Road-class-aware switching LQR control for real-time active suspension systems using machine-learning-based road identification

This study presents and experimentally validates a road-class-aware switching linear quadratic regulator (S-LQR) for a quarter-car active suspension system operating under ISO 8608 road classes. Four road-dependent LQR gains are designed offline after a statistical comparison of seven meta-heuristic optimizers, from which Battle Royale Optimization provides the best overall ranking. Online, an acceleration-based support vector machine estimates the road class once per second, and a Statistical Decision-Making Module (SD-MM) forms the switching command from non-overlapping three-prediction windows. A common quadratic Lyapunov function certifies nominal exponential stability of the continuous-time modes and their exact 1-kHz sampled-data realizations under every locally finite switching sequence; thus, the 3-s SD-MM interval is not a dwell-time requirement. Relative to the manufacturer-provided fixed-gain Quanser LQR, the proposed S-LQR reduces the RMS suspension-performance measures by 13.04–66.83% under variable-speed roads and by 9.59–51.43% under constant-speed roads. The ISO 2631-1 W k -weighted acceleration RMS decreases from 0.5143 to 0.1845 m/s 2 . Under the primary Delay-On assessment, the SD-MM attains 87.50% accuracy; its stable-segment accuracy reaches 100%, with no steady false switches. Ablation, robustness, measurement-corruption, and timing analyses further define the benefits and limits of the architecture. The complete S-LQR task has a mean execution time of 6.639 μs, a 99th-percentile time of 15.4 μs, and no missed deadline across 10 repeated 144-s real-time experiments.