Uncertainty Quantification Based Health Diagnosis for Lithium-Ion Batteries under Different Operating Conditions

Estimating the state of health (SOH) of lithiumion batteries is a crucial aspect of battery management. Due to complex and dynamic nature of real-world tasks, lithium-ion batteries are under multiple working conditions. The degradation paths are different under those working conditions, which poses challenges for estimation models, such as model mismatch. In addition, uncertainties exist in data acquisition and estimation modeling, which require uncertainty expression in the estimation. In this paper, a neural architecture is proposed, combing neural processes (NPs) with long short-term memory (LSTM). On one hand, assuming the degradation paths under different working conditions follow some stochastic process, NPs estimate the underlying distribution through the relationships from the context. On the other hand, probabilistic representation enables the architecture to capture and quantify uncertainty in estimation. Besides, leveraging LSTM as encoder and decoder enhances the ability of the architecture to capture long-term temporal dependencies. The estimation model is optimized during training using the evidence lower bound (ELBO) as the loss function. Finally, the effectiveness of the proposed architecture is validated based on a capacity degradation dataset obtained from the laboratory.