Pith. sign in

REVIEW 3 major objections 5 minor 38 references

A battery-safety model that watches force as well as heat can warn of thermal runaway an average of 15.6 seconds before onset, more than doubling the lead time of temperature-only baselines.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 14:06 UTC pith:ECLIVH3P

load-bearing objection Worth engaging: real dataset and an interesting force-channel result, but the headline metrics as written are internally inconsistent and need correction before the numbers can be trusted. the 3 major comments →

arxiv 2607.18860 v1 pith:ECLIVH3P submitted 2026-07-21 cs.LG cs.AI

Regime-Aware Physics-Guided Early Warning of Lithium-Ion Battery Thermal Runaway Using Thermo-Mechanical Signals

Syed Sajid Ullah , Muhammad Zunair Zamir , Salman Khan This is my paper
classification cs.LG cs.AI
keywords battery thermal runawayearly warningphysics-guided neural networkstemporal convolutional networkfeature-wise linear modulationmulti-task learningleave-one-experiment-out cross-validationthermo-mechanical signals
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tries to prove that fusing mechanical signals—force, deformation—with temperature, voltage, and state of charge in a regime-aware neural network can provide earlier and more reliable thermal-runaway warnings under mechanical abuse. It reports a mean warning lead time of 15.6 seconds, a detection success rate of 0.92, and an experiment-level false-alarm rate of 2.7% across 30 controlled abuse tests. The load-bearing claim is that force is the critical precursor: removing it cuts lead time by 60.3%. A sympathetic reader would see this as evidence that battery safety systems should treat mechanical sensing as a first-class warning channel, not an afterthought.

Core claim

The paper's central claim is that early warning for lithium-ion thermal runaway under mechanical abuse is substantially improved by a two-stage, physics-guided model that first infers the battery's safety regime (safe, warning, danger) from force and deformation signals, then uses those regime estimates to condition a causal temporal convolutional network that ingests temperature, voltage, force, deformation, and state of charge. The joint model predicts regime class, the probability of thermal runaway within 60 seconds, and the remaining time to disaster. Across 30 leave-one-experiment-out folds, it achieves a mean warning lead time of 15.6 seconds with a detection success rate of 0.92, a f

What carries the argument

The key mechanism is the regime-aware conditioning cascade: a lightweight convolutional classifier (Stage 1) reads only the force and deformation channels and outputs safe/warning/danger probabilities; those probabilities then modulate a causal dilated TCN backbone through regime-dependent multiplicative gating. The backbone also receives state-of-charge through feature-wise linear modulation (FiLM), a per-channel scaling and shifting of hidden features, and uses physics-biased attention that adds indicator masks for temperature and force channels to the attention logits. The three task heads are trained jointly with a weighted loss, forcing a single representation to serve regime identifica

Load-bearing premise

Every reported metric rests on defining thermal-runaway onset as the moment surface temperature reaches 150 °C or rises faster than 3 °C/s, and safety regimes as fixed 60/120 °C thresholds; if these thresholds do not correspond to the point of irreversible failure in these abuse tests, the 15.6-second lead time is time-to-threshold, not time-to-actual-runaway.

