Anomaly detection of industrial motors under few-shot feature conditions based on causality

It is observed that previous research studies focusing on few-shot feature data for fault diagnosis or anomaly detection have a limitation, that is, feature extraction methods to solve few-shot feature data problems will also have scenarios where they may not always be applicable. In this paper, a motor anomaly detection model with generalization performance is proposed to meet the anomaly detection needs in the above scenarios. The model consists of a reinforcement unit and a diagnosis unit. Firstly, the reinforcement unit extracts the adjacent features with different timestamps through ensemble learning. Secondly, the temporal convolutional network (TCN) model is nested to increase the receptive field of the reinforcement unit. Additionally, a residual network is introduced to improve the generalization performance. Finally, features obtained from the reinforcement unit are used for final anomaly detection through neural networks in the diagnosis unit. Experimental results indicate that the proposed model achieve an anomaly detection accuracy of 97.96% in factory motor dataset, while the model has the superior generalization ability.