{"id":"f795234e-7805-4a86-99df-a243e858c50a","arxiv_id":"2607.25037","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"CNN-to-SNN conversion on sparse current-trajectory matrices yields estimated 11 µJ per OC-fault diagnosis (382× below a GPU CNN) at 100% accuracy on a lab inverter.","lead":"A CNN converted to a spiking network diagnoses open-circuit faults in three-phase inverters at an estimated 11 µJ per inference—hundreds of times less energy than a GPU CNN—while keeping full accuracy. That could finally fit data-driven diagnosis inside the sub-watt power budget of real converter controllers.","discovery_kind":"new_application","skeptic_critique":{"model":"moonshotai/kimi-k3","headline":"The 382× figure divides a modeled, partial-system Loihi estimate by a fully-measured GPU baseline, and the paper's own energy decomposition shows the claimed sparsity mechanism accounts for ~1% of the numerator — the \"11 µJ via event-driven sparsity\" claim is therefore doubly unverified.","rationale":"The reader's weakest_assumption identified the same core concern — modeled Loihi energy versus profiled GPU energy as an unfair proxy — so I agree, and I sharpen it with the internal inconsistency on mechanism attribution, which is checkable from the manuscript alone. Credit where due: the authors are unusually transparent about the limitation (explicit disclaimer in §4.4, E_neuron dominance acknowledged in §2.3 and Discussion, offline workflow stated, and they cite the board-level benchmarking methodology of [28]), and the accuracy/robustness core — 100% on 4,400 real-inverter samples, graceful degradation to 80% at 10 dB SNR, step-change handling — is not contested here. The concern targets only the headline energy framing, not the diagnostic validity. The reader's CONDITIONAL verdict is correctly calibrated: the paper is a useful applied contribution whose central quantitative claim awaits board-level validation, and whose mechanism narrative should be revised regardless of what the measurement shows, since E_neuron dominance is asserted in the text itself. No manufactured objection; both prongs trace to specific sections and tables. Verdict unchanged: CONDITIONAL, with the condition being exactly the board-level, full-system energy measurement the authors themselves promise as future work.","tokens_in":12908,"tokens_out":3854,"duration_ms":95261,"concrete_test":"Deploy the converted SNN on physical Loihi hardware (e.g., a Nahuku board) and measure energy per diagnosis with board-level power probes — the methodology of Ostrau et al. [28], which the paper already cites for its GPU procedure — at the intended 5 kHz sliding-window diagnosis rate, including static chip power and host-to-chip communication. In the same session, run one dense-input control (inverted trajectory matrix, ~94% nonzero elements). If board-measured full-system energy is ≈11 µJ and the dense control raises energy substantially, the claim stands. If full-system energy is ≥2× the estimate, or the dense control changes energy by <5%, the headline must be restated as a modeled, core-only figure and the sparsity mechanism dropped.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim asserts 11 µJ per diagnosis — 382× below a GPU CNN — \"by event-driven processing of sparse current-vector trajectory matrices.\" Two load-bearing weaknesses, both visible from the text itself.\n\n(1) Asymmetric accounting in the ratio. The numerator is spike counts × 23.6 pJ from a 2018 characterization paper [21] plus modeled neuron-update cost via NengoLoihi, with weights/data pre-loaded and host communication excluded (§4.4: \"estimated neuromorphic inference energy rather than full-system energy\"). The denominator is nvidia-smi full-GPU power (73–79 W, only 3.6 W idle subtracted) on a laptop RTX 4060 running tiny single-sample inferences of a 3-layer CNN — close to a least-favorable embedded baseline. Loihi static power (~tens of mW) and x86 host I/O are absent from the numerator; at the paper's own 5 kHz sliding-window diagnosis rate, static power alone would add several µJ per diagnosis. The 382× is a real ratio of two numbers, but neither number is the decision-relevant quantity for the stated sub-watt embedded application.