{"id":"3b599723-af6a-4c6f-ac00-5ede4d540737","arxiv_id":"2411.15191","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Hyperparameter tuning strongly affects bearing-fault-detection neural networks, and filtered-data experiments show that high-frequency noise is not the main reason wide convolutional kernels perform well.","lead":"This paper tests how much a neural network's settings matter for detecting faults in machine bearings, and finds that the right settings change with the data. It also shows that the usual explanation for why wide detection filters work, removing high-frequency noise, is probably wrong.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Resampling experiment conflates sampling rate with kernel duration; the headline negative result about sampling rate does not follow from the reported design.","rationale":"The reader's weakest_assumption exactly identifies the same confound: Section III-C5 keeps kernel sizes in sample count while resampling changes sampling rate, so kernel duration in physical time changes across conditions. The paper's own text in Section III-C5 acknowledges that resampling changes the time represented by each data point, and Section IV-B4 interprets the box plots as showing no trend toward smaller kernels being better under resampling. But the comparison is not apples-to-apples: a kernel of 256 samples at 3 kHz covers roughly 16 times the physical time span of a 256-sample kernel at 48 kHz. The apparent persistence of large-kernel preference could simply reflect that larger physical durations remain beneficial when the data is more compressed. The filtering experiment (Section IV-B5, Figure 7) does not share this confound because filtering holds both sampling rate and sequence length constant, so the high-frequency-noise half of the central claim is credible. The sampling-rate half, however, is the load-bearing part of the abstract's negative claim ('sampling rate and the presence of high-frequency noise are not the main reason' in Section II-C), and it is not established by the reported experiment. The paper partially acknowledges limitations in Section V (seven datasets, early stopping), but does not flag this confound, and the conclusion in Section V states both factors together as ruled out. A conditional verdict is appropriate because the fix is a well-defined experimental variant (scaling kernel size with resampling) rather than a theoretical error; the filtering result and the multiple-defaults analysis remain useful. Agreement with the reader is full: their weakest_assumption is precisely the resampling/duration confound I would stress-test.","tokens_in":31736,"tokens_out":2014,"duration_ms":17968,"concrete_test":"Re-run the resampling experiment with kernel sizes scaled to hold physical kernel duration constant across sampling rates: at half sampling rate use half the kernel size in samples, so the duration in milliseconds is the same. Compare the best kernel size and the accuracy gap between longest and shortest kernels across 48, 24, 12, 6, and 3 kHz versions. If the wide-kernel advantage disappears when kernel duration is held constant, the sampling-rate conclusion in Section IV-B4 is an artifact of the confound; if it persists, the conclusion is supported.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that sampling rate and high-frequency noise are not the main reason wide kernels outperform narrow kernels (Abstract; Section IV-B5; Section V). The filtering half of this claim is well supported: keeping time-series length and sampling rate fixed while low-pass filtering at 12 kHz down to 187 Hz leaves the wide-kernel advantage intact (Figure 7), so removing high-frequency content does not remove the need for wide kernels. However, the resampling half is confounded. In Section III-C5, windows of 4096 samples are resampled from 48 kHz to 24, 12, 6, and 3 kHz, and the grid search keeps kernel sizes fixed in sample count (16-256). Resampling changes the physical time span each kernel covers: at 48 kHz a kernel of 256 samples covers 5.33 ms, while at 3 kHz it covers 85.3 ms. The paper interprets the persistence of large-kernel preference under resampling as evidence that sampling rate is irrelevant, but an alternative explanation is that the relevant quantity is kernel duration, not sample count. If kernels had been scaled proportionally with resampling (e.g., halving kernel size when halving sampling rate), the wide-kernel advantage might have disappeared. Thus the conclusion that sampling rate is not responsible for wide-kernel preference does not follow from this experiment. The paper's own framing in Section III-C5 says resampling 'will take fewer data points to record the same back-and-forth vibrations,' which is precisely the confound: keeping kernel size in samples fixed while changing the sampling rate changes which physical vibrations each kernel sees. The claim that the wide-kernel advantage is not explained by sampling rate therefore rests primarily on the filtering experiment, which isolates spectral content, not sampling rate per se. A reader relying on the strong negative claim needs the resampling dimension settled; as reported it is not.