Research On Non-contact Measurement Of Disconnector Based On UAV And Deep Learning

In the electrified railroad system,the disconnecting switch,as one of the key components,functions in the contact network to turn the circuit between each segment of the contact network on and off,so as to meet the needs of maintenance and different types of power supply and increase the flexibility and safety of power supply.Due to external factors,harsh environment and other factors,the distance between the lead line of the disconnect switch and the angle is too close,which is lower than the requirements of safety regulations and may produce discharge phenomena,thus causing major accidents.At present,the safety distance measurement of the disconnecting switch lead line relies entirely on human power,which leads to low efficiency,high labor intensity,and even the safety of personnel may be threatened.Therefore,this thesis takes the disconnecting switch of the contact network as the research object,and identifies and locates the image of the disconnecting switch obtained by the UAV inspection as well as the non-contact measurement of the safety distance on the lead line.In this thesis,we propose a YOLOv5 s target detection network model based on the Ghost lightweight module to address the problem that the speed of disconnect switch identification and positioning cannot meet the engineering needs due to the huge amount of UAV inspection data and the high real-time requirement of fault maintenance.By replacing part of the Convolution operation with a linear transformation,the number of parameters and the computational effort are reduced,and the speed and efficiency of target detection are greatly improved.Second,to address the problem of insufficient accuracy of disconnecting switch recognition due to dynamic acquisition by UAV and complex environment along the contact network,an improved YOLOv5 s algorithm based on joint CA and BiFPN is proposed.the CA attention mechanism can use both inter-channel relationship and location information,so that the improved model can obtain the connection between different channels and make an accurate judgment of direction and location.The BiFPN feature fusion algorithm is introduced into the YOLOv5 s model,and the idea of adding jumping connections to the YOLOv5 s model is introduced to achieve repeated superposition between different feature layers,so that the global features do not only contain the underlying information but also the high-level semantic information,which enhances the model’s ability to fuse features at different scales,and the fast normalized feature fusion method of BiFPN is applied to the YOLOv5 s network.YOLOv5 s network to reduce the model speed as little as possible.The improved YOLOv5 s algorithm with the combination of the two optimization modules has significantly enhanced the extraction performance of important features,thus improving the detection model’s average accuracy for detecting images with cluttered backgrounds and images of low definition without much speed degradation.Then,the three optimization modules mentioned above are jointly utilized in the YOLOv5 s improvement,which achieves a joint improvement in the performance of the model in terms of accuracy and speed.Thirdly,after the rapid and accurate identification of the disconnector,for the non-contact measurement of the distance between the disconnector lead wire and the angle,this thesis fully considers the structural characteristics of the disconnector and uses the improved Canny operator to obtain better measurement results.Firstly,fast adaptive bilateral filtering is introduced to realize the smoothing operation of the disconnector image,and then Otsu adaptive selection of high and low thresholds is used to avoid the flexibility problem of manually determining the threshold,and the excellent performance of the improved Canny algorithm is proved by experiments.Then,using the monocular visual measurement method,the diameter of the lead wire is used as the standard part size to realize the non-contact measurement of the safety distance of the lead wire of the disconnector.Finally,the measurement results of the algorithm are compared with the accurate data measured manually and the error analysis proves that the measurement accuracy meets the actual requirements of the project,and the algorithm has good practicability.Finally,according to the overall process of non-contact measurement,the noncontact measurement software of the safe distance of the catenary was developed using the Python programming language.The main functions of the software are video frame reduction,target detection,distance measurement,etc.And the software is easy to operate,easy to get started and highly practical.