Magnetic Field Compensation Control for Spin-Exchange Relaxation-Free Comagnetometer Using Reinforcement Learning

The triaxial drift magnetic field compensation (TDMFC) is a prerequisite to maintaining the excellent per-formance of the spin-exchange relaxation-free comagnetometer (SERFCM). In this article, we develop a previously undescribed controller design architecture running the proposed constrained dynamic action space Q-learning (CDA-Q) algorithm using reinforcement learning (RL) to solve the TDMFC problem. The architecture contains two parts: offline training and online deployment. Specifically, the CDA-Q algorithm trains the agents with the simulated environment to produce the control strategies adopted in the online deployment. Numerical simulations verify the effectiveness of the obtained control strategies. Experimentally, the control strategies are deployed in the real-time control system achieving efficient and adaptive compensation of the triaxial drift magnetic field. Comparative experiments show that the proposed method is 67.56% more efficient than the existing method.

Automated summary

In this article, a previously undescribed controller design architecture using the constrained dynamic action space Q-learning algorithm is developed to solve the triaxial drift magnetic field compensation problem for the spin-exchange relaxation-free comagnetometer. The architecture consists of offline training and online deployment, where the control strategies are obtained through the CDA-Q algorithm trained with simulated environment. Numerical simulations and comparative experiments demonstrate the effectiveness and efficiency of the proposed method, which outperforms the existing method by 67.56%.