{"id":"f9cf303a-0065-45e7-9777-277deb993899","arxiv_id":"1909.10583","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"A multiclass support vector machine trained on simulated bus voltages detects and locates high impedance faults with 100% reported security and dependability in an IEEE 13-node distribution network.","lead":"This paper tests several standard machine learning classifiers on simulated voltage data from an IEEE 13-node power distribution feeder to detect and locate high impedance faults. The authors report that a multiclass support vector machine achieves perfect detection and classification in their simulations.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The evaluation protocol in §4 does not establish an independent test set, so the reported 100% dependability/security may be a training-set artifact rather than a validated result.","rationale":"The reader's conditional verdict is a fair response to the paper. My pass locates the load-bearing weakness one step earlier than the reader: before asking whether the Brahma two-diode HIF model faithfully represents real arcs, one must know whether the reported 100% values were computed on samples that were not used for training. As written, §4.1 (60 training vs 100 normal test samples with only 100 total normal observations), §4.3 (validation on training data), and §4.4 (unspecified 50-sample test per class) leave this ambiguous. This is an internal-validity concern rather than a disagreement with the use of ML for HIF detection; the algorithm exposition is standard and the simulation pipeline is reasonable. Credit is due for the comparison with prior methods and for the clear statement of the model parameters, but no code, data, or staged-fault validation is provided. The same lack of detail undermines the robustness sentence about load variation and capacitor switching, which is asserted without a reported experiment. I therefore keep the conditional stance: acceptance should require a clean, reproducible train/test protocol and ideally a held-out operating condition. I mark partial agreement because the reader emphasized physical model fidelity, while the more immediately decisive issue is the independence of the evaluation set; both concerns point to the same need for more rigorous validation before the broad claims in §5 are accepted.","tokens_in":11476,"tokens_out":5771,"duration_ms":62719,"concrete_test":"Obtain the code and data, or the exact train/test indices used in §4.4, and re-run the M-SVM evaluation under a strict split by simulation realization: train on one subset of runs and test on runs generated with different random seeds, reporting per-class accuracy and security/dependability. If any held-out run is misclassified, or if the original train/test indices overlap with training samples, the 100% values in Table 1 are an artifact of the split and the central claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim, stated in §4.4 and the conclusion, is that M-SVM detects and locates HIFs with 100% dependability and security and remains robust under load variation and capacitor switching. The evidence is the 400-sample simulated data set, but the manuscript never specifies a non-overlapping train/test split. §4.1 uses 60 normal samples for training and then tests on 100 normal samples, although §4 says the data set contains only 100 normal observations; §4.3 says the SVM classifier was 'validated on training data'; §4.4 tests 50 observations per class without stating that these observations were withheld during training. If these sets overlap, or if test samples are temporally adjacent to training samples from the same simulation runs, the reported 100% accuracy only shows that the classifier memorized location-specific voltage signatures. The robustness claim is also not supported by any described experiment: no load-variation or capacitor-switching scenario is reported, and Table 1 does not list operating conditions. Even granting the HIF-model fidelity, the evaluation protocol as written cannot certify the generality stated in the conclusion.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a data-driven scheme for detecting and locating high impedance faults (HIFs) in power distribution networks. It applies Principal Component Analysis with Hotelling's T2 statistics, Fisher Discriminant Analysis, binary Support Vector Machines, and Multiclass Support Vector Machines to bus voltage data from a simulated IEEE 13-node feeder. The HIF is modeled with the Brahma two-diode source-resistance model. The authors report that PCA detects HIFs but cannot classify them, FDA can both detect and locate faults, and M-SVM achieves 100% security and 100% dependability, with claims of robustness under load variation and capacitor switching.","tokens_in":11840,"tokens_out":3991,"duration_ms":43797,"significance":"If the reported results are valid, the proposed approach would be practically attractive because it uses measurements from existing CTs and PTs, does not require signal injection, and is computationally light enough for embedded relay implementation. The manuscript gives a clear description of the algorithmic pipeline, uses a recognized test feeder, and compares against earlier HIF detection methods in Table 1. However, the significance is currently limited by the simulation-only validation, the absence of a documented independent test set, and the lack of experiments supporting the robustness claims. The paper does not provide code or data, so reproducibility of the exact 100% figures cannot be independently checked.","major_comments":[{"comment":"The evaluation protocol does not establish an independent test set. The dataset is described