Unsupervised Signal Anomaly Transformer method: Achieving bearing life anomaly detection without the need for failure samples
Ping Yu,Mengmeng Ping,Jialin Ma,Jie Cao
DOI: https://doi.org/10.1016/j.engappai.2024.108940
IF: 8
2024-07-20
Engineering Applications of Artificial Intelligence
Abstract:Bearings are a crucial component, whose performance and lifespan directly affect the safety and operational efficiency of the entire mechanical system. However, in the case of new components or new process bearing equipment that lack complete life-cycle data, supervised learning methods relying on labels may fail. Therefore, a specialized unsupervised anomaly detection method for bearing signals has been proposed. Drawing on the ideas of the anomaly transformer, an unsupervised signal anomaly transformer method is proposed. This method optimizes the anomaly transformer based on signal input and uses a convolutional neural network (CNN) encoder–decoder structure to encode and reconstruct signals. Given the instability of symmetric Kullback–Leibler (KL) divergence in the overlapping area of probability distributions, this method uses the Wasserstein distance to measure the distance between two distributions. Additionally, a new metric is proposed for comparing the decoder-reconstructed signal to the original signal. To verify the effectiveness of this method, an objective evaluation was conducted using simulated signals, achieving an average accuracy of 99.54%. Furthermore, the 2012 IEEE and XJTU-SY real datasets were used for subjective evaluation of anomaly detection signals. Multiple results confirm that this method has strong competitiveness in bearing signal anomaly detection, significantly improving the prediction accuracy and reliability of anomalies in bearings without full life data.
automation & control systems,computer science, artificial intelligence,engineering, electrical & electronic, multidisciplinary