⚡ electrical engineering

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

This paper introduces a robust and explainable framework for 3D mode shape recognition that utilizes a Canonical Engineering Graph Representation to transform heterogeneous finite element and experimental data into a geometry-independent graph, enabling accurate, transferable, and physically interpretable predictions across diverse vehicle architectures without relying on manual inspection or identical mesh topologies.

Original authors: Tong Duy Son, Marc Brughmans, Andrey Hense, Kohta Sugiura, Sebastian Ciceo, Paolo di Carlo, Theo Geluk

Published 2026-07-03

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Tong Duy Son, Marc Brughmans, Andrey Hense, Kohta Sugiura, Sebastian Ciceo, Paolo di Carlo, Theo Geluk

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a computer to recognize how a car's body (specifically the "Body-in-White," or the bare metal frame before paint and seats are added) vibrates. Engineers call these vibrations "mode shapes." Some vibrate like a twisting pretzel (torsion), some like a diving board (bending), and others like a drum skin (pumping).

Currently, experienced engineers have to look at complex 3D computer models and manually say, "Ah, this one is a twisting mode." This is slow, expensive, and hard to scale.

This paper proposes a new way to teach computers to do this automatically, but with a twist: instead of teaching the AI to look at the raw, messy details of the car's metal mesh, the authors teach it to look at the car's soul (its structural meaning).

Here is the breakdown of their solution using simple analogies:

1. The Problem: The "Pixel" vs. The "Concept"

Imagine you have two photos of a house.

Photo A is a high-resolution picture of a brick house.
Photo B is a sketch of a wooden house.

If you try to teach a computer to recognize "a house" by looking at the individual pixels (bricks vs. wood grain), the computer gets confused. It thinks Photo A and Photo B are completely different things because the "pixels" (the specific metal nodes in the car model) don't match up.

In the car industry, every new car model has a different "mesh" (a digital grid of millions of tiny points). Old AI methods tried to learn from these specific grids. If you trained the AI on Car Model A, it would fail miserably on Car Model B because the grid looked different, even if the cars were built the same way.

2. The Solution: The "Canonical Engineering Graph"

The authors created a new way to represent the car, which they call a Canonical Engineering Graph.

Think of this like translating a complex novel into a simple summary of its main characters.

Instead of looking at millions of tiny metal points, the AI looks at named structural regions: "The Roof Rails," "The Pillars," "The Floor," and "The Side Sills."
These regions are like the "main characters" of the car. Whether the car is made of steel or aluminum, or has a different number of bolts, the "Roof Rail" is always the "Roof Rail."

They connect these regions with lines (edges) that represent how they physically talk to each other (e.g., "The roof is connected to the pillars"). This creates a map that is the same for every car, regardless of how the underlying metal was cut or measured.

3. How the AI Learns: The "Smart Detective"

Once the car is translated into this simple map of regions, the AI (a Graph Neural Network) gets to work.

The Detective: The AI acts like a detective looking at the "crime scene" (the vibration).
The Clues: It doesn't just look at one spot; it looks at how the "Roof" moves compared to the "Floor."
The Logic: If the Roof and Floor move up and down together, the AI learns this is a "Pumping" mode. If the left side goes up while the right goes down, it's a "Twisting" mode.

Crucially, the AI is also given "engineering cheat sheets" (regional descriptors). It doesn't just guess; it combines its own learning with rules engineers already know (e.g., "If the floor vibrates heavily, it's likely a bending mode").

4. Why This is a Big Deal

The paper tested this on four different car programs (some from computer simulations, some from real physical tests).

The Result: The AI achieved nearly 100% accuracy in identifying the vibration types.
The Superpower: Because it learned the "concepts" (Roof, Pillar, Floor) rather than the "pixels" (specific metal nodes), it could take what it learned from Car A and apply it perfectly to Car B, even though Car B had a completely different shape and measurement setup.
The Explanation: Unlike "black box" AI that just gives an answer, this system can point to the map and say, "I know this is a twist because the Pillars and Side Sills are moving in opposite directions." This matches how human engineers think.

Summary

The paper argues that to make AI useful for engineers, we shouldn't just feed it more data or bigger computer models. Instead, we need to translate the data into the language engineers speak.

By turning messy, complex 3D car models into a simple, consistent map of structural parts (like a family tree of the car's bones), they created an AI that is:

Robust: It works on different cars without needing to be retrained from scratch.
Explainable: It tells you why it made a decision using terms engineers understand.
Transferable: Knowledge gained from one car program can be instantly used on the next.

In short, they taught the AI to stop counting pixels and start understanding the story of the car's structure.

