{"id":"4d35351e-f926-447a-a196-978712b139e1","arxiv_id":"2504.18105","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"An MLP estimated induction motor winding and bearing temperatures from drive signals with MSE 0.31, beating a linear model (1.94) and a CNN (0.54) on a single representative test profile.","lead":"This paper compares three machine learning models (linear regression, a multilayer perceptron, and a convolutional network) for estimating the internal temperature of an induction motor from signals already collected by the motor drive. The multilayer perceptron gave the lowest errors in the tested operating profile, suggesting that temperature monitoring and protection might be possible without adding extra sensors.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported evidence for 'high precision' rests on a single operating profile; the 18-fold leave-one-out results are never aggregated, so the representative-profile MSE may be a best-case outlier.","rationale":"The reader's weakest_assumption identifies the same load-bearing concern: the quantitative evidence for the central claim comes from a single operating profile, and the paper does not provide results or variance measures for the other 17 leave-one-out test folds. I agree with that assessment. The central claim in the introduction and conclusion is that the ML models estimate internal temperatures with high precision and perform satisfactorily under fast transient conditions. Section IV-C explicitly describes an 18-fold leave-one-out evaluation, so the experimental design to support that claim exists, but Section V reports detailed results for only one profile. Without per-fold aggregate metrics, the reader cannot distinguish between a typical result and a cherry-picked best case. The concern is not internal inconsistency or a modeling error; it is that the reported evidence underdetermines the strength of the claim. The paper also omits Table III and figure contents in the extracted text, but the missing aggregate results are the primary gap. A concrete check—reporting the distribution of MSE across all 18 folds—would settle whether the representative profile is indeed representative. This check does not require new experiments, only a re-analysis of existing results. Since the verdict was already conditional on this concern, my stress-test pass does not move the verdict; it confirms the conditionality is appropriate.","tokens_in":8482,"tokens_out":2383,"duration_ms":27819,"concrete_test":"Report the per-fold test MSE for all 18 leave-one-out profiles for the linear model, MLP, and CNN in a supplementary table, including median and interquartile range. If the representative profile's MLP MSE of 0.31 falls below the 25th percentile of the 18-fold distribution, or if the median MLP MSE is substantially larger than 0.31, the claim of high-precision estimation across transient conditions should be revised or explicitly restricted to the reported profile.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's quantitative claim of high-precision temperature estimation is supported only by results on one operating profile. Section IV-C states that all eighteen generated profiles are used as test data at least once in a leave-one-out scheme, but Section V reports detailed results for 'one representative operating profile' and provides no aggregate statistics across the other seventeen folds: no mean or median MSE, no variance, and no worst-case profile. The quoted values (linear MSE 1.94, MLP 0.31, CNN 0.54) therefore establish performance on a single, possibly favorable example rather than across the tested transient conditions. The load-bearing premise is that this profile is representative; if it was selected because it gives favorable errors, the headline claim overstates typical performance. This is a reporting/cherry-picking risk rather than a modeling flaw: the ML pipeline is plausible and consistent with prior work, but the evidence as presented does not demonstrate generalization across the full set of profiles.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops and compares three machine-learning models (elastic-net-regularized linear regression, a two-layer MLP, and a temporal CNN) for estimating stator-winding and bearing temperatures of an induction motor from drive-available signals (speed, current, shell temperature). Data were collected on a lab test bench under eighteen operating profiles with slow, medium, and fast dynamics; a leave-one-profile-out scheme is used for evaluation. A hyperparameter search is performed for each model. In Section V, quantitative results are reported for one representative profile, giving MSE values of 1.94 (linear), 0.31 (MLP), and 0.54 (CNN). The paper concludes that neural networks estimate the motor's internal temperatures with high precision, including under transient conditions.","tokens_in":8693,"tokens_out":3879,"duration_ms":38011,"significance":"If the reported performance is representative, the results are practically significant: a data-driven model using only signals already available in a drive could replace dedicated internal temperature sensors, with errors mostly within a few degrees Celsius. The leave-one-profile-out protocol is a sound approach for assessing