黑料福利网

Topic 

Integrating AI with Control Theory for Complex Mobility Systems

Abstract

The synthesis of robust control strategies for modern mobility systems鈥攕panning maritime, rail, and automotive sectors鈥攔equires a delicate balance between mathematical rigor and real-time feasibility. While classical methodologies like Pontryagin鈥檚 Minimum Principle (PMP) and quasi-Linear Parameter-Varying (qLPV) control provide powerful frameworks for optimization and stability, their application is often constrained by two primary factors: the prohibitive computational cost of online optimization and the inherent difficulty of modeling high-order physical nonlinearities.

This talk presents three distinct applications where Artificial Intelligence (AI) is employed to address these specific challenges, serving as a bridge between control theory and operational requirements.

• Vessel Energy Management: We explore how the computational burden of solving optimal control problems online can be bypassed. By training a neural network to learn optimal State-of-Charge (SOC) trajectories generated offline via PMP, we enable real-time reference generation for a qLPV tracking controller.
• Railway Emission Mitigation: In scenarios where the physics of particle emissions are too complex for traditional analytical modeling鈥攕pecifically due to intricate tribology鈥攚e utilize Long Short-Term Memory (LSTM) networks. This data-driven model allows for the prediction and minimization of emissions where physical models are currently unavailable.
• Autonomous Vehicle State Estimation: To improve upon simplified bicycle models without resorting to overly burdensome nonlinear state-space representations, we demonstrate a hybrid observer approach. Here, a simple physical model is augmented with a neural network to capture residual dynamics, significantly enhancing estimation accuracy for real-time control. Autonomous Vehicle State Estimation: To address the limitations of simplified bicycle models, we present a hybrid observer. By augmenting a simple physical model with a neural network to capture residual complex dynamics, we achieve high-fidelity state estimation without the overhead of a massive nonlinear state-space representation.

Ultimately, this work illustrates that the synergy of AI and traditional control does not require abandoning physical intuition. Instead, it allows researchers to maintain the stability and structure of classical control while leveraging AI to handle the residual complexity of the real world.

About the speaker

S茅bastien Delprat is a Professor at the University of Polytechnique Hauts-de-France (UPHF) and a researcher at the LAMIH laboratory, a joint unit of the French National Centre for Scientific Research (CNRS). With a career rooted in control theory and numerical optimization, his work focuses on developing energy management strategies for complex systems. He has developed several experimental platforms, ranging from hybrid and autonomous vehicles to fuel cell test benches. His recent research includes fuel cell control, battery health estimation, and energy management for intermediate vehicles that accounts for human fatigue.

In addition to his research, he serves as the Director of the FRA TTM (the Transport and Mobility Research Federation) within the CNRS and is the Scientific Coordinator for the CPER RITMEA, a seven-year project gathering 26 laboratories to address mobility challenges in the Hauts-de-France region.

Your host

Theo Hofman, Full Professor at the department of Mechanical Engineering host Professor S茅bastien Delprat.

is required but free of charge.