What would settle it

Take the trained model to an independent mechanical-abuse dataset where the true thermal-runaway onset is labeled by direct evidence of internal short circuit, venting, or rupture rather than by temperature thresholds; if the model's warnings at the claimed 15.6-second lead time do not precede those direct failure markers, or if false-alarm rates jump when measured against that ground truth, the central claim collapses.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If the claim holds, battery management systems operating at 1–10 Hz sampling can act on a 15-second warning to trigger cooling, disconnection, or occupant alert before thermal runaway becomes irreversible.
  • Force and deformation sensors should be treated as primary inputs in abuse-prone battery installations, since omitting force reduces warning lead time by more than half.
  • The regime-aware architecture shows a way to suppress false alarms by gating out anomalous patterns while the cell is in a safe thermal regime.
  • The monotonic improvement with state of charge suggests that low-SOC thermal runaway is intrinsically harder to predict; systems should allocate more sensing or accept shorter lead times there.
  • The reported 69.6% lead-time improvement over the best baseline indicates that existing temperature-centric warning pipelines are leaving substantial warning margin on the table under mechanical abuse.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The temperature-threshold definition of thermal-runaway onset may understate or overstate the model's true operational lead time: if mechanical signals predict irreversible failure earlier than the 150 °C or 3 °C/s threshold, the real time-to-rupture could be longer; if those thresholds are earlier than actual failure, the headline lead time would not transfer to real-world events.
  • Force dominance is likely specific to mechanical-abuse scenarios; under overcharge, external heating, or field aging, voltage, gas, or impedance could be equally or more informative, meaning the architecture's channel conditioning should be retrained per abuse mode rather than assumed universal.
  • The 30-experiment dataset with 20 thermal-runaway events leaves wide confidence intervals; a practical deployment path would need online calibration and uncertainty-aware thresholds, especially for low-SOC events where the model already shows weaker performance.
  • A testable extension is to use the Stage-1 regime predictions as a standalone precursor detector: if force-based safe/warning/danger transitions consistently precede the temperature-defined regime boundaries, that would confirm mechanical signals carry independent early-warning information.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes a two-stage, regime-aware, physics-guided neural framework for early warning of lithium-ion battery thermal runaway under mechanical abuse. Stage 1 is a lightweight CNN that infers safe/warning/danger regimes from force and deformation; Stage 2 is a causal TCN backbone conditioned on SOC via FiLM, physics-biased attention, and regime gating, with three outputs: regime logits, TR probability within a 60 s horizon, and time-to-disaster. Evaluation uses leave-one-experiment-out cross-validation on 30 mechanical-abuse tests (20 TR, 10 non-TR) at 2 Hz sampling. The reported headline results are F1=0.89, high-temperature prediction RMSE=12.3°C, mean warning lead time 15.6 s, DSR=0.92, and experiment-level FAR=2.7%, with an ablation showing that removing the force channel reduces lead time by 60.3%.

Significance. The topic is important and the general design is promising: fusing force/deformation with thermal and electrical signals is a sensible route to earlier warning under mechanical abuse, and the paper explicitly avoids label leakage by defining ground-truth regimes from temperature only (Section III). The leave-one-experiment-out protocol, the causal architecture, the lightweight deployment profile (156K parameters, 22 ms inference), and the force-channel ablation are tangible strengths. However, several headline metrics in Table III and Section V.F are internally inconsistent or architecturally unsupported: the FAR values are not multiples of 1/30 under the stated definition, DSR=0.92 is impossible for 20 TR experiments if it is a per-experiment detection rate, and the reported temperature RMSE has no corresponding output head in Section IV. The significance of the empirical claims is therefore conditional on correcting and re-auditing the evaluation section.

