Research On Defect Detection Algorithm Of Catenary Fasteners Based On Deep Learning

Maintaining the good working condition of the catenary and its auxiliary components is the basic guarantee for the safe operation of high-speed trains.Since various positioning bearing supports of the catenary are continuously affected by electrical shock and mechanical stress,this greatly increases the probability of damage to their components.If the positioning bearing support fails,it will directly damage the catenary equipment and even cause high-speed trains the sudden stop and threaten the safety of the people in the carriage.The timely detection of defects on the contact network is essential to ensure the safe operation of the railway.Existing methods for detecting contact network defects can be divided into: manual inspection,non-contact physical inspection and non-contact image technology inspection.However,manual inspection is not only time-consuming but also highly dependent on the inspection experience of the inspection personnel.It is difficult to detect subtle defects by non-contact physical inspection,and it is difficult for these two methods to achieve satisfactory defect detection results.This article makes full use of the combination of image processing technology and deep learning to develop a method of using image processing technology based on deep learning to solve the problem of fastener defects in high-speed rail catenary.Firstly,the method of acquiring high-speed rail catenary image data,the characteristics of the image,and typical defect categories are analyzed.For the collected catenary image quality,there are about 25% heterogeneous images.The heterogeneous image rejection method saves computer resources and improves the detection rate for subsequent defect detection.Aiming at the imbalance of positive and negative samples caused by few defective samples,this paper proposes four effective ways of increasing negative samples.Secondly,to analyze the particularity of defect detection of catenary fasteners,which is mainly reflected in the fact that the defects are often attached to specific components.Defects,the experimental results show that the cascaded network structure is superior to other comparison algorithms in detecting defects of high-speed rail catenary fasteners.The detection accuracy is improved from,for example,59.8% of faster RCNN to 86.7%.Then,in view of the small visual difference between different defects of the same kind of components,the defect feature interface is not obvious,a inverse attention mechanism is proposed and combined with the feature pyramid module,the focus feature extraction method is studied,and the defect area features are enlarged and focused The experiment proves that the inverse attention module effectively amplifies the difference between defects and improves the accuracy of defect detection by 0.7%.Finally,the small and medium targets in the fastener defect detection are analyzed to be sensitive to the intersection-overunion,and an elliptical detection label is proposed to replace the usual rectangular frame label.The experiment verifies that the new detection label proposed in this paper does improve the accuracy of the detection of fastener small parts degree.