A novel multi-task learning framework for bearing remaining useful life prediction under variable operating conditions: Integrating deep hybrid model and invariant consistency strategy

Current existing rolling bearing remaining useful life (RUL) prediction methods often suffer from performance degradation under variable operating conditions due to spurious correlations between source and target domains. To address this, a multi-task learning framework is proposed. Specifically, this paper constructs a deep hybrid model as the shared feature extractor. Based on the idea of bottom-level feature sharing, we integrate the primary regression task and auxiliary tasks for modeling. First, we adopt invariant consistency learning (ICON) to mitigate spurious correlations between the two domains and extract effective domain-invariant features. Second, to further reduce the adverse effects of spurious correlations, we add white noise to generate augmented samples. This approach expands the sample distribution and improves the model’s adaptability to signals under various operating conditions. Finally, we carry out joint optimization and training of the model by using original samples, augmented samples and the two types of tasks. To validate the effectiveness of the proposed method, experiments are conducted on two rolling bearing datasets and compared with other methods. The results demonstrate that the proposed method is capable of effectively predicting the RUL of bearings under varying operating conditions.