Defect Detection Of Track Components Considering Sample Imbalance And Model Acceleration Research

By the end of 2021,the national railway operating mileage has exceed 150,000 kilometers,and the urban rail transit operating mileage reached 8,708 kilometers.The rail transit plays an increasingly important role in the modern comprehensive transportation system,which leads to an increasing demand for the safety detection of the track infrastructure.Rails,sleepers and fasteners are important components of the track,and their service status directly affects the operation safety of the train.How to detect the defects of each component in advance through effective detection methods is the key research direction of track inspection.However,the existing detection technology still has problems such as low recognition rate and slow speed.Therefore,this thesis takes the track components as the research object,studies the method of defect detection based on deep learning,and focuses on solving the problems of sample imbalance and model complexity in defect detection,to achieve real-time and accurate edge defect detection.The main work of this thesis is as follows:First of all,the accurate recognition and location of track components is the premise and foundation of high quality defect detection.Therefore,based on the discussion of the main target detection algorithm and the instance segmentation algorithm,this thesis proposes an optimal algorithm which is suitable for the recognition and location of track components,the model is accelerated by the pruning algorithm based on BN layer coefficient and the TensorRT inference engine to realize the accurate and fast component location.Secondly,in order to solve the problem of unbalanced samples,this thesis analyzes the defect characteristics of sleeper,fastener and rail,then proposes the corresponding defect detection methods:1)Aiming at the characteristics of unbalanced sleeper defect samples and subtle sleeper crack defects,a sleeper defect detection method based on deep ensemble learning is proposed.A variety of deep networks are used to expand sleeper defect characteristics,and an integrated classifier is constructed to share individual risks and improve the accuracy of sleeper defect detection;2)In view of the scarcity of fastener defect samples and the obvious geometric features of fasteners,a defect judgment criterion of instance segmentation+geometric features is proposed.The instance segmentation is responsible for locating the fasteners,and then extracting the fastener mask for contour search and calculating the pixel length for comparison.Quantitatively analyze the pixel length difference of the fasteners on both sides,determine the fastener defects in a quantitative way,which suppress the negative impact caused by the scarcity of defective samples;3)Aiming at the characteristics of unbalanced rail defect samples and easily mixed defect categories,a cascade detection network of instance segmentation+defect sub-classification is proposed to achieve dual-resolution and high-resolution rail defect detection.In the instance segmentation stage,the Focal Loss category loss function is introduced to balance positive and negative samples,and the lightweight classification network MobileNet v2 is used in the defect subdivision stage to achieve fast classification.This method effectively improves the detection rate of rail defects and the accuracy of subdivision classification.Finally,the track components defect detection algorithm proposed in this thesis is deployed to the Jetson AGX Xavier embedded device.The final experiment shows that the method in this thesis has higher detection accuracy and faster detection speed,which can meet the requirements of real-time defect detection.