{"id":"b210770a-9d74-4b10-95e9-3c5cea855688","arxiv_id":"2507.14181","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"SSFL-DCSL combines Laplace-weighted pseudo-labels, local and global contrastive losses, and momentum-updated prototypes to train personalized fault diagnosis models across clients with few labels.","lead":"This paper proposes a semi-supervised federated learning framework for industrial fault diagnosis that combines pseudo-label weighting, dual contrastive losses, and prototype sharing. The authors report accuracy gains over several baseline methods on four bearing and pump datasets when only 10 percent of data are labeled.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline accuracy gain is not attributable: SSFL-DCSL is tested as personalized local models after client fine-tuning while FedAvg-style baselines are reported as a single global model, and the claimed 1.15–7.85% range is not reproduced by the χ=10% rows of Tables IV/VI.","rationale":"I selected the evaluation-protocol confound as the load-bearing issue because the abstract's quantitative claim is the paper's central assertion, and that assertion cannot be assessed if the comparison is apples-to-oranges. The reader's weakest_assumption identifies the same issue, and I agree. The paper otherwise has some independent support: public benchmark datasets, multiple trials with error bars, a component-wise ablation table, and a communication-cost advantage that follows from prototype-only exchange. Those do not neutralize the confound, however, because the reported gains may reflect local fine-tuning and personalization rather than the proposed TLAW and dual-contrastive components. The numerical mismatch between the abstract and the best-baseline gaps shown in Tables IV and VI strengthens the concern, since the claimed range cannot be reconstructed from the reported 10% label-rate rows. The Appendix A proof issue (the step equating the sample mean with the condition P(max(p)<μt)=0.5, and the lower-bound derivation) is a real correctness defect in the theoretical narrative, but it is secondary because the headline claim is empirical. The system-model inconsistency between full model averaging in Section II and prototype-only communication in Algorithm 1 and Section III-C is part of the same evaluation concern: with no global model to test, SSFL-DCSL is a personalized method, and comparing it to global FedAvg without an equivalent fine-tuning protocol is not a controlled test. A single re-run with uniform fine-tuning for all baselines would settle whether the claimed advantage is genuine. Since the reader already recommended REJECT and this concern supports that judgment, I recommend no change to the verdict.","tokens_in":22294,"tokens_out":6027,"duration_ms":70853,"concrete_test":"Re-run the full comparison at χ=10% under a single common protocol: after federated training on the same Dirichlet(0.5) partitions, apply the §III-A fine-tuning stage to every baseline (FedAvg-Supervised, FedAvg/FedProx-FixMatch/UDA, Fed-SSMPN, FedCon, FedCD) on its own labeled local data before local testing, then average over five trials as in §IV-A4. If the per-dataset best-baseline gaps are no longer in the claimed 1.15–7.85% range, or if SSFL-DCSL's advantage falls within the baselines' error bars, the central claim is an artifact of asymmetric evaluation. The same run should report exactly which dataset and trial produce the 1.15% and 7.85% endpoints.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of a 1.15–7.85% accuracy gain at 10% label rate is not established because the evaluation confounds the proposed method's personalization with its algorithmic components. Section III-A states that SSFL-DCSL fine-tunes each client on its locally labeled data before testing, but Section IV-A4 describes no equivalent fine-tuning for FedAvg-Supervised, FedAvg/FedProx-FixMatch/UDA, Fed-SSMPN, FedCon, or FedCD. Algorithm 1 (lines 9–10) communicates only prototypes, with no global-model averaging, contradicting the θG update in Section II; SSFL-DCSL is therefore effectively a personalized local-model method, while the FedAvg-style baselines report a single shared global model. The reported gains could then come from per-client adaptation rather than from TLAW, LCL/GCL, or