Interpretable Hybrid Vision Transformer Architectures for MIMO-Based Indoor Localization Using Synthetic Spatial Representations

Liu J.; Castillo-Cara M.; García-Castro R.; Orozco-Barbosa L. · IEEE Internet of Things Journal (2026) · DOI: 10.1109/jiot.2026.3696106

Resumen

Channel state information (CSI) from massive multiple-input multiple-output (MIMO) systems provides fine-grained spatial- and frequency-domain information for indoor localization, yet learning effective representations from CSI remains challenging, particularly when only a limited number of antennas is available, as is common in practical deployments. In this work, we investigate the use of synthetic image generation methods, originally proposed for tabular data, to transform CSI vectors into structured image-like representations that can be processed by vision models while preserving feature relationships induced by the transformation. Within this setting, we introduce HyViT, a hybrid architecture that combines a vision transformer (ViT) branch operating on synthetic CSI images with a multilayer perceptron branch operating on the raw CSI vector. We evaluate the proposed approach on a large-scale MIMO indoor localization benchmark spanning multiple antenna topologies and reduced-antenna configurations. The results show that HyViT consistently improves localization accuracy over classical regression baselines and prior vision-based models under the same evaluation protocol. In particular, it achieves state-of-the-art performance using only 8 or 16 antennas, without data augmentation, and attains accuracy competitive with previously reported results obtained with 64 antennas. We also analyze model interpretability using attention attribution metrics across spatial regions, antenna groupings, and modalities. These findings provide empirical support for the effectiveness and interpretability of combining tabular data into synthetic image transformations with hybrid vision-tabular architectures for CSI-based indoor localization.

Instituciones: Universidad de Lima