major comments (3)
  1. [Section V.F, Table III, Abstract] The reported FAR and DSR values are arithmetically inconsistent with the stated evaluation protocol. Section V.F defines FAR as the percentage of held-out experiment folds with at least one false warning episode; with 30 folds, any such percentage must be an integer multiple of 1/30 ≈ 3.33%. The reported 2.7%, and indeed all baseline FAR values (10.7%, 9.3%, 7.0%, 6.3%), are impossible under this definition. Likewise, with only 20 TR experiments, a per-experiment detection success rate can only be a multiple of 0.05, so 0.92 (i.e., 18.4 events) cannot be a per-experiment DSR. If DSR and FAR are instead window-level or episode-level statistics, that must be stated explicitly and the metrics renamed or redefined, because the abstract and Table III present them as experiment-level results. This directly affects the credibility of the lead-time claim.
  2. [Section IV.C/D, Section V.F, Table III] The reported temperature-prediction RMSE (12.3°C, Table III) has no corresponding output head in the architecture. Section IV describes three outputs: regime logits (f_r), TR detection probability (f_d), and time-to-disaster regression (f_t). None of these predicts a temperature trajectory; f_t outputs a scalar in seconds, not degrees Celsius. Yet Section V.F says RMSE and MAE are reported in degrees Celsius for the predicted high-temperature trajectory, and Table III lists RMSE for all models. Either the architecture must include a temperature-prediction head with an appropriate loss, or the RMSE/MAE/R² claims must be removed. As written, the second headline metric is architecturally undefined.
  3. [Section III, Eq. (1)-(2), Section VII] The ground-truth definition of thermal-runaway onset and the safety regimes uses fixed thresholds (T_TR=150°C, dT/dt=3°C/s, T_safe=60°C, T_danger=120°C) applied identically across SOC levels and loading protocols. The paper states these are canonical, but the operational meaning of 'lead time' is time-to-threshold, not necessarily time-to-irreversible failure. If these thresholds are not calibrated to the actual onset of irreversible runaway in these abuse tests, the 15.6 s headline would not transfer to a real BMS. At minimum, the authors should provide a sensitivity analysis of lead time and DSR/FAR to these thresholds, or explicitly justify the fixed thresholds with experimental evidence for the tested cells.
minor comments (5)
  1. [Section IV.C.2] The receptive field is stated as 61 time steps, 'spanning approximately 10 s of history.' With the stated 2 Hz sampling rate, 61 time steps correspond to 30.5 s, not ~10 s. Please correct the time calibration.
  2. [Fig. 11 vs. Section VI/Abstract] The per-fold lead-time distribution in Fig. 11 reports mean = 15.4 s, while the text, abstract, and Table III report 15.6 s. Please reconcile the exact mean value across the 20 TR experiments.
  3. [Section V.E and Introduction] The passage stating 'a same-input reimplementation is planned before journal resubmission' is editorial and self-referential, and should be removed. It also indicates that a potentially relevant companion-study comparison is missing; please either include the comparison or describe it in neutral, non-submission-specific language.
  4. [Section V.F] Detection success rate (DSR) is never formally defined. Please state whether DSR is the fraction of TR events detected, the fraction of TR windows correctly classified, or another quantity, and define the unit of analysis for accuracy, F1, RMSE, and MAE (windows vs. experiments).
  5. [Eq. (22)] The indicator function in the regime classification loss is written '⊮[r=c]' instead of '1[r=c]' or 'I[r=c]'. This is a typographical issue but should be corrected for clarity.

Circularity Check

0 steps flagged

No circularity: the paper is an empirical supervised-learning study whose labels and metrics are defined externally by temperature thresholds, and no claim reduces to its inputs.

full rationale

The paper does not claim a first-principles derivation. Thermal-runaway onset (Eq. 1) and safety-regime labels (Eq. 2) are defined a priori from temperature thresholds (150 °C, 3 °C/s, 60 °C, 120 °C); force and deformation are explicitly described as model inputs only and are not used to define ground truth, preventing label leakage. The Stage 1 regime classifier learns to map mechanical signals to these temperature-defined regimes, which is a supervised prediction task rather than a self-definitional step. The TTD target is the remaining time until the externally defined onset, and the lead-time metric is the standard difference between the model's first correct warning and that onset. No fitted parameter is renamed as a prediction: DSR, FAR, lead time, and RMSE are all reported on held-out LOEO folds, and ablations are empirical comparisons. The 'physics-guided' components (SOC-FiLM, attention bias, gating) are learned architectural conditioning mechanisms, not equations that reduce to their outputs. The paper's self-referential note about a companion study is explicitly non-load-bearing and is not used to justify any result. Concerns about arithmetic consistency of DSR/FAR and the undefined temperature-prediction head are internal-consistency or correctness issues, not circularity, and are outside the scope of this pass. The derivation chain is self-contained as an empirical evaluation.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The central claim rests on temperature-threshold definitions, hand-set hyperparameters, and a private 30-experiment dataset. The model itself introduces no new physical entities, but the term 'physics-guided' is supported only by learned conditioning and attention biases, not by physical equations. Because no code/data are released, the fitted model cannot be audited.

