A convolutional neural network model for SOH estimation of Li-ion batteries with physical interpretability

Abstract

Previous machine learning models for state-of-health (SOH) estimation of Li-ion batteries have relied on prescribed statistical features. However, there is little theoretical understanding of the relationships between these features and SOH degradation patterns of the batteries. This study proposes a convolutional neural network model to estimate the future SOH value of Li-ion bat-teries in the early phases of qualification tests. First, capacity degradation data are transformed into two-dimensional images using recurrence plots and Gramian angular fields, highlighting the time-series features of the data. Second, five types of convolutional neural network models are developed to estimate the SOH values of Li-ion batteries for a certain cycle. Here, class activation maps are generated to present how the models arrive at their conclusions. Finally, the perfor-mance and reliability of the developed models are assessed under various experimental condi-tions. The proposed approach has the following two advantages: it automatically extracts important temporal features from the capacity degradation data for SOH estimation, and obtains the contribution of each temporal feature with respect to the estimation process. The experi-mental results on 379Li-ion batteries confirm that the proposed approach can reduce the time required for qualification tests to 50 cycles, under a 6% mean absolute percentage error.