Synthetic Data Generation for improving Deep Learning-based 5G Indoor Positioning

Hannes Vietz, Manuel Hirth, Sebastian Baum, Michael Weyrich

Abstract: Gathering sufficient labeled training data to effectively train a high-performing deep learning model can be particularly challenging in the realm of industrial automation. Depending on the data type, this may require expensive interruptions to production processes or similar disruptions on the factory floor for data collection. It is often uncertain which data types are crucial for enhancing the performance of the trained model. For vision models, factors such as specific viewing angles or lighting conditions may be important, while for models utilizing radio signals, unique reflections generated by moving metal surfaces could be significant. Moreover, data labeling is expensive as it is primarily conducted manually by human workers. This paper demonstrates how to automatically generate relevant, labeled synthetic training data to boost a neural network's accuracy for deep learning-based 5G indoor positioning tasks. We reveal that employing this generated synthetic data to train a convolutional neural network can improve its median positioning accuracy by a notable 25%.