Acoustic explosion and rocket signatures from surface and airborne smartphones

Abstract

Rocket launches produce complex acoustic signatures with large amounts of energy in low (<300 Hz) and infrasonic (<20 Hz) frequency ranges. These acoustic signatures carry information about their source as they propagate through the atmosphere, which, if correctly detected and classified, can aid in monitoring efforts. Many infrasonic rocket signatures have been successfully collected after propagating great distances (>1000 km), but most of these signatures are highly attenuated, yielding broad insight but little detail into the nature of the launch that produced them. In addition, very few of these data are available to the public, limiting the rate of progress in the field of rocket acoustics.The overarching goal of the work covered in this dissertation is to develop a fast and reliable method of detecting acoustic rocket launch signatures. Results from each of the three content chapters contribute towards this goal. In Chapter 2, a surface chemical explosion signal is observed from an ascending balloon in the middle stratosphere. From this case of a stationary surface source and an ascending airborne receiver, we can gain insight into the inverse case of an ascending airborne source (a rocket) and a stationary surface receiver by invoking reciprocity. In addition, a comparison of low-cost, low-maintenance sensors (smartphones) with traditional infrasound sensors is made. In Chapter 3, a dataset of acoustic rocket signatures collected on smartphones at estimated propagation distances of 10-70 km is presented and released to the public, accompanied by preliminary analysis of the chronology and time-frequency characteristics observed in the dataset. In Chapter 4, this dataset is used in concert with two other open-access datasets to train and test machine learning models for near-real-time detection of acoustic rocket launch signatures on mobile platforms. The best performing model showed promising results, with an overall accuracy of 97% and a false positive rate of <1%. Measures to improve the model’s suitability for persistent monitoring are discussed, implemented, and evaluated, resulting in an estimated effective true positive rate of 99% and false positive rate of 0.07%.