A new anomaly detection approach based on siamese network with fast gradient sign algorithm for fault diagnosis of rolling bearing

Abstract Anomaly detection is a crucial step in rolling bearing health management. Recently, unsupervised learning models driven solely by normal data for anomaly detection have received great attention. For the sake of improving the accuracy of anomaly detection in bearing diagnosis, this paper proposes an anomaly detection method based on Zero-Shot Learning. The fast gradient sign method (FGSM) is utilized to generate abnormal samples, which are then classified utilizing the Siamese network, the Zero-Shot Learning problem can be changed into a traditional supervised learning problem. To validate the effectiveness of this method in anomaly detection, this paper utilizes two datasets for processing and analysis. The experiment show that this method performs well in tag classification and fault trend prediction. Furthermore, compared to the traditional semantic discrimination method of Zero-Shot Learning, the proposed method is more practical and straightforward.