Research On Application Of Object Detection Technology In Fault Diagnosis Of Transmission Tower Components

At present,the inspection of power pole towers has gradually changed from manual inspection to unmanned aerial vehicle inspection.It analyzes and recognizes the positions and defect categories of components in the images by manually interpreting many images taken by the drone,which is time-consuming and laborious.Machine vision-based target detection technology has been studied for many years,but due to the complex background in the tower image,the small size of the bolts of the metal parts,and the existence of many similar parts,the image recognition technology for the bolt defects of the small parts of the transmission line tower still challenging.This paper first constructs a pole tower bolt defect image dataset containing various scenarios and uses the most popular target detection method Faster RCNN to successfully prove the feasibility of using deep learning to solve the transmission line defect component identification task.By analyzing the process of Faster RCNN algorithm recognition and detection,the author found the shortcomings of the framework in the bolt component recognition task,and creatively proposed targeted solutions such as multi-layer feature fusion,local context information fusion,and an attention mechanism.The scheme achieves much higher accuracy of defect bolt recognition than Faster RCNN.Details as follows:(1)Because parts such as tower bolts account for a small proportion in the whole picture,and Faster RCNN algorithm only extracts foreground targets on the feature map output by the last layer of convolution,so the small target defective parts are missed.In this paper,a combination of deep feature maps and shallow feature maps in the feature extraction network is used to construct multiple new feature maps of different sizes,and then the foreground extraction network is used to perform the foreground target on multiple new features maps of different sizes.Extraction,thus greatly reducing the problem of missed inspection of small size defective bolts.At the same time,to solve the problem that the new feature map obtained by the fusion of shallow feature maps has an insufficient feature extraction due to fewer convolutional layers,this paper adds several residual block units to the foreground target classification and a position regression network,which makes up for the shortcomings caused by the fusion of multi-layer convolution feature maps.(2)The Faster RCNN algorithm only considers the information within the target box when classifying and regressing the foreground target,which causes the problem of misjudgment of any defective parts.In this paper,the target box and the context information surrounding the box are used to construct a recognition network that considers the local context information of the component,thereby solving the impact of complex backgrounds and similar components.(3)Aiming at the problem of equal distribution of attention resources when CNN uses convolution operation to extract features,this article uses the attention mechanism to strengthen the most characteristic areas in the picture,while suppressing unimportant areas to make limited attention resources better,the ground is assigned to the recognition of key point information,so that the model has focused like human visual attention.Experiments show that the model will have a better defect-recognition effect after adding the attention mechanism.Through evaluation,the optimal model designed in this paper improves the AP(Average Precision)value of bolts with defects by 11.4 points and 18.4%compared with the commonly used target detection method Faster RCNN.