Efficient Air Quality Index Prediction on Resource-Constrained Devices using TinyML: Design, Implementation, and Evaluation

These days, mobile/handheld air quality monitoring devices primarily sense pollutants and send them to the cloud without performing intelligent processing, such as soft calibration, sensor fault detection, or any onboard predictions. This paper explores the feasibility of harnessing Tiny Machine Learning (TinyML) to develop a compact system for inferring environmental pollution levels at a device’s location. Our primary challenge lies in fitting a trained machine-learning model within the device’s limited memory without compromising its accuracy or inference speed. Building upon the correlation between the Air Quality Index (AQI) and various meteorological factors, we propose a TinyML-based framework that predicts AQI using temperature, humidity, internet-sourced meteorological data, and temporal information collected through a web crawler. Our contributions encompass sensor activation, system design on a Raspberry Pi Pico W board, and hyperparameter tuning of machine learning models, culminating in an efficient XGBoost implementation within the 2MB memory constraint, achieving 75.2% accuracy with a 1615μs latency. Moreover, an accuracy of 60% was observed when a real-time validation was carried out. As per our knowledge, this system is the first-of-its-kind device that tries to embed some intelligence into a mobile air quality sensing device.