Rapid Detection and Classification of Aerosol Events Based on Changes in ParticleSize Distribution

摘要

A methodology is presented that allows aerosol particle size data to be used toindicate when an abnormal aerosol event may be occurring. Such data can be collected from an array of commercially available particle counter-sizers. The methodology employs two main elements: a detection element that recognizes when an aerosol concentration spike event is occurring; and a classification element that classifies aerosol events as normal (e.g. dust kicked up by wind gust or generated by normal vehicular activity) or abnormal (e.g. mistakenly released non-indigenous aerosol material). The detection element is based on observation of statistically significant rises in the aerosol concentration level, during an appropriate time interval. The classification element uses a new three dimensional feature space that highlights relevant differences in the aerosol particle size distribution function. The classifier adapts to the local environment by learning the region of the feature space that is occupied by normal aerosol events. Observations which then fall significantly outside this region are classified as abnormal. The methodology was developed using a set of atmospheric aerosol data containing over 600,000 observed aerosol particle size distributions, under both normal conditions, and with intentionally introduced abnormal aerosol. An implementation of the methodology is described. Many abnormal aerosol events in the data set are demonstrated to be distinguishable from normally occurring aerosol events.