Voltage relaxation-based state-of-health estimation of lithium-ion batteries using convolutional neural networks and transfer learning

Convenient and accurate state-of-health (SOH) estimation of lithium-ion batteries (LIBs) is crucial for the security of energy storage systems. However, it is a challenging task to estimate the SOH of LIB due to complex cycling conditions and limited training data. The inputs of most existing methods cannot always be satisfied under complex cycle conditions, as cycle conditions may change anytime, especially during dynamic discharge processes. Thus, we propose a new end-to-end SOH estimation method based on relaxation voltage that is not dependent on specific cycling conditions. Specifically, the relaxation voltage profiles at the end of fully charging are input to a one-dimensional convolutional neural network (CNN) to estimate SOH directly. Transfer learning is adopted to leverage the source domain knowledge to the target domain to solve the issue of limited data. Moreover, the most promising CNN hyperparameters are determined automatically by the Bayesian optimization algorithm (BOA) during the pre-training and transfer learning. The accuracy and robustness of the proposed method are verified on two publicly available datasets consisting of 121 and 4 commercial cells, respectively, with a real-driving discharge profile. The root-mean-square errors of the proposed method are 0.0128 and 0.0092, respectively, with only 1.5 % and 10 % training data from the two target domains. The method has a high potential for online applications with preferable accuracy and computational performance. Our work highlights the effectiveness and generalizability of the end-to-end LIBs SOH estimation method based on easily accessible relaxation voltage profiles.