PhD Student: Ben Halstead. Pollution from wood burners has serious health implications for residents of rural towns, even in developed countries. Monitoring the level of airborne particulate matter, PM2.5, in these areas often requires making inferences about missing or corrupted readings. Air Quality inference in these cases poses two key challenges. Firstly, air quality displays non-linear spatio-temporal relationships dependent on many factors. Secondly, these factors can evolve over time, changing the distribution of data. For example, changing wind directions can have a large impact on which neighboring sensors are most relevant to inference. Methods incorporating environmental factors to capture these changes, e.g. weather, traffic and points of interest, have found success in urban environments. However, many locations only have access to few if any of these features, thus, inference methods must employ alternate approaches to detect and adapt to changes. We propose a data stream based system, called AirStream, to infer missing PM2.5 levels that is able to detect and adapt to changes in unknown features. We deployed our approach on two air quality studies in New Zealand rural towns, and also tested it on a Beijing benchmark data set. We found gains in inference performance comparing AirStream against seven baseline methods. We further investigate the relationship between the changes we detected and changes in underlying weather conditions. We discovered a strong predictive link between the state of our system and current meteorological conditions. This project is part of Royal Society Marsden Fast-Start. Supervisors: Assoc Prof Yun Sing Koh, Dr Pat Riddle, Prof Mykola Pechenizkiy (TU/e Eindohoven), Prof Albert Bifet (Waikato).
Keywords: Air Pollution, Data Stream Mining, Continual Learning
In partnership with Dr Guy Coulson and Gustavo Olivares | NIWA