Pith. sign in

REVIEW 5 major objections 7 minor 28 references

Ultra-Low-Energy Open-Circuit Fault Diagnosis for Three-Phase Inverters

T0 review · 5 major / 7 minor · reviewed 2026-07-31 · grok-4.5

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A spiking network diagnoses inverter open-circuit faults at 11 µJ each—382× less energy than a GPU CNN—while keeping 100% accuracy.

desk verdict 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. read the letter →

arxiv 2607.25037 v1 pith:QMU6OSZW submitted 2026-07-27 eess.SY cs.SY

Xiaoyi Lei , Fanfu Wu , Yunting Liu This is my paper · ORCID
classification eess.SYcs.SY
keywords neuromorphiccomputingspikingneuralnetworkthree-phaseinverteropen-circuitfaultcurrent-vectortrajectoryCNN-to-SNNconversionembeddeddiagnosisenergy-efficientinference
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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.

What would settle it

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.

Watch

Extended reading notes

Core claim

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.

Load-bearing premise

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.

Editorial extensions

If this is right

  • 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.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • 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.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 7 minor

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.

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 (5)
  1. [§2.3–2.4, Eq. (1), Tables 1–2; Abstract; Discussion] 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
  2. [§4.4, Table 1, Abstract (382× claim)] 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
  3. [§4.4 (energy estimation paragraph)] 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.
  4. [§4.4 (dataset construction); §2.3] 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.
  5. [§2.5, §4.4, Discussion] 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.
minor comments (7)
  1. [Table 1] 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.
  2. [Table 3] Column alignment appears broken in the FPGA row ('Medium Medium – High Low High' for four columns). Please reformat.
  3. [§4.1, Eq. (2)] 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.
  4. [§2.5] 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.
  5. [§2.3, Fig. 3c] 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.
  6. [Data/Code availability] '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.
  7. [References] 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.

Circularity Check

1 steps flagged · score 1.0 of 10

No derivation-by-construction circularity; energy and accuracy claims are measured (or externally modeled), not forced by fitted inputs or self-definition.

  1. self citation load bearing [§2.2 Fig. 2a; §4.2–4.4; Table 7; refs to [22]]
    "The CNN-based OC fault diagnosis framework is set as the baseline. ... Both SNN and CNN adopt the same architecture utilized in [22]. ... trajectory matrices following [22] and (2)."

    The converted network, trajectory-matrix encoding, and GPU CNN comparator are taken from the authors’ own prior work [22]. This is ordinary baseline reuse, not a circular derivation: accuracy and energy of the SNN are still measured independently and are not implied by [22] alone. Flagged only as minor self-citation scaffolding, not as a step that forces the central 11 µJ / 382× / 100% claims.

full rationale

The paper’s load-bearing results are empirical: diagnostic accuracy on held-out lab trajectory matrices (balanced/unbalanced, step changes, SNR sweeps) and inference energy from either nvidia-smi (GPU baselines) or NengoLoihi spike counts × a literature Loihi energy-per-spike model. Nothing in the chain equates a claimed prediction to a quantity already fixed by definition or by a fit to the same target. CNN→SNN conversion follows standard rate-based constraints (zero bias, ReLU→spikes, no pooling); the firing-rate scale is a calibrated hyperparameter chosen so SNN accuracy can reach the trained CNN ceiling, which is ordinary experimental practice rather than a fitted-input-called-prediction. Self-citation of the authors’ prior CNN/trajectory pipeline [22] supplies the baseline architecture and encoding but does not underwrite a uniqueness claim or force the 11 µJ / 382× figures. Asymmetric energy accounting and the dominance of sparsity-independent E_neuron are methodological/attribution concerns, not circular reductions. Score 1 only for minor non-load-bearing dependence on [22] as the converted baseline.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The claim rests on standard power-electronics and SNN conversion practice plus a few tuned knobs and an external chip energy model. No new physical entities. Load-bearing modeling choices are the binary trajectory encoding, conversion constraints, firing-rate scale, and the Loihi per-spike/neuron energy model used in place of board measurement.

