurtosis) of the wavelet packet coefficients to produce a suitable feature vector. Initial results indicate that very high identification rates are possible with the SVM able to classify PD signals with a slightly higher accuracy than a PNN.

Pattern recognition of partial discharge (PD) has various beneficial applications in academia and industry. However, it is hard and expensive to obtain extensively annotated PD data to build a high-performance classification model. Given there is potential to use PD data more effectively, this paper presents a novel data generative model based on generative adversarial networks (GAN). GAN are able to learn deep feature representations of existing PD signals and synthesizes more extensive UHF PD signals. An original PD data set of UHF PD signals is established by the partial discharge experiment. With the result of the original PD signals dataset as baseline, we evaluated the performance of some classifiers on the augmented dataset. The results show that UHF PD classification model benefits from GAN-based data augmentation techniques. Clearly, the GAN-based model has good potential in industry diagnosis of PD, especially when the data are sparse.

Capacitor voltage transformers (CVTs) composed of aluminum foil/oil-impregnated paper insulation system have been operated over 30 years in substations. For the condition monitoring and diagnosis of aged CVTs, partial discharge (PD) detection is expected. However, PD characteristics from PD inception and progress to breakdown (BD) have not yet been clarified. Thus, PD characteristics of a fundamental insulation model composed of aluminum foil/oil-impregnated paper insulation systems are discussed in this paper. Experimental results under a fixed applied voltage revealed that PD characteristics can be classified into 3 periods. In Period 1, PD signals were intermittently generated mainly around the peak phase of applied voltage. In Period 2, PD signals were continuously generated around zero crossing (void discharge) to peak phase. In Period 3, PD signals larger than 10,000 pC were generated. Together with the above electrical PD measurement, the optical measurement of PD light emission images, bubble behavior, and their time transmission were also carried out in order to discuss the mechanism of PD progress.

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.

The cable attachment interface has always been a weak insulation link in cable systems. In this paper, the growth characteristics of electric tree at the interface was studied at 70 °C, and the growth morphology and partial discharge (PD) signals of electric trees were collected simultaneously. The cross-linked polyethylene (XLPE) and silicone rubber (SIR) were pressed together by a pressing device, and AC high pressure was applied to a steel needle that is vertically inserted 1mm from the XLPE-SIR interface, with a pin plate spacing of 2mm. It was shown that there was a growth pattern of "trigger-rapid growth-stagnation-breakdown" in the electric tree at the interface. It was found that the amplitude and frequency of partial discharge are closely related to the growth stage of electric tree. The relationship between the growth characteristics of electrical tree and the partial discharge signal at the interface was established, which provided the experimental basis for the online monitoring technology to judge the insulation state of cable accessories more accurately.

Today the sensitive measurement of partial discharge (PD) signals in XLPE cable can be an important element for quality assurance and diagnostics in the factory and during the on-site test procedure. The results clearly indicate that PD analysis of XLPE cable is an excellent online tool to assess its dielectric condition. However, as the serious interference in-field, the acquired PD signals may be interrupted by periodic interferences and white noises, easily. This paper describes an intelligent morphological filter algorithm to remove the periodic pulse interference and white noise contained in the PD signals. The algorithm consists of two stages of morphological processing. The performance of the algorithm is evaluated with simulated and real PD signals. Compared with classical wavelet, there are substantial improvements in the indexes of average error, root mean square (rms) error and maximum error, especially in reducing distortion of the baseline waveform.

The converter transformer is one of the most critical equipment in high-voltage direct current (HVDC) transmission systems. Valve-side windings withstand complex stresses combined with AC, DC, and pulse voltages. Oil-paper insulation is used within most of HVDC converter transformers. Partial discharge (PD) in oil-paper insulation is a major reason for the development of internal faults in HVDC converter transformer. The current paper presents the development process for the PD of oil-paper insulation under pulsating DC voltage combined with AC and DC voltages at different temperature. In the experiments, the artificial insulation defect models were designed to generate PD signals under pulsating DC voltage detected by a Rogowski coil sensor. The Statistical histograms of φ-q-n based on these PD signals were constructed. The variations in initial discharge voltages and φ-q-n of the PDs with pulsating DC voltages generated from different temperature were analyzed. Experimental results showed that the initial discharge voltages and PD phase-resolved distributions of different temperature were different from one another. The findings of the current study are useful in further research on the mechanism and pattern recognition of PDs in converter transformers.

Metal particles and bubbles are two typical impurities in transformer oil that can cause partial discharge (PD) and even electric breakdown. The PD characteristics are closely related to the motions of these impurities in flowing transformer oil. In this paper, the motions of impurities were obtained by a camera and a simulation model at first. Then, we employed three methods, namely, antennas, a non-inductive resistor, and the method recommended by IEC 60270, to measure the PD signals simultaneously. The breakdown experiments of transformer oil containing the metal particles and bubbles were also performed. In terms of these, single PD waveforms, statistical characteristics, and breakdown voltage were obtained. The results showed that the metal particles could excite ultrahigh frequency (300 MHz-3 GHz) PD signals, whereas bubbles only excited signals in the very high-frequency band (30-300 MHz). Moreover, the PDs induced by the metal particles had a higher repetition rate and a smaller amplitude than the PDs induced by bubbles. In addition, the metal particles easily caused PD, whereas bubbles easily led to the breakdown. Possible reasons for different discharge characteristics between the two impurities were discussed on the basis of the motions of impurities and the distortions of electric field.

