{"id":"bf23b51e-5f78-4d3f-a2ac-7db6aa032220","arxiv_id":"2507.15239","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"XSEI combines SHAP and occlusion sensitivity with manually defined ground-truth features to score how much arc fault diagnosis models actually use fault-related signal features.","lead":"This paper proposes a score called XSEI that uses explainable AI tools to check whether arc fault diagnosis models really focus on fault-related signal features rather than just matching patterns. The authors test the score across many machine learning and deep learning models and conclude that high accuracy alone does not guarantee trustworthy fault detection.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Ground-truth definition for deep models is not operational: Equation (4) uses exact equality on noisy continuous signals, and the paper supplies no threshold, so the reported deep-model XSEI scores cannot be reproduced from the text alone.","rationale":"The reader's weakest_assumption is that the ground-truth correctness definition is unvalidated, and that for deep models Equation (4) is not operational without a threshold. My stress-test pass converges on this as the most load-bearing concern, because both XSEI variants (Eq. 5 and Eq. 6) are defined as overlap ratios against this ground truth; if the ground truth is arbitrary or under-specified, the entire evaluation framework measures only agreement with the authors' choices. The feature-pool ground truth S is at least a finite, explicit set, so its arbitrariness is visible and testable (e.g., by adding a sixth candidate feature and seeing scores shift). The deep-model rule is worse: on real current waveforms with noise, x_n != x_hat_n is either always true or never true depending on preprocessing that is not reported. This is an internal operational gap, not merely a disagreement with external consensus. I do not see an internal inconsistency in the score definitions themselves, and the experimental campaign is broad, but the lack of a threshold/region protocol blocks independent reproduction of the central quantity. I also note the paper's own Section V limitation is about model-XAI combinations, not about this ground-truth gap, and the fast-drop and average-pooling conclusions in Sections IV-C/IV-E inherit the same ground-truth dependence. Therefore the verdict should remain CONDITIONAL, with the condition made explicit: release the ground-truth construction protocol and demonstrate threshold stability.","tokens_in":18080,"tokens_out":2474,"duration_ms":21596,"concrete_test":"Request the exact protocol for building r from Equation (4): the region length N, the alignment of x and x_hat, the difference metric, and the threshold used to declare x_n != x_hat_n. Then recompute one deep-model row (e.g., AE at SNR=-3 in Table III or the ArcNN occlusion example in Figure 4c) under a difference threshold swept over at least three orders of magnitude around the reported operating point. If the XSEI score changes by more than 0.2 across reasonable thresholds, the deep-model ground truth is not stable enough to support the claimed ranking of models.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that XSEI measures whether a model has found the real arc fault. For deep models (Eq. 6), the ground truth is r_n = 1 if x_n != x_hat_n, 0 otherwise, where x and x_hat are normal and arc-fault current signals segmented into N regions (Eq. 4). In continuous noisy measurements, exact equality has probability zero; every region will be marked 1 under any reasonable measurement noise, or no region will be marked 1 if the signals are aligned and quantized coarsely. Thus the ground-truth mask r is essentially undefined without a threshold on the difference magnitude, noise level, or segmentation choice, and the paper gives none. The reported deep-model scores (Tables II-V, e.g., AE 0.56, LCNN 0.22 at 5e-3 ms) therefore rest on an unspecified preprocessing decision. Additionally, Section III-A motivates S = {Variance, Entropy, Range, RMS, Integral} by assertion rather than by a falsifiable derivation or external standard, so for feature-pool models the same concern appears as: XSEI measures agreement with a hand-picked five-feature set, not with an independently validated ground truth of arc faults. The authors do flag a limitation in Section V (combinations of classifiers and XAI require expertise), but not this operational gap.