Early fault identification method of distribution cable based on PSO-DBN

With the acceleration of urbanization, underground distribution cables have been widely adopted due to their economic efficiency and minimal spatial footprint. However, early-stage arc faults caused by insulation defects exhibit transient and concealed characteristics, while traditional detection methods suffer from insufficient identification accuracy due to susceptibility to fault types, locations, and noise interference. To address this, this paper proposes an early-stage cable fault identification method based on the fusion of a Particle Swarm Optimization (PSO) algorithm and a Deep Belief Network (DBN). The PSO optimizes the number of hidden-layer neurons in the DBN, and the model performance is enhanced through layer-wise pre-training and fine-tuning strategies. Based on a modified IEEE 13-node hybrid distribution network model, 8,856 samples-including half-cycle faults, multi-cycle faults, and motor-starting scenarios-were simulated to evaluate identification performance under varying noise levels. Experimental results demonstrate that the PSO-DBN achieves a fault recognition accuracy of 98.87%, surpassing the conventional DBN by 8.24%. Moreover, it maintains over 90% accuracy under 2-10 dB noise interference, significantly outperforming Extreme Learning Machine (ELM) and Probabilistic Neural Network (PNN) methods. These findings validate its advantages in feature extraction, nonlinear data processing, and robustness.