Low Power Wide Area Networks: System Analysis and Optimization

With the introduction of LoRa, a race started to become the ultimate Low Power Wide Area Network. Competitors, like Sigfox and Weightless, 3GPP and Ingenu, all of these competitors proposed new technologies trying to capture this narrow market segment as quickly as possible. However, which technology to choose? Which one performs best? A wide range of varieties to choose from and even more marketing efforts confusing customers.

This work tries to answer the question of what is a good low power wide area network. What criteria it should meet and how to optimize these networks. More specifically, the focus in this work lies on LoRaWAN, Sigfox and Weightless, all contending for spectrum below 1 GHz.

Six contributions have been achieved throughout this work. The first one is the understanding of the playfield. Many different LPWANs have been proposed. Their internal operation is critical to assess their performance in various scenarios. This high-level analysis allowed us to narrow down the long range problem and to see opportunities in low power wide area networks.

LoRaWAN is selected as an ideal candidate to study long range communication thanks to the spread spectrum technique and the open-source nature of the MAC protocol. This protocol is studied and dissected in detail, to implement as detailed as possible within the network simulator ns-3. This implementation allows further and more detailed studies on and with this protocol. This study is my second contribution.

Using the realistic network simulator, I did several experiments to assess the actual performance limits of LoRaWAN. These tests included network density, multiple gateways, acknowledgments and interference. These tests, my third and fourth contributions, show that LoRaWAN is vulnerable to interference and scalability is difficult.

My fifth contribution is the optimization of spreading factor distribution in these LoRaWAN cells. An analytical derivation showed a theoretical spreading factor distribution that improved the packet error rate performance with 50%. This analytical derivation assumed no interference. This assumption has been validated, and indeed, the channel is not highly occupied, but the dynamics of interference in the sub-GHz bands need to be taken into account. Based on this measurement, I proposed a machine learning approach, based on multi-armed bandits, to dynamically assign spreading factors.

However, this protocol is constrained in mobility and downlink messages. As a result, my sixth contribution is presented: a new medium access layer protocol that targets maximal scalability without relying on the allocation of spreading factors, while still being backward compatible.

In this work, I show that expecting high reliability of low power wide area networks is impossible due to the dynamics of the system, the random access of the protocol and the restricted time on air for gateways. Targeting millions of devices is extremely challenging.