{"id":"03b03aee-e3cb-4bd8-b981-5fb883b8d258","arxiv_id":"2505.06285","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"FE-MCFormer, a time-frequency transformer-style network with a learnable Fourier reconstruction layer, achieves state-of-the-art fault diagnosis accuracy on three rotating machinery datasets under strong noise.","lead":"This paper presents FE-MCFormer, a deep learning model that combines learnable frequency-domain filtering with multi-scale attention to detect rotating machinery faults under heavy noise. It reports higher diagnostic accuracy than seven existing methods on bearing, gearbox, and real compressor data at noise levels down to -10 dB.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (19) defines accuracy as precision, so the reported accuracy tables are unverifiable without code or a corrected metric; this supports the reader's conditional verdict.","rationale":"The reader's weakest assumption is exactly the load-bearing issue I find: the only formal metric definition in the paper, Eq. (19), is precision, while the abstract and Section 3.2.2 report accuracy numbers such as 99.11%, 98.01%, and 94.74% and even claim robustness down to -10 dB SNR. These numbers are the entire empirical basis for the claimed novelty, superiority, and the plug-and-play value of FAREL. With no code or per-class counts released, the reader cannot determine whether the reported values are accuracy or precision, so the central claim is currently unverifiable as written. I do not accuse the authors of using the wrong formula deliberately; a typo is plausible. But under the reviewing rule that in-manuscript statements count as evidence, the metric definition must be corrected or substantiated before the state-of-the-art claim can be accepted. The Gaussian noise generalization concern is real but secondary: it is a limitation shared by most noise-robustness studies and does not invalidate the controlled comparison. The apparent selection of the scaling factor gamma on the test set via Table 4 is also secondary but worth a condition. For these reasons, the appropriate verdict remains conditional, matching the reader's judgment: accept only after the metric is verified, code or confusion matrices are released, and variance or significance evidence is added.","tokens_in":18197,"tokens_out":5936,"duration_ms":60934,"concrete_test":"Obtain or reimplement the evaluation script and run it on the PU dataset with the same five train/test splits, computing both standard multiclass accuracy (correct/total) and precision (TP/(TP+FP)) at SNR = -2, -4, -6, -8, and -10 dB. Compare the published mean-acc values in Table 3 with both metrics. If the published values match the precision formula but not accuracy, or if the metric choice changes the ranking versus Li-convformer or ResNet50 at any SNR, the headline claim is not supported as written. Alternatively, if the authors release code or per-class confusion matrices, recompute the reported entries from those artifacts; a match with standard accuracy would resolve the concern.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.1 defines the evaluation metric as Acc = TP/(TP+FP) x 100% (Eq. 19), which is precision, not accuracy. Every headline number in Tables 3, 4, 6, and 9, and the associated claims of superiority over seven baselines including the FAREL plug-in gains, is reported as mean-acc or max-acc under this definition. In a multi-class setting (18 classes in Case 1, 5 classes in Cases 2 and 3), precision is not the correct classification rate; unless the implementation actually used (TP+TN)/(TP+TN+FP+FN), the central accuracy claim is unsupported. The paper supplies no code, no raw confusion matrices beyond one illustrative case, and no per-class TP/FP counts, so the discrepancy cannot be resolved from the manuscript. This is a definitional gap rather than proof of wrong numbers, but it is load-bearing because all comparative and ablation conclusions are expressed in this metric. A secondary related weakness is that only mean and max over five runs are reported, without variance or significance tests; even with a corrected metric, small margins at some SNR levels, e.g., 99.11 versus 97.53 at -2 dB, need variance evidence before claiming consistent superiority.