Sound-Event Detection of Water-Usage Activities Using Transfer Learning

Area of Science:

• Engineering
• Computer Science
• Acoustics

Background:

• Accurate detection of daily activities is crucial for smart home applications.
• Existing sound event detection methods may lack specificity for distinct water usage sounds.
• Privacy concerns often limit the deployment of audio-based monitoring systems.

Purpose of the Study:

• To propose a novel sound event detection method for identifying specific bathroom activities (showering, flushing, faucet usage).
• To develop a privacy-preserving system using edge computing.
• To evaluate the performance of the proposed method and explore its potential applications.

Main Methods:

• A two-stage approach utilizing YAMNet for general water sound detection and a modified network (W-YAMNet) for specific activity classification.
• Transfer learning was employed to train W-YAMNet, adapting it to unique acoustic characteristics of bathrooms.
• Parameter analysis, including audio clip length, was conducted to optimize performance.
• Implementation on a Raspberry Pi-based edge computer for enhanced privacy.

Main Results:

• The proposed method demonstrated promising results in detecting and classifying bathroom water usage sounds from 10-min audio segments.
• The W-YAMNet model was successfully adapted to individual bathroom acoustics.
• The edge computing implementation ensured data privacy.

Conclusions:

• The developed sound event detection system effectively identifies specific bathroom activities.
• Further enhancements in accuracy are possible.
• The system holds potential for health and safety monitoring of elderly individuals living alone.