{"id":"03cf2f97-ed2a-4794-a527-9ac68493ffd2","arxiv_id":"2507.19168","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"An unsupervised framework detects, clusters, and explains circuit breaker faults from vibration and acoustic signals using only healthy data for training.","lead":"This paper proposes a system that learns to spot faults in high-voltage circuit breakers using only healthy examples, then groups different faults and highlights which parts of the vibration signal are responsible. It is aimed at online monitoring without labeled fault data, and it is demonstrated on a lab circuit breaker with spring and damper faults.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"XAI diagnostics are validated only for faithfulness to the pseudo-label classifier, not for physical correspondence to seeded fault causes; the central diagnostics claim therefore rests on an unmeasured link.","rationale":"The detection and segmentation results are internally credible: 98.21% of faulty samples are detected on the presented dataset, K-means reaches ARI 0.9172, and the sensor ablation study is informative. There is no reported false-positive rate or variance across repeated runs, and all data come from one CB, so deployment-level claims are not yet fully established. The stationarity assumption in Section 3.2 is a real robustness concern, but the authors explicitly acknowledge healthy variation in Figure 9d and propose including conditioning variables as inputs, so that issue is partially self-identified. The more load-bearing gap is the diagnostics claim: the only quantitative validation of the XAI output is faithfulness to a classifier trained on clustering pseudo-labels, which is circular for the purpose of diagnosing physical faults. The qualitative comparison to spectrogram differences is not enough, and no unseen-fault-type experiment supports the conclusion in Section 6. I therefore keep the reader's CONDITIONAL verdict, but for a reason different from the one flagged as the weakest assumption: the missing link between attribution maps and physical fault causes, rather than the temporal stability of the healthy distribution, is the load-bearing point. No code release or formal verification is claimed, so the validation rests entirely on the experimental dataset.","tokens_in":23885,"tokens_out":4467,"duration_ms":45310,"concrete_test":"Compute, for every faulty sample in the dataset, the overlap, such as IoU or Pearson correlation, between the binarized Integrated Gradients attribution map and the binarized pixel-wise healthy-versus-faulty spectrogram difference map, reported separately per condition. If the mean overlap is not significantly above a random-positive-pixel baseline, the XAI maps do not track the known fault-induced signal changes and the diagnostics claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of contribution 2 is that the XAI component provides fault diagnostics, namely indications of aged or faulty components, without ground-truth labels. For that claim to hold, the Integrated Gradients attributions computed on the pseudo-label classifier must highlight time-frequency regions that correspond to the physical effects of the seeded faults, such as spring tension and damper viscosity, rather than merely features that separate clusters. The validation in Section 4.2 and Figure 12 only establishes faithfulness: occluding high-attribution pixels changes the classifier output more than random occlusion. Because the classifier in Section 3.3 is trained on the very pseudo-labels produced by the clustering step, faithfulness confirms only that attributions are salient to a model of the clusters; it does not confirm that the highlighted regions diagnose components. The qualitative alignment with pixel-wise healthy-versus-faulty spectrogram differences in Figure 11, rows b3 to e3 and b4 to e4, is suggestive and shown only for the most representative sample per condition, with no quantitative agreement reported. Section 6 further claims diagnostics for fault types never observed, but no leave-one-fault-out experiment is presented. If the attribution maps are driven by global amplitude shifts or cluster-boundary artifacts rather than component-specific signatures, the fault diagnostics claim collapses into a re-description of the cluster separation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an unsupervised condition-monitoring framework for high-voltage circuit breakers. A convolutional autoencoder (CAE) is trained on healthy-only time-frequency spectrograms; faulty samples are detected when the mean absolute reconstruction residual exceeds a fixed threshold tau = mean + 3 sigma. Faulty samples are then segmented by clustering CAE latent features with K-means, OPTICS, or a self-organizing map, yielding pseudo-labels. A shallow classifier is trained on those pseudo-labels and Integrated Gradients is applied to produce time-frequency 'diagnostics matrices' intended to explain cluster assignments and support fault diagnostics. The framework is evaluated on laboratory data from one CB under healthy, spring-tension, and damper-viscosity