Partial Discharge Pattern Recognition Based on Swin Transformer for Power Cable Fault Diagnosis in Modern Distribution Systems

As critical infrastructure in modern distribution systems, power cables face progressive insulation degradation from partial discharge (PD), while conventional recognition methods struggle with feature extraction and model generalizability. This study develops an integrated experimental platform for PD pattern recognition in power cable systems, comprising a control console, high-voltage transformer, high-frequency current transformer, and ultra-high-frequency (UHF) signal acquisition equipment. Four distinct types of discharge-defective models are constructed and tested through this dedicated high-voltage platform, generating a dataset of phase-resolved partial discharge (PRPD) spectra. Based on this experimental foundation, an improved Swin Transformer-based framework with adaptive learning rate optimization is developed to address the limitations of conventional methods. The proposed architecture demonstrates superior performance, achieving 94.68% classification accuracy with 20 training epochs while reaching 97.52% at the final 200th epoch. Comparisons with the original tiny version of the Swin Transformer model show that the proposed Swin Transformer with an adaptive learning rate attains a maximum improvement of 6.89% over the baseline model in recognition accuracy for different types of PD defect detection. Comparisons with other deeper Convolutional Neural Networks illustrate that the proposed lightweight Swin Transformer can achieve comparable accuracy with significantly lower computational demands, making it more promising for application in real-time PD defect diagnostics.