Fault Recognition Method for Substation Equipment Based on Lifelong Learning

This paper proposes a fault recognition method for substation equipment based on lifelong learning, which facilitates the continuous expansion to new devices and new faults. Firstly, we classify faults based on equipment and create an expert network which can distinguish detailed fault types for each equipment class. A gating network is learned to automatically identify the current equipment and match it with the corresponding expert network, and achieve end-to-end equipment and fault recognition tasks. The feature extraction part of the expert network adopts a pre-trained residual network and incorporates spatial and channel attention modules to enhance the capability of fault identification. During the training process of the expert network, we save a small number of samples for each fault class and use the proposed loss function to reduce the forgetting problem during the lifelong learning process. Finally, we validate the effectiveness of the proposed approach by conducting comparative experiments on the recognition performance and scalability of the network.