Deep learning-based FPGA steady-state temperature field reconstruction for nuclear safety level I&C systems

Abstract

Obtaining the steady-state temperature distribution on the field programmable gate array (FPGA) chip is the basis for establishing an effective thermal monitoring mechanism to ensure the reliable operation of the nuclear safety level instrumentation and control (I&C) system. In this work, a U-Net-based deep learning method to reconstruct the temperature field of FPGA using limited measurement points is proposed. Firstly, the environmental test under accident conditions (ambient temperature of 55 ◦C) of I&C equipment is conducted, and finite element analysis (FEA) is used to simulate the thermal behavior of the control protection cabinet, and validated against experimental results. Subsequently, based on the FEA model, temperature distributions under varying power consumption distributions and heat dissipation configurations are manually generated. Using random sampling, 200 sparse datasets with 8 measurement points are constructed. A U-Net-based model is developed and compared with traditional multi-layer perceptron (MLP) and graph neural network (GNN) models. Results show the U-Net achieves the lowest reconstruction error, with a maximum absolute error of only 0.1828 %, demonstrating superior accuracy and robustness. This method provides an effective solution for FPGA temperature field reconstruction in complex engineering scenarios, which is beneficial for improving the reliability and stability of nuclear safety level I&C systems.