A rolling bearing fault diagnosis framework under variable working conditions considers dynamic feature extraction

Abstract

As a key component in industrial machinery, rolling bearings usually operate at variable speeds. The features in the signal are dynamic due to speed changes, with complementarity and correlation embedded in the different features. However, the utilization of this complementarity and correlation in mining dynamic signal features has been neglected, leading to reduced accuracy of fault classification models and less adaptability to variable working conditions. To address this problem, a multi-scale asymmetric feature reproduction plots-shifted window transformer (MAFRP-ST) framework of rolling bearing fault diagnosis is proposed under variable speed conditions in this study. Specifically, the framework includes dynamic feature capture and dynamic feature learning modules. The dynamic feature capture module is designed to convert signals into multi-scale asymmetric feature reproduction plots (MAFRP) containing features in the time–frequency domain, allowing the deeper dynamic features to be better captured by exploiting the complementarities and correlation. In the dynamic feature learning module, a shifted window (Swin) transformer adapted to dynamic features at different scales is developed, calculating local attention according to the window size in each layer and incrementally increasing the receptive field layer by layer. Compared with recently proposed similar methods, the MAFRP-ST framework improves diagnosis accuracy by about 4.1% and 2.2% on average across two datasets, respectively, and better robustness to noise is demonstrated.