fault conditions. The reported fault detection rate for faulty samples is 98.21%, K-means clustering achieves an adjusted Rand index of 0.9172, and the XAI faithfulness experiment shows attribution-based occlusion changes classifier outputs more than random occlusion.","tokens_in":24116,"tokens_out":4674,"duration_ms":50770,"significance":"If the results hold, the framework would be a useful step toward monitoring CBs without requiring fault labels during training, which is a real operational constraint. The manuscript has concrete strengths: fault detection and segmentation are evaluated against external ground-truth labels with multiple metrics; three clustering algorithms are compared in both offline and online settings; the sensor ablation study is informative; and the online clustering experiments demonstrate that the pipeline can operate incrementally. The central caveat is that the XAI component is validated only as being faithful to a classifier trained on pseudo-labels, not as identifying physical fault causes. That gap, plus the unquantified stability of the healthy distribution and an inconsistency between two reported clustering scores, currently limits the strength of the contribution.","major_comments":[{"comment":"The fault detection claim relies on a fixed threshold tau = mean + 3 sigma computed on healthy training residuals, with the stated assumption that the healthy condition remains stable over time and that healthy-to-healthy deviations are smaller than healthy-to-faulty deviations. This assumption is load-bearing for the real-world detection claim, but it is not demonstrated with data. The manuscript itself notes that seasonal temperature, gas pressure, and interrupted current levels can affect healthy signals. Please provide a quantitative analysis of healthy-to-healthy residual variability across operating or environmental conditions, and report the false-positive rate on the healthy test samples, so that the detection threshold can be assessed.","section":"Section 3.2, Eq. (4); Section 5.1, Figure 6"},{"comment":"The XAI validation establishes only faithfulness of the Integrated Gradients attributions to the classifier C_theta_c, which is trained on pseudo-labels obtained from the same clustering step. It therefore does not demonstrate that the highlighted time-frequency regions correspond to the physical effects of the seeded faults, such as spring tension or damper viscosity. The qualitative comparison in Figure 11 is shown for one representative sample per condition and is not quantitatively evaluated. Moreover, the conclusion in Section 6 that the approach achieves diagnostics 'even if the fault type has not yet been observed' is not supported by any leave-one-fault-out experiment. Please add a quantitative link between attributions and known physical fault signatures, or a leave-one-fault-out study, and otherwise temper the diagnostics claims to cluster explanation.","section":"Section 3.3, Section 4.2, Eq. (13), Figure 12"},{"comment":"Tables 3 and 5 report different adjusted Rand index values for what appears to be the same configuration: K-means with K = 5 using all four sensors gives ARI = 0.9172 in Table 3 and ARI = 0.9045 in Table 5. No explanation is given for this discrepancy. Please clarify whether the sensor influence study uses a different data split, a different random initialization, or a different evaluation protocol; this is needed for reproducibility.","section":"Section 5.2, Table 3 and Section 5.5, Table 5"},{"comment":"The paper reports that approximately 98.21% of faulty samples are detected, i.e., a false-negative rate of 1.79%, but it does not report how many healthy test samples exceed the threshold. Since the test set includes healthy samples and the detection threshold is derived from healthy training residuals, the false-positive rate is essential for interpreting the detection performance. Please report the confusion between healthy and faulty samples at the chosen threshold.","section":"Section 5.1, Figure 6"}],"minor_comments":[{"comment":"For the SOM clustering result, only the homogeneity score is reported among the external clustering metrics, while Table 3 reports ARI, completeness, and v-measure for K-means and OPTICS. Please report the full set of external metrics for SOM to enable a direct comparison.","section":"Section 5.2, Figure 8"},{"comment":"There is a typo: 'back-box model' should be 'black-box model'.","section":"Section 2.3"},{"comment":"The faithfulness plot would be easier to interpret with error bars or confidence intervals, since the averaging is over a small number of samples (approximately 340 operations in total).","section":"Section 5.4, Figure 12"},{"comment":"The manuscript states that the architecture is selected from a grid search but does not report the search space or the selection criterion. Adding this information would improve reproducibility.","section":"Section 4.3"},{"comment":"The sentence 'The direction of the accelerometer installation does not show a significant difference based on this experimental dataset' is contradicted in part by the reported metrics, where