Research On Migration Learning Based Image Recognition Method For Weld Defects

In weld defect detection,traditional NDT methods are mainly implemented through X-ray image detection,which is susceptible to subjective factors and has a huge determination workload,and cannot meet the actual inspection needs of automated welding lines.Deep learning has been widely introduced into automatic weld defect detection technology,as a large amount of marked defect data is difficult to obtain in practice,resulting in deep learning-based weld defect detection technology requiring a large amount of marked defect data for model training cannot be engineered.Therefore,in order to solve the above problems,this thesis proposes a migration learning based weld defect identification method,whose main research elements are as follows.(1)The weld seam acquisition process,mechanistic knowledge and operational data resources are analysed.The process and principles of weld seam data acquisition are introduced,and the mechanism of weld defect generation is analysed.The data characteristics,defect categories and defect characteristics of the captured weld defect images are analysed,and the defect images are annotated according to the specific requirements of the defect detection algorithm for the data set,and the evaluation indicators for defect detection are determined.(2)Study and analyse the characteristics of weld defect data,and then propose a weld defect detection method based on the SSD algorithm according to the analysis results.In view of the small amount of data in the defect image,the imbalance of defect categories,the large scale of defects,and the high proportion of small defects and large scale defects,this thesis uses the geometric transformation method to balance the difference in the number of defect categories,uses the Mosaic-based data augmentation method to enhance the complexity of the defect image,and uses the generative adversarial network to generate the defect image to further balance the various defect categories and increase the amount of data;on this basis Based on the SSD algorithm,the weld defect detection network is built,the attention mechanism module and feature fusion module are added to the network to enhance the feature extraction capability of the network model for small defects,and the NMS method in the network is replaced with soft-NMS to enhance the detection capability of the model for porosity defects,and finally the basic model for weld defect detection is constructed.(3)In order to further improve the detection accuracy of defect detection,two migration learning methods are proposed on the basis of the weld defect detection base model.The first method is a migration learning method based on sample selection,which complements the sample distribution by extracting samples from the source domain and adding them to the target domain,thereby improving the detection performance of the model;the second method is a model-based migration learning method,which migrates the parameters of the base detection model through a parameter fine-tuning strategy under supervised learning.The results of the experimental validation of this method demonstrate that the proposed method not only solves the problem of obtaining data from a large number of marked defects,but also effectively improves the detection accuracy of weld defects.