{"id":"d60cccc6-0a17-4eab-adf8-eafb201fb49e","arxiv_id":"2501.16377","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A pipeline combining PSO-optimized variational mode decomposition with a CNN-LSTM network estimates battery state of health with reported MAPE as low as 0.26% on one NASA battery, but the evaluation uses non-causal decomposition and lacks a trivial baseline.","lead":"A battery health estimator that splits capacity curves into three frequency components and then runs them through a CNN-LSTM reports state-of-health errors as low as 0.26% on a NASA battery, beating six comparison methods. The evaluation, however, uses the full capacity curve to set up the decomposition, so the results may be optimistic.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Non-causal VMD leaks future capacity information into the input features, so the reported 0.26% MAPE does not support the claimed online SoH estimation accuracy.","rationale":"The reader's weakest assumption correctly identifies the non-causal VMD decomposition as the fragile premise. My stress-test confirms this is load-bearing: the reported 0.26% MAPE is computed with features that contain future information, so the headline performance claim is not evidence of an online estimator. I also flag a second leakage the reader did not emphasize: the VMD parameter α is optimized per battery on the full capacity curve, meaning the test battery's decomposition parameters are also informed by future data. This strengthens the concern without changing the verdict. The paper has some genuine independent value: the use of PSO for VMD parameters is clearly scoped, the authors honestly state they claim no novelty for PSO, and the comparison of meta-heuristic convergence is useful. However, the experimental protocol does not support the deployment-oriented accuracy claim. The reader's CONDITIONAL verdict is appropriate: the paper needs a causal re-evaluation before the central claim can be accepted. My partial agreement reflects the additional parameter-leakage issue, not a disagreement about the main concern.","tokens_in":9263,"tokens_out":3100,"duration_ms":31743,"concrete_test":"Re-run the full evaluation with strictly causal features. For each test battery and each test time t, apply VMD only to the prefix f(1:t) (or a trailing window of fixed length), using α fixed from the training batteries only, then train CNN-LSTM and compute the metrics in Table I. Also include a persistence baseline that predicts SoH(t) as the last observed capacity. If the B0005 MAPE rises substantially above 0.26% or no longer beats VMD-Transformer under the same causal protocol, the leakage concern is confirmed and the central claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that OSL achieves the lowest battery SoH estimation errors (0.26% MAPE on B0005, Table I) rests on a non-causal feature construction. VMD is applied once to the entire capacity curve before any windowing (Section III.C, Fig. 4). The variational problem in Eq. (1) is solved globally, so every IMF value at time t depends on the whole signal, including future cycles. The CNN-LSTM then consumes these IMFs over the window f(t-T_in)...f(t-1), as described in Section II.D. Thus the input features at time t encode information from cycles after t, making the evaluation an offline curve-fitting exercise rather than an online estimation of the kind motivated in Section II.A ('OSL processes and analyzes the signal f(·) up to time slot t'). A second, related leakage is that the optimal VMD parameter α is obtained by running PSO on each battery's full capacity curve (Section III.C), so the test battery's α is tuned with access to its complete future. The claim that 'OSL outperforms the comparison algorithms in all the twelve tested cases' is therefore not supported for the practical BMS deployment scenario the paper argues for. Correcting this requires a causal decomposition and training-only parameter selection; until then, the reported accuracy advantage is not credible.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes OSL, a battery state-of-health (SoH) estimation method that combines particle-swarm-optimized variational mode decomposition (VMD) with a two-stage CNN-LSTM. The authors optimize VMD parameters K and alpha using envelope entropy, decompose the battery capacity curve into three IMFs, and feed windows of these IMFs into a CNN followed by an LSTM to estimate SoH. Experiments on the NASA battery dataset with four batteries are reported, and the paper claims that OSL achieves the best accuracy among the compared methods, including a MAPE of 0.26% on battery B0005 and best results across all twelve reported metric/battery combinations.","tokens_in":9529,"tokens_out":2428,"duration_ms":23056,"significance":"If the reported accuracy were achievable in a genuinely online setting, the proposed combination of optimized decomposition and multi-stage learning would be a practically useful contribution to battery health estimation. The paper also contains a useful negative result: the authors honestly observe that different meta-heuristics converge to the same VMD parameters, so the choice of optimizer is not the source of performance gains. The comparison