The analysis of partial discharge (PD) signals has been identified as a standardized diagnostic tool in monitoring the condition of different electrical apparatus nowadays. In this paper, we propose a novel data-driven approach to detect PD pulses in power cables using one-dimensional convolutional neural networks (CNNs), a successful deep neural network approach. Applying this deep learning method, an end-to-end framework has been proposed considering the propagations of PD signals and noises in power cables. The proposed method uses PD pulses as input, automatically extracts meaningful features for waveforms of PD pulses, and finally detects PD. Most of the existing methods, which use traditional classifiers, such as support vector machines (SVMs) and multi-layer perceptron (MLP) have mainly focused on improving feature representation and extraction manually for this task. However, the proposed CNN-based detection algorithm captures important latent features for waveforms of PD pulses with the help of its automatic feature extraction capability from raw inputs. Our experimental results show that the proposed method is better than conventional SVMs and achieves 97.38% and 93.23% detection accuracies in end-to-end settings based on our theoretical model-generated and empirical real-world PD signals, respectively.

Partial discharge detection is a crucial method for assessing the insulation condition of on-board high voltage cables. Due to the variability of the external environment and the multi-factor interference in the signal transmission process, it is difficult to obtain a more ideal partial discharge signal, which causes difficulties in the evaluation of cable insulation status. This paper proposes a method to suppress partial discharge narrow-band interference on the base of singular spectrum analysis, which achieves the analysis and prediction of one-dimensional signals by singular value decomposition (SVD) based on a matrix of specific matrices constructed from time serials., so that trend terms, periodic terms and noise can be decomposed from a time series. Featured in this article, we simulate the ideal partial discharge signal containing narrowband noise, and then apply this method to reduce the noise. The noise reduction results demonstrate that the proposed approach can be effective remove the narrow-band interference of partial discharge of vehicle-mounted high-voltage cables, and the research results of this paper are effective in extracting the partial discharge signal of vehicle-mounted high-voltage cables under narrow-band interference, which has certain guiding significance for determining the insulation state of vehiclemounted cables.

Recent studies have shown that wavelet transform can effectively be used for noise reduction in the context of partial discharge (PD) signal detection and classification. Several thresholding approaches for wavelet denoising have been reported in the literature. In this study, a novel wavelet threshold estimation method, named energy conservation‐based thresholding (ECBT), is introduced. The proposed thresholding function is capable of conserving a significant portion of the original signal energy, while the threshold value is determined based on the relative difference between the original and noisy signal energies. The proposed method is first applied to PD signals contaminated with different levels of simulated noise. Results show that ECBT produces a denoised PD signal with higher signal‐to‐noise ratio (SNR) and less distortion than PDs produced by the existing wavelet methods. Then, ECBT is modified to address actual PD signals corrupted with real noise, where a robust SNR estimation method is derived to estimate the noise level embedded in the measured PD signals. The denoised PD signals indicate that the proposed method yields higher reduction in noise levels than other methods.

Cellulose particles present a significant concern within the oil‐paper insulation of transformers, posing potential risks to insulation performance. Under the influence of the electric field, the movement of cellulose particles can compromise the transformer's insulation, leading to potential failure. An experimental platform was established to synchronously record particle motion images, partial discharge (PD) pulses, and electric voltage waveforms in oil, aiming to observe the PD characteristics resulting from particle motion under alternating current (AC) voltage and investigate the relationship between different particle motion modes, motion positions, and PD signals. The findings reveal that the phase distribution of PD signals is correlated with the particle motion mode. Specifically, the phase distribution of PD pulses during the back‐and‐forth motion mode is between 4°–94° and 182°–275°. In the suspended oscillation motion mode, the PD pulses phase is concentrated between 20°–84° and 203°–268°. The generation of PD pulses is closely linked to the particle's motion position. PD pulses occur when the particle remains on the electrode during the back‐and‐forth motion mode, generally, PD pulses rarely occur during the jumping process between the two electrodes. In the suspended oscillation motion mode, PD pulses occur when the particle moves upward, but generally do not occur during downward movement. Furthermore, the Pulse Sequence Analysis technique was used to employ the PD characteristics caused by particle motion in transformer oil. The simulation calculations of the electric field distribution for two different particle motion modes show that the particle's motion can cause distortion of the electric field distribution, leading to the generation of PD. The study of the PD characteristics at different particle motion modes and positions obtained contributes to a deeper understanding of the PD induced by cellulose particle motion under AC voltage and provides a reference for the insulation evaluation of transformers.