Sub-6G Domain-Adaptive Non-Contact Sensing Using Transformer and Prototypical Neural Network

In recent years, Sub-6G non-contact sensing using Channel State Information (CSI) has garnered significant attention and yielded promising results. Particularly, as deep learning advances, an increasing number of studies are employing neural network models for non-contact sensing. However, most deep learning methods only consider short-range correlation among features. So when a model trained in a specific domain is applied to new scenarios, such as different rooms or users, its recognition performance deteriorates significantly. Therefore, this article focuses on Sub-6G non-contact sensing by incorporating the transformer method with long-range correlation. Specifically, a novel domain-adaptive non-contact sensing method for gesture recognition (DAGR) is proposed by integrating the Vision Transformer (ViT) with the Prototypical Neural Network (PNN) and adaptive few-shot learning. The method includes a prototypical feature extractor based on ViT, a domain-adaptive prototypical classifier based on PNN, and an optimally-trained domain-adaptive prototypical feature space. DAGR performs classification by assessing the similarity between the features of CSI samples and the prototypes of different labels, instead of directly utilizing CSI signals. Simultaneously, employing the fast Fourier transform, the time delay of the signal is addressed in the frequency domain to yield higher-quality data for subsequent analysis in this paper. Through conducting various experiments on multiple datasets, the results show that within the domains, the method introduced in this study achieved accuracies exceeding 94% across all three datasets; In cross-domain scenarios, the method outperforms traditional models. Furthermore, compared to other small-sample learning techniques, the proposed DAGR demonstrates superior recognition capabilities.