Statistical Alignment-Based Metagated Recurrent Unit for Cross-Domain Machinery Degradation Trend Prognostics Using Limited Data

Nowadays, machinery degradation trend prognostics have made significant progress with sufficient and related data. Limited data may weaken the generalizations of artificial intelligence-based prognostics, which could worsen under the domain shift phenomenon. To overcome the dual challenges and further abstract the prior degraded knowledge from historical data, a novel prediction architecture entitled as statistical alignment-based metagated recurrent unit, profoundly integrating deep learning, knowledge transfer, and metalearning, is proposed in this study and simultaneously features domain adaptation ability and desired generalizations under limited data. To be specific, a designed high-order measure is first developed to align training and testing data distribution. Then, the well-known encoder-decoder framework is employed as the base model, and subtask & cross-subtask learning architectures are developed to distill sensitive prognostics agents for easily generalizing to new domains using limited data. Finally, FEMTO-ST bearing data sets and industrial data sets collected from petrochemical pumps are conducted to prove their effectiveness.