ADAPTIVE VECTOR QUANTIZATION OPTIMIZED BY GENETIC ALGORITHM FOR REAL-TIME DIAGNOSIS THROUGH FUZZY-SETS

The accurate diagnosis of accidents in a nuclear power plant has funda mental importance for decision making necessary to mitigate their cons equences for the power plant as well as for the general public, on the basis of emergency planning. Two main characteristics should be achie ved in this kind of diagnostics, namely, real-time features and adapti ve capacity. The first characteristic gives the operators the possibil ity of predicting degraded operations and monitoring critical safety f unctions related to that specific situation. The second one allows the system to be able to deal with accidents that were not incorporated i n the training sample set, in which case the operators are unprepared because they were not trained to face an event that they did not obser ve even in simulator training. The Three Mile Island accident is a cla ssic one to demonstrate that these kinds of events are possible. Sever al methodologies have been tried to match those characteristics. While the first one is achieved through the permanent evolution of new fast er processors, the second one can only be achieved through the simulat ion of human cognitive processes, which show higher adaptive behavior Our model utilizes a neural network, fuzzy sets, and a genetic algorit hm to simulate that behavior. We have used a neural network activated by an additive model and trained with an unsupervised competitive trai ning law Once trained with three accidents (steam generator tube ruptu re, blackout, and loss-of-coolant accident), a synaptic matrix was obt ained, in which the elements represent the interchanging weights betwe en neurons in the concatenated input/output space and the competitive neurons that fight to encode the input-output vector. This kind of com petition establishes a statistical classification of the state variabl es, changing with time, clustering them in centroids labeling the kind of accident for which variables are being sampled. Thus, the accident identification is done in real time with the synaptic matrix. However the centroids are located in the same time value, in view of the fact that the neural network algorithm treats the variable time as an inde pendent one. Therefore, a genetic algorithm is applied to a fuzzy syst em formed by the partition of the variables space with fuzzy sets dete rmined by the neural network centroids, in order to estimate the optim al positions in the time variable where the fuzzy system centroids mus t be located As a consequence, the diagnostic can be done in represent ative regions of each accident with maximum confidence.