Aleksandra Malkova

[ABSTRACT]

The positioning of connected objects has become a major enabling feature for a variety of applications and services in the Internet of Things (IoT) domain. To avoid prohibitive power consumption and hardware complexity at the IoT nodes to be localized, one alternative to the conventional Global Positioning System (GPS) consists in making opportunistic use of location-dependent radio metrics, out of the sensor data packets that may be transmitted from these end-devices to their serving gateways.

In this thesis, we focus on the Received Signal Strength Indicator (RSSI) radio metrics, which are available by default in Long Range (LoRa) low data rate transmis- sions, while targeting fingerprinting-based localization for the final application. So as to accurately position LoRa nodes at the city scale through fingerprinting, one first needs to build an accurate prior RSSI map of the radio environment (offline), typically based on a few real field measurements. This prior radio map is subsequently used as a reference (for comparison) during the online localization phase. One main goal of this work is hence to obtain an exhaustive and accurate prior map, given sparse and non-uniformly distributed RSSI measurements, while applying advanced machine learning approaches.

Firstly, we discuss the main general challenges of the map reconstruction itself, in light of our specific fingerprinting-based localization context. In particular, by means of theoretical bounds characterizing the best RSSI-based positioning accuracy in case of both fingerprinting (i.e., considering a prior radio map) or parametric positioning (i.e., considering a link-wise range-dependent power path loss model), we illustrate and discuss the impact of RSSI dynamics as a function of space with respect to the final positioning performance. Further, we also describe and analyze the experimental datasets used in our study, including a dataset related to the Grenoble city area, which is still under extension, while devising metrological and pre-processing aspects. These experimental data have been exploited to feed both the underlying radio models with realistic parameters, as well as the learning algorithms used for map reconstruction with realistic inputs.

Secondly, we briefly recall the main ideas behind supervised and semi-supervised learning approaches, their main working assumptions, as well as their classical models. In particular, we present a Neural Architecture Search (NAS) approach, which is a recent technique allowing to automatically find a Neural Network (NN) model with an optimized architecture suited to a given problem.

We then cover a first application of this NAS to the RSSI map reconstruction problem, in combination with data augmentation techniques, where only a limited amount of labeled input data is available for learning.

Finally, taking also advantage from having access to side (meta-)information about the local area (e.g., city layout, terrain elevation, gateway location, etc...), we solve out the generalization problem for each city (typically, over gateways). The NAS algorithm is then applied again to find the NN model with the best architecture for each of the supposed settings, depending on the amount of such prior side information. On this occasion, we show for instance that using additional views improves the final accuracy of the RSSI map reconstruction, especially in sub-areas close to the gateways where larger variations of the average received signal power are usually observed (ultimately, with a prominent beneficial impact onto positioning performance accordingly).