PTA. The precise range is also not recoverable from Tables IV and VI at χ=10%: the per-dataset best-baseline gaps are PU 2.64 (vs FedCon), MFPT 3.06 (vs FedCD), CWRU 2.31 (vs FedCD), and CP 6.58 (vs FedCD), spanning 2.31–6.58, not the claimed 1.15–7.85. These two facts together make the empirical claim unverifiable as stated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a semi-supervised federated learning framework, SSFL-DCSL, for intelligent fault diagnosis. The method combines a truncated-Laplace adaptive sample weighting function (TLAW) to down-weight low-confidence pseudo-labels, a dual contrastive loss made of a local contrastive loss (LCL) and a global contrastive loss (GCL), and prototype-based aggregation with momentum updates. Experiments are reported on the PU, MFPT, CWRU, and a factory chemical-pump dataset at 10%, 20%, and 40% label rates, including ablations, computational efficiency measurements, and a client-dropout study. The headline claim is a 1.15% to 7.85% accuracy improvement over state-of-the-art methods when only 10% of the data are labeled.","tokens_in":22613,"tokens_out":11111,"duration_ms":124542,"significance":"The problem addressed is relevant: label-scarce and heterogeneous federated fault diagnosis is a realistic industrial setting, and the prototype-only communication idea is attractive for privacy and bandwidth. If the claimed accuracy gains were established and the comparison were fair, SSFL-DCSL would be a meaningful contribution. The paper also provides a stepwise ablation study in Table V and evaluates on multiple datasets. However, the central empirical claim is not reproducible from the reported tables, and the experimental comparison appears to mix personalized local models with global-model baselines. In addition, the theoretical support for TLAW is not rigorously established. Until these load-bearing issues are resolved, the contribution cannot be assessed on the evidence presented.","major_comments":[{"comment":"The abstract claims a 1.15% to 7.85% improvement over state-of-the-art methods at a 10% label rate, but this range is not recoverable from the reported results. At 10% label rate, Table IV gives best-baseline gaps of 2.64 points on PU (93.23 vs. FedCon at 90.59), 3.06 on MFPT (93.05 vs. FedCD at 89.99), and 2.31 on CWRU (91.50 vs. FedCD at 89.19); Table VI gives 6.58 on CP (86.01 vs. FedCD at 79.43). The observed range is therefore 2.31 to 6.58, not 1.15 to 7.85. Please state exactly which comparisons produce the reported endpoints.","section":"Abstract and Tables IV and VI"},{"comment":"The evaluation does not isolate the proposed algorithmic components from personalization. Section III-A states that during the fine-tuning stage each client fine-tunes the model using its locally labeled data, and Section IV-A4 says results are averaged over clients, but no analogous fine-tuning is reported for FedAvg-Supervised, FedAvg/FedProx-FixMatch/UDA, Fed-SSMPN, FedCon, or FedCD. Moreover, Algorithm 1 only exchanges prototypes (lines 9-10) and never updates a global model, so SSFL-DCSL as implemented is effectively a personalized local-model method. The reported gains could therefore stem from per-client adaptation rather than from TLAW, LCL/GCL, or PTA. A controlled comparison with the same fine-tuning protocol applied to all methods, or to none, is needed.","section":"Section III-A, Section IV-A4, and Algorithm 1"},{"comment":"The proof of Theorem 1 assumes 'without loss of generality' that exactly half of the samples satisfy max(p_i) < mu_t because mu_t is the sample mean of max(p) and hence 'P(max(p) < mu_t) = 0.5'. This conflates the sample mean with the median; for a finite sample, it is not true that exactly half of the observations fall below the mean. Equations (23)-(27) rely on splitting the sum at U/2, so the stated lower bounds on f(p) and g(p) are not established. A corrected proof, or an explicitly justified distributional assumption, is required before Theorem 1 can be used to support TLAW.","section":"Appendix A, proof of Theorem 1"},{"comment":"There is an internal inconsistency in the description of the federated algorithm. Section II specifies a standard FL loop in which local models are initialized by the global model and aggregated as theta_G <- sum_k (N_k/N) theta_k, but Algorithm 1 never performs this update: lines 9-10 only update global prototypes and local prototype sets, and Section III-C also describes only prototype communication. The role of the global model theta_G is therefore unclear. The authors should state clearly whether SSFL-DCSL is a personalized prototype-sharing method and, if so, rewrite the system model accordingly.","section":"Section II and Algorithm 1"}],"minor_comments":[{"comment":"The abstract says experiments are conducted on two publicly available datasets, while Section IV-A1 describes three openly accessible datasets (PU, MFPT, CWRU) plus the factory CP dataset; please correct the count.","section":"Abstract vs. Section IV-A1"},{"comment":"The client-dropout study in Table VII reports only the proposed method without a baseline subjected to the same dropout pattern, so the claim of high tolerance to stragglers is not comparative.","section":"Section IV-E"},{"comment":"The symbol N_i is used both for the negative-pair set in Eq. (10) and for the total number of class-j samples in Eq. (16), and P_i/P_j denote both positive-pair sets and prototypes; please use distinct symbols.","section":"Notation, Eqs. (10)-(11) and (16)"},{"comment":"References [27] and [30] appear to be the same paper (same title, same volume, same pages, same authors) with different issue months; please deduplicate.","section":"References"}],"recommendation":"reject","confidential_remarks":"To the editor: the central accuracy claim is contradicted by the paper's own tables, and the comparison appears to mix personalized local models with global-model baselines. Correcting these issues would require re-running the full experimental suite and rewriting the proof of Theorem 1, which goes beyond a minor revision. If the authors address these points in a substantially revised version, a resubmission could be reconsidered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper is worth a serious look but not worth trusting as-is. The core contribution — TLAW weighting, dual local/global contrastive loss, and momentum prototype aggregation — is a genuine combination I haven't seen in the SSFL-fault-diagnosis literature. The ablation study is well structured and shows each component adds something, with the full model hitting 93.23/93.05/91.50/86.01% on PU/MFPT/CWRU/CP at 10% labels. That's a real empirical result on four datasets, including a proprietary pump dataset.\n\nHowever, three things keep me from accepting the claims at face value.\n\nFirst, the abstract says 1.15% to 7.85% improvement over SOTA at 10% labels. The tables at χ=10% give best-baseline gaps of 2.64, 3.06, 2.31, and 6.58 for PU, MFPT, CWRU, and CP respectively. That's 2.31–6.58, not 1.15–7.85. The range is off, and I can't recover the claimed numbers from the reported results. This is a reporting error, but it's the abstract's central claim, so it matters.\n\nSecond, the proof of Theorem 1 in Appendix A has a bad step. The authors assert that because mu_t is the sample mean of max(p), P(max(p) < mu_t) = 0.5. That's only true if the sample mean equals the median, which doesn't hold in general. The lower bound f(p) ≥ lambda_max/2 is not established by that argument. The TLAW heuristic might still work empirically — the ablations suggest it does — but the theoretical justification as written is invalid.\n\nThird, and most serious, the evaluation protocol. Section III-A describes a fine-tuning stage where each client adapts its model on local labeled data before testing. Section IV-A4 says results are averaged over clients, but there's no equivalent fine-tuning step for the FedAvg/FedProx baselines. Algorithm 1 communicates only prototypes, not model weights, so SSFL-DCSL is effectively a personalized method, while the baselines are reported as a single global model. The accuracy gains could come from per-client adaptation rather than from TLAW, LCL/GCL, or PTA. The ablations partly mitigate this — they show incremental gains within the SSFL-DCSL framework — but the comparison to external baselines is not apples-to-apples as reported.\n\nAlso minor: the description of FedCon as 'FedAvg-FixMatch with group normalization and group averaging' doesn't match the FedCon paper's actual contrastive framework.