| The transmission line inspection robot is an efficient and safe way to replace manual inspection of transmission lines,with the advantages of low operation and maintenance costs,fast inspection speed and high safety.With the development and application demonstration of inspection robot technology,factors such as long inspection routes and variable environments bring higher challenges to the automation and intelligence of inspection robots.To improve the automation and intelligence of the inspection robot,thesis focuses on the robot target detection and obstacle path planning based on the self-developed "Flying-Walking Power Transmission Line Inspection Robot(FPTLIR)",including the architecture and implementation of the robot platform,an auto-synthesis dataset approach for fitting recognition using prior sequential data,a fitting recognition method of deep-attention YOLOv5 based on prior synthetic dataset,and an obstacle path planning of inspection robot.(1)Designing robotic platforms.The working environment and inspection principle of FPTLIR are analyzed,based on which the hardware architecture and software architecture of FPTLIR are proposed,including the composition of FPTLIR flight mechanism and walking mechanism,and the implementation of system hardware and software architecture,providing a platform for the experimental verification of key core algorithms of FPTLIR.(2)Proposing auto-synthesis dataset method for fitting recognition using a priori sequential data.The method formulates synthesis rules based on four types of a priori series data: a priori fitting series,a priori model series,a priori geographical series,and a priori time series.Then,a dataset containing images and labels is outputted by processing the fitting models and panoramic mapping(HDRI)backgrounds according to the synthesis rules by Blender and Open CV technologies.The synthetic dataset is used to train the YOLOX neural network model,which achieved 98.38% m AP5095 when tested on a real dataset.experimental results verify the effectiveness of the proposed automatic synthetic dataset algorithm.(3)Proposing a deep-attention YOLOv5 fitting recognition method based on prior synthetic datasets.The method proposes a deep-attention module that is loaded after the head of the YOLOv5 neural network.The neural network is frozen during training and does not need to participate in the training process,or it can be added directly after a trained model.In this section,test field tests and line tests are conducted,and the test field tests showed that the method of recognition of the fitting improved by 5.2% over the standard YOLOv5 model m AP5095,validating the feasibility and effectiveness of the deep attention YOLOv5.Four types of insulator data are collected by FPTLIR walking on 10 k V transmission lines in line tests and a test set is produced.The trained deep-attention YOLOv5 model achieves 64.6% m AP5095 on the test set,which is a 3.2% improvement over the standard YOLOv5 model.Also compared with other attention mechanisms,the results show that the proposed method has better recognition accuracy and faster inference speed.(4)an obstacle path planning of inspection robot.The robot model and world environment are built through Gazebo,while the robot model communicates with PX4 to control robot motion and pass sensor data.Among them,the planned trajectory of robot motion is acquired in real-time based on ROS and the trajectory data is sent to PX4 to control the robot over the obstacles.In the ROS system,the EGO-planer algorithm is used for real-time local dynamic path planning.The robot position is provided by VINS-Fusion processing binocular and IMU data,the target is recognized by the deep attention YOLOv5 algorithm,and the viewpoint coordinates are output for the EGO-planer algorithm to plan the motion path.Simulation experiments verify the feasibility of the algorithm in this chapter. |
PDF Full Text Request