Research On Fitting Detection Of Transmission Line Based On Knowledge Reasoning

With the development of economy and the acceleration of power grid construction,the stable operation of transmission lines,as an important carrier of power transmission,has been paid more and more attention by power inspection departments.Note that,fittings,as an important accessory of transmission lines,have become a hot issue in current intelligent inspection tasks due to their complex working environment and increased maintenance requirements.With the development and breakthrough of digital image processing technology,the inspection technology of transmission line fittings based on aerial image processing has been applied in various power grids and subsidiaries.As the basis of fitting defect detection and condition monitoring,highprecision fitting detection technology has played an important role in ensuring the safe operation of powerConsidering the maturity of deep learning technology,fitting detection methods based on image processing and object detection have made great progress.However,the existing deep learning-based fitting detection algorithms are mainly based on appearance information design models.Due to the installation characteristics of transmission line fittings and the shooting method of aerial images,detection algorithms are often limited to detection difficulties such as small object size,complex background,data imbalance,and mutual occlusion.The existence of these problems reduces the accuracy and robustness of object detection.In view of this,this paper explores how to integrate various knowledge forms in the power field into the current detection paradigm in a lightweight and efficient way to simulate the human reasoning process,and explore how knowledge reasoning assists object detection technology to achieve that the knowledge,data,and models are simultaneously driven to solve difficult problems in computer vision detection.This paper takes the transmission line fitting detection model integrating knowledge reasoning as the main research line,and conducts a series of researches from the feature enhancement of spatial knowledge,the adaptive extraction of context information,and the instantiated expression of object relations.Specifically,the innovative achievements of this paper mainly include the following aspects:1.Firstly,this paper studies how to solve the problems of different object scales and mutual occlusion in the multi-fitting detection task.To this end,this paper proposes a multi-fitting detection method(Scale Constraint and Spatial Information RegionConvolutional Neural Network,SCSI R-CNN)for transmission lines with scale constraints and spatial information.The model uses the Faster R-CNN algorithm as the baseline framework,and the k-means++ algorithm is carried out on the cluster analysis of object size in the transmission line fittings dataset,and the anchor-box scale constraints suitable for the detection task are obtained through the CH index comparison.Then,based on the relative geometric features of the spatial layout of the fittings,the latent space module is used to enhance the input visual features to characterize the spatial layout between fittings and achieve high-precision object detection.2.Then,this paper studies how to solve the problem of multi-object detection in aerial images of transmission lines,which is tiny-size and susceptible to complex background interference.To this end,we implement a Context-based Graph Reasoning Model(CGRM)by integrating mechanisms such as multi-scale context and self-attention graph learning.The model consists of a Graph-based Global Context(GGC)sub-network and a Multi-scale Local Context(MLC)sub-network.To capture local contextual information,the MLC sub-network exploits the internal and external contextual information of each specific proposal at multiple scales.To obtain global contextual features,the GGC sub-network effectively extracts a pixel-level global map based on a multi-head attention mechanism,and further utilizes a Graph Convolutional Neural Network(GCN)for inference.Then,the global context and local context are fused together to realize the task of fitting detection.The model also provides a new idea for the task of small object detection.3.Finally,this paper studies the pain point of multi-fittings inspection in transmission lines combined with the unbalanced-sample and small-scale problems.To this end,we use prior knowledge to assist target detection through the analysis of the fixed structure of the fitting and the mining of the corresponding relationship.Then a variety of object relations are embedded into the object detection network and a hybridknowledge based object detection model(Hybrid Knowledge-based R-CNN,HK R-CNN)is proposed.Firstly,the expression is instantiated by using the structure combination rules of the transmission line fittings extracted by the data-driven method.Secondly,the Position-Sensitive Score Map(PSSM)is used to represent the relationship between the fittings’ structures.Finally,the knowledge integration module is embedded into the object detection model based on graph learning to achieve feature enhancement and effectively improve the detection accuracy of multiple fittings.