Multi-level early warning of thermal runaway based on internal pressure-temperature fusion for lithium-ion batteries

Timely and accurate early warning of thermal runaway is the trump card for preventing safety accidents of lithium-ion batteries. Due to the strong concealment, conventional warning methods based on electrical and thermal signals struggle to detect minor changes, which cause short warning times and low accuracy. Hence, an intelligent hierarchical early warning method based on internal pressure signals is proposed. The millimeter-scale gas-heat sensor is embedded into the battery without significant damage to precisely capture internal pressure-temperature. Through both normal and micro-overcharge cyclic tests, the coupled evolution law of electrical (voltage)- thermal (temperature)- gassy (pressure) is detailed analyzed to illustrate the feasibility and advancement. The Gate recurrent unit network is constructed and trained with the pressure and temperature as input and different TR risk levels as output. Attention mechanism is introduced to learn the importance of different variables to enhance the adaptability and interpretability. Results from various micro-overcharge cyclic experiments demonstrate that the proposed method achieves classification accuracy of 98.2% and average early warning time of 1.51 hours. Compared with other three methods based on temperature, gas, stress, respectively, the accuracy increased by 3.2% and the time extended by 157.14%, which provides a powerful support for active safety protection. • Millimeter-scale sensor is embedded in the battery for in-situ monitoring of internal temperature and pressure. • The gradual decrease in the minimum internal pressure during normal cycling is attributed to gas consumption caused by side reactions. • The greater the micro-overcharge degree,the acceleration effect of decomposition reactions are more pronounced to cause fewer TR cycles. • A multi-level early warning method based on internal pressure and temperature is developed using Attention-GRU network with high accuracy of 98.2%