{"id":"819f05c6-d20f-477b-827e-261791e3376f","arxiv_id":"2412.05901","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A three-layer Self-ONN classifies induction-motor faults from thermal images with accuracy comparable to much larger CNN models and far fewer parameters.","lead":"This paper tests a compact neural network called a Self-ONN for spotting faults in induction motors from thermal images. On a dataset with healthy, misaligned, and broken-rotor motors, the three-layer model matches larger CNNs in accuracy while using far fewer parameters.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'surpasses EfficientNet B1' claim is unsupported: unshuffled contiguous folds create temporal leakage from 4-fps recordings, and the comparison changes preprocessing and training protocol; the 0.010 accuracy gap needs a controlled rerun.","rationale":"The paper's central empirical claim is conditional on the evaluation being sound, and the weakest point is the unshuffled contiguous split combined with protocol mismatch. The reader's weakest_assumption captures the protocol mismatch; I partially agree but would promote the temporal-correlation leakage to the primary concern because it threatens absolute accuracy, not just the comparison. Since no code/data or session IDs are provided, the only way to settle is a controlled rerun. If the rerun reproduces the gap, the concern is resolved; if not, the paper needs major revision. This keeps the reader's CONDITIONAL verdict.","tokens_in":10637,"tokens_out":7728,"duration_ms":81474,"concrete_test":"Re-run the Q=2 Self-ONN and EfficientNet B1 with a grouped 5-fold split: assign all frames from each continuous 30-second recording session to the same fold, use the [5] global min-max normalization, AdamW (LR=0.00003) and the reference early-stopping schedule, and report per-fold accuracies with a paired significance test. If Q=2 does not stay above B1 by more than the fold noise, the claimed superiority is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline claim that Q=2 Self-ONN surpasses EfficientNet B1 rests on accuracy numbers that are not a valid estimate of generalization. Section III.A says the 5 folds are created 'without shuffling the data,' preserving class proportions by contiguous blocks. Thermal frames are recorded at 4 fps in ~30-second sessions, so consecutive frames are near-duplicates. Putting contiguous blocks into different folds places near-identical frames from the same session on both sides of train/test boundaries, inflating the 0.942 accuracy. The comparison to [5] is additionally uncontrolled: the reference used random folds, global min-max scaling, AdamW at 3e-5 and early stopping patience 10; here per-image min-max normalization (Eq. 6), Adam at 1e-3, patience 5 and LR decay are used. The reported 0.010 accuracy advantage is smaller than the per-fold std (0.070/0.086), and no paired significance test is given. Q is also selected after inspecting test-fold results. The within-paper Q=2 vs Q=1 comparison is suggestive, but without a session-grouped split and matched protocol it cannot support superiority over the external CNN baselines.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a 2D Self-Organized Operational Neural Network (Self-ONN) with three operational layers for classifying thermal images of squirrel-cage induction motors into healthy, misalignment, and broken-rotor classes. The authors compare their model against CNN baselines from the reference study [5], report that a Self-ONN with Taylor order Q=2 achieves 0.942 ± 0.070 accuracy versus 0.932 ± 0.086 for EfficientNet B1, and highlight reductions in trainable parameters and inference time. The paper also includes a within-paper comparison of Q=1 (a 2D CNN) and Q=2..5 Self-ONNs under a shared protocol, where Q=2 outperforms Q=1 (0.942 versus 0.887).","tokens_in":10803,"tokens_out":2729,"duration_ms":27772,"significance":"If the central comparison were valid, the result would be practically significant: it would show that a shallow, low-parameter operational network can match or exceed much larger CNNs on a thermal-imaging fault-diagnosis task, with clear deployment advantages for edge monitoring. The paper deserves credit for a self-contained within-paper comparison of Q=1 and Q=2..5 models, for reporting per-fold standard deviations, and for quantifying parameter counts and inference durations. However, the headline claim of surpassing EfficientNet B1 depends on cross-protocol comparisons to numbers taken from [5], and the data-splitting procedure raises a leakage concern; these issues make the magnitude and validity of the reported advantage uncertain.","major_comments":[{"comment":"The 5-fold split is described as preserving class proportions 'without shuffling the data.' Because the thermal camera records at 4 fps over roughly 