{"id":"cbd0cc27-557c-4187-a655-162de3f0f71c","arxiv_id":"2412.11123","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A bidirectional transition-based Lempel-Ziv complexity measure is introduced and reported to achieve the highest fault classification accuracy on two bearing datasets.","lead":"The authors define a new complexity measure for time series that tracks how patterns change from one moment to the next, then use it to classify bearing faults. They report that the new measure separates fault types more accurately than three existing Lempel-Ziv complexity methods.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'significantly outperforming' claim rests on single-split point estimates, dataset-specific (m,c) settings, and an unreported hierarchical layer K, with no statistical significance testing.","rationale":"The reader's weakest_assumption already pointed to parameter and split selection, so this stress-test agrees partially. My concern sharpens it in three ways: the absence of any statistical test for the 'significantly outperforming' claim, the unreported hierarchical layer K that makes the feature dimension irreproducible, and the per-dataset (m,c) choices that suggest tuning. I also considered the algorithmic soundness of BT-DELZC: the definition in Equations 9-15 is internally consistent, and the use of in/out transition sequences is a reasonable extension of DELZC. No fatal mathematical flaw emerged. The simulated experiments (Sections 3.2-3.4) support robustness but do not by themselves establish classification superiority. Therefore the central empirical claim is plausible but not yet evidenced to the standard implied by 'significantly outperforming.' Since the reader already issued CONDITIONAL, my recommendation is UNCHANGED: the paper should be accepted only if the authors add the repeated-split and sensitivity analyses described above.","tokens_in":16989,"tokens_out":2902,"duration_ms":28141,"concrete_test":"Re-run the full pipeline (hierarchical decomposition, PCA, KNN/DT/FNN) under repeated stratified 10-fold cross-validation (or 100 random 70/30 splits) on both datasets, with the same K and (m,c) for all four methods. Report mean ± std accuracy and McNemar's test between BT-DELZC and each baseline for every classifier. Additionally, sweep K over {1,2,3,4} and (m,c) over the grid used in Section 3.1 for both datasets. If BT-DELZC remains significantly better (p<0.05) across all configurations and splits, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that BT-DELZC achieves the highest accuracy across both bearing datasets, significantly outperforming LZC, PLZC, and DELZC. The most load-bearing weakness is that this conclusion rests on point estimates from a single 70/30 split (Section 4.2, Tables 3-4; Section 4.3, Table 6) with no repeated resampling, no confidence intervals, and no paired significance test. The reported margins vary by classifier and dataset, and with only 270 (PU) or 360 (CWRU) test samples, split-to-split variance could be substantial. In addition, the proposed method uses different parameters for the two datasets (PU: m=3, c=5; CWRU: m=4, c=4) without justification or sensitivity analysis, so the advantage could reflect per-dataset tuning. Finally, the hierarchical layer K, which controls the number of features fed into PCA, is never stated in Section 4.1: Step 2 says 'according to the selected layer K' but no K value appears in the real-data experiments. This makes the feature set irreproducible and leaves open the possibility that K was chosen favorably. The paper's use of 'significantly' is descriptive, not statistical; no test accompanies the accuracy comparisons. If the accuracy gap persists under repeated resampling and a K/m/c sensitivity sweep, the claim would be supported; otherwise, the reported superiority may be an artifact of evaluation protocol.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes BT-DELZC, a Lempel-Ziv complexity variant that combines dispersion-pattern-based DELZC with a weighted bidirectional transition network, and appends hierarchical decomposition to extract multi-frequency features. The method is evaluated on four simulated experiments (parameter analysis, robustness to noise, chirp signals, and a mix process) and on two public fault-bearing datasets (Paderborn University and Case Western Reserve University) using KNN, decision tree, and feedforward neural network classifiers. The authors report that BT-DELZC consistently yields the highest accuracy among LZC, PLZC, DELZC, and BT-DELZC, and claim that it 'significantly outperforms' the existing methods.","tokens_in":17286,"tokens_out":6926,"duration_ms":53717,"significance":"If the empirical claims are substantiated, BT-DELZC would be a useful addition to the family of Lempel-Ziv complexity measures, particularly for bearing fault diagnosis, and the combination of transition networks with LZC is a reasonable novelty. The paper has several strengths: the algorithm is defined step-by-step in Section 2.3; the simulated experiments include quantitative