Abstract
Accurate Remaining Useful Life (RUL) prediction for Lithium-ion batteries is critical for system safety, yet its efficacy is frequently limited by data scarcity in industrial contexts. The robustness of hybrid architectures combining Convolutional Neural Networks (CNNs) with sequential models, a potential solution, has not been systematically evaluated. This study addresses this knowledge gap by first using a CNN to derive low-dimensional feature representations from full charge-discharge cycles. We then systematically assess the performance of five prominent sequential models (GRU, LSTM, Transformer, Neural ODE, and Transformer (Pre-LN)) on these features under progressively severe data scarcity (0%, 10%, 30%, and 50% cycle removal). Based on leave-one-out cross-validation on the NASA and CALCE datasets, the analysis demonstrates that the CNN-based feature extraction significantly enhances the robustness of all tested sequential models. Furthermore, recurrent networks such as GRU and LSTM, possessing strong sequential inductive biases, consistently outperform more complex architectures under data-constrained conditions. This research validates a robust predictive methodology and provides practical insights for developing reliable RUL predictors for industrial applications where data is sparse.