Cascaded multitask learning for adaptive remaining useful life prediction for rolling bearings with multiple fault modes

Abstract Accurately predicting the remaining useful life (RUL) of rolling bearings is critical in condition monitoring and measurement for predictive maintenance. However, bearings often experience multiple fault modes (FMs), each producing distinct degradation trajectories and failure risks that complicate accurate RUL prediction. To address this issue, we propose a mixture-of-experts cascaded multitask learning framework that adaptively predicts RUL by leveraging FM-specific degradation features. Specifically, both FM recognition and RUL prediction are performed on the basis of a shared backbone network to learn common representations. FM recognition is formulated as a multilabel classification task, and the resulting outputs are used as gating signals to dynamically activate the corresponding expert networks in the RUL prediction module. Each expert is specialized for a particular FM and is selectively activated by the gating mechanism to provide adaptive RUL prediction. Furthermore, a weighted piecewise loss function is introduced to incorporate asymmetric risk control, prioritizing the penalization of RUL overestimation. By cascading FM recognition with RUL prediction, the proposed model effectively integrates shared representations with FM-specific knowledge to improve predictions under multiple fault scenarios. Experimental results based on the bearing dataset demonstrate that the proposed approach achieves superior predictive accuracy and robustness compared with that of state-of-the-art methods.