⚡ electrical engineering

Crack Detection Methods of Metal Plates Based on Time-Series Acoustic Images and Deep Learning Algorithm

This paper proposes a crack detection method for metal plates that converts multi-channel acoustic signals into time-series acoustic images and utilizes a lightweight transformer model to achieve highly accurate and efficient detection of cracks with varying positions and lengths.

Original authors: Xiaozheng Zhang, Jia Li, Lu Zhu, Shuai Wang

Published 2026-09-09

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Xiaozheng Zhang, Jia Li, Lu Zhu, Shuai Wang

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

Metal plates are the silent workhorses of modern engineering, forming the skins of aircraft, the frames of bridges, and the hulls of ships. Over time, these structures endure stress, vibration, and impact, which can cause tiny fractures to appear. If left unnoticed, these cracks can grow, leading to catastrophic failures. For decades, engineers have relied on methods that require touching the metal or using expensive, bulky equipment to find these flaws. However, a new approach suggests that the metal itself might be able to tell us where it is broken, simply by listening to the sound it makes. When a metal plate vibrates, it radiates sound waves into the air. The pattern of this sound changes depending on the health of the metal; a crack disrupts the flow of energy, creating a unique acoustic signature. By capturing these sound patterns and analyzing them with advanced computer programs, it is possible to locate damage without ever touching the surface.

A team of researchers at Hefei University of Technology has developed a method to do exactly this, turning sound into a visual map that a computer can read. Instead of listening to a single microphone, they used an array of twenty-four microphones arranged in a grid to capture sound from a vibrating metal plate. This setup allowed them to record not just the loudness of the sound, but how the sound waves moved across the surface over time. The researchers took these recordings and transformed them into a sequence of images, creating what they call "time-series acoustic images." Imagine a movie where each frame shows the sound pressure on the plate at a specific moment, with colors representing the intensity of the sound. By stacking twenty of these frames together, they created a three-dimensional data set that captures both the spatial layout of the sound and how it evolves over time. This visual representation makes it possible to see exactly where the sound field is distorted by a crack.

To test this idea, the team first ran detailed computer simulations of metal plates with cracks of different lengths and at different positions. They generated thousands of these time-series acoustic images, creating a dataset that included plates with no cracks, plates with small cracks, and plates with larger cracks. They then fed these images into three different types of artificial intelligence models to see which one could best identify the damage. The first model was a standard deep learning network designed to process video-like data. The second was a more complex model known as a Transformer, which is famous for its ability to understand relationships between different parts of a sequence. The third was a streamlined version of the Transformer, designed to be much smaller and faster while keeping the same intelligence.

The results showed that while the standard network could find the cracks most of the time, it struggled with accuracy, correctly identifying the damage in only 95 percent of the test cases. The larger Transformer model, however, achieved perfect accuracy, correctly identifying every single crack in the simulation. The most significant finding, though, came from the lightweight version of the Transformer. Despite having far fewer internal components and requiring much less computing power, this smaller model also achieved 100 percent accuracy. It proved that the system did not need a massive, resource-heavy computer to work; a simpler, more efficient design was sufficient to spot the subtle distortions caused by a crack. The researchers confirmed this by visualizing how the models "saw" the data, showing that the lightweight model could clearly separate the sound patterns of a cracked plate from those of a healthy one.

To ensure this method worked in the real world, not just in a computer, the team built a physical experiment. They used a steel plate measuring 450 millimeters by 450 millimeters and created actual cracks in it using a laser cutter. They placed the plate on a frame and vibrated it, using their microphone array to capture the sound. Just as in the simulations, they converted the sound data into time-series acoustic images and fed them into their lightweight model. The system successfully identified cracks of varying lengths and at different locations on the plate with 100 percent accuracy. The model was able to distinguish between a crack near the edge and one in the middle, as well as tell the difference between a 10-millimeter crack and a 20-millimeter crack. This demonstrated that the method is robust enough to handle real-world variations in sound and structure.

The study concludes that listening to a metal plate and turning that sound into a moving picture is a highly effective way to find cracks. The researchers found that the acoustic field changes in a predictable way when a crack is present, and these changes are distinct enough for a computer to recognize. While the current tests were conducted in a controlled, quiet environment, the success of the lightweight model suggests that the technology could eventually be adapted for use in noisy industrial settings. The ability to detect damage with such precision using a simple, non-contact method offers a promising path toward safer and more efficient maintenance of critical metal structures. The work confirms that the sound a structure makes is a reliable indicator of its health, and with the right tools, we can learn to read that language with perfect clarity.

