Accelerating Analysis of Ultrasonic Test Data in Carbon Fibre Reinforced Polymers with Vision Foundation Models

The aerospace industry is increasingly reliant on Carbon Fibre Reinforced Polymer (CFRP) components. The safety critical aspect of these components and complex manufacturing process entail a necessity for thorough non-destructive evaluation (NDT) inspection. Due to manipulator driven NDT delivery, the bottleneck of NDT has shifted to data interpretation. Despite the promise of artificial intelligence methods, widespread reliability concerns have stalled automation of inspection.

This work proposes vision foundation models for automated ultrasonic NDT data analysis. The approach leverages the generalisability of transformer architectures in absence of high volumes of data. Two vision foundation model encoders are fused with the weights kept frozen. To accommodate two ultrasonic data modes - B-scan and C-scan - the architecture employs a shared encoder and specialised decoder heads per domain. The model is fine-tuned through adapter blocks in a supervised manner on limited data, to perform pixel-wise segmentation of features such as sub-surface defects and structural echoes. This approach enables efficient detection, and characterisation of defects. An inherent advantage of this approach is cross-modal consistency; quantified via intersection between projected B-scan and C-scan predictions, providing an internal validation mechanism that improves confidence of detection.

Data in this work is acquired with a KUKA KR90 manipulator delivering phased array ultrasonic roller probe inspection of industrial grade CFRP components. Training dataset consists of inspection data from samples containing manufactured flat-bottom holes of varying sizes and a total of 2848 B-scans. Initial inference on a sample of 1348 B-scans with 25 defects shows all defects successfully detected and a mean dice score of 80% against 6dB threshold ground truth labels. These results imply excellent potential for detection and characterisation of defects. Further work will focus on comparative studies and generalisable defect segmentation on complex and challenging component geometries, more representative of industry settings.