Proposal of a Data-Driven Weakly Supervised Learning Method for Operating Mode Classification of Fuel Cell Garbage Trucks

Yida Bao･Xiang Zhang･Yiyuan Fang･Wei-hsiang Yang･Yushi Kamiya (Waseda University)

This study proposes a versatile weakly supervised deep learning framework for operating mode classification, which enables statistical feature analysis of Fuel Cell (FC) refuse trucks. By fusing unsupervised clustering of CAN signals with expert knowledge, the method automatically generates large-scale training data, eliminating the need for expensive manual annotation. The proposed model achieves high classification accuracy while reducing annotation costs by over 75%. This framework enables efficient analysis for massive real-world driving data, establishing a foundational technology that has an ability to contribute to the broader design and control strategies of next-generation commercial vehicles.

The structure of "anticipatory mechanisms" in autonomous vehicles

Masao Ito (NIL)

Previous research on autonomous vehicles and ethics has revealed that "anticipation mechanisms" (AM) are important for safety. They also have the potential to reduce the burden of scenario-based testing. This paper presents the results of an investigation into patterns for incorporating AM into autonomous vehicles.

PhysicsAI: Accelerating Automotive Design with Graph Neural Network-Based CFD and NVH Engineering

Son Tong･Marc Brughmans･Andrey Hense･Lester Deleon･Theo Geluk (Siemens Digital Industries Software)

PhysicsAI delivers fast physics predictions enabling engineering teams to generate design variations rapidly. PhysicsAI learns physics behavior using Graph Neural Networks (GNNs) trained on mesh geometries and CAD models data. Engineers can explore various design variations, optimize parameters, and accelerate innovation. We present two applications: (1) External aerodynamic drag prediction using CFD simulation data, achieving high accuracy while reducing computation time from hours to minutes; (2) Vibration mode shape recognition and classification for NVH optimization, demonstrating expert-level accuracy on complex automotive structures. Validation from comprehensive automotive datasets will be presented.

Considerations regarding the safety assurance of AI-based Automated Driving Systems

Olaf OP DEN CAMP･Jan-Pieter Paardekooper (TNO)

In the development of Automated Driving Systems manufacturers and AV-developers make more and more use of AI-based systems. In some cases, even an end-to-end (E2E) AI approach is followed in which no longer a distinction is made between perception, path planning and actuation in the ADS of the vehicle. The paper presents considerations regarding the safety assurance of AI-based systems. The vulnerabilities of AI-based systems and the negative impact of these vulnerabilities on safety assurance will be discussed. It will be shown how the design of AI-based systems can be improved to allow for proper safety assurance.