WiPure: Structurally Symmetric Cross-Attention with Fixed-Wavelet, Trainable Band Processing for Interference-Robust WiFi Activity Recognition

WiFi channel state information (CSI)-based human activity recognition remains challenging in dynamic indoor environments because water flow and non-target motion perturb the same multipath channel as target activities. We present WiPure, a dual-stream model that combines CSI amplitude and adjacent-antenna phase difference through fixed Symlet-4 decomposition, trainable per-band filtering, multi-scale temporal pooling, and structurally symmetric bidirectional cross-attention. WiPure was evaluated on 2000 samples covering five activities, five volunteers, two rooms, and paired interference states. It achieved 95.4±0.7% accuracy under bathroom water flow interference and 91.8±1.1% under laboratory non-target person interference, outperforming the strongest comparator by 1.1 and 2.4 percentage points, respectively. Five-fold leave-one-subject-out evaluation yielded 82.6±1.7% accuracy with a 95% confidence interval of 80.4–84.8%, demonstrating transfer to unseen participants. Fusion architecture comparisons and complete factorial ablations further confirmed the advantages of reciprocal cross-modal interaction and the combined WTM–MPM–BCM design. Without fine-tuning, WiPure achieved 78.5±1.8% and 76.0±2.3% accuracy in the two cross-environment directions. These results support multi-resolution amplitude–phase modeling as an effective approach to interference-robust and privacy-preserving indoor WiFi sensing.