Line-of-Sight Stabilization Control of an Electro-Optical Platform Based on an Improved TD3 Algorithm

The azimuth axis of a two-gimbal electro-optical stabilized platform is susceptible to nonlinear friction torque, external disturbances, and parameter variations during line-of-sight stabilization. To address these issues, an improved Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm incorporating multi-step returns and fuzzy dynamic action limiting, termed MSF-TD3, is proposed. In practical control, insufficient sensitivity to long-term control effects and pronounced steady-state output fluctuations may limit convergence and stability. A multi-step return mechanism is introduced to enhance the utilization of future reward information, while fuzzy adaptive action limiting constrains the control output according to the operating state. Simulation results show that, in variable-amplitude and variable-frequency target tracking, the root-mean-square error (RMSE) of MSF-TD3 is reduced by approximately 15.2%, 88.6%, and 87.2% compared with standard TD3, proportional-integral-derivative (PID) control, and extended sliding-mode disturbance observer (ESMO)-based control, respectively. In stepped target switching under compound disturbances, the corresponding reductions are approximately 1.7%, 16.1%, and 8.3%. Moreover, the root-mean-square (RMS) control input is reduced by approximately 69.7% relative to TD3. The results demonstrate the effectiveness of the proposed method in improving tracking accuracy, disturbance rejection, and control smoothness.