A Rapid Recognition Method for Transformer Partial Discharge Fault Patterns Based on Improved DSCNN

Pattern recognition of partial discharge (PD) is a key focus in the condition assessment of power equipment. The strong feature extraction capabilities demonstrated by deep learning in the fields of image and signal processing have made it a new trend in PD pattern recognition research. This paper proposes a method for PD pattern recognition based on an improved Deep Separable Convolutional Neural Network (DSCNN). The method first introduces an improved model of the DSCNN, secondly, utilizes a PD experimental detection platform to obtain Ultra-High Frequency (UHF) signals from partial discharges, and employs the S-transform to generate the time-frequency spectrum corresponding to the time-domain signals; then, it uses the improved DSCNN model to classify the time-frequency spectrum for the identification of PD types. The results show that the PD pattern recognition model proposed in this paper has a good recognition accuracy; compared to traditional deep learning models, the optimized model proposed in this paper reduces the scale of parameters and inference time, which is conducive to improving fault diagnosis speed.