Evaluating Generative Learning Approaches for Cylindrical Battery Thermographic Image Synthesis

Thermographic image synthesis of cylindrical batteries enables the estimation of surface temperature distribution without the need for additional thermal sensors. However, accurately estimating the two-dimensional surface temperature is challenging due to the uneven thermal distribution under varying current rates and environmental conditions. This study proposes a hybrid approach that integrates a physical model with generative learning techniques. Specifically, an equivalent circuit model (ECM) is employed to generate physical features from electrical signals, thereby enhancing the interpretability of data-driven methods. Two widely used generative learning models, generative adversarial networks (GANs) and conditional diffusion models (CDMs), are evaluated for thermographic image generation. Experimental results demonstrate that the GAN and CDM approaches achieve root mean squared errors (RMSE) of 1.48° and 1.63°, respectively, in the two-dimensional temperature estimation.