robustness checks with repeated samples and Wilcoxon rank-sum tests; and the real-world evaluation uses two standard public benchmark datasets and multiple classifiers. These strengths make the paper a candidate for publication after the evaluation protocol is made statistically rigorous and the reporting of experimental settings is completed.","major_comments":[{"comment":"The central claim that BT-DELZC 'significantly outperforms' LZC, PLZC, and DELZC is not supported by any statistical test. All reported accuracy rates in Tables 3, 4, and 6 are point estimates from a single 70/30 split. With only 270 (PU) or 360 (CWRU) test samples, split-to-split variance can be substantial, and the margins over the best baseline in some cases are small (e.g., CWRU KNN: 99.72% vs. 98.61%). The authors should report repeated random splits or k-fold cross-validation with means, standard deviations or confidence intervals, and apply a paired test such as McNemar's test between BT-DELZC and each baseline. Without this, the word 'significantly' in the abstract and conclusion is not justified.","section":"Section 4.2, Tables 3–4; Section 4.3, Table 6"},{"comment":"The hierarchical layer K is never specified for the real-world experiments. Section 4.1, Step 2 says 'according to the selected layer K', but Sections 4.2 and 4.3 do not state the K value used; the number of features fed into PCA depends on K. This makes the experimental results irreproducible and leaves open the possibility that K was chosen favorably. The authors must report K for both datasets and, ideally, provide a sensitivity analysis over K (e.g., K = 2, 3, 4) to demonstrate that the reported accuracies are not an artifact of this choice.","section":"Section 4.1, Step 2; Sections 4.2 and 4.3"},{"comment":"The hierarchical decomposition matrix in Eq. (17) is dimensionally inconsistent with the stated scale. The text claims that 'the scale of Q_K^ε becomes N - 2^K + 1 × N - 2^{K-1} + 1', but composing the operators in Eq. (16) K times should yield an (N-K)×N matrix (or similar, depending on the alignment convention), not the dimensions given. For K=2, the claimed scale would be (N-3)×(N-1), which does not match the result of applying two operators to X. Please clarify the exact construction of Q_K^ε, correct the matrix and dimension statement, and provide the precise matrix form used in the experiments.","section":"Eq. (17) and the surrounding text"},{"comment":"The parameters (m,c) differ between the two datasets (PU: m=3, c=5; CWRU: m=4, c=4) without a systematic justification or sensitivity analysis. While the same parameters are used for all methods within each dataset, the per-dataset choice must be shown not to favor BT-DELZC. The authors should provide a sensitivity study (e.g., a grid over m in {3,4,5} and c in {3,4,5}) on at least one dataset to demonstrate that the reported performance advantage is robust to parameter settings rather than a result of per-dataset tuning.","section":"Sections 4.2 and 4.3"}],"minor_comments":[{"comment":"The table header reads 'DETLZC' instead of 'BT-DELZC'; please correct the typo.","section":"Table 6 caption"},{"comment":"The statement 'Consequently, ∑_{j=1}^K 2^j components including the original time series X, can be obtained' is mathematically inconsistent: the sum equals 2^{K+1}-2, and including X gives 2^{K+1}-1 components. Please rephrase to accurately count the components.","section":"After Eq. (18)"},{"comment":"The index range for dispersion patterns is stated as 'i = 2, 3, ..., N − 1', but the actual range should account for the embedding dimension and delay, i.e., i = 2, ..., N − (m−1)τ − 1, as used in Eq. (14). Please correct for consistency.","section":"Section 2.3, Step 2"},{"comment":"The entries '**' are not defined in the table or caption, and the caption contains a typo ('statictical'). Please define the meaning of '**' and correct the typo.","section":"Table 1"},{"comment":"The reference to Bandt and Pompe appears as 'Bandt and Pompe[]' with an empty bracket; please insert the proper citation.","section":"Section 2.2.1"},{"comment":"The paper claims 'the first attempt to integrate transition network methods with LZC approaches.' The authors should verify that no prior work combines transition networks (even without the bidirectional weighting) with LZC-based metrics, and temper the claim if any exists.","section":"Section 5"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a reasonable engineering contribution, but the evaluation protocol is the main weakness. I would recommend requesting a revision that adds repeated resampling, confidence intervals or paired significance tests, reports the hierarchical layer K, and includes a parameter sensitivity analysis. The paper is within the scope of the journal, and the methodological novelty is sufficient for consideration after these revisions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"BT-DELZC is a real, precisely defined extension of DELZC: it takes dispersion-pattern embedding, builds per-pattern in/out link sequences, computes LZC on each, averages the two, and weights by pattern frequency. That combination is new relative to the cited literature, and the simulated experiments (robustness across noise types and lengths, chirp, mix process) do what they should to show the measure behaves. The Wilcoxon tests in Section 3.2 are a genuine attempt at statistical support. So the paper has a solid core and the authors know the prior art.