Research on Magnetorheological Semi-Active Suspension Control Using RBF Neural Network-Tuned Active Disturbance Rejection Control

Magnetorheological (MR) semi-active suspensions offer clear advantages in improving ride comfort and handling stability, yet their engineering applications are often hindered by strong nonlinear hysteresis of the damper, the randomness of road excitations, and the reliance on manual tuning of controller parameters. To address these issues, this paper proposes an integrated framework of “experimental modeling–semi-active implementation–adaptive control.” First, characteristic tests of the MR damper are conducted, based on which a current-dependent Bouc–Wen forward model is established. Tianji’s Horse Racing Optimization (THRO) is then employed for parameter identification to reproduce the hysteresis behavior accurately. Second, a back propagation (BP) neural network-based inverse current model is developed to achieve rapid mapping from “desired damping force” to “driving current,” enabling semi-active actuation. Furthermore, a radial basis function (RBF) neural network is embedded into the active disturbance rejection control (ADRC) structure to estimate the system Jacobian online and to tune key extended state observer (ESO) gains in real time, forming the proposed RBF-ADRC strategy and thereby enhancing disturbance observation and compensation capability. Simulation results under pulse-road and Class-C random-road excitations show that, compared with the passive suspension, the proposed method reduces the root mean square error values of sprung-mass acceleration, suspension dynamic deflection, and tire dynamic load by 25.14%, 18.71%, and 11.61%, respectively, while also outperforming skyhook control and fixed-gain ADRC. Frequency-domain results further show stronger attenuation in the low-frequency band relevant to body vibration. Under pulse excitation, RBF-ADRC yields smaller peak and trough body accelerations and faster post-impact recovery. Under ±30% sprung-mass variations, it achieves the best worst-case and fluctuation-range robustness among the compared strategies and remains close to offline retuning. These results demonstrate that the proposed method improves both control performance and robustness while reducing the need for repeated manual calibration.