Blind Calibration of Air Quality Wireless Sensor Networks Using Deep Neural Networks

TL;DR

This paper explores deep learning methods for blind calibration of wireless air quality sensor networks, effectively reducing calibration errors and enabling remote, cost-effective calibration without reference data.

Contribution

It introduces new deep learning models tailored for calibrating mixed static and mobile sensors using weather data, extending previous models and demonstrating significant error reduction.

Findings

Models reduce calibration error by an order of magnitude.

Deep learning is effective for remote sensor calibration.

Weather data improves calibration accuracy.

Abstract

Temporal drift of low-cost sensors is crucial for the applicability of wireless sensor networks (WSN) to measure highly local phenomenon such as air quality. The emergence of wireless sensor networks in locations without available reference data makes calibrating such networks without the aid of true values a key area of research. While deep learning (DL) has proved successful on numerous other tasks, it is under-researched in the context of blind WSN calibration, particularly in scenarios with networks that mix static and mobile sensors. In this paper we investigate the use of DL architectures for such scenarios, including the effects of weather in both drifting and sensor measurement. New models are proposed and compared against a baseline, based on a previous proposed model and extended to include mobile sensors and weather data. Also, a procedure for generating simulated air quality data is presented, including the emission, dispersion and measurement of the two most common particulate matter pollutants: PM 2.5 and PM 10 . Results show that our models reduce the calibration error with an order of magnitude compared to the baseline, showing that DL is a suitable method for WSN calibration and that these networks can be remotely calibrated with minimal cost for the deployer.