UAV-based wind turbine blade inspection using ECMRI-YOLO11: addressing environmental complexities through advanced detection efficiency

We present a lightweight and high-precision multitarget defect detection model for wind turbines. The model is suitable for different light intensity irradiation or multi-angle shooting and can also distinguish among different types of blade defects. The model, named Efficientformerv2 backbone, CARAFE and mixed local channel attention (MLCA), RFCBAMConv, and Inner-IoU–You Look Only Once (ECMRI-YOLO), is designed for wind farm defect detection. The model addresses the lack of accuracy in blade defect identification due to complex weather conditions and differences in light intensity resulting from unmanned aerial vehicles (UAVs) shooting from multiple angles. It enables real-time detection speed on a lightweight basis. The efficacy of the model is substantiated by a series of rigorous ablation and comparison tests with leading algorithms in the field. The number and positioning of the MLCA mechanisms are optimized, ensuring precise identification of defects. The convolutional structure of RFCBAMConv is shown to surpass the accuracy of other literature convolutional structures through the utilization of the control variable method. The enhanced model enhances accuracy while concurrently reducing model parameters and floating point operations. The experimental findings demonstrate that the mAP50 score of the enhanced model reaches 86.1%, representing a 3.2% enhancement over the baseline model, and the detection speed is augmented by 24.2%. In addition, the model’s performance in complex lighting scenarios is notably enhanced, both for single and multiple targets, near and far views. The prediction map exhibits improved accuracy in displaying target defects, and the severity of false and missed detections is addressed. These observations collectively indicate that the model possesses the potential to be implemented in an intelligent UAV inspection system for wind turbine blades.