Novel dataset and model for restroom sound event classification

Area of Science:

• Artificial Intelligence
• Machine Learning
• Environmental Monitoring

Background:

• Accurate classification of hygiene and water usage events in restrooms is crucial for resource management and public health.
• Existing methods often lack privacy preservation or fine-grained event detection capabilities.

Purpose of the Study:

• To develop a novel, privacy-preserving deep learning framework for classifying fine-grained hygiene and water-usage events in restrooms.
• To create and release a comprehensive dataset for training and evaluating such models.

Main Methods:

• Utilized stereo audio recordings from diverse bathroom environments.
• Transformed audio into triple-channel Mel spectrograms using a 1D-CNN.
• Employed RegNetY-008 architecture with semi-supervised learning, pseudo-labeling, and data augmentation (XY masking, horizontal CutMix).
• Developed an ensemble model combining multiple RegNetY-008 networks.

Main Results:

• Achieved 97.8% accuracy and a 0.966 macro-averaged F1-score across acoustically distinct environments.
• Robustly identified 11 distinct events, including faucet usage, toilet flushing, and handwashing.
• Demonstrated strong generalization performance.

Conclusions:

• The proposed privacy-preserving deep learning framework offers accurate and robust classification of restroom events.
• The publicly available dataset will facilitate future research in intelligent, privacy-conscious restroom monitoring.