generalization to unseen operating conditions, and the comparison of a linear baseline against nonlinear networks is appropriate. The lower MSE of the MLP relative to the linear model is consistent with the expected nonlinearity of motor thermal dynamics. However, the evidence presented in the manuscript is incomplete in two load-bearing respects: only one test profile is analyzed in detail, with no aggregate statistics across the eighteen folds, and key tables describing the experimental setup and final model configurations are referenced but not present in the manuscript text.","major_comments":[{"comment":"The central claim of 'high precision' (Introduction and Conclusion) is supported only by results on 'one representative operating profile' in Section V. Section IV-C states that all eighteen generated profiles were used as test data at least once in a leave-one-out scheme, but Section V provides no aggregate statistics for the other seventeen folds: no mean or median MSE, no variance, no worst-case error, and no per-fold table. If the selected profile is not typical, the quoted values (linear 1.94, MLP 0.31, CNN 0.54) overstate typical performance. Please report the full distribution of per-profile metrics, or explicitly justify the representativeness of the selected profile with quantitative evidence.","section":"Section V, Section IV-C"},{"comment":"The manuscript text refers to Table I (motor specifications), Table II (model inputs and outputs), and Table III ('the finally selected model configurations'), but none of these tables appears in the manuscript. Without the final hyperparameters (regularization penalty and mixing ratio for the linear model, layer sizes and dropout for the MLP, layer count, filter sizes, dilation rate, and input sequence length for the CNN), the exponentially weighted moving average span values, and the exact input/output variable list, the experiments cannot be reproduced or independently checked. Please include the full content of all three tables.","section":"Table III and Section IV-B"},{"comment":"The hyperparameter search is described only in broad terms: 'a lot of parameters are fine-tuned' and 'another small set of profiles has been taken away as the validation set.' The number of validation profiles, the split strategy, and the search protocol (grid or random search, number of trials, early-stopping criterion) are not specified. Because the final models are selected by this search, the absence of these details weakens the claim that the reported configurations are optimal and makes the comparison between models less interpretable.","section":"Section IV-C"}],"minor_comments":[{"comment":"There is a typo immediately before Eq. (1): 'I ts mathematical formulation' should read 'Its mathematical formulation.'","section":"Section II-A"},{"comment":"The term 'multilayer perception' should be 'multilayer perceptron,' and 'Convolutional neuron network' should be 'Convolutional neural network.'","section":"Throughout"},{"comment":"The sentence 'eight span values ranging' is incomplete; the actual span values are not listed.","section":"Section IV-B"},{"comment":"Figures 7–9 show error and input signals, but axis labels and units are not visible in the manuscript text; please ensure all axes are labeled.","section":"Section V"},{"comment":"The statement that the MLP 'shows an obvious difference between slow and fast dynamics' is not quantified; a per-segment error comparison would make the claim about transient performance precise.","section":"Section V-B"},{"comment":"References [1] and [13] are the same work; please consolidate to avoid duplication.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The manuscript as provided lacks Tables I–III, which may be a formatting artifact of the submission, but as presented the experimental reproducibility is compromised. The central risk to the paper's conclusion is the single-profile reporting in Section V; if the authors cannot provide aggregate leave-one-out results, the claim of 'high precision' should be weakened or qualified. The overall approach is sound and the leave-one-out protocol is appropriate, so I do not see a fundamental flaw that would necessitate rejection, provided the missing evidence is supplied."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper is a straightforward engineering comparison of linear, MLP, and CNN models for estimating induction-motor winding and bearing temperatures from drive-available signals. Good points first: the leave-one-profile-out protocol is the right way to test generalization, the feature expansion with exponentially weighted moving averages is a sensible and small practical contribution, and the reported numbers on the shown profile are in line with what you'd expect — MLP beats linear, CNN is in between. The experimental data are real lab data, and the authors only use signals a drive would have, which is the right constraint for deployment.