Use of principal component analysis for data reduction for training neural networks

Author

Glen Gray Paschal

Date of Award

12-1996

Degree Type

Thesis

Degree Name

Master of Science

Major

Nuclear Engineering

Major Professor

Robert E. Uhrig

Abstract

The use of neural networks has been effective in identifying relationships between plant data and a desired output. As the amount of data presented to a neural network increases the difficulty in training the networks increases. Principal component analysis is a technique available to reduce the size of data and thus may help the use of neural networks.

Principal component analysis is a method of systematically reducing a large number of dependent and independent variables to a smaller, conceptually more coherent set of variables. This smaller set of uncorrelated variables still retains most of the information of the original data set. This lends itself well as a preprocessing tool for neural networks where large data sets are involved. The smaller data set can be used to train neural networks quicker due to its reduced size. Uncorrelated variables provide the networks with simpler relationships and reduced noise which further improves performance.

An application of this technique is presented using data from TVA's Kingston Fossil Plant. A comparison is made between the use of raw data and data preprocessed with use of principal component analysis to train a neural network. The principal component analysis proves to be a very useful tool in reducing the number of inputs to a neural network and improving its performance.

Recommended Citation

Paschal, Glen Gray, "Use of principal component analysis for data reduction for training neural networks. " Master's Thesis, University of Tennessee, 1996.
https://trace.tennessee.edu/utk_gradthes/10930