\n\n(2) The paper's own accounting contradicts the causal attribution. §2.3 and §2.4 state E_neuron is ~160× E_synaptic and \"remains independent of input conditions\"; Table 2 shows identical 1.1×10⁻⁵ J under balanced and unbalanced loading. Since membrane updates occur every simulated timestep for all neurons regardless of spikes, ~99% of the estimated energy is sparsity-independent. The introductory and Discussion narrative — energy saved \"by exploiting the sparse structure of trajectory matrices\" — is thus unsupported by the paper's own numbers: the reduction versus GPU is a platform effect (neuromorphic vs. clock-driven GPU), not the proposed sparse event-driven mechanism. This also undercuts the argument for preferring SNNs over sparse convolution [14,15], which the introduction dismisses on data-movement grounds.","agreement_with_reader":"agree"},"referee_report":{"model":"moonshotai/kimi-k3","summary":"The manuscript proposes an event-driven neuromorphic pipeline for open-circuit (OC) fault diagnosis in three-phase inverters. A CNN trained on binary current-vector trajectory matrices (112×112, ~6% nonzero) is converted to a spiking neural network under explicit compatibility constraints (zero bias, ReLU, no pooling) and evaluated offline in NengoLoihi. On laboratory data covering 22 operating modes, the authors report 100% diagnostic accuracy on 4,400 balanced-load test samples and under 88% PCUR unbalanced loading, robustness to current-amplitude steps and injected noise down to 20 dB SNR, and an estimated energy of 11 µJ per diagnosis, presented as a 382× reduction over a GPU-based CNN (laptop RTX 4060, nvidia-smi profiling). The energy figure is model-based: recorded spike counts multiplied by a 23.6 pJ/spike constant from Davies et al. [21], plus modeled neuron-update costs, with weights/data pre-loaded and host communication excluded.","tokens_in":13359,"tokens_out":4068,"duration_ms":148800,"significance":"If the claims are properly scoped, this is a useful contribution: to my knowledge it is the first SNN-based OC fault-diagnosis framework for three-phase inverters with a neuromorphic energy model, and the experimental campaign is unusually broad for this subfield (22 modes on real hardware, extreme unbalance, step transients, SNR sweeps). The conversion constraints are stated explicitly, the firing-rate/accuracy/energy trade-off is characterized (Fig. 3c), and the authors are commendably transparent in §4.4 and the Discussion that the Loihi number is an estimate, not a board-level measurement. The falsifiable, quantitative comparison against eight GPU baselines is also a strength. However, the headline 11 µJ / 382× claim currently rests on asymmetric accounting and a causal attribution (sparsity-driven event-driven savings) that the paper's own energy decomposition contradicts; the significance therefore hinges on reframing rather than on new experiments beyond the authors' stated plans.","major_comments":[{"comment":"The central causal claim — that the energy reduction is 'mainly enabled by matching the sparse structure of current-vector trajectory matrices with event-driven computation' (Discussion; echoed in the Abstract) — is contradicted by the paper's own accounting. §2.3 states E_neuron is ~160× E_synaptic and 'largely independent of input sparsity'; §2.4 states E_neuron 'remains independent of input conditions'; and Tables 1 and 2 report an identical 1.1×10⁻⁵ J under balanced and unbalanced loading. Since Loihi updates every neuron's membrane state at every simulated timestep regardless of spikes, ~99% of the estimated energy is sparsity-independent and not event-driven in any meaningful sense. The measured advantage therefore comes from Loihi's low per-update cost, not from exploiting trajectory sparsity. The authors should either (a) reframe the mechanism claim accordingly, or (b) provide di","section":"§2.3–2.4, Eq. (1), Tables 1–2; Abstract; Discussion"},{"comment":"The 382× ratio divides two non-comparable quantities. The numerator excludes Loihi static power (tens of mW), the x86 host, and all I/O, with weights and data pre-loaded (§4.4: 'estimated neuromorphic inference energy rather than full-system energy'); at the stated 5 kHz sliding-window diagnosis rate, Loihi static power alone would add several µJ per diagnosis. The denominator is full-GPU nvidia-smi power (73–79 W, only 3.6 W