PaperlayerView exact revision on arXiv

Source-linked AI summary

Deep Learning-Based Detection of Electrical Faults and Power Quality Disturbances in Aerospace Power Systems

Ian C. Guzmán, Radu Babiceanu, Berker Peköz

arXiv:2609.10479v1eess.SPcs.LGeess.SY

TL;DR

Aircraft power-quality and fault-diagnosis methods are largely designed for 50–60 Hz grids, leaving high-frequency aerospace systems insufficiently addressed. The paper develops a hardware-aware deep-learning framework using Boeing 787-inspired simulation, waveform augmentation, and model comparison, with a Compact ResNet achieving high accuracy at low complexity and supporting FPGA deployment. The evaluation is simulation-based and requires independent operating-condition validation before broader deployment generalization can be established.

  • Problem

    Aircraft power-system diagnosis lacks specialized evidence for 400 Hz networks because existing methods primarily target conventional 50–60 Hz systems.

  • Method

    The paper combines high-fidelity aircraft-power-system simulation, signal and GAN augmentation, and comparative evaluation of compact deep-learning architectures for FPGA deployment.

  • Results

    The Compact ResNet reaches 96.94% test accuracy with 175,685 parameters, while retaining 96% validation accuracy near epoch 13 and 71.26% accuracy at 0 dB SNR.

  • Takeaways & Limitations

    The findings support simulation-based, accelerator-level feasibility of compact deep-learning models for embedded aircraft electrical health monitoring.

  • Takeaways & Limitations

    Reported metrics reflect held-out stochastic realizations from the same simulated data-generation process, so independent operating-condition validation remains necessary.

Abstract

from arXiv · show

More Electric Aircraft require fast and reliable monitoring of high-frequency electrical networks, yet most power quality disturbance and fault diagnosis methods are developed for conventional 50 or 60 Hz grids. This work presents a hardware-aware deep learning framework for multiclass detection of electrical faults and power quality disturbances in a 400 Hz aerospace power system. A high-fidelity simulation model inspired by the Boeing 787 electrical architecture generates voltage and current waveforms for 21 normal, disturbance, switching, open-circuit, and short-circuit conditions. Two datasets, each containing 73,500 samples, are formed from one-dimensional time-series signals and short-time Fourier transform time-frequency representations. Signal-processing augmentation, domain randomization, and class-specific generative adversarial networks increase waveform diversity, and the time-series dataset is released through IEEE DataPort. We compare 1D and 2D convolutional neural networks, long short-term memory networks, CNN-LSTM hybrids, ResNet, MobileNet, and VGG models under common training conditions. A compact ResNet provides the best accuracy-complexity tradeoff, achieving 96.94 percent software test accuracy with 175,685 parameters. After 8-bit quantization and deployment on a Xilinx Zynq UltraScale Plus MPSoC ZCU102, the model achieves 95.87 percent accuracy and a measured mean neural-network accelerator latency of 6.90 ms per input record. The results establish simulation-based, accelerator-level feasibility for embedded edge AI in aircraft electrical health monitoring and motivate future end-to-end data acquisition and experimental validation.

I. Introduction

More Electric Aircraft create complex, high-frequency electrical networks that require specialized monitoring and fault diagnosis. This work addresses the gap by proposing lightweight deep-learning models and FPGA deployment for aircraft power-system health monitoring.

  • Motivation: Aircraft electrification increases electrical-network complexity and reliability challenges while motivating continuous monitoring for safe and efficient operation.The paper links monitoring with early fault detection, predictive maintenance, and reduced downtime.
  • Research gap: Aircraft power-system diagnosis remains underdeveloped because most existing studies use steady-state or low-frequency signals and conventional-grid methods.The paper identifies differences in system dynamics, environmental factors, and real-time constraints as barriers to direct transfer.
  • Approach: The proposed framework performs automated multiclass detection of electrical faults and power quality disturbances in aerospace power systems.It combines high-fidelity signal generation, data augmentation, GANs, and evaluation of CNN, LSTM, and hybrid architectures.
  • Contributions: The paper contributes a Boeing 787-inspired dataset, lightweight real-time models, and an FPGA implementation designed for aerospace hardware constraints.The contributions cover dataset generation, model evaluation, and hardware resource utilization.

B. PQDs in aircraft power systems

Power-quality-disturbance research largely targets 50–60 Hz utility systems, whereas aircraft networks operate at 400 Hz and may exhibit different signal characteristics. The section motivates deep-learning and FPGA approaches suited to high-frequency, real-time aircraft monitoring.