Technical Summary: Crack Detection Methods of Metal Plates Based on Time-Series Acoustic Images and Deep Learning Algorithm

Problem Statement
Metal plates are critical structural components in engineering, yet they are susceptible to cracking due to processing errors, external impacts, high-intensity vibration, and fatigue. Undetected cracks can lead to equipment failure or catastrophic structural damage. While existing detection methods (sensor-based, infrared, ultrasonic, radiographic) exist, they often suffer from high dependence on equipment and operators, low efficiency, high costs, and limitations regarding contact measurements. Furthermore, traditional acoustic signal analysis using single microphones struggles to accurately detect cracks of varying positions and lengths. This study addresses the need for a non-contact, efficient, and accurate method to detect cracks in metal plates by leveraging the temporal and spatial characteristics of radiated sound fields.

Methodology
The proposed methodology integrates acoustic signal acquisition, image generation, and deep learning classification through the following steps:

Data Acquisition and Image Generation:
- Excitation and Measurement: Metal plates are excited by forces of varying frequencies. A microphone array (24 microphones in experiments, simulated in numerical studies) captures multi-channel sound pressure signals.
- Time-Series Acoustic Images: The multi-channel signals are converted into three-dimensional time-series acoustic images. These images represent sound pressure amplitude maps on a measurement plane at sequential time instants (20 frames per sample).
- Visualization: Amplitude values are mapped to the RGB color space (0–255) to normalize data and enhance texture contrast, resulting in samples of size $3 \times 20 \times 200 \times 200$ (channels $\times$ time $\times$ height $\times$ width).
- Simulation vs. Experiment: Numerical simulations using the acoustic finite element and boundary element methods generated the initial dataset. Physical experiments were conducted using Q235 steel plates with laser-cut cracks of varying lengths (10, 15, 20 mm) and positions, measured in an anechoic chamber.
Deep Learning Model Construction:
Three neural network architectures were constructed and compared to classify crack conditions (crack-free vs. cracks at different positions/lengths):
- 3D CNN: A standard 3D Convolutional Neural Network designed to capture spatiotemporal correlations in the acoustic image sequence.
- Transformer: A baseline model utilizing 3D convolutional feature extraction followed by a standard Transformer encoder (2 layers, 4 attention heads) to enhance feature representation.
- Lightweight Transformer: A streamlined architecture designed to reduce computational load. It features fewer convolutional blocks (4 vs. 5), reduced channel dimensions (max 256 vs. 512), a single Transformer layer with 2 attention heads, and a reduction in token count (32 vs. 196).
Training and Validation:
- Data Augmentation: To prevent overfitting and improve robustness, the training set underwent random cropping, spatial flipping, slight rotation, and the injection of Gaussian white noise with varying Signal-to-Noise Ratios (SNRs).
- Evaluation: Models were trained on a dataset split into training, validation, and test sets. Performance was evaluated using accuracy, loss curves, T-SNE feature visualization, and metrics including Precision, Recall, and F1-score.

Key Results

Model Performance:
- The 3D CNN achieved a detection accuracy of 95.0% but exhibited signs of overfitting (high training accuracy vs. lower validation accuracy).
- The Transformer model achieved 100.0% accuracy with robust generalization but required significant computational resources (28.8M parameters, 110MB memory).
- The Lightweight Transformer also achieved 100.0% accuracy. Crucially, it reduced the parameter count to 3.5M and memory usage to 13.5MB, significantly improving computational efficiency while maintaining perfect classification performance.
Feature Separation: T-SNE visualizations confirmed that the Lightweight Transformer effectively separated features for crack-free plates and plates with cracks at different positions and lengths, showing tight clustering and clear separation between classes.
Experimental Verification: Physical experiments on steel plates with varying crack lengths and positions confirmed the simulation results. The Lightweight Transformer successfully detected all crack conditions with 100.0% accuracy on the experimental test set.

Significance and Claims
The paper claims that the proposed method offers a significant advancement in non-destructive testing by:

Spatiotemporal Representation: Demonstrating that converting acoustic signals into time-series acoustic images effectively captures the evolution of sound fields, revealing physical mechanisms where structural cracks cause scattering and diffraction that manifest as specific distortions in the acoustic image.
Efficiency and Accuracy: Establishing that the Lightweight Transformer model achieves a dual breakthrough by maintaining 100% detection accuracy while drastically reducing model size and computational power consumption compared to standard Transformers and 3D CNNs. This makes the method suitable for real-time inference in edge computing environments.
Non-Contact Detection: Validating a method that eliminates the need for contact measurement, thereby offering a safer, more convenient, and cost-effective solution for monitoring metal plate health.

Limitations and Future Work
The authors acknowledge that the current study focuses primarily on open cracks and was conducted in an ideal anechoic environment. The generalization of the method to complex crack types (e.g., micro-fatigue cracks) and its robustness in industrial environments with complex background noise require further investigation. Future research will focus on optimizing anti-noise algorithms and expanding the model to handle diverse crack morphologies.

More like this