Predicting LoRaWAN Behavior: How Machine Learning Can Help

Abstract

Large scale deployments of Internet of Things (IoT) networks are becoming reality. From a technology perspective, a lot of information related to device parameters, channel states, network and application data are stored in databases and can be used for an extensive analysis to improve the functionality of IoT systems in terms of network performance and user services. LoRaWAN (Long Range Wide Area Network) is one of the emerging IoT technologies, with a simple protocol based on LoRa modulation. In this work, we discuss how machine learning approaches can be used to improve network performance (and if and how they can help). To this aim, we describe a methodology to process LoRaWAN packets and apply a machine learning pipeline to: (i) perform device profiling, and (ii) predict the inter-arrival of IoT packets. This latter analysis is very related to the channel and network usage and can be leveraged in the future for system performance enhancements. Our analysis mainly focuses on the use of k-means, Long Short-Term Memory Neural Networks and Decision Trees. We test these approaches on a real large-scale LoRaWAN network where the overall captured traffic is stored in a proprietary database. Our study shows how profiling techniques enable a machine learning prediction algorithm even when training is not possible because of high error rates perceived by some devices. In this challenging case, the prediction of the inter-arrival time of packets has an error of about 3.5% for 77% of real sequence cases.

A Comprehensive Survey of Data-Driven Solutions for LoRaWAN: Challenges & Future Directions

“…The proposed approach not only reduces collisions (≈56%) and delays but also enhances network throughput. The study conducted in [46] has also shared a similar objective of mitigating LoRaWAN challenges using ML models. Herein, Cuomo et al have investigated how ML models can be utilized to enhance LoRaWAN performance.…”

A Survey on LoRa for Smart Agriculture: Current Trends and Future Perspectives

“…This implies that devices may not occupy the ISM band for more than 36 seconds per hour, forbidding the transmission of new packets when this limit is attained [123]. Machine learning can be applied to model and analyze technical problems, improving the scalability of LoRa networks and predicting network congestion [124]. Further developments could include enhanced ADR mechanisms, optimization of GW locations, and interference cancellation techniques [125].…”

LoRaWAN Meets ML: A Survey on Enhancing Performance with Machine Learning

“…Their proposed method is simple, scalable, and efficient for reducing collision probabilities in LoRaWAN. Similarly, ED profiling using K-means was utilized in [ 135 ] to predict the behavior of LoRaWAN traffic. The authors grouped the EDs using the same SF and packet size and trained DT and Long Short-Term Memory (LSTM) using unsupervised traffic pattern classification methods.…”