Real time prediction of internal temperature of heat generating bodies using neural network

Abstract

Heat generating bodies are typically immersed in surrounding fluid that acts as a coolant, either though forced or natural convection. For safe operation, the heat generated should be transferred to the surrounding fluid at a rate that is required to maintain the internal temperature below a critical value. In cases such as nuclear fuel rods and lithium-ion batteries, it is very difficult to install conventional measuring devices like thermocouples and pyrometers inside the heated equipment to monitor the internal operating temperature and trigger warnings when it reaches the critical point. It is however relatively easy to measure the temperature of the surrounding coolant in which they are immersed. From these measurements, the temperature distribution inside the equipment can be estimated using inverse techniques. Inverse analysis can generate data that provides insight into thermal behavior of equipment and help detect overheating events so that corrective action may be taken. A generic and novel framework for conducting inverse analysis to predict internal temperature of cylindrical heat generating bodies has been developed and validated with experimental setup. Simulation-based inverse analysis techniques require large computation effort for complex geometries and high-fidelity models, which make them impractical for real time applications. To circumvent this, a neural network-based model was utilized for predicting temperature inside the system in lieu of a high-fidelity simulation model. Training data was generated through high fidelity simulation using OPENFOAM with an axi-symmetric model. The approach was applied to an experimental setup using a cartridge heater as a representation of a heat generating rod. The coolant height and temperature measured in fluid region was given as input to a trained neural network model to predict surface temperature and inside temperature of the heat generating rod. Inversely predicted temperature was in good agreement with experiment data over wide range of power input. In addition, temperature obtained from this model was used to display in real time on an augmented reality device, which could allow field workers to easily monitor in the field. Novelty of this framework is that the measured coolant temperature in which the heat generating body is immersed is used as input for the rapid and accurate inverse prediction of internal temperature.