Deep reinforcement learning attitude control of stabilized platform for rotary steerable system based on extended state observer

This is the authors' abstract. We don't add key points for this paper.

A Deep Deterministic Policy Gradient with Extended State Observer (DDPG_ESO) method is proposed to address control accuracy and robustness issues in the attitude control of stabilized platforms for rotary steerable systems. A nonlinear extended state observer is developed to estimate the total system disturbance, which is then integrated into a deep reinforcement learning framework to enhance the controller's adaptive capabilities. The controller is trained using variable amplitude and frequency external disturbances to improve its ability to suppress these disturbances. The superiority of the DDPG_ESO controller is validated through simulation experiments, demonstrating rapid tracking of toolface angles, reduced overshoot, and maintained steady-state errors. Compared to Proportional Integral Derivative (PID), Active Disturbance Rejection Control (ADRC), and Deep Deterministic Policy Gradient (DDPG) controllers, the DDPG_ESO algorithm exhibits higher robustness and adaptability under varying operational conditions, effectively mitigating the impact of external disturbances and modeling errors. The results indicate potential for practical application in the drilling industry