Abstract: This paper investigates the feasibility of acoustic emission detection technology for ultra-early identification of the incubation stage of thermal runaway in lithium-ion batteries. An experimental platform was established, integrating various monitoring devices including charge-discharge equipment, an acoustic emission system, a battery voltage acquisition system, and a battery temperature acquisition system. By acquiring the elastic wave characteristics generated inside lithium-ion batteries during normal charge-discharge cycles and during thermal runaway induced by overcharging, a neural network model was constructed for training and identification. The results show that， during normal charge-discharge, the acoustic emission signal amplitude is distributed within ±0.04 V, with a dominant frequency of 220 kHz and characteristic frequencies at 60， 100, 180 kHz; during overcharging, multiple outlier points exceeding this range appear in the signal amplitude, and a new characteristic frequency emerges at 20 kHz. Using a convolutional neural network to process and train the grayscale images of spectral information, the model achieves a classification accuracy of 99.87% for acoustic emission signals and maintains a high accuracy in cross-validation. This acoustic emission monitoring method can identify and differentiate the early, middle, and late stages of the thermal runaway incubation process in lithium-ion batteries, thereby enabling monitoring and early warning of thermal runaway.

Key words: lithium-ion battery, thermal runaway, acoustic emission, convolutional neural network