Reinforcement learning-based control co-design of digital twin-enabled full-vehicle active suspension systems.

Area of Science:

• Automotive Engineering
• Control Systems
• Artificial Intelligence

Background:

• Active suspension systems are crucial for vehicle dynamics but limited by static designs.
• Digital Twins (DTs) and Reinforcement Learning (RL) offer potential for real-time optimization.
• Integrating DTs and RL for co-design of physical and control systems is a key challenge.

Purpose of the Study:

• To present a novel RL-based Control Co-Design (CCD) framework for full-vehicle active suspensions.
• To jointly optimize suspension hardware and control policies using multi-generation design and DT concepts.
• To address partial observability and data uncertainty in active suspension control.

Main Methods:

• Developed a DT-enabled Control Co-Design (CCD) framework integrating Deep Reinforcement Learning (DRL).
• Employed automatic differentiation within DRL for joint optimization of physical components and control policies.
• Incorporated quantile learning for uncertainty quantification and adaptive model updating.

Main Results:

• Achieved personalized optimization of active suspension systems for mild and aggressive driving settings.
• Demonstrated significant reductions in control efforts (58% mild, 12% aggressive).
• Showcased substantial improvements in ride comfort (17% mild, 28% aggressive).

Conclusions:

• The DT-enabled CCD framework with DRL and uncertainty-aware updating effectively optimizes active suspensions.
• The multi-generation design approach enables self-improving systems across the vehicle lifecycle.
• Personalized optimization enhances performance for diverse driving behaviors and conditions.