A novel magnetorheological semi-active suspension control using sliding window Z-score and soft actor-critic based on reinforcement learning

Magnetorheological damper (MRD) based semiactive suspensions improve ride comfort through real time damping control. Yet MRD nonlinearity, hysteresis, and changing road excitations fundamentally limit the performance consistency and generalization of conventional controllers. Reinforcement learning (RL) methods reduce model dependence, but state distribution shifts and inadequate online normalization during off policy training often degrade performance on unseen roads. This study proposes SAC–SWZ, combining sliding window Z score (SWZ) normalization with soft actor critic (SAC). An experimental frequency dependent forward MRD model and a lightweight neural network inverse model are embedded in training to represent actuator nonlinearities and constraints. Finite window statistics enable SWZ to standardize suspension states online and mitigate shifts across road classes and vehicle speeds. Across nine unseen road and speed conditions, SAC–SWZ outperformed classical semi-active controllers and conventional SAC variants, reducing sprung mass acceleration RMS by 14%–23% versus SH–ADD with better frequency domain attenuation. Hardware in the loop (HIL) tests on a digital signal processor confirmed real time feasibility, ride comfort and road holding gains, and lower actuation demand. HIL coupled simulations without retraining examined fixed policy sensitivity to sprung mass variations and actuator and model output perturbations. Results show online normalization is critical to RL controller generalization and support SAC–SWZ for practical, scalable embedded intelligent MRD control.