Contact Wire Clamp Nut State Recognition Method Based on Improved YOLOv7 Algorithm
A modified you only look once version 7(YOLOv7)algorithm is proposed to address the issues of poor performance of traditional deep learning algorithms in dealing with small targets such as suspension wire clamp nuts in railway overhead contact systems,as well as slow and difficult manual inspection of suspension wire clamp nuts.Firstly,the algorithm integrates convolution and self attention mechanisms at the end of the feature extraction net-work of the original you only look once version 7 model,giving the algorithm the benefits of attention and convolu-tion,as well as the minimum computational complexity compared to pure convolution or attention,improving the speed of defect detection;Then,an space to depth convolution module is introduced at the output end of the feature extraction network,replacing the pooling layer with a spatial to depth layer and the cross row convolution layer with a non-cross row convolution layer to enhance the algorithm's ability to recognize nut defect states;Finally,a new lightweight coordinate attention mechanism for mobile networks was added to the output layer,which can generate direction awareness and position sensitivity attention maps.These maps can be applied complementarily to the input feature maps,which is more conducive to the recognition of small targets such as suspension wire clamps and nuts in the contact network.The simulation experiment results show that the algorithm has a recognition accuracy of over 90%and an average detection precision(mAP@0.5)value of 98.5%for the status of contact wire suspension clamp nuts on the uncut suspension wire dataset.The improved YOLOv7 algorithm is verified to be capable of identifying the status of contact wire suspension clamp nuts more accurately while maintaining detection speed.
catenary suspension wireidentification of the status of the suspension wire clamp nutsyou only look once version 7(YOLOv7)self attention and convolutional mixing(ACmix)space to depth(SPD)
NETL
NSTL
万方数据
维普
