Recent BMS-relevant papers with a practical engineering inspection lens—not a publication-quality score.
Paper summaries are discovery records. Dataset, method and reproducibility must be studied before a result supports engineering work.
IEEE ICETCI 2026
The paper describes an MoE-LSTM framework using large-scale real-vehicle charging data and a tiered feature system, with robustness tests under additive noise.
Promising for cloud health diagnostics, but the engineering review should inspect vehicle diversity, label construction, segment leakage, unseen-chemistry performance and compute cost.
International Journal of Electrochemical Science
The study compares multiple machine-learning methods and feature groups, reporting that feature choice changes model performance and that some models remain robust when noisy features are introduced.
Useful for feature-downselection, but transfer across temperature, duty cycle, chemistry and sensor quality remains the decisive production question.
Mathematics
The review organizes SOH methods around learning formulation and highlights uncertainty, calibration, distribution shift, robustness and physical validity.
This is directly aligned with a AI4BMS evaluation checklist for deciding whether an ML result is ready for engineering use.
Cell Reports Physical Science
The work extracts changes in time-series characteristics and feeds them to an XGBoost regressor for SOH prediction on an experimental NMC/graphite aging dataset.
The approach may reduce feature complexity; deployment review should test sensitivity to sampling, drive-cycle excitation, partial charging and cell-to-pack variability.
Scientific African
The paper uses a CNN-BiLSTM-attention pipeline, pre-training on NASA aging data and adapting selected layers to second-life residential microgrid data.
The explicit attention to distribution shift and evaluation leakage is valuable; teams should still inspect module history, target-data volume and uncertainty at operating boundaries.
Journal of Power Sources
The work proposes a model-based power-SOH estimator built around surface-temperature behavior during charging, motivated partly by independent battery-passport monitoring.
It suggests an independent observation path, but assumptions around charging profile, thermal boundary conditions, sensor placement and cell construction must be challenged.
Journal of Power Sources
The paper investigates an acoustic, deep-learning-based detection model intended to identify early-stage thermal runaway before conventional signals become decisive.
Potentially useful as an independent safety channel, but sensor integration, nuisance events, pack acoustics, aging and diagnostic coverage require extensive validation.