Optimization Method for Breakdown Diagnosis Based on Pre-Breakdown Current and Improved Residual CNN in Vacuum Circuit Breaker

Accurate and fast breakdown diagnosis was important for vacuum circuit breaker insulation improvement. This paper proposed an optimization method to improve both the breakdown diagnosis accuracy and efficiency based on pre-breakdown current and improved residual convolutional neural network (CNN). The pre-breakdown current was only obtained to diagnose breakdown instead of traditional whole voltage and current for the first time, which made the effective features more obvious by removing the ineffective features. Based on the pre-breakdown current features, an improved residual CNN was applied, which used convolutional block attention module to enhance the focus on effective features and residual module to enhance the extraction of effective features. K-Fold cross-validation study and ablation study were carried out. The results showed that the proposed method was robust during the training. The breakdown diagnosis accuracies with the proposed method (90.15%, 95.19% and 92.78% respectively) were the highest in ablation study. The corresponding evaluation metrics including precision, recall and F1-score were also the highest, which verified its effectiveness. Moreover, the time and parameters consumption were reduced to 20% and 15% respectively. Both the breakdown diagnosis accuracy and efficiency were improved effectively by the proposed method, which could have a promising application in breakdown diagnosis for vacuum circuit breaker.