{"id":"2184ea04-945e-4762-8f06-51866a756897","arxiv_id":"2502.05846","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"HAVOK analysis of simulated 12 kV distribution currents detects high-impedance arc faults in 0.45 ms, but the latency claim conflicts with the paper's own embedding-window requirement.","lead":"This paper applies HAVOK, a data-driven method from nonlinear dynamics, to detect high-impedance arc faults in simulated 12 kV power distribution lines, reporting detection in 0.45 ms. The work is notable because it claims fast, interpretable fault detection, but the speed claim conflicts with the method's own embedding requirements and all results are simulation-only.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 0.45 ms detection claim is not causally realizable as described: q=40 at 20 kHz requires a 2 ms (and per §3.1, a 20 ms) embedding window, and Algorithm 1 uses a non-causal central-difference derivative.","rationale":"I agree with the reader's weakest_assumption, and the full text strengthens it. The contradiction is internal, not a matter of consensus: §3.1 explicitly states the embedding-window rule, and Algorithm 1 explicitly uses a central-difference derivative, which is non-causal by construction. Since the headline latency is both the paper's main novelty and the basis of the 99.36% comparison, the central claim cannot be accepted as stated. I do not see a second concern that would change this: the simulation-only validation and the lack of train/test separation are real but secondary; the HAVOK feature idea may still be salvageable with a causal sliding window, honest latency accounting that includes the 2 ms (or 20 ms) window, and a held-out evaluation. No machine-checked proof or code release is provided, so the internal timing check is the decisive test.","tokens_in":14730,"tokens_out":4157,"duration_ms":40022,"concrete_test":"Implement a strictly causal version of Algorithm 1: at each new sample t_k, build the Hankel matrix from only the last 40 samples (and separately from the last 400 samples per the one-cycle rule), compute the SVD, and evaluate the threshold rule on the resulting v_r. Record the first t after fault injection at 0.2 s. If the first detection is ≥2 ms (or ≥20 ms), or never occurs, the 0.45 ms claim is refuted. As a second, independent check, truncate the simulation at 0.20045 s and recompute the forcing coordinate; if the value at 0.20045 s changes, the original detection was non-causal.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central contribution is the 0.45 ms latency (Abstract; §4.1; Table 2, Case A), used to claim 99.36% improvement over [37]. At the stated 20 kHz sampling, 0.45 ms is 9 samples. But Algorithm 1 must first form a Hankel matrix by stacking q=40 delayed versions of the current, and §3.1 states the embedding window should span at least one 50 Hz cycle (20 ms). Even the minimal causal construction q·Δt = 2 ms exceeds 0.45 ms; with the paper's own one-cycle rule it is 20 ms. Moreover, Algorithm 1 step 4 computes derivatives of V with a fourth-order central difference, which requires future samples, so the forcing coordinate at t=0.20045 s depends on data after the claimed detection time. The reported 0.20045 s crossing is therefore an offline post-hoc property of the full simulation, not a causal real-time alarm. This invalidates the speed headline and the 99.36% comparison derived from it. The accuracy claim is also weakened by threshold fitting on the same 50 simulations with no held-out set, but the timing contradiction is the load-bearing defect.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a high-impedance arc fault (HIAF) detector for 12 kV, 50 Hz distribution systems based on Hankel alternative view of Koopman (HAVOK) analysis. A Hankel matrix is built from 40 time-delayed copies of the measured current, decomposed by SVD, and a linear regression model is fit to the leading eigen-time-delay coordinates; the last coordinate is interpreted as a forcing operator whose magnitude is thresholded to distinguish arc faults from non-arcing disturbances and other faults. The authors report detection times as low as 0.45 ms, a 99.36% improvement over the harmonic-randomness and waveform-distortion method of [37], and 100% detection accuracy with 0% false positives across 50 simulations per scenario covering different grounding surfaces, load types, and noise levels. Validation is entirely by PSCAD simulation, and the paper acknowledges the need for future hardware-in-the-loop and field testing.","tokens_in":14937,"tokens_out":3692,"duration_ms":38004,"significance":"If the reported performance were supported by the described algorithm, the paper would offer a practically relevant, interpretable alternative to black-box AI fault detectors, and the comparative evaluation against [37] on the same simulated data is a useful feature. The paper also honestly acknowledges its