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FE-MCFormer, a transformer-inspired architecture for rotating machinery fault diagnosis under strong noise. It introduces a Fourier Adaptive Reconstruction Embedding Layer (FAREL) that performs learnable spectral filtering, and a Multiscale Time-Frequency Fusion (MSTFF) module combining multiscale convolutional attention with frequency-domain reconstruction. The method is evaluated on three datasets (PU bearing, SEU gearbox, and a real-world centrifugal compressor dataset) with Gaussian white noise added at SNR levels from -2 to -10 dB, compared against seven deep learning baselines, and is supported by ablation studies and interpretability visualizations. The reported results indicate consistent accuracy improvements, particularly from the FAREL component, which is also presented as a plug-and-play enhancement for existing CNN models.","tokens_in":18343,"tokens_out":6395,"duration_ms":53585,"significance":"If the reported results are accurate, the paper makes a useful empirical contribution: a novel architecture that improves noise robustness, and a plug-and-play component (FAREL) that consistently improves existing models such as WDCNN. The experiments are extensive, include a real-world dataset, and the architecture is described in sufficient detail for reproduction. The inclusion of the FAREL ablation with WDCNN-FAREL provides a falsifiable claim about the component's generality. However, the central metric-definition issue and the lack of variance reporting currently prevent full verification of the quantitative claims, and the interpretability evidence is post-hoc rather than independent verification of physical mechanisms.","major_comments":[{"comment":"The evaluation metric is defined as Acc = TP/(TP+FP) × 100%, which is precision, not accuracy. All reported mean-acc and max-acc values in Tables 3, 4, 6, and 9, as well as the abstract headline numbers, depend on this definition. In the multi-class settings of this paper (18 classes in Case 1, 5 classes in Cases 2 and 3), accuracy should be computed as (TP+TN)/(TP+TN+FP+FN). Unless the implementation actually used a correct accuracy formula, the claimed superiority over the seven baselines is unverified. Please correct the definition or provide code, per-class confusion matrices, or TP/FP counts to resolve this discrepancy.","section":"Section 3.1, Eq. (19)"},{"comment":"Only the mean and maximum over five repeated runs are reported, with no standard deviation, confidence intervals, or significance tests. Several comparative margins are small at high SNR (e.g., Table 3, SNR=-2: FE-MCFormer 99.11% vs. Li-convformer 97.53%; Table 6, SNR=-8: 98.52% vs. 97.86%). Without variance information or significance testing, the claim of consistent superiority is not statistically supported. Please report per-run results, error bars, or perform statistical tests for the key comparisons.","section":"Section 3.1, experimental settings"},{"comment":"The text states that 'FE-MCFormer achieves diagnosis accuracies of 100%, 100%, 94.93%, 98.83%, and 95.51% in five trials,' but Table 6 lists mean accuracies of 95.51%, 98.52%, 99.79%, 100%, and 100% for SNR values from -10 dB to -2 dB. The five numbers in the text do not match the table entries, creating an internal inconsistency in the reporting of the main Case 2 result.","section":"Section 3.3.2, after Table 6"},{"comment":"The claim that 'smaller scaling factors γ achieve the better performance across all noise levels' is contradicted by Table 4, where FE-MCFormer-0.2 (73.58%) and FE-MCFormer-0.3 (86.41%) outperform FE-MCFormer-0.1 (72.99% and 86.31%) at SNR=-10 and SNR=-8, respectively. The text acknowledges these exceptions but still draws the opposite conclusion, so the sensitivity analysis interpretation needs to be reconciled with the reported data.","section":"Section 3.2.3, Table 4"}],"minor_comments":[{"comment":"The sentence fragment 'various tasks and data properties' appears without a verb; please rewrite the sentence.","section":"Section 2.1, after Eq. (5)"},{"comment":"The abbreviation 'MSA' is used to refer to the multiscale attention mechanism, while the Introduction uses 'MHSA' for multi-head self-attention; please disambiguate these terms.","section":"Sections 2.2.1 and 3.2.3"},{"comment":"The table header contains formatting artifacts ('Methodsssssss', '𝐽1ssssss𝐽2'); please clean up the table formatting.","section":"Table 7"},{"comment":"The phrase 'the 7200 and 3300 