Measurement-Driven Uncertainty Modeling and Error Analysis for AI-Based 5G RSS/IMU Indoor Localization

Abstract

Reliable indoor localization using commercial fifth-generation (5G) received-signal-strength (RSS) measurements is challenged by device orientation, fingerprint (FP) ambiguity, and deployment geometry. This article presents a measurement-driven framework for 5G RSS/inertial measurement unit (IMU) indoor localization and retrospective error interpretation. Modality-specific Transformer encoders are first pretrained. An IMU-guided cross-attention module then aligns short-term motion features with radio FP contexts. A heteroscedastic likelihood head jointly estimates the 2-D position and input-dependent uncertainty. To investigate the spatial origins of residual errors, diagnostic variables are constructed from the measured FP field and base station (BS) geometry, including the effective residual noise floor, residual spatial variation, RSS gradient, visible-cell number, an empirical FP-field Fisher-information indicator, geometric dilution of precision (GDOP), and user-equipment (UE) orientation. A spatially varying ridge model is used to describe their location-dependent associations with localization error. Experiments are conducted in a hall dataset with one-sided nonideal geometry and a meeting-room dataset with different propagation conditions. In the hall testing set, the proposed method achieves a root-mean-square error (RMSE) of 1.99 m, reducing RMSE and trajectory-jump rate by $5.7{\,}\%$ and $43.3{\,}\%$ , respectively, relative to the direct RSS/IMU Transformer. The pretrained representations also improve limited-label learning and hall-to-meeting-room adaptation. The predicted uncertainty provides effective risk ranking in the hall. In retrospective error analysis, GDOP exhibits the strongest and a comparatively stable local association with localization error. UE orientation produces the most reproducible coefficient patterns across model seeds and neighborhood settings. Residual spatial variation shows a strong but neighborhood-scale-sensitive association. These results demonstrate that the proposed measurement-driven fusion method improves localization accuracy. The spatial analysis provides a location-dependent explanation of the localization errors produced by the neural network.