Abstract:The frequent occurrence of early cable faults easily leads to the occurrence of permanent cable faults, causing serious impacts on the stable operation of the power grid. In order to accurately identify early cable faults before permanent failures occur, a method based on an improved deep residual contraction network for early cable fault identification is proposed in this paper. Firstly, an improved complete ensemble empirical mode decomposition with adaptive noise method was employed for fault signal processing, and correlation coefficients were utilized for filtering IMF components. Then, the composite multiscale permutation entropy of the IMF component was calculated for further feature extraction, constructing a feature dataset. Finally, the improved deep residual shrinkage network, incorporating an enhanced shrinkage module, multi-scale convolutional layer, Self Attention, and SimAM attention mechanism, was employed for early cable fault identification experiments. The algorithm is demonstrated by experimental results to accurately identify early cable faults and exhibit a certain degree of anti-interference capability.