Estimation of Air Quality Parameters using Lightweight Machine Learning on Low-cost Edge-IoT Architectures

Abstract

The vast increase in connected Internet of Things (IoT) devices have revolutionised how data are processed. This fact, coupled with the current trend from cloud to edge computing paradigms, has resulted in the need for efficient and reliable data processing near to data sources using resource-constrained devices. In this article, low-cost edge-IoT architectures are implemented to deploy lightweight Machine Learning (ML) models for air quality estimation, such as Polynomial Regression and Artificial Neural Networks (ANN). ML models are deployed in wireless centralised and distributed parallel architectures with common modules such as sensor fusion for luminosity, temperature, humidity, CO2, and other gases. The centralised architecture uses a Graphic Processing Unit (GPU) and the Message Queuing Telemetry Transport (MQTT) protocol, but low-performance processing devices and the Message Passing Interface (MPI) protocol are used in the distributed one. The training and testing of models are attained with appropriate datasets obtained from multiple peak, step, and transient test cases for each air quality parameter. The results for temperature forecasting, and similar ones for other parameters, supports that the distributed parallel architecture could achieve a slightly better estimation metrics and a better performance in power consumption compared to the centralised architecture despite using low-cost general purpose devices.