Nuclear power plant status diagnostics using simulated condensation : an auto-adaptive computer learning technique

The application of artificial neural network concepts to engineering analysis involves training networks, and therefore computers, to perform pattern classification or function mapping tasks. This training process requires the near optimization of network inter-neural connections. A new method for the stochastic optimization of these interconnections is presented in this dissertation. The new approach, called simulated condensation, is applied to networks of generalized, fully interconnected, continuous perceptrons. Simulated condensation optimizes the nodal bias, gain, and output activation constants as well as the usual interconnection weights. This approach is shown to provide a more complete and faster optimization than standard techniques such as backpropagation. The method also adjusts the network nodal architecture to give faster learning and better generalization results. Because of this dynamic node architecture feature, the appropriate network architecture need not be known before the learning process is initiated. The simulated condensation methodology simply builds the network required for the learning task at hand. In this work, the simulated condensation network paradigm is applied to nuclear power plant operating status recognition. A set of standard problems such as the exclusive-or problem and others are also analyzed as benchmarks for the new methodology. The objective of the nuclear power plant accident condition diagnosis effort is to train a network to identify both safe and potentially unsafe power plant conditions based on real time plant data. The data is obtained from computer generated accident scenarios. A simulated condensation network is trained to recognize seven nuclear power plant accident conditions as well as the normal full power operating condition. These accidents include, hot and cold leg loss of coolant, control rod ejection and steam generator tube leak accidents. Twenty-seven plant process variables are used as input to the neural network. Results show the feasibility of using simulated condensation as a method for diagnosing nuclear power plant conditions. The method is general and can easily be applied to other types of plants and plant processes.