Handling Device Heterogeneity for Deep Learning-based Localization

TL;DR

The paper tackles device heterogeneity in deep-learning–based outdoor cellular localization using RSS fingerprinting. It analyzes traditional RSS-mapping methods and introduces two DL-based strategies—transfer learning and multitask learning—to enable cross-device adaptation with limited data on new devices. Across four Android devices and four testbeds, the proposed heterogeneity techniques yield substantial gains, achieving median localization errors as low as $24\mathrm{ m}$ and improvements up to $>1176\%$ over baselines without heterogeneity handling. These methods demonstrate practical, scalable deployment of DL-based localization in heterogeneous device ecosystems.

Abstract

Deep learning-based fingerprinting is one of the current promising technologies for outdoor localization in cellular networks. However, deploying such localization systems for heterogeneous phones affects their accuracy as the cellular received signal strength (RSS) readings vary for different types of phones. In this paper, we introduce a number of techniques for addressing the phones heterogeneity problem in the deep-learning based localization systems. The basic idea is either to approximate a function that maps the cellular RSS measurements between different devices or to transfer the knowledge across them. Evaluation of the proposed techniques using different Android phones on four independent testbeds shows that our techniques can improve the localization accuracy by more than 220% for the four testbeds as compared to the state-of-the-art systems. This highlights the promise of the proposed device heterogeneity handling techniques for enabling a wide deployment of deep learning-based localization systems over different devices.

Figures (4)

• Figure 1: The proposed deep learning localization model. The input is the RSS coming from different cell towers in the environment (features). The output is a probability distribution over different reference locations in the area of interest. Gray-shaded neurons represent examples of nodes that have been temporary dropped-off to increase the model robustness and avoid over-training.
• Figure 2: Transfer learning deep model structure. Input is the cell towers RSS information. Gray layers are responsible for extracting the relation between the RSS input features. The output layer is trained once as a part of the network using a specific device (solid layer) with a large number of samples. Then, the model weights are tuned using a small samples from the different devices (dashed layers).
• Figure 3: Multitask learning deep model structure. Input is the cell towers RSS information from different devices. Gray layers are responsible for extracting the relation between the RSS input features. The output layers (dashed layers) are trained simultaneously using a small number of samples from each phone type.
• Figure 4: Effect of different device heterogeneity techniques on localization accuracy.