Noninvasive Acoustic Recognition of Water Flow Sources for Human Activity Monitoring in Smart Homes

This paper presents a noninvasive system for identifying water flow sources with the final goal of supporting human activity recognition (HAR) in activities of daily living (ADL). The system employs a single microphone to capture ambient sounds within a room and detects active water sources based on their acoustic signatures. The audio signals are converted into time-resolved spectrograms, which are processed via time- and frequency-domain convolution and subsequently classified using a neural network. This approach enables both the identification of specific water sources, even combined, and the measurement of their usage duration with an overall accuracy of 90.6%. The study focuses on four common bathroom fixtures: toilet, bidet, shower, and washbasin. The proposed system is adaptable to various environments, requires no modifications to plumbing infrastructure, making it suitable for smart home and digital health applications.