MicroCrackAttentionNeXt: Advancing Microcrack Detection in Wave Field Analysis Using Deep Neural Networks through Feature Visualization

Fatahlla Moreh, Yusuf Hasan, Bilal Zahid Hussain, Mohammad Ammar, Sven Tomforde

Abstract

Structural health monitoring relies heavily on the early detection of microcracks to prevent catastrophic failure. Wave-field analysis offers a non-destructive path to detection, yet it is hampered by the high dimensionality of spatio-temporal data and severe class imbalance. We present MicroCrackAttentionNeXt, a deep learning architecture that utilizes hybrid feature extraction and attention mechanisms to segment microcracks with high precision. By leveraging Manifold Discovery and Analysis (MDA), we visualize the network's internal logic, demonstrating how specific architectural choices like GeLU activation functions enhance the detection of rare crack patterns. Our model achieves an 86.85% detection accuracy and a Dice Similarity Coefficient of 0.86, significantly outperforming existing benchmarks.

Introduction

Material fatigue and the resulting microcracks pose significant risks to critical infrastructure, such as aerospace components and engine parts. Traditional inspection methods often fail to identify these microscopic fissures, especially within complex geometries. Wave-field analysis, which involves propagating waves through a material and analyzing the resulting disruptions, has emerged as a powerful alternative. However, the data generated is inherently complex—spatio-temporal in nature—and characterized by extreme class imbalance, where crack-related signals may constitute less than 5% of the total dataset. Standard neural networks often fail in such scenarios, defaulting to majority-class predictions to achieve deceptively high accuracy scores.

Methodology

MicroCrackAttentionNeXt is an asymmetric encoder-decoder network specifically engineered for this challenge. The architecture employs a multi-stage approach to feature extraction. Initially, 1D-CNN layers are used to isolate temporal features from the wave propagation data. This is followed by spatial analysis using Squeeze-and-Excitation (SE) modules to recalibrate channel-wise features and attention layers to help the model focus on critical time steps where wave-crack interaction is strongest.

The architecture of MicroCrackAttentionNeXt, featuring an encoder-decoder structure with integrated attention and SE modules.

To diagnose and optimize the model, we utilized Manifold Discovery and Analysis (MDA). Unlike traditional visualization tools like t-SNE, MDA preserves the geodesic distances of high-dimensional feature maps, allowing for a more accurate representation of the underlying data structure. This diagnostic tool was instrumental in evaluating the impact of various activation functions on the model's ability to distinguish cracks from intact material.

Experimental Results and Discussion

The model was evaluated using a simulated wave-field dataset consisting of 2,000 time steps recorded across an 81-sensor grid. The results indicate that the combination of Gaussian Error Linear Unit (GeLU) activation and Combined Weighted Dice Loss (CWDL) yielded the highest performance. GeLU's smooth, probabilistic gating mechanism proved superior to ReLU and SELU in managing the information flow for minority classes.

MDA visualization of feature manifolds, illustrating the separation between crack and non-crack classes across different model configurations.

The optimized MicroCrackAttentionNeXt achieved an overall accuracy of 86.85% and a Dice Similarity Coefficient (DSC) of 0.86. Notably, MDA plots revealed that while SELU (Scaled Exponential Linear Unit) tended to dampen the signals of rare patterns (the cracks), GeLU allowed the model to maintain focus on these critical features. The visualization showed a clear, arch-like manifold transition from "no crack" to "crack" in well-trained models, providing a qualitative "proof of learning" that accuracy metrics alone cannot convey.

Conclusion and Future Work

MicroCrackAttentionNeXt demonstrates that combining specialized architectural components with advanced visualization techniques can significantly improve microcrack detection in imbalanced datasets. While the current model provides high-accuracy detection, the resolution of the segmentation output is limited by the sparse sensor grid. Future research will explore the use of Generative Adversarial Networks (GANs) for super-resolution to upscale low-resolution masks and the integration of Mamba architectures to better capture long-range temporal dependencies in wave propagation.

Summary updated

Abstract

Micro Crack detection using deep neural networks(DNNs) through an automated pipeline using wave fields interacting with the damaged areas is highly sought after. However, these high dimensional spatio-temporal crack data are limited, moreover these dataset have large dimension in the temporal domain. The dataset exhibits a pronounced class imbalance, with crack pixels accounting for an average of only 5% of the total pixels per sample. This severe imbalance presents a challenge for deep learning models when dealing with various microscale cracks, as the network tends to favor the majority class, often resulting in reduced detection accuracy. This study proposes an asymmetric encoder–decoder network with Adaptive Feature Reutilization Block for micro-crack detection. The impact of various activation and loss functions were examined through feature space visualisation using manifold discovery and analysis (MDA) algorithm. The optimized architecture and training methodology achieved an accuracy of 87.74%.
BibTeX
@misc{moreh2024microcrackattentionnext,
  author = {Fatahlla Moreh and Yusuf Hasan and Bilal Zahid Hussain and Mohammad Ammar and Sven Tomforde},
  title = {{MicroCrackAttentionNeXt: Advancing Microcrack Detection in Wave Field Analysis Using Deep Neural Networks through Feature Visualization}},
  year = {2024},
  eprint = {2411.10015},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2411.10015},
}
Figure