{"id":"4ee1ea41-1829-4398-a132-9c29a5b3ec00","arxiv_id":"2505.23097","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A bi-residual network with intra-layer shortcuts and multi-scale convolutions improves synchronous motor fault diagnosis accuracy on low-resolution noisy data by about 1 to 3 percent over ResNet18.","lead":"This paper proposes Bi-ResNet, a shallow neural network with within-layer shortcut connections (intra-links) and multi-scale convolution blocks for diagnosing electrical faults in synchronous motors. It claims this design extracts high-frequency fault features without adding parameters, improving accuracy on low-resolution, noisy signals compared with standard CNNs and ResNet18.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The empirical ranking may be driven by temporal leakage: the split granularity (sample vs experiment) is unspecified, and the reported margins (e.g., 94.15% vs 93.87% at SNR -5) are small enough that within-experiment correlation could explain them.","rationale":"The paper's core empirical assertion is that Bi-ResNet outperforms ResNet18 and other baselines on low-resolution, noisy synchronous motor fault data, and that intra-links are the cause. The evidence is a set of single-run test accuracies (Tables IV, V, VI, VIII, IX) with no variance estimates. The most damaging ambiguity is the split granularity: Section IV.B reports only a 0.8:0.1:0.1 ratio, while Section IV.A describes each fault experiment as a one-second event with 10,000 samples. In time-series classification, splitting at the sample or window level mixes overlapping segments of the same event between training and test, which artificially boosts accuracy and compresses differences between models. The reported margins are small enough (0.28 percentage points at SNR -5; 0.65 pp at 20 ms with noise) that they could easily be an artifact of such leakage or of random seed variation. The absence of error bars or multiple seeds means the reader cannot assess this. The occlusion experiment (Figure 8) is qualitative, with a single example, and does not establish the high-frequency localization claim quantitatively. The ablation in Table IX shows a 0.07 pp drop when the spatial-temporal block is removed, which is within noise and thus does not support the necessity of that component; however, that is secondary. The load-bearing fix is to re-evaluate with an experiment-level split and repeated seeds. If the margin persists, the central claim is supported; if not, the paper requires major revision. Therefore the verdict remains conditional pending this check.","tokens_in":12856,"tokens_out":6701,"duration_ms":62660,"concrete_test":"Re-run the low-resolution noisy comparisons (Table IV noise rows, Table V) with a grouped split: assign all 10,000 samples of a given fault experiment to the same fold, then split 0.8/0.1/0.1 at the experiment level, repeating with 10 random seeds and reporting mean±std. If Bi-ResNet's accuracy margin over ResNet18 falls within the seed noise or reverses, the headline claim is not supported by the current evidence.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing assumption is that the 0.8:0.1:0.1 split is performed at the experiment level. Section IV.B states only the ratio, not the split unit, while Section IV.A describes each fault experiment as a one-second event containing 10,000 samples. If the split is applied to individual samples or windows, the same fault event contributes highly correlated, overlapping segments to both training and test sets. This temporal leakage would inflate all reported accuracies and, more importantly, could erase the small margins that carry the central claim: at SNR -5 and 10 ms, Bi-ResNet scores 94.15% versus 93.87% for ResNet18; at 20 ms with noise, 94.82% versus 94.17%. No error bars, confidence intervals, or multiple-seed results are reported, so the observed ranking cannot be distinguished from split-induced correlation or random seed variation. Because the paper's conclusion that intra-links provide a high-frequency extraction advantage rests on these comparative tables, the missing split granularity is load-bearing.