Deep Machine Learning for Acoustic Path Length Characterization of Materials Using Acoustic Diffraction

Acoustic non-destructive testing is used in many fields for detection of objects or features of interest. This detection typically requires the decision of an experienced technician, and often the evaluation varies depending on the technician’s experience. This evaluation becomes even more challenging as the object decreases in size. In this study, we propose a Convolutional Neural Network (CNN) to categorize and approximate acoustic anomalies with an eye toward future application to micro-scale specimens such as biofilms. Data are generated by emitting a continuous sound wave at a controlled frequency of 2 MHz through metallic specimens of varying heights each containing an anomaly in the form of a hole of equal radius. Data are collected as the transmitted signal is sampled at several lateral locations on the opposite side of the specimen.  We have developed both a categorical and regression based CNN to analyze the acoustic signal in the frequency domain. Both CNNs take spectrograms representing the change in the amplitude, phase, or both over multiple observation points as input. The first CNN classifies the specimen in regards to four acoustic path length categories: small, medium, large, and largest. The second CNN analyzes the specimens and produces an estimation of the acoustic path length of the anomaly in radians. Both models perform with high accuracy, the categorical achieving upwards of 95% and the regression predicting within a fraction of a radian. With the performance of these models, we demonstrate that utilizing the transfer function 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.