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a soft evaluation indicator, XSEI, meant to assess whether AI-based arc fault diagnosis models actually rely on the true arc-fault features rather than on spurious patterns. For feature-pool models, XSEI is the Jaccard overlap between a hand-selected set of five features (Variance, Entropy, Range, RMS, Integral) and the top-5 SHAP features of the model. For deep learning models, XSEI is the Jaccard overlap between a ground-truth region mask defined by Eq. (4) and the region mask produced by occlusion sensitivity. The authors report test accuracy and XSEI scores for many ML and DL models over two experimental arc-fault datasets at several sample times and noise levels, and they propose a lightweight balanced neural network (LBNN) with average pooling. They conclude that high test accuracy does not guarantee trustworthiness, that a rapid drop in accuracy under lower precision is a sign that the model has found the key features, and that average pooling is preferable for arc fault diagnosis.","tokens_in":18441,"tokens_out":3688,"duration_ms":37553,"significance":"If the proposed evaluation indicator is valid, it would fill a genuine practical need: accuracy alone does not tell an engineer whether a fault detector is using diagnostic features or dataset artifacts. The paper deserves credit for using real arc-fault experiments following UL 1699 and GB 31143, for testing across many models and two datasets, and for presenting a simple, reproducible scoring formula. However, the central validity of XSEI rests entirely on the correctness of the authors' ground-truth definition of an arc-fault explanation. The deep-learning ground truth in Eq. (4) is not operational as stated (exact equality on noisy continuous signals), and the feature-pool ground-truth set is asserted rather than validated against any external standard. The evidence used to support XSEI is largely read from the same XSEI computations, making the validation circular. These issues are load-bearing for the paper's main claim.","major_comments":[{"comment":"The deep-learning ground truth r_n = 1 if x_n != x_hat_n is not operational for continuous noisy measurements. For two separately recorded current signals, exact equality has probability approximately zero under measurement noise, so r_n would be 1 for essentially every region; if the signals are aligned and quantized coarsely, r_n could be 0 everywhere. No threshold on the difference magnitude, no noise model, and no segmentation rule are supplied. Because Eq. (6) defines the XSEI score for all deep models, the scores reported in Tables II, III, IV, and V cannot be reproduced from the text and may be artifacts of an unspecified preprocessing decision.","section":"Section III-A2, Eq. (4)"},{"comment":"The ground-truth feature set S = {Variance, Entropy, Range, RMS, Integral} is selected by assertion rather than derived or externally validated. The text states that these features are 'normally' efficient for arc faults and then fixes them as ground truth; no evidence is given that exactly these five features, and not alternative candidates such as zero-crossing rate, spectral flatness, or wavelet coefficients, are the causal arc-fault features. Consequently, for feature-pool models, Eq. (5) measures agreement with a hand-picked set rather than with an independently established physical ground truth. A concrete external gold standard (expert-labeled fault segments, controlled fault injection, or a physics-based simulator) is needed to break this circularity.","section":"Section III-A1"},{"comment":"The heuristic that 'a rapid drop in test accuracy is a sign that the model has found the key features' is introduced as an interpretation of the same XSEI results and is then used as supporting evidence for XSEI and as the motivation for the average-pooling architecture in Section IV-E. No independent test, statistical significance, or out-of-sample prediction is provided; the fast-drop pattern is identified qualitatively in selected models (SEmodel, LightGBM, AE, ArcNN). This is circular validation: the phenomenon offered to validate the indicator is read off the indicator's own output. The paper should specify how the fast drop is measured and test it as a falsifiable prediction on held-out data.","section":"Section IV-C and Conclusion (ii)"},{"comment":"The paper does not specify how the occlusion-derived region mask r_Occlusion,i is computed from the occlusion sensitivity heatmap. No threshold on the responsibility score in Eq. (2), no mask size or stride for M_j, and no rule for converting continuous attribution values into the binary mask used in Eq. (6) are given. Without these procedural details, the deep-model XSEI scores are not reproducible, since the same heatmap can yield many different binary masks under different thresholds.","section":"Section