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Bi-ResNet, a residual network with intra-layer shortcut connections (intra-links) and a multi-scale spatial-temporal convolution block, for synchronous motor electrical fault diagnosis. The central claim is that intra-links act as a parameter-free high-frequency feature extractor, enabling a shallow network to match or exceed a deep ResNet18 on noisy, low-resolution fault data. Experiments compare six models across sample times (1-20 ms) and SNR levels (-5 to 5 dB) on a motor fault dataset, plus ablation studies varying the intra-link count n and the presence of the spatial-temporal block. Reported results show Bi-ResNet achieving the highest test accuracy in most settings, e.g., 94.15% vs 93.87% for ResNet18 at SNR -5 (Table V).","tokens_in":13111,"tokens_out":7605,"duration_ms":69078,"significance":"If the central claims hold, the parameter-free intra-link design would be an interesting contribution for resource-constrained fault diagnosis, and the systematic evaluation across multiple resolutions and SNR levels is a strength. The paper provides a real-life dataset, multiple baselines, and ablation experiments, and the intra-link idea is simple and potentially extensible. However, the significance is currently limited by the absence of statistical rigor (no error bars, confidence intervals, or multiple-seed runs), an unspecified data-split granularity that creates a temporal-leakage risk, and a high-frequency mechanism that is asserted rather than directly measured. The reported margins over ResNet18 are often small (e.g., 0.28 percentage points at SNR -5), which makes the comparative ranking vulnerable to implementation and evaluation details.","major_comments":[{"comment":"The dataset split is stated only as a 0.8:0.1:0.1 ratio (Section IV.B), with no indication of whether the split is performed at the experiment level (one-second events) or at the sample/window level. Section IV.A states that each experiment contains 10,000 samples from the same one-second fault event. If the split is sample- or window-level, highly correlated or overlapping segments from the same event appear in both training and test sets, inflating all reported accuracies and potentially erasing the small margins over ResNet18 (e.g., 94.15% vs 93.87% at SNR -5 in Table V). No error bars, confidence intervals, or multiple-seed runs are reported anywhere in the paper, so the comparative ranking cannot be distinguished from split-induced leakage or random seed variation. This issue is load-bearing because the paper's central empirical claim is precisely this ranking.","section":"Section IV.B, Section IV.A"},{"comment":"The text states that 'Bi-ResNet takes the lead by the largest margin, i.e., surpassing the second place ResNet18 by over 1% and the third place by about 3%'. However, Table V shows that at SNR -5 the margin over ResNet18 is 0.28 percentage points (94.15% vs 93.87%), and margins over 1% occur only at SNR -3 and -1. The claim as written overstates the advantage in the most challenging condition, which is exactly the condition the paper emphasizes.","section":"Section IV.D.1, Table V"},{"comment":"The paper repeatedly states that intra-links are a 'built-in high-frequency extractor' (Abstract, Section I, Section III.A) and that residual learning captures high-frequency components, but no frequency-domain analysis is provided to support this claim. The occlusion experiment (Fig. 8) demonstrates feature localization, not frequency extraction, and the ablation in Table IX shows only a 0.07 percentage point difference (99.14% vs 99.07%) when the spatial-temporal block is removed, which is within likely noise. To substantiate the central mechanism, the authors should measure the frequency content of learned features (e.g., via spectral analysis of feature maps or input-output sensitivity) and compare intra-link variants against a baseline with matched capacity.","section":"Section IV.C, Section IV.D.2, Section IV.E.2"},{"comment":"Eq. (4) defines the intra-link pre-activation g_j^i as depending on g_{j+1}^i, creating a sequential chain within a layer. The paper does not specify how this chain is computed in the forward pass (e.g., order, boundary handling) or how gradients flow through this recurrence during backpropagation. In addition, the described mode n=1 ('neurons connected two by two') does not match the chain structure in the left panel of Figure 3, which appears to show a full sequential mode. This ambiguity prevents reproduction of the architecture and makes it unclear whether the reported improvements are due to the intra-link connectivity itself or to implementation details.","section":"Section