the vertical and axial accelerometers outperform the horizontal one by approximately 0.03-0.04 in ARI. Please rephrase to describe the observed differences accurately.","section":"Section 5.5, Table 5"}],"recommendation":"major_revision","confidential_remarks":"The detection and segmentation parts of the paper are well supported by the reported experiments, but the diagnostics contribution, which is a central claimed novelty, is not yet validated against physical fault causes. The authors should either add a validation that links the attribution maps to the seeded fault mechanisms or clearly reframe the XAI component as explaining cluster assignments rather than diagnosing components. The inconsistency between Tables 3 and 5 should also be resolved before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the unsupervised detection and clustering part is a solid, honest engineering contribution on a real dataset; the XAI-guided diagnostics claim is the soft spot, because the validation only shows faithfulness to the pseudo-label classifier, not physical correspondence to the seeded spring/damper faults.\n\nWhat is actually new: combining a healthy-only CAE, latent-space clustering, and a classifier-plus-Integrated-Gradients scheme for CB vibration/acoustic monitoring, with a new lab dataset. The pieces are mostly from prior work, and the paper cites the relevant one. The detection result (98.21% of faulty samples above a healthy-only threshold) and the segmentation results (K-means ARI 0.9172, SOM homogeneity 0.9564) are supported on this dataset. The online clustering demonstration and the sensor ablation are useful and not mere tack-ons. The paper is also transparent about its limitations, including the single-CB evaluation and the stationarity assumption.\n\nThe load-bearing issue is in the diagnostics claim. Training a classifier on the clustering's pseudo-labels and then applying Integrated Gradients produces explanations of the classifier, not diagnoses of components. The faithfulness test in Figure 12 only shows that high-attribution pixels are important to the classifier; it doesn't show they point to the physical effects of low spring tension or degraded damper. The qualitative agreement with pixel-wise healthy-vs-faulty spectrogram differences in Figure 11 is suggestive, but it is one representative sample per condition and no quantified agreement. The conclusion goes further and says the framework can diagnose faults never observed; no leave-one-fault-out experiment supports that. This needs to be fixed either by softening the claims or by adding validation against the known fault causes.\n\nThe other concerns are milder. The fixed threshold tau = mean + 3 sigma assumes the healthy distribution is stable; the paper acknowledges environmental variation but doesn't demonstrate that healthy-to-faulty deviations always dominate. And with 30 operations per condition and one CB, the numbers are point estimates with no variance or false-positive rate. That is acceptable for a lab study, but worth stating.\n\nBottom line: the detection and segmentation claims hold up. This deserves serious peer review, not a desk reject. I'd recommend the referee ask for a revised diagnostics section — either a leave-one-fault-out check or a clearer statement that the XAI step explains clusters and only suggests possible fault signatures.","headline":"Solid unsupervised detection and clustering on real CB data; the XAI 'diagnostics' claim is the soft spot because it is validated only against the pseudo-label classifier, not against the physical fault causes.","tokens_in":24672,"tokens_out":3239,"would_cite":true,"duration_ms":35433,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper shows that a healthy-data-only autoencoder, clustering, and Integrated-Gradients explanations can detect breaker faults, separate fault types, and indicate likely faulty mechanisms without any fault labels.","keywords":["high-voltage circuit breaker condition monitoring","unsupervised fault detection","fault segmentation","fault diagnostics","convolutional autoencoder","vibration signals","time-frequency spectrograms","explainable artificial intelligence"],"falsifier":"Collect healthy vibration and acoustic recordings from breakers in service across different seasons, gas pressures, and interrupted-current levels and feed them through the same healthy-trained CAE; if the mean residual distribution or its spread shifts so that a substantial fraction of healthy operations exceed $μ+3σ$, the fixed-threshold detection claim fails. A laboratory version would vary chamber gas pressure and temperature across healthy operations and measure how often the healthy-only threshold is crossed.","tokens_in":1742,"feed_emoji":"⚡","tokens_out":14314,"duration_ms":214320,"temperature":0.7,"pith_summary":"High-voltage circuit breakers fail rarely but with costly consequences, and most monitoring today relies on physical parameters such as gas pressure with preset thresholds, which miss many mechanical faults. This paper argues that an unsupervised pipeline can replace supervised fault diagnostics: a