Abstract
The increasing demand for enhanced energy efficiency and occupant comfort in buildings has led to a focus on developing occupant-centric building control systems. A key aspect of this study is to obtain real-time and accurate information related to occupants' activities. This study presents WiSA, an innovative, WiFi-based framework for recognizing the intensity of occupants' activities. WiSA offers non-intrusive, personalized activity recognition while minimizing privacy risks through the use of federated learning. We developed a specialized deep learning neural network for classifying activities into different intensity levels (light, moderate, vigorous), incorporating a tailored fine-tuning strategy to address the complexities of federated learning. Our dataset, exhibiting non-identically and independently distributed (No-IID) characteristics, was derived from the activities of 15 occupants in two rooms. This dataset was instrumental in demonstrating WiSA's effectiveness. The results show that WiSA achieved an impressive 98.0% average accuracy across all clients without necessitating the upload of annotated local data. The personalized fine-tuning strategy enhanced the performance of federated learning models by an average of 18.7%, markedly increasing adaptability to No-IID data. Additionally, WiSA displayed substantial robustness in handling the resolution of CSI frames. This approach offers a privacy-conscious method for building managers to gather information on occupant activities, enabling the development of more effective building operation strategies.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have