against several baselines, including EMD-based and Transformer-based alternatives, is informative. However, the significance of the central performance claim is currently undermined by a causal-leakage issue in the evaluation protocol, so the contribution cannot yet be taken at face value.","major_comments":[{"comment":"The VMD decomposition is applied to the entire battery capacity curve before any input window is formed. As stated in Section II.C and shown in Fig. 4, the three IMF values at time t are computed from the full capacity curve, including future cycles, because the variational problem in Eq. (1) is solved globally over the whole signal. The CNN-LSTM then consumes IMFs over the window t-T_in to t-1, as described in Section II.D. Therefore the input features at time t encode information from cycles after t, making the reported RMSE/MAE/MAPE an offline curve-fitting result rather than an online SoH estimation of the kind motivated in Section II.A. The authors should either provide a causal decomposition (e.g., using only data up to t) or explicitly restrict the claims to offline analysis; without this, the headline accuracy numbers do not support the claimed real-world BMS integration.","section":"Section II.C, Section II.D, Fig. 4"},{"comment":"The VMD parameters K and alpha are optimized per test battery using its full capacity curve. For battery B0005, for example, alpha=30 is selected by PSO running on the entire B0005 capacity signal, which includes future information and the label being predicted. In the leave-one-battery-out setup, this constitutes leakage: the test battery's decomposition parameters are not selected from training data alone. The authors should select parameters using only the training batteries or via nested cross-validation, and report whether the accuracy advantage persists under that protocol.","section":"Section III.C"},{"comment":"The abstract states that OSL demonstrates 'a mean error of just 0.26%', but the 0.26% MAPE is achieved only on battery B0005. The mean MAPE across the four batteries in Table I is approximately 0.33%, and the conclusion similarly refers to '0.26% MAE' while the table reports 0.26% as MAPE. The summary statistics and the metric used should be reported consistently and should not conflate a single-battery result with a dataset-wide mean.","section":"Table I and Abstract"},{"comment":"Results for SAGPCN, BiLSTM, and VMD-Transformer are taken from previous publications [2], [4], and [10], while the other methods are implemented by the authors. If those external results use different train/test splits, input features, or evaluation protocols than the authors' own experiments, the comparison in Table I is not apples-to-apples. The paper should state whether the same leave-one-battery-out protocol and capacity-only input were used for all compared methods, or should rerun the comparisons under a unified protocol.","section":"Section III.D.2, Table I"}],"minor_comments":[{"comment":"There are several typographical errors, including 'o ptimal' in the abstract and 'B00005' instead of 'B0005' in Section III.C; these should be corrected.","section":"Abstract and Section I"},{"comment":"Fig. 3 reports convergence of envelope entropy for battery B0005 only, and the text says results for other batteries are 'nearly the same'. Showing the actual convergence curves or reporting the obtained K and alpha for all batteries in a table would make the parameter-optimization claim more reproducible.","section":"Section III.C, Fig. 3"},{"comment":"The paper says the CNN input is 'the three IMFs from time t-T_in to time t-1' but does not clearly state what happens at the beginning of the series, when fewer than T_in previous IMFs exist; a description of the padding or truncation strategy would improve clarity.","section":"Section II.D"},{"comment":"Reference [12] misspells 'Prognostics' as 'Progmostics', and reference [13] is listed as a 2025 arXiv paper without a DOI; the citation details should be checked.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal, and the authors are commendably candid about the limited role of meta-heuristic choice. My main concern is the non-causal VMD decomposition and per-test-battery parameter tuning, which jointly undermine the central performance claim. If the authors can re-run the evaluation with a causal decomposition and training-only parameter selection, the contribution would be much stronger. I would not recommend rejection at this stage, but the leakage issue must be resolved before the reported accuracies can be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing to know: this is a competent, honest engineering paper whose central accuracy claim is likely inflated by causal leakage. The 0.26% MAPE in the abstract is for battery B0005 only, and the input IMFs at time t are computed by VMD over the whole capacity curve, including future cycles. So as written, the experiment tests offline curve fitting, not the online estimation the paper motivates.