Joint Remaining Useful Life Prediction and Capacity Estimation of Lithium-Ion Batteries Using Partial-Charging Data
Khoa Tran, Ho-Si-Hung Nguyen, Phone Wai Yan Moe, Hung-Cuong Trinh, Thi-Hoang-Giang Tran
cs.LG
Quick verdict
Cross-expert FiLM framework for joint battery RUL/capacity prediction from partial charging; thorough ablations, honest limits, but fusion gains are marginal.
Summary
This paper presents a cross-expert neural framework that jointly predicts remaining useful life (RUL) and estimates capacity of lithium-ion batteries using only partial-charging data. A RUL Expert encodes long-term degradation via a pretrained GRU-autoencoder, 2D-CNN, and temporal GRU over ten 10-min charging segments sampled (stride 3) from a 30-cycle window; a Capacity Expert feeds 28 statistical descriptors from 40-min segments of ten consecutive cycles through a 2D-CNN and Transformer. A FiLM module lets the short-term representation modulate the long-term one before a shared two-output regression head, with training split into three stages and both experts frozen at fusion. On the public Severson (124 cells) and Ma (77 cells) LFP datasets, the reference model reports RUL RMSE of 143.69 and 161.10 cycles and capacity RMSE of 12.36 and 7.28 mAh — the lowest RUL RMSE among compared methods, though not the best capacity accuracy (the rerun GRU reaches 9.39 mAh on Dataset I; quoted Ma2022 reports 2.57 mAh on Dataset II). Ablations on Dataset I cover inputs, windows, encoders, backbones, and nine fusion operators; FiLM's RUL gain over simple feature concatenation is negligible (143.69 vs 143.57 cycles), its advantage lies in capacity. Limitations: LFP-only cells, batch-shift evaluation, no significance tests, unclear ensemble treatment of baselines, and 5-epoch training stages.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Partial-charging segments with segment-local charge integration suffice for joint RUL prediction and capacity estimation, requiring no measured historical full-cycle capacity.
- Pairing a long-term view (10-min segments over a 30-cycle window) with a short-term view (40-min segments over ten consecutive cycles) captures complementary degradation information that improves both tasks over either standalone expert on Dataset I.
- The three-stage training strategy (supervised autoencoder pretraining, independent expert pretraining, frozen-expert fusion) preserves expert-specific knowledge and yields the lowest RUL RMSE among compared methods on both datasets (143.69 and 161.10 cycles).
- FiLM provides the best joint RUL-capacity trade-off among nine tested fusion operators with frozen experts, though feature concatenation and attention pooling match or slightly beat its RUL RMSE.
Strengths
Deployment-oriented input design
Segment-local charge integration (Eq. 5) resets accumulated charge at the segment start, removing the pre-segment measurement dependence the authors correctly diagnose in Ma et al. [17]; neither branch needs historical full-cycle capacity, only segments starting above 3.1 V (Sec. I-II).
Honest negative reporting
The text explicitly concedes MSFEH's better Dataset I RUL MAPE (9.82% vs 10.91%), the GRU baseline's better capacity RMSE (9.39 vs 12.36 mAh), and that on Dataset II several methods beat its 7.28 mAh (Sec. IV-A/B) — rare candor in this literature.
Ablation suite mirrors the architecture construction
Batch-shift evaluation protocol
Expert naming backed by evidence
Weaknesses
Fusion gain on the primary objective is within noise
No inferential statistics anywhere
Very short Stage II/III training (5 epochs each)
Possible ensemble asymmetry vs baselines
Dataset II weaknesses unexplored
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Preliminary Review: Joint RUL Prediction and Capacity Estimation from Partial-Charging Data
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What input data does this battery RUL and capacity model need?
Only partial-charging measurements: a nominal 10-minute charging segment per cycle (sampled over a 30-cycle window) for the long-term branch, and nominal 40-minute segments over the ten most recent consecutive cycles for the short-term branch, both starting at the first sample above 3.1 V. It does not require measured full-cycle capacity, complete charge-discharge records, or accumulated charge referenced to the start of the full charging session.
How accurate is the proposed method?
On two public LFP datasets (Severson 2019, 124 cells; Ma 2022, 77 cells) it reports mean RUL RMSE of 143.69 cycles (Dataset I) and 161.10 cycles (Dataset II) and capacity RMSE of 12.36 and 7.28 mAh respectively. It has the lowest RUL RMSE among the compared methods, but not the best capacity accuracy: a rerun GRU baseline reaches 9.39 mAh on Dataset I, and the quoted Ma2022 study reports 2.57 mAh on Dataset II.
Is code or data available for this paper?
The two evaluation datasets are public (Severson et al. 2019; Ma et al. 2022). The paper does not link a public code repository; it states that code inquiries may be directed to AIWARE Limited Company via aiware.website. Architecture tables and hyperparameters are described in enough detail for reimplementation.
Does the FiLM fusion mechanism actually help?
Modestly, and mainly for capacity. On Dataset I, FiLM improves both mean errors over the standalone experts (RUL 143.69 vs 147.08 cycles; capacity 12.36 vs 14.41 mAh), but simple feature concatenation achieves a slightly lower RUL RMSE (143.57) with worse capacity (15.12 mAh). FiLM's clear advantage over all nine tested fusion operators is capacity RMSE (12.36 mAh); its RUL edge over simpler fusion is within the reported between-cell variability.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 2, 3, and 6 (radar profiles, trajectory plots, before/after FiLM scatters) are not analyzable from the extracted text; only captions and inline descriptions were available.
- The abstract embedded in the metadata is truncated mid-sentence, but the full-text extraction contains the complete abstract.
- Rows for MSFEH, structural pruning, and Ma2022 are the original papers' self-reported numbers; this review could not independently verify their protocol comparability.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.