Semi-supervised self-organizing maps for post-processing dynamic scenarios of an integrated deterministic and probabilistic safety analysis

The safety analysis of dynamic systems is challenged by the need of implementing efficient methods for accidental scenarios generation that would account for timing of failure events, and methods for the thereby combinatorial increase of accidental scenarios. Post-processing is aimed at retrieving safety relevant information regarding the system behavior, by classifying the scenarios in safe, Near Misses (NMs), failed and Prime Implicants (PIs) that are scenarios with no, little, considerable and large risk associated to their occurrence, respectively. In this context, three novel approaches has been proposed for the post-processing of dynamic scenarios. In this thesis, we propose the use of a Semi-Supervised Self Organizing Map (SSSOM) in a novel learning scheme based on the Manhattan distance, for post-processing the multi-valued dynamic scenarios that are collected during an Integrated Deterministic and Probabilistic Safety Analysis (IDPSA) of a dynamic system. We exploit different features of the SSSOM (the stand-alone SSSOM ,the MQE based SSSOM, the Baricenter based SSSOM, the Minimum neuron based SSSOM and the Maximum neuron based SSSOM) through which we define three alternatives strategies: a stand-alone SSSOM, a locally weighted ensemble of SSSOMs and a decision tree based on an ensemble of SSSOMs. The rationale behind the three alternative strategies will be given, together with the comparison and the analysis of the results of their application to the U-Tube Steam Generator (UTSG) accidental scenarios. It is shown, how the stand-alone SSSOM is able to group the dynamic scenarios in four groups and how the total classification performances are improves by recurring to the ensemble approaches.