A Novel Approach for Outdoor Signal Strength Prediction Using Machine Learning

• Published in: IEEE Vehicular Technology Conference pp. 1 - 6
• Main Authors: Jassim, Mostafa, Kurner, Thomas
• Format: Conference Proceeding
• Language: English
• Published: IEEE 07.10.2024
• Subjects: Accuracy; Adaptation models; Buildings; Channel Modelling; Data models; Machine Learning; Ray tracing; Regression; Signal Strength; Training; Urban areas; Vehicular and wireless technologies; Wireless networks
• ISSN: 2577-2465
• DOI: 10.1109/VTC2024-Fall63153.2024.10757720

This paper presents a novel approach to predict radio signal strength in wireless networks combining geometry extraction techniques and Machine Learning (ML). Traditional methods whether empirical or deterministic have their own ad-vantages and disadvantages, with empirical methods often falling short with accuracy, while deterministic models can be resource intensive, especially with large scenarios. Our approach proceeds towards a hybrid implementation of environment description, exploring the complex relationships between the different environmental factors and signal strength, harnessing the power of ML. We employ LightGBM, a powerful tree-based learning algorithm with a gradient boosting framework, to predict signal strength based on a set of features. These features include transmitter height, location coordinates, distance, buildings and terrain height. In our study, labels represent the actual signal strength in different scenarios, we utilized a unique approach of training our model using two distinct sources of labels: Ray traced and measurement data. The ray traced labels were extracted from a large variety of urban cities, consisting of high building density and diverse infrastructures. The results demonstrate that our hybrid approach was able to provide promising results with a standard deviation of approximately 1.2 dB in comparison with ray tracing, while utilizing a fraction of the resources. Additionally, the model was able to adapt to the measurement data, offering high accuracy results with a standard deviation of approximately 2.6 dB.