A Dynamic Sensitivity Assessment Continual Learning Framework for Remaining Useful Life Prediction of Bearing

In the practical industry, constructing a comprehensive training dataset for bearing degradation data is exceptionally challenging due to prolonged collection cycles and diverse degradation modes, which hinders the engineering implementation of deep learning-based remaining useful life (RUL) prediction methods. To address the aforementioned limitations, this article establishes a novel continual learning framework named dynamic sensitivity assessment continual learning framework (DSACL) for RUL prediction, with core innovations including omni-dimensional dynamic convolution (ODConv) and sensitivity assessment. The ODConv strengthens multi-scale feature extraction capabilities through a multi-dimensional attention mechanism, while sensitivity assessment prevents alterations to importance weights through assessing the sensitivity of output functions to parameter variations. Meanwhile, a novel RUL prediction neural network model based on ODConv is constructed for the neural network model of DSACL. Experimental validation on the XJTU-SY dataset demonstrates that DSACL surpasses state-of-the-art methods in prediction performance. Moreover, the introduction of a forgetting mechanism achieves an effective balance between model capacity and prediction accuracy.