idle subtracted) on a laptop RTX 4060 performing single-sample inference of a small 3-layer CNN — close to a least-favorable baseline, and the paper's own Introduction cites accelerator results of 7.94–337 mJ [10,11] and model-based diagnosis at 0.12 mJ [13], against which the 11 µJ advantage shrinks to roughly one order of magnitude. The 382× should either be removed from the Abstract or explicitly re-scoped ('vs. full-GPU power on a laptop RTX 4060, model-based Lo","section":"§4.4, Table 1, Abstract (382× claim)"},{"comment":"The Loihi energy model is described inconsistently: NengoLoihi's energy.summary is said to account for 'dynamic power consumption of the neuromorphic cores during spike processing,' but the next paragraph states energy is computed as recorded spike counts × 23.6 pJ from [21]. The 23.6 pJ figure in Davies et al. is workload-dependent, and the dominant E_neuron term must come from a separate per-neuron-update constant that is never given. Please state exactly which energy terms are included (synaptic events, neuron updates, per-op constants and their sources), whether they are NengoLoihi defaults or values from [21], and report the spike counts and simulated timesteps per inference so the 11 µJ figure is independently reproducible.","section":"§4.4 (energy estimation paragraph)"},{"comment":"The train/test split protocol is not specified. With one-fundamental-cycle sliding windows advancing by a single sample (§2.5), adjacent trajectory matrices are highly correlated; if the 22,000 matrices per loading condition were split randomly at the cycle level, near-duplicates of test trajectories appear in training, inflating the reported 100% accuracy. Please describe the split (per-cycle random, contiguous blocks, or session-held-out) and, if it was random, add a session- or day-held-out evaluation. This is load-bearing for the '100% accuracy' half of the central claim.","section":"§4.4 (dataset construction); §2.3"},{"comment":"No detection-latency figure is given, which is a key metric for fault diagnosis. Two design choices contribute delay: transient windows are labeled healthy, deferring decisions until steady-state trajectories fill the window (§4.4), and Loihi timesteps are simulated rather than physical time (Discussion). At minimum, report the number of simulated timesteps per inference, the wall-clock mapping NengoLoihi assumes, and the diagnosis delay in fundamental cycles after fault onset, so readers can judge the latency/energy trade-off.","section":"§2.5, §4.4, Discussion"}],"minor_comments":[{"comment":"Header typo: 'T raining Energy'. Also, the '20,000× reduction relative to GPU-based SNN simulators' is a strawman comparison — GPU simulation of spiking dynamics is known to be inefficient and the MB baseline [23] only reaches 78.32% accuracy; consider softening or contextualizing this number.","section":"Table 1"},{"comment":"Column alignment appears broken in the FPGA row ('Medium Medium – High Low High' for four columns). Please reformat.","section":"Table 3"},{"comment":"The floor notation applied to the vector I_{α,β}(t_i) is ambiguous; please clarify that the row/column indices are computed componentwise and define the normalization (the '/4' factor) and its relation to rated current.","section":"§4.1, Eq. (2)"},{"comment":"The explanation of the apparent THD increase (8-bit transmission quantization) is plausible but unsupported; a sentence quantifying the quantization step relative to the harmonic amplitudes, or a reference, would help.","section":"§2.5"},{"comment":"The accuracy plateau near 80% for 200–300 Hz followed by 100% at 330 Hz is a sharp transition; please add axis units and indicate the selected operating point on the figure, and comment on sensitivity of the result to the ±10% choice of scaling factor.","section":"§2.3, Fig. 3c"},{"comment":"'Available upon reasonable request' is weak for a paper whose central claim is a quantitative energy estimate; depositing the conversion scripts, spike-count logs, and nvidia-smi traces in a public repository would substantially strengthen reproducibility.","section":"Data/Code availability"},{"comment":"The 'first literature' claim (§1) should be tempered or supported with a sentence distinguishing this work from event-driven power-system applications [16–20]; also reference [22] is the authors' own prior pipeline — the dependence of the present results on that architecture choice deserves one sentence.