Robust and Explainable 3D Mode Shape Recognition Using Region-Aware Graph Neural Networks

In automotive NVH (Noise, Vibration, and Harshness) development, engineers must frequently classify the vibration mode shapes of Body-in-White (BiW) structures. Traditional methods rely on manual visual inspection or geometry-dependent AI models that struggle to generalize across different vehicle architectures, mesh discretizations, and sensor layouts. This paper addresses the challenge of creating an AI framework that is robust, explainable, and transferable across heterogeneous engineering datasets.

The researchers propose a Canonical Engineering Graph Representation. Instead of training models directly on raw finite element (FE) meshes—which are highly sensitive to node numbering and mesh density—the authors transform various FE models and experimental measurements into a common semantic graph. In this graph, nodes represent persistent structural regions (e.g., roof rails, pillars, side sills) and edges encode engineering-informed physical relationships. The framework uses a region-aware graph attention network (GAT) combined with engineering-informed regional descriptors to classify mode shapes. This approach decouples the learning process from the underlying numerical discretization, allowing the model to focus on structural semantics.

The proposed framework demonstrates high classification accuracy and strong cross-vehicle transferability, even under conditions of severe label scarcity. By testing on four different vehicle programs (including both simulation and experimental data), the authors show that the model maintains performance despite significant differences in vehicle geometry and sensor configuration. Furthermore, the model provides physical interpretability; because the graph nodes correspond to recognizable structural components, the system can highlight which specific regions contribute to a classification, allowing engineers to verify the model's reasoning against their own domain knowledge.