Machine Learning for Estimation of Fiber Optics Temperature Field Sensing in a Thermal Hydraulic Flow Loop

Fiber optic distributed temperature sensors are potentially promising options for nuclear reactor thermal-hydraulic system because they provide information about fluid temperature field. Compared to an array of point sensors (e.g. thermocouples), fiber optic sensors based on Rayleigh backscattering offer temperature sensing with higher spatial density and faster response time. Data measured with fiber optic sensor is a 2D matrix, where one dimension is length along the fiber, and another dimension is time of measurements. When exposed to high temperature and ionizing radiation environment in a nuclear reactor, fiber optic silica material degrades over time. The objective is to use machine learning methods to self-monitor or validate distributed measurements to detect early signs of failure. In our approach, we use long short-term memory (LSTM) neural networks trained on prior history of measurements to predict the next data point in real time. The basis set of fiber optic segments is identified by calculating correlation coefficients between nearest neighbors. Performance of this approach is studied using dynamic temperature field data sets obtained with single mode 1550nm optical fiber installed in a water flow loop with a thermal mixing Tee. This study also aims to determine the minimal number of training fiber segments for reliable estimation of measurements made with other segments of the fiber.