Heuristic learning parameter identification for surveillance and diagnostics of nuclear power plants

A new method of heuristic reinforcement learning has been developed for parameter identification purposes. In essence, this new parameter identification technique is based on the idea of breaking a multidimensional search for the minimum of a given functional into a set of unidirectional searches in parameter space. Each search situation is associated with one block in a memory organized into cells, where the information learned about the situations is stored (e.g. the optimal directions in parameter space). Whenever the search falls into an existing memory cell, the system chooses the learned direction. For new search situations, the system creates additional memory cells. This algorithm imitates the following cognitive process: (i) characterize a situation, (ii) select an “optimal" action, (iii) evaluate the consequences of the action, and (iv) memorize the results for future use. As a result, this algorithm is "trainable" in the sense that it can learn from previous experience within a specific class of parameter identification problems. From the mathematical point of view, the algorithm utilizes the inversion of a newly introduced concept of "fuzzy maps" between two spaces: the feature space and the parameter correction space. However, this operation is performed by the cognitive process described above rather than by mathematical manipulations. The main advantages of the new method are: (i) because of the minimum amount of computations, the parameter identifications proceed faster than by the usual methods. and (ii) the parameter search can proceed automatically without the user-program interaction. The new algorithm was validated both analytically and via extensive computer simulations, utilizing a model of a U-tube steam generator of the type used in pressurized water reactors. A "real world" application was implemented whereby analog signals from an experimental pressure loop were utilized as inputs to the present learning algorithm. After a training period, during which the algorithm was shown the normal behavior of the loop, the system was able to diagnose failures in the loop and to accurately determine the parameters characterizing its normal behavior.