Mining Limited Data Sufficiently: A BERT-inspired Approach for CSI Time Series Application in Wireless Communication and Sensing

Channel State Information (CSI) is the cornerstone in both wirelesscommunication and sensing systems. In wireless communication systems, CSIprovides essential insights into channel conditions, enabling systemoptimizations like channel compensation and dynamic resource allocation.However, the high computational complexity of CSI estimation algorithmsnecessitates the development of fast deep learning methods for CSI prediction.In wireless sensing systems, CSI can be leveraged to infer environmentalchanges, facilitating various functions, including gesture recognition andpeople identification. Deep learning methods have demonstrated significantadvantages over model-based approaches in these fine-grained CSI classificationtasks, particularly when classes vary across different scenarios. However, amajor challenge in training deep learning networks for wireless systems is thelimited availability of data, further complicated by the diverse formats ofmany public datasets, which hinder integration. Additionally, collecting CSIdata can be resource-intensive, requiring considerable time and manpower. Toaddress these challenges, we propose CSI-BERT2 for CSI prediction andclassification tasks, effectively utilizing limited data through a pre-trainingand fine-tuning approach. Building on CSI-BERT1, we enhance the modelarchitecture by introducing an Adaptive Re-Weighting Layer (ARL) and aMulti-Layer Perceptron (MLP) to better capture sub-carrier and timestampinformation, effectively addressing the permutation-invariance problem.Furthermore, we propose a Mask Prediction Model (MPM) fine-tuning method toimprove the model's adaptability for CSI prediction tasks. Experimental resultsdemonstrate that CSI-BERT2 achieves state-of-the-art performance across alltasks.