Journal of Korean Institute of Architectural Sustainable Environment and Building Systems ISSN:1976-6483(Print) 2586-0666(Online) 한국건축친환경설비학회논문집

This study proposes a Convolutional Neural Network (CNN) model for classifying indoor occupant activities—boiling, frying, and vacuuming—using Mel-spectrograms derived from audio data. Compared to traditional CO₂ or vision-based approaches, the proposed method offers a non-intrusive and privacy-preserving alternative suitable for smart environment control. To enhance interpretability, Gradient-weighted Class Activation Mapping (Grad-CAM) and Fast Fourier Transform (FFT) analyses were applied. Grad-CAM visualizes the time-frequency regions the model focuses on during classification, while FFT quantitatively identifies key spectral patterns. The combined analysis provides insight into the model’s decision-making process, validating that it relies on meaningful audio features, such as the 100–200 Hz range for vacuuming. The approach demonstrates not only high accuracy (95.0%) but also explainability, reinforcing trust in audio-based behavior recognition. These findings lay a foundation for practical applications in intelligent indoor systems, contributing to occupant-aware environmental control and automation technologies.