free parameters (4)
  • firing-rate scaling factor = 330 Hz
    Hand/accuracy-selected scale controlling spike rate vs energy; fixed at 330 Hz after accuracy–energy sweep (Fig. 3c, Table 7). Central 100% accuracy and reported µJ both depend on this choice.
  • trajectory matrix resolution D = 112
    112×112 binary grid sets neuron count and thus dominant E_neuron; chosen following prior CNN setup, not derived.
  • CNN training hyperparameters = lr=3e-4, epochs=10, filters=32-64-128
    Learning rate 3e-4, batch 16, 10 epochs, filter counts 32/64/128 (Table 7) determine the weights transferred to the SNN and thus accuracy ceiling.
  • Loihi energy-per-spike constant = 23.6 pJ/spike
    23.6 pJ/spike from Davies et al. [21] multiplies spike counts to produce the headline 11 µJ; not re-measured here.
assumptions (5)
  • domain assumption Binary α-β trajectory matrices from Clarke-transformed currents uniquely encode the 22 single/double switch OC modes for classification.
    §4.1 and Eq. (2); inherited from [22]. Accuracy claims assume these geometries remain separable under the tested loads and noise.
  • domain assumption CNN with zero biases, ReLU only, and no pooling converts to a rate-coded SNN whose outputs match CNN class decisions at sufficient firing rate.
    §4.3 conversion constraints following Diehl-style rate learning [25].
  • ad hoc to paper NengoLoihi spike-count × characterized Loihi synaptic/neuron energies equals relevant neuromorphic inference energy for comparison to GPU nvidia-smi energy.
    §4.4 explicitly notes no on-chip power monitor and excludes communication energy; this comparability axiom carries the 382× claim.
  • ad hoc to paper Transient windows during mode/amplitude changes may be labeled healthy so diagnosis waits for steady-state trajectories.
    §4.4 training protocol; affects how step-change robustness in Fig. 4 is interpreted.
  • standard math Integrate-and-fire / Loihi spiking ReLU dynamics with default threshold adequately implement the converted network.
    Eqs. (5)–(6) and NengoLoihi default neuron model (§2.3).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Ultra-Low-Energy Open-Circuit Fault Diagnosis for Three-Phase Inverters." pith.science (2026). https://pith.science/paper/QMU6OSZW

@misc{pith2026260725037,
  author       = {Pith},
  title        = {Pith review of: Ultra-Low-Energy Open-Circuit Fault Diagnosis for Three-Phase Inverters},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QMU6OSZW}},
  note         = {Machine review of arXiv:2607.25037}
}
read the original abstract

Embedded fault diagnosis in three-phase inverters must satisfy the sub-watt power budget of converter control hardware, but conventional convolutional neural network (CNN)-based methods require dense multiply-accumulate operations and impose substantial inference energy. This work proposes an event-driven neuromorphic framework for energy-efficient open-circuit (OC) fault diagnosis. A CNN trained on current-vector trajectory matrices is converted into a spiking neural network (SNN) and evaluated using the NengoLoihi framework with Loihi-based neuromorphic energy estimation. By exploiting the sparse structure of trajectory matrices, the SNN activates computation only in informative regions instead of processing the full feature map densely. Experiments on a three-phase inverter platform show that the proposed method achieves 11 microjoules per diagnosis, corresponding to a 382 times inference-energy reduction compared with a GPU-based CNN, while maintaining 100% diagnostic accuracy. Robustness is further validated under unbalanced loading, current amplitude step changes, and injected measurement noise.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

28 extracted references · 2 linked inside Pith

  1. [22]

    & Liu, Y

    Lei, X., Wu, F. & Liu, Y. An online convolutional neural network-based method for open-circuit fault diagnosis in three-phase inverters under extremely unbalanced loading condition.IEEE Trans. Power Electron.1–14 (2026)

    2026

  2. [21]

    Davies, M.et al.Loihi: A neuromorphic manycore processor with on-chip learning.IEEE Micro38, 82–99 (2018)

    2018

  3. [13]

    & Fuchs, F

    Rothenhagen, K. & Fuchs, F. W. Performance of diagnosis methods for IGBT open-circuit faults in three-phase voltage source inverters for ac variable speed drives. Proc. Eur. Conf. Power Electron. Appl. (EPE) (2005). Pp. 10–pp

    2005

  4. [1]

    & Gou, B

    Xia, Y., Xu, Y. & Gou, B. A data-driven method for IGBT open-circuit fault diagnosis based on hybrid ensemble learning and sliding-window classification. IEEE Trans. Ind. Informat.16, 5223–5233 (2019)

    2019

  5. [2]

    Xia, Y. & Xu, Y. A transferrable data-driven method for IGBT open-circuit fault diagnosis in three-phase inverters.IEEE Trans. Power Electron.36, 13478–13488 (2021)

    2021

  6. [3]

    & Xie, M

    Cai, B., Zhao, Y., Liu, H. & Xie, M. A data-driven fault diagnosis methodology in three-phase inverters for PMSM drive systems.IEEE Trans. Power Electron. 32, 5590–5600 (2016)

    2016

  7. [4]

    Gou, B., Xu, Y., Xia, Y., Deng, Q. & Ge, X. An online data-driven method for simultaneous diagnosis of IGBT and current sensor fault of three-phase PWM inverter in induction motor drives.IEEE Trans. Power Electron.35, 13281–13294 (2020). 16

    2020

  8. [5]

    & Chen, Y

    Wang, H., Zhang, C., Zhang, N., Chen, Y. & Chen, Y. Fault diagnosis for IGBTs open-circuit faults in high-speed trains based on convolutional neural network. Proc. Prognostics Syst. Health Manag. Conf. (PHM) (2019). Pp. 1–8

    2019

Show all 28 references
  1. [6]

    & Huang, Y

    Hang, J., Shu, X., Ding, S. & Huang, Y. Robust open-circuit fault diagnosis for PMSM drives using wavelet convolutional neural network with small samples of normalized current vector trajectory graph.IEEE Trans. Ind. Electron.70, 7653–7663 (2023)

    2023

  2. [7]