in §4 as containing 100 normal observations and 300 fault observations, but §4.1 trains PCA on 60 normal samples and tests on 100 normal samples plus 100 faulty samples; §4.3 states that SVM classifiers were \"validated on training data\"; and §4.4 tests 50 observations per class without stating that these observations were withheld during training. If the test samples overlap the training samples, or are temporally adjacent samples from the same simulation runs, the reported 100% dependability and security only show that the classifier memorized the training data. Please specify the exact train/test split and rerun the evaluation with disjoint samples, ideally with repeated stratified cross-validation and confidence intervals.","section":"Section 4, especially §4.1, §4.3, and §4.4"},{"comment":"The conclusion that M-SVM is \"robust against capacitor and load switching transients\" is not supported by any experiment described in the paper. No load-variation or capacitor-switching scenario is reported in Section 4, and Table 1 lists only security and dependability percentages without any description of the operating conditions tested. Either add experiments that explicitly vary load and include capacitor switching, with results, or remove this robustness claim from the abstract, Section 4.4, and the conclusion.","section":"§4.4, Table 1, and Section 5"},{"comment":"All training and test data are generated from a single HIF model, the Brahma two-diode model, with parameters Vp, Vn, Rp, and Rn that are said to be \"tuned according to test feeder\" but are not validated against staged-fault or field data. The reported perfect scores therefore mainly demonstrate that the classifier separates samples of this particular simulation model. To support the paper's claim of detecting \"all types of HIF,\" validation on independent HIF data—for example, staged-fault recordings or a structurally different HIF model—is needed, or the claims must be explicitly limited to the modeled fault characteristics.","section":"§2.2 and §4"},{"comment":"The M-SVM result is reported as 100% on 200 test observations (50 per class) with fixed hyperparameters (Gaussian RBF kernel with sigma = 0.5, penalty factor 10, regularization parameter lambda = 1) and no sensitivity analysis or uncertainty quantification. Because the test set is small and the hyperparameters are tuned on the same simulated data, the 100% figure should be accompanied by confidence intervals, repeated cross-validation results, or a sensitivity study over hyperparameter values before it can support the strength of the central claim.","section":"§4.4"}],"minor_comments":[{"comment":"Equation (1) writes 1/sqrt(1-n) X = U Sigma V^T; the sign under the square root is presumably a typo, and the expression should be 1/sqrt(n-1) X = U Sigma V^T.","section":"§3.1, Eq. (1)"},{"comment":"The denominator in Eq. (12) is n(n-1), but consistency with Eq. (4) suggests it should be n(n-a); please check the formula and the degrees of freedom.","section":"§3.2, Eq. (12)"},{"comment":"The abstract says data from voltage and current sensors are used, but the input matrix in §4 is described only in terms of 29 bus voltage variables; clarify whether current measurements are actually used or remove the reference to current sensors.","section":"Abstract and §4"},{"comment":"The name of the statistic is misspelled as \"Hotteling's\" in several places; the standard spelling is \"Hotelling's.\"","section":"Throughout"},{"comment":"The introduction refers to \"Section II,\" \"Section III,\" etc., but the actual sections are numbered 1 through 5; renumber the cross-references.","section":"Section 1"},{"comment":"Table 1 compares the proposed method with prior methods without stating the test conditions, data sources, or fault scenarios used for each entry; without common test conditions, the comparison should be interpreted cautiously.","section":"Table 1"},{"comment":"The caption says the M-SVM predicted labels are shown on a 2-D plane, but no projection method for the 29-dimensional data is described in §4.4; please clarify how the 2-D plot was obtained.","section":"Figure 19"}],"recommendation":"major_revision","confidential_remarks":"The evaluation gap identified in the major comments is substantial: as written, the 100% results may be artifacts of training-set overlap or of single-model simulation. If the authors can provide a clean train/test split, cross-validated results, and experiments for the claimed load and capacitor switching robustness, the paper could become a useful contribution; otherwise the claims should be substantially narrowed. I see no evidence of misconduct, but the manuscript needs a rigorous re-analysis before it can be considered for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a competent write-up of a routine machine-learning exercise, but the reported 100% numbers are not believable because the paper never establishes that test samples were withheld from training. The stress-test note is on target.\n\nWhat is actually new: applying multiclass SVM (RBF kernel) to locate three HIF positions on the IEEE 13-node feeder, alongside PCA and FDA baselines. The methods are textbook and the comparison is not novel. Credit where due: the simulation model follows Brahma's HIF model, the description of PCA/FDA/SVM is accurate, and the paper is clearly organized.