III.A, Eq. (4)"},{"comment":"The text claims that 'with the increase of n, the test accuracy also increases in all models', but Table VIII shows non-monotonic behavior: e.g., CNN-2 peaks at n=2 (96.94%) and drops to 96.14% at n=4, and Bi-ResNet peaks at n=2 (99.32%) before declining to 98.89% at n=4. The subsequent explanation that 'the potential computing power is dispersed into useless high-frequency component extraction' is post-hoc and not supported by any measurement. The asserted trade-off between intra-link number n and input data complexity is not established by the data presented.","section":"Section IV.E.1, Table VIII"}],"minor_comments":[{"comment":"There are several typos and wording issues: 'PRMOF' in the Table I caption, 'SRNs' in the Table V caption, 'bath-normalization' in the Figure 5 description, 'access' instead of 'assess' in the opening of Section IV, and 'trade-off trade-off' in the Conclusion.","section":"Global"},{"comment":"The paper states that each experiment contains 10,000 samples and 10 features, but it does not describe how these samples are formed into input examples (window length, stride, normalization, and whether windows overlap). This information is essential for reproducibility and for interpreting the leakage risk.","section":"Section IV.A"},{"comment":"Table I lists only CNN-1, CNN-2, CNN-3, Bi-ResNet, and ResNet18; the two baselines CNN-4 and LSTM are omitted. Their parameter counts should be included, or the table caption should explain why they are absent.","section":"Table I"},{"comment":"The notation 'Bi-residual block (*2)' appears four times with filter sizes 32, 64, 128, and 256. Please clarify whether each stage contains two repeated blocks, and specify the number of intra-links (n) used in each stage.","section":"Table III"},{"comment":"The paper counts 'five advanced CNN-based models and two baselines', but the model list includes CNN-4, which is also a CNN baseline. The taxonomy of advanced models versus baselines is confusing and should be clarified.","section":"Abstract and Section IV"},{"comment":"The 'spatial-temporal' block is implemented with 1D convolutions; the term 'spatial' is not defined. Please clarify how the spatial dimension is represented (e.g., treating feature channels as spatial locations).","section":"Section III.B"},{"comment":"The paper motivates the work by resource-constrained environments, but it does not report training/inference time or FLOPs for the compared models. Adding a computational-cost comparison would strengthen the practical claims.","section":"Section IV.B"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's empirical foundation is fragile. The most important issue is the unspecified split granularity: if the authors confirm the split is at the sample or window level rather than at the experiment level, the reported accuracies would be inflated by temporal leakage and the paper would require substantial re-analysis. Even if the split is correct, the lack of error bars or multiple seeds makes the small margins over ResNet18 uninterpretable. The high-frequency narrative is not backed by direct evidence, and the dataset is from the same research group (Ref. [30] includes two co-authors of this paper), which is acceptable but should be acknowledged. If the split turns out to be sample-level, I would recommend rejection; otherwise, major revision with the requested experiments and clarifications seems appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The real news here is the application of intra-links, borrowed from [24], to 1D motor fault signals, plus a multi-scale spatial-temporal block. The paper claims this lets a small network (1.05M params) match or beat ResNet18 (11.2M) on low-resolution, noisy data. That is a practical claim worth checking, and the paper does some things well: it compares against five CNN baselines and two simple ones, runs ablations on the intra-link count n, and shows the intra-link idea transfers to other architectures. The parameter efficiency and the performance at 10ms and 20ms sampling are genuinely interesting. The soft spots are real, though. First, the paper never says whether the 0.8:0.1:0.1 split is at the experiment level or the sample level. Section IV.A says each experiment is one second of data with 10,000 samples; a sample-level split would let the same fault event appear in both training and test, which would inflate all accuracies