Fault Assessment of Circuit Breakers Based on Wide-Area Generalization Information and Improved ACGAN

Switching mechanical faults pose a significant risk, characterized by an evolution from mild to severe states. Accurate identification of these faults can greatly improve the operational reliability of conventional circuit breakers (CCBs). In light of this, a fault degree assessment method is proposed based on the wide-area information fusion and improved deep learning models. First, the switching fault-related vibration signal segments are segmented to reduce the interference of invalid components caused by the multisource nature of the vibration signals. Meanwhile, the wide-area information of fault signal segments is fused to synthesize color image samples to enrich the fault characterization information and minimize the limitations of single domain analysis. Second, targeting at the dataset class imbalance, the size of samples is expanded using the Wasserstein auxiliary classifier generative adversarial network with gradient penalty (WACGAN-GP) to balance the dataset. It aims to effectively generate sample data highly similar to the real samples and improve the robustness of fault degree assessment. In addition, the MDA-ResNeXt model is constructed by incorporating the multitask learning (MTL) model into the improved residual network (ResNet) and dual attention network (DANet) to enhance the fault degree assessment by mutual learning of fault classification and degree assessment tasks. The fault degree assessment example analysis shows that this method exhibits higher classification accuracy and operational efficiency compared to other methods.