Noninvasive acoustic time-of-flight measurements in heated, hermetically-sealed high explosives using a convolutional neural network

• Published in: Machine learning with applications 2022-09, Vol.9, p.100391, Article 100391
• Main Authors: Greenhall, John, Zerkle, David, Davis, Eric Sean, Broilo, Robert, Pantea, Cristian
• Format: Article
• Language: English
• Subjects: Acoustic measurement in high explosives; Acoustic time-of-flight; Convolutional Neural Networks; Nondestructive evaluation; OTHER INSTRUMENTATION; Wave interference
• ISSN: 2666-8270; 2666-8270
• DOI: 10.1016/j.mlwa.2022.100391

We present a data-driven technique for measuring the time-of-flight through material sealed within a container. Time-of-flight measurement provides a noninvasive means of quantifying the sound speed profile within a material by transmitting an acoustic burst and then measuring the time required for the burst to arrive at an opposing receiver. In a hermetically-sealed cylindrical container, a portion of the acoustic energy propagates through the material as a bulk wave, while the remainder of the acoustic energy propagates around the container walls as guided waves. As a result, interference from the guided waves obscures the bulk arrival, inhibiting measurement of the sound speed. The technique uses a Convolutional Neural Network (CNN) to identify critical features in the measured waveforms and identify bulk wave arrivals. We demonstrate this time-of-flight measurement technique on high explosive-filled containers as they are heated from room temperature to detonation. This is a particularly challenging application for acoustic time-of-flight measurements as the high explosives have significant sound speed gradients as they undergo heating, and they lead to significant attenuation of the bulk wave, as opposed to the guided waves, which do not suffer significant attenuation. We characterize the performance of the CNN as a function of the high explosive temperature and as a function of the CNN hyperparameters. We then provide physical insight into the error trends. •Noninvasive inspection of sealed high explosive containers.•Interference between waves in sample and in container walls inhibits measurements.•Convolutional Neural Networks for distinguishing between interfering waves.