    IEEE Trans

    Luo, W.et al.Enhanced 1D convolutional neural network based open-circuit fault diagnosis and hybrid fault-tolerant control for three-level NPC converters. IEEE Trans. Instrum. Meas.(2025)

    2025

  3. [8]

    Wang, H.et al.Global feature perception network for open-switch fault diagnosis in grid-connected photovoltaic inverters under imbalanced data conditions.IEEE Trans. Instrum. Meas.(2025)

    2025

  4. [9]

    POWEREST: Power estimation tool

    Texas Instruments. POWEREST: Power estimation tool. https://www.ti.com/ tool/POWEREST (2025). Accessed: Mar. 4, 2026

    2025

  5. [10]

    Chen, Y.-H., Krishna, T., Emer, J. S. & Sze, V. Eyeriss: An energy-efficient reconfigurable accelerator for deep convolutional neural networks.IEEE J. Solid- State Circuits52, 127–138 (2016)

    2016

  6. [11]

    & Sze, V

    Chen, Y.-H., Yang, T.-J., Emer, J. & Sze, V. Eyeriss v2: A flexible accelerator for emerging deep neural networks on mobile devices.IEEE J. Emerg. Sel. Topics Circuits Syst.9, 292–308 (2019)

    2019

  7. [12]

    Hadidi, R.et al.Characterizing the deployment of deep neural networks on commercial edge devices. Proc. IEEE Int. Symp. Workload Characterization (IISWC) (2019). Pp. 35–48

    2019

  8. [14]

    Spatially-sparse convolutional neural networks.arXiv preprint arXiv:1409.6070(2014)

    Graham, B. Spatially-sparse convolutional neural networks.arXiv preprint arXiv:1409.6070(2014)

    2014 arXiv

  9. [15]

    & van der Maaten, L

    Graham, B. & van der Maaten, L. Submanifold sparse convolutional networks. arXiv preprint arXiv:1706.01307(2017)

    2017 arXiv

  10. [16]

    & Chaudhuri, N

    Mahapatra, K., Lu, S., Sengupta, A. & Chaudhuri, N. R. Power system dis- turbance classification with online event-driven neuromorphic computing.IEEE Trans. Smart Grid12, 2343–2354 (2021). 17

    2021

  11. [17]

    & Wang, D

    Yang, X. & Wang, D. Reinforcement learning for dynamic event-driven control of multi-machine power systems.IEEE Trans. Circuits Syst. II, Exp. Briefs72, 1413–1417 (2025)

    2025

  12. [18]

    Power Electron.40, 563–573 (2025)

    Liu, X.et al.Combining data-driven and event-driven for online learning pre- dictive control in power converters.IEEE Trans. Power Electron.40, 563–573 (2025)

    2025

  13. [19]

    Power Electron.40, 4914–4926 (2025)

    Liu, X.et al.Event-driven based reinforcement learning predictive controller design for three-phase NPC converters using online approximators.IEEE Trans. Power Electron.40, 4914–4926 (2025)

    2025

  14. [20]

    Liu, X., Fang, L. Q. a. Y., Wang, K., Li, Y. & Rodr ´ ıguez, J. FCS-MPC of power converters: A data-driven model-free reinforcement learning solution.IEEE Trans. Power Electron.41, 8438–8450 (2026)

    2026

  15. [23]

    & Strauss, R

    Arena, P., Cal ` ı, M., Patan´ e, L., Portera, A. & Strauss, R. A mushroom bodies inspired spiking network for classification and sequence learning. Proc. Int. Joint Conf. Neural Networks (IJCNN) (2015). Pp. 1–8

    2015

  16. [24]

    Ottati, F.Efficient Deep Learning Inference: A Digital Hardware Perspective— Evaluating and Improving Performance and Efficiency of Artificial and Spiking Neural Networks Hardware Accelerators. Ph.D. thesis, University of Bologna (2024)

    2024

  17. [25]

    U.et al.Fast-classifying, high-accuracy spiking deep networks through weight and threshold balancing

    Diehl, P. U.et al.Fast-classifying, high-accuracy spiking deep networks through weight and threshold balancing. Proc. Int. Joint Conf. Neural Networks (IJCNN) (2015). Pp. 1–8

    2015

  18. [26]

    Power Electron.(2025)

    Lu, F.et al.A novel simultaneous diagnosis method for IGBT open-circuit faults and current sensor faults of three-phase SPWM inverter.IEEE Trans. Power Electron.(2025)

    2025

  19. [27]

    S., Salem, A., Huynh, V

    Mahmoud, M. S., Salem, A., Huynh, V. K. & Robbersmyr, K. G. Robust self- augmented open-circuit fault diagnosis of three-level inverters for EV powertrains. IEEE Trans. Transp. Electrific.11, 10250–10261 (2025)

    2025

  20. [28]

    & R¨ uckert, U

    Ostrau, C., Klarhorst, C., Thies, M. & R¨ uckert, U. Benchmarking neuromorphic hardware and its energy expenditure.Front. Neurosci.16, 873935 (2022). 18

    2022

Pith tools

Reviewed July 31, 2026 · model on record in the stance chip above.