Skip to main content
U.S. Department of Energy
Office of Scientific and Technical Information

Reactor diagnostics rule generation through pattern recognition

Thesis/Dissertation ·
OSTI ID:6076287

A systematic generation of shallow knowledge for nuclear power plant transient diagnostics through pattern recognition is investigated. An entropy minimax algorithm is used to partition a database of simulated transient events into classes of events of similar characteristics, which are then represented as production rules for diagnostics. Pattern recognition involving feature selection and pattern discovery is used to choose N best features and to discover patterns associated with each event class. The selection of the best features is attained by discarding redundant and non-discriminatory features. The information-theoretic entropy is used to discover the patterns by searching through a partitioned N-dimensional feature space, populated with events of the database, to locate subspaces that discriminate among the event classes. The feature space is partitioned by decomposing the N-dimensional partitioning problem into N independent one-dimensional problems. Transient events are represented by patterns containing common characteristics of time-varying features over the diagnostic time. An algorithm that can update the patterns in an incremental fashion with minimum effort, as new data are obtained, has also been developed. The Midland Nuclear Power Plant Unit II Simulator is used to generate 144 single-failure events. Based on these events twenty-five production rules are generated, representing a two-level hierarchical knowledge structure along the critical safety function approach. These rules subsume the corresponding rules at intermediate time interval, thereby providing diagnostic capability at any time throughout the transient.

Research Organization:
Michigan Univ., Ann Arbor, MI (USA)
OSTI ID:
6076287
Country of Publication:
United States
Language:
English