Publikasjonsdetaljer
Del av: Intelligent Information and Database Systems. 15th Asian Conference, ACIIDS 2023, Phuket, Thailand, July 24–26, 2023, Proceedings, Part I. (Springer, 2023)
ISBN: 9789819958337
Doi: doi.org/10.1007/978-981-99-5834-4_20
Arkiv: hdl.handle.net/10037/32928
Sammendrag:
<p>Monitoring air pollution is a critical step towards improving public health, particularly when it comes to identifying the primary air pollutants that can have an impact on human health. Among these pollutants, particulate matter (PM) with a diameter of up to 2.5 μm (or PM2.5) is of particular concern, making it important to continuously and accurately monitor pollution related to PM. The emergence of mobile low-cost PM sensors has made it possible to monitor PM levels continuously in a greater number of locations. However, the accuracy of mobile low-cost PM sensors is often questionable as it depends on geographical factors such as local atmospheric conditions. <p>This paper presents new calibration methods for mobile low-cost PM sensors that can correct inaccurate measurements from the sensors in real-time. Our new methods leverage Neural Architecture Search (NAS) to improve the accuracy and efficiency of calibration models for mobile low-cost PM sensors. The experimental evaluation shows that the new methods reduce accuracy error by more than 26% compared with the state-of-the-art methods. Moreover, the new methods are lightweight, taking less than 2.5 ms to correct each PM measurement on Intel Neural Compute Stick 2, an AI-accelerator for edge devices deployed in air pollution monitoring platforms.