Partial Discharge Pattern Recognition Based on Time‐Frequency Multi‐Scale Residual Attention Network

Current partial discharge (PD) recognition models are often constrained by limitations such as insufficient recognition accuracy, limited processing speed, and procedural complexity. To mitigate these limitations, time‐frequency multi‐scale residual attention network (TFMRAnet) is designed to analyse PD signal, which comprises a multi‐scale residual attention‐based adaptive denoising network, a frequency‐domain recognition network, and a decision fusion module based on Dempster–Shafer (D–S) evidence theory. Specifically, the multi‐scale residual attention adaptive denoising module is used to extract multi‐scale features by dilated convolutions of different scales and accelerate training convergence by residual connections. Moreover, a GAM attention mechanism and an adaptive soft‐thresholding function are used for denoising, which preserves PD information and amplifies cross‐dimensional global interactions, thereby improving the performance of the network model. In frequency domain recognition networks, frequency domain features are extracted by performing Fourier transforms on PD signals. The recognition results of the time‐domain and frequency‐domain models are deeply combined through the D‐S evidence theory to improve the multidimensional recognition capability for PD patterns. Finally, a PD experimental platform was built to create four representative PD fault models and generate PD datasets corresponding to distinct fault types. Results from comprehensive comparative and ablation experiments demonstrate that the proposed model achieves perfect recognition accuracy (100%) on the collected PD dataset and exhibits the fastest inference speed (0.643 ms). These findings underscore its significant potential for practical engineering applications.