and could easily explain the small margins over ResNet18 (e.g., 94.15% vs 93.87% at SNR -5). This is not a minor omission; it is load-bearing for the main empirical claim. Second, no error bars, confidence intervals, or multiple seeds are reported, so we cannot distinguish real gains from seed noise. Third, the high-frequency extraction mechanism is asserted, not demonstrated. The occlusion experiment is suggestive, but there is no spectral analysis or anything that directly shows intra-links are extracting high-frequency fault features rather than just acting as a regularizer. Fourth, the ablation in Table IX shows the spatial-temporal block contributes almost nothing (99.14% vs 99.07%), which undercuts one of the two claimed novelties. I do not think the paper is a waste of time. The core idea, that recycling neurons within a layer can substitute for depth, is plausible and the empirical pattern, while fragile, is consistent. But as written, the missing split granularity and missing error bars make the accuracy ranking impossible to trust. The authors need to state the split unit, run multiple seeds, and ideally add a frequency-domain analysis to back the high-frequency story. Who is this for? Researchers working on resource-constrained motor fault diagnosis, and anyone interested in network architecture tricks that avoid adding parameters. It deserves a serious referee round, but with the expectation of major revision. I would not cite it in its current form, and I would bring it to a reading group mainly to discuss the leakage issue as a cautionary example. Recommendation: send to peer review, but the reviewers should insist on split-level clarification and error bars before any acceptance.","headline":"A plausible, incremental fault-diagnosis architecture whose central accuracy claims rest on an unstated train/test split choice and no error bars; worth a careful referee but not ready as reported.","tokens_in":733,"tokens_out":839,"would_cite":false,"duration_ms":27473,"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 shallow residual network with intra-layer shortcuts extracts high-frequency fault features from low-resolution noisy signals without adding parameters, and edges out ResNet18 on synchronous-motor data.","keywords":["Bi-residual neural network","Synchronous motor electrical fault diagnosis","Intra-linked layer","Residual learning","High-frequency feature extraction","Low-resolution noisy data","Spatial-temporal convolution"],"falsifier":"Hold out entire one-second fault events rather than individual time samples when constructing the 0.8:0.1:0.1 split, and repeat each model run across at least five random seeds. If Bi-ResNet's 94.15% versus 93.87% edge over ResNet18 at SNR -5 collapses within seed-to-seed variation, the claimed parameter-free high-frequency advantage is not established; inspecting whether any single event contributes both training and test samples would confirm whether temporal leakage is present.","tokens_in":1704,"feed_emoji":"⚡","tokens_out":1940,"duration_ms":94152,"temperature":0.7,"pith_summary":"The paper tries to establish that a deliberately shallow residual network, Bi-ResNet, can pull high-frequency fault features out of low-resolution, noisy motor current and voltage signals as effectively as much deeper networks, by adding intra-layer shortcuts (called intra-links) and a multi-scale spatial-temporal convolution block. If true, this gives resource-constrained industrial diagnostics an end-to-end classifier whose accuracy holds when sampling is coarse and noise is strong, without needing wavelet-style preprocessing. The empirical claim is comparative: Bi-ResNet reports the best test accuracy across noise levels and sample rates, for example 94.15% at SNR -5 with 10 ms sampling versus 93.87% for ResNet18, while using roughly one tenth of ResNet18's parameters. The paper also argues the intra-link mechanism is generic, improving accuracy when grafted onto other CNN and LSTM baselines, and that their number trades off against input complexity.","feed_headline":"Small net with intra-layer links beats ResNet18 on noisy motor faults","feed_subtitle":"On low-resolution noisy motor data, a 1M-parameter model edges out 11M-parameter ResNet18 by reusing same-layer neurons.","key_machinery":"The load-bearing mechanism is the intra-linked layer: within one network layer, one neuron's pre-activation receives the previous layer's