eddy current probes' should specify what these numbers refer to (e.g., model numbers or measurement ranges).","section":"Section 3.4.1"},{"comment":"Reference [3] has a publication year of 2026, which is in the future; please verify the citation details.","section":"References"},{"comment":"The abstract claims 'stable and interpretable diagnostic performance ... down to -10 dB SNR', but in Case 1 the mean accuracy at -10 dB is 72.99%, which is substantially lower than at higher SNR; please clarify what is meant by 'stable'.","section":"Abstract and Section 3.2.2"}],"recommendation":"major_revision","confidential_remarks":"The main barrier is the accuracy/precision definition in Eq. (19). If the authors correct this and provide code or confusion matrices to verify the reported numbers, the paper could be acceptable. The lack of variance reporting is also important but fixable with standard statistical reporting. The internal inconsistency in the Case 2 results text should be straightforward to address."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Zoe,\n\nThe paper you'll want to know about is a classic case of a good engineering idea wrapped in a sloppy metrics section. FE-MCFormer combines learnable frequency-domain filtering (FAREL), multiscale convolutional attention (MSCAL), and a time-frequency fusion feedforward network (TFFN) into a transformer-ish architecture for machinery fault diagnosis under strong noise. The FAREL plug-in is the most interesting piece: adding it to WDCNN gives consistent gains across three datasets, including a real centrifugal compressor case. That's a genuine, reproducible-sounding contribution.\n\nThe experiments are extensive: three datasets, seven baselines, ablations, sensitivity analysis on the scaling factor, and t-SNE and attention visualizations. The ablations show each component earns its keep, and the relative ordering of methods is stable across SNR levels. This is careful work in the tradition of the subfield.\n\nThe problem is that the central metric is misdefined. Eq. (19) states Acc = TP/(TP+FP) * 100%, which is precision, not accuracy. The paper then reports mean-acc and max-acc tables with numbers like 99.11% at -2 dB. If those numbers are actually accuracy computed correctly, the formula in the paper is a typo and the claim might hold. If they are precision, then every headline number is mislabeled and the state-of-the-art claim is unsupported. The paper provides no code, no confusion matrices beyond one case, and no per-class TP/FP counts, so a reader cannot tell which is true. This is a load-bearing ambiguity, not a cosmetic issue.\n\nSecondary issues: only mean and max over five runs are reported, with no variance or significance tests; small gaps (e.g., 99.11 vs 97.53 at -2 dB) need variance evidence. The scaling factor gamma is effectively selected on the test set through the ablation table, which inflates the reported performance. And the noise model is additive Gaussian white noise, which is standard but not industrial reality.\n\nI'd send this to peer review, but the reviewers should demand a corrected metric, error bars, and code or data. If the metric is fixed and the numbers hold, this is a useful contribution. As written, I wouldn't cite the accuracy values.","headline":"Solid engineering paper with a load-bearing metric problem: Eq. (19) defines accuracy as precision, so the headline accuracy tables are unverifiable.","tokens_in":18967,"tokens_out":2310,"would_cite":false,"duration_ms":20685,"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":"FE-MCFormer proposes a learnable Fourier reconstruction layer plus multiscale time-frequency fusion to keep rotating-machinery fault diagnosis accurate and interpretable at signal-to-noise ratios down to -10 dB, outperforming seven…","keywords":["fault diagnosis","rotating machinery","time-frequency transformer","Fourier adaptive reconstruction","noise robustness","interpretability","multiscale convolution","vibration signal processing"],"falsifier":"Recompute every reported result with the standard accuracy formula $\\frac{TP+TN}{TP+TN+FP+FN}$ on the same test splits; if the corrected percentages fall substantially or the method no longer leads at low SNR, the central claim fails. A second, complementary test is to run FE-MCFormer on field vibration data with colored or non-stationary noise; if its margin