Uni-Mamba: A unified framework for cross-material and cross-condition battery state-of-health estimation via domain-adaptive state space models

Accurate state-of-health (SOH) estimation across different battery chemistries and operating conditions remains challenging because material properties, cycling protocols, and long-term temporal evolution strongly influence degradation features extracted from measurable signals. These differences lead to substantial distribution shifts between source- and target-domain batteries, thereby limiting the cross-domain transferability of conventional data-driven models. To address this issue, this paper proposes Uni-Mamba, a cross-domain SOH estimation framework that integrates health-feature representation, hierarchical temporal modeling, and subdomain distribution alignment. First, statistical, curve-variation, and entropy-based health features are derived from voltage and current measurements to characterize complementary aspects of degradation. A Hybrid Causal-Attention Block (HCAB), consisting of multi-scale causal temporal convolutions and low-rank bilinear feature projection, is then employed to capture local degradation dynamics and long-range feature dependencies. On this basis, a Mamba backbone based on a selective state-space model is introduced to model historical degradation sequences while maintaining favorable computational efficiency for long-sequence inputs. Under the unsupervised domain adaptation setting, Local Maximum Mean Discrepancy (LMMD) is further employed to align the feature distributions of labeled source-domain samples and unlabeled target-domain samples within corresponding aging subdomains. Extensive experiments are conducted on six public battery datasets using multiple random seeds, together with systematic ablation and comparative studies. The results demonstrate that Uni-Mamba achieves accurate and stable SOH estimation under both cross-condition and cross-domain settings, with the mean absolute error (MAE) remaining below 3.1% across all evaluated cross-domain target batteries.