⚡ electrical engineering

NDT Data Fusion for Post-Impact Damage Characterization in a Discontinuous Long Fiber Reinforced Polymer Composite with Stochastic Meso-Structure

This study demonstrates that pixel-level data fusion of ultrasonic testing, infrared thermography, and X-ray micro-computed tomography effectively overcomes the limitations of individual non-destructive techniques to provide a more complete and conservative characterization of impact damage in Prepreg Platelet Molded Composites with complex stochastic meso-structures.

Original authors: Marco Didone, Rachel Lear, David Jack, Sergii G. Kravchenko

Published 2026-08-05

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Marco Didone, Rachel Lear, David Jack, Sergii G. Kravchenko

Original paper licensed under CC BY 4.0 (https://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 a detective trying to solve a mystery inside a very strange, jumbled house. This house isn't built with neat, straight bricks; instead, it's constructed from thousands of tiny, randomly tossed wooden planks glued together. This is what scientists call a "discontinuous long fiber composite"—a super-strong material used in airplanes and cars because it's light and tough. But because the "planks" (fibers) are thrown in random directions, the house is full of hidden nooks, crannies, and weak spots that behave unpredictably.

Now, imagine someone throws a heavy ball at this house. It might not leave a giant hole, but it could cause invisible cracks and layers to peel apart inside, like a sandwich where the filling has shifted but the bread looks fine. This is called "low-velocity impact damage." The problem is, if you only look at the house with one pair of eyes, you might miss the damage. One pair of eyes might see a shadow but not the crack; another might see a heat spot but not the deep split. To solve the mystery, you need to combine clues from different sources—like using a flashlight, a thermal camera, and an X-ray machine all at once. This is the heart of the research: figuring out how to mix these different "super-senses" to see the whole picture of the damage without tearing the house apart.

The Paper's Mission: Mixing Super-Senses to See the Invisible

In this study, researchers Marco Didone, Rachel Van Lear, David Jack, and Sergii G. Kravchenko tackled the tricky job of inspecting these random-plank houses after they've been hit. They used a special material called Prepreg Platelet Molded Composites (PPMCs), which are made by compressing tiny, randomly oriented strips of carbon fiber. When these materials get hit, the damage is messy and three-dimensional, spreading in weird directions that a single inspection tool often misses.

The team decided to play a game of "combine your clues" using three different Non-Destructive Testing (NDT) methods. Think of these as three different detectives:

Ultrasonic Testing (UT): This is like sending out sound waves (sonar) to listen for echoes. If there's a crack or a delamination (layers peeling apart), the sound bounces back differently. However, just like in a noisy room, sometimes the sound gets blocked by the top layers, hiding what's happening deeper down.
Infrared Thermography (IRT): This detective uses heat. They warm up the surface and watch how it cools down. If there's a crack underneath, the heat gets stuck or flows differently, creating a hot or cold spot on the surface. But this detective mostly sees what's near the surface and might miss deep secrets.
X-ray Micro-Computed Tomography (µCT): This is the high-tech X-ray vision. It takes thousands of pictures from different angles to build a 3D model of the inside. It's great for seeing the shape of the damage, but it can't always see cracks that are squeezed so tight they look like solid material, and it takes a long time to scan big pieces.

The Big Discovery: The "Union" is the Answer

The researchers didn't just look at the results from these three detectives separately; they tried to fuse them together. They used two main strategies:

Feature Integration: This is like taking the final "yes/no" maps from each detective and laying them on top of each other. If any detective says, "There's damage here," they mark it as damaged.
Feature Classification: This is more like mixing the raw data (the greyscale images) mathematically before making a decision, using rules like "take the average" or "take the maximum."

What They Found

The study found that Feature Integration was the clear winner for understanding the damage. When they combined the maps, they created a "conservative upper-bound estimate" of the damage. In plain English, this means they drew a safety net around the damage that was bigger than what any single tool saw, ensuring they didn't miss anything.

Here are the specific numbers they found:

The µCT (X-ray) saw the smallest area, covering 5.3% of the plate.
The UT (Sound) saw a larger area of 7.1%.
The IRT (Heat) saw 7.5%.
When they combined all three into one fused map, the total projected damage area was 8.5%.

This 8.5% isn't necessarily the "perfect" truth, but the authors suggest it is a very safe, complete picture of the damage. They noted that 99.66% of the damage seen by the X-ray was included in this combined map. This proves that by using all three tools, you catch almost everything the X-ray sees, plus the extra bits that the sound and heat tools found that the X-ray missed.

Why One Detective Isn't Enough

The paper explicitly argues against relying on just one method.

The X-ray (µCT) missed some tight cracks because the two sides of the crack were touching, leaving no air gap for the X-rays to see. It also couldn't see everything because of the weird shape of the plate.
The Sound (UT) sometimes saw a bigger area than reality because the sound waves get blocked by top layers (shadowing), making the damage look like a solid blob rather than a specific shape.
The Heat (IRT) was great at seeing near-surface issues but couldn't see deep down.

The researchers also tested the "Feature Classification" method (mixing the raw images mathematically). They found this approach was less stable and more sensitive to noise. For example, the "Difference" method (looking only where the tools disagreed) showed a tiny area of 0.34%, while the "Maximum" method showed 6.15%. These numbers were all over the place compared to the clean, logical result of the Feature Integration method. The paper suggests that for these messy, random-plank materials, simply combining the final maps (Integration) is much better than trying to mathematically blend the raw images (Classification).

The Takeaway

The authors conclude that for these complex, jumbled materials, you can't trust a single "super-sense." Instead, you need to listen to all of them. By fusing the data, you get a clearer, more complete picture of the damage. The X-ray gives you the 3D shape, the sound gives you the deep layers, and the heat gives you the surface details. When you put them together, you get a "conservative" estimate that is safe and reliable, making sure that no hidden damage is left behind in these high-tech materials.

