3-Dimensional residual neural architecture search for ultrasonic defect detection

TL;DR

This paper introduces a 3D residual neural architecture search method for ultrasonic defect detection in composites, leveraging synthetic volumetric data and domain-specific augmentation to improve accuracy and efficiency.

Contribution

It presents a novel 3D neural architecture search approach tailored for ultrasonic defect detection, incorporating synthetic data and augmentation techniques for enhanced performance.

Findings

Best model achieved 100% accuracy.

Neural architecture search outperformed hand-designed models.

Fully convolutional layers outperformed max pooling.

Abstract

This study presents a deep learning methodology using 3-dimensional (3D) convolutional neural networks to detect defects in carbon fiber reinforced polymer composites through volumetric ultrasonic testing data. Acquiring large amounts of ultrasonic training data experimentally is expensive and time-consuming. To address this issue, a synthetic data generation method was extended to incorporate volumetric data. By preserving the complete volumetric data, complex preprocessing is reduced, and the model can utilize spatial and temporal information that is lost during imaging. This enables the model to utilise important features that might be overlooked otherwise. The performance of three architectures were compared. The first two architectures were hand-designed to address the high aspect ratios between the spatial and temporal dimensions. The first architecture reduced dimensionality in the time domain and used cubed kernels for feature extraction. The second architecture used cuboidal kernels to account for the large aspect ratios. The evaluation included comparing the use of max pooling and convolutional layers for dimensionality reduction, with the fully convolutional layers consistently outperforming the models using max pooling. The third architecture was generated through neural architecture search from a modified 3D Residual Neural Network (ResNet) search space. Additionally, domain-specific augmentation methods were incorporated during training, resulting in significant improvements in model performance for all architectures. The mean accuracy improvements ranged from 8.2% to 22.4%. The best performing models achieved mean accuracies of 91.8%, 92.2%, and 100% for the reduction, constant, and discovered architectures, respectively. Whilst maintaining a model size smaller than most 2-dimensional (2D) ResNets.