Accurate monitoring of temperature distribution in lithium-ion power batteries is critical for advanced thermal management. However, nonlinear dynamics and sparse sensor deployment pose significant challenges for model-based temperature regulation. To address this, we propose a data-driven modeling framework based on eigenmode approximation for real-time full-state temperature field reconstruction under sparse sensing conditions. The framework integrates three key components: offline training, sensor optimization, and online learning. First, proper orthogonal decomposition (POD) extracts dominant eigenmodes from offline temperature field data. Next, the orthogonal-triangular factorization with column pivoting (QR-CP) algorithm optimizes sensor placement and derives low-order representations of the temperature field. Leveraging these results, a machine learning model dynamically estimates unknown system states and reconstructs the full temperature field from sparse measurements. Experimental validation confirms the framework’s robustness in prediction accuracy and its ability to reduce sensor requirements.