| With the deepening of marine research and the improvement of marine resources development and so on,as an autonomous marine observation platform,the Underwater glider(UG)is one of the most cost-effective equipment for mobile observation in the present stage of marine observation due to the advantages of low energy consumption,long voyage,and relatively inexpensive cost,etc.The UG can play a greater role in the field of marine research.UG formation system can play a greater role in the field of ocean research because its comprehensive capability,range of use,and effect are far more than that of a single machine.In this context,the research on formation path planning is carried out for the large-scale and autonomous observation tasks of glider formation to provide a theoretical basis and technical support for the planning under different ocean current environments.Based on the hierarchical control system,using methods such as multi-body dynamics and deep learning,fixing other ocean factors such as wind and waves,and according to the progressive relationship of known,uncertain to unknown environment,this thesis studies the path planning of the UG formation in the ocean current environment,and form a set of planning layer oriented technological solutions.The main contents of this thesis are as follows:Firstly,the structurally flexible glider formation is modeled in generalized coordinates using the Kane equation.Based on the hierarchical control system,the planning layer of the glider formation and the single-machine control are regarded as mutually independent research.Based on the glider mathematical model,the relationship between gliders is constructed by the continuous potential field,and the Kane equation is used to avoid the calculation of the interaction force between the gliders under the premise of formation consistency,to give the kinematics and dynamics model of glider formation in generalized coordinates.The virtual leader method is used to construct the formation as a whole,and the step input is adopted to change the formation shape,which realizes the transformation of the formation and avoids the calculation of the leader velocity so that the formation is both stable and flexible.The formation architecture and kinematic model are provided for subsequent research.Secondly,the Rapidly-exploring Random Tree(RRT)algorithm is improved by using the guidance of ocean currents for the path planning of glider formation in the known ocean current environment.The ocean current dataset is established to provide an environment for the simulation test,and then the evaluation indexes of glider formation path planning in the ocean current are given.The RRT algorithm is improved with the help of the ocean current velocity and direction of the ocean current,and the algorithm uses the target potential field and the ocean current direction to guide the path trial direction of RRT under the premise of maintaining the exploration ability,which reduces the sampling area of the RRT nodes,lowers the random range of exploration,improves the searching speed of the optimal path in the environment,and enables the algorithm to utilize the ocean current to optimize the path at the same time.Simulation experiments show that the improved RRT based on energy strategy optimization reduces the number of cycles by 7.45%,path length by 5.84%,and improves the energy utilization index by 9.09% compared to the original algorithm;the improved RRT* reduces the number of cycles by 7.06%,path length by 4.56%,and improves the energy utilization by11.43%.The simulation verifies the effectiveness of the improved RRT algorithm using ocean currents for path planning in a known environment.Then,a dual-branch network architecture with deep feature extraction capability of Convolutional Neural Network(CNN)and global planning is proposed for the path planning of glider formation in the environment of globally observable but uncertain future ocean current speed and direction.The Deep Convolutional Neural Network for ocean current environment(Doc-CNN)can utilize known ocean current data to guide glider path planning by implicitly predicting the future.The inputs of the algorithm include glider positions and dynamic ocean current data,branch 1 is used to extract global features,branch 2 is used to extract local features,and the fusion of the features implicitly predicts the state of the next cycle.Based on the prediction,the algorithm outputs planning commands for the gliders,realizing the construction of formation,obstacle avoidance,and path planning.Simulation experiments show that the planning success rate of Doc-CNN in the ocean current dataset is about 82.5%,which is better than that of Deep Convolutional Neural Network with dual branches(DB-CNN)of 68.4%;the energy utilization index of Doc-CNN is better than that of DB-CNN and Value Iterative Network,the average energy utilization index of Doc-CNN is 2.56% and 8.11% higher than that of DB-CNN and Value iteration Network(VIN).Doc-CNN improves the success rate and adaptability of glider formation planning in uncertain ocean currents,and completes the progression process of glider formation planning from known to uncertain environments,and perfects the planning ability and environmental adaptability of UG formations.The simulation verifies the effectiveness of the ocean current implicit prediction ability and path planning ability of the dual-branch Doc-CNN.Finally,for UG formation path planning in the ocean current environment with unknown global speed and direction,a deep reinforcement learning planning method is designed based on the processing capability of Deep Deterministic Policy Gradient(DDPG)for the unknown environment by considering the motion characteristics of the gliders in the ocean current environment.By integrating the planning strategy and the collision risk assessment strategy,a new Markov Decision Process(MDP)is designed and constructed,which includes the characteristics of glider formation movement,collision avoidance,flexible formation,etc.,and can improve energy utilization capability.Through the flexible reward functions and the depth strategies gradient,the behavior strategies are generated,judged,and updated,so that the glider formation can achieve formation maintenance,formation transformation,obstacle avoidance,and path planning in the unknown ocean current environment.Simulation experiments show that the episode used by DDPG to achieve stability in formation planning training is lower than that of Deep Q-network(DQN)and Policy Gradient(PG);the average reward of formation planning approaches 27,which is 18 and 10 higher than that of DQN and PG,respectively;and the algorithm has a good generalization ability in unknown ocean environments.DDPG extends the path planning of glider formation to the unknown ocean current environment,completes the recursive process of known,uncertain,and then unknown,and improves the environmental adaptability of UG formation.The simulation verifies the effectiveness of the improved DDPG planning algorithm in unknown ocean currents and proves the effectiveness of MDP.In summary,other marine environment parameters are fixed,this thesis studies the path planning of UG formation for the ocean current environment with increasing unknown and uncertain factors.Considering the formation consistency to establish the glider formation model,utilizing the ocean current to improve the speed and range,and proposing the improved RRT,Doc-CNN,and improved DDPG algorithms for the known,uncertain,and unknown ocean current environments,respectively.The visualization simulation results show that the algorithms proposed in this thesis are suitable for their use environments,can cover most of the use cases of glider formation path planning in ocean currents,and have good engineering application prospects. |
PDF Full Text Request