[Paper Review] Improving the Performance of Deep Learning for Wireless Localization

Ramdoot Pydipaty, Johnu George+2 more|arXiv (Cornell University)|Jun 16, 2020

Indoor and Outdoor Localization Technologies15 references4 citations

TL;DR

This paper improves deep learning for wireless localization by introducing automatic hyperparameter optimization and a hybrid data augmentation technique using autoencoders and naive sampling. The approach reduces mean localization error by 15% (from 23.2 ft to 19.7 ft) on a real-world BLE RSSI dataset, with all code publicly released for reproducibility and reuse.

ABSTRACT

Indoor localization systems are most commonly based on Received Signal Strength Indicator (RSSI) measurements of either WiFi or Bluetooth-Low-Energy (BLE) beacons. In such systems, the two most common techniques are trilateration and fingerprinting, with the latter providing higher accuracy. In the fingerprinting technique, Deep Learning (DL) algorithms are often used to predict the location of the receiver based on the RSSI measurements of multiple beacons received at the receiver. In this paper, we address two practical issues with applying Deep Learning to wireless localization -- transfer of solution from one wireless environment to another \emph{and} small size of labelled data set. First, we apply automatic hyperparameter optimization to a deep neural network (DNN) system for indoor wireless localization, which makes the system easy to port to new wireless environments. Second, we show how to augment a typically small labelled data set using the unlabelled data set. We observed improved performance in DL by applying the two techniques. Additionally, all relevant code has been made freely available.

Motivation & Objective

• To address the challenge of deploying deep learning models in dynamic, real-world wireless environments where manual tuning is impractical.
• To improve model generalization in low-data regimes by augmenting small, imbalanced labeled datasets using unlabeled data.
• To enable rapid, automated adaptation of deep learning models to new indoor layouts without retraining from scratch.
• To develop a scalable, open-source solution for wireless localization that is both accurate and portable across different environments.

Proposed method

• Applying automatic hyperparameter optimization to a deep neural network (DNN) for RSSI-based localization, using Bayesian optimization to tune learning rate, batch size, and number of hidden units.
• Using a standard autoencoder to learn the distribution of unlabeled RSSI measurements and generate synthetic samples for under-represented locations.
• Implementing a naive data augmentation method that generates new samples via uniform sampling between min and max RSSI values for each beacon at each location.
• Combining the naive and autoencoder-based augmentation techniques into a hybrid approach to maximize labeled data diversity and coverage.
• Training and evaluating DNNs with Adam optimizer on the augmented datasets, comparing performance across original, naive-augmented, autoencoder-augmented, and hybrid-augmented data.
• Validating results on a real-world BLE fingerprinting dataset collected in a library environment with 1420 original labeled samples and 5191 unlabeled samples.

Experimental results

Research questions

• RQ1 Can automatic hyperparameter optimization improve the performance of deep learning models in wireless localization without manual tuning?
• RQ2 Can unsupervised data augmentation using autoencoders effectively improve model generalization in low-data regimes for RSSI fingerprinting?
• RQ3 Does combining naive and autoencoder-based augmentation yield better performance than either method alone?
• RQ4 To what extent can these techniques reduce localization error in real-world indoor environments with sparse and imbalanced data?

Key findings

• The hybrid data augmentation approach reduced the mean localization error from 23.2 ft (original data) to 19.7 ft, representing a 15% improvement.
• The autoencoder-based augmentation alone achieved a mean error of 20.7 ft, outperforming the naive approach (21.9 ft).
• The naive augmentation method increased the labeled dataset from 1420 to 1608 samples, improving performance over the original data.
• The hybrid method generated 346 new labeled samples (188 from naive + 158 from autoencoder), resulting in the largest improvement in model accuracy.
• The use of automatic hyperparameter optimization enabled faster, more scalable model adaptation across different indoor environments.
• All code and models were released publicly on GitHub, enabling reproducibility and further research in open-source wireless localization.