Research on Many-Objective Parameter Optimization of Variable-Speed Axial Blood Pump Controller Based on Deep Reinforcement Learning

Axial blood pumps serve as vital auxiliary therapeutic devices for patients with end-stage heart failure, and parameter optimization of the controller is critical to improve system performance. Existing optimization methods cannot satisfy the parameter optimization requirements of variable-speed axial blood pump controllers in terms of optimization accuracy. Therefore, this paper investigates a many-objective optimization method adapted to the operating characteristics of blood pumps. Firstly, a many-objective optimization model is established for the controller. To efficiently solve the proposed model, an optimization algorithm integrating deep reinforcement learning, named DQN-NSGA-CT, is developed. On the basis of the population evolution state, the optimal strategy learned by DQN dynamically adjusts the crossover probability and mutation probability, which adaptively balances population diversity in the early iteration stage and convergence speed in the later iteration stage. Meanwhile, a novel environmental selection strategy is used to reconcile population convergence and diversity. To select the best compromise solution, an entropy weight–Copula–TOPSIS comprehensive evaluation method is proposed, which realizes objective weight assignment, objective correlation correction and multi-attribute ranking. Experimental results verify the efficiency of the DQN-NSGA-CT algorithm in solving the many-objective optimization model of the controller. The research addresses the many-objective optimization problem of variable-speed axial blood pump control.