[2A5] Neural architecture search for a 3D CNN to classify defects from volumetric ultrasonic testing data of composites
S McKnight¹, C MacKinnon¹, E Mohseni¹, S Pierce¹, V Tunukovic¹, C MacLeod¹ and T O’Hare²
¹University of Strathclyde, UK
²Spirit AeroSystems, UK
Carbon fibre-reinforced polymers (CFRPs) are increasingly employed in both civilian and military aerospace industries due to their exceptional physical properties, such as high specific strength and corrosion resistance. However, the growing utilisation of composite components necessitates extensive non-destructive testing (NDT) inspections during the manufacturing process, with ultrasonic testing (UT) being the most commonly employed method. Automating this testing process poses significant challenges and can become a major bottleneck for large-scale manufacturing. While the deployment of phased array probes can be automated using robotic inspection techniques, it results in the generation of substantial amounts of data. Despite advancements, the interpretation of this data remains primarily reliant on skilled operators in industrial settings. This manual interpretation process is not only time-consuming but also introduces the potential for human errors. Deep learning (DL) methods offer an exciting possibility to help address the automated interpretation of NDT data interpretation. The scarcity of reliable training data in NDT poses significant challenges when it comes to training and testing DL algorithms. Despite this limitation, there has been a growing body of research exploring the use of DL for ultrasonic testing (UT) in NDT, particularly in the automated detection and characterisation of defects. These studies typically work with B- or C-scan images. The former sacrifices spatial information about the defect, while the latter retains spatial information but compresses the acoustic response data and requires manual preprocessing to remove the front and backwall responses. Unfortunately, this preprocessing step eliminates useful features, as the absence of backwall data can also lead to missed defect indications. This work presents a novel method for automated inspection of full volumetric ultrasonic data using three-dimensional convolutional neural networks (3D-CNNs). The method reduces the need for manual processing (such as gating) by detecting and classifying full ultrasonic volumetric data. The proposed research contributes a new technique for generating full volumetric synthetic UT data, allowing for training of a 3D-CNN with vastly reduced preprocessing. In addition, the use of domain-specific augmentation methods for training, which significantly increase classification performance, is introduced. A neural architecture search is performed on a ResNet-based search space that was modified to account for 3D volumetric data. The resulting model showed impressive classification results when trained on augmented synthetic data and tested on data experimentally gathered from manufactured defects. The model successfully detected all back-drilled hole defects.