\n\nThe soft spot is exactly where the stress-test note lands: the bearing-diagnosis claim of “significantly outperforming” is not statistically supported. Tables 3 and 6 are single 70/30 split point estimates, with no repeated resampling, no confidence intervals, no paired tests. The margins are large enough that the effect could be real, but with 270–360 test samples per dataset, split-to-split variance may be nontrivial. The dataset-specific (m,c) settings (m=3,c=5 for PU; m=4,c=4 for CWRU) are presented without sensitivity analysis, and the hierarchical layer K is never stated in Section 4 at all. That last omission is simple to fix, but it makes the feature set irreproducible and leaves open the possibility of favorable selection. Minor: Eq. (17) has a typo-level dimensional inconsistency in the matrix size description, but that should not hold up review.\n\nThe paper is an incremental but genuine step, not a conceptual leap. It is coherent on its own terms, and the circularity burden is low: the measure is defined a priori. The main weakness is experimental rigor, not the method itself. I would send this to a serious referee and ask for repeated resampling or confidence bounds, a sensitivity sweep over (m,c,K), and a clear statement of all experimental parameters. If the accuracy gap holds under that, it is a useful contribution to the LZC-variant literature.","headline":"Genuine LZC variant with a clean definition, but the top-line accuracy claim in the bearing experiments is not backed by significance tests or a reported K value.","tokens_in":17799,"tokens_out":3070,"would_cite":true,"duration_ms":28135,"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":"This paper introduces BT-DELZC, a Lempel-Ziv complexity measure that counts new patterns in the in-link and out-link sequences of each dispersion pattern, and reports that it outperforms LZC, PLZC, and DELZC on two fault bearing datasets.","keywords":["Lempel-Ziv complexity","transition network","dispersion entropy","hierarchical decomposition","fault bearing diagnosis","nonlinear time series","Markov chain","complexity measure"],"falsifier":"Run the two bearing experiments under repeated stratified cross-validation over a grid of embedding dimension, class count, and hierarchical layer values, and compare the distributions of accuracy; if BT-DELZC's margin over DELZC falls within the cross-validation noise for most settings, the central claim fails.","tokens_in":16790,"feed_emoji":"🛠️","tokens_out":6128,"duration_ms":46922,"temperature":0.7,"pith_summary":"This paper proposes a new Lempel-Ziv complexity measure, BT-DELZC, that captures not only which dispersion patterns appear in a time series but how those patterns transition into one another. Prior LZC variants such as PLZC and DELZC symbolise the series and count new patterns in the symbol sequence, ignoring the dynamics between consecutive embedding vectors. BT-DELZC builds a weighted bidirectional transition network over dispersion patterns, computes the DELZC of each pattern's in-link and out-link sequences, and weights these by pattern frequency. The authors report that the resulting measure is more robust across noise, more sensitive to gradual frequency changes, and achieves the highest fault-classification accuracy on both real-world bearing datasets they test.","feed_headline":"Signal-transition metric tops prior LZC in bearing-fault tests","feed_subtitle":"Adding in/out-link transition patterns to dispersion Lempel-Ziv complexity lifts fault-classification accuracy on two benchmark datasets.","key_machinery":"The load-bearing construction is the weighted bidirectional transition network over dispersion patterns. Each embedding vector of the time series is assigned a dispersion pattern; for every pattern, the algorithm collects the sequences of patterns that precede and follow it (excluding self-loops), computes a DELZC value on each of those two link sequences, averages them, and weights the result by the pattern's relative frequency. A hierarchical decomposition with averaging and differencing operators feeds multiple frequency components of the signal into this calculation, and PCA compresses the resulting feature vectors before classification.","core_discovery":"The central claim is that incorporating transition dynamics into Lempel-Ziv complexity improves feature extraction from nonlinear time series. Concretely, the paper claims that BT-DELZC outperforms LZC, PLZC, and DELZC in distinguishing chaotic from noisy signals, in tracking the frequency sweep of a chirp signal and the stochastic-to-periodic drift of a mix process, and in classifying nine fault types from one bearing