30-second sessions, adjacent frames are near-duplicates. Placing contiguous blocks from the same recording session into different folds means that near-identical frames appear on both sides of the train/test boundary, typically inflating cross-validation accuracy. This directly affects the reported 0.942 ± 0.070 figure and the comparison with [5]. The authors should use a session-grouped or temporally separated split and report the resulting accuracy.","section":"Section III.A"},{"comment":"The comparison to the CNN baselines from [5] is uncontrolled. The reference study used random folds, global min-max scaling with dataset-wide extrema, the AdamW optimizer with learning rate 3e-5, and early-stopping patience 10, whereas the present study uses contiguous unshuffled folds, per-image min-max normalization (Equation 6), Adam with learning rate 1e-3, and early-stopping patience 5 with a learning-rate reduction schedule. Any of these differences could contribute to the 0.010 accuracy gap (0.942 vs 0.932), which is smaller than the reported per-fold standard deviations. The claim that the Self-ONN 'surpassed' EfficientNet B1 therefore needs either a matched re-implementation of the baselines under the same protocol or a paired significance test; as written, the claim is unsupported.","section":"Section IV, Table III"},{"comment":"The Taylor order Q is selected after inspecting the test-fold results, with Q=2 reported as the best among Q=1..5. Because Q controls model capacity and multiple Q values are evaluated on the same test folds, the reported 0.942 accuracy is a selected maximum and is likely to be optimistically biased. Q should be selected on validation folds, or the paper should explicitly present all Q results as exploratory and avoid claiming that the best Q value represents expected performance.","section":"Section IV, Table III"}],"minor_comments":[{"comment":"Equation (1) is described as giving 'the input feature map' but x_l_k(m,n) is the pre-activation output of the neuron; the terminology should be corrected to avoid confusion between input and output feature maps.","section":"Section II"},{"comment":"The domain/codomain notation for the nodal and pool operators is typeset in a way that is hard to parse; please restate the notation with clear dimensions, for example by writing the nodal operator as acting on the K×K patch and the pool operator as reducing that patch to a scalar.","section":"Section II, Equation (2)"},{"comment":"There is a typo: 'misaligment' should be 'misalignment'.","section":"Conclusion"},{"comment":"The header 'Standart' should be 'Standard'.","section":"Table II"},{"comment":"The phrase 'To further the improve performance of CNNs' is ungrammatical; it should read 'To further improve the performance of CNNs'.","section":"Introduction"},{"comment":"The inference-time comparison reports average duration over 100 runs on a single GPU but does not specify batch size, framework, or whether the same software and input pipeline were used for all models; these details should be added for the comparison to be reproducible.","section":"Section IV"}],"recommendation":"major_revision","confidential_remarks":"The core methodological gap is that the headline superiority claim rests on an uncontrolled comparison to numbers imported from [5], combined with a data split that risks frame-level leakage. The within-paper Q=1 versus Q=2 comparison is more defensible and could be the basis for a revised claim, but the external comparison should either be re-run under a matched protocol or softened to 'comparable accuracy with fewer parameters.'"},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a plausible but under-validated application of 2D Self-ONNs to thermal-image motor fault diagnosis. The within-paper Q=1 vs Q=2 comparison is real evidence, but the headline claim that it beats EfficientNet B1 rests on an uncontrolled comparison and a fold-splitting choice that likely leaks temporal duplicates.\n\nWhat's actually new: no one has applied 2D Self-ONNs to thermal images of induction motors before. The paper gives a clear description of the generative neuron formulation, a clean three-layer architecture, and reports parameter counts and inference times. The self-contained comparison between Q=1 (plain CNN) and Q=2 (Self-ONN) shows a consistent improvement across all metrics, and the confusion matrices show fewer healthy-as-broken-rotor errors. That part is credible and worth building on.