Online health estimation strategy with transfer learning for operating lithium-ion batteries

Complex power supply operation conditions complicate the degradation process of lithium batteries, which makes the charge–discharge cycle incomplete and the maximum available capacity not easily accessible. Besides, data-driven methods suffer from limited adaptation and possible overfitting. This paper proposes an online health estimation strategy with transfer learning for estimating the state of health (SOH) of batteries under varying charge–discharge depths and current rates. It aims to alleviate the difficulty in estimating SOH for operating batteries, and broaden the application range of the training model. The core of this strategy is a two-domain transfer CNN-LSTM model that estimates targets by transferring the battery degradation trends of multiple constant conditions. First, health indicators (HIs) with relatively high correlations and wide application ranges are extracted from the voltage and current data of the daily charge process. Then HI-based source domain selection criteria are designed. Since the battery experiences full and incomplete-discharged cases leading to various aging rates, a two-domain transfer CNN-LSTM model is designed. Each subnet includes a CNN and an LSTM to accomplish feature adaptation and time series forecasting. The weights of the sub-nets are updated online to track the drift of the time series covariates. Finally, the proposed strategy is verified on target batteries with varying cut-off voltages and currents, which demonstrates notable accuracy and reliability.