Abstract
Initiating event identification is essential in managing nuclear power plant (NPP) severe accidents. To facilitate the identification, a two-stage feature extraction scheme that incorporates the proposed sensor type-wise block projection (stBP) and deflatable sequential forward selection (dSFS) is utilized to elicit the discriminant information in the data from various NPP sensors. Based on the idea of stBP, the primal features can be extracted without breaking the intrinsic spatial structure among the multi-channel data of specific sensor types. The extracted features are then subject to further dimensionality reduction by selecting the sensors that are most relevant to the events under consideration. Exhaustive search is not feasible in this selection, and a combinatorial optimization technique is required to find suitable solutions. Unlike the original sequential forward selection, dSFS includes a sensor deflation scheme allowing sensors in the preselected set to be recursively refined. Results from detailed experiments with data generated from a simulator of Taiwan Maanshan NPP illustrate the efficacy of the proposed scheme, achieved a recognition rate of 95.36 % that is higher than those obtained using the features from the existing methods.