A Fault Diagnosis Method for Power Electronic Circuits Based on GADF Coding and Channel Split Residual Network

With the objective to achieve an accurate diagnosis of commonly encountered fault problems in power electronic circuits and to prevent the occurrence of catastrophic failures, a soft fault diagnosis model for dc-dc converter based on the gram angle difference filed (GADF) and the channel split residual network (CSRN) is proposed. Firstly, the original fault data is coded by GADF to generate the fault data set. Secondly, channel split preprocessing (CSP) and Darwinian selection mechanism (DSM) are introduced to enhance the residual network and further construct the CSRN. CSP emphasizes the important channel features in the fault feature images by achieving a brute-force enhancement of the channel information and highlighting subtle features. DSM retains the key features and removes the secondary features during network training. The GADF-CSRN diagnosis model proposed in this paper is verified by the fault data collected from a dc-dc converter. The result showed that a diagnosis accuracy of 97.92% is achieved, surpassing the accuracy of conventional method by at least 1.85%, while exhibiting better noise resistance.