benchmark and twelve from another, under KNN, decision-tree, and feedforward-neural-network classifiers, with accuracy gains of several percentage points in most comparisons.","pith_inferences":["The same bidirectional transition construction could be applied to permutation-based LZC or other symbolisation schemes, not just dispersion patterns; the paper does not test these alternatives.","Excluding self-loops may change the complexity estimate for highly repetitive or strongly chaotic signals; the paper flags this as future work but does not quantify the effect.","The accuracy results rest on a single 70/30 train-test split per dataset; repeated stratified cross-validation across parameter choices would show whether the reported margin is stable.","If the embedding dimension, class count, time delay, and hierarchical layer were tuned per dataset, part of the reported advantage could come from that tuning rather than from the transition network itself."],"forward_implications":["On both real-world bearing datasets, BT-DELZC yields the highest accuracy for every classifier tested (KNN, DT, FNN), with the largest single gain over DELZC reaching about 6.7 percentage points.","The method separates fully chaotic logistic maps from white noise at lower signal-to-noise ratios than LZC, PLZC, or DELZC in most tested hierarchies.","BT-DELZC tracks gradual dynamics better than the existing measures: it produces the smoothest decline in the mix process and the clearest frequency-change signal in the chirp experiment.","Because the measure is weighted by pattern frequency, common dispersion patterns dominate the complexity value, which the authors argue makes the measure more stable across noise realisations.","The hierarchical decomposition stage multiplies the number of features per sample, and the paper shows that classification accuracy stays high across different numbers of PCA components."],"supporting_citations":[{"why":"Defines the original Lempel-Ziv complexity algorithm on finite sequences, which BT-DELZC extends.","marker":"(Lempel and Ziv, 1976)"},{"why":"Introduces dispersion entropy, which supplies the dispersion-pattern alphabet used in BT-DELZC.","marker":"(Rostaghi and Azami, 2016)"},{"why":"Proposes DELZC, the base complexity measure that BT-DELZC modifies by adding transition links.","marker":"(Li et al., 2022b)"},{"why":"Introduces the ordinal pattern transition network that motivates the link-based dynamic framework.","marker":"(Small, 2013)"},{"why":"Provides the hierarchical decomposition model used to extract multi-frequency features.","marker":"(Jiang et al., 2011)"},{"why":"Supplies the overlapped sliding-window hierarchical decomposition variant adopted to stabilise high-frequency components.","marker":"(Zhao et al., 2023)"},{"why":"Supplies one of the two fault-bearing datasets used in the classification experiments.","marker":"(Lessmeier et al., 2016)"},{"why":"Provides the other fault-bearing dataset and a benchmark study for rolling-element bearing diagnostics.","marker":"(Smith and Randall, 2015)"},{"why":"Demonstrates a prior combination of transition networks with entropy measures, which this work extends to LZC-based metrics.","marker":"(Zhang and Shang, 2021)"}],"fun_headline_variants":["Transition-aware LZC metric beats prior on bearing faults","BT-DELZC: new complexity metric outdoes LZC in diagnosis","Bidirectional transition LZC improves fault classification","Markov-based LZC variant tops old measures on bearings","New LZC with transition patterns wins fault-bearing tests"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The results assume that the parameter settings (embedding dimension, class count, time delay, hierarchical layer) and the single 70/30 train-test split are unbiased and representative, so that the measured accuracy differences reflect genuine method performance rather than sampling noise or favourable tuning.","fun_headline_variants_meta":{"raw":{"variants":["Transition-aware LZC metric beats prior on bearing faults","BT-DELZC: new complexity metric outdoes LZC in diagnosis","Bidirectional transition LZC improves fault classification","Markov-based LZC variant tops old measures on bearings","New LZC with transition patterns wins fault-bearing tests"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000167,"raw_usage":{"total_tokens":1233,"prompt_tokens":900,"completion_tokens":333,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":516,"completion_tokens_details":{"reasoning_tokens":254}},"tokens_in":516,"tokens_out":333,"duration_ms":3833,"temperature":1.0,"reasoning_tokens":254,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T15:16:15.513807+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the two bearing experiments under repeated stratified cross-validation over a grid of embedding dimension, class count, and hierarchical layer values, and compare the distributions of accuracy; if BT-DELZC's margin over DELZC falls within the cross-validation noise for most settings, the central claim fails.","supporting_citations":[],"review_version":1}