Deep learning techniques for in-core perturbation identification and localization of time-series nuclear plant measurements

The use of machine learning in the field of reactor safety and noise diagnostics has recently seen great potential given the advancements made in computational tools, hardware and noise simulations. In this work we demonstrate how deep neural networks, specifically recurrent and convolutional neural networks can be trained in a synthetic setting and aligned to operate on real plant measurements to recover perturbation type and origin location from time-series signals. We first utilize the vast quantities of synthetic data generated from the extended SIMULATE-3K codes, simulating a Swiss 3-loop pre-KONVOI reactor to train our networks under a variety of differing perturbation settings. Additionally, we extend these approaches to operate in the setting of unsupervised real plant measurements, where information about the true perturbation characteristics is unknown. As such, we show the applicability of a self-supervised domain adaptation approach to correctly align the representations learned by the neural network between both the synthetic and real detector readings to more concretely classify and localize perturbation. We validate our approaches under a number of experimental analyses showing successful performance in both simulated and synthetic domains.