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies hyperparameter sensitivity of a wide-kernel CNN for bearing fault detection across seven benchmark datasets, two additional architectures (LSTM, transformer), and manipulated copies of the CWRU dataset. It proposes a sequence of 'multiple defaults' for efficiently tuning the architecture on new data, and it uses resampling and low-pass filtering to investigate whether sampling rate or high-frequency noise explains why wide first-layer kernels are often preferable to narrow ones. The central negative claim is that, contrary to earlier speculation, high-frequency noise is not the main reason wide kernels outperform narrow kernels; a secondary claim is that sampling rate is also not responsible. The filtering experiment is a clean falsifiable test and supports the noise-related claim; however, the resampling experiment is confounded because kernel sizes are kept fixed in sample count while the physical duration they cover changes with sampling rate.","tokens_in":32048,"tokens_out":6166,"duration_ms":61310,"significance":"If the negative result on high-frequency noise holds, it is a useful contribution: it refutes a commonly repeated explanation for wide-kernel CNNs and redirects attention to other mechanisms such as implicit regularization or training dynamics. The filtering experiment is well designed and gives a concrete, falsifiable test, and the multiple-defaults procedure with leave-one-out evaluation is a practical and reproducible way to give tuning guidance across datasets. The main weakness is that the sampling-rate half of the central claim currently rests on an experiment that conflates sample count with kernel duration, so the scope of the paper's headline conclusion is wider than the evidence supports.","major_comments":[{"comment":"The resampling experiment does not support the conclusion that sampling rate is irrelevant to the wide-kernel preference. The grid search keeps kernel sizes fixed in sample count (16-256) across resampling levels, so each kernel covers a different physical duration at each rate: a 256-sample kernel covers 5.33 ms at 48 kHz but 85.3 ms at 3 kHz. The persistence of the large-kernel advantage under resampling is exactly what would be expected if the relevant quantity were kernel duration or the number of signal cycles spanned, rather than sample count. To test the sampling-rate hypothesis, the authors should either scale kernel sizes inversely with the resampling factor to keep duration constant, or re-analyze the existing grid by converting kernel sizes to physical durations and checking whether optimal duration is invariant across rates. Until then, the resampling half of the central claim remains ambiguous.","section":"Section III-C5 / Figure 5"},{"comment":"The filtering experiment is convincing for the noise hypothesis: removing high frequencies while holding sampling rate and window length fixed does not remove the wide-kernel advantage, so pre-filtering does not make narrow kernels viable. However, the Abstract and Conclusion state the joint claim that 'neither the sampling rate nor noise is responsible' for the wide-kernel preference. The noise part is supported, but the sampling-rate part currently rests on the confounded resampling experiment described above. The wording should be scoped to what is demonstrated, or the resampling experiment should be extended with a duration-matched control.","section":"Abstract / Section IV-B5 / Section V"}],"minor_comments":[{"comment":"The filtering thresholds '750 kHz', '375 kHz', '187 kHz', '93 kHz', and '46 kHz' should be labeled in Hz, not kHz; as written they exceed the Nyquist frequency of the 48 kHz signal.","section":"Table 4"},{"comment":"The phrase 'it will take fewer data points to record the same back-and-forth vibrations' accurately describes the confound: after resampling, each sample covers a larger physical time step. Consider making this explicit and discussing its consequences for comparing kernel sizes in samples.","section":"Section III-C5"},{"comment":"The correlation analysis uses Pearson correlations on accuracy scores that are heavily ceiling-affected. The caveat about ceiling effects is appreciated, but a rank-based correlation would be more robust for the claim that hyperparameter rankings are preserved across resampling conditions.","section":"Section IV-B4 / Figure 4"},{"comment":"Minor typographical issues include 'University of Conneticut' and 'the data are collected with a sampling frequency' in Appendix A; these should be corrected.","section":"Appendix A / Appendix B"},{"comment":"No code or data availability statement is included. Given that the paper relies on a large grid search and a reproducible multiple-defaults procedure, adding a link to the code and processed data would strengthen the practical contribution.