Mechanism-data-driven control strategy for active suspension systems: Integrating deep reinforcement learning with differential geometry to enhance vehicle ride comfort

Long-standing research on vehicle comfort optimization has centered on active and semi-active suspension control using experience-based or optimization-based algorithms. However, these methods often require substantial engineering resources and pose challenges in acquiring theoretical knowledge. The emergence of advanced Artificial Intelligence (AI), particularly data-driven approaches, has transformed how engineers tackle knowledge-intensive tasks like suspension control. Yet, the interpretability challenges of data-driven methods limit their widespread use in engineering. This study proposes a mechanism-data-driven active suspension control strategy that integrates Differential Geometry (DG) and Deep Reinforcement Learning (DRL) to achieve theoretical fusion of mechanism and data models. A DRL control architecture (DGRL) based on DG theory is introduced, enabling mechanism-level analysis of suspension control and dividing the control strategy into mechanism and data models. For the data model, a DRL optimal control framework is constructed, incorporating the Twin-Delayed Deep Deterministic policy (TD3) with an expert-guided soft-hard module (TD3-SH) and the Deterministic Experience Tracing (DET) mechanism. This effectively explores and utilizes the knowledge in massive data. Simulation results show that the DGRL strategy outperforms baseline algorithms such as Deep Deterministic Policy Gradient (DDPG), TD3, Linear Quadratic Regulator (LQR), Model Predictive Control (MPC), and TD3-SH by 75.8%, 65.5%, 77.5%, 56.3%, and 46.5%, respectively. In complex environments with varying road features and considering the domain randomization of the suspension system, the DGRL strategy can improve ride comfort by up to 85%, demonstrating its robustness and significant potential for widespread application in industrial and real-world scenarios.