Multi-modal early warning system for lithium-ion battery thermal runaway via SDP image fusion and deep residual learning

Abstract Lithium-ion batteries have become the dominant technology in rapidly advancing energy storage systems, owing to their high energy density and long cycle life, which are critical for applications ranging from portable electronics to electric vehicles and grid-scale storage. However, under harsh conditions, such as battery short-circuits caused by needlesticks, battery fires hinder large-scale commercial applications. This study zeroes in on the critical issue of thermal runaway (TR), leveraging data from mechanical-induced internal short circuits obtained at the Oak Ridge National Laboratory. Diverging from the traditional methodologies that rely on static voltage and temperature thresholds, this research introduces an innovative SDP-ResNet50-TR network. This network amalgamates voltage, temperature, and load signals into a visual format, facilitating image-based recognition. This approach adeptly captures the nuanced fluctuations in voltage and temperature, significantly advancing the warning time. The study further extends its scope to the diagnosis and classification of ferrous lithium phosphate (LFP) batteries, which are notoriously difficult to discern, particularly in the realms of low and moderate TR conditions. The SDP-ResNet50-TR network achieves an impressive 97.3% accuracy in distinguishing between calculated hazard severity (CHS) 13.8 and CHS27.1 states, underscoring its efficacy in identifying subtle yet critical battery health indicators. Concluding with an application to a diverse dataset encompassing multiple lithium-ion battery configurations, the method autonomously classifies TR states. The method achieves a validation accuracy of 93.21%, demonstrating robust generalizability across diverse conditions.