| 초록 |
Diagnosing concurrent multi-accident scenarios in nuclear power plants is critical for safety but faces a fundamental data scarcity bottleneck: the number of possible accident compositions grows combinatorially, making exhaustive simulation infeasible. We propose a concept-vocabulary compositionality (CVC) framework, which learns a discrete vocabulary of transient concepts-per physical subsystem via vector quantization-from single-accident simulation data and then diagnoses novel multi-accident compositions at inference through concept-level compositional reasoning, without requiring any multi-accident training data. The key innovation is CVC-based pairwise differential scoring with temporal phase profiles, a zero-shot method that identifies the concurrent accidents that best explain an observed concept histogram by comparing it with the additive superposition of single-accident prototypes. On the Compact Nuclear Simulator (CNS) for a 900 MWe pressurized water reactor-three accidents (loss-of-coolant accident, steam generator tube rupture, main steam line break), 400 single-accident training runs and 500 genuinely coupled multi-accident test runs-the best seed-ensemble configuration achieves 97.0% composition diagnosis accuracy (Clopper-Pearson 95% CI [96.0%, 97.8%]), and the discrete 192-concept vocabulary provides statistically significant advantages over continuous-feature baselines (Wilcoxon signed-rank p < 0.01, Cohen's d approximate to 1.0) with constructive seed-ensembling that continuous representations cannot support. To assess generalization on a substantially harder problem, we further validate the framework on an independent Generic PWR (GPWR) simulator with four accident types, six pairwise compositions, and a three-accident composition. On this harder task the Conditional CVQ - the same CVC pipeline combined with a calibrated conditioning input - attains 0.997 six-class and 0.951 all-composition zero-shot accuracy. In an apples-to-apples comparison without conditioning, the marker-free CVC representation (0.72 six-class) outperforms a strong Transformer encoder (0.51 six-class) by 21 percentage points despite equal single-accident competence, indicating that discrete, superposable representations enable zero-shot compositional generalization more effectively than supervised continuous embeddings. Concept fingerprint analysis confirms that each accident type develops physically interpretable concept signatures in distinct reactor subsystems, enabling traceable compositional reasoning for safety-critical deployment.
|
EXPERT SYSTEMS WITH APPLICATIONS, v.332, 2027-01
Keywords Vector quantization Compositional generalization Zero-shot learning Multi-accident composition Nuclear power plant diagnosis Concept vocabulary