Research On X-Ray Image Defect Recognition Methods For Transmission Lines Based On Deep Learning

In the power industry,the quality of strain clamps in overhead transmission lines is directly related to the safety of transmission lines.How to ensure the quality and safety of strain clamps has become an u rgent problem to be solved.With the development of X-ray non-destructive testing,X-ray technology to collect digital images of tension clips for non-destructive testing has gradually been widely used.However,the obtained X-ray images still need to be judged by professionals whether they are defective,resulting in low detection efficiency and easy to be interfered with by human factors.Because of these problems,this paper applies the deep learning method to t he defect detection of strain clamps to solve the drawbacks of manual inspection.The main contents include:(1)Defect annotation is performed on the acquired strain clamp X-ray images.A dataset for object detection and image segmentation tasks is made i n the corresponding format to complete the p reparation for the subsequent implementation of defect detection and defect dimension measuremen t.(2)Aiming at the characteristics that the position of crimping defects in the collected X-ray images of the strain clamps occupy a small proportion in the image and are closely arranged,a "two-step method" detection strategy is adopted,The crimping part is first extracted,and then the crimp defect is detected from the crimping site.Model comparison experiments we re carried out on the tensile wire clip data set to verify that the adopted detection strategy has the advantages of high detection accuracy and fast detection speed compared with the traditional direct detection method.(3)Research on defect size measurement method based on convolutional neural network.First,the corresponding relationship between the actual size and the image pixels is established.Then the deep learning image segmentation algorithm is used to segment the X-ray image of the tensile clip to the idea of the defective part.In this process,several traditional image segmentation methods are mainly compared and tested,and the U-Net with better effect is determined as the defect segmentation model.Based on the U-Net model,the original convolution layer is replaced by dilated convolution to improve the segmentation accuracy.Finally,the edge extraction algorithm is used to extr act the outline of the segmented defect,and the minimum circumscribed rectangle method is used to measure its size,which provides data reference for further r epair work in the project.In this paper,the deep learning method is applied to defect detectio n and defect size measurement of strain clamps.According to the characteristics of the collected data sets,the detection method and the structure of the bench mark model used are improved,and the "qualitative," "Positioning," and "quantitative" tasks provide a specific reference for subsequent related research.