Rapid Training Strategies for Fault Detection in Similar Electromechanical Systems

Abstract

A lot of research has been carried out on failure detection to monitor the quality of electromechanical systems in changing environments. Data that has already been collected may show new signal patterns if the operation of the equipment has changed or if it is newly installed. Therefore, data may not be sufficient because new data must be collected. If failure detection is performed with insufficient data, it may be difficult to obtain good performance. To solve these problems, this study proposes a rapid training strategy that identifies the core parts of existing failure detection models and uses them to train on data collected from new equipment or equipment with different operation. It also proposes a method to measure the similarity between data from existing equipment and data collected from new equipment or equipment with different operation. If the two data are dissimilar, the performance of transfer learning may be adversely affected, so we propose a method to measure this. The proposed method is explained and demonstrated through a case study of bearing vibration signals and Gearbox Simulator.