III-B and Algorithm 1"}],"minor_comments":[{"comment":"The abstract contains the typo 'the the correct explanation'.","section":"Abstract"},{"comment":"The model listed as 'FW A-1DCNN' in Table I is referred to as 'IFW A-1DCNN' everywhere else; the table header should be corrected.","section":"Table I"},{"comment":"The text 'SRN=-5' is a typo and should read 'SNR=-5'.","section":"Section IV-C"},{"comment":"The conclusion says 'seven traditional machine learning methods and four deep learning methods', but Section IV lists nine ML models and five DL models; the counts are inconsistent.","section":"Section V"},{"comment":"In Algorithm 1, 'T op5' appears with an erroneous space, and the formatting of the g(f) vector and of the occlusion score accumulation is corrupted; these should be cleaned up.","section":"Algorithm 1"},{"comment":"The sentence 'when we mask all the signal, keeping nothing but Rmask' is confusingly worded and seems to contradict the condition in Eq. (3); the relationship between the responsibility score in Eq. (2) and the equivalence in Eq. (3) should be stated more carefully.","section":"Section II-B"},{"comment":"The text refers to Example 1 and Example 2, but the figure panels are labeled (a)-(d) without explicit cross-referencing; the reader must infer which panel corresponds to which example.","section":"Figure 3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript addresses a relevant practical problem and the experimental effort is real, but the reviewer's main concern is the status of the 'ground truth' explanation. Eq. (4) is non-operational without a threshold, and the feature-pool ground truth is an unvalidated assertion. The validation loop is circular because the fast-drop heuristic is justified by the same XSEI scores it is claimed to support. With a concrete thresholding procedure, an external gold standard, and a non-circular predictive test of the fast-drop heuristic, the paper could be publishable; without those, the central claim is not established."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The XSEI paper is worth taking seriously: it tries to answer a real question—does an arc fault classifier actually look at the fault, or just at something that correlates with it?—and it does so with a clean, simple overlap metric applied across a wide sweep of ML and deep models, two datasets, multiple sample times and noise levels. That is a legitimate contribution to the arc fault diagnosis literature, and the paper is honest about the expertise needed to combine XAI methods with classifiers.\n\nThe soft spots are real and load-bearing. For deep models, the ground-truth mask r_n = 1 if x_n != x_hat_n, on continuous noisy measurements, is not operational: exact equality has probability zero, so every region or no region gets flagged depending on how the signals happen to be quantized. No threshold, segmentation rule, or noise tolerance is given, which means the deep-model scores in Tables II–V cannot be reproduced or compared from the text alone. For feature-pool models, the five-feature ground-truth set S is asserted, not derived; XSEI is then measuring agreement with the authors' chosen features, not with an independently established ground truth about where arc faults actually occur. The validation loop is partly circular for the same reason: the qualitative occlusion plots and the 'fast drop means trustworthy' pattern are read from the same calculations the paper uses to define the score. The paper does flag a limitation in Section V about the expertise required for classifier–XAI combinations, but it does not address these operational gaps.\n\nThe 'fast drop' observation is interesting, but it is not established as a general criterion. It could be an artifact of aliasing or of sensitivity to downsampling rather than evidence of finding the 'real' fault. The average-pooling study is suggestive but covers only two architectures, without statistical error bars.\n\nNone of this kills the central idea. The metric is well-defined once the ground truth is specified, and the experimental sweep is broad. With an operationalized ground truth (threshold plus sensitivity analysis), an external gold standard or known fault-location benchmark, error bars, and released code and data, this could become a useful practical tool.