Lightweight Network based on Depth-separable Convolution for 5G Indoor Positioning

The importance of 5G positioning lies in its ability to provide high-precision location accuracy, impacting network optimization, emergency response and safety, augmented reality and virtual reality, and commercial applications. 5G positioning offers accurate location information, fostering innovation and improving user experience and quality of life. The importance of 5G CSI fingerprint positioning for machine learning-based methods has led to rapid development in recent years. However, existing models from computer vision and natural language processing lack adaptation to the physical meaning of 5G CSI, resulting in suboptimal performance. In this paper, we propose a lightweight convolutional neural network (CNN) based 5G positioning method that considers spatial and frequency domain modeling. The method utilizes interpretable CNN weights to explore dependencies on frequency domain information across different base stations. Performance validation is conducted on three real-world 5G positioning datasets with varying signal-to-noise ratios.