\n\nThe soft spots are load-bearing. Section 4 says 400 observations: 100 normal, 300 faulty across three locations. Section 4.1 says training uses 60 normal samples and testing uses 100 normal samples—that is only possible if test and training sets overlap, since only 100 normal samples exist. Section 4.3 states the SVM was 'validated on training data,' which is invalid as a generalization check. Section 4.4 tests 50 observations per class (200 total) but the data set has only 100 normal samples; again, overlap or repeat runs are the only way to get 50 normal test samples. So the 100% security/dependability could easily be the classifier memorizing the simulation runs, and the paper does not rule that out.\n\nThe robustness claim is even thinner. The abstract and conclusion say the method handles load variation and capacitor switching, but no such experiment is described anywhere in Section 4. Table 1 compares performance indices without listing operating conditions, so the comparison is not meaningful.\n\nThe HIF model itself is a single source-diode-resistance model with parameters 'tuned according to test feeder.' That is fine for a proof of concept, but it undercuts the claim to detect 'all types' of HIF. No staged-fault data, no code/data release, no error bars.\n\nAll of this adds up to a conditional proof of concept, not a validated protection scheme. The paper is readable and the basic pipeline is sensible, so it could be useful as a starting point for a student or a literature review. But the central claim as written is not supported. I would desk-reject this version, with an invitation to resubmit using a properly specified non-overlapping train/test split, cross-validated final results, and at least one scenario with load/capacitor disturbance actually reported.","headline":"Routine SVM application with a fatal evaluation gap: no evidence of a clean train/test split, so the 100% results are likely training-set artifacts.","tokens_in":12276,"tokens_out":2882,"would_cite":false,"duration_ms":30681,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A multiclass support vector machine trained on bus voltages detects high-impedance faults and locates the faulted section of a distribution feeder, hitting 100% dependability and security in the paper's 400-sample test.","keywords":["High Impedance Fault","Support Vector Machine","Multiclass SVM","Fault Detection","Fault Location","Principal Component Analysis","Fisher Discriminant Analysis","13-node distribution feeder"],"falsifier":"Run the trained M-SVM on voltage measurements from staged high-impedance faults on a real distribution feeder, with load and capacitor-switching events included; any missed fault lowers dependability below 100%, and any trip on a switching transient lowers security below 100%.","tokens_in":11236,"feed_emoji":"⚡","tokens_out":10754,"duration_ms":99225,"temperature":0.7,"pith_summary":"High-impedance faults draw current well below normal load levels, so conventional overcurrent protection misses most of them, and their arcing signatures resemble load and capacitor switching. This paper tries to show that a data-driven classifier can solve both problems at once. Using bus-voltage observations from a simulated 13-node distribution feeder, the authors train a multiclass support vector machine to distinguish the healthy state from high-impedance faults at three locations. They report that the classifier detects every fault and never mistakes a non-fault event for one, giving 100% dependability and 100% security, and that it keeps this accuracy when loads vary and capacitors switch. If true, existing voltage and current sensors plus a numerical relay could detect and isolate these faults without new hardware or signal injection.","feed_headline":"Multiclass SVM finds and locates high-impedance faults at 100%","feed_subtitle":"Bus voltages from a 13-node feeder let the classifier locate the fault despite load and capacitor switching.","key_machinery":"The central machinery is the multiclass support vector machine (M-SVM), a supervised classifier that separates data classes with a maximum-margin hyperplane in a kernel feature space; here it uses a Gaussian radial-basis kernel with $\\sigma = 0.5$ and penalty $C = 10$, labeling four classes (healthy plus three fault locations). It runs directly on 29-dimensional bus-voltage measurements, so no separate feature extraction is needed. For comparison the paper also uses principal component analysis with the $T^2$ statistic for detection and Fisher discriminant analysis with its discriminant function for classification; M-SVM is the mechanism that carries the simultaneous detect-and-locate claim.","core_discovery":"On the paper's own terms, the central discovery is that multiclass support vector machines outperform both principal component analysis and Fisher discriminant analysis for high-impedance fault diagnosis: PCA detects the fault but cannot tell fault locations apart, FDA can locate faults, and M-SVM both detects and locates them with perfect scores. The result is obtained by training on 29 bus-voltage variables from a 13-node distribution feeder model, with a Gaussian radial-basis kernel and a cross-validated penalty factor, and testing on 400 observations covering the normal state and three fault positions. The paper also claims the classifier keeps its perfect accuracy under load variation and capacitor switching, and that it can be implemented as a numerical relay using measurements already available in distribution networks.","pith_inferences":["Editorial inference: because all 400 samples come from one two-diode simulation model, the 100% figures measure separation on that model; real-world performance should be expected to drop unless the model's randomness matches staged-fault data.","Editorial inference: the approach naturally extends to other feeder topologies and more fault locations, but the training data would need to cover those cases; nothing in the paper shows transfer across networks.","Editorial inference: a strong testable extension is to feed the trained classifier with recorded staged-fault voltage waveforms; if the arc's real randomness differs from the simulation, accuracy will reveal it."],"forward_implications":["A protective relay could use the existing voltage and current measurements to trip on high-impedance faults without injecting any test signal, so power quality is not degraded.","Load changes and capacitor switching will not trigger false trips, because the classifier was tested on those transients.","The faulted section can be identified, not just the presence of a fault, which enables targeted isolation rather than shutting down the whole feeder.","The M-SVM result matches the 100% dependability and security of mathematical-morphology methods while outperforming wavelet, time-frequency, and morphological-gradient methods in the comparison table.","The classifiers are low-order and computationally light, so they can run on embedded numerical-relay hardware."],"supporting_citations":[{"why":"Supplies the two-diode source-diode-resistance HIF model from which all training and test samples are generated.","marker":"Brahma, 2013"},{"why":"Establishes that conventional overcurrent protection detects only about 17% of HIFs, defining the problem the classifier must solve.","marker":"Jones, 1996"},{"why":"Provides the characterization of arcing, asymmetry, and randomness that the simulation model is built to reproduce.","marker":"Ghaderi et al., 2016"},{"why":"Supplies the PCA-based fault-detection procedure and $T^2$ statistic used for the first detection method.","marker":"Jamil et al., 2015"},{"why":"Gives the F-distribution threshold formula used to turn $T^2$ values into fault decisions.","marker":"Chiang et al., 2000"},{"why":"Provides the SVM maximum-margin classification theory behind the binary and multiclass classifiers.","marker":"Burges, 1997"},{"why":"Supplies the kernel-trick feature-space mapping that lets the RBF-kernel SVM separate nonlinear HIF patterns.","marker":"Nayak, 1998"},{"why":"Is the mathematical-morphology method in the comparison table that the proposed M-SVM matches at 100% dependability and security.","marker":"Gautam and Brahma, 2012"}],"fun_headline_variants":["Multiclass SVM perfectly detects and locates high-impedance faults","SVM scheme isolates high-impedance faults on IEEE 13-node feeder","Multiclass SVM outperforms PCA and FDA for high-impedance faults","Perfect detection and location of high-impedance faults with M-SVM"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the two-diode simulation model reproduces real high-impedance arc behavior, since all 400 training and test samples come from that single model.","fun_headline_variants_meta":{"raw":{"variants":["Multiclass SVM perfectly detects and locates high-impedance faults","SVM scheme isolates high-impedance faults on IEEE 13-node feeder","Multiclass SVM outperforms PCA and FDA for high-impedance faults","Perfect detection and location of high-impedance faults with M-SVM"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000349,"raw_usage":{"total_tokens":1857,"prompt_tokens":847,"completion_tokens":1010,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":463,"completion_tokens_details":{"reasoning_tokens":929}},"tokens_in":463,"tokens_out":1010,"duration_ms":8620,"temperature":1.0,"reasoning_tokens":929,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:16:46.956719+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the trained M-SVM on voltage measurements from staged high-impedance faults on a real distribution feeder, with load and capacitor-switching events included; any missed fault lowers dependability below 100%, and any trip on a switching transient lowers security below 100%.","supporting_citations":[{"cited_title":"Detection of high impedance fault in power distri- bution systems using mathematical morphology,","cited_arxiv_id":null,"evidence_quote":"Supplies the two-diode source-diode-resistance HIF model from which all training and test samples are generated."},{"cited_title":"High impedance fault detection implementation is- sues,","cited_arxiv_id":null,"evidence_quote":"Establishes that conventional overcurrent protection detects only about 17% of HIFs, defining the problem the classifier must solve."},{"cited_title":"High impedance fault detection: A review,","cited_arxiv_id":null,"evidence_quote":"Provides the characterization of arcing, asymmetry, and randomness that the simulation model is built to reproduce."},{"cited_title":"Fault diagnosis of pakistan research reactor-2 with data-driven techniques,","cited_arxiv_id":null,"evidence_quote":"Supplies the PCA-based fault-detection procedure and $T^2$ statistic used for the first detection method."},{"cited_title":"Advanced Textbooks in Control and Signal Processing","cited_arxiv_id":null,"evidence_quote":"Gives the F-distribution threshold formula used to turn $T^2$ values into fault decisions."},{"cited_title":"A Tutorial on Support Vector Machines for Pattern Recognition,","cited_arxiv_id":null,"evidence_quote":"Provides the SVM maximum-margin classification theory behind the binary and multiclass classifiers."},{"cited_title":"Support vector machines and machine learning on docu- ments,","cited_arxiv_id":null,"evidence_quote":"Supplies the kernel-trick feature-space mapping that lets the RBF-kernel SVM separate nonlinear HIF patterns."},{"cited_title":"Detection of high impedance fault in power distribution systems using mathematical morphology","cited_arxiv_id":null,"evidence_quote":"Is the mathematical-morphology method in the comparison table that the proposed M-SVM matches at 100% dependability and security."}],"review_version":1}