\n\nI would send this to peer review: it deserves a serious referee and a request for major revision. It is aimed at arc fault diagnosis practitioners and researchers working on XAI for industrial diagnostics. I would not cite it in its current form.","headline":"A useful XAI-based trust metric for arc fault classifiers, but the ground-truth definition is too loose to reproduce; worth reviewing with major revision.","tokens_in":18913,"tokens_out":2425,"would_cite":false,"duration_ms":28261,"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":"This paper proposes XSEI, a soft evaluation indicator that measures whether an arc fault diagnosis model's explanations overlap with a defined physical ground truth of arc faults, arguing that high accuracy alone does not certify that a…","keywords":["arc fault diagnosis","explainable artificial intelligence","soft evaluation indicator","SHAP","occlusion sensitivity","feature extraction score","lightweight neural network","model trustworthiness"],"falsifier":"A decisive test would be to low-pass filter the arc-fault current so that the high-frequency components defining the ground-truth regions are removed; if a high-XSEI model still classifies faults accurately, the score does not measure reliance on the defined fault features.","tokens_in":17863,"feed_emoji":"⚡","tokens_out":6427,"duration_ms":65081,"temperature":0.7,"pith_summary":"The paper's central claim is that classification accuracy is not enough to certify an arc fault diagnosis model, because a model can reach high accuracy by exploiting dataset patterns unrelated to the fault. It proposes XSEI, a soft evaluation indicator that compares a model's explanation with a predefined ground truth of arc fault features: the five hand-picked features Variance, Entropy, Range, RMS, and Integral for feature-pool models, and the signal regions that change between normal and fault currents for deep models. The score is a Jaccard overlap between the model's attribution and this ground truth, so values near 1 mean the model's explanation coincides with the presumed physical fault. The paper argues these scores can guide model selection and ensemble weighting and reports experiments on two datasets with varying sample times and noise levels.","feed_headline":"New soft score reveals which AIs truly find arc faults","feed_subtitle":"Accuracy alone can't certify an arc-fault model; XSEI checks explanations against physical ground truth.","key_machinery":"The central object is a Jaccard-style soft score $g(f_i) \\in [0,1]$ that measures overlap between an XAI attribution and the paper's ground-truth fault explanation. For feature-pool models the attribution comes from SHAP values, and for deep models it comes from occlusion sensitivity maps; the ground truth is either the five-feature set $S$ or the region-difference vector $r$. This score turns an explanation heatmap into one comparable number, which is what lets the paper rank models and propose the soft score as a basis for selection and ensemble weighting.","core_discovery":"On its own terms, the paper establishes a definition of what a correct explanation of an arc fault should look like, then measures models against it. For feature-pool models the correct explanation is the set $S = \\{\\text{Variance}, \\text{Entropy}, \\text{Range}, \\text{RMS}, \\text{Integral}\\}$; the model's explanation is the top-5 SHAP features, and the score is $g(f_i) = |S \\cap S_{\\mathrm{SHAP},i}| / |S \\cup S_{\\mathrm{SHAP},i}|$. For deep models the correct explanation is the regions $r$ where the arc fault current differs from the normal current, and the model's explanation is the regions $r_{\\mathrm{Occlusion},i}$ found by occlusion sensitivity; the score is $g(f_i) = |r \\cap r_{\\mathrm{Occlusion},i}| / |r \\cup r_{\\mathrm{Occlusion},i}|$. Across the tested models, SEmodel and the L2/L1 norm score highest among machine learning methods, while AE and ArcNN score highest among deep learning methods, and several top-scoring models show a sharp accuracy drop when sample time or noise changes, which the authors read as evidence that the models are genuinely keyed to the fault features rather than to general patterns.","pith_inferences":["The region-difference ground truth in Equation (4) presumes the normal and fault signals can be aligned region by region; if real deployments have phase shifts or timing jitter, the score would penalize models that locate the same physical fault in shifted regions.","XSEI as defined measures agreement with the authors' chosen feature set, so a model could score high by using those features as statistical proxies rather than as causal fault indicators; an intervention that removes the high-frequency content while preserving the proxy feature values would separate the two cases.","The same overlap-scoring scheme could be transferred to other industrial diagnostic tasks where a small physical ground-truth feature set is known, but the definition of that ground truth would be the main