Data Augmentation and Attention for massive MIMO-based Indoor Localization in Changing Environments

TL;DR

This work tackles millimeter-level indoor localization with massive MIMO CSI in dynamic environments, where models trained on static data struggle as blockages and channel conditions change. It proposes two data augmentations to mimic blocked antennas and inserts two Attention modules (subcarrier and antenna level) into a state-of-the-art CSI-based DL model. On static training data augmented with these techniques, the approach achieving 66 mm mean error in changing environments, a significant improvement over baseline results, and shows that training directly on changing data can reach around 12–13 mm. These results demonstrate strong generalizability to dynamic indoor scenarios and highlight practical implications for reliable, real-time localization in evolving spaces; code is open-source for reproducibility.

Abstract

The demand for high-precision indoor localization has grown significantly with the rise of smart environments, industrial automation, and location-aware applications. While massive Multiple-Input and Multiple-Output (MIMO) systems enable millimeter-level accuracy by leveraging rich Channel State Information (CSI), most existing solutions are optimized for static environments, where users or devices remain fixed during data collection and inference. Real-world applications, however, often require real-time localization in changing environments, where rapid movement, unpredictable blockages, and dynamic channel conditions pose significant challenges. To address these challenges, we introduce two data augmentation techniques designed to resemble blocked antennas, enhancing the generalizability of localization models to dynamic scenarios. Additionally, we enhance an existing Deep Learning (DL) model by incorporating attention modules, improving its ability to focus on relevant channel features and antennas. We train our model on data from a static scenario, augmented with the proposed techniques, and evaluate it on a dataset collected in changing scenarios. We investigate the performance enhancements achieved by the data augmentation techniques and the Attention modules, and observe a localization accuracy improvement from a mean error of 286 mm, when trained without Attention and without data augmentations, to 66 mm, when trained with Attention and data augmentation. This shows that high localization accuracy can be maintained in changing environments, even without training data from those scenarios.

Figures (6)

• Figure 1: Our proposed enhancement of the DNN in bast_csi-based_2019. We insert two Attention modules, shown in pink, to enable the model to focus on specific antennas and subcarriers. The rest follows the original model architecture.
• Figure 2: The setting of the nomadic dataset dataset. The black stars show the user positions, in blue the ULA, and in orange six trajectories of a human walking, each defining a new scenario in which data is collected. For the static scenario, no human is moving.
• Figure 3: Cumulative distribution function of DN's and ADN's localization error, both trained either without data augmentations (None), with the vanilla data augmentations (vanilla) or with random attenuation (RA).
• Figure 4: Mean error when evaluating the models on the changing scenarios. In blue, the DN, and in red, the ADN, both trained with either no data augmentations (None), vanilla data augmentations (Vanilla), random attenuation (RA), or directly on data from the changing scenarios (Upper Bound). The last case serves as an indication of how well the models can perform, and is only evaluated on a subset of the data, as the rest has been used for training.
• Figure 5: Example error curves of different models, trained with different data augmentation techniques, evaluated on the changing scenario where a human walks up and down between the ULA and the four users. ADN trained with random attenuation shows the most stable behavior among all four users (d).