A reinforcement learning-based approach to valve plate optimization for reducing flow ripple in an axial piston pump

Axial piston pumps are indispensable components in fluid power systems, yet mitigating flow ripple remains a persistent challenge. This study presents a reinforcement learning-based optimization framework for valve plate design, employing the Soft Actor-Critic (SAC) algorithm. Within a simulated fluid dynamic environment, the agent adaptively adjusts the valve plate geometry in response to real-time pressure and flow variations. To enhance learning efficiency, a Lyapunov-based reward function is introduced, ensuring stable and accelerated convergence toward the optimal control policy. Simulation results show that the proposed SAC approach reduces inlet and outlet flow ripples by 20.4% and 26.2% compared with the original design, while achieving substantially faster convergence than the Multi-Objective Genetic Algorithm (MOGA) and Multi-Objective Particle Swarm Optimization (MOPSO). Experimental validation further demonstrates reductions of 42.4% and 52.6% in inlet and outlet pressure ripples, respectively. Moreover, the proposed method exhibits strong generalization, enabling efficient optimization of diverse valve plate topologies using pre-trained agents. These results highlight reinforcement learning as a powerful and efficient paradigm for valve plate optimization, offering distinct advantages over conventional heuristic algorithms.