  • Aircraft-specific gap: Aircraft electrical systems operate at 400 Hz, so their PQDs may differ from 50–60 Hz grid disturbances in shape, duration, and amplitude.The higher frequency produces signal components and time-domain behavior not typically addressed in conventional grid studies.
  • Research landscape: Most PQD studies analyze conventional residential or renewable-energy systems rather than aircraft power systems.Existing work commonly applies CNN, LSTM, hybrid, GAN, and signal-processing methods outside the aerospace setting.
  • Methodological motivation: Traditional PQD pipelines rely on feature extraction before classification, making diagnosis potentially time-consuming and labor-intensive.Deep models can learn representative features directly from one-dimensional signals or time–frequency representations.
  • Hardware motivation: FPGAs offer parallel, high-speed, compact, and lower-power computation suitable for real-time aircraft fault-management systems.Prior FPGA work includes FFT, DWT, filter-bank, and comparator implementations for PQ-event analysis.

III. Modeling Methodology

The methodology models an aircraft electrical network with high-frequency AC and multiple DC buses, then records transient and fault behavior for learning-based diagnosis. The simulated system supports monitoring of faults and PQDs relevant to critical aircraft loads.

  • Operational purpose: The monitoring objective is to detect faults and PQDs early enough to support corrective action, preventive maintenance, and protection of vital aircraft systems.The paper connects this objective to avionics, navigation, flight control, communication, and component reliability.
  • System model: The modeled aircraft system includes a 400 Hz, 230 V AC main bus, 28 V and 270 V DC distribution lines, and a Boeing 787-inspired generation and distribution architecture.The model uses a 250 kVA three-phase synchronous generator, centralized distribution, switching devices, protective components, loads, a transformer, and TRUs.
  • Power distribution: The system supplies resistive, motor, auxiliary AC, battery-charging, and DC motor loads through transformer and transformer-rectifier stages.The 28 V and 270 V buses serve different load types after AC-to-DC conversion.
  • Diagnostic signals: Switching and fault events generate transient voltage and current signatures containing temporal and spectral information useful for diagnosis.The study records these behaviors at 33.3 kHz to capture transient details.

B. Simulated Operating Conditions

The study simulates 21 aircraft power-system classes under varied operating conditions and expands them through signal-processing and GAN-based augmentation. Domain randomization introduces waveform variability intended to reflect aircraft operating conditions more closely.

  • Simulated conditions: The simulations cover 21 fault, disturbance, switching, and operating-condition classes measured with a multisensor formulation.Most events use voltage waveforms, while selected generator faults use current waveforms; PQDs are measured at the main busbar.
  • Signal acquisition: Signals are sampled at 33.3 kHz, and the simulated fault and PQD waveforms are organized by event type and measurement location.The model includes natural TRU harmonics and places PQD measurements at the main busbar.
  • Operating variability: Domain randomization varies signal depth, duration, harmonic content, transient frequency and damping, noise, frequency deviation, switching transients, and sampling jitter.These variations are intended to increase signal diversity and better reflect aircraft operational variability.
  • Signal augmentation: Signal-processing augmentation applies rotation, scaling, frequency deviation, sampling jitter, and AWGN or color noise to generate additional waveform instances.The scaling and noise operations are used to modify signal intensity and impairment strength.
  • GAN augmentation: Class-specific GANs generate the remaining samples by learning each class’s underlying waveform distribution.This strategy is intended to preserve class-specific temporal and spectral characteristics while increasing dataset diversity.

A. Dataset

The dataset combines high-fidelity aircraft-power-system simulations with signal-processing augmentation and class-specific GAN synthesis. It covers 21 conditions and evaluates generated signals for spectral, temporal, and latent-space consistency.

  • Dataset construction: 73,500 samples across 21 classes combine raw simulations, signal-processing augmentation, and GAN-generated waveforms.The dataset was augmented before a 70%/10%/20% train-validation-test split, so evaluation uses related stochastic realizations from the same simulation process.
  • GAN-based data generation: Class-specific GANs synthesize signals intended to preserve each fault or disturbance type’s temporal and spectral characteristics.Each GAN was trained independently for one class, using latent noise to generate signals matching real-sample statistical properties.
  • GAN-based data generation: PSD comparisons show mean squared errors on the order of 10^-9 across all 21 classes, indicating close spectral agreement between real and generated signals.The analysis covers dominant 400 Hz aircraft-power-system features, including harmonics, flicker-induced sidebands, and broadband transients.
  • GAN-based data generation: Mean DTW distances between real and generated signals closely follow real-to-real baselines, with higher values for abrupt interruptions and fault transients.Smoother disturbances produce lower DTW values, while more abrupt waveform changes produce higher distances in both comparisons.
  • GAN-based data generation: UMAP projections show substantial overlap and well-defined clusters, with synthetic samples extending real-data manifolds without evidence of mode collapse.Together with the PSD and DTW results, this supports preservation of spectral, temporal, and structural characteristics while expanding intra-distribution diversity.

