Optimal control-guided reinforcement learning for turbojet engine speed regulation

Miniature turbojet engines are critical components in emerging low-altitude aerial systems, where accurate and adaptive control is required due to nonlinear and time-varying dynamics. This paper proposes a reinforcement learning-based control approach for turbojet engine speed regulation using proximal policy optimization. The controller is trained in simulation with an optimal control-inspired reward design incorporating tracking accuracy, control effort constraints, and action smoothness. The learned policy is directly deployed onto real engine systems without additional fine-tuning. Experimental validation is conducted on multiple engines to evaluate cross-engine generalization, and dynamic performance is further assessed under stepwise throttle inputs across a wide operating range. Both simulation and real-world results demonstrate that the proposed controller achieves consistently faster transient response than the onboard Proportional-Integral-Derivative (PID) controller while maintaining smooth and stable actuation. Under low spool speed operating conditions, the rise time is reduced to approximately 0.55 s, compared to 2.52 s for the PID controller, representing a reduction of approximately 78%. The controller also exhibits strong robustness to measurement noise and adapts effectively to unseen operating conditions. These results demonstrate the effectiveness of the proposed approach for real-time turbojet engine control and highlight the potential of reinforcement learning in safety-critical nonlinear systems.