Rolling mill fault diagnosis under limited datasets

Abstract

Sensor technology and deep learning have gained a lot of attention in the field of mill fault detection, which provides new possibilities for the condition monitoring of mills. The study provides a novel dual impact feature enhancement framework for rolling mill condition monitoring to address the issue of variable condition diagnosis with limited data. This feature enhancement framework is jointly guided by the multi-scale impact feature method and dual attention mechanism. Firstly, different multi-scale impact feature methods are developed for vibration and acoustic signals to fully exploit the impact features of signals. Secondly, coordinate attention is introduced for vibration signals, and a multi-level feature coding block is designed to effectively mine advanced impact features. Then, for acoustic signals, efficient channel attention is introduced and a multi-level feature coding block is designed to effectively mine advanced impact features. Finally, the effectiveness of the suggested approach is validated utilizing two experimental datasets. Experimental results reveal that the suggested approach outperforms seven current defect diagnosis approaches in performance.