A Transformer-based Domain Generalization Method for Equipment Remaining Useful Life Prediction

Abstract

Remaining Useful Life (RUL) prediction is crucial for the maintenance decision-making and operational safety of nuclear power systems. However, neural-network-based RUL prediction methods that emerged in recent years face the risk of insufficient generalization when confronted with operating-condition drift caused by power generation peak-shaving, individual differences among components, and so on, which is inconsistent with the high-safety and high-reliability requirements of nuclear power. To address this problem, and also to break through the limitation of conventional domain adaptation methods that rely on target-domain data, this paper proposes a Transformer-based domain generalization method for RUL prediction of nuclear power equipment. The core ideas of the proposed model are “paired learning” and “paired querying.” During training, the model uses a cross-attention mechanism to pair-wise learn the positional correspondence between local monitoring signals and the global degradation history to which they belong. During inference, the model pairs the online-monitored local signal one by one with every historical global signal in the training set and queries them, and computes similarity weights based on the Dynamic Time Warping (DTW) algorithm to obtain the final weighted prediction. Cross-domain generalization experiments on the C-MAPSS dataset show that the proposed method, without requiring any target-domain data, significantly outperforms multiple domain adaptation methods that do require target-domain data. This demonstrates that the extracted features, which embed positional information, possess stronger generalization capability and are well suited to nuclear-power scenarios with stringent safety and reliability requirements.