Robust capacity estimation with uncertainty quantification for li-ion batteries under temporal data masking challenges: A progressive learning approach

Accurate estimation of lithium-ion battery capacity is critical for reliability management but it faces challenges due to temporal data masking, a prevalent issue in real-world cloud applications causing time-series masking and data degradation. To address this, we propose a progressive learning framework that constructs a data-quality-aware learning pathway, enabling robust training solely on high-quality laboratory data by progressively generating and incorporating artificially masked low-quality samples. The framework integrates dynamic sampling and adaptive resampling strategies to enhance model robustness against data skewness. Additionally, uncertainty quantification with strong physical interpretability is efficiently achieved through implicit ensemble learning on homologous charging segments, avoiding the computational bottlenecks of Bayesian or ensemble-based methods. Validated on the LFP, NCA, and NCM datasets, our method achieves RMSEs of 0.2170 %, 0.1924 %, and 0.1326 % on clean data, respectively. When 50 % of the data is masked, the RMSEs increase only slightly, with the maximum absolute increase being just 0.0303 %, and the model maintains high accuracy even with masking ratios as high as 70 %. The framework also generalizes well across different deep learning architectures. This work bridges the gap between laboratory models and real-world deployment for battery management systems.