Condition-Aware Progressive Knowledge Distillation: Lightweight and Accurate Prediction of Lithium-Ion Battery Remaining Useful Life

Abstract

This study presents a condition-aware progressive knowledge distillation (CPKD) framework for developing lightweight models capable of accurately predicting the remaining useful life (RUL) of lithium-ion batteries. Within this framework, a teacher model is constructed by embedding a conditional variational autoencoder (CVAE) into a Mamba backbone. Through high-order compression and reconstruction of local–global battery features and the adoption of a dependency-aware loss function, the latent space is regularized to capture the underlying structure of degradation trajectories, enabling high-precision RUL prediction. Subsequently, a lightweight student predictor, realized as a CVAE enriched with operating-condition descriptors, is configured as the target model. Unsupervised clustering is employed in conjunction with a dual-path gating mechanism to activate an operating-condition monitor. While a multichannel topology constructed from depthwise separable convolutions (DSCs) is triggered under the synergistic action of the gating weights to extract discriminative features across varying operating regimes and to deepen the representation of complex degradation dynamics. To transfer domain knowledge, a stepwise parameter-freezing knowledge distillation (KD) strategy is devised whereby accuracy-controlled degradation knowledge is progressively distilled from the teacher, ensuring effective RUL prediction by the compact student. Validation indicates that the distilled lightweight model achieves average root-mean-square errors of 0.0163 on the MIT dataset and 0.0384 on a real-world EV dataset, while using only 0.167M parameters and requiring 13.88 ms per inference on embedded hardware.