free parameters (5)
  • TR onset criteria (T_TR, dT/dt_TR) = 150 °C, 3 °C/s
    Eq. (1); defines t_TR and hence every lead-time/DSR/FAR metric; chosen from literature, not fitted, but arbitrary within a range.
  • Regime thresholds (T_safe, T_danger) = 60 °C, 120 °C
    Eq. (2); creates ground-truth regime labels used to train Stage 1 and gate Stage 2; fixed across all experiments by assertion.
  • Prediction horizon H = 60 s
    Section III; labels for TR detection ('TR within H seconds'); directly controls detection difficulty and implied lead time ceiling.
  • Warning threshold = 0.5
    Section IV-F; TR warning issued when predicted probability ≥ 0.5; t_warn and thus lead time depend on this scalar.
  • Loss weights (λ1, λ2, λ3) and positive-class weight = 0.1/0.6/0.3, 10
    Eq. (21)-(23); chosen by hand; affect the trade-off between regime accuracy, detection, and TTD regression.
axioms (4)
  • domain assumption Temperature thresholds in Eq. (1)-(2) provide a valid, universal definition of thermal-runaway onset and safety regimes under mechanical abuse.
    All labels, lead time, and success/failure metrics are computed from these thresholds; no independent physical validation is given for the abuse tests studied.
  • domain assumption The 30-experiment dataset (20 TR, 10 non-TR) is representative enough for leave-one-experiment-out cross-validation to yield stable generalization estimates.
    LOEO on 30 experiments leaves one experiment per fold; the authors themselves note limited statistical power; per-fold metrics can swing by one experiment.
  • domain assumption Force and deformation are causally prior to temperature rise in these abuse tests, so they can act as precursors rather than correlates.
    Ablation shows in-sample importance, but no experiment or physics demonstrates that force always precedes temperature in uncontrolled field abuse.
  • domain assumption SOC is perfectly known and constant within each experiment; the SOC scalar used for FiLM conditioning is accurate.
    SOC values are approximate (10/50/90%); estimation error in field BMS would degrade the conditioning.

pith-pipeline@v1.3.0-alltime-deepseek · 14296 in / 14517 out tokens · 126691 ms · 2026-08-01T14:06:22.266289+00:00 · methodology

0 comments
read the original abstract

Thermal runaway in lithium-ion batteries poses a major safety risk to electric vehicles and energy storage systems. Current early-warning methods depend mainly on temperature and may therefore miss mechanical precursors that emerge before rapid heating. We introduce a regime-aware, physics-guided framework that integrates temperature, voltage, force, deformation, and state-of-charge measurements for early warning under controlled mechanical abuse. A lightweight convolutional classifier first infers safe, warning, or danger regimes from mechanical signals. These regime estimates then condition a causal temporal convolutional backbone through feature-wise linear modulation, physics-biased attention, and regime-dependent gating. Joint learning unifies regime identification, thermal-runaway detection, and time-to-disaster estimation. We evaluate the framework using leave-one-experiment-out cross-validation on 30 mechanical-abuse tests across state-of-charge levels of 10%, 50%, and 90% and two loading protocols. The method achieves an F1 score of 0.89, a high-temperature prediction root-mean-square error of 12.3 {\deg}C, a mean warning lead time of 15.6 s, a detection success rate of 0.92, and an experiment-level false alarm rate of 2.7%. Its lead time exceeds that of the strongest baseline by 69.6%. Removing force reduces the lead time by 60.3%, highlighting the value of mechanical precursors. These results support regime-aware thermo-mechanical fusion as a promising strategy for earlier and more reliable thermal-runaway warning under controlled abuse conditions.

Figures

Figures reproduced from arXiv: 2607.18860 by Muhammad Zunair Zamir, Salman Khan, Syed Sajid Ullah.

Figure 1
Figure 1. Figure 1: Timeline of thermal progression through the safety regimes to TR [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overall architecture of the proposed regime-aware, physics-guided TR early-warning framework. Stage 1 uses a lightweight CNN to estimate the [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 5
Figure 5. Figure 5: Experimental dataset overview, including SOC distribution, experiment [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Representative thermo-mechanical signal evolution during a thermal [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 4
Figure 4. Figure 4: Representative lithium-ion cells after mechanical-abuse-induced [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 7
Figure 7. Figure 7: Temperature and force trajectories across SOC conditions (90%, 50%, [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Overview of the preprocessing and window-generation pipeline. [PITH_FULL_IMAGE:figures/full_fig_p009_8.png] view at source ↗
Figure 10
Figure 10. Figure 10: compares warning lead time across models, while [PITH_FULL_IMAGE:figures/full_fig_p009_10.png] view at source ↗
Figure 10
Figure 10. Figure 10: Mean warning lead-time comparison across the proposed framework [PITH_FULL_IMAGE:figures/full_fig_p010_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Distribution of per-fold lead times for the proposed method across [PITH_FULL_IMAGE:figures/full_fig_p010_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Ablation analysis of lead-time and RMSE changes. [PITH_FULL_IMAGE:figures/full_fig_p011_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Heat-map summary of Wilcoxon signed-rank p-values comparing [PITH_FULL_IMAGE:figures/full_fig_p012_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Per-experiment lead time versus SOC under different loading [PITH_FULL_IMAGE:figures/full_fig_p012_14.png] view at source ↗
Figure 16
Figure 16. Figure 16: Actual-versus-predicted diagnostic plots for assessing temporal [PITH_FULL_IMAGE:figures/full_fig_p013_16.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

