State of Health Estimation for Lithium-ion Batteries Based on Mechanism Fundamental Learning under Variable Charging Strategies

This is the authors' abstract. We don't add key points for this paper.

Monitoring the State of Health (SOH) of lithium-ion battery is critical to ensure its safe use and prolong the lifespan. Conventional SOH estimation methods require specific charging strategies which brings barrier to practical applications. In this paper, Mechanism Fundamental Learning (MFL) model is proposed to estimate SOH of lithium-ion batteries using charging processes under variable charging strategies and sample segments. The mechanism is that battery voltage and current are significantly impacted by the unmeasurable polarization state, which is associated with the usage and charging process. The MFL model learns the underlying mechanism of battery and tracks the internal state by training current and voltage prediction tasks with Long-Short Term Memory models, and shares the extracted information with SOH estimation task by Multi-Task Learning. Finally the SOH is decoupled from the internal state carried by both hidden state and cell state of LSTM. It has assessed the SOH of batteries with a mean absolute error of 0.76% and root mean squared error of 1.04%.