Title: Reconstructing Unobservable Temperature Fields via Simulation-Aided Intelligent Sensing

Abstract: Monitoring the temperature distribution within components and sub-structures in real time presents a significant challenge for many systems, largely due to limitations on where sensors can be practically installed. Although machine learning (ML) is a versatile tool across numerous applications, its implementation for high-resolution thermal monitoring is often constrained by the scarcity of high-quality training datasets. To address this, this study introduces a novel methodology for creating datasets tailored for industrial use, leveraging randomized physics-based simulations. We validate this approach through a proof-of-concept hardware experiment, employing a neural network (NN) trained exclusively on synthetic data to reconstruct the internal temperature field from sparse sensors embedded within the hardware. The results indicate that NN-based reconstructions surpass Kriging in terms of robustness and facilitate real-time inference, thereby rendering the method ideal for the online monitoring of thermal states that are otherwise unobservable.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC