A sub-regional active learning-based hybrid network framework for remaining useful life prediction of offshore wind turbine gearboxes

A sub-regional active learning-based hybrid network (SRAL-HNet) framework is proposed for remaining useful life (RUL) prediction of offshore wind turbine gearboxes. First, a multi-scale sliding window strategy is employed to extract both short-term fluctuations and long-term degradation trends from the raw monitoring data. Then, the random forest is utilized for sensitivity features selection to eliminate redundant information. Subsequently, a bidirectional long short-term memory (BiLSTM) encoder and a fully connected (FC) regression head are combined to construct the BiLSTM-FC regression network. By combining health-index monotonicity analysis with percentile-based threshold segmentation, a sub-regional active learning method is developed to partition the sequence data into distinct regions. Consequently, the SRAL-HNet method is constructed for RUL prediction by integrating the sub-regional active learning model with the BiLSTM-FC regression network. To validate the feasibility and superiority of the proposed method, an offshore wind turbine SCADA dataset is used. The results demonstrate that the proposed model achieves accurate RUL predictions with a significant reduction in the number of samples required. Overall, the proposed framework is able to support offshore condition-based maintenance, risk-informed inspection planning, and life-extension decision-making in marine renewable energy systems.