\n\nSo who is this for? Researchers working on semi-supervised federated learning for fault diagnosis. The method is a reasonable engineering contribution and the empirical study is substantial, but the paper needs a major revision: fix the abstract numbers, fix or drop the Theorem 1 proof, and rerun the baselines with the same fine-tuning protocol. I'd send it to peer review rather than desk reject — it has enough substance to warrant referee time and the flaws are fixable in principle.","headline":"A genuinely new combination of TLAW weighting, dual contrastive losses, and momentum prototype aggregation, but the headline accuracy claim is not reproducible from the paper's own tables and the proof of Theorem 1 has an invalid step; the evaluation protocol also likely favors the proposed method via per-client fine-tuning.","tokens_in":23162,"tokens_out":3570,"would_cite":false,"duration_ms":38138,"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":"A semi-supervised federated framework that shares only compact prototypes can lift fault-diagnosis accuracy by 1.15% to 7.85% when just 10% of data are labeled.","keywords":["semi-supervised federated learning","intelligent fault diagnosis","contrastive learning","prototype aggregation","pseudo-label weighting","Laplace distribution","label scarcity","non-IID data"],"falsifier":"Run the strongest baselines with the same per-client fine-tuning and local test protocol described in Section III-A at a 10% label rate; if their accuracy matches SSFL-DCSL, the claim that the proposed components drive the gain is falsified. As a second test, replace the Laplace weighting with a fixed confidence threshold and check whether accuracy on the first benchmark drops by the reported 0.51–3.87 points.","tokens_in":22057,"feed_emoji":"⚙️","tokens_out":11406,"duration_ms":119193,"temperature":0.7,"pith_summary":"The paper aims to establish that federated intelligent fault diagnosis can be made accurate when each industrial client holds very few labeled samples and large amounts of unlabeled, non-identically distributed sensor data. Its framework, SSFL-DCSL, combines pseudo-label semi-supervised learning with contrastive representation learning, and communicates only compact per-class prototypes instead of full models. A truncated-Laplace sample-weighting function suppresses unreliable pseudo-labels, while a dual contrastive loss aligns features locally and against global prototypes. On three public benchmark datasets and a factory-collected pump dataset, the method reports accuracy gains of 1.15% to 7.85% over state-of-the-art baselines at a 10% label rate, with a more than 99% reduction in per-round communication. If these results replicate, factories could build shared diagnosis models from mostly unlabeled data with far less expert labeling and bandwidth.","feed_headline":"Federated fault diagnosis gains 1.15–7.85% at 10% labels","feed_subtitle":"Confidence-weighted pseudo-labels and prototype sharing close the label gap while cutting communication by 99%.","key_machinery":"The central object is the per-class prototype, the average embedding of all samples (labeled and pseudo-labeled) of a class within a client. The server fuses local prototypes as $\\tilde{P}_j = \\frac{1}{|\\mathcal{N}_j|}\\sum_{k \\in \\mathcal{N}_j}\\frac{|D^k_j|}{N_j} P^k_j$, then applies momentum $P^{t+1} = \\kappa P^{t} + (1-\\kappa)\\tilde{P}^{t+1}$. Clients pull back the global prototypes and use them in a global contrastive loss $L_{GC} = -\\sum_{i=1}^C \\log \\frac{\\exp(\\mathrm{sim}(P_i,\\tilde{P}_i)/\\tau_i)}{\\sum_{j \\neq i}\\exp(\\mathrm{sim}(P_i,\\tilde{P}_j)/\\tau_j)}$, while the local contrastive loss selects positives and negatives by pseudo-labels and uses a dynamic temperature. The TLAW weight $\\lambda(p)=\\lambda_{\\max}2b_t\\phi(\\max(p);\\mu_t,b_t)$ if $\\max(p)<\\mu_t$, and $\\lambda_{\\max}$ otherwise, where $\\phi(x;\\mu,b)=\\frac{1}{2b}e^{-|x-\\mu|/b}$ is the Laplace density and $\\mu_t,b_t$ are estimated by exponential moving average of batch confidence, controls how much each pseudo-label contributes to the unsupervised loss.","core_discovery":"The central claim is that SSFL-DCSL, a federated semi-supervised framework, achieves state-of-the-art intelligent fault diagnosis under severe