Physics-informed unsupervised battery state-of-charge estimation via temperature-conditioned Mamba and parallel Bayesian inference

In this paper, we propose Mamba-IPLS, a physics-informed, unsupervised algorithm for offline, label-free battery state-of-charge (SOC) trajectory reconstruction. The architecture integrates a Mamba backbone with a Temperature-Conditioned Modulator (TCM) to generate model parameters under different ambient temperature conditions. For a sequence of length N , this Mamba–TCM neural parameter generator requires O ( N ) total work. The generated parameters drive a physical state–space model, where the latent battery states are estimated using a Parallel Iterated Posterior Linearization Smoother (IPLS). Each IPLS iteration also requires O ( N ) total work, while its associative filtering and smoothing scans have O ( log N ) parallel span. Comparative analysis within the unified Deep Neural Network (DNN)-IPLS architecture first reveals that the Mamba backbone exhibits superior batch-processing speed and stability compared to Transformer and long short term memory variants, particularly under constrained parameter budgets. Building on this optimized architecture, experimental validation across diverse driving cycles and temperatures (−10 °C to 50 °C) demonstrates that the proposed algorithm significantly outperforms state-of-the-art data-driven baselines in estimation accuracy while requiring substantially fewer trainable parameters. Comprehensive ablation studies further confirm that the TCM and IPLS modules improve the robustness of SOC estimation.