In power systems, the accurate identification of partial discharge (PD) in cables is crucial for the early detection of insulation defects and for maintaining power grid reliability. However, the low amplitude of PD signals and their susceptibility to noise interference increase the complexity of the identification process. To address this challenge, this paper proposes a novel deep learning architecture: the Adaptive Residual Diffusion Denoising Morphological Attention Network (ARDDMA-Net). ARDDMA-Net employs a two-stage design. In the first stage, an adaptive residual diffusion denoising module dynamically suppresses noise, while a morphological attention mechanism preserves critical features of PD signals, such as pulse width and edge structures. This effectively mitigates the masking and misclassification of weak PD pulses caused by noise. In the second stage, a multi-scale attention-enhanced ResNet-1D convolutional neural network extracts multi-scale features and performs precise classification, thereby addressing the limitations of traditional convolutional neural networks (CNNs) in handling low signal-to-noise ratio (SNR) conditions and multi-scale characteristics. Experimental results demonstrate that ARDDMA-Net maintains stable recognition performance under various noise levels and data loss conditions, outperforming traditional machine learning methods and existing deep learning models in terms of identification accuracy.
Yang Zou (Mon,) studied this question.