38 extracted references · 3 linked inside Pith

  1. [1]

    A review of early warning methods of thermal runaway of lithium ion batteries,

    D. Kong, H. Lv, P. Ping, and G. Wang, “A review of early warning methods of thermal runaway of lithium ion batteries,”Journal of Energy Storage, vol. 64, p. 107073, 2023

    2023

  2. [2]

    Mitigation strategies for li-ion battery thermal runaway: A review,

    B. Xu, J. Lee, D. Kwon, L. Kong, and M. Pecht, “Mitigation strategies for li-ion battery thermal runaway: A review,”Renewable and Sustain- able Energy Reviews, vol. 150, p. 111437, 2021

    2021

  3. [3]

    Effect of thermal impact on the onset and propagation of thermal runaway over cylindrical li-ion batteries,

    Y . Liu, L. Zhang, Y . Ding, X. Huang, and X. Huang, “Effect of thermal impact on the onset and propagation of thermal runaway over cylindrical li-ion batteries,”Renewable Energy, vol. 222, p. 119910, 2024

    2024

  4. [4]

    Monitoring thermal runaway of lithium-ion batteries by means of gas sensors,

    X.-X. Wang, Q.-T. Li, X.-Y . Zhou, Y .-M. Hu, and X. Guo, “Monitoring thermal runaway of lithium-ion batteries by means of gas sensors,” Sensors and Actuators B: Chemical, vol. 411, p. 135703, 2024

    2024

  5. [5]

    Thermal runaway early warning and risk estimation based on gas production characteristics of different types of lithium- ion batteries,

    Y . Cuiet al., “Thermal runaway early warning and risk estimation based on gas production characteristics of different types of lithium- ion batteries,”Batteries, vol. 9, no. 9, p. 438, 2023

    2023

  6. [6]

    Reliable and early warning of lithium-ion battery thermal runaway based on electrochemical impedance spectrum,

    P. Dong, Z. Liu, P. Wuet al., “Reliable and early warning of lithium-ion battery thermal runaway based on electrochemical impedance spectrum,” Journal of The Electrochemical Society, vol. 168, p. 090529, 2021

    2021

  7. [7]

    Real-time overcharge warning and early thermal runaway prediction of li-ion battery by online impedance measurement,

    N. Lyu, Y . Jin, R. Xionget al., “Real-time overcharge warning and early thermal runaway prediction of li-ion battery by online impedance measurement,”IEEE Transactions on Industrial Electronics, vol. 69, no. 2, pp. 1929–1936, 2021

    1929

  8. [8]

    An online data-driven fault diagnosis and thermal runaway early warning for electric vehicle batteries,

    Z. Sun, Z. Wang, P. Liuet al., “An online data-driven fault diagnosis and thermal runaway early warning for electric vehicle batteries,”IEEE Transactions on Power Electronics, vol. 37, no. 10, pp. 12 636–12 646, 2022

    2022

  9. [9]

    Case study of an electric vehicle battery thermal runaway and online internal short-circuit detection,

    W. Gao, X. Li, M. Maet al., “Case study of an electric vehicle battery thermal runaway and online internal short-circuit detection,” IEEE Transactions on Power Electronics, vol. 36, no. 3, pp. 2452–2455, 2020

    2020

  10. [10]

    Developing an online data- driven approach for prognostics and health management of lithium-ion batteries,

    S. Khaleghi, M. S. Hosen, D. Karimiet al., “Developing an online data- driven approach for prognostics and health management of lithium-ion batteries,”Applied Energy, vol. 308, p. 118348, 2022

    2022

  11. [11]

    Model-constrained deep learning for online fault diagnosis in li-ion batteries over stochastic conditions,

    R. Chenet al., “Model-constrained deep learning for online fault diagnosis in li-ion batteries over stochastic conditions,”Nature Com- munications, vol. 16, p. 1537, 2025

    2025

  12. [12]

    Are transformers effective for time series forecasting?

