A synthetic dataset is sufficient to measure the height of thermally fatigued surface-breaking cracks using convolutional neural networks

Non-destructive evaluation (NDE) is used to detect and characterise defects in safety-criticalcomponents. This paper focuses on pulse-echo ultrasonic testing applied to the sizing of surfacebreaking crack-like defects. In the field, practitioners often useA-scans from single crystal, pulse-echo ultrasonic testing. Machine learning, and in particular neural networks, have huge potential for applications to NDE-type problems due to their ability to recognise patterns from signals. Accurate predictions require large, labelled databases yet there is a paucity of such ultrasonic data for surface-breaking thermally fatigued cracks. To address this deficit, this work demonstrates that a database composed entirely of simulated, syntheticA-scans provides promising predictions on measured data. Crack height predictions with a mean absolute error of around 0.124 mm are achieved for measured data (on cracks between 0.5 mm and 4 mm). When tested on simulated A-scans, both height and tilt angle predictions are extremely accurate (to within 0.035 mm and 0.41◦, respectively). To achieve such accuracy the in-situ inspection techniques were mimicked, with particular attention paid to the characteristics of the thermally fatigued crack species, and the single crystal transducer and its input signal. The details of the finite element simulations that generated the database are presented here. The work outlined in this paper shows that A-scans can be used to size defects using neural networks trained entirely on synthetic data, informed by experimental measurements, with potential applications for improving the efficiency of sizing cracks in-situ.