Three-Dimensional Residual Neural Architecture Search for Ultrasonic Defect Detection.

Area of Science:

• Materials Science
• Artificial Intelligence
• Non-Destructive Testing

Background:

• Experimental ultrasonic testing data acquisition for carbon fiber composites is costly and time-consuming.
• Existing deep learning methods often struggle with the complexities of volumetric ultrasonic data.

Purpose of the Study:

• To develop and evaluate a deep learning methodology for defect detection in carbon fiber-reinforced polymer (CFRP) composites using volumetric ultrasonic testing (UT) data.
• To address the challenges of limited experimental training data through synthetic data generation and advanced neural network architectures.

Main Methods:

• Utilized 3D convolutional neural networks (CNNs) for analyzing volumetric UT data.
• Extended synthetic data generation to incorporate complete volumetric information, reducing preprocessing.
• Compared three CNN architectures: a standard volumetric classifier, a modified architecture with cuboidal kernels, and a novel architecture discovered via neural architecture search (NAS).
• Incorporated domain-specific data augmentation techniques during training.

Main Results:

• The discovered NAS architecture achieved the highest performance, with a mean accuracy improvement of 22.4% after augmentation.
• This architecture outperformed the second-best model by 7.9% in mean accuracy.
• The model consistently detected all defects, maintained a smaller size than typical 2D ResNets, and had an inference time under 0.5 seconds.

Conclusions:

• Deep learning, particularly NAS-discovered 3D CNNs, offers a highly effective and efficient solution for automated defect detection in CFRP composites using volumetric UT data.
• Synthetic data generation and domain-specific augmentation are crucial for improving the performance and robustness of these models.
• The developed methodology significantly enhances the NDT process for composite materials.