New region-aware GNN automates 3D mode shape recognition across vehicles
A graph learning framework that works across different car models without retraining...
A team of researchers from academia and industry (including Tong Duy Son, Marc Brughmans, and others) has published a paper on arXiv presenting a robust and explainable AI approach for 3D mode shape recognition—a critical but manually intensive task in automotive NVH (Noise, Vibration, Harshness) development. Traditional methods rely on expert visual inspection or geometry-dependent AI that fails across different vehicle architectures, finite element meshes, and experimental setups. Their solution uses a Canonical Engineering Graph Representation combined with region-aware graph neural networks (GNNs). Instead of learning directly from vehicle-specific meshes, the system converts heterogeneous FE models and experimental measurements into a common graph where nodes represent semantically meaningful structural regions (e.g., panels, joints) connected by engineering-informed relationships. Graph attention learning and region-aware pooling capture structural interactions while preserving engineering semantics, enabling predictions that are both accurate and physically interpretable.
The framework was validated using FE and experimental datasets from four distinct vehicle programs under severe label scarcity. Results show high classification accuracy, cross-vehicle transferability (no need for identical mesh topology or sensor configurations), and physically meaningful explanations—predictions directly relate to engineering-defined structural regions used in NVH analysis. This work addresses a major pain point in automotive engineering: the reliance on manual, subjective visual inspection of mode shapes. Beyond mode shape recognition, the Canonical Engineering Graph Representation offers a reusable engineering abstraction for trustworthy and transferable AI across heterogeneous simulation and experimental workflows, potentially reducing development time and cost for vehicle programs by eliminating the need for retraining on each new architecture.
- Transforms heterogeneous FE models and experimental data into a common graph with semantics-driven region nodes
- Achieves high classification accuracy and cross-vehicle transferability across four vehicle programs without identical mesh topology
- Provides physically interpretable predictions by linking outputs to engineering-defined structural regions
Why It Matters
Automates manual NVH inspection, saving hundreds of engineering hours and accelerating vehicle development across diverse programs.