Risk-Based Clustering for Near Misses Identification in Integrated Deterministic and Probabilistic Safety Analysis

In Integrated Deterministic and Probabilistic Safety Analysis (IDPSA), safe scenarios and Prime Implicants (PIs), i.e., minimum combinations of failure events that are capable of leading the system into a fault state are generated by simulation. Post-processing is needed to extract relevant information from these scenarios. In this paper, we propose a novel post-processing method which resorts to a risk-based clustering method for identifying Near Misses among the safe scenarios, i.e., combinations of failure events that lead the system to a quasi-fault state, a condition close to accident. This is important because the possibility of recovering these combinations of failures within a tolerable grace time allows avoiding deviations to accident and, thus, reducing the downtime (and the risk) of the system. The early identification of Near Misses can, then, be useful for online integrated risk monitoring, for rapidly detecting the incipient problems and setting up the recovery strategy of the occurred failures. The post-processing risk-significant features for the clustering are extracted from: i) the probability of a scenario to develop into an accidental scenario, ii) the severity of the consequences that the developing scenario would cause to the system, iii) the combination of i) and ii) into the overall risk of the developing scenario. The optimal selection of the extracted features is done by a wrapper approach, whereby a Modified Binary Differential Evolution (MBDE) embeds a K-means clustering algorithm. The characteristics of the Near Misses scenarios are identified solving a multi-objective optimization problem, using the Hamming distance as a measure of similarity. The feasibility of the analysis is shown with respect to fault scenarios in a dynamic Steam Generator (SG) of a Nuclear Power Plant (NPP)