\n\nWhere it gets soft: Section V shows results for one 'representative' operating profile and nothing else. The paper says all eighteen profiles were used as test data at least once, but there are no aggregate statistics — no mean or median MSE, no variance, no worst case. So the 'high precision' claim in the introduction and conclusion is only actually demonstrated for one profile. If that profile was picked because it looks good, the quoted MSE values overstate typical performance. The stress-test note is correct. This is a reporting gap rather than a modeling flaw, and it's fixable, but it is the difference between a supported claim and a plausible one. The hyperparameter description also has holes (Table III's contents are largely missing from the text I have), and there is no code or data release, which is a real limitation for an empirical engineering paper.\n\nThe citation pattern is fair: the prior ML temperature-estimation work is properly acknowledged, and this paper does not oversell its novelty. No circularity that I can see: hyperparameters were chosen separately and the evaluation is on held-out profiles.\n\nBottom line: this is a competent, honest subfield paper. It deserves referee time, not a desk reject. But the review should require the authors to report leave-one-out aggregate results across all eighteen profiles and explain how the representative profile was selected. That is a reasonable revision, and with that addition the paper would make a useful data point for anyone building drive-integrated thermal monitoring.","headline":"The paper's ML comparison is competent, but the headline precision claim rests on a single operating profile; the missing aggregate leave-one-out results are a fixable but real reporting gap.","tokens_in":9206,"tokens_out":2690,"would_cite":false,"duration_ms":27436,"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 multilayer perceptron trained on drive signals can estimate induction motor winding and bearing temperatures with mean squared errors below one degree under the tested transients.","keywords":["induction motor","machine learning","temperature estimation","temperature monitoring","multilayer perceptron","convolutional neural network","transient operation","stator winding"],"falsifier":"Run the same leave-one-profile-out evaluation and report the mean and standard deviation of MSE across all eighteen folds; if the average is close to the single-profile values (linear 1.94, MLP 0.31, CNN 0.54), the central claim holds, whereas a large gap or an obviously favorable choice of reported profile would show the claimed precision does not generalize.","tokens_in":8282,"feed_emoji":"🔥","tokens_out":7773,"duration_ms":68182,"temperature":0.7,"pith_summary":"This paper tries to establish that data-driven models, trained on signals already available in a modern electric drive, can estimate internal motor temperatures accurately enough for commercial monitoring. The authors compare linear regression, a multilayer perceptron, and a temporal convolutional network on experimental data from a test bench running eighteen operating profiles of slow, medium, and fast dynamics. On the representative profile they report, the linear baseline has MSE 1.94, the MLP reaches 0.31, and the CNN 0.54. The claim matters because winding and bearing overheating is a major failure cause, and accurate temperature estimates without additional sensors would support condition monitoring and control optimization.","feed_headline":"Neural nets estimate induction motor temperatures within a few degrees","feed_subtitle":"Using only signals modern drives already collect, the best model cut error to about a sixth of a linear baseline.","key_machinery":"The central mechanism is a supervised regression pipeline built on drive-available inputs: motor speed, stator current, and shell temperature, each normalized and expanded with exponentially weighted moving averages over eight span values so that non-sequential models such as linear regression and the MLP still carry memory of past behavior. The MLP is configured with two hidden layers of 90 and 20 neurons, and the temporal CNN uses 1D convolution with small filters and dilation, followed by global pooling and a dense regression head. The load-bearing work is done by the feature-expansion step together with a leave-one-profile-out training scheme that removes each of the eighteen operating profiles from training once, so every profile serves as test data at least once.","core_discovery":"On the paper's own terms, the discovery is that a two-layer multilayer perceptron, fed with normalized speed, current, and shell temperature expanded by exponentially weighted moving averages, estimates stator winding and bearing temperatures under transient operation with high precision. It beats both a linear baseline (MSE 1.94) and a three-layer temporal CNN (MSE 0.54) on the reported operating profile, achieving MSE 0.31. The authors state that neural networks performed satisfactorily even under fast transient conditions and that errors remained mostly within a few degrees Celsius, which they regard as acceptable for commercial applications. The paper frames this as opening the way to deploying similar models in commercial drives using only signals the drive already collects.","pith_inferences":["The paper reports detailed errors for only one representative profile; the leave-one-profile-out design implies that the other seventeen folds also produced results, but their mean and spread are not shown, so the quoted MSE values should be read as one draw rather than a verified average.","The