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The filtering experiment is solid and the multiple-defaults contribution is useful, but the resampling confound directly affects the paper's headline conclusion about sampling rate. If the authors add a duration-matched analysis or explicitly rescope the conclusion to the noise result, I would be willing to support acceptance. I do not see a novelty or attribution issue."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know before you read this. First, the filtering experiments are the real contribution: low-pass filtering CWRU down to 187 Hz doesn't remove the wide-kernel advantage, which is a clean falsifiable test against the 'wide kernels are just low-pass filters' story. That negative result is new and should survive review. Second, the resampling experiment is confounded: kernel sizes are kept fixed in sample count while the sampling rate drops from 48 kHz to 3 kHz, so each kernel covers a longer physical time window. The paper interprets this as sampling rate being irrelevant, but it actually shows kernel duration matters. The confound is acknowledged in the stress-test note and also visible in the paper's own framing that resampling 'will take fewer data points to record the same back-and-forth vibrations.' So the strong claim that sampling rate is not responsible for wide-kernel preference should not rest on that experiment.\n\nWhat the paper does well: the multiple-defaults procedure is practical, the leave-one-out evaluation is honest about small sample size, and the comparison across LSTM and transformer architectures strengthens the point that hyperparameter tuning is broadly important. The authors clearly state limitations: domain-specific advice, only seven datasets, no statistically robust sample. That's good scientific practice.\n\nSoft spots: besides the resampling confound, accuracy scores come from single training runs per configuration—no repeated-run error bars—so differences of a few percent between kernel sizes may be noise. The leave-one-out estimate uses only seven folds, so the 92nd-percentile claim is fragile. There are also minor typos in Table 4 (kHz vs Hz) that should be fixed.\n\nWho is this for? Practitioners applying wide-kernel CNNs to vibration data, and researchers studying hyperparameter transfer. It deserves a serious referee, but the resampling section needs to be rewritten to either scale kernel sizes with the sampling rate or explicitly limit claims to what the experiment can support. With those changes, the filtering result alone is worth publishing.","headline":"The filtering experiments are a clean negative result that should survive review; the resampling experiment is confounded and should not be used to support claims about sampling rate.","tokens_in":32587,"tokens_out":1921,"would_cite":false,"duration_ms":18565,"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":"Contrary to earlier speculation, high-frequency noise is not why a wide convolutional kernel beats a narrow one in bearing-fault detection.","keywords":["bearing fault detection","wide-kernel CNN","hyperparameter tuning","time series classification","multiple defaults","resampling","low-pass filtering","convolutional neural network"],"falsifier":"Run the same grid search on CWRU at 48, 24, 12, 6 and 3 kHz with first-layer kernel sizes scaled so that each kernel covers the same physical duration in milliseconds (for example, 256 samples at 48 kHz would be matched by 128 samples at 24 kHz and 64 samples at 12 kHz). If, under constant-duration kernels, wide kernels stop beating narrow kernels or the optimal kernel size tracks the sampling rate, the paper's conclusion that sampling rate is not responsible would be overturned.","tokens_in":31566,"feed_emoji":"⚙️","tokens_out":9336,"duration_ms":95490,"temperature":0.7,"pith_summary":"This paper studies why the width of the first convolutional kernel matters so much for neural networks that detect damaged machine bearings from vibration signals. Using seven benchmark datasets, the authors confirm that poor hyperparameter choices can cut