Abstract
The coin-tap test is a convenient and primary method for non-destructive testing, while its manual on-site operation is tough and costly. With the help of the latest intelligent signal processing method, convolutional neural networks (CNN), we achieve an intelligent coin-tap test which exhibited superior performance in recognizing the defects. However, this success of CNNs relies on plenty of well-labeled data from the identical scenario, which could be difficult to get for many real industrial practices. This paper further develops transfer learning strategies for this issue, that is, to transfer the model trained on data of one scenario to another. In experiments, the result presents a notable improvement by using domain adaptation and pseudo label learning strategies. Hence, it becomes possible to apply the model into scenarios with none or little (less than 10%) labeled data adopting the transfer learning strategies proposed herein. In addition, we used a benchmark dataset constructed ourselves throughout this study. This benchmark dataset for the coin-tap test containing around 100,000 sound signals is published at https://github.com/PPhub-hy/torch-tapnet.
Introduction
The "coin-tap test" is a staple of Non-Destructive Testing (NDT). An inspector taps a surface and listens to the acoustic response; a "clear" ring indicates a solid structure, while a "dull" thud suggests internal defects like delamination or cavities. While effective, it is slow, subjective, and requires highly trained experts.
While modern AI can automate this by "listening" to the signals, standard Deep Learning models are data-hungry. In real-world industrial settings, it is often impossible to collect and label thousands of examples for every new material or part geometry. This paper addresses this "data shortage" by investigating how a model trained on one material (e.g., Acrylic) can be effectively adapted to another (e.g., Aluminum) using transfer learning.
Methodology
The researchers proposed a framework consisting of two main components: high-performance CNN architectures and a robust transfer learning strategy.
High-Performance CNNs
The study compared three modern architectures—Plain CNN, ResNet, and DenseNet—against a baseline model (TICNN). DenseNet, which utilizes dense connections to reuse features across layers, was found to be the most effective for acoustic signal classification.
Transfer Learning Strategy
To handle the transition between different materials (the "domain gap"), two specific techniques were employed:
- Domain Adaptation via MMD: Maximum Mean Discrepancy (MMD) was used to minimize the statistical distance between the feature distributions of the "source" material (labeled) and the "target" material (unlabeled).
- Pseudo-Label Learning: After an initial training phase, the model "guesses" labels for the unlabeled target data. The most confident guesses are then fed back into the training loop as true labels, helping the model self-correct and refine its decision boundaries.
Datasets
A massive new benchmark dataset was constructed for this study:
- Size: ~101,070 sound signals.
- Materials: Glass Fiber Reinforced Plate (GFRP), Acrylic, and Aluminum.
- Categories: 7 classes including Normal (N), large-area and bubble-like cavities (LC, BC), and various types of top/bottom disbonds.
- Availability: Hosted on GitHub.
Key Results and Insights
- Supervised Learning: DenseNet achieved 96.32% mean accuracy, significantly outperforming traditional machine learning methods like SVM (64.66%).
- Zero-Label Transfer: In scenarios with zero labeled data for the target material, the transfer learning approach (DA + PLL) boosted accuracy from 47.79% to 59.74%.
- Few-Shot Transfer: With just 10% labeled data in the target domain, the model reached 88.33% accuracy, making it viable for industrial applications with minimal calibration data.
- The "Domain Gap": Simply applying a model trained on one material to another results in poor performance because material properties fundamentally change the acoustic signature of a defect.
- The "Stubborn" Model: Pseudo-labeling makes the network more "stubborn," improving accuracy for well-learned categories but potentially reducing flexibility for tricky, small defects.
Conclusion
This work demonstrates that deep learning, combined with domain adaptation, can effectively automate the coin-tap test across different materials. While large-area defects are easily identified, small "bubble-like" defects remain a challenge due to their subtle acoustic signatures. Future work will explore adversarial training and physics-informed constraints to improve detection of these subtle anomalies.
Summary updated