Technical Summary: NDT Data Fusion for Post-Impact Damage Characterization in Discontinuous Long Fiber Reinforced Polymer Composites

Problem Statement
The characterization of impact damage in Prepreg Platelet Molded Composites (PPMCs) is complicated by their heterogeneous mesostructure, which arises from the compression molding of randomly oriented unidirectional prepreg platelets. This stochastic architecture leads to local variations in fiber orientation distribution (FOD), resin-rich zones, and inter-platelet interfaces. Under low-velocity impact (LVI), these factors result in complex damage morphologies, including irregular projected areas, non-uniform through-thickness propagation, and tightly closed cracks.

Single Non-Destructive Testing (NDT) modalities struggle to fully capture this damage state:

Ultrasonic Testing (UT): Suffers from shadowing effects where upper-layer delaminations obscure deeper damage, and is influenced by local anisotropy and attenuation due to FOD variations.
Infrared Thermography (IRT): Limited by depth sensitivity and surface emissivity; local FOD variations alter thermal conductivity, creating artifacts, while complex subsurface geometries can distort thermal anomalies.
X-ray micro-computed tomography ( $\mu$ CT): Provides detailed 3D reconstructions but is restricted to small samples due to acquisition time and size constraints. Crucially, $\mu$ CT cannot resolve cracks that are tightly closed or lack an air gap.

Relying on any single technique risks incomplete or biased assessments. While data fusion is established in other composite inspection contexts, its application to discontinuous fiber composites with stochastic mesostructures remains underexplored.

Methodology
This study applied a pixel-level data fusion framework to post-impact PPMC specimens (Hexcel IM7/8552, 1.76 mm thick) subjected to 5 J LVI. Three modalities were employed:

Ultrasonic Testing (UT): Pulse-echo testing using a 10 MHz transducer on both front and back surfaces to mitigate shadowing. Data was processed into depth-resolved C-scans (16 slices) and binarized using Otsu's thresholding.
Transient Infrared Thermography (IRT): Dual-sided scanning (front and back) using halogen excitation. Thermal images were processed using first-derivative analysis. Segmentation was performed via manual global thresholding, selected for its stability against meso-structural fluctuations compared to Otsu or Canny methods.
Micro-Computed Tomography ( $\mu$ CT): Two scans were performed: a lower-resolution full-plate scan and a higher-resolution scan of the impact zone. The volume was virtually sliced into 16 layers (0.1 mm thickness) for layer-by-layer analysis.

Data Fusion Strategies
Two distinct fusion approaches were implemented:

Feature Integration: Separate binary damage masks from UT, IRT, and $\mu$ CT were spatially registered and merged. The final fused area was defined as the union ( $A_{fused} = A_{IRT} \cup A_{UT} \cup A_{\mu CT}$ ), intended to provide a conservative upper-bound estimate of the projected damage area.
Feature Classification: Greyscale contrast maps from UT and IRT were fused before segmentation using pixel-wise operators: averaging, absolute difference, maximum, and Hadamard product.

Key Results

Projected Damage Area: The individual modalities yielded different projected damage areas (DPZ): $\mu$ CT (5.3%), UT (7.1%), and IRT (7.5%). The feature integration approach (union of all three) resulted in a DPZ of 8.5%, interpreted as a conservative upper-bound estimate. Notably, 99.66% of the $\mu$ CT-projected damage region was contained within the combined UT and IRT mask.
Spatial Agreement: Overlap metrics (IoU and Dice) indicated partial but meaningful agreement. UT and $\mu$ CT showed the highest overlap (IoU 0.76), while IRT showed lower overlap due to its sensitivity to near-surface thermal effects.
Feature Classification Performance: Classification-based fusion (averaging, difference, etc.) produced results highly sensitive to thresholding and meso-structure noise. The "difference" operator yielded the smallest area (0.34%), while the "maximum" operator yielded 6.15%. These methods were found less stable and less informative for area estimation than the feature integration approach.
Through-Thickness Morphology: Layer-by-layer comparisons revealed that $\mu$ CT offers the most direct representation of damage shape but underestimates tightly closed cracks. UT captures broader, depth-integrated regions but overestimates damage in shadowed zones. IRT provides complementary near-surface sensitivity. Destructive micrographs confirmed that no single modality captures all features; rather, they capture different aspects of the same damage field.

Significance and Claims
The authors claim that the primary value of this multi-modal fusion framework is improved interpretability and the reduction of risk in overlooking modality-specific indications, rather than an unqualified claim of superior detection accuracy.

Key contributions include:

Conservative Estimation: The fusion framework successfully provides a conservative upper-bound estimate of the projected damage area by explicitly combining the strengths of different sensing physics.
Explicit Disagreement: The approach makes areas of agreement and disagreement between modalities explicit, aiding in the morphological interpretation of complex damage.
Role Definition: The study clarifies the complementary roles of the techniques: $\mu$ CT serves as a non-destructive morphological reference (not absolute ground truth), UT provides depth-integrated sensitivity, and IRT contributes near-surface sensitivity.
Methodological Insight: The paper demonstrates that classification-based fusion strategies, often used for homogeneous composites, may not transfer directly to platelet-based systems without refinement, whereas feature integration (mask merging) offers a more transparent basis for interpretation in stochastic materials.

The work concludes that multi-method inspection strategies are necessary for post-impact damage evaluation in discontinuous long fiber composites, providing a foundation for future correlations between fused NDT descriptors and residual mechanical performance.

Technical Summary: NDT Data Fusion for Post-Impact Damage Characterization in Discontinuous Long Fiber Reinforced Polymer Composites

More like this