    A. Zeng, M. Chen, L. Zhang, and Q. Xu, “Are transformers effective for time series forecasting?” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 37, no. 9, June 2023, pp. 11 121–11 128

    2023

  13. [13]

    Cylindrical battery fault detec- tion under extreme fast charging: A physics-based learning approach,

    R. Firoozi, S. Sattarzadeh, and S. Dey, “Cylindrical battery fault detec- tion under extreme fast charging: A physics-based learning approach,” IEEE Transactions on Energy Conversion, vol. 37, no. 2, pp. 1241–1250, 2021

    2021

  14. [14]

    Multi-source domain transfer learning with small sample learning for thermal runaway diagnosis of lithium-ion battery,

    C. Dong and D. Sun, “Multi-source domain transfer learning with small sample learning for thermal runaway diagnosis of lithium-ion battery,” Applied Energy, vol. 365, p. 123248, 2024

    2024

  15. [15]

    Mechanism of heat transfer suppression and safety evaluation of high-performance aerogel insulation materials in the ther- mal runaway propagation of lithium-ion batteries,

    Y . Huanget al., “Mechanism of heat transfer suppression and safety evaluation of high-performance aerogel insulation materials in the ther- mal runaway propagation of lithium-ion batteries,”Energy, vol. 334, p. 137684, 2025

    2025

  16. [16]

    Non-uniform phase change material strategy for direc- tional mitigation of battery thermal runaway propagation,

    W. Zhanget al., “Non-uniform phase change material strategy for direc- tional mitigation of battery thermal runaway propagation,”Renewable Energy, vol. 200, pp. 1338–1351, 2022

    2022

  17. [17]

    An integrated methodology for dynamic risk prediction of thermal runaway in lithium-ion batteries,

    H. Meng, Q. Yang, E. Zio, and J. Xing, “An integrated methodology for dynamic risk prediction of thermal runaway in lithium-ion batteries,” Process Safety and Environmental Protection, 2023

    2023

  18. [18]

    Risk analysis of lithium-ion battery accidents based on physics-informed data-driven Bayesian networks,

    H. Menget al., “Risk analysis of lithium-ion battery accidents based on physics-informed data-driven Bayesian networks,”Reliability Engineer- ing and System Safety, vol. 251, p. 110294, 2024

    2024

  19. [19]

    Real-time knowledge- and data-driven reliability analysis for lithium-ion battery energy storage system by Bayesian fault propagation network,

    Y . Wanget al., “Real-time knowledge- and data-driven reliability analysis for lithium-ion battery energy storage system by Bayesian fault propagation network,”Applied Energy, vol. 402, p. 127013, 2026

    2026

  20. [20]

    A time series is worth 64 words: Long-term forecasting with transformers,

    Y . Nie, N. H. Nguyen, P. Sinthong, and J. Kalagnanam, “A time series is worth 64 words: Long-term forecasting with transformers,”arXiv preprint arXiv:2211.14730, 2022

    Pith/arXiv arXiv 2022

  21. [21]

    Timesnet: Temporal 2d- variation modeling for general time series analysis,

    H. Wu, J. Xu, J. Wang, and M. Long, “Timesnet: Temporal 2d- variation modeling for general time series analysis,”arXiv, Tech. Rep. arXiv:2210.02186, 2022

    Pith/arXiv arXiv 2022

  22. [22]

    iTransformer: Inverted transformers are effective for time series forecasting,

    Y . Liuet al., “iTransformer: Inverted transformers are effective for time series forecasting,” inInternational Conference on Learning Represen- tations, 2024

    2024

  23. [23]

    ETSformer: Exponen- tial smoothing transformers for time-series forecasting,

    G. Woo, C. Liu, D. Sahoo, A. Kumar, and S. Hoi, “ETSformer: Exponen- tial smoothing transformers for time-series forecasting,” inInternational Conference on Learning Representations, 2023

    2023

  24. [24]

    Non-stationary transformers: Exploring the stationarity in time series forecasting,

    Y . Liu, H. Wu, J. Wang, and M. Long, “Non-stationary transformers: Exploring the stationarity in time series forecasting,” inAdvances in Neural Information Processing Systems, 2022

