Measurement technology of wind turbine blades infrared defect with human feedback introduced
In the process of achieving unmanned and intelligent inspection of wind turbine blades,the preparation and labeling of data sets seriously affects the prediction accuracy of deep learning networks.This is particularly important in unmanned airborne infrared-based thermal imaging,unlike visible light inspection:the limited pixels of the infrared camera lead to problems such as fewer pixels on the inspection image,small temperature difference,and defect temperature characterization varying by defect type.The accuracy of network recognition relying on only one data calibration is low,and the difference in data set labeling brings serious impact on recognition accuracy.To this end,a human-in-the-loop-based method for intelligent infrared defect detection and measurement of wind turbine blades is proposed,which introduces human feedback into the network performance iteration and solves the problem of temperature characterization varying with species in defect area measurement by defect localization assumption.The experiments show that the introduction of human participation can effectively improve the recognition rate of wind turbine blade defect detection.In the homemade infrared dataset of wind turbine blades,The mAP@0.5 of the second data iteration increased by 22.94%compared with the first data annotation,and the mAP@0.5 of the third data iteration increased by 27.8%compared with the second data iteration.
NETL
NSTL
万方数据
维普
