This study proposes an improved deep deterministic policy gradient (IDDPG) controller for electromagnetic suspension systems to overcome the limitations of conventional maglev control strategies, particularly their dependence on precise mathematical models and challenges in real-world deployment. Leveraging reinforcement learning, the IDDPG approach achieves robust, model-free performance while meeting the stringent real-time requirements of magnetic suspension.

The system model is derived from electromagnetic force balance and Newtonian mechanics, yielding nonlinear coupled equations of coil current and air-gap displacement. These equations are linearized around the operating equilibrium to simplify controller design. Building on this foundation, the deep deterministic policy gradient (DDPG) algorithm is examined as a model-free actor–critic reinforcement learning method for continuous control. Recognizing its limitations in steady-state accuracy and transient response, we introduce a segmented inverse-proportional reward function that emphasizes small air-gap errors, accelerating convergence and improving response speed. To address hardware constraints, training is optimized by integrating network update latency and action–state delay into a unified control cycle, ensuring stable learning while reducing iteration time and execution delay on embedded platforms. The IDDPG controller is validated through simulations and hardware-in-the-loop experiments on a test rig replicating the suspension apparatus. Comparative studies with sliding mode control (SMC) and proportional–integral (PI) schemes demonstrate superior performance: steady-state error is reduced below 5% (vs. 31% with SMC and 12% with PI). Under parameter variations and disturbances, the controller maintains consistent performance with fixed hyperparameters, underscoring its robustness and generalization capability. Disturbance rejection tests further show that, compared to conventional PID control, IDDPG reduces overshoot by 51% and shortens adjustment time by 49%, yielding more stable levitation and lower mechanical stress. In summary, the IDDPG framework significantly improves control performance for electromagnetic suspension systems and expands the applicability of reinforcement learning in nonlinear control. By combining targeted reward function design, workflow optimization, and experimental validation, this work demonstrates a practical pathway toward deploying model-free, learning-based controllers in maglev and other precision suspension platforms.