over baselines disappears, the robustness claim is an artifact of white-noise injection.","tokens_in":17903,"feed_emoji":"⚙️","tokens_out":10930,"duration_ms":96932,"temperature":0.7,"pith_summary":"The paper claims that fault diagnosis of rotating machinery under strong noise can be made both accurate and interpretable by coupling a learnable frequency-domain reconstruction step with a multiscale time-frequency transformer. It proposes FE-MCFormer, whose Fourier adaptive reconstruction embedding layer (FAREL) multiplies the signal's complex spectrum by a learned weight, suppresses noise-dominated frequency bins, and maps the result back to time. On three datasets — a rolling bearing benchmark, a gearbox benchmark, and a real centrifugal compressor — the method reports higher accuracy than seven deep-learning baselines at signal-to-noise ratios down to -10 dB, with best bearing accuracy 99.11% at -2 dB and 94.74% at -6 dB. The authors also argue FAREL works as a plug-and-play module, improving a WDCNN baseline whenever added.","feed_headline":"Frequency-tuned transformer diagnoses faults down to -10 dB","feed_subtitle":"Learnable Fourier reconstruction suppresses noise while keeping fault harmonics, lifting accuracy to 99.11% at -2 dB SNR.","key_machinery":"The load-bearing mechanism is FAREL (Fourier adaptive reconstruction embedding layer), a learnable spectral gate: after a $1\\times 63$ convolution, the signal is transformed with the discrete Fourier transform, multiplied element-wise by a learnable complex weight $W$, transformed back with the inverse DFT, and added back to the original feature map with a scaling factor $\\gamma$. This makes the network learn which frequency bins carry fault information and which carry noise. The same Fourier adaptive reconstruction appears inside the time-frequency fusion network (TFFN), and the multiscale time-frequency fusion (MSTFF) module's multiscale convolutional attention layer (MSCAL) supplies the local time-domain branch with $1\\times 3$ and $1\\times 5$ kernels, softmax attention, and residual connections; distillation layers halve the time dimension to widen the receptive field.","core_discovery":"FE-MCFormer establishes that a transformer for machinery fault diagnosis does not need conventional multi-head self-attention to excel under heavy noise. The paper's central discovery is that replacing multi-head self-attention with a multiscale convolutional attention layer and inserting Fourier adaptive reconstruction at the embedding and feed-forward stages lets the network learn which frequency components carry fault information: it attenuates noise-dominated bins, preserves fault harmonics, and concentrates attention on impulsive fault regions. The evidence is the reported accuracy at SNR values from -2 to -10 dB on the bearing and gearbox datasets, plus feature-quality metrics and t-SNE and heatmap visualizations showing tighter clustering and more focused attention than baselines. On the real compressor data, the method reports 96.89% mean accuracy despite only 240 samples per class.","pith_inferences":["Although the paper only demonstrates FAREL as a plug-in for WDCNN, the same mechanism could be inserted into other CNN or transformer backbones; a systematic study across architectures would establish how general the improvement is.","The learned frequency mask that FAREL produces could be extracted and applied as a standalone pre-processing denoiser before any classifier, which the paper hints at but does not formalize.","Because all noise injections are Gaussian white noise, a direct test would add colored or non-stationary noise from variable-speed machinery; the 'strong noise' claim is only shown for white-noise contamination."],"forward_implications":["If the reported results hold, FE-MCFormer is the best among the compared methods on all three datasets at every SNR from -2 dB to -10 dB, with the largest gains at the lowest signal-to-noise ratios.","FAREL is claimed to be a general plug-and-play component: adding it to a WDCNN baseline raised mean accuracy by 4.31 to 9.92 percentage points across SNR levels on the bearing dataset.","The method maintains usable accuracy in extreme noise, for example 72.99% mean accuracy on the bearing dataset at -10 dB SNR, where the strongest transformer baseline reaches 60.65%.","On the small-sample real compressor dataset, FE-MCFormer