|关键词||巡线机器人 多任务学习 绝缘子分割 融合感知 联合标定|
|英文摘要||At present, regular inspections of power transmission lines are still based on manual inspections, which have high labor intensity, low inspection efficiency, and high safety risks. Some helicopters can improve the inspection quality, but the operating costs are expensive. Existing hang-up inspection robots have complex mechanisms, limited obstacle avoidance capabilities, and difficulty in getting on and off the line. Although the small multi-rotor UAV is flexible, it has a short flying duration and limited assistance.|
This paper proposes a hybrid inspections robot with a combination of flying and rolling. It can fly on and off the wire easily, moves on the lightning protection wire (overhead ground wire) and fly across the obstacles with potential long duration, which provides a new way for the inspection of transmission lines in harsh areas. In this paper, the full line, automation, and long-distance for hybrid inspections robot are studied around their environmental perception methods. The main contents are as follows:
Firstly, the design of the hybrid inspections robot and the joint calibration method of the sensors are studied. For the obstacle avoidance and online movement requirements of the hybrid inspections robots, a corresponding roller structure and a two-dimensional environment perception system combining the laser ranger finder (LRF) and the camera are designed to enable them to land and cross obstacles on the wire. The data delay model between LRF and swing servo motor, the mechanical parameter model of the swing LRF, and the relative pose model between the LRF and the camera are established. The coupled mechanical parameters and delay parameters are calibrated by the joint calibration method with a single or multiple plane. The simulation and experiment verify the accuracy and stability of the proposed method.
Secondly, the method of identifying the pose of linear objects based on neural network is studied. Aiming at the problem that the conventional box annotation in object detection cannot effectively represent the direction of the linear object, a multi-objective end-to-end neural network is constructed, and a corresponding loss function is designed based on the conditional generative adversarial nets (cGAN). In the two branches contained in the network, the location branch outputs the position and direction of the wire, and the saliency map branch outputs the saliency area of the wire in the image. Based on this, a dynamic segmentation method for conductors based on Census transformation is proposed, which makes the segmentation of conductor area more precise. It solves the problem of identifying the wire that from approximate straight line to spiral texture region in the process of the hybrid inspections robot approaching and landing on the wire.
Thirdly, the synthesis method of image segmentation samples and its training strategy are studied. Aiming at the scarce training samples in overhead lines, a method of synthesizing segmented samples using image segments is proposed. The synthesis process takes into account the diversity of targets and backgrounds respectively. Using the repeatability of insulators or wires in the image, the different sample segments are spliced and enhanced and then blended with the background image in proportion to the training sample, and the training strategies to control learning difficulty and segmentation effects using the transparency of the target object is proposed. It solves the problem of the scarcity of supervised image segmentation training samples and reduces the consumption of synthetic sample. The synthesis method and training strategy have achieved good training effect in various segmentation networks and have been verified in experiments.
Fourthly, the landing and obstacle crossing of hybrid inspections robot based on perceptual information fusion is studied. Aiming at the requirement of real-time detection of the wire pose at different distances, a neural network landing point detection method is proposed, which combines depth and image texture information. The depth information is also combined in the process of sub-image segmentation. The fusion method reduces the interference of rod-shaped objects and shorten the overall landing time. A compact three-dimensional convolutional neural network is constructed for the identification of online obstacles and key targets (insulators), and the methods for synthesizing three-dimensional point cloud samples and jitter blurred image samples are proposed to recognize various of online objects and the key targets in videos, which has yielded good results in landing and crossing obstacle experiments. At the same time, the experiment also analyzed the power consumption of the robot under different conditions, and verified its long-term potential.
Finally, it summarizes the work of this paper and points out the work that needs further development.
|常文凯. 混合式巡线机器人环境感知方法研究[D]. 北京. 中国科学院研究生院,2018.|