Demetres Evagorou

Comparison of two partial discharge classification methods

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.

Evaluation of Partial Discharge Denoising using the Wavelet Packets Transform as a Preprocessing Step for Classification

The identification of partial discharges in high voltage equipment has emerged as one of the most effective condition monitoring methods for assessing the integrity of the equipment under test. The fact that the application of PD monitoring methods is being applied online makes the measurements suffer from noise, inevitable at the measurement point, and reduces the sensitivity of the measurements. Signal processing methods to post process the measurements have been utilised, resulting not only in rejection of the noise and improvement of the sensitivity, but also in improved classification of the PD. A powerful noise rejection technique, the wavelet packets transform (WPT) has been extensively employed for the effective extraction of PD signals from noise. This technique is particularly useful in denoising signals which have transient characteristics. It expands the signal into different bases that are chosen adaptively according to a cost function, transforming the signal into a set of wavelet coefficients. The choice of a cost function has a significant effect on the compact representation of the signal. In this paper after the theory of wavelet packets is first briefly presented, and the denoising performance of the various wavelet packets parameters, such as the wavelet function, the thresholding type, and the cost function to be used is studied through the use of data acquired in a laboratory experimental environment for four types of discharges; namely the corona discharge in air, the internal discharge in oil, the floating discharge in oil and the surface discharge in air. The Symmlet wavelet has been compared with the Daubechies wavelet, both with 8 vanishing moments, the hard thresholding rule has been compared with the soft thresholding rule, and three cost functions have been compared as to their suitability for best basis expansion. Using some predefined criteria to assess their denoising performance the Symmlet 8 has been found to outperform the Daubechies 8 wavelet, the hard thresholding rule to yield better performance than the soft thresholding rule and the Shannon entropy cost function to perform better that the log energy and the norm energy cost functions.

Classification of Partial Discharge Signals using Probabilistic Neural Network

Partial Discharge (PD) classification in power cables and high voltage equipment is essential in evaluating the severity of the damage in the insulation. In this paper, the Probabilistic Neural Network (PNN) method is used to classify the PDs. After the algorithm has been trained it uses the input vector, which contains the features that would be used for classification, to calculate the probability density function (pdf) of each class and together with the assignment of a cost for a misclassification the decision that minimizes the expected risk is taken. The maximum likelihood training is employed here. The success of this particular method for classification is asserted. This method has the advantage over Multilayer Neural Network that it gives rapid training speed, guaranteed convergence to a Bayes classifier if enough training examples are provided (i.e. it approaches Bayes optimality), incremental training which is fast (i.e. additionally provided training examples can be incorporated without difficulties) and robustness to noisy examples. The results obtained here (99.3%, 84.3% and 85.5% for the corona, the floating in oil and the internal discharges respectively) are very encouraging for the use of PNN in PD classification.

A lower dimensional feature vector for identification of partial discharges of different origin using time measurements

Partial discharge (PD) classification into sources of different origin is essential in evaluating the severity of the damage caused by its activity on the insulation of power cables and their accessories. More specifically, some types of PD can be classified as having a detrimental effect on the integrity of the insulation while others can be deemed relatively harmless, rendering the correct classification of different PD types of vital importance to electrical utilities. In this work, a feature vector was proposed based on higher order statistics on selected nodes of the wavelet packet transform (WPT) coefficients of time domain measurements, which can compactly represent the characteristics of different PD sources. To assess its performance, experimental data acquired under laboratory conditions for four different PD sources encountered in power systems were used. The two learning machine methods, namely the support vector machine and the probabilistic neural network, employed as the classification algorithms, achieved overall classification rates of around 98%. In comparison, the utilization of the scaled, raw WPT coefficients as a feature vector resulted in classification accuracy of around 99%, but with a significantly higher number of dimensions (1304 to 16), validating the PD identification ability of the proposed feature. Dimensionality reduction becomes a key factor in online, real-time data collection and processing of PD measurements, reducing the classification effort and the data-storage requirements. Therefore, the proposed method can constitute a potential tool for such online measurements, after addressing issues related to on-site measurements such as the rejection of interference.

Performance of the Support Vector Machine Partial Discharge classification method to noise contamination using phase synchronous measurements

The Support Vector Machine (SVM) method has been used with success in classifying Partial Discharge (PD) data of different sources. In this work it was investigated whether the previous success of the Support Vector Machine (SVM) could be extended to the case where a PD measurement was corrupted by Additive White Gaussian Noise (AWGN). Data was collected from experiments using PDs of different sources under controlled laboratory conditions at the Tony Davies High Voltage Laboratory, University of Southampton. Artificial PD signals were injected into the HV electrode of a bushing and a high frequency current transformer (HFCT) was used to monitor the current between the tap-point and earth. The signals produced by four different artificial PD sources (corona discharge in air, floating discharge in oil, internal discharge in oil and surface discharge in air) were acquired using the peak detection mode of the oscilloscope and were processed to extract the feature that was used by each algorithm. The feature extraction algorithm involved the use of the Wavelet Packet Transform (WPT) on phase synchronous measurements corrupted by artificial AWGN. Once the SVM was trained using part of the data acquired in the laboratory then the remaining data was corrupted by noise of two different amplitudes, giving SNRs of 7 dB and 3dB. These noisy data were classified using the SVM and the classification results were recorded. This procedure validated the SVM as an effective classification method that can be trained using laboratory noise free PD signals which can subsequently be used to classify field on-line measurements that have been corrupted with noise.

Preliminary investigation on the feasibility of radiometric techniques to detect faults in buried cable joints

This paper investigates the preliminary use of radiometric techniques to the detection of PDs in buried cables, and in particular to cable joints. The transfer function from the source to the detector is a function of the propagation characteristics of the media involved. In the case of radiometric detection the inclusion of soil, in general a lossy and dispersive medium with frequency and content dependent characteristics, further contributes to signal attenuation. The work undertaken here examines whether a repetitive pulse of varying amplitude and frequency, injected into an experimental arrangement that simulates buried power cables, is being detected by two simple antennae above ground. Successful detection of the pulses showed the preliminary possibility of the use of such techniques in PD detection, which creates the need for further experiments and antenna designs to be explored.

Multisource PD identification based on phase synchronous and asynchronous data

Partial Discharge (PD) measurements in cables and their accessories play a fundamental role in Condition Based Monitoring (CBM) of High Voltage (HV) equipment. CBM monitoring has been enforced by utilities in the transmission and distribution (T&D) environment as part of a predictive maintenance program that aims to result in less unscheduled downtime and lower maintenance cost. Identifying the source of a PD rather than merely assessing its magnitude provides additional information that could enable more educated decisions concerning the integrity of the insulation to be made. In on-line scenarios the presence of multiple PD sources that are simultaneously active as well as the presence of interference, complicates the identification process. In this paper, the Density-Based Spatial Clustering of Applications with Noise (DBSCAN) algorithm has been employed to identify PDs of different sources. Phase synchronous measurements were acquired in the laboratory and pre-processed through a peak detection algorithm to extract the single pulses (phase asynchronous). To extract a feature the Wavelet Packet Transform (WPT) and Higher Order Statistics (HOS) were employed according to previous work by the authors. The feature was then analyzed by the Principal Component Analysis (PCA) for dimensionality reduction and study of different PD sources has been shown to form separate clusters. Application of this method on on-line data acquired from the network of the Electricity Authority of Cyprus (EAC) has demonstrated its potential use in PD identification and interference rejection.