Doctoral Dissertations
Date of Award
5-1992
Degree Type
Dissertation
Degree Name
Doctor of Philosophy
Major
Nuclear Engineering
Major Professor
Robert E. Uhrig
Committee Members
Belle R. Upadhyaya, Rafael B. Perez, Lefteri H. Tsoukalas. Tadeusz Janik
Abstract
A new neural network architecture called Lateral Feedback Neural Network is introduced in this dissertation. It is to modify the widely used backpropagation by introducing intra-layer connections to the hidden layer of the network. The intra-layer connections follow the competition mechanism of self-excitatory and neighborhood-inhibitory. The learning algorithms are developed by using Generalized Delta Rule to adapt all inter-layer and intra-layer connections and bias terms. The benchmark tests show that the Lateral Feedback network has advantages in learning speed, convergence, and stability over the original backpropagation for all tested cases. Lateral feedback network is applied to fault diagnoses for TVA's Watts Bar nuclear power plant together with the sensitivity analysis and genetic algorithms.
The sensitivity analysis is developed for both backpropagation and lateral feedback networks. It is the analog of the sensitivity analysis for ordinary functions. It treats a successfully trained network as a mapping function and develops the first derivatives of output variables with respect to input variables through the network learning algorithms and architectures. The information extracted from the sensitivity analysis can be used to rank input variables of a network into the order of importance. In the application to TVA's Sequoyah nuclear power plant thermal performance study, the sensitivity analysis provides the information on what may cause the deviation in the plant performance. In the application to fault diagnoses for TVA's Watts Bar nuclear power plant simulator data, the sensitivity analysis is used to guide the selection of optimal input variables for a set of small modular networks to monitor different accident scenarios.
Genetic algorithm is a special search algorithm based on the mechanisms of natural selection and natural genetics. It is a very efficient and robust search methodology which can perform optimal search in a large and complex search space where the conventional algorithms, such as calculus-based search and hill-climbing method, can not perform well. It is applied in this dissertation to neural networks to guide the search for optimal selection of input variables to simplify the network system. The goal of the search is to find an optimal combination of input variables for a neural network to reach fast training, accurate recall, and fewest possible inputs. The application to fault diagnoses for TVA's Watts Bar nuclear power plant simulator data shows that genetic algorithms do find the solutions for each modular network which monitors one assigned accident scenario.
Recommended Citation
Guo, Zhichao, "Nuclear power plant fault diagnostics and thermal performance studies using neural networks and genetic algorithms. " PhD diss., University of Tennessee, 1992.
https://trace.tennessee.edu/utk_graddiss/10897