Domain Generalized State of Health Estimation without Target Domain Data Based on Feature Disentanglement

Accurate State of Health (SOH) estimation is essential for the safe and stable operation of lithium-ion batteries in battery management systems (BMS). However, varying operating conditions (e.g., ambient temperatures and charging rates) lead to data distribution shifts, posing challenges for cross-domain SOH estimation. Most existing cross-domain SOH estimation methods rely on either labeled or unlabeled target data to improve generalization, increasing the storage burden on the BMS. To address this issue, this study proposes the Feature Disentangled Transformer-CNN Network (FD-TCNet), which employs a dual-branch architecture consisting of a CNN branch for local feature extraction and a Transformer branch for long-range dependency modeling. A feature disentanglement mechanism is further introduced through explicit task-specific constraints and mutual information minimization. This mechanism separates domain-invariant health features from condition-related interference features. As a result, the health-related features maintain consistent distributions across varying operating conditions, enabling cross-domain SOH estimation without target-domain data. Compared with the baseline methods, the proposed method achieves the maximum observed improvements, with RMSE reduced from 4.16% to 1.53% (63.15% reduction) and MAE reduced from 3.56% to 1.37% (61.46% reduction).