convolutional autoencoder trained only on healthy vibration and acoustic recordings detects deviations, clustering in the latent space separates fault types, and Integrated Gradients attributions through a classifier trained on cluster pseudo-labels show experts which time-frequency regions drive each cluster. On an experimental spring-drive breaker with four artificially induced spring/damper faults, the healthy-only detector catches 98.21% of faulty samples and K-means clustering reaches an adjusted Rand index of 0.9172. If this transfers to real substations, utilities could continuously monitor breakers and investigate emerging fault signatures without ever collecting a labeled failure dataset.","feed_headline":"Circuit-breaker faults detected and clustered with no fault labels","feed_subtitle":"A healthy-only autoencoder plus explainable attributions flags spring and damper wear before labeled failure data exists.","key_machinery":"The key machinery is a pseudo-label bridge that turns an unsupervised clustering result into a supervised explanation target. In step one, the convolutional autoencoder compresses each spectrogram into an 80-dimensional latent vector and the reconstruction residual with threshold $τ=μ+3σ$ is the fault detector. In step two, clustering on latent vectors produces pseudo-labels for every test sample, with K-means, OPTICS, and the self-organizing map as interchangeable instantiations. In step three, a one-layer softmax classifier is trained on the frozen latent features with those pseudo-labels as targets, which makes Integrated Gradients applicable; max-pooling the attribution maps yields the low-resolution diagnostics matrix. The pseudo-label bridge is what carries the argument: it converts the paper's unsupervised fault segmentation into a form that supervised XAI can interrogate, and it works regardless of which clustering algorithm produced the labels.","core_discovery":"On its own terms, the paper's central finding is that fault diagnostics for high-voltage circuit breakers can be decomposed into three stages that require no fault labels. A convolutional autoencoder is trained on log-Mel spectrograms of size $128×100×4$ (three accelerometer directions plus a microphone) using only healthy opening operations, and the mean absolute reconstruction error of a test sample is compared with $τ=μ+3σ$ computed from the healthy training residuals; on the experimental data this detects 98.21% of faulty samples. The encoder's 80-dimensional latent vectors are then clustered with K-means (K=5), OPTICS, or a self-organizing map, separating healthy and fault conditions; K-means reaches an adjusted Rand index of 0.9172. To make the clusters explainable, the paper trains a one-layer softmax classifier on the frozen latent features using cluster assignments as pseudo-labels, then applies Integrated Gradients to this classifier; the resulting attributions are max-pooled into a 4×5 diagnostics matrix that localizes cluster-driving signal content in time and frequency. An occlusion-based faithfulness check shows that removing the most attributed spectrogram regions changes the classifier output more than removing random regions, supporting the claim that the explanations track the features that actually determine cluster membership, and the matrices show physically coherent signatures for spring-tension and damper-viscosity faults.","pith_inferences":["Editorial inference: the explanations are only as trustworthy as the clusters that generate the pseudo-labels; when clustering merges two fault subtypes, as low-spring samples with different damper viscosities do here, the attribution maps may highlight what separates clusters rather than what isolates the physical defect.","Editorial inference: the paper's own observation that healthy samples split into subclusters by experimental day points to a direct upgrade: replacing the fixed threshold with a healthy model conditioned on measured temperature, gas pressure, and interrupted current, so the detector adapts to normal environmental drift.","Editorial inference: because the pipeline needs only healthy recordings, a natural next test is fleet-level deployment in which one healthy model built from several same-type breakers serves as the baseline and any breaker whose latent trajectory drifts is flagged for inspection."],"forward_implications":["Fault monitoring can begin with only healthy recordings from an in-service breaker; no artificially induced faults or ground-truth labels are needed to train the detector.","Unknown or previously unseen fault types still get separated into clusters, and XAI explanation maps can indicate whether two clusters share a common culprit, such as low spring tension with different damper states.","The pseudo-label classifier is clustering-agnostic, so practitioners can choose K-means, OPTICS, or a self-organizing map and still obtain attributions for the clusters.","A single axial-direction accelerometer gives clustering performance close to the full four-sensor set, which lowers installation cost for real deployments.","Using all four sensors gives the best