Research On Detection Method Of Railway Diseased Fastener Based On Visual Image Processing

Fastener is the key component and important infrastructure of railway,which plays a very important role in mmaintaining the safety of train operation.However,with the long-term wheel and rail vibration and the influence of the external environment,fastener will appear damage and missing and other diseases,wh ich may be lead to major safety accidents.Therefore,it is necessary to conduct regular detection of fastener state to ensure the safety of railway operation.Based on the machine vision inspection technology and platform,this paper conducts research on the automatic detection method of diseased fasteners,focusing on the key problems and technicial difficulties faced in the process of achieving automatic detection.The main research results and innovations in the whole paper are listed as follows:(1)The basic idea and process of vision technology-based detection of diseased fastener is introduced,and it is pointed out that the existing methods of diseased fastener detection mainly follow the two-stage detection procedure and fast detection procedure to complete the detection task.The diseased fastener detection platform with a line-scan camera and a near-infrared laser light source for linkage imaging is described,as well as the indoor and outdoor testing environments.The practical problems and difficulties faced in realizing the automatic detection of diseased fasteners are analyzed,which lead to the research contents of the subsequent chapters.(2)For the problem that the vehicle vibration causes the deformation of railway image and thus affects the detection accuracy of the diseased fasteners,this paper proposes a deformed railway image correction method based on the consistency of imaging ratio.The proposed method adopts the micro-element idea,firstly obtains the sub-region images by U-Net segmentation,and then determines the deformation information by template comparsion.Finally,the correction task is completed based on the deformation information and deformation regions.The experimental results show that the proposed method has good correction performance for deformed railway images,and its correction performance has good stability for deformed railway images acquired under different railway lines and environments.(3)To address the problem that fastener region is difficult to locate accurately due to the polymorphic railway lines and complicated environments,this paper proposes multi-scale feature fusion-based detection network for fastener region localization.The proposed detection network is achieved by opt imizing Faster RCNN and which includes three aspects: firstly,the multi-scale feature fusion network is proposed to improve the feature learning capability of the detection network;secondly,K-means clustering algorithm is used to automatically determine the anchors to ensure the generated anchors are more similar to the real fastener regions;thirdly,the classification and regression module is removed from the detection network to speed up the localization.The experimental results show that the proposed detection network can accurately locate fastener regions in real railway lines,and its localization performance is stable and better than related methods in the literatures.(4)For the problem that the imbalanced between defective and normal fasteners affects the detection performance of diseased fasteners,this paper proposes an image generation-based detection method for diseased fasteners.The proposed method accomplishes detection task through two steps: diseased fastener image generation and fastener state recognition.For the image generation,a Four Discriminator Cycle-consistent Generative Adversarial Network is proposed,which fully combines the advantages of dual adversarial generative network and cycle consistent generative adversarial network to effectively solve the model collaps e caused by insufficient diseased fastener images,so as to generate high-quality diseased fastener images and realize the expansion of diseased fastener samples.For the fastener state recognition,it is achieved by using VGG16.The experimental results show that the proposed method can effectively solve the problem of imbalanced fasteners and improve the detection accuracy of diseased fasteners.(5)In order to fundamentally solve the problem of imbalanced between diseased and normal fasteners affects the detection performance of diseased fasteners,a diseased fastener detection method based on key region recognition is proposed in this paper.The proposed method first obtains the key region sub-images of fasteners based on the prior knowledge,and then recognizes the categories of key regions through the constructed region classification network.Finally,the detection of diseased fasteners is achieved by analyzing the recognition results using the built decision tree.The experimental results show that the proposed method can achieve stable and effective detection of diseased fasteners with only normal fasteners,while achieving excellent detection performance.This proves the effectiveness,reliability and thoroughness of the proposed method in solving the p roblem of imbalanced fasteners affecting the detection performance of diseased fasteners.(6)For the problem that poor performance in fast and real-time detection of diseased fasteners,this paper proposes a fast detection method of diseased fasteners based on the lightweight network.The proposed lightweight network is achieved by optimizing YOLOv4-tiny.Specially,the optimization consists of two aspects.Firstly,a new feature extraction module Res Block-N is proposed to replace the CSPBlock module in YOLOv4-tiny,which can further reduce network computation and improve the detection speed.Secondly,a large scale feature map(52 × 52)is added for prediction output,which can fusion the more features to ensure detection accuracy.The experimental results show that the proposed lightweight network can achieve fast and accurate detection of diseased fasteners,while its detection performance has good stability and is better than other state-of-the-arts.In summary,this paper aims at the actual problems in the p rocess of disease fastener detection for the real railroad scenario,mainly focusing on four aspects,i.e.,railway image deformation correction,fastener region accurate positioning,imbalanced disease fastener accurate detection and disease fastener fast detection and proposes the corresponding optimization methods.The proposed methods can effectively improve the accuracy,reliability and timeliness of the disease fastener detection.Meanwhile,they have greater theoretical significance and practical val ue for promoting the intelligent level of track inspection.