Machine Learning-Enabled Liquid Flow Front Estimation from Ultrasonic Guided Waves Signals

Abstract

Accurate monitoring of the liquid front during infusion remains a significant challenge in out-of-autoclave composite manufacturing, directly affecting process control, part quality, and manufacturing yield. This study investigates the use of ultrasonic guided waves for front localisation in a liquid-only infusion mould designed to produce traceable, and reproducible front geometry across subsequent trials, enabling consistent labelling and development of experimentally trained prediction models. A one-dimensional convolutional neural network (CNN) trained on time-domain amplitude scans achieves a mean absolute error of 7.71 ± 2.11 mm across five independently acquired experimental sets, outperforming the baseline energy-based functional approximation. The baseline method estimates position from the normalised waveform energy, capturing only the amplitude attenuation effects as the flow advances. In contrast, the CNN uses the unprocessed waveform, therefore exploiting richer physical information. Since building sufficiently large experimental training sets is costly and time-consuming, the paper also explores simulation-based data generation. A refined finite element simulation is developed to reproduce the fundamental modal behaviour of the experimental amplitude scans at substantially reduced computational costs, supporting the efficient generation of large synthetic training sets. To bridge the simulated and experimental domains, a conditional Generative Adversarial Network (GAN) is developed. When GAN synthetic data is combined with a limited experimental dataset, prediction error drops by 46% compared with training on the reduced experimental set alone. The results show that front estimation using machine learning, supported by GAN synthetic data, provides a promising and data-efficient route for monitoring resin flow in liquid composite moulding processes.

Keywords: Resin Infusion Process Monitoring, Ultrasonic Leaky Lamb Waves, Machine learning, Synthetic Data Augmentation, Finite Element Analysis, Generative Adversarial Network