Deep Machine Learning for Path Length Characterization Using Acoustic Diffraction

Abstract

Many fields now perform non-destructive testing using acoustic signals for the detection of objects or features of interest. This detection requires the decision of an experienced technician, which varies from technician to technician. This evaluation becomes even more challenging as the object decreases in size. In this paper, we assess the use of both traditional signal-processing machine learning algorithms, Long Short-Term Memory (LSTM), as well as Convolutional Neural Network (CNN) architectures to approximate acoustic anomalies with an eye toward micro-scale applications such as application to biofilms. The probing signal is generated using a continuous sound wave emitted at controlled frequencies of 1 and 5 MHz through metallic specimens of varying heights each containing an anomaly in the form of a hole. Data are collected as the transmitted signal is sampled at several locations as the wave travels through the specimen. We have developed both a CNN and an LSTM architecture for frequency-domain feature detection and approximation. The CNN models, one for phase and one for amplitude data, take short-distance Fourier transforms (SDFTs) representing the change in the signal over multiple observation points as input. The LSTM model takes the change in phase or amplitude points at each lateral location as a comma-separated value (CSV) input. The models analyze the frequency and spatial changes experienced by each specimen and produce an estimation of the acoustic path length of the anomaly in radians. The models are evaluated using mean-square error and the R-square statistic. All models perform with a fairly high R-square score, the amplitude CNN and LSTM models achieving upwards of a 99% fit and the phase CNN achieving a 97% fit on average for the predicted values. With the performance of these models, we demonstrate that utilizing the transfer function phase and amplitude data to analyze acoustic diffraction patterns leads to the ability to extract, with great precision, features in the input signal that describe the nature of the anomaly.