output plus the activated output of a neighboring neuron in the same layer, so the layer reuses its own information through a same-layer shortcut. With $m$ neurons in a layer and a hyperparameter $n$ controlling how many such connections exist, no new weights are introduced, only data paths, which the paper calls the network's height and a built-in sub-signal extractor. The second piece is the embedded spatial-temporal convolution block, four parallel 1D convolutions with kernel sizes 3, 5, 8 and 16 followed by a global 1D convolution, supplying multi-receptive-field temporal features for the intra-links. The name Bi-ResNet refers to this double residual structure: outer residual blocks between layers plus inner residual connections within a layer.","core_discovery":"The central discovery is that residual shortcuts need not run only between layers; placing shortcuts between neurons in the same layer, termed intra-links, turns a shallow network into a built-in high-frequency extractor at zero additional parameter cost. Bi-ResNet combines these intra-links with an embedded spatial-temporal convolution block, made of parallel 1D convolution kernels of sizes 3, 5, 8 and 16 followed by a global 1D convolution, inside residual blocks. On the authors' synchronous-motor fault dataset, the resulting network, with about 1.05 million parameters, tracks or beats ResNet18, which has about 11.2 million parameters, and the gap grows when data are downsampled and white noise is added. Occlusion experiments show Bi-ResNet concentrating its attention on the fault-onset region where ResNet18 spreads attention over normal signal, supporting the claim that intra-links locate high-frequency fault components rather than merely improving classification.","pith_inferences":["Editorial inference: if intra-links act as a frequency prior, Bi-ResNet's margin over ResNet18 should be largest for fault classes with abrupt, localized transients such as opened phases and short circuits, and smallest for slow-drift faults; this can be tested class by class on the same dataset.","Editorial inference: the reported single-run accuracies carry no error bars, so the decisive comparison is whether the SNR -5 gap of 0.28 percentage points survives multiple random seeds and an experiment-level split; such a test would support the parameter-free high-frequency claim far more strongly than the current tables.","Editorial inference: the same intra-link construction should transfer to other one-dimensional industrial signals such as vibration, acoustic emission, or partial-discharge waveforms, where high-frequency burst features matter, but the paper only demonstrates motor current and voltage."],"forward_implications":["On 10 ms sampled noisy data, Bi-ResNet holds 99.14% test accuracy while the best non-residual CNN sits at 96.86%, so coarse sampling hurts models without residual shortcuts more than it hurts Bi-ResNet.","Across SNR levels from -5 to 5 dB, Bi-ResNet ranks first in every condition; at SNR -5 it reaches 94.15% versus 93.87% for ResNet18 and about 92.34% for the best non-residual CNN.","Grafting intra-links onto CNN-1, CNN-2 and CNN-3 raises their accuracy by 0.46 to 1.97 percentage points at SNR -1 with 10 ms sampling, so the mechanism is portable across architectures.","Ablation shows accuracy improves as the intra-link count rises from $n=0$ to $n=2$ and then declines by $n=4$, indicating a trade-off between the number of intra-links and input signal complexity.","Because intra-links add no weights, the paper attributes the gains to architecture rather than model capacity, which is the core reason shallow networks can approach deep-network performance."],"supporting_citations":[{"why":"Supplies the residual block and the ResNet18 architecture that Bi-ResNet extends and is compared against.","marker":"[13]"},{"why":"Inspires the intra-link idea by showing how same-layer shortcuts can recombine frequency information; the paper adapts this into intra-linked layers.","marker":"[24]"},{"why":"Provides the synchronous-motor fault dataset and the CNN-4 and LSTM reference classifiers used as baselines.","marker":"[30]"},{"why":"Supports the claim that residual learning gravitates toward high-frequency components of images, motivating the high-frequency extractor design.","marker":"[19]"},{"why":"Defines the CNN-1 baseline, a 1D end-to-end CNN for motor drive fault diagnosis.","marker":"[7]"},{"why":"Defines