Data-driven short circuit resistance estimation in battery safety issues

作     者：Yikai Jia Jun Xu Yikai Jia;Jun Xu

作者机构：Department of Mechanical Engineering and Engineering ScienceThe University of North Carolina at CharlotteCharlotteNC 28223USA Vehicle Energy&Safety Laboratory(VESL)North Carolina Battery ComplexityAutonomous Vehicle and Electrification(BATT CAVE)Research CenterThe University of North Carolina at CharlotteCharlotteNC 28223USA School of Data ScienceThe University of North Carolina at CharlotteCharlotteNC 28223USA 

出 版 物：《Journal of Energy Chemistry》 (能源化学（英文版）)

年 卷 期：2023年第79卷第4期

页      面：37-44页

核心收录：

学科分类：0808[工学-电气工程] 08[工学] 081104[工学-模式识别与智能系统] 0837[工学-安全科学与工程] 0811[工学-控制科学与工程] 0812[工学-计算机科学与技术（可授工学、理学学位）] 

基　　金：supported by the U.S.Department of Energy’s Office on Energy Efficiency and Renewable Energy(EERE)under the Advanced Manufacturing Office award number DE-EE0009111。 

主　　题：Lithium-ion battery Safety risk Internal short circuit Short circuit resistance Convolutional neural networks 

摘      要：Developing precise and fast methods for short circuit detection is crucial for preventing or mitigating the risk of safety issues of lithium-ion batteries(LIBs).In this paper,we developed a Convolutional Neural Networks(CNN)based model that can quickly and precisely predict the short circuit resistance of LIB cells during various working conditions.Cycling tests of cells with an external short circuit(ESC)are produced to obtain the database and generate the training/testing samples.The samples are sequences of voltage,current,charging capacity,charging energy,total charging capacity,total charging energy with a length of 120 s and frequency of 1 Hz,and their corresponding short circuit resistances.A big database with~6×10^(5)samples are generated,covering various short circuit resistances(47~470Ω),current loading modes(Constant current-constant voltage(CC-CV)and drive cycle),and electrochemical states(cycle numbers from 1 to 300).Results show that the average relative absolute error of five random sample splits is 6.75%±2.8%.Further parametric analysis indicates the accuracy estimation benefits from the appropriate model setups:the optimized input sequence length(~120 s),feature selection(at least one total capacity-related variable),and rational model design,using multiple layers with different kernel sizes.This work highlights the capabilities of machine learning algorithms and data-driven methodologies in real-time safety risk prediction for batteries.