Abstract:Tower barrel system (including tower barrel, bolt) as an important basic component of the normal operation of wind turbine, it is very important to effectively identify cracks and other faults that affect its safety. In view of the obvious crack representation, low recognition and poor contrast, this paper proposed an improved YOLOV7-Seattention algorithm model based on YOLO series algorithms, and compared it with a variety of algorithm models such as Faster R-CNN, RFCN, SSD, YOLOv5, YOLOv7, etc. The three indexes of Recall, Precision and Average Precision were comprehensively evaluated. The results show that the improved YOLOV7-SEattention model shows significant advantages in the surface crack detection of the tower barrel system and has higher accuracy and reliability than the original YOLOv7 and other algorithm models in the crack detection of the fan tower barrel system. The average precision (AP) of tower barrel crack detection was increased by 2.6% to 83.7%, and the average accuracy of bolt crack detection was increased by 4% to 84.3%. The improved model in this paper can accurately and efficiently detect the surface cracks of the tower system, reduce the operation and maintenance cost, and improve the benefit of the wind farm.