obstacle.","Using XSEI as ensemble weights may over-reward models that match the chosen ground truth and under-reward models whose complementary errors would improve the ensemble; weighting should be validated against ensemble accuracy, not just soft scores."],"forward_implications":["If XSEI is right, a model with high accuracy but low soft score should be viewed with suspicion even when its test accuracy is excellent.","The soft score can be used as a basis for model selection when several models have comparable accuracy, as in the reported tables.","The scores can serve as ensemble weights, so models whose explanations match the ground truth contribute more to the combined decision.","A rapid drop in test accuracy as sample time lengthens or noise rises becomes a positive signal that the model relies on genuine fault features.","Average pooling is a better architectural choice than max pooling for locating arc faults, based on the soft score comparisons."],"supporting_citations":[{"why":"Supplies the SHAP attribution method used to extract top-5 features for feature-pool models.","marker":"[26]"},{"why":"Provides the Shapley value feature-attribution estimation framework behind the SHAP scores.","marker":"[27]"},{"why":"Supplies the occlusion sensitivity method used for deep model region attribution.","marker":"[28]"},{"why":"Discusses the feature-removal strategies that SHAP-based attribution depends on.","marker":"[29]"},{"why":"One source for the ground-truth feature candidates used in the feature-pool definition.","marker":"[15]"},{"why":"Provides a public arc fault dataset and features that inform the ground-truth feature set and the autoencoder baseline.","marker":"[16]"}],"fun_headline_variants":["XSEI score tests if arc-fault AIs really see faults","Soft indicator exposes which arc-fault models are trustworthy","New metric judges arc-fault AI by correct explanations","Arc fault AI trust: a soft score that checks reasoning","Score reveals if arc-fault models key on actual faults"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper assumes its hand-picked feature set and region-difference rule are the true explanation of arc faults, and in particular that high-frequency current distortion is the causal mechanism; without a threshold for what counts as $x_n \\neq \\hat{x}_n$ in noisy data, the deep-learning ground truth is not operational on its own.","fun_headline_variants_meta":{"raw":{"variants":["XSEI score tests if arc-fault AIs really see faults","Soft indicator exposes which arc-fault models are trustworthy","New metric judges arc-fault AI by correct explanations","Arc fault AI trust: a soft score that checks reasoning","Score reveals if arc-fault models key on actual faults"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000373,"raw_usage":{"total_tokens":1996,"prompt_tokens":949,"completion_tokens":1047,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":565,"completion_tokens_details":{"reasoning_tokens":965}},"tokens_in":565,"tokens_out":1047,"duration_ms":8897,"temperature":1.0,"reasoning_tokens":965,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:37:28.359680+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive test would be to low-pass filter the arc-fault current so that the high-frequency components defining the ground-truth regions are removed; if a high-XSEI model still classifies faults accurately, the score does not measure reliance on the defined fault features.","supporting_citations":[{"cited_title":"Algorithms to es- timate shapley value feature attributions,","cited_arxiv_id":null,"evidence_quote":"Provides the Shapley value feature-attribution estimation framework behind the SHAP scores."},{"cited_title":"Visualizing and understanding convo- lutional networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the occlusion sensitivity method used for deep model region attribution."},{"cited_title":"The many shapley values for model explanation,","cited_arxiv_id":null,"evidence_quote":"Discusses the feature-removal strategies that SHAP-based attribution depends on."},{"cited_title":"Series ac arc fault detection method based on hybrid time and frequency analysis and fully connected neural network,","cited_arxiv_id":null,"evidence_quote":"One source for the ground-truth feature candidates used in the feature-pool definition."},{"cited_title":"Intever public database for arcing event detection: Feature analysis, benchmark test, and multi-scale cnn application,","cited_arxiv_id":null,"evidence_quote":"Provides a public arc fault dataset and features that inform the ground-truth feature set and the autoencoder baseline."}],"review_version":1}