Intelligent Servo Algorithm Based On Reinforcement Learning And Centralized Computing Research

Motor control has always been favored in scientific and industrial fields.In recent years,Model Predictive Control(MPC)has been widely used in the field of power electronics and electric drive because of its simple principle and easy realization of multi-variable control.However,the control performance of MPC depends on the accurate system model and system parameters,and the mismatch of motor parameters will seriously affect the steady-state performance of the system;and it also has the disadvantage that the weight coefficient of the value function under multi-objective is difficult to determine.In response to the above problems,this paper uses reinforcement learning to optimize the traditional model predictive control algorithm;and proposes to use the EtherCAT field bus to build a centralized computing platform to reduce the cost of artificial intelligence practical applications.The main research work and achievements of this paper are as follows:(1)Combined with the characteristics that the selection of the weight coefficient will affect the total harmonic distortion(THD)of the threephase current,this paper proposes to use the Deep Deterministic Policy Gradient(DDPG)algorithm to realize the self-tuning of the value function weight coefficient.The simulation experiment of the proposed method is completed in Matlab/Simulink,and the comparison experiment with the same heuristic algorithm is carried out to verify the effectiveness of the algorithm.The experimental results show that this algorithm can realize the self-tuning of the weight coefficient and effectively reduce the current THD value.And compared with the chaotic particle swarm,the method proposed in this paper has better search accuracy and higher efficiency.(2)In order to solve the problem of decreasing control accuracy caused by the mismatch of model parameters,an adaptive extended observer based on reinforcement learning is proposed to reduce the dependence of model predictive control on parameter accuracy.According to the system model,the mathematical expression of prediction error is deduced,and the influence of prediction error on system performance is analyzed through experiments.According to the analysis,the Extended State Observer(ESO)was used to obtain the inductance error information to correct the inductance in the model predicted torque control,and combined with the Twin Delayed Deep Deterministic Policy Gradient(TD3)algorithm to achieve ESO Parameter self-tuning improves the robustness and accuracy of ESO.The simulation results show that the proposed method can effectively correct the inductance parameters and improve the anti-interference of the system.(3)A centralized computing framework for artificial intelligence algorithms applied to motor control is proposed to effectively solve the cost problem of artificial intelligence algorithms applied in industrial fields.The open source solution SOME implements the EtherCAT master station on the Linux operating system,designs a process synchronization scheme to complete the data exchange between the master station process and the Tensor Flow process,and modifies the EtherCAT process data object table to complete the data exchange between the master station and the slave module and motor drive.Through this platform,the validity and feasibility of the intelligent algorithm proposed above are further verified.Since the solution realizes the space-time multiplexing of the computing power of the graphics card,the application cost can be effectively reduced.After a brief cost comparison,the additional cost of the traditional solution when applied to 100 motors is 4.5 times that of this solution.The research results of this paper can provide useful reference and technical support for artificial intelligence combined with the traditional control method of PMSM.