TimeFi: Level-wise Contrastive Learning for Robust Wireless Sensing via Time-Warping Perturbations

Wireless sensing-based human activity recognition (HAR), which uses RF signals such as Wi-Fi CSI and UWB radars, has attracted increasing attention from both academia and industry. However, RF-based HAR often suffers from substantial cross-domain performance degradation because user habits, motion execution rates, and dynamic environments introduce domain shifts that are difficult to model. To mitigate this challenge, we propose TimeFi as a complementary framework for cross-domain wireless HAR that combines class-aware alignment and temporally informed augmentation. Specifically, TimeFi adopts a two-level contrastive strategy, in which the first-level objective preserves view consistency across augmentations, while the second-level class-aware objective promotes class-level alignment and reduces domain discrepancy. TimeFi further introduces a spectrogram-oriented time-warping augmentation that applies perturbations along the temporal axis to mimic nonlinear variations in activity speed while preserving the frequency layout of RF spectrograms, thereby improving robustness to temporally structured variability. In addition, we design a multi-level fusion module that decouples and hierarchically integrates temporal, frequency, and channel information to capture global dependencies in wireless signals. Experiments on multiple public wireless sensing benchmarks show that TimeFi consistently outperforms strong baselines in unseen-user and unseen-environment settings, demonstrating that class-aware alignment together with temporally informed augmentation effectively improves cross-domain robustness in wireless sensing-based HAR.