A multi-task prediction network with dynamic convolution and self-supervised contrastive learning in mechanical remaining useful life prediction

Remaining useful life (RUL) prediction and health state (HS) assessment are two key tasks in prognostics and health management (PHM). However, existing studies mostly investigate them independently, ignoring their intrinsic correlation, which reduces prediction efficiency and may affect the accuracy of maintenance decision-making. This paper proposes a multi-task prediction network based on dynamic convolution and self-supervised contrastive learning (DCSCL-MTPN). Under a multi-task learning framework, it constructs an HS assessment task and multiple RUL prediction branches, and selects the corresponding RUL branch according to the HS assessment results, thereby improving the specificity and prediction accuracy of the model. First, a dynamic convolution feature extraction module is established as the underlying structure of the multi-task framework, and shared features for HS assessment and life prediction are extracted by incorporating a dynamic adaptive strategy; subsequently, an input and feature self-supervised contrastive learning module (IFSCL) is constructed to align input signals and high-dimensional features in the latent representation space, ensuring distribution consistency between the high-dimensional features and the original signals, thereby preserving key degradation information and improving the generalization ability of the model. In addition, a homoscedastic uncertainty method is introduced to adaptively balance the loss weights among tasks, ensuring that each task can be collaboratively optimized. Finally, experiments are conducted on the XJTU-SY bearing dataset and a private bearing performance degradation dataset. The results show that the proposed method outperforms multi-task comparison methods such as ACNN-SAE and MSFMTP, with average accuracy improvements of 10.37% and 7.05%, respectively.