Research On The Inspection Algorithm Of Railway Fasterner Based On Distribution Learning

The railway fastener is the part of the rail fixed on the rail pillow,which is used to ensure the safe operation of the train.If fasteners are lost or broken,it will lead to the occurrence of major accidents,so it has great significance to detect the state of the fastener.Traditional inspection of railway fasteners is carried out by manual inspection.However,with the rapid development of Chinese railway industry,manual inspection cannot meet the daily inspection of fastener.Computer vision-based inspection of fasteners has been paid more attention.Based on computer vision technology,this paper studies fastener detection.The main research contents are as follows:(1)Aiming at the problem that the existing tag distribution construction method cannot be directly used in fastener detection,a semantic polynomial construction method of fastener based on weakly supervised learning was proposed.Firstly,the convolutional feature was extracted by fine-tuning the convolutional neural network.Then,according to the feature size,molecular blocks were divided and the sub-block features are represented as a Gaussian mixture model.Finally,semantic polynomials were calculated by using the sub-block Gaussian mixture model as the label distribution of the fastener image.According to the visualization experiment of the distribution of fasteners,the rationality of the distribution of fasteners label constructed by this method proposed in this paper was verified.Theoretical analysis and experiment show that this method can express the semantic information of fastener image effectively and has high description accuracy.(2)Aiming at the problems of weak adaptability and high false detection rate of common fastener detection algorithms,combined with the above construction method of fastener label distribution,a learning model of fastener state distribution based on hierarchical semantic polynomial DS(Dempster Shafer,DS)fusion was proposed.Based on the analysis of the semantic differences described by the convolution features of different dimensions,the model constructs multi-layer semantic polynomials by referring to the image semantic model,then fuses the multi-layer semantic polynomials by DS evidence theory,finally obtains the state distribution of the fastener.The experimental results show that the sample distribution of the fastener fusion with multi-layer semantic polynomials was more accurate than that with single-layer semantic polynomials,the semantic information of the image was more complete,the over-fitting phenomenon is reduced.The false detection rate was 1.9%,and the missed detection rate was 2.3%.Theoretical analysis and experiments show that,compared with other fastener detection models,the proposed model can effectively reduce the misdetection of fasteners and has strong adaptability.