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The experimental accuracy work is solid and the topic fits the journal, but the Abstract's headline numbers (11 µJ, 382×) currently oversell what the manuscript's own §2.3–2.4 decomposition and §4.4 disclaimers support. All requested changes are reframing, additional disclosure, and modest additional analysis — no new hardware is strictly required — so major revision rather than rejection seems right. The 'first SNN-based OC fault diagnosis' novelty claim is plausible but rests partly on the authors' own prior CNN pipeline [22]; the editor may wish to check overlap with that companion paper."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful core here is experimental, not architectural novelty. They take their own trajectory-matrix CNN, convert it with the usual Diehl-style constraints (zero bias, ReLU, no pooling), and show 100% accuracy on 22 OC modes from a physical three-phase inverter—balanced, 88% PCUR unbalanced, amplitude steps, and noise down to ~20 dB. That package is real lab work and is what the paper actually adds relative to the cited CNN baseline and generic CNN-to-SNN literature.\n\nWhat it does well: clear conversion rules, explicit architecture, held-out trajectory tests, and honest discussion that Loihi energy is estimated (spike count × 23.6 pJ plus neuron updates) rather than board-measured, and that the workflow is offline with pre-loaded weights. Tables 1–2 and the SNR/step figures are readable and internally consistent. Citations are appropriate; self-use of [22] is normal given that it supplies the encoding and CNN.\n\nThe soft spot is the load-bearing energy claim, and it is visible from their own text. Numerator is modeled neuromorphic inference energy excluding host I/O and static power; denominator is nvidia-smi on a laptop 4060 with only idle subtracted. That ratio is 382×, but it is not a fair full-system or sub-watt embedded comparison. More importantly, they report E_neuron ~160× E_synaptic and identical 11 µJ under balanced and unbalanced loads, with membrane updates every timestep independent of spikes. So ~99% of the estimated energy does not come from the sparse event-driven mechanism the abstract and intro advertise. The reduction is mostly “Loihi model vs GPU,” not “sparsity saved the MACs.” That also weakens the brush-off of sparse convolution. Minor relative to that: firing-rate scale is tuned to hit the CNN ceiling; no public code/data.\n\nWho it is for: power-electronics people who care about data-driven OC diagnosis under tight auxiliary budgets, and anyone tracking whether SNNs help sparse sensing on converters. Not a methods breakthrough in neuromorphic computing.\n\nI would send it to referees. Ask them to force a rewrite of the causal energy story, demand board-level or at least static-inclusive accounting, and keep the accuracy results—which are the part that holds up. Worth engaging if you work this application; cite the accuracy/robustness package, not the 382× line, until someone measures a chip.","headline":"Solid lab accuracy on a real inverter; the 382× energy headline is a platform comparison under asymmetric accounting, and the paper’s own numbers show sparsity is not what drives most of the estimated energy.","tokens_in":14460,"tokens_out":608,"would_cite":false,"duration_ms":11965,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"A spiking network diagnoses inverter open-circuit faults at 11 µJ each—382× less energy than a GPU CNN—while keeping 100% accuracy.","keywords":["neuromorphic computing","spiking neural network","three-phase inverter","open-circuit fault","current-vector trajectory","CNN-to-SNN conversion","embedded fault diagnosis","energy-efficient inference"],"falsifier":"Measure wall-plug or chip power on actual Loihi (or an equivalent neuromorphic/FPGA board) while streaming live inverter currents through the same SNN and check whether energy per diagnosis stays near 11 µJ at 100% accuracy.","tokens_in":14186,"feed_emoji":"⚡","tokens_out":827,"duration_ms":15951,"temperature":0.7,"pith_summary":"Converter controllers have only a sub-watt auxiliary budget, yet CNN