A Multimodal Deep Learning Framework for Overcharge‐Induced Thermal Runaway Prediction in Prismatic Lithium Iron Phosphate Battery

With the widespread adoption of lithium iron phosphate (LFP) batteries in energy storage and electric vehicles, overcharge‐induced thermal runaway (TR) is recognized as a critical threat to battery safety. However, conventional voltage/temperature threshold‐based detection methods suffer from delayed responses, proving inadequate for early‐warning applications. To address these shortcomings, this research employs a common prismatic LFP battery, collecting multimodal data (gas emissions—H 2 /CO/HF, voltage, and temperature) through five stepwise overcurrent overcharging experiments. With the only consistent quantitative association among multimodal data, the time stamp was chosen to synchronize temporal alignment to enable both representing multimodal measurements at any given time point as a unified numeric array suitable for deep learning and quantitatively benchmarking algorithmic superiority. Based on this, a hybrid deep learning framework is developed to extract spatiotemporal features from multimodal electrochemical time‐series data, using integrated transformer layers and multiscale convolutional layers. Compared with experimental measurements, the proposed multimodal early‐warning model is demonstrated to surpass traditional voltage/temperature change rate threshold methods. Validation results indicate that the warning time was advanced by an average of 289 s across all overcharging conditions—ranging from 188 s (1.5 times current overcharge) to 492 s (0.5 times current overcharge)—which represents a 37.2% improvement. These results provide an intelligent safety monitoring strategy beyond single‐parameter thresholds and a time‐aligned multimodal learning paradigm that can be generalized to broader electrochemical safety diagnostics.