PaperlayerView exact revision on arXiv

Source-linked AI summary

PhyMamba: Physics-Modulated Mamba for Robust Battery Health Prognostics

Sara Sameer, Yunyi Zhao, Wei Zhang, Minggang Zeng, Wenqing Li, Man-Fai Ng, Yonggang Wen

arXiv:2608.27978v1cs.LGeess.SY

TL;DR

Long-horizon SoH forecasting from BMS signals is difficult under operating-condition variability and sensor noise, while explicit electrochemical parameter identification can require intrusive measurements. PhyMamba addresses this with a two-stage physics-modulated Mamba that learns aging features and regulates forecasting dynamics, achieving a 31.8% overall mean MAE reduction against diverse ML baselines across three public datasets and multiple horizons.

  • Problem

    Long-horizon battery health forecasting from BMS signals is challenging because operating conditions and sensor noise affect prognostic modeling, while electrochemical parameter identification can require intrusive measurements.

  • Method

    PhyMamba uses a lightweight Mamba encoder and aging-parameterization module to construct physics-informed features, then injects them into a Mamba forecasting backbone’s internal state-space dynamics.

  • Results

    31.8% overall mean error reduction in MAE against diverse ML baselines across three public datasets and multiple forecast horizons.

  • Takeaways & Limitations

    PhyMamba achieves the best aggregated performance while balancing prognostic accuracy and model size for battery health prognostics.

Abstract

from arXiv · show

Battery health prognostics is a core function in battery management systems (BMSs), yet long-horizon health forecasting from BMS signals remains challenging due to operating-condition dependency and sensor noise. In this paper, we propose PhyMamba, a two-stage physics-modulated Mamba framework that integrates electrochemical aging into sequence modelling. PhyMamba does not require explicit identification of internal aging parameters, which often relies on intrusive measurements. In stage-1, a lightweight Mamba encoder first processes BMS signals and produces a latent representation that is transformed via an aging parameterization module, into physics-informed aging features. In stage-2, a customized Mamba forecasting backbone performs multi-cycle prediction, where physics is tightly integrated to regulate the model's internal temporal updates toward degradation-consistent evolution. Experiments on three public datasets under multiple forecast horizons show that PhyMamba achieves the best aggregated performance, with an overall mean error reduction of 31.8% compared with a diverse range of baselines. PhyMamba also offers an optimized accuracy-efficiency trade-off, which supports practical deployment for robust battery health prognostics.

I. INTRODUCTION

Battery health prognostics supports BMS monitoring, but existing data-driven and physics-based approaches face complementary limitations. PhyMamba addresses this gap by integrating electrochemical aging physics into Mamba’s internal sequence updates for efficient long-horizon SoH forecasting.

  • Motivation: Battery health prognostics is a core BMS function for monitoring degradation and supporting reliable battery operation.BMSs continuously measure signals such as voltage and current and produce state estimates.
  • Motivation: SoH is defined as the ratio of current usable capacity to rated capacity, with EoL commonly associated with SoH below 80%.
  • Related Work: Existing prognostic models are largely data-driven, learning mappings from BMS signals or derived features to health indicators.Representative approaches include regression, LSTM, Transformer-based, diffusion, and time-frequency neuromorphic models.
  • Related Work: Electrochemical models offer mechanistic fidelity, but parameter extraction and real-time integration can require intrusive measurements and expensive computation.Consequently, electrochemical analysis is often performed offline under calibrated settings.
  • Contribution: PhyMamba uses electrochemical aging physics to regulate Mamba’s internal state evolution instead of appending aging parameters as ordinary inputs.This design exploits Mamba’s explicit state-evolution mechanism to encourage degradation-consistent cycle updates.
  • Contribution: PhyMamba is a two-stage framework that constructs physics-informed aging features before performing multi-cycle SoH forecasting with a conditioned Mamba backbone.The paper evaluates the framework across three public datasets and multiple forecast horizons.

A. Physics-Guided Learning for Battery SoH Forecasting

Physics-guided battery prognostics combines operational measurements with degradation knowledge, while sequence models provide efficient long-context processing. PhyMamba extends this direction by embedding aging information directly into selective state-space dynamics.