label scarcity and non-IID client data by combining three design choices: a truncated Laplace-based adaptive sample weighting (TLAW) function that down-weights low-confidence pseudo-labels, a dual contrastive loss made of a local sample-level contrast term and a global prototype-level alignment term, and prototype aggregation with momentum that replaces full-model communication with per-class feature averages. The authors report that with only 10% of data labeled, SSFL-DCSL improves accuracy by 1.15% to 7.85% over the compared state-of-the-art methods across three public benchmark datasets and a chemical-plant pump dataset, while reducing per-round communication by more than 99%. The ablation study supports the claim that each component contributes, with the largest single gains coming from the global contrastive loss and the Laplace weighting at low label rates.","pith_inferences":["The reported comparison may not be apples-to-apples: the paper evaluates SSFL-DCSL as personalized models fine-tuned on each client's local labeled data, while FedAvg-style baselines are reported through their shared global model; giving baselines the same local fine-tuning could shrink the 1.15%–7.85% gap and would isolate the proposed components' true contribution.","The Laplace-distribution assumption on model confidence is testable: if the empirical histogram of $\\max(p)$ over unlabeled batches deviates strongly from Laplace, the guaranteed lower bounds on pseudo-label quantity and quality in Theorem 1 may not hold in practice.","Prototype-only communication reduces bandwidth but changes the privacy surface: per-class mean embeddings can reveal which fault classes exist and roughly how clients cluster, so quantifying the information leakage and adding formal privacy protection is a natural extension.","Because the local contrastive loss selects positives and negatives using pseudo-labels, wrong pseudo-labels also corrupt pair selection; an extension would apply the same confidence-based weighting when sampling pairs, not only when weighting the unsupervised loss."],"forward_implications":["At a 10% label rate, SSFL-DCSL reportedly matches or exceeds supervised federated training with 20% labels, so a factory could roughly halve its labeling effort for the same diagnostic accuracy.","Because clients exchange only prototype vectors (~0.06 MB per round) instead of full model weights (~8.6 MB), per-round communication drops by more than 99%, which matters for bandwidth-limited industrial links.","The method tolerates client dropout: losing one client's prototype lowers accuracy from 96.71% to 96.09% at a 40% label rate on the largest public benchmark, and losing three clients still leaves 93.57%.","The ablation path shows each proposed component contributes, with the global contrastive loss improving one benchmark by 13.03 points at 10% labels and TLAW adding 0.51–3.87 points, so the headline result is attributed to the full combination."],"supporting_citations":[{"why":"Supplies the FedAvg server-client averaging scheme that defines the supervised baseline and the FixMatch/UDA federated combinations.","marker":"[8]"},{"why":"Supplies FixMatch's confidence thresholding and weak/strong augmentation, the consistency baseline that TLAW's weighting replaces.","marker":"[31]"},{"why":"Supplies FedProx's proximal term, used in the FedProx-FixMatch and FedProx-UDA heterogeneity baselines.","marker":"[32]"},{"why":"Supplies FedCon, the instance-level contrastive federated semi-supervised baseline whose objective the dual contrastive loss extends.","marker":"[33]"},{"why":"Supplies FedCD, a dual-teacher class-aware distillation baseline that is among the strongest comparisons in the experiments.","marker":"[34]"},{"why":"Supplies SSMPN, the centralized momentum-prototype semi-supervised fault-diagnosis method extended into a federated baseline and related to prototype-based alignment.","marker":"[26]"},{"why":"Supplies unsupervised data augmentation, the consistency-regularization basis for the FedAvg-UDA and FedProx-UDA baselines.","marker":"[42]"},{"why":"Supplies the pseudo-labeling paradigm and the ramp-up schedule for the unsupervised loss weight used in local training.","marker":"[52]"},{"why":"Supplies the weak and strong time-series augmentation