A stochastic approach to accident identification in nuclear power plants

Identification of the types of accidents and proper actions is required at an early stage of an accident in nuclear power plants. The accident of the plant can be identified by their symptom patterns related to the principal variables and operating status of major equipment. The patterns are identified by the Self-Organizing Feature Map (SOFM), unsupervised artificial neural network, for feature mapping algorithm and the Hidden Markov Model (HMM), a stochastic technique for solving the time series problem. The off-line data from a compact nuclear simulator are vector quantized by SOFM clustering algorithm. The HMM is created for each accident from a set of training data which are the result of vector quantization. The accident identification is decided by calculating which model has the highest probability for given test data. The system uses a left-to-right model including 6 states and 16 input variables to identify 7 types of accidents and the normal state. The HMM is trained by the maximum-likelihood estimation method which uses forward-backward algorithm and Baum-Welch re-estimation algorithm. The optimal path for each model at the given observation is found by Viterbi algorithm, and then the probability of optimal path is calculated. The simulation results show that the proposed system identifies the accident types correctly. It is also shown that the diagnosis is performed well for incomplete input observation caused by sensor fault or malfunction of certain equipment. (author). 13 refs, 4 figs, 2 tabs