DPMA: Self-Supervised Dual-Path Meta Alignment Network for Remaining Useful Life Prediction With Limited Data and Unknown Working Conditions

Degradation state assessment and remaining useful life (RUL) prediction are crucial for formulating maintenance strategies under variable operating conditions. Current cross-condition RUL prediction methods primarily depend on extensive labeled source and target domain data for domain alignment. However, in practice, the amount of labeled data is limited, and these data inevitably exhibit low-quality attributes due to anomalies or missing values, leading to an increase in the distributional differences of degraded data under variable working conditions. This situation makes the RUL prediction methods difficult to obtain satisfactory domain alignment and generalization performance. To address this challenge, a new self-supervised dual-path meta alignment network is proposed for RUL prediction under limited data and unknown operating conditions. In the dual-path structure, the model utilizes Bayesian inference to learn degradation features and treat them as hidden variables, thus enhancing the model’s ability to quantify the uncertainty of sharing knowledge across different tasks. Then, the distribution of potential representations of the dual paths is aligned using a domain space alignment technique to improve the feature information for domain adaptation. In addition, the shared knowledge and task-specific information in the dual paths are integrated into a lightweight self-attention network, which allows for exploring long-term dependencies in related tasks and capturing richer domain information. Finally, extensive experiments on engine and bearing degraded data are conducted to verify the superior performance of the proposed model.