Indoor Positioning with AI/ML Using Simulated 5G Data

Jonathan Nilsson () and Philip Söderbom ()

Abstract

With the future technologies and Indusry 4.0, there is need for robust and ac- curate positioning. The previous solutions using triangulation or angle of arrival estimation is not suitable for a clutter-dense factories. This thesis project used the AI/ML models CNN and ResNet for positioning in a clutter sparse respective clutter dense scenario to investigate if these models are robust and give a better position accuracy then the legacy solutions.
The factories simulate production areas, assembly areas, beam structures and robots in the Ericsson state-of-the-art version of NVIDIA’s Omniverse. The chan- nel impulse response is simulated using a Ray-tracer tool.
When using the CNN in the clutter-sparse factory, the positioning error went from 4m using the legacy solution to below 1m in the baseline factory. After applying changes to the baseline, the position error was still below 2m.
For the ResNet in the clutter-dense baseline scenario, the position error went from 17m to around 1m. For the modified baseline, the position error went up to 5m, but also below 1m when smaller changes were done.
To improve the ResNet model, mixed training with the baseline scenario modi- fications with added AGVs and forklifts was used. When applying this, the position error went down to below half a meter for the legacy scenario, but almost stayed unchanged for the more modified baselines.
This shows a that a CNN solution gives a low and robust position accuracFy while a ResNet gives good position accuracy, but is more sensitive to changes in the environment.