Witnessing its large swaths of success in various fields, digital twins (DTs) are considered a promising scheme for 6th Generation (6G) cellular systems, showing a leading edge in networking and communication modelling. However, another 6G core property ofhigh-precision positioningcan hardly be supported by existing 6G DT solutions due to the lack ofenvironmental modellingandsignal interactions with physical scenes. This shortcoming yields a series of challenges in 6G DT-enabled positioning, including positioning data acquisition, accuracy enhancement, and continuous optimization. In this regard, we propose a novel paradigm of localization-oriented DT (LocDT) with a compound architecture of 7 sub-DT layers to characterize the 6G integrated-localization-and-communication (ILAC) feature. LocDT starts from a physical environment sublayer to mirror 6G signal interactions within a real-world scenario, along with an ILAC baseband sublayer and a channel frequency Polar-coordinate (CFP) image construction method to provide finer-grained fingerprints. Furthermore, insight from LocDT reveals an interesting phenomenon: the channel features of Line-of-Sight (LoS) / None-Los (NLoS) gNodeBs makedifferentiated-contributionsto positioning accuracy, especially in wide-existingpartial-LoS-coveragescenarios. Benefiting from this, a DT-driven Artificial Intelligence (AI) positioning model, SSI-Net, is designed with a device-attention mechanism, achieving complementary improvements in accuracy. Evaluation results show LocDT and SSI-Net’s advantages from a position-of-strength in accuracy and time overhead, outperforming state-of-the-art models.

Location-based services (LBSs) provide necessary infrastructure for daily life, from bicycle sharing to nursing care. In contrast to traditional positioning methods such as Wi-Fi, Bluetooth, and ultra-wideband (UWB), fifth-generation (5G) networking is defined as a paradigm ofintegrated sensing and communication(ISAC). With its advantages of wide-range coverage and indoor-outdoor integration, 5G is promising for high-precision positioning in indoor and urban canyon environments. However, 5G location studies face great obstacles due to the lack of commercialized 5G ISAC base stations that support positioning functions as well as publicly available datasets. In this paper, we first propose a dataset generation method, the Multilevel Feature Synthesis Method (Multilevel-FSM), to obtain positioning features. In particular, the features of a multiple-input multiple-output (MIMO) channel are flattened into a single image to increase the information density and improve feature expression, and data augmentation is performed to provide stronger robustness to noise. Subsequently, we devise a specially designed deep learning positioning method, Multipath Res-Inception (MPRI), trained on the proposed dataset to enhance positioning accuracy. Finally, the results of extensive experiments conducted in two typical 5G scenarios (indoors and urban canyon) show that Multilevel-FSM and MPRI outperform state-of-the-art works in accuracy, time overhead and robustness to noise.