Masters Theses

Ying Liu

Date of Award

12-1991

Degree Type

Thesis

Degree Name

Master of Science

Major

Nuclear Engineering

Major Professor

Belle R. Upadhyaya

Committee Members

R. E. Uhrig, L. Tsoukalas

Abstract

The on-line measurement of chemical mixture composition under different operating conditions is an important problem in many industries. Effective control of industrial processes based on the measurement of mixture composition will result in a reduction in the overall energy consumption. Raman and infrared spectroscopy are often used to estimate the composition of chemical streams. The inelastic scattering of photon energy to or from the energy levels near infrared (NIR) range is called the Raman effect. The Raman spectroscopy has several advantages over the infrared spectroscopy, especially for aqueous solutions. Multiple linear regression is one of the methods currently used in spectroscopic analysis. This thesis presents a new approach which utilizes a hybrid signal preprocessing and artificial neural networks for chemometric data analysis. This method provides a general relationship between spectral signatures and percent composition of chemical samples. This approach can be easily extended to power plant applications such as lubrication oil analysis, effluent gas analysis, chemistry of reactor coolant analysis and boiler water chemistry analysis, application in the nuclear medicine field and others.

A multi-layer perceptron, with a back-propagation algorithm to train the network connection weights, was utilized in this study. Different approaches were examined to reduce the estimation error and the learning time. Preprocessing the learning data is necessary if the data contain a large number of training patterns. A Kohonen network may be invoked as a preprocessor to cluster the patterns into different classes with associated networks. This reduces the learning time by decreasing the number of training patterns necessary to train a network. Since the architecture of the network affects learning convergence, different numbers of hidden nodes are examined in order to obtain the best network performance. Experimental results were used to explore the issue of the proper number of hidden nodes for the networks used in the present application. The network with the smallest estimation error was used to determine the optimal number of hidden nodes. The network performance was also evaluated using both one and two hidden layers. The average estimate from an ensemble of networks, that are trained with different initial conditions, was used to improve the overall estimation error. Sensitivity of the network estimation for uncertainties in the input pattern, and due to regional perturbation of spectral signatures were also studied. Analysis of network connection weights was another subject of this study. Studies were also carried out to determine the behavior of the convergence of connection weights during training. A statistical study was performed to study the distribution of connection weights.

Several spectral pre-processing approaches were used to enhance the sensitivity of composition estimation. These include spectral averaging, bias removal by differencing, and the use of different target vectors.

The results of this research and development demonstrated the feasibility of applying the neural networks technology to chemical composition analysis. Application to the estimation of power plant variables and process parameters are also presented. A set of guidelines for developing and applying neural networks for chemometric analysis were developed as part of this thesis.

Download
Files over 3MB may be slow to open. For best results, right-click and select "save as..."

Share

COinS