Wind turbine blade rotational condition monitoring based on RBs-YOLO deep learning model

Abstract

Blades serve as the primary elements for converting mechanical energy into electrical energy in wind turbines, making their operational parameters crucial for structural safety evaluation. The traditional monitoring methods for the rotating blades require labor-intensive attachment of artificial markers, and also face challenges in image background processing. This paper proposes a visual monitoring method based on a deep learning model for monitoring the operational parameters of wind turbine blades in rotational conditions. Firstly, a deep learning-based image fusion algorithm is proposed, where the fusion factor is set to 0.7 to achieve precise segmentation of various components of wind turbines, enhance features, and eliminate background interference. Secondly, the introduction of a rotating bounding box detection method proposes a deep learning model, RBs-YOLO (Rotating Boxes − You Only Look Once), specifically designed for blade monitoring in operational conditions, enabling target-free detection of rotating blades with an accuracy of over 98 %. Subsequently, the Deep sort algorithm is integrated for target encoding, enabling full-field displacement output for different blades. Test validation using wind turbine blades in rotational conditions at different rotational speeds demonstrates that this method has a time-domain error within 0.1 mm and achieves accurate identification of operational parameters such as blade trajectories, distance, and time–frequency domain characteristics at different rotational speeds. Finally, field monitoring was conducted on a 2.0 MW wind turbine at a wind farm in the western of China, validating the method’s capability for full-field monitoring of wind turbine blades in rotational condition without the use of artificial markers.