Integrating Path-Loss Modeling and Machine Learning for Localisation within LoRaWANs

ElSabaa, AlaaAllah (2026) Integrating Path-Loss Modeling and Machine Learning for Localisation within LoRaWANs. Doctoral thesis, Birmingham City University.

Preview

Text AlaaAllah Ahmed ElSabaa PhD Thesis_Final Version_Final Award July 2026.pdf - Accepted Version Download (9MB)

Abstract

Internet of Things (IoT) technologies and applications have surged across various environments; urban and remote environments have amplified the need for energy-efficient and accurate localisation techniques. Low-Power Wide-Area Networks (LPWANs) operate in-dependently for extended periods, making understanding the communication channel between end nodes and gateways crucial for network range, performance, and profitability. This research addresses the challenge of node localisation in LoRaWAN-based IoT networks by proposing a data-driven, GPS-independent framework utilising Long Short-Term Mem-ory (LSTM) neural networks, constrained to single-gateway, sector-level classification across three angular sectors.

The study begins with the design, deployment, and analysis of a real-world LoRaWAN network across a defined urban area in Birmingham, UK. Through extensive field experimentation, empirical path-loss models were derived to characterise the radio environment, capturing spatial signal attenuation under diverse conditions. The Birmingham-specific regression model yielded a path-loss exponent of n = 2.52 and intercept B = 130.85 dB, outperforming classical models (Okumura-Hata and COST-231) with a mean absolute error of 9.93 dB. These models formed the basis for generating synthetic Received Signal Strength (RSS) values, augmenting the original dataset and addressing the imbalance in sector-wise data distribution. Statistical validation confirmed KDE divergence values below 1.71 × 10−4 between real and synthetic distributions, ensuring the fidelity of synthesised samples for training.

Subsequently, a deep learning localisation model was developed using LSTM networks, trained solely on time-series RSS (Received Signal Strength) and SNR (Signal-to-Noise Ratio) values of sequentially received packets. The proposed LSTM framework achieved an overall classification accuracy of 72.1% (single split) and 68.8% ± 2.2% across five-fold cross-validation, with balanced accuracies of 0.72–0.80 per sector. The architecture outperformed a CNN–LSTM hybrid (65% accuracy) and static baseline models; including SVM (30.9%), kNN (27.0%), Logistic Regression (30.0%), and HMM (46.1%). Statistical validation via McNemar’s test confirmed the significance of improvement over the CNN–LSTM baseline (χ2 = 9.26, p = 0.0023). Noise-injection experiments demonstrated accuracy stability at approximately 58% under Gaussian RSSI perturbations of up to ±5 dB, confirming resilience to real-world signal fluctuations.

This study presents a unified methodology for coarse-grained, single-gateway localisation using only sequential RSSI and SNR measurements, without relying on dense infrastructure or GPS hardware, providing an extensible foundation for next-generation IoT and smart-city deployments.

Item Type:	Thesis (Doctoral)
Dates:	Date Event 27 July 2026 Accepted
Uncontrolled Keywords:	IoT, ML, LSTM, LoRaWAN, Pathloss modeling, WSN, LP-WANs.
Subjects:	CAH10 - engineering and technology > CAH10-01 - engineering > CAH10-01-08 - electrical and electronic engineering CAH11 - computing > CAH11-01 - computing > CAH11-01-01 - computer science CAH11 - computing > CAH11-01 - computing > CAH11-01-05 - artificial intelligence
Divisions:	Architecture, Built Environment, Computing and Engineering > Engineering Doctoral Research College > Doctoral Theses Collection
Depositing User:	Louise Muldowney
Date Deposited:	19 Aug 2026 08:40
Last Modified:	19 Aug 2026 08:40
URI:	https://www.open-access.bcu.ac.uk/id/eprint/17172

Actions (login required)

View Item