Neural network recognition and classification of aerosol particle distributions measured with a two-spot laser velocimeter

Abstract

This paper describes the use of a neural computational network model for pattern recognition and classification of aerodynamic particle size distributions associated with a number of environmental, bacterial, and artificial aerosols. The aerodynamic particle size distributions are measured in real time with high resolution using a two-spot He-Ne laser velocimeter. The technique employed here for the recognition and classification of aerosols of unknown origin is based on a three-layered neural network that has been trained on a training set consisting of 75 particle size distributions obtained from three distinct types of aerosols. The training of the neural network was accomplished with the back-propagation learning algorithm. The effects of the number of processing units in the hidden layer and the level of noise corrupting the training set, the test set, and the connection weights on the learning rate and classification efficiency of the neural network are studied. The ability of the trained network to generalize from the finite number of size distributions in the training set to unknown size distributions obtained from uncertain and unfamiliar environments is investigated. The approach offers the opportunity of recognizing, classifying, and characterizing aerosol particles in real time according to their aerodynamic particle size spectrum and its high recognition accuracy shows considerable promise for applications to rapid real-time air monitoring in the areas of occupational health and air pollution standards.