A NOVEL APPROACH FOR BREATHING SOUND DETECTION SYSTEM USING DEEP LEARNING

Abstract

In this study a groundbreaking work on the detection of human breathing, a critical physiological function traditionally monitored by healthcare professionals is presented. The research introduces an innovative acoustic-based approach that leverages a deep learning model to discern the ultra-low intensity sound of respiration from ambient noise within a room. Utilizing a microphone, the model captures air pressure variations induced by respiratory activities, which are then transformed into a spectrogram image classification problem. The methodology involves applying a Fast Fourier Transform to the .wav file recordings, converting the time-series data into a frequency domain to visualize the subsonic breathing rates as spectrograms. These representations serve as inputs for a pre-trained ResNet18 model, enriched with additional layers through transfer learning, to identify breathing sounds with remarkable accuracy. The study displayed a significant 93% accuracy in detecting respiration within a 3-meter radius and has further developed into a multi-class classifier capable of estimating the distance from the sound source with an 87% accuracy. A bespoke dataset comprising 1700 instances with 13 classes has been curated for breath detection. The implications of this research are profound, offering substantial advancements in non-invasive monitoring techniques. The model’s potential applications extend to disaster response and security monitoring, where it can significantly improve human detection in environments with poor visibility or restricted access. This thesis delineates the challenges, potential applications, and future research directions, marking a pivotal contribution to the interdisciplinary fields of biomedical engineering, electronics, and machine learning.