A Numerical Simulation Method Of Fish Adaption Behavior Based On Deep Reinforcement Learning And Fluid-structure Coupling Method

Fish,as aquatic organisms proficient in utilizing the principles of fluid dynamics in nature,have been the subject of extensive research over the past few decades.Their flexible and agile swimming behavior and remarkable swimming efficiency are unmatched by any artificial propulsion systems.Researchers worldwide are currently devoted to unraveling the fluid mechanics behind fish locomotion,aiming to provide theoretical foundations and references for the study of intelligent underwater equipment and the conservation of aquatic organisms.Among various research approaches,numerical simulation methods have gained popularity due to their advantages such as controllable simulation environments and abundant information acquisition.However,traditional numerical simulation methods can only simulate the passive motion of fish bodies,as fish behaviors are predetermined through computer programming.Simulated fish lack the active perception,decision-making,and interaction capabilities with the environment,there is a need for further advancements in tools and methodologies to address fish locomotion.Therefore,departing from the limitations of traditional research tools,this study integrates artificial intelligence techniques with computational fluid dynamics methods to conduct interdisciplinary research and applications.The main objectives and conclusions of the present work are outlined as follows:(1)A fish autonomous swimming simulation platform was designed through selfprogramming.The platform consists of three basic modules: flow field environment,fish body,and fish brain.The basic functionalities of each module were implemented using the Immersed Boundary-Lattice Boltzmann(IB-LB)fluid-structure coupling algorithm and deep reinforcement learning methods.An intelligent interface for data exchange was proposed to ensure the integration of the entire platform and the intelligence and continuity of the training process.(2)With the established smart fish autonomous swimming simulation platform,biological and behavioral guidelines from real fish were referenced to carefully design the state,action,and reward of the intelligent biomimetic fish.Extensive training was conducted for fish predation behavior in various flow field environments,gradually increasing the difficulty.In each derived class experiment of predatory swimming,the intelligent biomimetic fish consistently reached the food location following the ideal path,indicating good convergence of the fish brain decision-making model and reliable algorithms.(3)By incorporating lateral-line machine with the macro-action system,a method for flow field recognition and memory transfer in the intelligent biomimetic fish was proposed.This method enabled the adaptation of existing swimming strategies in unfamiliar flow field environments.It was introduced into the simulation platform,resulting in an intelligent fish locomotion controller that is more robust,exhibits stronger generalization ability,and simulates behavior more closely to real fish in the natural environment.Simulation of swimming under three different flow velocities of the Kárman vortex street was performed,comparing the control effects on the fish body with and without activating the system.The results demonstrated a significant enhancement in the fish’s self-maintenance of motion capability and motion stability in turbulent flow fields when the system was activated,and existing memories could be transferred to different flow fields.The first item mentioned above falls under the category of method exploration,providing a foundation and guarantee for the subsequent applied research.The latter two items represent applied innovation,supporting and validating the methods presented earlier.This thesis addresses existing problems by introducing deep reinforcement learning algorithm and combining it with traditional computational fluid dynamics methods to provide more advanced and powerful tools and methods for the study of fish awareness and behaviour.Moreover,this work also provides technical support for the development of ecological digital twins in various scenarios,such as intelligent channels,intelligent fishways,intelligent rivers,and intelligent water power stations.