Autonomous Obstacle Avoidance Planning Methods For UUV Based On Deep Learning

As important underwater unmanned equipments,Unmanned Underwater Vehicles(UUVs)are widely used in Marine scientific research,Marine resource survey,and Marine safety.With the deepening of human exploration of the ocean,the operating environment of UUV is gradually characterized by unreachable operators and weak communication signals.The autonomous obstacle avoidance planning of UUV has gradually become an important technology to ensure its safe navigation,and also an important manifestation of its autonomous ability.The development of autonomous UUV obstacle avoidance technology is restricted by the unknown obstacle movement and the uncertainty of underwater detection in complex operating environments.The autonomous obstacle avoidance planning system of UUV is a dynamic nonlinear system.Deep learning technology can establish dynamic nonlinear description between the input space and the target space through continuous training.It has strong adaptability to the disturbance in the input space,and establishes the state space to describe the problem autonomously.Therefore,based on deep learning technology,this thesis conducts in-depth research on UUV autonomous obstacle avoidance planning method.Firstly,the problem of UUV obstacle avoidance planning based on traditional methods is studied,and a UUV autonomous obstacle avoidance planning teacher system based on traditional obstacle avoidance planning methods such as the improved artificial potential field method,dynamic window method,and fuzzy logic algorithm is established.An obstacle avoidance planning method based on hybrid quantum ant colony optimization is proposed.Ant positions are represented by quantum bits to enlarge the search space.Quantum rotation operation is used to update pheromones to increase the diversity of population positions and avoid premature convergence.And the local search method is introduced to optimize the planning results.This algorithm is combined with the historical obstacle information to avoid the oscillation of the output control,path redundancy and local minimization caused by the traditional obstacle avoidance mode of "real-time sensing real-time planning" in the teacher system.Secondly,aiming at the unknown and uncertainty of underwater moving obstacles,the target state estimation of UUV is investigated to estimate the motion state of the moving obstacle in UUV autonomous obstacle avoidance planning.In order to establish the mapping between the target measurement state space and motion state space,a deep neural network framework based on the Gated Recurrent Units(GRUs)is built.Samples are extracted from historical observation data are used to approximate the distribution of observation.Then the current states of the particles are predicted by fully trained deep neural networks.The target state estimation method based on GRUs particle filter overcomes the uncertainty of target state estimation caused by the complexity of target kinematics and the non-linearity of target system.Thirdly,the autonomous obstacle avoidance planning for UUV in two-dimensional plane is addressed.The nonlinear and non-Markov state equations of UUV autonomous obstacle avoidance planning system in plane are constructed.Deep neural networks based on Recurrent Neural Networks(RNNs)are established to describe the state space of this problem,and establish the mapping between input space and state space is established.Thus,the UUV autonomous obstacle avoidance planning for depth determination is realized.This method improves the sensitivity of traditional obstacle avoidance planning methods to sonar measurement noise,enhances the adaptability of obstacle avoidance planning to measurement noise and environment.By comparing the performance of three classical RNNs in this problem,it is found that GRU achieves better performance of obstacle avoidance with simpler structure and fewer parameters.Finally,the problem of autonomous obstacle avoidance planning for UUV in 3D dynamic space is investigated.In view of the insensitivity of RNNs to spatial information,a threedimensional autonomous obstacle avoidance planning method for UUV based on Convolutional Recurrent Neural Network(CRNN)is proposed.The proposed method uses CNN to extract features from sonar detection data,and RNN to carry out obstacle avoidance planning according to sonar features.The experimental results show that the CRNN-based UUV autonomous obstacle avoidance planning method still performs well under the condition that the environment is more complex and the detection noise is larger than that of the training environment,which greatly improves the autonomous ability of UUV obstacle avoidance planning.