Deep Learning-Based Detection of Electrical Faults and Power Quality Disturbances in Aerospace Power Systems
Ian C. Guzmán, Radu Babiceanu, Berker Peköz
eess.SP · cs.LG · eess.SY
Quick verdict
Reports 96.94% 21-class fault/PQD accuracy and 6.9 ms FPGA latency, but accuracy rests on leakage-uncontrolled simulation of a 400 Hz aircraft power system.
Summary
The paper presents a hardware-aware deep learning framework for classifying electrical faults and power quality disturbances (PQDs) in a 400 Hz More-Electric-Aircraft power system, where most prior work targets 50/60 Hz grids. A high-fidelity simulation inspired by the Boeing 787 architecture (230 V AC bus, 115 V auxiliary bus, 28/270 V DC via TRUs, 250 kVA generator, 33.3 kHz sampling) generates 21 operating classes — normal operation, PQDs, switching transients, and component open/short-circuit faults. Two 73,500-sample datasets (raw 1D waveforms and STFT images) are built using signal-processing augmentation, domain randomization, and class-specific GANs; the 1D dataset is released on IEEE DataPort. Twelve architectures (1D/2D CNNs, LSTMs, hybrids, ResNet, MobileNetV2, VGG) are trained under identical conditions (Adam, LR 1e-3, 20 epochs); a compact 2D ResNet with 175,685 parameters reaches 96.94% test accuracy versus 97.29% for an 8.3M-parameter 1D-CNN. INT8-quantized deployment on a Xilinx ZCU102 retains 95.87% accuracy with 6.90 ms mean (14.40 ms worst-case) accelerator latency, using 15.52% DSP, 63.71% BRAM, and 93.77% CLB LUTs. Main limitations: all signals are simulated with no experimental waveform validation shown; validation is not parent-waveform-disjoint, which the authors concede may inflate accuracy; results are single-run without variance or significance testing; and the 93.77% LUT use leaves full DAQ/STFT integration as future work.
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 2
● low 2
5 strengths · 0 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A 21-class aerospace fault/PQD dataset generated from a Boeing-787-inspired 400 Hz simulation (73,500 samples per representation; 1D version released on IEEE DataPort) enables deep-learning training for aircraft power monitoring.
- A compact 2D ResNet with 175,685 parameters achieves 96.94% test accuracy — second only to an 8.3M-parameter 1D-CNN at 97.29% — giving the best accuracy-complexity tradeoff among 12 architectures.
- INT8 quantization and ZCU102 deployment retain 95.87% accuracy at 6.90 ms mean (14.40 ms worst-case) accelerator latency, using 15.52% DSP, 63.71% BRAM, and 93.77% CLB LUTs.
- Classification degrades gracefully under additive noise (>96% at 10 dB SNR, 71.26% at 0 dB), indicating reliance on stable temporal and spectral structure.
- Lightweight deep learning can deliver accurate multiclass classification under aerospace-relevant constraints — demonstrated at accelerator level only, with full system integration explicitly deferred.
Strengths
Complete simulation-to-hardware validation chain
The paper actually deploys and measures, not just projects: INT8 quantization on a ZCU102 with post-quantized accuracy 95.87%, mean latency 6.90 ms, worst case 14.40 ms, and post-synthesis resources (391/2,520 DSP, 581/912 BRAM, 256,992/274,080 LUTs) — Tables V and VI.
Public 21-class dataset for an underexplored domain
73,500 one-dimensional samples spanning 21 fault/PQD classes at 400 Hz, a regime where public data is scarce; the released dataset enables external benchmarking of 1D models.
Quantitatively supported efficiency argument
Unusually candid limitation analysis
Noise-robustness sweep with graceful degradation
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper targets multiclass detection of electrical faults and power-quality disturbances (PQDs) in 400 Hz More-Electric-Aircraft power networks, a setting Sec. II argues is underexplored relative to 50/60 Hz utility grids. Three contributions are claimed (Sec. I): a comprehensive simulated dataset from a Boeing-787-inspired model, lightweight DL models for real-time fault/PQD classification, and an FPGA implementation under aerospace constraints. The headline numbers are 96.94% software test accuracy over 21 classes with a 175,685-parameter compact 2D ResNet (Table III) and a measured INT8 deployment on a Xilinx ZCU102 at 95.87% accuracy with 6.90 ms mean accelerator latency (Table VI). Claims are mostly well-scoped: the discussion (Sec. IV) explicitly confines the metrics to intra-distribution performance.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 0 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What accuracy does this aircraft fault detection model achieve?
A compact 2D ResNet (175,685 parameters) reaches 96.94% software test accuracy across 21 operating classes on a simulated 400 Hz aircraft power dataset. After INT8 quantization on a Xilinx ZCU102 FPGA, accuracy is 95.87% with a measured mean accelerator latency of 6.90 ms per record.
Is this paper based on real aircraft data?
No. All signals come from a high-fidelity simulation inspired by the Boeing 787 electrical architecture (230 V, 400 Hz AC bus with TRU-fed 28/270 V DC buses). The authors state that independent operating-condition validation remains necessary, and the accuracy may be optimistic because the evaluation is not parent-waveform-disjoint.
Is the dataset from this paper publicly available?
The one-dimensional time-series dataset (73,500 samples, 21 classes) is publicly released on IEEE DataPort. The STFT (time-frequency) dataset used by the deployed 2D model, the GAN generators, and the simulation model are not stated as released.
Can this model run in real time on aircraft hardware?
The FPGA accelerator averages 6.90 ms per input record (worst case 14.40 ms; about 145 records/s), processing STFT images. However, the accelerator alone occupies 93.77% of the chip's logic resources, so full integration of data acquisition, STFT computation, and communication is explicitly left as future work, and no concrete avionics timing deadline is specified.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Four '[...sampling gap...]' regions skip the end of Sec. II, part of the dataset-construction/GAN text (Sec. III-C–IV-A), the per-class metric discussion, and the tail of Sec. V-D through Sec. VI conclusions; judgments about those parts rest only on visible fragments.
- The FPGA latency description cuts off mid-sentence at 'transmitting batches of 1000'; the full measurement protocol (repetitions, percentiles) is not visible.
- References [55]–[74] are not visible, so completeness of related-work coverage cannot be judged.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.