\n\nWhere it gets soft: the comparison to the [5] baselines is not apples-to-apples. Different normalization (per-image vs global), different optimizer (Adam vs AdamW), learning rate, early stopping, and validation split protocol. The accuracy gap is 0.010, smaller than the reported std of 0.070/0.086, and no significance test is given. More importantly, the folds are created 'without shuffling the data.' Given 4-fps capture in ~30-second sessions, consecutive frames are near-duplicates. Contiguous blocks across train/test folds means the model is likely seeing near-identical frames on both sides, which inflates the 0.942 number. The Q selection also appears to happen after seeing all Q results; reporting all Qs is fine, but claiming superiority based on the best one needs a held-out selection criterion.\n\nThe inference time and parameter counts are nice, but the precision of those numbers doesn't rescue the accuracy comparison.\n\nBottom line: the methodological core (Self-ONN as a shallow replacement for deep CNNs) is plausible, but the evidence as presented does not support the 'surpasses EfficientNet B1' claim. The paper needs a session-grouped split, a matched training protocol, and significance testing. That is a real revision, not a desk reject. I'd send it to review with a request for those fixes, and I'd point the authors to the leakage concern explicitly.","headline":"Plausible efficiency result for 2D Self-ONNs on thermal motor images, but the headline accuracy claim over EfficientNet B1 is not supported by the current experimental design.","tokens_in":11446,"tokens_out":2826,"would_cite":false,"duration_ms":25670,"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":"Shallow 3-layer Self-ONN beats EfficientNet on thermal motor fault diagnosis","keywords":["thermal imaging","fault diagnosis","induction machines","Self-Organized Operational Neural Networks","convolutional neural networks","broken rotor","misalignment","edge computing"],"falsifier":"Run the same 5-fold protocol with a single preprocessing pipeline and optimizer for both a plain 2D CNN (Q=1) and a Self-ONN (Q=2) matched in parameter count, and check whether the accuracy gap observed here (0.942 vs 0.887) persists; if it shrinks to less than one standard deviation, the attributed non-linearity benefit is not the cause.","tokens_in":10354,"feed_emoji":"🔥","tokens_out":5400,"duration_ms":49461,"temperature":0.7,"pith_summary":"The paper sets out to show that a very shallow 2D Self-Organized Operational Neural Network can diagnose induction-motor faults from thermal images at least as accurately as much deeper CNN models. Its best configuration, a three-operational-layer Self-ONN with Taylor-polynomial order Q=2, reaches 0.942 ± 0.070 accuracy on the healthy/misalignment/broken-rotor classification task, marginally above the best CNN baseline, EfficientNet B1 at 0.932 ± 0.086. That accuracy comes with only 294,083 trainable parameters and a 0.61 ms average inference time, versus 6,517,027 parameters and 14.97 ms for EfficientNet B1. If the comparison holds, the practical payoff is a non-contact condition-monitoring system that runs in near real time on edge hardware without sacrificing diagnostic accuracy.","feed_headline":"Shallow 3-layer net beats EfficientNet on thermal motor faults","feed_subtitle":"A 294k-parameter Self-ONN hits 0.942 accuracy in 0.61 ms, topping a 6.5M-parameter CNN.","key_machinery":"The load-bearing mechanism is the self-organized operational (generative) neuron, which computes pre-activations as $f_{\\psi_k^l}(w_k^{(Q)}(r,t), y^{l-1}(m+r,n+t)) = \\sum_{q=1}^Q w_k^{(Q)}(r,t,q)\\,(y^{l-1})^{q}$, realized as Q ordinary 2D convolutions applied to powers of the layer input and summed. The order Q is a hyperparameter controlling the polynomial degree; Q=1 reduces the layer to a standard convolution, while Q>1 injects non-linearity inside the neuron, and all weights are learned by standard backpropagation. This mechanism is what lets the three-layer network express richer mappings than a depth-matched CNN.","core_discovery":"The central claim is that the generative neuron, whose nodal operator is a Taylor-polynomial approximation of degree Q over the input patch, can substitute for the depth of conventional CNNs in this diagnostic task. With Q=2, eight filters per layer, and kernel sizes 5×5, 3×3, and 2×2, the network reaches the best mean accuracy, F1, and recall among all compared models, and its confusion matrix shows zero errors between misalignment and broken rotor classes. The same network forced to Q=1, which is mathematically a plain 2D CNN, drops to 0.887 ± 0.106 accuracy, so the paper attributes the gain to the non-linear operational neurons rather than to the particular filter sizes or pooling scheme.","pith_inferences":["The accuracy gap over EfficientNet B1 may be partly an artifact of protocol differences: the paper uses stratified folds created without shuffling, per-image min-max normalization, Adam at 0.001, and a different early-stopping schedule than the baseline's AdamW at 0.00003; a matched-protocol rerun is needed to know how much of the 0.010 accuracy difference is due to the Self-ONN itself.","The Q parameter looks like