ML4POS

The ability to locate assets is a key prerequisite for many logistics and Industrie 4.0 applications. Fraunhofer IIS draws on a deep well of object positioning expertise and experience. Many applications rely on radiolocating solutions that generally measure the transit time of mobile objects’ radio signals. The accuracy of a positioning system depends largely on the extent to which fusion and system parameters are fine-tuned for different and sometimes changing environments.

Fraunhofer IIS is striving to develop more robust positioning methods with this in mind. Its researchers aim to tap the powers of machine learning to teach the system to calculate positions automatically based on empirical data sourced from reference measurements taken by a positioning system. Providing the learning material for the ML4POS positioning system, this raw data serves to train deep neural networks that can replace some or perhaps even all conventional computational positioning techniques.

Radio signals’ multipath propagation is big problem for positioning systems because this frequently occurring phenomenon can skew the result. A convolutional neural network (CNN) learns about the given environment’s characteristics and how it causes radio waves to propagate via multiple paths. This sort of training requires what are called ground truth positions – that is, positions that have been determined by reliable methods with conventional positioning equipment such as a temporarily installed or mobile laser measuring system. Once this data has been uploaded to calibrate the system, it no needs longer to reference these ground truth positions to determine locations. The system does have to be adapted specifically to the target environment – this hybrid combination of machine learning and conventional positioning techniques takes care of that. This way, it can determine an assets’ position precisely even in environments with a lot of multipath propagation. And that relegates the painstaking effort of fine-tuning a radiolocating system by hand to the dustbin of history.

Deep learning works in the real world – it can actually serve to compute positions. These two diagrams show a multipath tunnel in our L.I.N.K. test and application center. This setup models a logistics environment for measurement purposes. The positioning tag roams the shop floor in a rectangular pattern, crossing the tunnel several times. The researchers deliberately positioned the walls so that towards the end of the tunnel few antennas have line-of-sight contact with the radio transmitter.
 

This technology is mature enough to be deployed for various applications. One example is virtual-reality applications that require reliable means of locating assets, particularly for mobile systems. Networks with long short-term memory (LSTM) have also been put to use in projects. They served to support and replace conventional positioning methods such as error-mitigating Kalman filtering.

Fraunhofer IIS offers very good infrastructure for integrating machine learning into object positioning systems. Experts and researchers in both fields collaborate to this end. Our specialized team can work with you to develop a solution that fits your needs.