Transmission Line Defect Detection Based On Neural Network

In recent years,the UAV technology is widely applied for transmission line detection.The electric power inspection efficiency has been greatly improved.Meanwhile,the inspection mode has been changed from the wide range convention inspection to the well-directed one which can detect defects of important equipment of the electric transmission line and contribute to improve the decision-making level of power grid operation.Insulators and pins are important electrical components,but also vulnerable to damage.And they are more likely to be defective in the daily operation of power grid,so it is very necessary to study the defect detection algorithm.The traditional defect detection is time-consuming and difficult to ensure the consistency and accuracy because of the manual evaluation method for UAV aerial image evaluation.Compared with traditional image processing technology,deep learning method has become a common approach today due to its high detection precision and speed.According to the actual work requirements,this thesis proposes an online defect detection algorithm based on FF-YOLO(Feature Fusion-YOLO)and an improved defect detection algorithm based on Faster RCNN.FF-YOLO is used for online detection,which balances speed and accuracy,and the improved Faster RCNN is used for offline detection with high accuracy.For online detection tasks,a defect detection algorithm,FF-YOLO is proposed to detect multi-scale targets.In order to improve the detection effect of small targets,this model adopts a feature fusion backbone network FF-Darknet,and the Spatial Pyramid Pooling(SPP)module is added to enrich the feature expression ability.Meanwhile,the model reduces the mutual inhibition among hierarchical features by improving the feature pyramid prediction module.In order to improve the stability of anchor frame generation before training,a K-mean++algorithm based on Euclidean distance is proposed.Aiming at the problem of poor detection effect because of the large scale gap,a new loss function optimization model was adopted in FF-YOLO.Besides,for the purpose of balancing proportion of data samples,data amplification methods such as small sample over-sampling and background fusion are used to improve the model performance.The thesis proposes an improved Faster RCNN algorithm,since offline detection tasks generally pay more attention to detection accuracy.Pyramid RPN anchor point generation method is proposed in this model,which improves the detection accuracy of small targets.Besides,in order to locate the target accurately,this thesis uses the floating point bilinear interpolation method to aggregate regional features.The experimental results show that the improved Faster RCNN algorithm can significantly improve the detection effect of targets of different scales.In summary,this thesis proposes a detection algorithm FF-YOLO and an improved Faster RCNN detection algorithm according to the different focus of online and offline detection tasks.FF-YOLO detection algorithm ensures good real-time performance and improve the detection accuracy to 88.41% m AP.The detection accuracy of the improved Faster RCNN algorithm is91.30% m AP.The experimental results show that the algorithm proposed in this paper can be applied to the actual transmission line inspection task.