B. Classification Model Development

The study compares diverse neural architectures under common training conditions and introduces a compact STFT-based ResNet. The Compact ResNet nearly matches the best accuracy with far fewer parameters, supporting efficient FPGA deployment.

  • Model development: The evaluation covers 1D and 2D CNNs, LSTMs, CNN-LSTM hybrids, ResNets, MobileNetV2, VGG16, VGG19, and denoising CNNs using common metrics and training settings.Recorded metrics include parameter count, accuracy, precision, recall, F1 score, and inference time per sample.
  • Compact ResNet: Compact ResNet uses three residual layers with 16, 32, and 64 channels on two-dimensional STFT inputs, followed by adaptive average pooling.Its residual connections support feature extraction while maintaining a low parameter count.
  • Model comparison: 96.94% test accuracy with 175,685 parameters gives Compact ResNet the strongest accuracy-complexity balance among the evaluated models.The 1D-CNN reaches 97.29% but requires 8.34 million parameters, only 0.35 percentage points above Compact ResNet.
  • Classification results: 96.96% precision, 96.94% recall, and 96.95% F1 score indicate balanced Compact ResNet classification, while the 1D-CNN scores slightly higher with a much larger model.The results also show that greater architectural complexity does not necessarily improve predictive performance.
  • Efficiency analysis: 175,685 parameters make Compact ResNet the smallest evaluated architecture, while VGG16 uses approximately 138 million and ResNet50 more than 23 million without meaningful accuracy gains.The accuracy-versus-size distribution places Compact ResNet in the desirable region of high accuracy and minimal parameter requirements.
  • Deployment implications: Compact ResNet’s compact structure facilitates FPGA resource allocation and low-latency, energy-efficient inference for real-time aircraft-system deployment.The paper links lower parameter requirements to more efficient use of DSP slices, LUTs, and BRAM than larger alternatives.

D. Evaluation of Compact ResNet: Learning Curves and Confusion Matrix

The Compact ResNet converges rapidly and maintains stable, balanced classification across all 21 classes, with strong diagonal performance and limited class confusion.

  • Training accuracy exceeds 90% by the fifth epoch, while validation accuracy stabilizes between 95% and 96% after approximately epoch 13.The small training–validation gap supports stable convergence and limited overfitting.
  • More than 650 of 700 test samples are correctly classified for most classes, producing strong diagonal dominance in the confusion matrix.Errors are limited and concentrated among neighboring classes, with no dominant error pattern.
  • Precision exceeds 92% and recall exceeds 91% for every class, while most classes achieve recall above 95%.The per-class results indicate a favorable precision–recall balance without a corresponding increase in false alarm rates.
  • The false negative rate ranges from 0.29% to 8.29% across the 21 classes.Further reduction of the worst-class false negative rate under independent experimental conditions remains a future direction.
  • 96.94% software test accuracy is achieved by the Compact ResNet with consistently strong classification across all 21 classes.The model shows high recall, low false negative rates, and high F1-scores across the complete class set.

E. Evaluation of Model Robustness to Noisy Inputs

The Compact ResNet retains strong classification performance under controlled additive Gaussian noise, with accuracy declining progressively as signal quality worsens.

  • 96.04% accuracy is maintained at 10 dB SNR, while 92.74% remains at 5 dB SNR under additive white Gaussian noise.The evaluation applies additional inference-time noise across SNR levels from 0 dB to 20 dB.
  • 71.26% accuracy is achieved at 0 dB SNR, where signal and noise powers are equal.Correct classification remains possible despite severe signal corruption.
  • Accuracy declines progressively as SNR decreases, without an abrupt collapse across the evaluated noise range.At higher SNR levels, accuracy remains between 96.41% and 96.93%.
  • The results support resilience to adverse operating conditions relevant to aerospace sensor measurements affected by interference and environmental disturbances.Accuracy remains above 92% at 5 dB SNR and above 96% at 10 dB SNR.
  • The robustness test adds controlled inference-time noise beyond the AWGN already used during training augmentation.This design evaluates behavior under explicitly controlled and increasingly adverse signal conditions.

F. Discussion

Across architectures, compact convolutional models provide strong fault and disturbance classification while better satisfying complexity and deployment constraints than larger or hybrid alternatives.

