Research On Defect Detection Method Of Transmission Line Bolt Based On Knowledge Representation

Intelligentization is a prerequisite for building a strong,sustainable,economical,and safe power grid,which is of great significance to national energy security.Automatic inspection of transmission lines is one of the main technologies for building smart grids,and its development has received extensive attention from experts and scholars in recent years.As bolts are widely used in transmission lines and play an important role in connection and fastening,how to detect bolt defects in time is a point that needs to be solved urgently.This paper mainly studies the detection of defective bolts in aerial images of transmission lines.Bolts have the following obvious characteristics in aerial images: 1)Small proportion in the image,fewer effective features that can be extracted by the machine;2)There are many types of bolt defects,and the difference between the classes is small,but the structural features are obvious.Recognition depends on fine-grained information;3)There is a latent relationship between different defects in bolts,and the model needs to learn the knowledge relevance between labels.In response to the above characteristics,this paper first constructed a multi-label recognition data set for bolt surface state.It includes three types of label for pin,nut and slim.Based on this data set,this paper proposes a multi-label recognition network for bolt surface state based on the joint visual-semantic knowledge for the problem that the relationship between bolt labels is difficult to learn effectively.The network designs a novel cascaded self-attention classifier based on knowledge relevance,which can jointly model the learned bolt visual representation and bolt semantic representation,filling the visual-semantic gap.Under the verification of multiple types of evaluation indicators,the network has significant advantages compared with the classic multi-label recognition network,with a label-level accuracy rate of 93.63% and an image-level accuracy rate of83.21%.Aiming at the problem that bolt structured and fine-grained features cannot be extracted effectively,this paper proposes a multi-label recognition network for bolt surface state based on the joint vision-location knowledge.This network is designed with a special bolt grid feature extractor and is combined with LSTM(Long Short-Term Memory).It successfully learns the fine-grained local features of bolts.Experiments show that the label-level accuracy rate of this network reaches 93.04%,and the image-level accuracy rate reaches 83.07%.Finally,this paper unifies the first two networks,and introduces the EfficientDet detection model,which realizes the bolt feature extraction from coarse-grained to fine-grained,and constructs a novel cascade detection framework for bolts on the transmission line fittings based on complex knowledge.Applying this method to the actual transmission line automated inspection task can significantly reduce the labor burden and time cost of the inspection task,.It is one of the effective ways to realize the intelligent power grid.