Incipient Fault Detection Based on Multiscale Time Series Feature Extraction
Wang Chengcheng, Keming Sheng, Zhen Liu, Jinjiang Wang, Min Wang · IEEE Transactions on Instrumentation and Measurement · 2025
As the rapid development of industrial manufacturing with increasing level of automation, the reliability and performance requirements of systems have been receiving more and more attention. The occurrence of fault usually leads to substantial financial losses and even monstrous disasters in production process. However, incipient fault detection in manufacturing process is usually limited by two aspects: 1) lack of fault data and 2) hidden abnormal features in time series. Therefore, an incipient fault detection method which is called autoencoder multiscale temporal convolutional network (AMTCN) based on multiscale time series feature extraction is proposed in this article. By autonomously learning distinctive features from the normal samples, the AMTCN circumvents the subjectivity and time-consuming process of manual feature extraction. Moreover, the AMTCN effectively captures long-term dependencies in the temporal dimension while extracting multiscale information in the feature space through the integration of autoencoder (AE) structure and multiscale temporal convolutional layers. Ultimately, a maintenance platform of industrial instruments and meters is constructed and the process data from the platform is utilized to demonstrate the effectiveness and efficiency of the proposed approach for incipient fault detection.