PRA-Flow: A Dynamic-Window Transformer Network for Optical Flow Estimation in Offshore Wind Turbine Maintenance

Offshore wind power generation is recognized as an indispensable component of renewable energy systems. However, growing concerns regarding operational safety have emerged with the increasing deployment of offshore wind turbine foundations. Current maintenance methodologies predominantly rely on vessel-based approaches, where technicians access turbine equipment via gangways, rendering the secure transfer of personnel onto turbine structures a critical operational requirement. Conventional stabilization platforms for offshore wind operation-and-maintenance vessels predominantly employ six-degrees-of-freedom platforms controlled by classical control strategies, yet these mechanical systems frequently encounter limitations in response speed, positioning accuracy, and cost-effectiveness. To address these challenges, a novel Propagation-aware Residual-corrected Adaptive Flow Estimation (PRA-Flow) optical flow estimation method based on the GMFlow algorithmic framework is proposed. The methodology leverages the inherent advantages of feature matching combined with computer vision techniques, where global image correspondence is achieved through Transformer architectures. Three core enhancements are implemented: dynamic window partitioning, occlusion-aware propagation, and recurrent iterative refinement of the baseline algorithm. These innovations not only improve estimation accuracy but also maintain computational efficiency. For comprehensive evaluation, the WPOF+ dataset is constructed, encompassing four complex operational scenarios (offshore maintenance, base commissioning and maintenance, gangway maintenance, and nearshore maintenance) and five evaluation conditions (reference maritime scenes and four synthetic conditions: Dark, Snow, Fog, and Rain). Comparative experiments and ablation studies demonstrate that the proposed method achieves an overall mean EPE of 1.878, compared with 1.891 for RAFT, corresponding to a 0.70% reduction, while ranking second in the runtime comparison. These results demonstrate the effectiveness of PRA-Flow in optical flow estimation under complex offshore environments.