Self-Attention based Semi-Supervised Learning for Time-varying Wi-Fi CSI-based Adjoining Room Presence Detection

Abstract

Device-free indoor human presence detection problem has been studied in recent years based on supervised learning and wireless signals such as channel state information (CSI). Thanks to the abundant spatial information of CSI, we can perform indoor presence detection more accurately. Nevertheless, the CSI is susceptible to humidity, temperature, and even machine restarts, resulting in unexpected changes and prediction failures of the trained model. The most intuitive approach to resolve this time-varying phenomenon is to retrain the model by recollecting and labeling CSI data. However, it is time- and labor-consuming to label the dataset every time we retrain the model. Our proposed self-attention based semi-supervised learning for adjoining room presence detection (SAS-PD) system provides an alternative solution to retrain the detection model without any effort to label data and consequently overcome the time-varying issue. The proposed teacher/student training strategy can effectively classify four classes based on a single permanent labeled CSI dataset and recollected unlabeled CSI datasets. In addition, the positional encoder is adopted to enhance time domain correlations of CSI data. Experimental results show that our proposed SAS-PD system can provide enhanced presence detection accuracy under time-varying environments, and it can almost reach the upper bound of the training algorithm with supervised learning.