Multiscale Feature Extension Enhanced Deep Global–Local Attention Network for Remaining Useful Life Prediction

With the development of sensing techniques, multisensor long-sequence monitoring signals are available to provide abundant information from sensory and temporal dimensions. It raises high demands for remaining useful life (RUL) prediction methods on effectively mining degradation information from enormous data. However, sensor sensitivity variations under different temporal scales and degradation stages are not explicitly considered. The local feature representation and its adaptive integration with global dependencies remain unexplored. These deficiencies limit the representational power and prediction accuracy of the existing prognostic models. Therefore, a multiscale feature extension enhanced deep global–local attention network (MS-DGLAN) is proposed for RUL prediction. First, the designed multiscale dynamic time warping feature extension dynamically identifies multisensor degradation sensitivity and fully exploits transient and temporal-average signal features. Furthermore, the deep global–local attention network (DGLAN) achieves parallel extraction and adaptive interactivity of coarse-to-fine level features, leading to complete high-level representations. Hereinto, the developed multidimensional attention and local-focused mask endow DGLAN with featurewise long-term dependency modeling and content-dependent detail perception abilities, respectively. MS-DGLAN synergistically considers salient sensory information identification and global–local temporal feature integration, thus processing multisensor long-sequence data more effectively. Moreover, it has high parallel computing efficiency and is more interpretable for its fully attention-based structure. Case studies on four turbofan engine datasets are conducted to demonstrate the superiority of the proposed method.