Computational methods for Loss-Of-Flow Accident (LOFA) Precursors identification in a simplified Superconducting Magnet Cryogenic Cooling Circuit for Nuclear Fusion Application

In nuclear fusion systems, such as the International Thermonuclear Experimental Reactor (ITER), plasma is magnetically confined with Superconductive Magnets (SMs) that must be maintained at cryogenic temperature to preserve their superconductive properties by a Superconducting Magnet Cryogenic Cooling Circuit (SMCCC). The Loss-Of-Flow Accident (LOFA) must be avoided, because it endangers the ability of the SMCCC to keep the SMs cooled. In this respect, an approach to promptly identify LOFA precursors (i.e., those component failures leading to a LOFA) is here developed, based on an On-line Supervised Spectral Clustering (OSSC) method embedding the Fuzzy C-Means (FCM) algorithm. The approach is applied to the simplified cryogenic cooling circuit of a single module of the ITER Central Solenoid (CS), whose behaviour in normal and abnormal conditions is simulated by the validated deterministic 4C code. Results show that the approach elaborated recognises timely several LOFA precursors and identifies most of the components failed. On the other hand, in some cases, it detects LOFA precursors in scenarios with no LOFA and identifies as precursors some components that are not actually failed. On one side, this conservatively increases the safety of the SMCCC (by overestimating the number of failed components to be inspected); on the other side, it reduces its availability (due, e.g., to unnecessary inspection procedures). An attempt to reduce this overestimation is thus made by improving the quality of the “maps” used for training the LOFA precursors identification approach. To this aim, an adaptive meta-model is constructed that mimics the behaviour of the detailed, long-running 4C code, but with a reduced computational cost. This allows a significantly larger number of simulation runs and, thus, a deeper exploration of the possible abnormal conditions of the system to be used for learning the rules of precursors identification. The “additional” information thereby collected is employed to refine the LOFA precursors maps and to possibly enhance the performance of the proposed approach.