Mechanical Behavior and Failure Prediction of Cylindrical Lithium-Ion Batteries Under Mechanical Abuse Using Data-Driven Machine Learning

Mechanical failure prediction of lithium-ion batteries (LIBs) can provide important maintenance information and decision-making reference in battery safety management. However, the complexity of the internal structure of batteries poses challenges to the generalizability and prediction accuracy of traditional mechanical models. In view of these challenges, emerging data-driven methods provide new ideas for the failure prediction of LIBs. This study is based on an experimental data-driven application of machine learning (ML) models to rapidly predict the mechanical behavior and failure of cylindrical cells under different loading conditions. Mechanical abuse experiments including local indentation, flat compression, and three-point bending experiments were conducted on cylindrical LIB samples, and mechanical failure datasets for cylindrical cells were generated, including displacements, voltages, temperatures, and mechanical forces. Six ML models were used to predict the mechanical behavior of cylindrical batteries, four metrics were used to evaluate the prediction performance, the coefficients of determination of eXtreme Gradient Boosting (XGBoost) regression and random forest were 0.999, and the root-mean-square errors (RMSE) were lower than 0.015. It is shown that the integrated tree models tested in this study are suitable for the failure prediction of LIBs under the conditions of mechanical abuse. Also, the random forest prediction model outperforms other ML prediction models with the smallest RMSE values of 0.005, 0.0149, and 0.007 for local indentation, flat compression, and three-point bending, respectively. This work highlights the capability of ML algorithms for LIB safety prediction.