the CNN-2 baseline that combines 1D-CNN with LSTM.","marker":"[23]"},{"why":"Defines the CNN-3 baseline using 2D-CNN with GRU.","marker":"[29]"},{"why":"Represents the wavelet-preprocessing route that Bi-ResNet aims to replace with end-to-end high-frequency extraction.","marker":"[4]"}],"fun_headline_variants":["Zero-param intra-links give small nets deep-net accuracy on noisy faults","Same-layer shortcuts turn shallow nets into high-frequency extractors for free","Intra-links: no extra params, but shallow Bi-ResNet beats deep ResNet18 on noise","Bi-ResNet's intra-links boost high-frequency fault features without extra params","Zero-cost intra-links let tiny Bi-ResNet rival 11M-param ResNet18 on noisy data"],"cache_read_input_tokens":15872,"weakest_assumption_plain":"The load-bearing premise is that the 0.8:0.1:0.1 data split separates independent motor-fault events; if instead it cuts each one-second event's 10,000 time samples between training and test, temporally correlated samples leak across the split and inflate every reported accuracy, shrinking the true margin that Bi-ResNet's architecture actually earns.","fun_headline_variants_meta":{"raw":{"variants":["Zero-param intra-links give small nets deep-net accuracy on noisy faults","Same-layer shortcuts turn shallow nets into high-frequency extractors for free","Intra-links: no extra params, but shallow Bi-ResNet beats deep ResNet18 on noise","Bi-ResNet's intra-links boost high-frequency fault features without extra params","Zero-cost intra-links let tiny Bi-ResNet rival 11M-param ResNet18 on noisy data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000605,"raw_usage":{"total_tokens":2808,"prompt_tokens":917,"completion_tokens":1891,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":533,"completion_tokens_details":{"reasoning_tokens":1779}},"tokens_in":533,"tokens_out":1891,"duration_ms":12181,"temperature":1.0,"reasoning_tokens":1779,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:53:03.497956+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Hold out entire one-second fault events rather than individual time samples when constructing the 0.8:0.1:0.1 split, and repeat each model run across at least five random seeds. If Bi-ResNet's 94.15% versus 93.87% edge over ResNet18 at SNR -5 collapses within seed-to-seed variation, the claimed parameter-free high-frequency advantage is not established; inspecting whether any single event contributes both training and test samples would confirm whether temporal leakage is present.","supporting_citations":[{"cited_title":"Amplitude-phase recombination: Rethinking robustness of convolutional neural networks in frequency domain,","cited_arxiv_id":null,"evidence_quote":"Inspires the intra-link idea by showing how same-layer shortcuts can recombine frequency information; the paper adapts this into intra-linked layers."},{"cited_title":"A public data-set for synchronous motor electrical faults diagnosis with cnn and lstm reference classifiers,","cited_arxiv_id":null,"evidence_quote":"Provides the synchronous-motor fault dataset and the CNN-4 and LSTM reference classifiers used as baselines."},{"cited_title":"High Frequency Residual Learning for Multi-Scale Image Classification","cited_arxiv_id":"1905.02649","evidence_quote":"Supports the claim that residual learning gravitates toward high-frequency components of images, motivating the high-frequency extractor design."},{"cited_title":"Dnocnet: A novel end-to-end network for induction motor drive systems fault diagnosis under speed fluctuation condition,","cited_arxiv_id":null,"evidence_quote":"Defines the CNN-1 baseline, a 1D end-to-end CNN for motor drive fault diagnosis."},{"cited_title":"An efficient approach based on a novel 1d-lbp for the detection of bearing failures with a hybrid deep learning method,","cited_arxiv_id":null,"evidence_quote":"Defines the CNN-2 baseline that combines 1D-CNN with LSTM."},{"cited_title":"Fault diagnosis of asynchronous motors based on 2dcnn-gru network optimization,","cited_arxiv_id":null,"evidence_quote":"Defines the CNN-3 baseline using 2D-CNN with GRU."},{"cited_title":"Diagnostic of combined mechanical and electrical faults in asd-powered induction motor using modwt and a lightweight 1-d cnn,","cited_arxiv_id":null,"evidence_quote":"Represents the wavelet-preprocessing route that Bi-ResNet aims to replace with end-to-end high-frequency extraction."}],"review_version":1}