Research On Bolt Multi-attribute Classification Method Based On Multi-label Learning

With the deepening of the power system reform,the construction of a new generation of power systems in our country has developed rapidly.Transmission equipment is one of the important components of the power system,the intelligent identification and analysis of its defects have become an important research direction.Bolts are the most numerous fastening parts on the transmission line which are prone to failure.It has long been a difficult problem for the power system to realize automatic and efficient defect identification.It is of great practical significance to use computer vision and artificial intelligence technology to identify bolt defects accurately.Based on the research on bolts,this paper proposes the idea of visual separability for bolt attributes,adopts a bolt multi-attribute classification method based on multi-label learning,improves the feature extraction ability of deep learning model and realizes more accurate bolt attribute classification,and provides a new solution for the identification of bolt defects.First,in the traditional bolt defect detection task,the direct use of the object detection model to detect the bolt defect leads to the shortcoming of low accuracy.In view of this shortcoming,this paper puts forward the idea of visual separability of bolt attributes,and determines whether the bolt has pin hole,shim and nut,and so on,which belong to the category of visual separable.Therefore,bolt identification is regarded as a multi-attribute classification problem,a bolt multi-attribute classification dataset is constructed based on 2000 images,1500 images of which are used for model training and500 images for model testing.Secondly,compared with the general image multi-label classification task,the overall features of the objects to be detected in the multi-attribute classification task of transmission line bolts are relatively similar,and the classification model needs to capture the key local features that can distinguish different categories of attributes.In order to efficiently and accurately locate the key local areas that contain detailed semantic information and get rid of the heavy workload of label production,this paper innovatively proposes to use the NTS-Net model to classify bolt attributes.In this model,a multi-agent cooperative self-monitoring mechanism is adopted to complete the efficient and accurate positioning of key local areas of bolts,and then the key local features and global features are integrated to achieve better classification.Through experimental verification,NTS-Net model obviously achieved better classification accuracy than NS-Net model,Res Net,VGG and other basic models.Finally,aiming at the problems of small difference of bolt targets and diverse shape changes,this paper adds deformable convolution module to the feature extraction network of NTS-Net model to improve the modeling ability of the model for spatial information.Compared with the traditional convolution method,the deformable convolution module has better geometric deformation modeling ability and better screening ability for the bolt target.In addition,in order to consider the different effects of different local features on different attribute labels,channel attention mechanism is introduced in the fusion process of local features and global features to extract channel weights of features and obtain key channel features to improve the multi-label classification effect.The experimental verification shows that the average classification accuracy rate of the proposed model on the multi-attribute classification data set of bolts is 84.5%,which is 10%-20% higher than that of the traditional multi-label classification method.In addition,the Grad-CAM algorithm is used in the qualitative experiment to visually analyze the key areas of bolt image classification,which improves the interpretability of the model.