California Scientific

California Scientific
1000 SW Powell Ct
Oak Grove, MO 64075
California Scientific  *  BrainMaker Neural Network Software  *  Predict Forecast Classify Stocks Bonds Markets Commodities Diagnose Medical

BrainMaker Neural Network Software

Using a Neural Network to Measure Air Quality

Researchers Eugene Yee and Jim Ho at the Defense Research Establishment Suffield, Chemical & Biological Defense Section, in Alberta, Canada have trained a neural network to recognize, classify and characterize aerosols of unknown origin with a high degree of accuracy. Their results hold considerable promise for applications to rapid real-time air monitoring in the areas of occupational health and air pollution standards.

Their research applied a neural network to the recognition and classification of environmental, bacterial, and artificial aerosols on the basis of the aerodynamic particle size distribution. Because of their variability, aerosols are difficult to recognize using conventional pattern recognition techniques. However, the health effects posed by airborne industrial, bacterial, and viral particles depend critically on the ability to recognize, characterize and classify these particles on the basis of their particle size distribution functions.

The input data was constructed from aerodynamic particle size distribution functions (PSDF) obtained from 11 different aerosol populations. The PSDF's were measured with an aerodynamic particle size which determines the aerodynamic diameter of individual aerosol particles by measuring the transit time of the particles between two spots generated by a laser velocimeter that employs a polarized laser light source. Size distributions were classified into 11 categories depending on the source of the aerosol particles generating the distribution.

It was found that a recognition rate of 100% was obtained for the training set using neural networks with three or more hidden units and that there was a smaller number of passes through the training data with an increase in the number of hidden units in the network. There was virtually no increase in the learning times of the networks with more than 10 hidden neurons. In addition, the performance of the networks did not deteriorate when the number of hidden units was increased beyond 10.

Experiments were also conducted to study the performance characteristics of the neural network as a function of the quality of data used for the training set and the test set and of the inclusion of random noise in the connection strengths of the trained network. Results showed that the neural network was more suited than conventional methods for classification of signals from systems where one is confronted with ignorance of the statistical characteristics of the noise corrupting the signals.