Machine learning for acoustic sensing in challenging environments

Abstract

Acoustic measurements provide unique capabilities that make them attractive for a wide range of monitoring, characterization, and classification applications where other sensing techniques based on electromagnetic waves or radiography have physical limitations. One of the main challenges in real-world applications of acoustic sensing is the information-rich nature of the measurements. These measurements contain vast amounts of information about the system being investigated, but it can be difficult to extract the useful aspects of the data from the complex raw data. Thus, machine learning (ML) techniques have been gaining popularity rapidly in acoustics, by identifying complex correlations in the measurements that are not apparent from traditional techniques. In this presentation, I focus on performing acoustic measurements in challenging environments, where data is highly complex and/or noisy. I discuss how ML can be applied to the main categories of acoustic data, i.e., time-series, frequency-spectra, and numeric features, and provide examples where each category is useful in a different challenging environment, such as in materials with thermal gradients, multi-component systems with complex vibrational coupling, and in fast-moving streams of flowing biomass.