TY - JOUR
T1 - State-of-health estimation of Li-ion batteries in the early phases of qualification tests
T2 - An interpretable machine learning approach
AU - Lee, Gyumin
AU - Kim, Juram
AU - Lee, Changyong
N1 - Publisher Copyright:
© 2022 Elsevier Ltd
PY - 2022/7/1
Y1 - 2022/7/1
N2 - Reducing the time and cost associated with lithium-ion (Li-ion) battery qualification tests is critical to developing electronic devices and establishing their quality assurance policies. In this study, we develop an interpretable machine learning model for estimating the future state-of-health (SOH) of Li-ion batteries in the early phases of qualification tests. First, a window-moving technique is used to extract the statistical features that represent battery capacity-fading behaviors over certain cycles. Second, a machine learning model is developed to estimate a battery's future SOH value at a certain cycle. Third, the performance and reliability of the machine learning model are assessed using multiple experiments with varying forecast horizons for SOH estimation. Finally, the SHapley Additive exPlanation (SHAP) method is applied to the model to identify which statistical features are important when estimating a battery's SOH value. The experimental results confirm that the proposed approach can reduce the time required for qualification tests to 100 cycles, i.e., less than a month in practice, with less than a 5% mean absolute percentage error (MAPE) and a 0.002 mean squared error (MSE). The results of model interpretation by SHAP demonstrate that the changes in the SOH values of Li-ion batteries are more important than the values themselves to the SOH estimation. Moreover, the SOH degradation trends near the 100th cycle during the qualification tests are proved to have a significant impact on the future SOH values of the batteries.
AB - Reducing the time and cost associated with lithium-ion (Li-ion) battery qualification tests is critical to developing electronic devices and establishing their quality assurance policies. In this study, we develop an interpretable machine learning model for estimating the future state-of-health (SOH) of Li-ion batteries in the early phases of qualification tests. First, a window-moving technique is used to extract the statistical features that represent battery capacity-fading behaviors over certain cycles. Second, a machine learning model is developed to estimate a battery's future SOH value at a certain cycle. Third, the performance and reliability of the machine learning model are assessed using multiple experiments with varying forecast horizons for SOH estimation. Finally, the SHapley Additive exPlanation (SHAP) method is applied to the model to identify which statistical features are important when estimating a battery's SOH value. The experimental results confirm that the proposed approach can reduce the time required for qualification tests to 100 cycles, i.e., less than a month in practice, with less than a 5% mean absolute percentage error (MAPE) and a 0.002 mean squared error (MSE). The results of model interpretation by SHAP demonstrate that the changes in the SOH values of Li-ion batteries are more important than the values themselves to the SOH estimation. Moreover, the SOH degradation trends near the 100th cycle during the qualification tests are proved to have a significant impact on the future SOH values of the batteries.
KW - Interpretable machine learning
KW - Li-ion battery
KW - Qualification test
KW - SHapley Additive exPlanation method
KW - State-of-health estimation
UR - http://www.scopus.com/inward/record.url?scp=85125845066&partnerID=8YFLogxK
U2 - 10.1016/j.eswa.2022.116817
DO - 10.1016/j.eswa.2022.116817
M3 - Article
AN - SCOPUS:85125845066
SN - 0957-4174
VL - 197
JO - Expert Systems with Applications
JF - Expert Systems with Applications
M1 - 116817
ER -