reports 96.89% mean accuracy, suggesting the architecture transfers to field data with limited training samples.","The interpretability claims imply the model's attention and reconstructed spectra can be read as physical evidence: noise bands are attenuated and harmonics at $1\\times f$, $2\\times f$, and $3\\times f$ are enhanced."],"supporting_citations":[{"why":"Supplies the rolling-bearing vibration dataset used in Case 1, with 18 health states and 64 kHz sampling; the central accuracy comparisons are run on this data.","marker":"[29]"},{"why":"Supplies the gearbox dataset used in Case 2, with five fault classes and two operating conditions; supports the claim of cross-dataset generality.","marker":"[34]"},{"why":"Provides the Li-convformer transformer baseline and the between-class and within-class covariance metrics used to judge feature quality.","marker":"[33]"},{"why":"Provides the MSCNN-LSTM multiscale convolutional baseline that FE-MCFormer is compared against.","marker":"[30]"},{"why":"Provides the MA1DCNN multi-attention one-dimensional convolutional baseline.","marker":"[31]"},{"why":"Provides the open benchmark that supplies the ResNet50 and DenseNet baselines used in the comparison.","marker":"[32]"}],"fun_headline_variants":["Fourier attention beats self-attention for fault diagnosis under noise","No self-attention needed: Fourier filter lifts fault diagnosis accuracy","FE-MCFormer: Fourier filtering makes transformer diagnosis noise-robust","Replace self-attention with Fourier for fault detection down to -10 dB"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The headline percentages are only as strong as the accuracy metric in Eq. (19), which defines Acc as $TP/(TP+FP)$ — precision, not accuracy — and the noise-robustness results are only as strong as the assumption that Gaussian white noise added via Eq. (20) represents real industrial strong noise.","fun_headline_variants_meta":{"raw":{"variants":["Fourier attention beats self-attention for fault diagnosis under noise","No self-attention needed: Fourier filter lifts fault diagnosis accuracy","FE-MCFormer: Fourier filtering makes transformer diagnosis noise-robust","Replace self-attention with Fourier for fault detection down to -10 dB"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00059,"raw_usage":{"total_tokens":2739,"prompt_tokens":888,"completion_tokens":1851,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":504,"completion_tokens_details":{"reasoning_tokens":1775}},"tokens_in":504,"tokens_out":1851,"duration_ms":13575,"temperature":1.0,"reasoning_tokens":1775,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:34:38.383666+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute every reported result with the standard accuracy formula $\\frac{TP+TN}{TP+TN+FP+FN}$ on the same test splits; if the corrected percentages fall substantially or the method no longer leads at low SNR, the central claim fails. A second, complementary test is to run FE-MCFormer on field vibration data with colored or non-stationary noise; if its margin over baselines disappears, the robustness claim is an artifact of white-noise injection.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the rolling-bearing vibration dataset used in Case 1, with 18 health states and 64 kHz sampling; the central accuracy comparisons are run on this data."},{"cited_title":"Highly accurate machine fault diagnosis using deep transfer learning","cited_arxiv_id":null,"evidence_quote":"Supplies the gearbox dataset used in Case 2, with five fault classes and two operating conditions; supports the claim of cross-dataset generality."},{"cited_title":"Bearing fault diagnosis base on multi-scale cnn and lstm model","cited_arxiv_id":null,"evidence_quote":"Provides the MSCNN-LSTM multiscale convolutional baseline that FE-MCFormer is compared against."},{"cited_title":"Understanding and learning discriminant features based on multiattention 1dcnn for wheelset bearing fault diagnosis","cited_arxiv_id":null,"evidence_quote":"Provides the MA1DCNN multi-attention one-dimensional convolutional baseline."},{"cited_title":"Deeplearningalgorithmsforrotatingmachineryintelligentdiagnosis: An open source benchmark study","cited_arxiv_id":null,"evidence_quote":"Provides the open benchmark that supplies the ResNet50 and DenseNet baselines used in the comparison."}],"review_version":1}