Research Report

← Home

Research Report: GNN-Based SHM for Payload Fairings

1. Target Structure Specification: JAXA Epsilon/H3 Fairing

1.1 Material Composition

Research confirms that modern payload fairings for Epsilon (Enhanced/S) and H3 launch vehicles are Carbon Fiber Reinforced Polymer (CFRP) Sandwich Structures.

• Face Sheets: CFRP (Carbon Fiber/Epoxy). High specific stiffness.
• Core: Aluminum Honeycomb. High crushing strength and stiffness-to-weight ratio.
• Manufacturing: Often Out-of-Autoclave (OOA) or filament wound (for monolithic parts), but fairings are typically large sandwich shells made in halves. Kawasaki Heavy Industries (KHI) and Beyond Gravity (formerly RUAG) are key manufacturers.

1.2 Geometric Parameters (Estimates for Simulation)

• Diameter:
  • Epsilon: ~2.6m (Max diameter).
  • H3: ~5.2m (Standard fairing).
• Structure:
  • Cylindrical section + Ogive nose.
  • Construction: Two half-shells (Clamshell) held by separation bolts.

Implication for Modeling:

• The simulation MUST model a sandwich structure (Shell-Solid-Shell or Homogenized Shell).
• Anisotropy: CFRP face sheets are highly anisotropic. Wave velocity depends on propagation direction relative to fibers.
• Curvature: The cylindrical geometry causes mode conversion and focusing/defocusing of guided waves, unlike flat plates.

---

2. Physics of Guided Waves in Honeycomb Sandwich

2.1 Propagation Characteristics

• Leaky Waves: In sandwich structures, wave energy often leaks from the skin into the core and the opposite skin.
• High Attenuation: The honeycomb core (and adhesive layer) acts as a damper. High-frequency waves (>100 kHz) attenuate rapidly compared to monolithic plates.
• Modes:
  • Global Modes: Flexural waves of the entire sandwich (low frequency).
  • Skin Modes: Rayleigh-like waves confined to the skin (high frequency). This is preferred for detecting skin-core debonding.

2.2 Defect Signature (Debonding)

• When the skin separates from the core (debonding), the local boundary condition changes from "supported" to "free".
• Trapped Energy: Waves entering the debonded region often get "trapped," causing amplitude amplification and ringing (standing waves) within the debonded skin.
• Velocity Change: Phase velocity decreases in the debonded region (thin plate vs. sandwich).

---

3. GNN Methodology for Curved Shells

3.1 Graph Construction (Mesh-to-Graph)

Standard CNNs fail on curved surfaces due to distortion in 2D projection. GNNs operate directly on the non-Euclidean manifold.

• Nodes: Finite Element nodes or sensor locations.
• Edges: Connected if geodesic distance < threshold $R$, or k-Nearest Neighbors.

3.2 Feature Engineering for Curvature

To make the GNN "geometry-aware," node features must include:

1. Coordinates: $(x, y, z)$ (Global position).
2. Surface Normals: $(n_x, n_y, n_z)$. The change in normal vectors between neighbors encodes local curvature.
   • Insight from StructGNN: Explicitly encoding structural properties (stiffness, mass) into the graph improves physics compliance.
3. Material Direction: Vector defining the fiber orientation $(v_1, v_2, v_3)$ at each node.

3.3 Architecture Candidates

• StructGNN / Physics-GNN: Architectures that mimic the mass-stiffness matrix operations.
• GraphSAGE / GAT: Standard inductive frameworks. Good for learning local wave scattering patterns.
• Kernel GNN: specialized for continuous manifolds.

---

4. Sim-to-Real Domain Adaptation

4.1 The Problem

FEM data is "clean" and idealized. Real experimental data has:

• Sensor coupling variability.
• Environmental noise.
• Manufacturing variability (thickness variations, resin rich zones).

4.2 Strategy: Transfer Learning

• Source Domain: Abaqus Simulation (Rich labels, infinite data).
• Target Domain: Open Guided Waves Dataset (Limited labels, real physics) / Future Epsilon Test Data.
• Method: Domain Adversarial Neural Networks (DANN) or Transfer Component Analysis (TCA).
  • Align the feature distributions of FEM and Experiment in a latent space.
  • Train the classifier on FEM features, ensuring they are indistinguishable from Experimental features.

---

5. Next Steps (Actionable)

1. Refine Abaqus Model: Switch to "Composite Layup" tool in Abaqus to model Skin-Core-Skin sandwich explicitly.
2. Frequency Selection: Target 50-100 kHz range where "Skin Modes" begin to dominate but attenuation is manageable.
3. GNN Prototype: Build a simple GraphSAGE model using $(x, y, z, \text{time_series})$ as input to predict defect location.