  • All evaluated models exceed 94% test accuracy, but their differences become clear when classification performance is considered jointly with model complexity and deployment constraints.This comparison motivates prioritizing accuracy–complexity balance rather than accuracy alone.
  • 96.94% accuracy is achieved with 175,685 parameters, compared with 97.29% for the 1D-CNN using more than 8.3 million parameters.The 0.35-percentage-point accuracy difference accompanies a complexity reduction exceeding an order of magnitude.
  • ResNet50 and VGG16 add substantial parameters without corresponding classification gains, while convolutional–recurrent hybrids do not outperform purely convolutional models.The discussion attributes the compact model’s efficiency to residual connections, stable training, and feature reuse.
  • The Compact ResNet maintains balanced precision, recall, and F1-score across all 21 classes, with strong confusion-matrix diagonal concentration and limited neighboring-class dispersion.High recall is particularly relevant to reducing missed fault detections in safety-critical systems.
  • At 10 dB SNR accuracy exceeds 96%, at 5 dB SNR it remains above 92%, and at 0 dB SNR it is 71.26%.The progressive degradation under controlled noise supports tolerance to signal corruption, while the authors note that the evaluation remains within the simulated operating envelope.
  • The reported robustness reflects intra-distribution performance within simulated aircraft-power-system conditions and does not establish deployment-level generalization.Independent operating-condition validation remains necessary because related-sample overlap can produce optimistic estimates.

V. FPGA Implementation

The Compact ResNet is selected for FPGA deployment because it combines time–frequency feature extraction with a substantially smaller model footprint and hardware-suitable efficiency.

  • The software evaluation covered all candidate architectures and both time–frequency and time-series approaches, while this section focuses on FPGA deployment.The deployment discussion therefore narrows from comprehensive software comparison to the selected model’s hardware implementation.
  • The Compact ResNet is selected for FPGA deployment because it balances classification performance with architectural efficiency on STFT inputs.Its two-dimensional processing maps naturally to spatial feature extraction in hardware.
  • The Compact ResNet uses far fewer parameters than ResNet18, ResNet50, and VGG16, which require approximately 11.3 million, 23 million, and more than 138 million parameters.This footprint reduction supports deployment under constrained hardware resources.
  • Time-series models such as 1D-CNN, 1D-CNN-LSTM, and 1D-ResNet18 provide strong temporal feature extraction but typically require greater complexity for comparable accuracy.The Compact ResNet instead exploits two-dimensional time–frequency structure with a reduced model footprint.
  • The demonstrated combination of predictive performance, noise robustness, memory efficiency, and compactness supports FPGA deployment for onboard aerospace monitoring.The stated deployment context has limited hardware resources and requires reliable real-time fault detection.

B. Hardware Setup

The FPGA implementation uses a MATLAB-controlled deep-learning processor on a Xilinx Zynq UltraScale+ MPSoC ZCU102, with AXI-based data movement and dedicated memory, convolution, and profiling modules. Post-synthesis results show low DSP usage, moderate Block RAM use, and high LUT utilization, supporting accelerator-level feasibility while leaving full system integration unresolved.

  • Platform and Processor: The MATLAB-controlled processor targets a Xilinx Zynq UltraScale+ MPSoC ZCU102 and uses Ethernet for host-to-FPGA communication.MATLAB DL HDL Toolbox deploys the neural network through a dedicated deep-learning processor IP core.
  • Data Movement and Computation: AXI4 interfaces connect external DDR memory with processing modules for input signals, weights, outputs, activations, and timing data.The architecture includes memory-access arbitration, a profiler interface, and a convolution kernel receiving weights and activations through dedicated interfaces.
  • Resource Utilization: 15.52% of DSP slices, 63.71% of Block RAM, and 93.77% of CLB LUTs were used in post-synthesis resource analysis.The design used 391 of 2520 DSP slices, 581 of 912 Block RAM units, and 256,992 of 274,080 CLB LUTs.
  • Integration Boundary: High LUT utilization demonstrates accelerator-level feasibility but does not establish that DAQ, STFT, communication, and supervisory functions fit on the same FPGA.Full integration requires separate resource-budget, placement-and-routing, and timing-closure analysis.

D. FPGA Inference Performance

The selected deep-learning model was evaluated on the FPGA for accuracy and inference latency under 8-bit quantization. It retained 95.87% accuracy with 6.90 ms mean latency, while measurements reflect only the neural-network accelerator supplied with host-generated STFT outputs.

  • Measurement Scope: Reported latencies cover only the DL IP core because the host computer provides STFT outputs, excluding acquisition and STFT processing overhead.An integrated sensor, data-acquisition, STFT, and FPGA pipeline would incur additional latency.
  • Inference Results: The worst-case onboard latency was 14.40 ms per sample, occurring for class 11 and remaining tightly bounded across classes.Latency was estimated from batches of 1000 samples per class.
  • Model Selection: The compact ResNet was selected after software evaluation for its balance between model complexity and classification performance before FPGA deployment.The implementation focused on real-time feasibility under aerospace hardware constraints.
  • Future Work: Future work includes integrating the full DAQ/STFT/DL pipeline and validating the approach experimentally with real-flight data.The authors also identify additional fault types and onboard health-monitoring integration as future directions.
Loading 2609.10479v1…