Predicting the impact of IoT devices on Radio Frequency Noise in South African environments using machine learning

Abstract

The increasing number of Radio Frequency (RF) devices proliferating in our environment inspired this research. Forecasts suggest that the Internet of Things (IoT) industry will populate society with an ever-increasing number of RF-operated applications such as smart homes, remote surveillance, intelligent vehicles, tracking, smart grid, remote metering, innovative health, and smart cities. This research aimed to investigate whether the presence of IoT-like radiations could influence the levels of ambient RF noise. With that in mind, the study required surveying by collecting real-world ambient radio noise data in target urban, suburban, and industrial environments over the Industrial Scientific Medical (ISM) bands, such as 433 MHz, 868 MHz, and 2.4 GHz. At the time of this survey, the IoT industry was still in the infancy stage in South Africa. Therefore, the exercise necessitated two series of survey campaigns. The first set of measurements had as its primary mission to assess the existing levels of ambient RF noise in selected candidate sites considering their early IoT development phase. Subsequently, this phase helped to verify and validate that the research deployed appropriate equipment, hardware, and software for collecting environmental radio noise data. This study designed a Radio Noise Surveying System (RNSS) using softwaredefined radio techniques with the Universal Software Radio Peripheral (USRP) and the GNU Radio platforms as part of the equipment. The simulation and test results agreed that the RNSS performed adequately, and that all system was suitable for radio noise surveying. This first phase also helped to confirm that the post-processing methods of importing and transforming raw data into clean data and the applied calibration techniques were correct. Exploratory data analysis with these baseline measurements revealed ambient radio noise data characteristics, for example, their extensive data volume. One of the remarkable findings was that, out of six candidate sites, the Steeve Biko Campus showed the highest levels of ambient radio noise compared to the rest, irrespective of frequency bands. The second measurement trial, over five candidate sites, envisaged assessing the direct contribution of IoT operations. Therefore, the exercise necessitated environments populated with IoT devices. Hence, this research created IoT radio noise generators (ING) to produce intentional RF emissions in the ISM bands to imitate the presence of IoT devices in selected environments. The research underwent a complete product design cycle covering conceptualisation, component selection, schematic and PCB design, board assembly, firmware development, and functional testing. Survey campaigns deployed forty-five ING units, of which fifteen covered each of the three frequencies of interest. Data analysis exploited the elements of descriptive statistics to understand the characteristic nature of data emanating from ambient RF measurements. Concerning the central question of this research, exploratory results revealed that 80% of analysed cases show an increase in environmental radio noise levels, with a conclusion that ambient radio noise levels were directly proportional to radio activities in given environments. This finding forewarns that the proliferating presence of IoT products will directly influence ambient radio noise levels. Finally, this study applied Machine Learning techniques to develop linear regression models to predict the levels of ambient RF noise. The research developed a computer application as Radio Noise Predictor (RNP) software installable in Windows PC. Based on models produced in this research, the RNP application allows interested users to estimate the radio noise levels from a selected environment and frequency.