A Substation Line Fault Diagnosis Method Based on Analog–Switch Fusion Features and a Dual-Channel One-Dimensional Convolutional Neural Network

Abstract

To address the limited use of relay-protection action information in existing line fault diagnosis methods, this paper proposes a diagnostic approach based on analog - switch fusion features (ASFF) and a dual-channel one-dimensional convolutional neural network (DC1D-CNN). CEEMDAN-HHT is used to extract transient time - frequency features from voltage and current signals, while the rising sequence position coefficient, falling sequence position coefficient, and action time proportion coefficient are used to describe protection switching signals. After principal component analysis, the two feature groups are fed into the two channels of the DC1D-CNN. Batch normalization, an attention mechanism, and dropout are then used for feature learning and fault classification. Experimental results show that the proposed model achieves a maximum accuracy of 96.67% on a mixed multi-type fault dataset, while all evaluation metrics remain above 93% across five independent trials, indicating strong diagnostic accuracy and generalization capability.