Accurate online capacity estimation of lithium-ion batteries represents a critical technological challenge for ensuring the safe operation of Battery Management Systems (BMS). Conventional data-driven methods often suffer from model overfitting and ambiguous decision-making when addressing the nonlinear time-varying characteristics of battery degradation trajectories. To overcome these limitations, this study proposes an interpretable capacity estimation framework based on multi-source heterogeneous feature fusion. The methodology first extracts multi-scale health features from three aspects: charge-discharge cycle profiles, incremental capacity (IC) curves, and differential thermal voltammetry (DTV) curves. Pearson correlation analysis and principal component analysis (PCA) are then applied sequentially to eliminate redundant features and compress dimensionality. A robust XGBoost regression model is subsequently trained on the optimized feature set for capacity estimation, with SHapley Additive exPlanations (SHAP) values systematically quantifying feature contributions to model decisions. Experimental validation on NASA's lithium-ion battery dataset demonstrates the framework's superiority. The results show that the proposed method can significantly improve the accuracy of capacity estimation, and the root mean square error and average absolute error of the leave-one-out cross-validation method (LOOCV) are less than 2%.

lithium-ion battery, multi-source feature fusion, XGBoost, capacity estimation, interpretability