feature expansion uses eight exponentially weighted moving-average spans, which suggests the model's memory length is a key tuning factor; one could test whether the MLP's advantage over the CNN shrinks or grows with the number of spans.","If extended beyond the reported test bench, the approach could be checked for robustness to motor-to-motor variation, cooling condition changes, and aging, none of which the current experimental design varies.","A controlled comparison matching model size and training budget would clarify whether the MLP's edge over the CNN comes from the architecture or from the hyperparameter search, a distinction the paper does not separate."],"forward_implications":["If the reported results hold, a two-layer MLP can reduce temperature-estimation error by roughly a factor of six compared with a linear baseline (MSE 0.31 versus 1.94) on a profile mixing slow and fast transients.","Because the same input set, speed, current, and shell temperature, is already logged by modern drives, the approach could be deployed without installing internal temperature sensors.","Because the CNN performs slightly worse than the MLP despite seeing sequences, the explicit memory features appear to carry most of the temporal information needed for this task.","Accurate internal temperature estimates would let a drive monitor winding and bearing health continuously and act before thermal limits are exceeded, supporting condition monitoring and control optimization."],"supporting_citations":[{"why":"Supplies the failure statistics showing bearing and stator failures exceed 40% and 38%, motivating temperature monitoring.","marker":"[1]"},{"why":"Provides the three-way taxonomy of thermal monitoring methods that frames the paper's approach category.","marker":"[2]"},{"why":"Shows supervised machine learning being applied to induction motor condition monitoring, grounding the problem statement.","marker":"[4]"},{"why":"Demonstrates deep neural networks estimating magnet temperatures in an automotive drive, the main precedent for NN temperature estimation.","marker":"[6]"},{"why":"Extends the precedent to convolutional and recurrent networks for permanent-magnet motors, supporting the CNN choice.","marker":"[7]"},{"why":"Combines ML with lumped parameter thermal networks, showing data-driven adjustment of thermal models is feasible.","marker":"[10]"},{"why":"Provides an MLP-based winding temperature estimator for a BLDC motor, the closest architectural antecedent.","marker":"[13]"}],"fun_headline_variants":["Neural net cuts motor temp error to 0.31 MSE","Simple MLP beats CNN and linear models for motor temperature","Motor temps estimated within a few degrees using existing drive data","Data-driven MLP predicts motor winding and bearing temperatures"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the single operating profile used for the reported results is representative of all eighteen test profiles, since the quoted MSE values would overstate typical performance if that profile was selected for its favorable errors.","fun_headline_variants_meta":{"raw":{"variants":["Neural net cuts motor temp error to 0.31 MSE","Simple MLP beats CNN and linear models for motor temperature","Motor temps estimated within a few degrees using existing drive data","Data-driven MLP predicts motor winding and bearing temperatures"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000608,"raw_usage":{"total_tokens":2779,"prompt_tokens":841,"completion_tokens":1938,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":457,"completion_tokens_details":{"reasoning_tokens":1870}},"tokens_in":457,"tokens_out":1938,"duration_ms":13766,"temperature":1.0,"reasoning_tokens":1870,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:23:56.829958+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same leave-one-profile-out evaluation and report the mean and standard deviation of MSE across all eighteen folds; if the average is close to the single-profile values (linear 1.94, MLP 0.31, CNN 0.54), the central claim holds, whereas a large gap or an obviously favorable choice of reported profile would show the claimed precision does not generalize.","supporting_citations":[{"cited_title":"A Review of Thermal Monitoring Techniques for Radial Permanent Magnet Machines,","cited_arxiv_id":null,"evidence_quote":"Provides the three-way taxonomy of thermal monitoring methods that frames the paper's approach category."},{"cited_title":"Supervised Machine Learning Algorithm Selection for Condition Monitoring of Induction Motors,","cited_arxiv_id":null,"evidence_quote":"Shows supervised machine learning being applied to induction motor condition monitoring, grounding the problem statement."},{"cited_title":"Deep Residual Convolutional and Recurrent Neural Networks for Temperature Estimation in Permanent Magnet Synchronous Motors,","cited_arxiv_id":null,"evidence_quote":"Extends the precedent to convolutional and recurrent networks for permanent-magnet motors, supporting the CNN choice."},{"cited_title":"Machine Learning for Sensorless Temperature Estimation of a BLDC Motor,","cited_arxiv_id":null,"evidence_quote":"Provides an MLP-based winding temperature estimator for a BLDC motor, the closest architectural antecedent."}],"review_version":1}