a practical tuning dial: rather than scaling network depth or width, an engineer could adjust the Taylor order to trade accuracy against parameter count, and this paper's Q=4 and Q=5 results show the trade-off is not monotonic.","The same three-layer Self-ONN could be tested on other rotating-machine thermal datasets (e.g., bearing faults under variable load) to see whether the advantage persists beyond this one motor and camera setup."],"forward_implications":["A Self-ONN with Q=2 offers a drop-in replacement for EfficientNet B1 on this dataset with roughly 22 times fewer parameters and 24 times faster inference, making edge deployment practical.","Every Self-ONN variant (Q=2 to Q=5) beats the Q=1 CNN trunk, so on this task adding neuron-level polynomial non-linearity helps more than adding layers.","The near-perfect separation of misalignment and broken rotor (zero cross-class errors in the Q=2 confusion matrix) suggests thermal signatures of these two faults are highly distinguishable once the network has enough non-linear capacity.","Because the model is shallow and fast, it can be stacked into multi-device monitoring systems where one thermal camera per motor is not affordable."],"supporting_citations":[{"why":"Supplies the thermal image dataset, the experimental setup, and all CNN baseline results that the proposed Self-ONN is compared against.","marker":"[5]"},{"why":"Introduces Self-ONNs with generative neurons, the architecture family and the Taylor-polynomial nodal operator used here.","marker":"[15]"},{"why":"Defines operational neural networks with nodal and pool operators, the precursor framework that Self-ONNs simplify and extend.","marker":"[19]"},{"why":"Demonstrates 1D Self-ONNs for broken rotor bar detection, the closest prior application motivating the use of Self-ONNs in induction-motor diagnosis.","marker":"[16]"}],"fun_headline_variants":["3-layer Self-ONN rivals deep CNNs on thermal motor faults","Thermal imaging + Self-ONN: shallow net, top fault accuracy","Non-linear neurons let 3-layer Self-ONN match deep CNN accuracy","Self-ONN's generative neurons pack CNN depth into shallow net","Self-ONN: 3 layers, non-linear neurons, deep CNN-beating thermal fault diagnosis"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim that the Self-ONN surpasses the CNN baselines rests on the assumption that the different data splitting, normalization, optimizer, and training schedules used in the two studies do not materially affect the comparison; if they do, the 0.010 accuracy gap could be protocol, not architecture.","fun_headline_variants_meta":{"raw":{"variants":["3-layer Self-ONN rivals deep CNNs on thermal motor faults","Thermal imaging + Self-ONN: shallow net, top fault accuracy","Non-linear neurons let 3-layer Self-ONN match deep CNN accuracy","Self-ONN's generative neurons pack CNN depth into shallow net","Self-ONN: 3 layers, non-linear neurons, deep CNN-beating thermal fault diagnosis"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000862,"raw_usage":{"total_tokens":3738,"prompt_tokens":943,"completion_tokens":2795,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":559,"completion_tokens_details":{"reasoning_tokens":2692}},"tokens_in":559,"tokens_out":2795,"duration_ms":18987,"temperature":1.0,"reasoning_tokens":2692,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T20:12:51.175259+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same 5-fold protocol with a single preprocessing pipeline and optimizer for both a plain 2D CNN (Q=1) and a Self-ONN (Q=2) matched in parameter count, and check whether the accuracy gap observed here (0.942 vs 0.887) persists; if it shrinks to less than one standard deviation, the attributed non-linearity benefit is not the cause.","supporting_citations":[{"cited_title":"Unraveling induction motor state through thermal imaging and edge processing: A step towards explainable fault diagnosis,","cited_arxiv_id":null,"evidence_quote":"Supplies the thermal image dataset, the experimental setup, and all CNN baseline results that the proposed Self-ONN is compared against."},{"cited_title":"Self-organized operational neural networks with gener- ative neurons,","cited_arxiv_id":null,"evidence_quote":"Introduces Self-ONNs with generative neurons, the architecture family and the Taylor-polynomial nodal operator used here."},{"cited_title":"Operational neural networks,","cited_arxiv_id":null,"evidence_quote":"Defines operational neural networks with nodal and pool operators, the precursor framework that Self-ONNs simplify and extend."},{"cited_title":"Improved detection of broken rotor bars by 1-d self-onns,","cited_arxiv_id":null,"evidence_quote":"Demonstrates 1D Self-ONNs for broken rotor bar detection, the closest prior application motivating the use of Self-ONNs in induction-motor diagnosis."}],"review_version":1}