  • Physics-Guided Learning: Model-based battery prognostics provides physical interpretability but often requires accurate parameter identification and detailed electrochemical measurements unavailable from standard BMS logs.
  • Physics-Guided Learning: Data-driven methods learn degradation patterns directly from operational voltage, current, and temperature measurements.They include classical machine learning, deep neural networks, and Gaussian-process regression for SoH and RUL prediction.
  • Physics-Guided Learning: Physics-informed learning incorporates degradation knowledge through constrained losses, physically meaningful latent variables, or hybrid battery models.
  • Research Gap: Many existing physics-informed approaches use physics mainly as an external regularizer or auxiliary loss.
  • SSMs and Mamba: SSMs encode sequence history in a latent state whose transitions and skip connections control information retention and propagation.This structure suits continuous signals and long-range time-series dependencies.
  • SSMs and Mamba: Mamba combines input-dependent selective state-space dynamics with recurrent processing and linear sequence-length scaling.Its recurrent hidden state avoids a growing key-value cache during autoregressive inference.
  • PhyMamba: PhyMamba extracts an electrochemical aging representation from BMS observations and embeds it into selective SSM dynamics rather than using physics only as auxiliary input.This retains Mamba’s efficiency while regulating temporal updates according to degradation information.

III. PHYMAMBA: METHODOLOGY

The methodology section introduces PhyMamba as a two-stage framework for constructing physics-informed representations and forecasting battery health.

  • III. PHYMAMBA: METHODOLOGY: PhyMamba comprises problem formulation, physics-informed aging feature construction, and physics-modulated forecasting.
  • III. PHYMAMBA: METHODOLOGY: The framework’s methodology is organized around integrating aging information into the forecasting process.
  • III. PHYMAMBA: METHODOLOGY: The section presents the proposed framework rather than a standalone data-driven forecasting model.

A. Battery Health Prognosis and System Overview

Battery prognosis models use historical BMS time series to estimate current health and predict future degradation. PhyMamba implements this task with a two-stage architecture that first extracts structured aging features and then forecasts SoH over a future horizon.

  • A. Battery Health Prognosis and System Overview: Battery health models process raw or derived BMS features from multiple charge/discharge cycles and use historical health targets such as capacity or SoH.
  • Problem Formulation: At cycle c, the model input X_c contains a fixed-length window of T past cycles with feature dimension d.
  • Problem Formulation: The forecasting task predicts SoH for the next H cycles, from c+1 through c+H.
  • Problem Formulation: Model parameters are learned by minimizing a loss between predicted and ground-truth SoH values.Mean squared error is given as an example loss.
  • System Overview: PhyMamba first extracts a compact physics-informed representation and then supplies it to a second stage for multi-cycle horizon forecasting.

B. Physics-based Aging Feature Construction

PhyMamba constructs physics-informed aging features from BMS logs without directly measuring internal electrochemical states. A lightweight Mamba encoder infers BoL-aligned latent parameters, which a differentiable aging model evolves into features for forecasting.

  • BMS measurements indirectly reflect internal aging, while electrochemical parameters such as SEI thickness and electrode surface area often require intrusive characterization.
  • Stage-1 encodes voltage, current, and time-related measurements with lightweight Mamba layers and maps the representation to six inferred BoL parameters.
  • The aging module represents each channel as a BoL baseline minus accumulated degradation driven by SoH change and cycle progression.
  • Learnable occurrence degrees weight different aging phenomena as multiplicative contribution factors while preserving a shared physics-motivated degradation form.
  • The resulting six-channel vector is passed to forecasting as a physics-informed representation learned from BMS data rather than direct electrochemical-state measurements.

C. Forecasting with Physics-Modulated Mamba

The forecasting stage uses aged electrochemical features to modulate Mamba’s selective state-space dynamics. Physics-conditioned updates adapt memory, writing, and readout behavior to degradation severity while producing multi-cycle SoH predictions.

  • The forecasting module maps the current aged feature vector to future health predictions over horizon H.
  • The forecasting backbone uses Mamba’s linear sequence-length scaling and content-dependent selection for long cycle histories.
  • Small step sizes retain memory, whereas large step sizes update states faster and discount older information.
  • The physics-modulated step size adds α|u_t| to the standard Mamba pathway, making temporal updates depend on the degradation-conditioned representation.
  • Stronger degradation increases effective write strength, while the readout adapts to different degradation compositions and a residual path directly carries aged physics features to the output.
  • PhyMamba injects aging features into Mamba’s internal state-space dynamics rather than treating them as ordinary input variables.

IV. EXPERIMENTAL STUDY

The experimental section evaluates PhyMamba on public battery degradation datasets, comparing baselines and conducting ablation studies.

  • The evaluation covers public battery degradation datasets, baseline comparisons, and ablation studies.

