| In recent years,with the continuous changes in the external environment of power transmission channels,there have been frequent incidents of external damage to power transmission channels,and the problem of external damage has gradually become a severe test for grid companies to ensure power supply.In order to ensure the safety of people,equipment,and power grids,the power grid company has spent a lot of financial and material resources on the maintenance and repair of power transmission channels through manual inspections and video surveillance.Due to the limitations of manual operation,in the inspection and monitoring process,there are problems such as low flexibility and high missed detection rate,and it is impossible to solve the hidden dangers of external damage in the transmission channel in a timely and effective manner.With the rapid development of feature extraction technology based on deep learning,the target detection technology based on deep learning has made rapid breakthroughs and is widely used in industry,agriculture,medical and other fields.The target detection technology based on deep learning can extract high-level semantic information from natural images,and accurately locate the position and category of the target.Compared with traditional manual inspection,this method has high detection accuracy and stability,fast processing speed,etc.Advantage.In this paper,the target detection technology is applied to the intelligent inspection of the transmission channel,and the hidden danger detection model of the transmission channel is constructed.The model detects four types of hidden danger targets in the data of the multi-channel camera in the transmission channel:excavators,tower cranes,bulldozers,and cranes,and can help inspectors find the hidden danger information in the transmission channel in time,and deal with them in time to ensure The daily stable operation of the transmission channel.Taking into account the problem of many images to be detected at the same time in practical applications,this paper uses a combination of a lightweight feature extraction network and a one-stage lightweight detection network to improve the detection efficiency of the model,and combines data processing technology to improve the quality of the data set to improve the detection accuracy.This article mainly completes the following tasks:（1）This article cleans and analyzes the transmission channel hidden danger data set,and preprocesses the target data set from the perspectives of image compression,image contrast,noise processing and data enhancement.The experimental results show that the preprocessing operation used in this paper can effectively improve the accuracy of the detection model in the target scene.（2）According to the requirements of the model application scenarios for the detection speed,this paper uses the lightweight feature extraction network MobileNetV3 as the backbone network,and the one-stage lightweight detection network SSDLite as the detection model.In addition,this article also uses K-means clustering to set the anchor box of the detection model.The experimental results show that although the SSDLite detection model using a lightweight feature extraction network has a slight decrease in detection accuracy,the detection speed has been greatly improved,and the anchor box can be set through K-means to effectively improve the detection accuracy.,Suitable for landing scenes.This article also uses the inference acceleration module TensorRT to accelerate the model,which further improves the detection efficiency of the model.（3）This paper designs a prototype system for the detection model（only includes basic functions and reserves interfaces）,uses the C#WPF framework to build the interface program on the PC side,and uses the Python language to provide detection services on the back end,and system integration of the model.And at the system level,the detection strategy of parallel detection is adopted to further improve the detection efficiency of the system. |
PDF Full Text Request