Global temperature reconstruction of equipment based on the local temperature image using TRe-GAN

Abstract

During contemporary industrial design and production, traditional numerical simulation is incapable of predicting the temperature field of undetectable surfaces with high accuracy, rapidity and board applicability. In this paper, we present a temperature reconstruction generative adversarial network (TRe-GAN), which can rapidly predict the temperature field of all equipment surfaces through one readily observable two-dimensional temperature image rather than specific thermodynamic parameters. Compared with previous surface temperature prediction networks, TRe-GAN reconstructs surface temperature field via vertex temperatures instead of surface texture, which innovatively incorporates both computer vision semantics and heat transfer semantics. Conditions inputs and an optimization output module were designed to strengthen vertex association and mitigate over-fitting. An orthogonal experimental design scheme considering heat transfer theory is employed to construct reliable data sets representative enough for heat transfer phenomenons. Model details, such as loss function and normalized scheme, are discussed to improve the overall temperature accuracy for thousands of nodes (MAPE = 0.93K). TRe-GAN trained in numerical-simulation data performs well in predicting the global temperature field of the measured experimental data (MAPE=0.89%), which proves that our network effectively avoids over-fitting and can predict measured data from easily available simulated data. Therefore, TRe-GAN has great application potential in producing or designing industrial equipment and target characteristics prediction on the battlefield.