Research On Key Technologies Of Intelligent Detection Of Steel Rail Surface Defects Based On Machine Vision

Rail transportation serves as an essential form of public transit,and the failure of railway tracks can lead to severe disruptions in train operations,jeopardizing the safety of passengers and their belongings.Therefore,timely intervention in the early stages of failure occurrence is crucial.Accurate and reliable real-time detection of surface defects on steel tracks,along with early intervention to suppress defect progression,represents an effective approach to prevent railway accidents.Currently,two widely adopted methods for inspection include manual inspection and track vehicle patrols.However,these existing methods suffer from issues such as poor conditions for nighttime inspections,low inspection frequency,and constraints imposed by complex and extreme weather conditions.Additionally,existing defect detection systems still face challenges such as limited coverage range,inadequate fault warning,complex data processing and analysis,and excessive reliance on manual assistance.These systems are subjective,and the detection results heavily depend on the experience,technical expertise,physical and mental condition,and work environment of the inspectors.To overcome these challenges,this study proposes a deep learning-based intelligent detection system for railway track surface defects and investigates key technologies involved.The system achieves accurate extraction of tracks under various challenging conditions,defect classification,precise defect segmentation,and predictive identification capabilities,even in the presence of strong noise.Furthermore,real-world testing of the system is conducted,and experimental results demonstrate the practicality and effectiveness of the proposed methods.The main contributions of this study are summarized as follows:(1)A machine vision and neural network-based method is proposed for accurate extraction of rail defects under complex noise backgrounds.The method effectively identifies rails by combining morphological filtering and probabilistic Hough transform algorithms to eliminate significant noise in the original images.It also addresses the issue of excessive false edges generated by the Canny algorithm during rail edge extraction.The proposed method sequentially applies thresholding and discrete methods to obtain the true edges of the rails.Additionally,an improved cross-entropy loss function,considering both recall rate and precision rate,is designed.A highperformance rail surface morphology classifier(Track CNN)is established based on a convolutional neural network.The results of the field image experiments show that the proposed algorithm in this paper has a good performance in the comprehensive aspects of detection efficiency,real-time,accuracy and robustness.(2)A pixel-level segmentation network called Rail Surface Defect Pixel-Level Segmentation Network(RPLSN)is proposed to address real-time and accurate segmentation of rail surface defects under complex operating conditions and lighting variations.The method concatenates features in the channel dimension to form thicker features,facilitating the propagation of more texture information of rail defects in highresolution layers.Dropout is applied to weaken weak correlations learned during the convolution process,allowing the convolution blocks to share a set of weights,reducing redundant calculations and model complexity.Performance experiments are conducted under multiple models,datasets,and operating conditions to demonstrate the adaptability and superiority of the proposed model.(3)Given the varying service years of rail surfaces and the dispersed distribution and diverse forms of different surface defects,the recognition performance is susceptible to environmental factors such as weather and lighting.To address these challenges,a recognition and prediction framework for rail defect risk level assessment(RLRNM)is proposed.An improved residual network model is proposed for rail defect feature extraction,and multiple trainings are conducted to enhance stability and robustness.A comparison is made between the proposed model and popular feature extraction networks in terms of accuracy loss value,and root mean square error(RMSE).Performance experiments are conducted under multiple models,datasets,and operating conditions,demonstrating the advancement of the proposed method.(4)To meet the requirements of daily safety monitoring and management of rail transportation equipment,an intelligent detection system for rail surface defects is designed.The system comprises several components,including the rail surface defect detection module,online updating module for detection models,train data interaction module,rail defect localization system,power supply module,and anomaly handling module.Field and on-site tests are conducted on the system and its corresponding devices.Detailed explanations are provided for the key components of the system,and the software and hardware implementation of the deep learning models are described.In summary,for the problems of defect extraction,segmentation and recognition prediction of multi-sample and multi-working condition steel rail samples,the research is carried out in terms of visual localization,sequential dynamic thresholding,sample classification network,accurate segmentation network and defect recognition prediction network,respectively.The methods and passed the experimental validation.The obtained results provide a certain basis and reference value for the research direction based on visual defect detection.