simulation-only scope and the need for real-world validation. However, the central latency claim is internally contradicted by the method's own embedding-window requirements and by the non-causal derivative computation in Algorithm 1, and the accuracy claims rest on thresholds calibrated on the same simulations used for evaluation. As presented, the main quantitative contributions are not supported, and the significance of the work is therefore substantially reduced.","major_comments":[{"comment":"The claimed 0.45 ms detection time in Case A is not causally realizable with the described algorithm. Section 3.1 states that the total embedding window (embedding dimension times sampling interval) should span at least one 50 Hz cycle, i.e., 20 ms, and the paper uses q=40 at a 20 kHz sampling rate, which already gives a 2 ms Hankel window. Even the minimal 2 ms window exceeds the reported 0.45 ms detection latency, and the paper's own one-cycle rule implies a 20 ms window. Furthermore, Algorithm 1 step 4 computes derivatives of V using a fourth-order central difference, which requires future samples, so the forcing-coordinate value reported at t=0.20045 s depends on current data after the claimed detection time. The reported crossing is therefore an offline, post-hoc property of the full simulation record, not a causal real-time alarm. This invalidates the headline latency and the 99.36% speed comparison in Section 5 and Table 3.","section":"§3.1, §4.1, Algorithm 1"},{"comment":"The detection thresholds in Algorithm 1 (0.06–0.18 for arc faults, >0.2 for other faults, <0.045 for non-arcing disturbances) were selected after \"numerous experiments\" on the same simulations that are then used to report 100% accuracy and 0% false positives across 50 simulations per scenario. No held-out dataset, cross-validation procedure, or separate calibration/test split is described, and the 50 simulations per scenario are not characterized in terms of parameter distributions, seeds, or confidence intervals. The reported accuracy is therefore in-sample calibration performance, not an estimate of detection performance on unseen conditions, so the robustness claim in Section 4 is not supported.","section":"§4 and Algorithm 1"},{"comment":"The algorithm does not describe a causal, sliding-window implementation that would allow detection at a well-defined time after fault onset. The Hankel matrix in Eq. (15) is formed from the entire simulated record, the SVD is global, and the forcing operator is extracted from the full-data SVD. There is no procedure for updating the SVD or the forcing coordinate as new samples arrive, and no computational-cost analysis for real-time operation. Without such a causal formulation, the reported detection times in Table 2 are not meaningful as alarm latencies.","section":"§4.1 and Algorithm 1"}],"minor_comments":[{"comment":"","section":"Eq. (4)"},{"comment":"","section":"Eq. (12)"},{"comment":"","section":"Algorithm 1, step 6"},{"comment":"","section":"Table 3"},{"comment":"","section":"Section 4.7"},{"comment":"","section":"References"}],"recommendation":"reject","confidential_remarks":""},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe thing to know: this is the first paper to apply HAVOK (Hankel alternative view of Koopman) to high-impedance arc fault detection in medium-voltage distribution networks. That is real novelty. The authors simulate a 12 kV PSCAD system, extract a forcing operator, and show thresholds that separate arc faults from load switching and line-ground faults. If the method worked as claimed, it would be a useful, interpretable alternative to black-box classifiers.\n\nThere is also credit to give: the arc model is described carefully, the test scenarios cover different grounding surfaces, noise, and induction motor loads, and the comparison against [37] is a reasonable benchmark. The idea of using the intermittent forcing in HAVOK as an arc signature is plausible.\n\nThe soft spots are not minor. The headline latency of 0.45 ms is contradicted by the paper's own machinery. With q=40 at 20 kHz, the Hankel embedding window is 2 ms, and the paper itself says the window should span at least one 50 Hz cycle, i.e., 20 ms. Either way, the forcing coordinate at t=0.20045 s cannot exist without samples after that time. On top of that, Algorithm 1 computes derivatives of V with a fourth-order central difference, which uses future samples. So the reported detection time is an offline property of the full simulation, not a causal real-time alarm. That load-bearing claim, and the 99.36% speed-up derived from it, do not hold as stated.\n\nThe accuracy claims also rest on thresholds fitted to the same 50 simulations per scenario, with no held-out test set and no error bars. The conclusion acknowledges only simulation-based validation, which is honest, but it doesn't fix the timing problem.