Time Reduction for Collecting Fingerprint Data in Indoor Positioning Systems with Generated Synthetic Data by Ensemble Models and GANs

Abstract

Nowadays, the demand for IPS is growing due to the increasing need for accurate indoor location services in applications. The IPS fingerprint techniques are widely popular because they offer high accuracy. However, the process of collecting fingerprint data is labor-intensive and time-consuming. This study aims to alleviate the burden of data collection by generating synthetic data using Machine Learning (ML) and Generative Adversarial Networks (GANs). To create ML synthetic data, we used a dataset containing RSSI values and coordinates. Various regression models were trained using Randomized Search for hyperparameter tuning. The best models were then combined into an ensemble method using Voting Regressor. This ensemble model was used to predict RSSI values for new, synthetic coordinates generated around each reference point, forming the synthetic dataset. We combined synthetic data with actual data from the IPS fingerprint RSSI collecting from the mobile application to create three new datasets with varying ratios of actual to synthetic data from 90:10 to 10:90. These combined datasets were used to train models including Random Forest, Decision Tree, Linear Regression, Gradient Boosting, and K Nearest Neighbors. Our results indicate that models trained on combined datasets significantly reduce the mean distance error (MDE) compared to those trained solely on actual data. This improved performance, however, comes with trade-offs in terms of slightly increased training time, prediction time, and memory usage during both training and prediction phases.