Implementation Of Neural Network And Wavelet Transform For Recognition Of GIS Signals

Partial discharge measurements provide vital information on the integrity of GIS insulation. However, noises present in the signals limit the accuracy of diagnoses from such measurements. This limitation can cause delay in appropriate remedial measures to be taken, leading to further deterioration of GIS insulation or a total breakdown.Corona discharges are a form of stochastic pulse-shape noises occurring outside the GIS. It is difficult to distinguish and suppress due to similarities between PD and corona. Methods based on neural networks were proposed to classify PD and stochastic pulse-shape noises. These methods however do not provide an effective feature extraction process, leading to difficulties and inefficiency in designing the neural networks.A method based on wavelet packet transform is proposed here for feature measurements and extractions of partial discharge and corona.In this paper, a method employing wavelet analysis is proposed for feature extraction of PD and corona signals. Based on the extensive investigation of PD analysis and measurement in GIS, wavelet packet transform is used to capture the time-frequency characteristics of the above-mentioned signals. Furthermore, Fisher's discriminant is employed to extract the feature vectors to be used as the input of pattern recognition system. Subsequently, a MLP model is trained and tested to classify PD and corona. Experimental results show that the feature extraction approach remarkably reduces the dimensionality of the input vector while the characteristics of the signals have been reserved. This simplifies the design of neural network and improves the precision of recognition.