\n\nThis paper deserves a serious referee because the core application is new and the underlying idea might be salvageable with a proper causal implementation and honest latency accounting. But as written, the central contribution is unsupported. I would not cite it in its current form, and I would send it back for major revision before any acceptance.","headline":"A genuinely new HAVOK application to MV arc fault detection, undermined by a detection-time claim that contradicts the method's own window and causality requirements.","tokens_in":15575,"tokens_out":2647,"would_cite":false,"duration_ms":24772,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a Koopman-based forcing signature can flag high-impedance arc faults in 12 kV distribution systems in 0.45 ms, with 100% simulation accuracy.","keywords":["Arc fault detection","High-impedance arc faults","Medium voltage distribution","Hankel alternative view of the Koopman","Koopman operator","Forcing operator","Power system safety","PSCAD simulation"],"falsifier":"Re-run the HAVOK pipeline on the same 20 kHz current traces with a causal constraint: the Hankel matrix may contain only samples up to the candidate detection time, and the embedding window must span at least one 50 Hz cycle (20 ms) as the paper prescribes. If the forcing burst first crosses the arc threshold at 0.45 ms under that constraint the latency claim is confirmed; if the earliest compliant detection occurs only after the window has filled, the headline latency is not supported.","tokens_in":14480,"feed_emoji":"⚡","tokens_out":7315,"duration_ms":69095,"temperature":0.7,"pith_summary":"This paper sets out to show that high-impedance arc faults in medium-voltage (12 kV) distribution networks can be detected much faster and more transparently than existing methods allow. It claims that the HAVOK forcing operator, a quantity extracted from the nonlinear current waveform by Koopman analysis, produces a clean, interpretable fault signature within 0.45 ms of fault onset. That would matter because arc faults are hard to see: their currents are weak, noisy, and easily confused with normal load changes or other faults. A detector that is both fast and explainable could let protection systems respond in real time and give engineers a physical reason to trust each alarm.","feed_headline":"Arc faults caught in 0.45 ms by Koopman forcing signature","feed_subtitle":"Reads current waveforms to flag high-impedance arc faults in 12 kV grids—100% accuracy in simulation.","key_machinery":"The central object is HAVOK (Hankel alternative view of the Koopman), a data-driven method that converts a nonlinear time series into a linear model with an intermittent forcing term. Its machinery: build a Hankel matrix of 40 time-shifted current samples, take its singular value decomposition, keep the leading $r$ eigen-time-delay coordinates, and fit $\\frac{d}{dt}v(t)=A v(t)+B v_r(t)$ to the first $r-1$ coordinates, where the leftover coordinate $v_r$ is the forcing operator. The forcing operator carries the nonlinear signature of the arc; bursts that push it into the arc threshold band are the detection signal, and the threshold bands are what separate arc faults, non-arcing disturbances, and other faults.","core_discovery":"The central claim is that the forcing operator of HAVOK distinguishes arcing from non-arcing events in a simulated 12 kV, 50 Hz distribution system. On the paper's own terms: by stacking time-shifted current samples into a Hankel matrix, applying SVD, and fitting a linear model to the leading eigen-time-delay coordinates, the last coordinate acts as an intermittent forcing term; arc faults appear as forcing bursts whose magnitude lies between ±0.06 and ±0.18, while load switching stays below ±0.045 and line-to-ground faults exceed ±0.2. The paper reports detection latencies from 0.05 ms (high-current arc) to 0.65 ms (arc with noise) across grounding surfaces and load types, with the headline low-current case detected at 0.45 ms, 99.36% faster than the 70 ms harmonic-randomness and waveform-distortion baseline, and 100% accuracy with 0% false positives over 50 simulations per scenario.","pith_inferences":["An editorial extension: the reported 0.45 ms latency implies a decision from roughly 9 post-fault samples at 20 kHz, yet the paper's own embedding rule calls for a window of at least one 20 ms cycle; a causal, streaming implementation would need to show whether the forcing signature can be estimated reliably before the full embedding window has elapsed.","An editorial extension: the threshold bands (±0.06 to ±0.18 for arcs) are calibrated on simulated cases, so transferring to real feeders would likely require per-network calibration; one testable path is to derive bands from a healthy-signal baseline and then measure separation under staged faults.","An editorial extension: because HAVOK bursts flag any strong intermittent nonlinearity, the method may also detect incipient cable faults, transformer inrush, or other