Improving pattern recognition accuracy of partial discharges by new data preprocessing methods

• 18 samples are constructed to study the internal PDs in solid and oil insulations, corona, and surface PDs measured in a high voltage laboratory.
• Three novel preprocessing methods are proposed to extract the discriminative features from measured raw data in each class.
• Three scenarios including five classes, one scenario including three classes, and one scenario including seventeen classes are defined for pattern recognition.
• FFBP, RBF, and NPRTool neural networks are trained in each of the five scenarios with extracted feature vectors by each of the three new preprocessing methods.
• The results are superior with those achieved from traditional PRPD method.

In this paper, raw data of partial discharges (PDs) in solid, oil, and air insulation materials are measured experimentally in a high voltage laboratory for 18 samples. Then, three new methods for preprocessing the data based on first, second, and infinite signal norms and besides autocorrelation function (ACF) are proposed. Eventually, feed-forward back propagation (FFBP), radial basic function (RBF) neural networks, and neural network pattern recognition toolbox (nprtool) are used to recognize the patterns of the processed data. The results of the new methods are compared with phase resolved partial discharge (PRPD) method which is common in previous studies. Thanks to the new preprocessing methods, correlation factor in FFBP network, error value in RBF network, and classification percentage in nprtool become 0.9867, 0.0001 and 96.4%, respectively. Moreover, it is concluded that PDs process is a stationary random process which can be estimated by Gauss–Markov process.