AUV pipeline looping method based on deep reinforcement learning of image features

Abstract

The invention discloses an AUV pipeline tracking method based on deep reinforcement learning of image features. First, the AUV tracking control problem is modeled as a Markov decision-making process of continuous state and continuous action; secondly, the control strategy is abstracted as a mapping from AUV observation state (image captured by the camera) to motion action, and expressed by deep neural network; finally , using the proximal policy optimization (PPO) method to autonomously collect data and train a deep neural network, and finally obtain an end-to-end tube-based control strategy with a certain generalization ability. Simulation results show that the invention can effectively control the tube-following action of the AUV, and has strong generalization ability for new and unknown tube geometry structures. This method is an end-to-end (end-to-end) visual tube motion control method, which does not need to know the kinematics / dynamics model of the AUV, and does not need manual feature extraction.

Description

technical field [0001] The invention relates to the field of intelligent marine equipment, in particular to an AUV pipeline tracking method based on deep reinforcement learning of image features, which has good self-adaption ability, strong generalization ability, and low software and hardware costs. Background technique [0002] Currently, oil, gas and electric companies have extensive networks of underground pipelines or cables. Since these pipelines and cables are widely distributed in the ocean, specially trained operators must use Remote Operated Vehicle (ROV) to perform daily inspection and maintenance work. Not only is the labor cost and maintenance cost high, it takes a long time, but it is also highly restricted by the weather conditions on the ground and the ocean. Therefore, the industry's demand for automatic detection, measurement and maintenance of underwater pipelines is growing and becoming increasingly urgent, and the emergence of autonomous underwater vehi...

AI Technical Summary

Problems solved by technology

Although the above methods do not require pre-acquisition of an accurate dynamic model in a complex seabed environment, they assume that the position and velocity of the AUV are known and are sensitive to the accuracy of the position and velocity information

On the one hand, in reality, it is very difficult for us to obtain the position and pose information of AUVs. Even if the information can be measured by precision sensors, the price and cost will be very high; on the other hand, the underwater environment is complex and changeable, and the acquisition It is also difficult to guarantee the accuracy and reliability of the AUV position and pose information

[0006] Therefore, the existing AUV pipeline tracking methods based on dynamic models and AUV pipeline tracking methods based on reinforcement learning strategies have many constraints and limitations, and there is still a long way to go before they can be put into practical use.

Images