Nuclear plant diagnostics using neural networks with dynamic input selection

The work presented in this dissertation explores the design and development of a large scale nuclear power plant (NPP) fault diagnostic system based on artificial neural networks (ANNs). The viability of detecting a large number of transients in a NPP using ANNs is demonstrated. A new adviser design is subsequently presented where the diagnostic task is divided into component parts, and each part is solved by an individual ANN. This new design allows the expansion of the diagnostic capabilities of an existing adviser by modifying the existing ANNs and adding new ANNs to the adviser;This dissertation also presents an architecture optimization scheme called the dynamic input selection (DIS) scheme. DIS analyzes the training data for any problem and ranks the available input variables in order of their importance to the input-output relationship. Training is initiated with the most important input and one hidden node. As the network training progresses, input and hidden nodes are added as required until the networks have learned the problem. Any hidden or input nodes that were added during training but are unnecessary for subsequent recall are now removed from the network. The DIS scheme can be applied to any ANN learning paradigm;The DIS scheme is used to train the ANNs that form the NPP fault diagnostic adviser. DIS completely eliminates any guesswork related to architecture selection, thus decreasing the time taken to train each ANN. Each ANN uses only a small subset of the available input variables that is required to solve its particular task. This reduction in the dimensionality of the problem leads to a drastic reduction in training time;Data used in this work was collected during the simulation of transients on the operator training simulator at Duane Arnold Energy Center, a boiling water reactor nuclear power plant. An adviser was developed to detect and classify 30 distinct transients based on the simulation of 47 scenarios at different severities. This adviser was then expanded to detect and classify a total of 36 transients based on the simulation of 58 transient scenarios. The noise tolerant characteristics of the adviser are demonstrated