Indoor positioning with AI/ML using simulated 5G data

Sammanfattning

With future technologies and Industry 4.0, the need for robust and accurate positioning to optimize productivity and industrial operations increases. The previous positioning methods using triangulation or angle-of-arrival are not good enough, especially not for clutter-dense factories with poor line-of-sight conditions. For this thesis project, a convolutional neural network model was trained to better estimate time-of-arrival and classify line-of-sight for a clutter-sparse simulated factory. For the clutter-dense factory, a residual network fingerprinting model was trained to map channel impulse responses to a position. Both the clutter-sparse and clutter-dense factories were modified by moving and rotating machines as well as adding robots in order to investigate the robustness to changes in the factory environment. The factories simulated production areas, assembly areas, beam structures and robots using the Ericsson state-of-the-art version of Nvidia Omniverse. The fingerprinting model was also used in a real scenario from Mobile World Congress 2024, where a robot drove two routes in an open office environment. The results show that both neural networks gave a positioning error below 1 m. For the modified scenarios, the convolutional neural network was more robust than the residual network fingerprinting model. The modifications mainly impacted positioning errors locally where changes were introduced.