fault diagnosers burn millijoules to joules per inference because they dense-convolve sparse current-trajectory images. This paper trains a CNN on binary α–β trajectory matrices of a three-phase inverter, converts it to a spiking network, and runs the SNN under a Loihi-style energy model. Because only the few nonzero trajectory pixels generate spikes, most of the matrix never triggers computation. On a laboratory inverter the method reports 11 µJ per diagnosis and 100% accuracy across 22 healthy and open-circuit modes, including extreme unbalanced load, current steps, and added measurement noise. If the energy numbers hold on real neuromorphic or FPGA hardware, data-driven diagnosis becomes compatible with the power envelope of ordinary converter controllers.","feed_headline":"Inverter faults diagnosed at 11 µJ—382× less energy","feed_subtitle":"Spiking net keeps 100% accuracy on sparse current trajectories inside a converter's power budget","key_machinery":"CNN-to-SNN conversion of sparse current-vector trajectory matrices: nonzero pixels alone drive integrate-and-fire spikes, so computation is confined to the informative trajectory instead of the full dense feature map.","core_discovery":"Converting a trajectory-matrix CNN into an event-driven spiking network and estimating energy on a Loihi model yields 11 µJ per open-circuit diagnosis—a 382× reduction versus the same CNN on a laptop GPU—while retaining 100% accuracy on thousands of balanced and unbalanced test cases.","pith_inferences":["The same sparse-trajectory encoding could extend to short-circuit or sensor-fault classes if their α–β signatures remain geometrically distinct.","If neuron-update energy were made fully event-driven, the remaining gap to purely spike-based theoretical energy would shrink by roughly two orders of magnitude.","A sliding-window real-time port on an FPGA controller would let detection latency be measured in physical milliseconds rather than Loihi time steps."],"forward_implications":["Data-driven open-circuit diagnosis can fit inside the ~0.9 W auxiliary budget of typical converter controllers.","Event-driven SNNs become a practical route for any diagnosis task whose features are naturally sparse images or trajectories.","FPGA implementations that keep weights and neuron state in distributed on-chip RAM can emulate the same compute-in-memory advantage without specialized neuromorphic silicon.","Energy per inference stays nearly constant across balanced and unbalanced loads because neuron-update cost dominates and is architecture-fixed."],"fun_headline_variants":["11 µJ open-circuit diagnosis via event-driven SNN","SNN cuts inverter fault energy 382× vs GPU CNN","Sparse trajectory SNN: 100% OC accuracy at 11 µJ","Loihi-estimated SNN diagnoses inverter faults at 11 µJ","Event-driven neuromorphic OC diagnosis stays in sub-watt budget"],"cache_read_input_tokens":128,"weakest_assumption_plain":"That multiplying recorded spike counts by a published energy-per-spike figure plus modeled neuron updates is a fair stand-in for real embedded inference energy, even though no board-level Loihi power was measured and the workflow is offline with pre-loaded data.","fun_headline_variants_meta":{"raw":{"variants":["11 µJ open-circuit diagnosis via event-driven SNN","SNN cuts inverter fault energy 382× vs GPU CNN","Sparse trajectory SNN: 100% OC accuracy at 11 µJ","Loihi-estimated SNN diagnoses inverter faults at 11 µJ","Event-driven neuromorphic OC diagnosis stays in sub-watt budget"]},"model":"grok-4.5","effort":"low","cost_usd":0.001788,"raw_usage":{"total_tokens":811,"prompt_tokens":715,"num_sources_used":0,"completion_tokens":76,"cost_in_usd_ticks":17884000,"prompt_tokens_details":{"text_tokens":715,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":20,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":715,"tokens_out":76,"duration_ms":2452,"temperature":1.0,"reasoning_tokens":20,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-31T02:47:21.145646+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Measure wall-plug or chip power on actual Loihi (or an equivalent neuromorphic/FPGA board) while streaming live inverter currents through the same SNN and check whether energy per diagnosis stays near 11 µJ at 100% accuracy.","supporting_citations":[],"review_version":1}