pair used to create the two views in the local contrastive loss.","marker":"[53]"}],"fun_headline_variants":["Federated fault diagnosis: 99% less communication, 1.15–7.85% accuracy gain","Dual contrastive learning improves federated fault diagnosis with few labels","Semi-supervised federated diagnosis: 10% labels, 99% less communication","Fault diagnosis via federated learning: dual contrastive loss and soft labels"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The results rest on comparing each client's personalized local model, fine-tuned on its own labeled data, against baselines measured by their single shared global model; if that asymmetry is removed, the reported gains may shrink.","fun_headline_variants_meta":{"raw":{"variants":["Federated fault diagnosis: 99% less communication, 1.15–7.85% accuracy gain","Dual contrastive learning improves federated fault diagnosis with few labels","Semi-supervised federated diagnosis: 10% labels, 99% less communication","Fault diagnosis via federated learning: dual contrastive loss and soft labels"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000769,"raw_usage":{"total_tokens":3459,"prompt_tokens":1047,"completion_tokens":2412,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":663,"completion_tokens_details":{"reasoning_tokens":2320}},"tokens_in":663,"tokens_out":2412,"duration_ms":21190,"temperature":1.0,"reasoning_tokens":2320,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:01:35.800445+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the strongest baselines with the same per-client fine-tuning and local test protocol described in Section III-A at a 10% label rate; if their accuracy matches SSFL-DCSL, the claim that the proposed components drive the gain is falsified. As a second test, replace the Laplace weighting with a fixed confidence threshold and check whether accuracy on the first benchmark drops by the reported 0.51–3.87 points.","supporting_citations":[{"cited_title":"Communication-efficient learning of deep networks from decentralized data,","cited_arxiv_id":null,"evidence_quote":"Supplies the FedAvg server-client averaging scheme that defines the supervised baseline and the FixMatch/UDA federated combinations."},{"cited_title":"Fixmatch: Simplifying semi- supervised learning with consistency and confidence,","cited_arxiv_id":null,"evidence_quote":"Supplies FixMatch's confidence thresholding and weak/strong augmentation, the consistency baseline that TLAW's weighting replaces."},{"cited_title":"Federated optimization in heterogeneous networks,","cited_arxiv_id":null,"evidence_quote":"Supplies FedProx's proximal term, used in the FedProx-FixMatch and FedProx-UDA heterogeneity baselines."},{"cited_title":"FedCon: A Contrastive Framework for Federated Semi-Supervised Learning","cited_arxiv_id":"2109.04533","evidence_quote":"Supplies FedCon, the instance-level contrastive federated semi-supervised baseline whose objective the dual contrastive loss extends."},{"cited_title":"Fedcd: Federated semi-supervised learning with class awareness balance via dual teachers,","cited_arxiv_id":null,"evidence_quote":"Supplies FedCD, a dual-teacher class-aware distillation baseline that is among the strongest comparisons in the experiments."},{"cited_title":"Semisupervised momentum prototype network for gearbox fault diagnosis under limited 13 labeled samples,","cited_arxiv_id":null,"evidence_quote":"Supplies SSMPN, the centralized momentum-prototype semi-supervised fault-diagnosis method extended into a federated baseline and related to prototype-based alignment."},{"cited_title":"Unsupervised data augmentation for consistency training,","cited_arxiv_id":null,"evidence_quote":"Supplies unsupervised data augmentation, the consistency-regularization basis for the FedAvg-UDA and FedProx-UDA baselines."},{"cited_title":"Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the pseudo-labeling paradigm and the ramp-up schedule for the unsupervised loss weight used in local training."},{"cited_title":"Time-series representation learning via temporal and contextual con- trasting,","cited_arxiv_id":null,"evidence_quote":"Supplies the weak and strong time-series augmentation pair used to create the two views in the local contrastive loss."}],"review_version":1}