Research On High Speed Rail Fastener Detection Algorithm Based On Deep Learning

Railway fastener is an important components to ensure the safe operation of trains.With the rapid development of China’s high-speed railway technology,designing a system that can realize the automatic detection of track status has become a strong demand from railway related departments.In recent years,more and more automatic detection algorithms based on computer vision have been proposed.Compared with the traditional manual inspection method,the fastener detection method based on computer vision is faster and more accurate.However,due to changes in lighting and various failure modes of fastener images,it is difficult for the features extracted from traditional visual algorithms to describe the true content of the image.This thesis applies deep learning to the positioning and status recognition of fasteners.The main research contents are as follows:(1)Aiming at the problem of slow speed and low accuracy in fastener positioning by traditional visual methods,an improved Single shot multibox detector deep learning fastener positioning algorithm Improved＿SSD is proposed.The algorithm first uses the residual network to update the original feature extraction network,which increases the depth of the network while improving the ability to capture features.Second,it uses an expanded convolution method to expand the receptive field of the network and increase the robustness of the algorithm Third,the original suppression method is replaced by the non-maximum weighted suppression method designed in this paper to improve the accuracy of the final output frame position.Finally,the position verification method is used to quickly determine whether there are missing fasteners in the image.Theoretical analysis and experiments show that Improved＿SSD effectively reduces the missed positioning rate and the mislocated rate,the positioning speed is faster,and it has strong robustness to light changes,which can meet the practical engineering application needs of high-speed fastener positioning.(2)Aiming at the problem that the existing fastener recognition algorithm is difficult to accurately distinguish the failure types of fasteners,a high-speed rail fastener state recognition algorithm FLDF that combines local features and depth features is proposed.The local features of the algorithm are extracted using the improved SIFT operator DR＿SIFT combined with the bag-of-words model,and the deep features are extracted using the VGG-16 network.First,determine the number of words in the bag-of-words model and the specific feature layer in the deep network model through experiments;second,fuse feature extraction methods on the basis of the two models.During the fusion process,the weight ratio experiment is used to Determine the weight of the fusion of the two features,and normalize the feature vectors at the same time;finally input the feature vectors into the vector machine for classification.The new algorithm not only retains the characteristics of the SIFT local feature extraction process that does not deform the brightness changes,and has a unique feature,but also complements the deeper feature of deeper semantic analysis of fasteners in the global feature.Theoretical analysis and experiments show that FLDF can effectively identify the types of defective fasteners.Compared with mainstream fastener state recognition algorithms,the fastener detection algorithm in this paper has a higher accuracy and is more robust to illumination.