Learning Motion-Induced Channel Dynamics with Multiresolution Multiplex Graphs for Wi-Fi CSI-Based Human Activity Sensing

Abstract

Highlights What are the main findings? Spectral alignment and relative channel-response fields reduce acquisition-dependent input mismatch. Multiresolution multiplex graphs encode signed subcarrier relations within and across temporal resolutions. What is the implication of the main finding? Signal constraints and relational inductive biases are learned within one CSI sensing framework. The measured complete model has a cross-domain mean of 48.81%, compared with 43.97% for the strongest of ten locally reevaluated baselines. Abstract Wi-Fi channel state information (CSI) supports human activity recognition from motion-induced changes in indoor propagation, yet device, environment, and user changes perturb both channel responses and subcarrier relations. We present a model that aligns valid subcarriers across inputs and combines CSI response features with a multiresolution graph. Temporal decomposition produces residual components at multiple resolutions and a smooth component. The graph propagates information along positive and negative relations between subcarriers within each resolution and uses signal energy to guide propagation across resolutions. A gate incorporates the graph representation into the response classifier. A controlled multipath study further tests the physical interpretation of signed amplitude correlations. Experiments on CSI-Bench yield weighted-F1 scores of 95.96%, 50.16%, 44.08%, and 52.19% on the four protocols, with a cross-domain mean of 48.81%, exceeding the strongest of ten locally reevaluated baselines by 4.84 points.