A. Experimental Setup

Experiments use three public battery degradation benchmarks with cellwise data splits and history-window samples for horizon-based forecasting. Inputs contain voltage, current, and timestamps, and training uses end-to-end optimization with early stopping.

  • Experiments use the MATR, NASA, and CALCE public battery degradation benchmarks.
  • Input samples use a fixed history window of cycle-indexed BMS logs, with tensor shape X ∈ R^128×100×3 for voltage, current, and timestamps.
  • Cellwise splitting keeps batteries disjoint across training, validation, and test sets.
  • Forecast horizons are configured by experiment, including H = 10 and H = 30.
  • Training uses Adam with learning rate 0.01 and weight decay 1 × 10^-5 for up to 500 epochs with patience-30 early stopping.

B. Comparison Study

PhyMamba is evaluated across three public datasets and multiple forecast horizons, where it delivers the strongest overall and consistently competitive performance. Its gains extend beyond vanilla Mamba and remain evident against diverse comparison models.

  • PhyMamba achieves the lowest overall mean MAE of 1.26 and mean MSE of 0.09 across the evaluated datasets and horizons.
  • PhyMamba remains consistently competitive across MATR, NASA, and CALCE rather than excelling only on one benchmark or forecast horizon.
  • 98.2%, 30.5%, and 81.4% are the higher mean MAEs of LSTM, iTransformer, and SimpleTM compared with PhyMamba, respectively.Their mean MAEs are 2.50, 1.65, and 2.29, respectively.
  • 18.0% and 16.3% are the higher overall mean MAE and MSE of vanilla Mamba relative to PhyMamba, respectively.Vanilla Mamba records overall mean MAE 1.49 and MSE 0.10.
  • 2.3% is TimeMixer’s higher mean MSE than PhyMamba, while its mean MAE is 20.4% higher.

C. Ablation Studies of PhyMamba

Ablation studies show that both the two-stage design and physics modulation improve multi-cycle SoH forecasting. The framework also exposes accuracy-efficiency trade-offs among alternative encoders and backbones.

  • 21.3% is the MAE degradation when the physics-modulated forecasting module is replaced by a vanilla Mamba backbone.The MAE increases from 0.24 to 0.30, indicating additional benefit from stage-2 physics modulation.
  • 2.3× is the single-stage baseline’s MAE relative to full PhyMamba, with errors increasing to 0.56.The comparison tests the value of learning a physics-informed representation before forecasting.
  • Vanilla Mamba without stage-2 physics modulation has only around one-quarter of PhyMamba’s efficiency across the reported settings.
  • LSTM can work for short horizons but degrades for longer horizons, whereas iTransformer is more horizon-stable but much larger and less efficient.
  • PhyMamba primarily uses Mamba in both stages to reduce implementation complexity, while hybrid encoder designs remain possible but are outside the paper’s focus.

D. ECM-based Versus Aging Model Physics

The study compares ECM-based physics with electrochemical aging physics in battery prognostics. PhyMamba consistently outperforms the selected ECM-based baseline PACE on the MATR dataset at the tested horizons.

  • ECM variables have explicit physical meaning and can be constructed accurately from BMS signals, while aging parameters more directly describe electrochemical degradation but are harder to reconstruct.
  • PhyMamba therefore learns latent physics representations because reconstructing electrochemical aging parameters from BMS data is complex, coupled, and potentially lower-fidelity.
  • PhyMamba consistently outperforms PACE on MATR for forecast horizons H = 30 and H = 50.PACE is selected as a strong ECM-based baseline because its public implementation performs competitively against mainstream methods.

V. CONCLUSION

PhyMamba addresses reliable, efficient long-horizon battery health forecasting under operating-condition variability through a two-stage physics-guided design. Across three public datasets and multiple horizons, it achieves the best aggregated performance while balancing accuracy and model size.

  • PhyMamba targets reliable long-horizon battery health forecasting under condition variability while remaining efficient for BMS deployment.
  • The two-stage design uses aging-informed features to provide structured degradation information to a forecasting backbone.
  • Across three public datasets and multiple horizon settings, aging-informed representation improves robustness relative to direct end-to-end learning from BMS terminal signals.
  • Tightly coupling physics representation with forecasting dynamics yields additional gains beyond using physics as a static conditioning signal.
  • 31.8% overall mean MAE reduction gives PhyMamba the best aggregated performance against diverse machine-learning baselines while preserving an accuracy–model-size balance.
Loading 2608.27978v1…