Research On Intelligent Magnetic Particle Testing Method For Bogie Frame Based On YOLO-CET

As a key load-bearing part of high-speed train operation,the frame is prone to fatigue fracture when operating under high load conditions for a long time.Magnetic particle testing,as an important tool for analyzing surface defects of structures,currently remains mostly in the stage of traditional manual observation of damage,with problems such as missing defects,difficulty distinguishing authenticity,and background misjudgment.Therefore,there is an urgent need for an intelligent framework defect detection method to achieve efficient and accurate identification of defects.Therefore,this article improves an intelligent defect recognition algorithm based on the YOLO-CET object detection model for flaw detection images,achieving the detection and recognition of linear cracks,bubbles,and oxidation defects in the framework.A lightweight Co TNet feature extraction network based on the CSP module is proposed to address the issues of model redundancy and recognition difficulty in handling true and false defects using the YOLO-v5 algorithm.At the same time of residual feature learning and multiscale feature fusion in the algorithm model,expand the algorithm Receptive field to complete the efficient identification of true and false defects.In response to the problem of weak feature information capture ability for small defects in the Neck end of the model,an efficient channel attention mechanism is introduced to fuse the position weight in space with the target weight in the channel,thereby accurately completing the classification and localization of the regression box and improving the global feature capture ability of small target defects.In response to the problem of weak generalization ability and poor robustness of the algorithm model,it is proposed to replace the original three types of Head heads in the YOLOv5 model with four Transformer Prediction Heads at the model output.The TPH four head structure is used to alleviate the negative impact of scale variance caused by different features in defect images,improve the recognition accuracy of small defect targets and the model generalization ability,The recognition effect of small cracks with high density concentration is good.Experimental testing was conducted using 1300 self-developed defect datasets,and the results showed that compared with YOLO-V5,the YOLO-CET model improved the average accuracy of the model by 33.8%,the F1 index by 0.26,and the floating-point computation increased by only 1.5B.This model effectively improves the problems of difficult identification of authenticity,missed detection of small defects,and background misjudgment in automatic detection,meeting the requirements of surface defect detection for bogie frames.Finally,the integrated deployment of the YOLO-CET detection model and the development of a flaw detection image detection system client have enabled intelligent and efficient analysis and recognition of structural defects.