Domain-Adaptive Mixture-of-Experts for Cross-Dataset Lithium-Ion Battery State-of-Health Prediction via Adaptive Strategy Selection

Abstract

Accurate cross-dataset state-of-health prediction for lithium-ion batteries remains challenging due to distribution shifts arising from diverse cathode chemistries, operating temperatures, and charge–discharge protocols across heterogeneous battery fleets. Drawing upon established machine learning paradigms, this study tailors a Domain-Adaptive Mixture-of-Experts (DA-MoE) framework to the battery prognostic context, automatically selecting the optimal domain adaptation strategy for each target domain through a physics-aware, lightweight linear gating network comprising merely 32 learnable parameters. The framework integrates a shared Transformer-based backbone with four adaptation strategies spanning the full spectrum of target-domain information utilization, namely zero-shot transfer, Test-Time Adaptation, Fine-Tuning, and Model-Agnostic Meta-Learning. A comprehensive evaluation on 564 battery cells from seven publicly available datasets under Leave-One-Domain-Out Cross-Validation protocol demonstrates that the proposed framework achieves an average coefficient of determination of 0.864 with perfect oracle strategy alignment under full domain training and maintains competitive generalization at an average R2 of 0.795 when each target domain is held out during gating network training. Hard argmax selection consistently outperforms weighted fusion across all seven domains with an average margin of +0.027 in R2, confirming that the four adaptation strategies compete rather than cooperate in this application context. A feature ablation analysis identifies sample count as the dominant determinant of strategy selection with performance degradation of ΔR2 = −0.182 upon removal, followed by the early-cycle degradation slope and early-cycle nonlinearity index as secondary signals, all of which are computable at deployment time without future ground-truth SOH information. The proposed framework provides a practically deployable solution for battery management systems operating across heterogeneous fleets with minimal computational overhead and strong cross-dataset generalization capability.