IBSS: Intelligent Behavior Sensing System Based on Comparative Language-CSI Learning

In recent years, the integration of communication and sensing has become a hot research field, among which human behavior sensing (HBS) based on Channel State Information (CSI) is particularly active. However, current methods often simplify the sensing task as a prediction problem for fixed category labels, that is, establishing a direct mapping relationship between input features and preset labels through training. Although this paradigm performs well in specific single task scenarios, the model’s generality and task generalization ability are limited. Inspired by Contrastive Learning and the multi-modal Contrastive Language-Image Pretraining model (CLIP), this paper proposes a novel intelligent behavior sensing system (IBSS), based on improved Contrastive Language-CSI Learning. IBSS converts behavior labels into supervised text in natural language form, and uses Transformer model to construct text encoder and CSI encoder respectively. Among them, the text encoder extracts task related semantic features from the supervised text generated by discrete labels, while the CSI encoder extracts feature representations related to sensing tasks from CSI data. By contrastive learning and training, the features of the two modalities are mapped into a joint feature space, and human behavior sensing is achieved by capturing the similarity between CSI features and text features. In addition, thanks to the advantages of the Transformer model in capturing global contextual information between tasks, the proposed method is not only suitable for single-task sensing, but can also be extended to multi-task sensing scenarios with correlation. To verify the effectiveness of the proposed method, we conducted experimental evaluations on public dataset and self-built dataset. The experimental results show that under the same training conditions, the average recognition accuracy of our method in gesture sensing and position sensing tasks reaches over 95%. Compared with the existing state-of-the-art methods, the recognition performance is effectively improved. This achievement provides theoretical support and technical foundation for the application of wireless signal based HBS technology in a wider range of scenarios.