\n\nWhat is actually new is modest: PSO-optimized VMD combined with a two-stage CNN-LSTM. The authors explicitly disclaim novelty for using PSO on VMD parameters and even show that several metaheuristics converge to the same optimum. That is a genuinely useful negative result and the write-up is refreshingly clear about it. The paper is well structured, the comparisons are reasonably broad, and the claim that signal decomposition helps is supported by the consistent gap between LSTM and VMD-LSTM. The detailed PSO settings and the decomposition figures help reproducibility.\n\nThe soft spots are load-bearing, though. Equation (1) is a global variational problem, and Figure 4 shows IMFs computed across all cycles. When you feed a window ending at t-1, the IMF values inside that window already contain information from after t. The PSO optimization of alpha also uses the full test battery curve, so the test battery's decomposition parameters are tuned with future access. That is a second leakage path. The reported advantage over VMD-LSTM and VMD-Transformer could shrink or disappear under a causal decomposition with training-only parameter selection. A persistence baseline is also missing; for slowly decaying capacity, a last-value extrapolation often gives very low error. The abstract's \"mean error of just 0.26%\" is misleading because it is a single-battery result, not a mean over the four batteries. Minor issues: no code or data release, no confidence intervals over repeated training runs.\n\nWho is this for? Researchers working on battery state-of-health estimation with signal decomposition. The paper is a useful cautionary example of how easy it is to leak future information in feature construction, and the metaheuristic comparison is worth having on record.\n\nRecommendation: send it to peer review as a borderline-major revision. The topic is relevant and the authors are honest, but the evaluation protocol must be fixed before the accuracy claim is credible. I would not cite the 0.26% number until the authors rerun with causal VMD, train-only parameter tuning, a persistence baseline, and honest per-battery averages in the abstract.","headline":"Competent and honest engineering paper, but the headline 0.26% error is inflated by a non-causal VMD decomposition that leaks future capacity information into the input.","tokens_in":10055,"tokens_out":1786,"would_cite":false,"duration_ms":18033,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"PSO-tuned VMD plus CNN-LSTM estimates battery health to 0.26% mean error.","keywords":["state of health estimation","variational mode decomposition","particle swarm optimization","CNN-LSTM","battery capacity degradation","signal decomposition","battery management system"],"falsifier":"Run OSL with a strictly causal VMD, refitting the decomposition online using only cycles up to t on the same four batteries and comparing MAPE; if the advantage over baselines shrinks or the 0.26% best-case error rises substantially, the non-causal preprocessing is the real source of the accuracy gain.","tokens_in":9053,"feed_emoji":"🔋","tokens_out":4250,"duration_ms":36714,"temperature":0.7,"pith_summary":"This paper proposes OSL, a battery state-of-health estimator that first decomposes the capacity curve into intrinsic mode functions using variational mode decomposition whose parameters are tuned per battery by particle swarm optimization, then feeds windows of those IMFs through a two-stage CNN-LSTM. The central claim is that this combination of optimized signal decomposition and multi-stage learning yields the lowest estimation errors among all compared methods on a public four-battery aging dataset, with a best mean absolute percentage error of 0.26% on one cell and the best score in all twelve reported battery-metric cases. The authors argue this matters because practical battery management systems need accurate, low-cost health estimates, and their pipeline is simple enough, using capacity alone as input, to be deployed.","feed_headline":"Decompose the capacity curve, then learn: battery SoH error drops to 0.26%","feed_subtitle":"On four public aging cells, PSO-tuned VMD with CNN-LSTM beats all baselines in every metric.","key_machinery":"The load-bearing mechanism is PSO-optimized variational mode decomposition. VMD solves a constrained variational problem to split the capacity signal into K intrinsic mode functions with compact, minimally overlapping frequency bands; its two most influential parameters, K (the number of IMFs) and alpha (the balance between band compactness and reconstruction fidelity), are tuned per battery by particle swarm optimization using Shannon entropy of the IMF envelopes as the fitness function. The resulting three IMFs become the input to a two-stage CNN-LSTM in which 1D convolutions extract spatial patterns and LSTM captures temporal dependencies.","core_discovery":"OSL's discovery is that treating battery capacity as a signal with separable frequency content, rather than as a raw time series, materially improves state-of-health estimation. With VMD decomposition optimized per battery (optimal K=3 and battery-specific alpha found by PSO using envelope entropy as the fitness), the three IMFs separate the long-term degradation trend, intermediate variations, and high-frequency noise. Feeding windows of these IMFs into a convolutional stage for spatial features and then an LSTM stage for temporal dependencies yields mean absolute errors below 