EMSTS-Pos: AI-driven indoor positioning with enhanced-multipath space-time structural features in O-RAN system

Driven by the integration of wider bandwidths and massive antenna arrays, mobile cellular networks demonstrate significant potential for supporting high-precision positioning functionalities. This article introduces the EMSTS (Enhanced-Multipath Space-Time Structural) feature set, which is a substantial extension of our prior conference work SPSTS(Strongest-Path Space-Time Structural) feature set. Our solution achieves robust and precise positioning in real-world indoor environments. The core methodology involves extracting the relative time-delay differences using uplink Sounding Reference Signals(SRS). Although the baseline SPSTS-Pos achieved superior positioning accuracy on both validation datasets and real-time system, with 50 th and 90 th percentile CDF errors of 0.054 m and 0.228 m, its sole reliance on the strongest MPC resulted in inadequate tail performance, leading to a maximum absolute error (MAXE) reaching 2.743 meters and the 95 th percentile CDF error reaching approximately 0.7 meters.To resolve this reliability issue, we propose the EMSTS-Pos method with three major technical enhancements: 1) The feature set is expanded to EMSTS by incorporating the relative indices of the 1 st , 2 nd , and 3 rd strongest MPC to enhance spatial redundancy. 2) A custom, knowledge-driven channel attention mechanism is introduced to apply a non-random weight bias, prioritizing the most discriminative path information. 3) We utilize comprehensive data robustness methods, including temporal forward filling for raw data outlier suppression and Gaussian noise regularization on coordinates. The EMSTS feature maintained the strong average performance of the SPSTS baseline while achieving a substantial enhancement in CDF95% of 0.45 m and MAXE of 1.10 m, resolving the critical ’tail effect’ issue.