accuracy dramatically and that the best settings differ between datasets; the first-layer kernel size is the most data-dependent choice. Their central finding is negative: low-pass filtering that strips high frequencies from the signal does not remove the advantage of wide kernels, so high-frequency noise is not the main reason wide kernels beat narrow ones. That conclusion matters because it redirects research toward other mechanisms, such as implicit regularization or training dynamics, and toward practical tuning guidance for new data. The paper's concrete deliverable is a small ranked list of hyperparameter defaults, built by exhaustive search, that reaches high percentile performance on unseen datasets.","feed_headline":"Wide kernels keep their edge after noise is filtered out","feed_subtitle":"A seven-dataset bearing study refutes the noise-filtering story behind wide-kernel CNNs.","key_machinery":"The central object is the first-layer kernel of a 1D wide-kernel CNN: a convolutional filter whose length in samples (16, 32, 64, 128 or 256) is the hyperparameter whose optimal value shifts most across datasets. The argument is carried by manipulated copies of the CWRU vibration dataset. Resampling condenses the signal while keeping kernel sizes fixed in sample count, and low-pass filtering removes high frequencies while keeping sample rate and window length identical; grid searches over the three most important hyperparameters on each version produce the box plots and correlation matrices that distinguish sampling-rate effects from spectral-content effects. A separate exhaustive search over the completed seven-dataset grid produces a ranked 'multiple defaults' list, a small set of hyperparameter combinations to try in sequence on new data.","core_discovery":"On the paper's own terms, the discovery is that the standard explanation for wide-kernel convolutional networks in bearing fault detection—that a long filter averages over high-frequency noise—is not supported. In experiments on the CWRU benchmark, copies of the data were low-pass filtered at cutoffs from 12 kHz down to 46 Hz and resampled from 48 kHz to 3 kHz; in all but the most destructive filtering conditions, first-layer kernels of 128 or 256 samples continued to outperform short kernels. Because the wide-kernel advantage survives the removal of high frequencies, the authors conclude that high-frequency noise is not the main reason wide kernels are preferable. They are left speculating that the advantage may be a generic architectural property such as regularization that helps training, while noting that this does not explain why the advantage reverses on the SEU dataset. The same experiments support a practical claim: resampling to 24 kHz caused no visible drop in accuracy, so the highest recording rates may be wasteful.","pith_inferences":["Not pursued in the paper: the resampling experiment conflates sample-count kernel size with physical kernel duration; a cleaner test would scale kernel length with the sampling period to see whether duration, not sampling rate, controls the wide-kernel advantage.","If the wide-kernel advantage is not about frequencies, the mechanism may be a longer receptive field acting as implicit regularization; this could be tested by comparing wide and narrow kernels on synthetic signals whose fault signature and noise are generated with known spectra.","The multiple-defaults procedure could be applied to the LSTM and transformer results to produce per-architecture default lists, since the same exhaustive-search principle only needs a completed performance table.","The filtering sweep suggests a cheap diagnostic: if aggressive low-pass filtering does not hurt accuracy, the fault signature is concentrated at low frequencies and high-rate acquisition is unnecessary; this could be checked on the other six benchmarks, not just CWRU."],"forward_implications":["Resampling a 48 kHz bearing signal to 24 kHz showed no accuracy loss, so recording at the highest rates may be unnecessary for this architecture.","No single hyperparameter configuration dominates all seven benchmarks; users should expect to re-tune first-layer kernel size when moving to new data.","The recommended tuning order for the wide-kernel CNN is first-layer kernel size, then first-layer filter count, then filter count in layers 3-5, because tuning kernel size later forces re-tuning of the others.","The susceptibility to bad hyperparameters is not specific to wide-kernel CNNs: both LSTM and transformer fault detectors varied from poor to near-perfect accuracy depending on configuration.","Multiple defaults, tried in sequence, reach the 92nd percentile of grid-search performance on held-out benchmarks, so