TL;DR
Detecting internal microcracks using seismic or ultrasonic wave propagation is difficult because high-dimensional temporal signals rarely map intuitively to tiny spatial flaws. This paper introduces MicroCrackAttentionNeXt, an asymmetric encoder–decoder architecture designed to process spatio-temporal wavefield sensor data into spatial crack segmentation masks. Supported by Manifold Discovery and Analysis (MDA) to visually inspect latent representations, the optimized model achieves an overall detection accuracy of 87.74% and a Dice Similarity Coefficient of 0.91.
What Problem Does This Paper Address?
In aerospace, materials science, and civil infrastructure, hidden microscale cracks can lead to catastrophic structural failure. Standard visual inspection cannot see inside opaque materials or complex geometries. While wave-based testing (sending stress waves through a medium and recording displacements) can detect internal flaws, deciphering the resulting numerical time series is notoriously difficult:
- Spatio-temporal complexity: Wavefields generate long, multi-channel numerical sequences ($2000$ time steps) across spatial sensor grids.
- Extreme class imbalance: Cracks occupy only about 5% of the material area on average. Small cracks ($<1\,\mu\text{m}$) cause minimal disruption to propagating waves, leading deep neural networks to favor the majority (healthy) class.
- Opaque latent spaces: Standard dimensionality reduction tools like t-SNE or UMAP struggle to provide useful diagnostic views for dense segmentation tasks.
Methods
The authors design MicroCrackAttentionNeXt, an asymmetric deep neural network tailored for temporal-to-spatial feature extraction:
- Temporal Encoder: To bypass the high computation of 3D/spatio-temporal convolutions, the network applies temporal max pooling (reducing $2000 \to 500$ steps) followed by four 1D convolutional blocks (using $(3,1)$ kernels). These blocks extract temporal patterns across channels without mixing spatial sensor coordinates prematurely.
- Attention & Squeeze-and-Excitation (SE): SE modules recalibrate channel-wise dependencies, while self-attention layers capture long-range temporal-spatial interactions, helping the model focus on the exact moments waves interact with cracks.
- Adaptive Feature Reutilization: Intermediate feature maps are upsampled and fed back into early convolutional stages via a feedback mechanism inspired by Self-Attention Modules (SAM), improving feature reuse.
- Full-Horizon Bottleneck & Spatial Decoder: A large temporal kernel ($31, 1$) collapses the remaining temporal dimension into a compact code. The decoder reshapes this code into a 2D spatial grid ($9 \times 9$) and progressively upsamples it using transposed convolutions to output a $36 \times 36$ binary segmentation mask.
- Manifold Discovery and Analysis (MDA): The authors apply MDA to project intermediate high-dimensional layer activations onto 2D manifold embeddings, visually evaluating feature continuity and class separation across various activations and loss formulations.
Datasets
- Lattice Element Method (LEM) Simulation Dataset (Wuttke et al., 2021): Numerical simulations of homogeneous 2D plates modeled with Voronoi–Delaunay lattices.
- An external dynamic force ($1000\text{ N}$) generates stress waves.
- A uniform $9 \times 9$ sensor grid records orthogonal displacements ($x, y$) over $2000$ time steps (input shape: $2 \times 81 \times 2000$).
- Target labels are $36 \times 36$ binary segmentation maps covering microcrack lengths ranging from $<1\,\mu\text{m}$ to $>4\,\mu\text{m}$.
Key Insights and Ideas
- Decoupling Time and Space: Processing the temporal dynamics first with 1D convolutions before reconstructing spatial representations dramatically reduces parameter overhead while outperforming joint spatio-temporal baselines.
- Why SeLU Underperforms: While Scaled Exponential Linear Units (SeLU) offer self-normalizing properties, the MDA feature manifolds revealed they compress activation variance so strongly that they "dampen" the subtle signatures of rare microcracks.
- The Superiority of GeLU + CWDL: Gaussian Error Linear Units (GeLU) provide smooth probabilistic gating that pairs best with Combined Weighted Dice Loss (CWDL). MDA visualizations confirmed that this combination produced tight, smooth, unbroken manifold trajectories with clear separation between cracked and intact regions.
Key Results
- Overall Accuracy: MicroCrackAttentionNeXt trained with GeLU and CWDL achieved 87.74% accuracy across all crack sizes ($>0\,\mu\text{m}$), climbing to 98.48% on cracks larger than $4\,\mu\text{m}$.
- Comparison to Prior Work: Outperforms the previous 1D-DenseNet benchmark (Moreh et al., 2024; 83.68% accuracy) while reaching a Dice Similarity Coefficient (DSC) of 0.91.
- Latent Coherence: MDA plots confirmed that unlike the baseline 1D-DenseNet—which exhibited broken, discontinuous clusters for crack-associated feature maps—MicroCrackAttentionNeXt maintained smooth, continuous manifold representations.
Limitations and Future Work
- Purely Synthetic Data: The model was tested entirely on idealized numerical LEM simulations; testing on noisy, real-world physical specimens remains an ongoing effort.
- Low Spatial Output Resolution: Predictions are limited to a coarse $36 \times 36$ grid. The authors plan to integrate super-resolution GANs (SR-GANs) to sharpen crack boundaries.
- State-Space Exploration: To better handle long-horizon wave sequences without information loss, future iterations aim to incorporate modern state-space architectures like Mamba into the encoder.
Summary updated