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

Problem Statement
Body-in-White (BiW) mode shape recognition is a critical task in automotive Noise, Vibration, and Harshness (NVH) development, traditionally reliant on manual visual inspection by experienced engineers. Existing automated approaches face three primary limitations:

Lack of Transferability: Current methods, often based on engineering heuristics, Modal Assurance Criterion (MAC), or geometry-dependent AI representations, struggle to generalize across different vehicle architectures, finite element (FE) meshes, and experimental measurement layouts.
Data Scarcity: Obtaining expert-reviewed mode labels is expensive and time-consuming, leading to severe label scarcity in industrial datasets.
Explainability Gap: Many AI models function as black boxes, failing to provide predictions that align with engineering intuition or the specific structural entities (e.g., pillars, side sills) used by engineers during modal assessment.

Consequently, valuable engineering knowledge remains fragmented, and AI models cannot be easily reused across historical data or different vehicle programs.

Methodology
The paper proposes a Region-Aware Graph Learning Framework built upon a Canonical Engineering Graph Representation. The core philosophy is to decouple engineering semantics from numerical discretization, allowing learning to occur on persistent engineering concepts rather than vehicle-specific mesh topologies.

The framework operates through four key stages:

Canonical Engineering Graph Representation:
- Heterogeneous FE models and experimental wireframes are transformed into a common semantic graph.
- Nodes represent persistent, engineering-defined structural regions (e.g., roof rails, A/B/C-pillars, side sills, floor structures) rather than individual FE nodes.
- Edges encode physically meaningful relationships, including structural adjacency, left-right symmetry, longitudinal load-path coupling, and vertical roof-floor coupling.
- Features: Node features are geometry-independent regional descriptors (mean/RMS displacement, dominant direction, signed vertical response, deformation energy) aggregated from the original data. Edge attributes capture interaction characteristics like phase agreement and coupling.
Region-Aware Graph Learning:
- A Graph Attention Network (GAT) propagates information between engineering regions to learn structural interactions associated with vibration modes.
- Engineering-Aware Feature Fusion: Instead of relying solely on learned embeddings, the framework concatenates the global graph embedding ( $z_{GNN}$ ) with analytical, engineering-informed regional descriptors ( $z_{eng}$ ) (e.g., floor/roof energy fractions). This fusion helps distinguish physically similar mode families.
Hierarchical Prediction:
- The model employs a hierarchical classification strategy mirroring engineering review processes: first predicting the dominant mechanism (Level 1: torsion, bending, pumping), then refining to subtypes (Level 2).
- Training uses a multi-task objective function balancing Level 1 and Level 2 cross-entropy losses.
Dataset Generation:
- To address label scarcity, the authors utilize a physics-guided data augmentation strategy. Starting from a small set of expert-reviewed modes, they generate structural variants via controlled stiffness modifications on concept beam models.
- Labels are transferred automatically using MAC within the augmented dataset, as the underlying engineering wireframe remains consistent.
- The benchmark dataset comprises four heterogeneous vehicle programs (two FE simulations, two experimental tests) with varying node counts and topologies, all mapped to the same canonical graph.

Key Contributions

Canonical Engineering Graph Representation: A novel representation that maps heterogeneous BiW models into a common semantic space independent of mesh topology and sensor configuration. This enables robust learning across different vehicle architectures and data sources.
Region-Aware Graph Learning Framework: A learning architecture that combines graph attention mechanisms with explicit engineering-informed regional descriptors to achieve explainable, hierarchical mode shape classification.
Demonstrated Cross-Vehicle Transfer: Validation showing that the framework achieves high classification accuracy and transferability across different vehicle programs, FE discretizations, and experimental layouts under severe label scarcity, without requiring identical mesh topologies.

Experimental Results
The framework was validated on a multi-vehicle benchmark containing 310 augmented training samples and a held-out test set from four distinct vehicle programs.

Accuracy: The multi-vehicle training approach achieved 100% Level-1 accuracy and 98.7% Level-2 accuracy (99.2% combined) on the held-out test set. In contrast, a model trained only on a reference vehicle achieved only 81.6% Level-2 accuracy on unseen vehicles.
Transferability: The model successfully generalized to vehicle programs with only 2–5 manually reviewed labels, demonstrating that the learned representation captures persistent engineering concepts rather than vehicle-specific numerical artifacts.
Explainability: Attention maps and feature importance scores directly correspond to recognizable structural regions (e.g., pillar-side sill connections for torsion, roof-floor coupling for bending), aligning with established engineering interpretations.

Significance and Claims
The paper asserts that the primary barrier to transferable Engineering AI is not the sophistication of the neural network architecture, but the representation of the engineering data. By shifting from geometry-dependent mesh discretizations to a canonical representation based on persistent engineering semantics, the framework enables:

Knowledge Reuse: Historical engineering knowledge can be accumulated and transferred across successive vehicle programs rather than being relearned from scratch.
Trustworthy AI: Predictions are physically interpretable, allowing engineers to verify decisions against known structural mechanisms.
Industrial Scalability: The approach reduces the need for extensive vehicle-specific retraining and large annotated datasets, making AI deployment feasible in resource-constrained industrial workflows.

The authors conclude that the Canonical Engineering Graph Representation serves as a reusable engineering abstraction, providing a foundation for trustworthy and transferable AI across heterogeneous simulation and experimental workflows in automotive development.