PULSE: physics-aware temporal embedding learning for domain adaptive wireless sensing

Abstract

We present PULSE, a lightweight domain-adaptive sensing framework that learns the physics-aware temporal embeddings, extracted from Wi-Fi channel frequency response (CFR). Unlike the existing approaches where the CFR is directly used as the input the tensor for Learning model, PULSE extracts temporal descriptors and embeds them through a light-weight 1D convolutional network that jointly learns discriminative representations and sensing semantics. The framework achieves over 99% accuracy while reducing 85% of the input tensor dimensionality and maintaining low inference latency, demonstrating suitability for real-time edge inference. Furthermore, through supervised contrastive pretraining and few-shot adaptation, PULSE generalizes effectively to unseen domains using only 5 s worth of labeled data, outperforming the state-of-the-art frameworks. We have extensively evaluated PULSE sensing framework with publicly available dataset of 20 different activities collected over multiple, subjects and propagation environment. For reproducibility, the code base will be made available at: https://github.com/rifatzabin/PULSE.