MPA-Net: A multi-fault diagnosis algorithm for wind turbine blades based on multi-path feature fusion

Abstract

Wind turbine blades, as critical components of modern wind power systems, are subjected to variable loads and harsh environmental conditions over extended periods, making them susceptible to surface contamination, leading-edge erosion, coating delamination, and other damage. Failure to detect these defects in a timely manner can reduce aerodynamic performance and increase operational risks. Existing detection methods often face challenges in balancing detection accuracy and model complexity in practical inspection scenes. To address this issue, this paper proposes the multi-path feature fusion network (MPA-Net) for detecting four categories of wind turbine blade defects. The backbone network integrates a multi-path feature fusion strategy with a multi-channel global residual module to enhance discriminative representation learning for heterogeneous defect patterns. The neck network introduces a partial cross-scale aggregation module, which improves fine-grained damage perception by fusing features of heterogeneous resolutions and cross-stage information. The detection head employs adaptive spatial feature fusion to dynamically align and integrate multi-scale features. Under the common five-run comparison protocol, MPA-Net achieves an mAP@0.5 of 0.906 ± 0.003, whereas YOLOv8n achieves an mAP@0.5 of 0.876 ± 0.004. The complete MPA-Net configuration contains 2.82 M parameters and requires 8.9 GFLOPs for a 640×640 single-image input. These results show that MPA-Net improves detection accuracy over YOLOv8n while retaining a compact model configuration under the reported experimental conditions.