Learning globally ordered and locally consistent degradation representations for remaining useful life prediction

Recent studies on remaining useful life (RUL) estimation have shown that deep neural networks can effectively extract informative features from sensor data, thereby improving the prediction performance. However, most existing methods rely solely on direct mapping between labels and data to construct the feature space, while ignoring the exploration of feature relationships. This study believes that strongly generalized degradation features should have two properties: global orderliness and local consistency. The former stems from the irreversibility of the degradation process, while the latter reflect the stability of the system state in a short period of time. In this work, a globally ordered and locally consistent representation learning (GOLCRL) method is proposed for RUL prediction. GOLCRL extracts degradation representations using stacked convolutional neural networks, integrating multi-scale convolution and channel attention mechanism to facilitate the information interaction across spatial and temporal dimensions. To refine the ordered relationships, GOLCRL regularizes the geometric structure of the feature space through supervised group contrastive learning and correlation-aware distribution alignment. Moreover, GOLCRL guides the label smoothing of neighboring samples in the feature space through a pseudo-labeling strategy, mapping them to a more coherent label region, thereby enhancing local consistency. Two case studies demonstrate that GOLCRL outperforms existing methods in terms of generalization capabilities, achieving more accurate RUL prediction results.