Design And Research Of Control System Of Catenary Insulator Water Flushing Device Based On Vision

In the research of visual servoing of KJ series live catenary insulator water flushing device,the insulator target continuous positioning method based on low level features is difficult to work in complex background images,and the existing vision controller methods need accurate system modeling.In order to improve the shortcomings of the existing research,the visual control system of the device was researched.Firstly,in order to determine the visual servo control strategy of the device,the operating conditions,mechanical structure and system control requirements of the device were analyzed.According to the analysis results,the image based visual servoing and the eye-in-hand are selected.At the same time,the monocular distance measuring model is determined to obtain the depth.On this basis,considering the problems of dual-rate control and attitude isolation,the overall design of the visual servo control system of the device was completed,and the architecture of the system hardware and software was designed,and the hardware selection of part of the system is completed.Secondly,in order to obtain the control deviation of the visual servo system of the device,continuous positioning of the insulator target is needed.As a preliminary task of insulator target continuous location,insulator target detection is studied in this paper.DPM(Deformable Part Model)and YOLOv3(You Look Only Once v3)were used to train and test the detection model respectively,and the results showed that the detection speed of both models was not ideal.In the improvement of DPM,the feature interpolation estimation is introduced to obtain the accelerated Cascade-DPM detection model.The test results show that the model greatly improves the detection speed,but the detection accuracy is reduced.In the improvement of YOLOv3,a lightweight YOLOv3 target detection model is obtained by combining the lightweight network architecture and the improved bottleneck layer module.The test results show that the model can greatly reduce the number of model parameters,calculation amount and model size,and improve the detection speed and accuracy of the model.In contrast,the lightweight YOLOv3 target detection models with different configurations can provide better detection performance than the accelerated Cascade-DPM model.Then,in addition to detecting the insulator,the identity of the detected target needs to be determined to accomplish the continuous positioning of the insulator target,which requires target tracking.In order to balance the tracking speed and accuracy,the integrated algorithm of insulator target detection and tracking is studied by combining multi-object and single-object tracking.The SORT(Simple Online and Realtime Tracking)multi-object tracker and the lightweight YOLOv3 detector are used to complete the integrated algorithm based on detection and tracking route.The test results show that the algorithm can track multiple targets quickly.The ECO-HC(Efficient Convolution Operators-Hand Craft)single target tracker and YOLOv3 detector were used to complete the undetected tracking route integrated algorithm.The test results show that the algorithm can accurately track a few targets simultaneously on the CPU computing platform.On this basis,the insulator contour fitting processing algorithm and the image processing software of the upper computer are designed to realize the calculation and data transmission of the deviation between the real time and the target features,which lays a foundation for the subsequent research of the vision controller method.Finally,image deviation can be obtained by the continuous positioning of the insulator target.In order to complete the control task,the algorithm of the IBVS controller needs to be studied.Considering the requirements of the system for the robustness of the control parameters,a dynamic quasi Newton visual servo controller was designed based on the adaptive square root untraceable Kalman filter model,and an adaptive neural network PID visual servo controller was designed based on the radial basis neural network.Through Peter Corke’s robot and vision MATLAB toolbox to build a model,the simulation comparison of the above two algorithms was completed.The results show that the latter calculation is simpler and the error is smaller.