transient events; the paper's separation of arc from non-arc depends on the chosen bands holding across those event classes."],"forward_implications":["A protection relay equipped with a HAVOK forcing-operator threshold could, in principle, trip within a fraction of a power cycle, far inside the tens-of-milliseconds windows of conventional harmonic and wavelet methods.","Because the fault signature is a dynamical quantity rather than a trained classifier's decision boundary, the method offers a physical explanation for each detection: the arc is an intermittent forcing event in an otherwise linear current response.","The same forcing-magnitude bands allow the method to reject load switching and line-to-ground faults without retraining, provided the bands hold for the network in question.","Operation under white Gaussian noise down to 60 dB SNR suggests the approach can tolerate realistic measurement noise, with detection delayed but still reliable."],"supporting_citations":[{"why":"Original HAVOK method that supplies the forcing-operator construction used for detection.","marker":"[11]"},{"why":"Baseline harmonic-randomness and waveform-distortion detector whose 70 ms detection time the paper claims to beat by 99.36%.","marker":"[37]"},{"why":"Distortion-controllable arc model that generates the PSCAD fault waveforms through DURATION, EXTENT, and OFFSET parameters.","marker":"[24]"},{"why":"Noise-robust Koopman spectral analysis that provides the Hankel-matrix delay-embedding and SVD pipeline with rank thresholding.","marker":"[32]"},{"why":"Supplies the high-impedance fault circuit structure with arc resistance in series with grounding resistance.","marker":"[7]"},{"why":"Long-arc modeling reference that defines the low-current arcing fault range used in Case A.","marker":"[34]"},{"why":"Evaluation of LV and MV arc parameters that defines the high-current arcing fault range used in Case B.","marker":"[35]"}],"fun_headline_variants":["Arc faults in 0.45 ms via Koopman forcing signal","HAVOK catches MV arc faults 99% faster than old method","Koopman analysis reveals arc faults in 0.45 ms","Interpretable arc fault detection: 0.45 ms, 100% accuracy","Medium-voltage arc faults detected 0.45 ms with HAVOK"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The headline 0.45 ms detection speed assumes the fault can be read off the first nine post-fault current samples, while the paper's own embedding rule says the method needs a window spanning at least one full 20 ms power cycle; if that window is truly required, the reported latency cannot be produced by the method as described.","fun_headline_variants_meta":{"raw":{"variants":["Arc faults in 0.45 ms via Koopman forcing signal","HAVOK catches MV arc faults 99% faster than old method","Koopman analysis reveals arc faults in 0.45 ms","Interpretable arc fault detection: 0.45 ms, 100% accuracy","Medium-voltage arc faults detected 0.45 ms with HAVOK"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000586,"raw_usage":{"total_tokens":2752,"prompt_tokens":943,"completion_tokens":1809,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":559,"completion_tokens_details":{"reasoning_tokens":1726}},"tokens_in":559,"tokens_out":1809,"duration_ms":12627,"temperature":1.0,"reasoning_tokens":1726,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T17:41:19.912178+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the HAVOK pipeline on the same 20 kHz current traces with a causal constraint: the Hankel matrix may contain only samples up to the candidate detection time, and the embedding window must span at least one 50 Hz cycle (20 ms) as the paper prescribes. If the forcing burst first crosses the arc threshold at 0.45 ms under that constraint the latency claim is confirmed; if the earliest compliant detection occurs only after the window has filled, the headline latency is not supported.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Baseline harmonic-randomness and waveform-distortion detector whose 70 ms detection time the paper claims to beat by 99.36%."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Distortion-controllable arc model that generates the PSCAD fault waveforms through DURATION, EXTENT, and OFFSET parameters."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Noise-robust Koopman spectral analysis that provides the Hankel-matrix delay-embedding and SVD pipeline with rank thresholding."},{"cited_title":"Zhang, Y","cited_arxiv_id":null,"evidence_quote":"Supplies the high-impedance fault circuit structure with arc resistance in series with grounding resistance."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Long-arc modeling reference that defines the low-current arcing fault range used in Case A."},{"cited_title":"Gaudreau, B","cited_arxiv_id":null,"evidence_quote":"Evaluation of LV and MV arc parameters that defines the high-current arcing fault range used in Case B."}],"review_version":1}