Diagnosis of distribution network fault using multiresolution S-transform and modified convolution neural network

This study presents a new modified convolution neural network (MCNN) based on a three-stage fault diagnosis framework. An improved multiresolution S-transform (MST) model is proposed initially to calculate the points of feed line fault initiation and recovery efficiently, considering the various sampling resolutions of feed line fault recording and large amount of fault information. First, the proposed fault detection model is robust even without a detection threshold; it achieves relatively high detection accuracy through adaptive adjustment of Gaussian window width. Second, the preprocessed fault waveforms are converted into time–frequency images. Third, a CNN with a parallel block is proposed as a robust classifier. This method can realize fast convergence by utilizing a new activation function and achieve high accuracy by extracting image features in a wide and short spatial range. Finally, simulation and real measurement data are leveraged in the testing phase to verify the performance of the proposed diagnosis method. The proposed models have great performance on the test database when evaluated using accuracy, recall, precision, and F1-score. Results show that the proposed framework obtains an average accuracy of 99.8 % and 98.3 % for simulation fault cases and real measurement data, respectively. The test results of MCNN are better than 1-D CNN and other well-known classifiers.

© 2017 Elsevier Inc. All rights reserved.