Deep Learning for Micro-Scale Crack Detection on Imbalanced Datasets Using Key Point Localization

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

Introduction: Addressing the "Needle in a Haystack" Problem

In structural health monitoring, internal micro-scale cracks represent a significant diagnostic challenge. Unlike surface-level defects, these cracks are invisible to traditional visual inspection methods, necessitating the use of seismic wave propagation analysis. However, applying deep learning to this domain introduces a severe class imbalance problem. In a typical material sample, the area occupied by a micro-crack is negligible compared to the healthy volume. Standard semantic segmentation models, which classify every pixel, often fail by converging to a "lazy" state where they predict no cracks at all, achieving high nominal accuracy while failing their primary objective.

Methodology: From Segmentation to Key Point Localization

To overcome the limitations of pixel-wise classification, this research introduces MicroCrackPointNet. The core innovation lies in reframing crack detection as a regression task rather than a segmentation problem.

Overview of the MicroCrackPointNet architecture, mapping seismic wave inputs to crack localization outputs.

The architecture incorporates several advanced deep learning techniques:

  1. Key Point Localization: Instead of labeling every pixel, the model predicts the $(x, y)$ coordinates of four key points defining a bounding box around the crack. This approach bypasses the class imbalance issue by focusing solely on the geometry of the defect.
  2. Wide Convolutional Blocks: Inspired by Inception modules, the network utilizes parallel convolutional branches with varying kernel sizes (1x1, 3x3, and 5x5). This allows the model to capture multi-scale features of the seismic signal simultaneously.
  3. Attention Mechanisms: Self-attention layers are integrated to help the network prioritize signal disruptions caused by cracks over background noise.
  4. Robust Optimization: The model is trained using Huber loss, which provides a balance between Mean Absolute Error (robustness to outliers) and Mean Squared Error (smoothness near zero), ensuring stable convergence.
Detailed structure of the parallel wide convolutional block inspired by Inception modules.

Data Acquisition: Synthetic Seismic Wave Simulation

Given the difficulty of obtaining labeled real-world data for internal micro-cracks, the study utilizes the dynamic Lattice Element Method (dLEM) to simulate seismic wave propagation. The input data consists of a 9x9 sensor grid recording wave displacement over 2000 time steps. Ground truth labels were generated by converting segmentation masks into normalized coordinate sets representing the crack boundaries.

Results: Accuracy and Computational Efficiency

The proposed model demonstrates significant improvements over traditional benchmarks in both localization precision and resource consumption:

  • Localization Performance: The model achieved an average Intersection over Union (IoU) of 0.511. For larger micro-cracks (>4 µm), the performance increased to an IoU of 0.631. The Mean Absolute Error (MAE) for coordinate prediction was 0.0731, indicating high spatial precision.
  • Efficiency: MicroCrackPointNet is remarkably compact, utilizing only 90 layers compared to the 444 layers in a standard 1D-DenseNet. This reduction in complexity led to a dramatic decrease in training time, from over 4.3 hours for the DenseNet benchmark to just 36 minutes for the proposed architecture.

Conclusion: Limitations and Future Work

While the key point localization approach successfully mitigates class imbalance, several areas for improvement remain. The current iteration is limited to detecting a single, rectangularly-bounded crack per sample. Real-world structural failures often involve multiple, irregularly shaped cracks. Future research will explore more flexible architectures, such as YOLO or Faster R-CNN, to accommodate multiple defects and more complex geometries, alongside testing on higher-resolution sensor arrays.

Summary updated

Abstract

Internal crack detection has been a subject of focus in structural health monitoring. By focusing on crack detection in structural datasets, it is demonstrated that deep learning (DL) methods can effectively analyse seismic wave fields interacting with micro-scale cracks, which are beyond the resolution of conventional visual inspection. This work explores a novel application of DL based key point detection technique, where cracks are localized by predicting the coordinates of four key points that define a bounding region of the crack. The study not only opens new research directions for non-visual applications but also effectively mitigates the impact of imbalanced data which poses a challenge for previous DL models, as it can be biased toward predicting the majority class (non-crack regions). Popular DL techniques, such as the Inception blocks are used and investigated. The model shows an overall reduction in loss when applied to micro-scale crack detection and is reflected in the lower average deviation between the location of actual and predicted cracks, with an average IOU being 0.511 for all micro cracks (> 0.00 µm) and 0.631 for larger micro cracks (> 4 µm).
BibTeX
@misc{moreh2024deep,
  author = {Fatahlla Moreh and Yusuf Hasan and Bilal Zahid Hussain and Mohammad Ammar and Sven Tomforde},
  title = {{Deep Learning for Micro-Scale Crack Detection on Imbalanced Datasets Using Key Point Localization}},
  year = {2024},
  eprint = {2411.10389},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2411.10389},
}
Figure