Enabling non-intrusive occupant activity modeling using WiFi signals and a generative adversarial network

Abstract

Occupant activity (OA) is a crucial prerequisite for providing energy-efficient and occupant-centric services in intelligent buildings. To understand OAs well, conventional technologies require either wearable equipment of smart devices or the deployment of specialized cameras, raising relevant issues of body and privacy-intrusion. Recent WiFi-based methods circumvent the above issues and characterize the OA patterns in a non-intrusive manner. However, they demand the cooperation of occupants for a sufficient amount of collected samples, otherwise, the performance might degrade significantly. In this paper, we propose a non-intrusive approach that models distinct OA patterns for comprehensive understanding. Based on the idea of generative adversarial network (GAN), our technical novelties are three-fold. First, we modify the working principle of vanilla GAN with external constraints to avoid brute-force generation. Second, we integrate a self-attention mechanism to establish contextual relations from both local and global perspectives. Third, to recover the informative details, we construct a powerful generator with deep residual convolutional operations. We conduct extensive experiments on a real dataset and visualize the generation results for intuitive evaluation. Numerical results compared with several state-of-the-art baselines further illustrate the superior performance of our proposed GAN approach.