Self supervised multi scale learning for WiFi CSI contactless human activity recognition with potential applications in health monitoring

WiFi Channel State Information (CSI) provides a privacy-preserving and contactless sensing modality for human activity recognition, but practical CSI recognition remains challenged by limited labeled data, heterogeneous motion dynamics, and signal perturbations. This study proposes CSI-MSRNet, a self supervised representation-learning framework that combines CSI-aware contrastive augmentation with a lightweight temporal and subcarrier multi scale encoder. The encoder models motion evolution at multiple temporal receptive fields and complementary frequency-selective responses across OFDM subcarriers, while Efficient Multi-Scale Attention is used only as a lightweight feature-refinement component rather than as the primary methodological novelty. Self supervised pretraining is restricted to the training partition, after which the encoder is fine tuned with different proportions of labeled samples. Experiments were conducted on the public UT-HAR and NTU-Fi HAR benchmarks using the SenseFi data partitions. Main recognition and low-label experiments were repeated five times with random seeds 42–46, and low-label subsets were independently sampled by stratified class sampling in each run. CSI-MSRNet achieved 98.76 ± 0.18% accuracy on UT-HAR and 99.85 ± 0.21% on NTU-Fi HAR. With only 1% labeled data, the corresponding accuracies were 73.6 ± 2.4% and 77.8 ± 2.8%, respectively. Controlled synthetic tests further indicated improved tolerance to amplitude noise and random subcarrier loss. These results should be interpreted as benchmark-level behavioral recognition rather than evidence of subject-independent generalization, clinical monitoring effectiveness, or embedded-device deployment. The framework therefore provides a computationally lightweight approach with potential applications in smart-home and health-related behavioral sensing.