In this paper, the spontaneous motion of the single joint is studied on the basis of the cortical neuron firing activity model, and the working principle of the closed-loop brain machine interface is analyzed from the perspective of the control theory. The Kalman filter and artificial neural network are used to design system decoders to replace the original system of spinal cord current, then the advantages and disadvantages of these two decoders are compared. Due to the dramatically decrease of the decoder in the absence of natural proprioception, the reinforcement learning algorithm(Actor-Critic) combined with the artificial neural network is used to design the PID controller, which can generate the stimulus signal to stimulate the neurons of the cerebral cortex, track cortical neuron firing activity with the natural proprioception and restore the performance of the decoder. Finally, the validity of the artificial sensing feedback signal design based on the reinforcement learning algorithm is verified by comparing with other control algorithms.