Large-scale wind turbine blade operational condition monitoring based on UAV and improved YOLOv5 deep learning model

Abstract

Monitoring the operational condition of wind turbine blades is a critical aspect of structural health monitoring, especially considering the challenges associated with traditional sensing techniques on rotating blades. This paper proposes an innovative method for monitoring large-scale wind turbine blades using an improved YOLOv5 (You Only Look Once) deep learning model that eliminates the use of manual markers. Firstly, the SE (Squeeze and Excitation) attention mechanism is added to the original YOLOv5 deep learning model to enhance the features of the target, while the pre-training model and freeze training strategy of migration learning are added to improve the model training speed and convergence efficiency. Additionally, a pre-trained model from transfer learning is integrated, along with a freeze training strategy, to expedite the training process and improve convergence efficiency. Secondly, the Deep sort algorithm is integrated seamlessly as a tracking mechanism to encode and track the targets detected by the improved YOLOv5 model. This enables the classification and coordinate output of selected targets across multiple blades, providing a comprehensive understanding of their operational condition. To validate the performance of the proposed SE_Tfreeze_YOLOv5 deep learning model, rigorous testing and assessments are conducted. The training loss, accuracy, and time of the model were evaluated and compared to several other models to demonstrate the superiority of the model. Laboratory tests were used to validate blade operating patterns at different rotational speeds, and the relationships between blade trajectories, spacing, and time–frequency information during blade operation were thoroughly discussed. To validate the practicality and reliability of the method, field monitoring measurements are performed on two 2 MW wind turbines located in a wind farm in western China. The monitoring demonstrates the capability of the vision method for remote, low-cost, high-precision, multi-point monitoring of wind turbine blades under operational conditions. The results of these monitoring are encouraging and indicate the potential of this approach for widespread application in the wind energy industry.