Defect Detection Of Contact Wire Clips Of Contact Wire Steady Arm Under Unbalanced Samples

The contact wire clip of contact wire steady arm is a key component that directly connects the positioner and the contact line.It has long been subjected to the impact force of the pantograph and the catenary as well as the oscillating tension of the contact line caused by the high-speed operation of the train.Frequent external force may cause the nut of the contact wire clip to loosen or even fall off,which seriously affects the safety and reliability of the catenary system.Looseness or falling-off of the contact wire clip nut may cause the pantograph to bow,shorten the service life of the skateboard,or cause a serious pantograph-catenary accident.Therefore,a reliable defect detection method is urgently needed to realize the timely and accurate detection of the defects of the contact wire clip and avoid serious accidents.The high-speed rail power supply safety detection and monitoring(6C)system gradually replaces the traditional manual inspection for the defect detection of catenary components.However,in order to realize the accurate detection of the state of the contact wire clip,the first thing to do is to complete the precise positioning of the contact wire clip of steady arm among the 4C image.However,the size of the 4C image is large while the contact wire clip bolts occupy a small proportion in the picture.Object detection is severely difficult in this situation.Moreover,the defect samples of contact wire clip collected on site are scarce,and it is very difficult to detect defects in the face of unbalanced data sets.To solve the above problems,the main work of this paper is as follows:Firstly,aiming at the small target positioning problem of contact wire clip bolts,a cascade positioning network of contact wire clip bolts based on improved YOLOv3 is proposed.The object detection range is reduced by the cascaded positioning method.The feature prediction scale of YOLOv3 is increased to 4 scales,and K-means clustering is performed on the size of the prior frame,which improves the effect of the network on small object positioning.The experiments show that: the cascade network has a good detection rate and accuracy rate for the contact wire clip bolts.Secondly,in view of the problem of scarcity of samples of contact wire clip defects,a detection method of contact wire clip defects based on geometric transformation selfsupervised learning is proposed.Firstly,the samples are preprocessed to enhance the image quality.Secondly,the influence of different geometric transformation pseudo-label construction methods on the defect detection performance is explored,and the optimal pseudolabel construction method for contact wire clip defects is determined.Experiments show that the method in this paper can effectively improve the detection rate and accuracy of contact wire clip defects without using any defect samples to participate in training.Finally,aiming at the imbalance of positive and negative samples in the detection of contact wire clip defects,a multi-task detection method for contact wire clip defects is proposed.The method in this paper constructs two tasks,supervised learning and selfsupervised learning,which respectively uses labeled samples and unlabeled samples to participate in network training.The two tasks share the same parameter sharing space and the parameters of the sharing space are updated by alternate training.Experiments show that the detection method of contact wire clip defects based on multi-task learning can improve the generalization performance of the network and effectively improve the reliability of the network for the detection of contact wire clip defects.