    2022

  25. [25]

    Crossformer: Transformer utilizing cross- dimension dependency for multivariate time series forecasting,

    Y . Zhang and J. Yan, “Crossformer: Transformer utilizing cross- dimension dependency for multivariate time series forecasting,” in International Conference on Learning Representations, 2023

    2023

  26. [26]

    Pathformer: Multi-scale transformers with adaptive path- ways for time series forecasting,

    P. Chenet al., “Pathformer: Multi-scale transformers with adaptive path- ways for time series forecasting,”arXiv preprint, vol. abs/2402.05956, 2024

    Pith/arXiv arXiv 2024

  27. [27]

    TSMixer: An all-MLP architecture for time series forecasting,

    S.-A. Chenet al., “TSMixer: An all-MLP architecture for time series forecasting,”Transactions on Machine Learning Research, 2023

    2023

  28. [28]

    Long-term forecasting with TiDE: Time-series dense encoder,

    A. Daset al., “Long-term forecasting with TiDE: Time-series dense encoder,”Transactions on Machine Learning Research, 2023

    2023

  29. [29]

    N-HiTS: Neural hierarchical interpolation for time series forecasting,

    C. Challuet al., “N-HiTS: Neural hierarchical interpolation for time series forecasting,” inProceedings of the AAAI Conference on Artificial Intelligence, 2023

    2023

  30. [30]

    Development of a robust early-stage thermal runaway detection model for lithium-ion batteries,

    W. C. Tamet al., “Development of a robust early-stage thermal runaway detection model for lithium-ion batteries,” National Institute of Standards and Technology, Tech. Rep., 2024

    2024

  31. [31]

    Study on thermal runaway characteristics of lithium batteries under high-rate charge/discharge and development of a deep learning-based early warning model,

    Y . Huanget al., “Study on thermal runaway characteristics of lithium batteries under high-rate charge/discharge and development of a deep learning-based early warning model,”Energy, vol. 334, p. 137676, 2025

    2025

  32. [32]

    Fault cause inferences of onboard lithium-ion battery thermal runaway using convolutional neural network,

    S. Wang, Z. Wang, Z. Zhang, and X. Cheng, “Fault cause inferences of onboard lithium-ion battery thermal runaway using convolutional neural network,”Energy, vol. 320, p. 135328, 2025

    2025

  33. [33]

    Fault detection for li-ion batteries of electric vehicles with feature-augmented attentional autoencoder,

    Y . Fanet al., “Fault detection for li-ion batteries of electric vehicles with feature-augmented attentional autoencoder,”Scientific Reports, vol. 15, p. 18534, 2025

    2025

  34. [34]

    V oltage faults diagnosis for lithium-ion batteries in electric vehicles using optimized graphical neural network,

    J. Ouyang, Z. Lin, L. Hu, and X. Fang, “V oltage faults diagnosis for lithium-ion batteries in electric vehicles using optimized graphical neural network,”Scientific Reports, vol. 15, p. 27328, 2025

    2025

  35. [35]

    An IMFO-LSTM-BiGRU combined network for long- term multiple battery states prediction for electric vehicles,

    Z. Wanget al., “An IMFO-LSTM-BiGRU combined network for long- term multiple battery states prediction for electric vehicles,”Energy, vol. 309, p. 133069, 2024

    2024

  36. [36]

    Deep learning and polarization equilibrium based state of health estimation for lithium-ion battery using partial charging data,

    T. Wanget al., “Deep learning and polarization equilibrium based state of health estimation for lithium-ion battery using partial charging data,” Energy, vol. 317, p. 134564, 2025

    2025

  37. [37]

    Thermal runaway mechanism of lithium ion battery for electric vehicles: A review,

    X. Fenget al., “Thermal runaway mechanism of lithium ion battery for electric vehicles: A review,”Energy Storage Materials, vol. 10, pp. 246– 267, 2018

    2018

  38. [38]

    Temporal convolutional network– transformer hybrid architecture with hippo optimization for lithium battery SOC estimation,

    L. Wu, Y . Wang, and L. Xing, “Temporal convolutional network– transformer hybrid architecture with hippo optimization for lithium battery SOC estimation,”World Electric Vehicle Journal, vol. 17, no. 5, p. 236, 2026

    2026