Dynamic-Residual-Attention KalmanNet: Unsupervised RUL Prediction for Lithium-Ion Batteries With Capacity Regeneration Handling
Abstract
The widespread industrial adoption of lithium-ion (Li-ion) batteries underscores the critical need for accurate state monitoring and virtual measurement of their remaining useful life (RUL). Since internal degradation states resist direct physical measurement, traditional estimators are widely used for online capacity tracking. However, their practical deployment is complicated by the rigid requirement for fixed noise covariance matrices and complex sensor uncertainties. To address these measurement challenges, an advanced state estimation framework termed dynamic-residual-attention KalmanNet with innovation anomaly detection (DRA-KN-IAD) is proposed. This approach integrates a dynamic residual attention (DRA) mechanism with a recurrent neural network (RNN) to adaptively learn the Kalman gain directly from measurement sequences, without explicitly estimating the process or measurement noise covariance matrices. To ensure broad practical applicability, an unsupervised training scheme, which is essentially state-label-free and based purely on observable capacity data, is adopted. A regeneration-handling mechanism is introduced to reduce the impact of transient capacity excursions on RUL prediction. Experiments on the NASA and CALCE datasets demonstrate the effectiveness of DRA-KN-Innovation anomaly detection (IAD).