WiFi Based Multi-Task Sensing via Selective Sharing Module

Boyu Yang,Ting Jiang

DOI: https://doi.org/10.1109/VTC2023-Spring57618.2023.10199587

2023-01-01

Abstract:WiFi-based sensing technology has become a very popular research area. However, the current major research focuses on single-task sensing, and the very few studies on WiFi multi-task sensing expose the problems of unbalanced sharing of information between multiple WiFi sensing tasks and unclear task relevance. To address these issues, this paper propose Wimuss, a WiFi-based multi-task sensing system that can perform three tasks simultaneously: activity recognition, identity recognition, and indoor localisation. The system first uses Convolutional Neural Networks and Bi-directional Long and Short Term Memory (CNN-BiLSTM) as the base model to simultaneously extract temporal and spatial features of WiFi data. Subnetworks are then constructed using a selective sharing module, and multiple subnetworks are combined with each other and trained together in conjunction with task-specific layers to achieve reasonable sharing of partial parameters of each task. We collect a large amount of data in real scenarios to fully verify the feasibility of the constructed multi-task sensing system. The evaluation results show that Wimuss achieves accuracy rates of 92.5%, 96.5% and 98.3% for activity recognition, identity recognition and indoor localisation respectively. More importantly, we also demonstrate that the application of selective sharing module in WiFi-sensing models outperforms other sharing mechanisms in multi-task learning, and effectively improves task accuracy and model generalisation performance.