Research Of Intelligent Detection Method For Pin Missing State Of Transmission Tower Bolts Based On Deep Learning

Transmission tower bolt is an important part of transmission tower,which is widely used in all kinds of important connection parts of transmission tower.The falling off of bolt and nut will cause structural loosening of transmission tower parts and other problems.In order to prevent bolts and nuts from falling off,the existence of pins has a very important effect on the safety of transmission lines.The increasingly large modern transmission network brings more and more heavy transmission line inspection tasks.In addition,limited by the timeliness and economy of modern remote manual detection technologies such as UAV and helicopter,it is urgent to develop a real-time detection method for the missing transmission tower bolts and pins.This article in view of the existing power transmission tower bolt,pin detection model generalization ability,the problem of insufficient,in the face of different bolt pin image characteristics of the suitable degree of samples with different characteristics,to imitate human cognitive model,based on attention mechanism and closed-loop feedback,is proposed based on an adaptive attention mechanism of bolt pin detection methods lack intelligence.The main work of this paper is as follows:(1)To solve the problem of insufficient aerial bolt pin image samples,affine transformation such as rotation,translation and clipping and generated adverse network were adopted to expand the aerial bolt sample set.In addition,in the face of the problems of overexposure,underexposure and image blurring caused by fog and turbulence,histogram equalization was applied to the aerial image to enhance the overall contrast of the image and provide a high-quality sample set for subsequent identification of bolt pin missing state.(2)For bolts high miss rate of the problems faced by small target detection,in the face of large amount of convolution computation of defects in deep learning,the aerial bolt pin images using noise disturbance instead of convolution operation,based on attention mechanism of iterations attention from rough to fine area,in the form of circulation layers of focus for the position of the target area more precise coordinates.In addition to solving the problem of missed detection rate,it provides accurate target image input for the subsequent pin missing identification model.(3)For deep softmax classifier in the neural network generalization ability is insufficient,with the aim of attention mechanism output image as input,the disturbance high-level abstract characteristics of neural network to extract the target image,the random configuration network classifier parameters adaptive setting,under the constraint of random input weights and deviation.A new random basis function in the network is generated according to the residual of the output pin missing result,which makes the classification model have the ability of ten thousand local approximation.In addition,the input target image changes when the focus level of the attention mechanism changes,so the network classifier is randomly configured to update the network structure and parameters adaptively to the diverse data sets at the input end,and the generalized classification criteria are generated.(4)In view of the traditional posterior statistical method cannot real-time evaluation missing bolt pin state of uncertainty as a result,the imitation of human from global to local repeated scrutiny than cognitive mode of thinking,and entropy theory,based on the generalized error build uncertain bolt pin form of entropy evaluation index focus on results,real-time evaluation training bolt pin attention is focused on the credibility of the results.A new bolt-pin missing detection model was generated by dynamically adjusting the level of the attention mechanism,and the training data set was reconstructed from the training samples with low reliability of the evaluation results of the focused target feature map.The deeper focus of the attention mechanism and the self-optimizing reconstruction of the classification criteria were carried out to re-identify the bolt-pin missing state.Finally,the bolt-pin missing detection model set based on the adaptive attention mechanism was used for multi-model fusion identification.The experimental results show that,compared with other target detection algorithms,the proposed method improves the cognitive accuracy of the model and enhances the generalization ability of the model,which has a good application value for fault inspection of transmission lines.