Machine-Learning-Based PID Gain Scheduling for DC Motor Speed Control Under Load Disturbances

This paper proposes a machine-learning-based PID gain scheduler for DC motor speed control under step changes in load torque. It effectively resolves a key limitation of conventional fixed-gain PID control, where the motor speed exhibits noticeable sag during load application and undesirable overshoot when the load is removed. A DC motor model is implemented in MATLAB/Simulink, and locally optimal PID gains are first obtained offline for multiple operating points defined by reference speed (ωref) and load torque (TL). These data form a supervised-learning dataset mapping (ωref, TL) to PID gains (Kp, Ki, Kd) and a k-nearest-neighbour (k-NN) regression algorithm is used online to select the gains corresponding to the closest operating condition. To avoid large transients when switching between gain sets, slew-rate limiting is applied to the scheduled gains, yielding smooth gain transitions suitable for embedded implementation. Simulation results show that the proposed ML-based gain scheduling significantly reduces speed sag during load application while keeping overshoot small when the disturbance is removed, outperforming a conventional fixed-gain PID controller.