Author

Evesque

Date of Award

4-2003

Degree Name

Master of Science

Department

Computer Science

First Advisor

Dr. Robert Trenary

Second Advisor

Dr. Donna Kaminski

Third Advisor

Dr. Elise deDoncker

Access Setting

Masters Thesis-Open Access

Abstract

Electronic noses are used to identify and characterize unknown odors in industry. Chemometrics and neural network algorithms are used as pattern recognition systems for these devices. Experimentation with Kohonen clustering as the pattern recognition system for electronic noses was not noted prior to 1997. [BEG] This thesis investigated the use of a Kohonen neural network algorithm as a clustering algorithm for electronic nose data using the chemometrics algorithms built into the electronic nose as a performance standard. A secondary aim was to improve the clustering and identification capabilities of the Kohonen network.

The unsupervised Kohonen network was not able to cluster the electronic nose data. Duplicating the data pre-processing performed by the electronic nose, fine-tuning the visualization of clustering data, and varying the learning and weight update rates offered minor improvements but did not allow for accurate identification of samples. Significant improvements were obtained when the network was changed to a semi-supervised network by incorporating average sensor values of the known odorant samples into the network weights. This improved the clustering effectiveness of the network to 60-70% compared to a 100% effectiveness of the chemometrics system built into the nose. Several avenues were identified for further study to improve the effectiveness of the neural network for use with the electronic nose.

Share

COinS