a practitioner can find a strong configuration quickly."],"supporting_citations":[{"why":"Original wide-kernel 1D CNN for raw vibration signals, promoted as having good anti-noise ability; the claim being tested originates here.","marker":"[35]"},{"why":"Explains how bearing vibrations travel through machine components and can carry irrelevant or misleading information, the rationale for the high-frequency-noise hypothesis.","marker":"[12]"},{"why":"Previous hyperparameter study of the same architecture that found first-layer kernel size matters, especially at high sampling rates; the paper's experiments are designed to test that connection.","marker":"[2]"},{"why":"Prior seven-dataset grid search whose results the paper re-analyses and extends; provides the accuracy table and the influence scores.","marker":"[1]"},{"why":"Introduces the multiple-defaults idea for hyperparameter transfer; the paper adapts it with an exhaustive search over completed grid results.","marker":"[8]"},{"why":"Supplies the LSTM architecture used to show hyperparameter sensitivity extends beyond the wide-kernel CNN.","marker":"[54]"},{"why":"Supplies the transformer architecture used in the same cross-architecture check.","marker":"[38]"}],"fun_headline_variants":["Wide kernels beat narrow ones even without high-frequency noise","High-frequency noise isn't the reason wide kernels win","Wide-kernel advantage survives low-pass filtering","Wide kernels stay superior after noise is removed","Wide kernels don't rely on high-frequency noise to win"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The weakest load-bearing premise is that a first-layer kernel with a fixed number of samples is the same object after resampling: the experiment keeps kernel size in samples while changing the sampling rate, so the physical time span the kernel covers changes together with sampling rate, and the claim that sampling rate is not responsible for the wide-kernel preference depends on that comparison being meaningful.","fun_headline_variants_meta":{"raw":{"variants":["Wide kernels beat narrow ones even without high-frequency noise","High-frequency noise isn't the reason wide kernels win","Wide-kernel advantage survives low-pass filtering","Wide kernels stay superior after noise is removed","Wide kernels don't rely on high-frequency noise to win"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000231,"raw_usage":{"total_tokens":1518,"prompt_tokens":1012,"completion_tokens":506,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":628,"completion_tokens_details":{"reasoning_tokens":429}},"tokens_in":628,"tokens_out":506,"duration_ms":5266,"temperature":1.0,"reasoning_tokens":429,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T17:36:52.361801+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same grid search on CWRU at 48, 24, 12, 6 and 3 kHz with first-layer kernel sizes scaled so that each kernel covers the same physical duration in milliseconds (for example, 256 samples at 48 kHz would be matched by 128 samples at 24 kHz and 64 samples at 12 kHz). If, under constant-duration kernels, wide kernels stop beating narrow kernels or the optimal kernel size tracks the sampling rate, the paper's conclusion that sampling rate is not responsible would be overturned.","supporting_citations":[{"cited_title":"Zhang, G","cited_arxiv_id":null,"evidence_quote":"Original wide-kernel 1D CNN for raw vibration signals, promoted as having good anti-noise ability; the claim being tested originates here."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Explains how bearing vibrations travel through machine components and can carry irrelevant or misleading information, the rationale for the high-frequency-noise hypothesis."},{"cited_title":"van den Hoogen, D","cited_arxiv_id":null,"evidence_quote":"Previous hyperparameter study of the same architecture that found first-layer kernel size matters, especially at high sampling rates; the paper's experiments are designed to test that connection."},{"cited_title":"Hudson, J","cited_arxiv_id":null,"evidence_quote":"Prior seven-dataset grid search whose results the paper re-analyses and extends; provides the accuracy table and the influence scores."},{"cited_title":"Pfisterer, J","cited_arxiv_id":null,"evidence_quote":"Introduces the multiple-defaults idea for hyperparameter transfer; the paper adapts it with an exhaustive search over completed grid results."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the LSTM architecture used to show hyperparameter sensitivity extends beyond the wide-kernel CNN."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the transformer architecture used in the same cross-architecture check."}],"review_version":1}