0.35% across the four cells, outperforming LSTM, BiLSTM, SAGPCN, EMD-LSTM, VMD-LSTM, and VMD-Transformer in every metric-battery combination.","pith_inferences":["The paper leaves unspecified whether VMD is fit on the full capacity history before windowing; a causal implementation that recomputes IMFs with data only up to time t would likely raise the reported errors, so the headline accuracy is an upper bound under real-time constraints.","A stronger test of the method would be training on a broader set of cycling protocols and cell chemistries, since the per-battery alpha values suggest sensitivity to aging mode.","The capacity-only input is a deliberate simplification; using voltage and temperature windows as additional signal channels could either enhance the decomposition or complicate it."],"forward_implications":["If OSL is right, signal decomposition plus two-stage learning is the performance frontier for capacity-only state-of-health estimation on this dataset.","VMD parameters must be optimized per battery; fixed parameters degrade results, but the choice of meta-heuristic used for tuning matters little.","Noise and rapid capacity-regeneration oscillations can be isolated into high-frequency IMFs, letting the learner focus on the degradation trend.","The same feature pipeline can be extended to future health forecasting by shifting the output window."],"supporting_citations":[{"why":"Supplies the four-battery public aging dataset used for all training and testing.","marker":"[12]"},{"why":"Defines variational mode decomposition, the core preprocessing method.","marker":"[8]"},{"why":"Introduces particle swarm optimization, the algorithm used to tune VMD parameters.","marker":"[11]"},{"why":"Provides the VMD-Transformer baseline results that OSL is directly compared against.","marker":"[10]"},{"why":"Provides the SAGPCN baseline, a recent non-decomposition method.","marker":"[2]"},{"why":"Provides the BiLSTM attention baseline results.","marker":"[4]"},{"why":"Provides empirical mode decomposition, the alternative decomposition method compared with VMD.","marker":"[7]"}],"fun_headline_variants":["Optimal signal decomposition + multi-stage learning: battery SoH error 0.26%","Split capacity signal into 3 IMFs, learn spatial+temporal: 0.26% error","VMD-optimized decomposition + CNN-LSTM: battery aging error 0.26%","Optimal VMD + dual-stage deep net = battery SoH error 0.26%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The VMD decomposition is computed from the entire capacity curve before any windowing, so the IMF values fed to the network at time t contain information from future cycles; if the pipeline were constrained to causal inputs only, the reported errors could be higher.","fun_headline_variants_meta":{"raw":{"variants":["Optimal signal decomposition + multi-stage learning: battery SoH error 0.26%","Split capacity signal into 3 IMFs, learn spatial+temporal: 0.26% error","VMD-optimized decomposition + CNN-LSTM: battery aging error 0.26%","Optimal VMD + dual-stage deep net = battery SoH error 0.26%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000983,"raw_usage":{"total_tokens":4124,"prompt_tokens":852,"completion_tokens":3272,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":468,"completion_tokens_details":{"reasoning_tokens":3173}},"tokens_in":468,"tokens_out":3272,"duration_ms":21537,"temperature":1.0,"reasoning_tokens":3173,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T15:14:44.795225+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run OSL with a strictly causal VMD, refitting the decomposition online using only cycles up to t on the same four batteries and comparing MAPE; if the advantage over baselines shrinks or the 0.26% best-case error rises substantially, the non-causal preprocessing is the real source of the accuracy gain.","supporting_citations":[{"cited_title":"Battery data set,","cited_arxiv_id":null,"evidence_quote":"Supplies the four-battery public aging dataset used for all training and testing."},{"cited_title":"A new optimizer using particle swarm theory,","cited_arxiv_id":null,"evidence_quote":"Introduces particle swarm optimization, the algorithm used to tune VMD parameters."},{"cited_title":"Edge–cloud collaborative estimation lithium-ion battery soh based on mewoa-vmd and transformer,","cited_arxiv_id":null,"evidence_quote":"Provides the VMD-Transformer baseline results that OSL is directly compared against."},{"cited_title":"Sagpcn: Self-attention graph pooling convolutional network for battery state of health estimation,","cited_arxiv_id":null,"evidence_quote":"Provides the SAGPCN baseline, a recent non-decomposition method."},{"cited_title":"An end-cloud collaboration approach for state-of-health estimation of lithium-ion batteries based on bi-lstm with collaboration of multi-feature and attention mechanism,","cited_arxiv_id":null,"evidence_quote":"Provides the BiLSTM attention baseline results."},{"cited_title":"A data-driven approach with uncertainty quantification for predicting future capacities and remaining useful life of lithium-ion battery,","cited_arxiv_id":null,"evidence_quote":"Provides empirical mode decomposition, the alternative decomposition method compared with VMD."}],"review_version":1}