Enabling Light-Weight AI-Based Positioning by Log Signature Transform of Wireless Channels

The conflict between increasing demands for high accuracy positioning information and poor performance of conventional schemes in non-line-of-sight (NLOS) scenarios motivates the study to find new promising approaches. Artificial intelligence or machine learning (AIIML) based schemes have shown great potential to reach this goal. However, one key point is to design suitable data input for enabling AIIML for positioning in practice. And considering the increasing bandwidth, sampling rate, and number of antennas in wireless cellular systems, the commonly used raw channel information (e.g., channel impulse response CIR), which could be very high-dimensional, may not be appropriate for the input of AI models and result in high model complexity and demand for massive training samples. Therefore, in this paper, the log signature transform based method, which utilizes iterated integrals of paths, is proposed for AI-based positioning with novel time-augmented path construction and feature derivation. The proposed method can drastically reduce the input feature dimensionality and consequently reduce the AI model complexity. Moreover, it can have better accuracy for limited training samples or/and apply simpler AIIML models such as multi-layer perceptron (MLP). The experimental results show that around 98 % reduction in both input dimensionality and computational complexity can be achieved without compromising the positioning accuracy.