Multi-granularity Cross-Domain Temporal Regression Network for Remaining Useful Life Estimation of Aero Engines

Remaining useful life (RUL) prediction is a crucial task for predictive maintenance of industrial equipment. Benefiting from advanced sensing and artificial intelligence technologies, data-driven RUL prediction methods based on multimodal data analytics achieved rapid development in recent years. However, traditional RUL prediction methods often fail to meet the demand and challenge of data distribution discrepancy under different working conditions. To solve this issue, a novel aero-engine RUL estimation approach is proposed based on multi-sensor fusion and deep transfer learning. First, a multi-granularity cross-domain temporal regression (MCDTR) network is constructed to learn effective degradation information via fusing a coarse-grained learn strategy executed on the source domain and a fine-grained update strategy applied to the target domain. With such a multi-granularity transfer strategy, this network can exploit robust temporal features for accurate RUL prediction. In addition, the uncertainty quantification of predictive results based on the bootstrap method is also examined to improve the reliability and stability of RUL prediction for industrial aero engines. Related comparative experiments on the N-CMAPSS 2021 Challenge Dataset suggest the effectiveness and robustness of the proposed approach, which provides a valuable reference for prognostics and health management in industrial applications.