Details

Abstrakt in Englisch

With the rising interest in integrating sensing functionalities into mobile wireless communication networks, and given the well-proven value of satellite-based positioning systems, this dissertation aims at exploring the potential of machine learning to contribute to the signal processing toolbox available to engineers for designing algorithms that enable precise and reliable positioning services indoors. The current techniques available for indoor positioning have come a long way, but an approach that is accurate to sub-meter levels in this case still remains elusive. The main reasons for this issue are: the hostile propagation characteristics of indoor environments, the difficulty in treating mathematical models that related the source position with the signal received at sensing units (SUs) and the mismatch of such models due to hardware impairments, the consequent computational complexity of proposed algorithms, and the absence of a straightforward parameter fusion mechanism.

To address these challenges, the work carried out in this dissertation proposes a neural network architecture that takes received signal strength (RSS) measurements from spatially distributed SUs as input and calculates the signal source coordinates in a supervised learning setting. RSS-based positioning is in focus due to its ability to deliver blind source localization, i.e., there is no positioning specific pilot signaling required, and the simple implementation complexity of the SUs needed to estimate the RSS.

The data-driven paradigm of machine learning employed is supervised learning, which turns the positioning algorithm also blind to propagation related variables, since these are embedded in the training data. Hence, within this dissertation, a positioning approach based on a fully connected deep neural network (DNN) is investigated and compared against classical model-based positioning algorithms both in numerical and experimental settings. The results suggest that the DNN scheme is able to provide source location estimates with comparable performance to the maximum likelihood approach, while incurring significantly less online computational costs. Moreover, the ability of the DNN to scale and address the simultaneous multi transmitter positioning problem is also studied.

In an attempt to push the boundaries of positioning algorithms, the usage of a new source of location information – the position information correlation matrix (PICM) – is also proposed. This matrix is used as input to a convolutional neural network (CNN) that, in turn, outputs the source coordinates. This concept can be exploited in scenarios where a network of synchronized SUs is available. In contrast to the blind positioning possible through the pure RSS-DNN approach, the characteristics of the transmit signal can be optimized for enhancing the position representation ability of the PICM. Thus, guidelines for selecting the system parameters are examined.

The performance of the PICM-CNN technique is investigated employing a ray tracing simulation tool, which enables a fair comparison against other positioning approaches that rely on different sources of position information, such as time difference of arrival (TDoA) and RSS, within a common simulation framework. The outcomes of the numerical study demonstrate that the PICM-CNN approach can deliver an accurate indoor positioning service advancing the current state of art, showcasing its potential as a promising alternative for the future integration of wireless communication and sensing.