Assessing and Improving the Measurement Quality of Low-Cost Air Quality Sensors Using Machine Learning Techniques

Air quality monitoring is an ever relevant field of study, with a large proportion of the global population living in areas with unsafe levels of pollution.Increased monitoring is required to determine the health impacts on a local population, but traditional equipment is cost-prohibitive, requiring investment to operate and maintain.This makes traditional monitors impractical to deploy in dense networks, rural areas or developing countries.Low-cost alternatives are used as a way to supplement traditional measurement methods, but have issues in their measurement that need to be addressed prior to deployment.This thesis aimed to recommend the best methodologies for an end-user of low-cost air quality monitors, with caveats.This was done by first comparing the state-of-the-art within the literature to identify trends and gaps within published studies.A suite of methodologies are then tested for both correcting off-the-shelf sensor system measurements and calibrating bespoke systems, using a range of locations and systems.Finally, a network re-calibration methodology was developed to reinforce spatially disperse low-cost sensor system measurements against baseline drift, reducing downtime and re-location costs.It is difficult to make conclusions based on the literature as the performance metrics used can vary with many variables other than the performance of a technique (length of study, average concentrations during training and testing, distribution of measured variables, etc).Variance in results between different sensor systems and co-location sites is often more than methodologies when both correcting and calibrating low-cost sensor system measurements.The network recalibration method reduced the Mean Absolute Error of a large proportion of systems tested and reduced the Mean Bias Error of most systems, but requires more work before being used with real-time measurements in a network.Though no single technique proved most effective throughout, this thesis serves as a useful guide to end-users of low-cost air quality sensor systems and recommends a range of techniques to test prior to deployment.