Deep Learning-Based Approach for Automatic Defect Detection in Complex Structures Using PAUT Data.

Area of Science:

• Materials Science
• Non-Destructive Testing
• Artificial Intelligence

Background:

• Automated defect detection in complex structures is crucial for ensuring material integrity.
• Traditional signal processing methods for phased array ultrasonic testing (PAUT) face limitations in noisy environments.
• Deep learning offers potential for enhanced accuracy in analyzing PAUT data.

Purpose of the Study:

• To develop and compare automated defect detection methods using PAUT data.
• To evaluate traditional signal processing and various deep learning architectures.
• To introduce and validate a novel Convolutional Attention Temporal Transformer for Sequences (CATT-S) model.

Main Methods:

• Implemented an improved signal-to-noise ratio algorithm with automatic depth gate calculation as a baseline.
• Developed and trained fully connected networks (FCN), convolutional neural networks (CNN), and the CATT-S model.
• Utilized diverse datasets including simulated CIVA data and real-world data from welded and composite specimens.

Main Results:

• The improved SNR algorithm showed robust flaw indication but struggled with noisy data.
• CNN achieved 94.9% test accuracy, effectively capturing local PAUT signal features.
• The CATT-S model outperformed baselines with 99.4% accuracy and a 0.905 F1-score, modeling both morphology and inter-beam dependencies.

Conclusions:

• Deep learning, particularly the CATT-S model, significantly improves automated defect detection in complex structures using PAUT.
• The CATT-S model's ability to capture fine-grained signal morphology and long-range dependencies is key to its superior performance.
• This integrated approach holds substantial practical potential for reliable and efficient non-destructive testing (NDT) of heterogeneous materials.