Ultrasonic Frequency Anomaly Localization with Machine and Deep Learning

Acoustic ultrasonic signals, characterized by frequencies exceeding 20 kHz, indicate electronic equipment malfunctions, electrical arcs, and pressurized volume integrity loss. Detecting and localizing these signals are crucial for preserving system integrity in space habitats and industrial settings. Traditional localization techniques, like Time Difference of Arrival (TDOA), face challenges when dealing with a broad frequency band. This research investigates the feasibility of employing machine learning, specifically an Artificial Neural Network (ANN), to predict the location of ultrasonic anomalies using Fast Fourier Transform (FFT) data from a linear array of MEMS microphones. A dataset obtained through a custom data collection setup undergoes rigorous preprocessing and analysis. The study compares the performance of various machine learning models, demonstrating that the ANN exhibits superior accuracy, precision, and recall. The results suggest that machine learning holds promise for effective ultrasonic anomaly localization in dynamic and mobile systems, paving the way for future developments.