separation, with all clustering metrics above 0.9, so sensor fusion remains the reference configuration."],"supporting_citations":[{"why":"Frames fault segmentation as unsupervised grouping of unknown fault types, the formulation the paper's Step 2 adopts.","marker":"[24]"},{"why":"Supplies the Integrated Gradients algorithm whose pixel-level attributions are used to explain cluster assignments.","marker":"[54]"},{"why":"Motivates applying Integrated Gradients to a classifier and linking classification outputs back to input regions, grounding the pseudo-label explanation design.","marker":"[80]"},{"why":"Provides the OPTICS density-clustering algorithm, one of the three segmentation methods demonstrating framework flexibility.","marker":"[77]"},{"why":"Provides the self-organizing map algorithm used for offline and online fault segmentation.","marker":"[78]"},{"why":"Shows a self-organizing map applied to large-scale vibration monitoring of aircraft engines, supporting the choice of SOM for mechanical vibration clustering.","marker":"[79]"},{"why":"Represents the supervised deep-learning CB fault-diagnosis approach that the paper's unsupervised method is designed to bypass.","marker":"[13]"},{"why":"Describes vibration-signal monitoring of a spring-drive high-voltage CB, the experimental measurement basis for the collected dataset.","marker":"[82]"}],"fun_headline_variants":["No-label AI detects, clusters, and explains breaker faults","Unsupervised autoencoder spots breaker faults and XAI names the part","Healthy-only training: label-free fault diagnostics for circuit breakers","Vibration data + XAI: no-fault-label breaker diagnostics"],"cache_read_input_tokens":26752,"weakest_assumption_plain":"The detection threshold assumes that healthy vibration and acoustic behavior stays stable over time and across operating conditions, so any deviation beyond a fixed cutoff is a fault; if ordinary healthy variation from temperature, gas pressure, or interrupted current is large enough, the detector would either miss faults or flood operators with false alarms.","fun_headline_variants_meta":{"raw":{"variants":["No-label AI detects, clusters, and explains breaker faults","Unsupervised autoencoder spots breaker faults and XAI names the part","Healthy-only training: label-free fault diagnostics for circuit breakers","Vibration data + XAI: no-fault-label breaker diagnostics"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000267,"raw_usage":{"total_tokens":1690,"prompt_tokens":1094,"completion_tokens":596,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":710,"completion_tokens_details":{"reasoning_tokens":523}},"tokens_in":710,"tokens_out":596,"duration_ms":6225,"temperature":1.0,"reasoning_tokens":523,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:58:58.145920+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Collect healthy vibration and acoustic recordings from breakers in service across different seasons, gas pressures, and interrupted-current levels and feed them through the same healthy-trained CAE; if the mean residual distribution or its spread shifts so that a substantial fraction of healthy operations exceed $μ+3σ$, the fixed-threshold detection claim fails. A laboratory version would vary chamber gas pressure and temperature across healthy operations and measure how often the healthy-only threshold is crossed.","supporting_citations":[{"cited_title":"Implicit supervision for fault detection and segmentation of emerging fault types with deep variational autoencoders,","cited_arxiv_id":null,"evidence_quote":"Frames fault segmentation as unsupervised grouping of unknown fault types, the formulation the paper's Step 2 adopts."},{"cited_title":"Axiomatic attribution for deep networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the Integrated Gradients algorithm whose pixel-level attributions are used to explain cluster assignments."},{"cited_title":"From classification to segmentation with explainable ai: A study on crack detection and growth monitoring,","cited_arxiv_id":null,"evidence_quote":"Motivates applying Integrated Gradients to a classifier and linking classification outputs back to input regions, grounding the pseudo-label explanation design."},{"cited_title":"Large-scale Vibration Monitoring of Aircraft Engines from Operational Data using Self-organized Models,","cited_arxiv_id":null,"evidence_quote":"Shows a self-organizing map applied to large-scale vibration monitoring of aircraft engines, supporting the choice of SOM for mechanical vibration clustering."},{"cited_title":"A novel u-net and capsule network for few-shot high-voltage circuit breaker mechanical fault diagnosis,","cited_arxiv_id":null,"evidence_quote":"Represents the supervised deep-learning CB fault-diagnosis approach that the paper's unsupervised method is designed to bypass."},{"cited_title":"Diagnostics of high voltage circuit breakers by monitoring of vibration signals,","cited_arxiv_id":null,"evidence_quote":"Describes vibration-signal monitoring of a spring-drive high-voltage CB, the experimental measurement basis for the collected dataset."}],"review_version":2}