Abstract
Predictive health management (PHM) plays a pivotal role in the maintenance of contemporary industrial systems, with the evaluation of the state of health (SOH) and the prediction of remaining useful life (RUL) constituting its central objectives. Nevertheless, existing studies frequently approach these tasks in isolation, overlooking their interdependence, and predominantly concentrate on single-condition settings. While Transformers have demonstrated exceptional performance in RUL prediction, their substantial parameter requirements pose challenges to computational efficiency and practical implementation. Further, multi-task learning (MTL) models often experience performance deterioration as a result of imbalanced weighting in their loss functions. To address these challenges, the MID-1DC+LRT model was proposed in the present study. The proposed model integrates a multi-input data 1D convolutional neural network (1D-CNN) and low-rank transformer (LRT) within an MTL framework. This model processes high-dimensional sensor data, multi-condition data, and health indicator data, optimizing the Transformer structure to reduce computational complexity. A homoscedastic uncertainty-based method dynamically adjusts multi-task loss function weights, improving task collaboration and model generalization. The results demonstrate that the proposed model significantly outperformed existing methods in SOH assessment and RUL prediction under multi-condition scenarios, demonstrating superior prediction accuracy and computational efficiency, especially in complex and dynamic environments.