Simulation trained CNN for accurate embedded crack length, location, and orientation prediction from ultrasound measurements

Abstract

Accurate quantitative characterization of crack length, location, and orientation are critical for the safety assessment of load bearing structures to avoid catastrophic structural failures. Ultrasound non-destructive evaluation is one of the key methods to detect and evaluate embedded flaws inside a material during fabrication or operation. Although significant progress has been made in developing advanced ultrasound sensors and signal data processing methods, current practices rely on human expertise to evaluate the ultrasound measurements, which leads to high uncertainty and errors in the predictions. Here we demonstrate that an embedded crack reflected ultrasound time signal contains complete information about the key characteristics of the crack, which can be accurately quantified using an optimally trained machine learning model. A lack of sufficiently large, well distributed, and suitably labeled datasets to train machine learning models continues to be a significant obstacle for evaluating non-visible cracks. To overcome this limitation, we demonstrate that our finite element simulation trained convolutional neural network (CNN) is able to accurately predict all three crack characteristics from experimentally measured ultrasound non-destructive test signals. We created a moderate size A-scan time signal simulation dataset (1200 scans) for three-dimensional (3D) elliptical penny-shaped cracks inside rectangular cuboid steel to train our CNN. Independent validation experiments were performed by conducting 21 ultrasound tests on 3D printed steel specimens containing a variety of embedded crack geometries. We show that our purely finite element simulation trained CNN accurately predicts crack length, crack location, and crack orientation from experimentally measured signals with an average error of 5.7%, 5.6%, and 8.4% for length, location, and orientation, respectively. This approach of utilizing simulation-based training of a neural network can be used in other applications where experimental data are not readily available.