Study On Train Precision Component Defect Detection System Based On Vision Analysis

The defects of train precision components will seriously affect the safe operation of rolling stock.The manual image discrimination methods commonly used today are extremely resource-intensive and have the potential for missed and false detections.With the increasing scale of the road network and vehicle ownership,the development of automatic defect detection systems is on the line.However,the extremely small size of precision components compared to a complete train body makes it difficult to locate them precisely.And the data will be significantly different with the change of acquisition stations and train models,which requires high robustness.The diversity of component types and the uncertainty of defects also make it impossible to establish an effective defect database.Meanwhile,the detection algorithm should also have a high processing speed to ensure timely maintenance and the overall system should be easy to deploy to ensure convenient application.Within the above context,in this paper a multi-type train precision component defect detection system under the framework of CLDD(Component Location and Defect Diagnosis)is proposed based on vision analysis.The main work of this thesis is as follows:(1)An image database of CR400 BF electric multiple unit series is established and a CPA(Component Pre-location Algorithm)based train precision component location network is proposed.Thanks to the consideration of visual integrity of components,the network enables the overall system to improve the F1-score of end cover,brake pad and cover plate components defect detection by 7.30%,8.41% and 3.81%,respectively.(2)A data augmentation scheme for each viewpoint is designed to satisfy the robustness of the system to multiple acquisition stations and a domain adaptive based object detection model training system is proposed to satisfy the robustness of the system to multiple models.(3)An image self-generation based defect diagnosis algorithm for train brake pad is proposed.By self-supervising the generation of defective samples and adding them to the training of the classifier,the algorithm results in an average improvement of 4.81% and5.24% in F1-score and AUC(Area Under Curve).In this paper,we also propose ISGAN(Image-Similarity Generative Adversarial Networks),which is applicable to the defect diagnosis of multi-type components.By automatically generating matching templates for the input images and using the similarity between them as a basis for defect diagnosis,the network achieves F1-scores of 84.55%,71.69%,and 88.22% for end cover,brake pad,and cover plate components,respectively.That means the network has the ability of automatic defect diagnosis for a variety of train precision components.(4)Channel pruning is performed to enrich the applicability of the system under the same hardware conditions.After compression,the number of model parameters and model volume are only 3.5%～12.5% of the original,the GPU memory consumption is only 50% of the original,the inference speed becomes 116% of the original,and the m AP-50 only decreases by 0.0004～0.0162.In summary,this paper carries out model compression and system deployment optimization on the basis of the train precision component location network and component defect diagnosis algorithm,establishes a defect detection system for train precision components,and designs a visualization interface.The system can complete the detection of a train with 16 carriages in 42.34 seconds and can detect very small defects on components with a size of only 3.5(?) of the original image size.There are 40 Figures,20 Tables and 73 References.