5G Indoor Positioning for Manufacturing using Convolutional Neural Networks

Abstract

Current disrupting trends in manufacturing like mobile robot-based intralogistics and digital twins require precise indoor positioning. 5G-based positioning has the potential to be more cost-efficient than other positioning systems due to shared hardware for positioning and communication. But first it has to achieve high positioning accuracy and stability. No positioning technique has yet been established as standard in the industry in combination with 5G because traditional positioning approaches can deliver unstable results when multi-path fading, reflections, or missing line-of-sight effects occur. Identifying promising techniques is still an open research topic. It is established in the literature that deep learning-based approaches can mitigate these effects better than traditional ones by learning unique fingerprints that correspond to precise positions. However, fully-connected neural networks tend to overtrain on these fingerprints which again leads to unstable positioning due to a lack of generalization. This can be explained by the inability of fully connected neural networks to learn structural interrelationships in high dimensional data. This is because their input data is flattened, which means the structure is removed. Convolutional neural networks are being widely used with image data and can learn structural interrelationships. This paper presents a technique that uses convolutional neural networks in combination with a channel-smoothing approach to position mobile robots in the research factory ARENA 2036. The results presented in this paper indicate an indoor positioning accuracy similar to other industrial positioning systems, with improved stability compared to a fully-connected approach.