AI-based Localization


5G and AI – Data-driven Localization to Enable Telecommunication Benefits of 5G and beyond

Tobias Feigl1,2
1Fraunhofer Institute of Integrated Circuits (IIS), Nuremberg, Germany
2Friedrich-Alexander-University Erlangen-Nuremberg (FAU), Germany

The problem of localization has always been discussed in detail in the field of wireless communication, also for future 5G networks. In the past few years, deep learning (DL) technology has developed rapidly. The high-dimensional modelling ability of DL makes it possible to solve localization problems in suboptimal scenarios that are difficult to handle with classic models. As a result, the researchers at Fraunhofer IIS have carried out extensive research on wireless localization based on classic model-driven methods and DL over the past decade.

Figure 1: DL-supported localization in a typical industrial scenario with 5G signals to stabilize communication and connectivity in harsh environments.

The communication industry is rapidly evolving towards wireless 5G technologies and beyond to meet the ever-increasing demands for higher data rates, improved quality of service and more precise localization. New applications require wireless connectivity with enormously increased data rates and significantly reduced latency, which are optimized based on the position of the end user. These requirements represent new challenges that can no longer be efficiently addressed by conventional approaches. Artificial intelligence (AI) is considered one of the most promising solutions for improving the performance and robustness of 5G and beyond. This is supported by the enormous amount of data in 5G networks and the availability of powerful data processing structures.

However, wireless localization systems, including 5G networks, are often faced with major challenges. This is particularly evident in industry-relevant applications such as asset tracking or in the tracking of industrial trucks, see Figure 1. Difficult indoor environments, e.g., production and machine halls, logistics centres, parking garages and construction sites, make localization even more challenging as metallic objects often cause signal reflections. The resulting multi-path situation in such environments often leads to errors when using classic methods, which limits these systems, see Figure 2.The researchers at Fraunhofer IIS [L1] are therefore researching AI-supported parameter estimators to optimize radio localization and to enable more precise position estimation, see Figure 3. As a result, a large number of research projects on AI-based communication technologies have developed at Fraunhofer IIS [L2] in the last ten years, which above all promise robust and precise localization as well as higher data rates through signal compression methods with affordable computing and implementation effort. To enable these improvements, different families of learning methods are researched at Fraunhofer. 

Figure 2 (left): Radio localization system using a classic Bayesian filter: under massive multipath conditions (up to the loss of visual contact between transmitter and receiver), a classic system for position estimation is subject to large errors or jumps in the position or failure of position calculation, since the time of arrival of the transmission signal, that is no longer on the direct path cannot be correctly determined.

Figure 3 (right): Results of a position determination using deep learning: the approach can use the multipath information in the antenna signals even in the case of massive shadowing and multipaths to determine the true position reliably and precisely [1,2]. In addition, recurrent neural networks learn the motion of the object, which results in even more precise motion models and hence, in highly accurate positions [3].

One focus is the deep coupling of supervised model- and data-driven methods in hybrid filters for tracking movement patterns and signal anomalies as well as their temporal, semantic, and spatial relationships to enable robust and highly precise localization in many difficult scenarios. That way, it was possible to research cutting-edge technologies that clearly surpass the state of the art. Processes are currently being researched to use simulators to facilitate data acquisition and to bring the DL processes to the scene of operation more cheaply, more specifically, and more quickly [1-3]. On the other hand, unsupervised learning is used to collect and identify unknown information directly on site and to categorize it interactively in order to broaden the knowledge horizon of the learning process.Another focus is the joint optimization of beamforming, power control, and interference coordination in a 5G radio network to improve communication performance for end users. A typical technique is deep reinforcement learning (RL) using deep Q networks and deep deterministic policy gradient methods. The use of supervised, unsupervised, and RL methods in combination with classic Bayesian estimation methods enables researchers to describe non-linearized relationships in radio channels to enable precise and robust localization and consequently higher data rates and improved quality of service, see Figure 3.


Links 

[L1] https://www.iis.fraunhofer.de/de/ff/lv/lok.html
[L2] https://www.iis.fraunhofer.de/de/ff/lv/dataanalytics.html

References

[1] Kram S., Stahlke M., Feigl T., Seitz J., & Thielecke J. (2019). UWB Channel Impulse Responses for Positioning in Complex Environments: A Detailed Feature Analysis. In: Sensors (pp. 1-27).
[2] Niitsoo A., Edelhäußer T., Hadaschik N., Eberlein E., & Mutschler C. (2019). A Deep Learning Approach to Position Estimation from Channel Impulse Responses. In Sensors (pp. 1-23).
[3] Feigl T., Nowak T., Philippsen M., Edelhäußer T., & Mutschler C. (2018). Recurrent Neural Networks on Drifting Time-of-Flight Measurements. In Proc. Intl. Conf. on Indoor Positioning and Indoor Navigation (pp. 101- 109).

Please contact

Dr.-Ing. Tobias Feigl
Member of the Precise Positioning and Analytics department at Fraunhofer IIS in Nuremberg, Germany.