Research On Motion Planning Based On Autonomous And Self-learning Behavior Agents For AUV

This dissertation, based on national important basic project of "Autonomous Underwater Vehicle technology", researches deeply on application of key technques of motion planning based on behaviorism agent method for autonomous underwater vehicle (short for AUV). The aim of research is to enhance the adaptability to environment, speediness of reaction and efficiency of decision during autonomous mission process for AUV.This dissertation focuses on the following problems:According to the difficulties of motion planning in uncertain environment, the behavior dynamics method is proposed and explained, which takes the behaveior of agent as a dynamic interaction process between AUV and environment. The horizontal autonomous behavior agent of AUV is designed by this method. The simulation results show the autonomous behavior agent reacts to unstructural environment fast, correctly and effectively.The depth of submergence and height to the bottom are required in vertical motion planning, while the unique depth keeping control and height keeping control can't satisfy both the requirements. The vertical autonomous behavior agent is designed by fuzzy theory to integrate the depth keeping behavior and height keeping behavior. The simulation results show that the vertical autonomous behavior agent can ensure the safty of AUV, and can satisfy the height requirement of acoustic devices. Besides, AUV doesn't change depth frequently for waving sea bed terrain, so the navigation is steady.In order to coordinate the horizontal and vertical autonomous behavior agents, the coordination principle based on mission decomposing and task execution language is proposed. The mission is decomposed in sequential tasks, and the task execution language of particular task is defined, which can convert the execution of task to synchronous autonomous behavior agents, and then coordinate behaviors according to priority rules. The 3-D simulation method and mission case are designed to verify the developed behavior system. The results of simulation show that the designing and coordination principles of autonomous behavior agent are correct and practical.According to the shortcomings of traditional reinforcement learning method when applied to AUV engineering, such as generalization problem, risks by trial-and-error and low learning efficiency, the neural network and case based Q-learning (NCQL) is proposed. The basic principle of NCQL is making use of neural network to solve generalization problem, and case based learning to make sure the convergency of learning process, to avoid the risks by trial-and-error and to enhance the learning efficiency. The elements of NCQL based self-learning behavior agent are realized. Simulation tests are carried out. From the results, we can conclude that the NCQL is well done in its convergency property, and the speed of convergency is fast. NCQL has the properties of on-line learning and self-adapt learning.According to the top designing principles of autonomous control system, the behavior agent integrated autonomous hybrid autonomous control system is proposed. The colored Petri net (CPN) is adopted to build formalized model for the architecture. The basic properties of the architecture, such as boundness, liveness and mutual exclusion, are analysed by mathematic methods of CPN. The formalized analyses results verify the correctness and validity of the properties above of the autonomous control architecture.In order to verify the proposed methods in this dissertation, the sonar lake test, small scaled simulated tests and semi-physical simulation tests are done respectively. This dissertation proposes a new digital filtering algorithm to process sonar information afterwards. The test scheme of obstacle avoidance based on sonar information is designed; the lake test is carried out to verify the obstacle avoidance behavior based on sonar information. A small scaled simulated environment of motion planning is designed. This dissertation designs small scaled simulation tests, which apply laser measurement device to simulate sonar of AUV. According to application background horizontal behavior agent is verified under different test schemes. Simulation tests for behavior agent integrated autonomous control system are carried out by semi-physical simulation platform. The tests results are accord with requirement of designing, the proposed methods are practical and effective. The research on autonomous or self-learning behavior agent based motion planning have important theory and practical meaning of long-range navigation and terrain scaning safely, effectively for AUV.