Precise location of failed fuel assembly in operating PWR-core based on fine nuclide evolution and deep metric learning

In Pressurized Water Reactors (PWRs), significant risks to operational safety and economic efficiency can be caused by fuel cladding failure. Traditional diagnostic methods can effectively detect the occurrence and can guess the size of breaches. But precise spatial localization remains a formidable challenge due to the complex, non-linear feature of fission products transport and the inherent “many-to-few” mappings from fuel rods in the core to sensors in the primary coolant loop.A novel localization method that integrates high-precision physical mechanism modeling with deep metric learning is proposed and validated in this paper. With the actual specific operating history, sub-rod subdivision nuclide evolution is calculated by using the software of Bamboo-Core and Bamboo-SFuel. Second, by using a simulation software named Bamboo-LoopN, the nuclide composition in primary loop after the breaching can be evaluated for each possible failure location. In this way, a data library can be constructed to quantitatively link the nuclide composition in primary loop (nuclide composition) to the location in the operating core (core coordinates). Third, a Deep Metric Learning network, optimized with a weighted cross-entropy loss based on a Gaussian kernel similarity matrix, is designed and trained to perform the mapping from nuclide composition to core coordinates. Fourth, once the actual nuclide composition in the primary loop can be measured after the occurrence of fuel failure, a failure probability distribution in the operating core can be obtained by using the Deep Metric Learning network to quantitatively locate the failed fuel assembly.The new method has been rigorously tested using the BEAVRS data and validated by an actual measured data from a commercial M310 unit. Results demonstrate its feasibility. What’s more, as demonstrated by the quantitative results, the new method can also accommodate as much as 30 % noise in the measured nuclide composition, which means that the right failed fuel assembly can also be located even when the noise in the actual nuclide composition measurement is 30 %. It can be concluded that the nuclide composition in fuel assemblies can be taken as the “fingerprint” to recognize the failed fuel assembly in operating PWR-core. It provides a robust, physics-based “prior-prediction” approach for the real-time integrity management of the operating PWR-cores.