Precise and fast safety risk classification of lithium-ion batteries based on machine learning methodology

Abstract

How to establish a fast and accurate model with physics implications still remains an unsolved headache for the community. Herein, we establish a battery safety risk classification modeling framework based on a machine-learning algorithm that can accurately and rapidly classify the potential safety risk level. The model can identify defective cells, cells with internal short circuits (ISCs), and cells with possible thermal runaway (TR) using a small portion of one cycle data. About 3 × 105 training samples are generated, covering a wide spectrum of State of Charge (SOC), short circuit resistance, and charging/discharging rate (C-rate)) ranges based on established electrochemo-mechanical model, serving as training and prediction datasets for the machine learning (ML) algorithm. Classifiers in the ML model demonstrates satisfactory performance and robustness. We further prove that only 5 min is needed for the model to classify the cells with over 95% accuracy with a voltage error<0.1 mV if the ISC resistance is smaller than 102Ω. This work highlights the promising future of combining physics-based numerical modeling and data-driven algorithm to provide satisfactory battery safety identification with limited data produced in a short cycling time.