面向航空电网故障检测的硬件感知深度学习框架

My Thoughts on Edge AI for Aircraft Electrical Systems

Okay, so here's what I've been working on. More Electric Aircraft, or MEA, demand rapid and dependable monitoring of their high-frequency electrical networks. However, most fault diagnosis and power quality disturbance methods are designed for those conventional 50 or 60 Hz grids. I'm presenting a hardware-aware deep learning framework designed specifically for multi-class detection of electrical faults and power quality disturbances within a 400 Hz aerospace power system.

We created a high-fidelity simulation model, inspired by the Boeing 787's electrical architecture. This model generated voltage and current waveforms covering 21 different conditions: normal operation, disturbances, switching events, and open/short circuits. We then formed two datasets, each containing 73,500 samples. One dataset uses time-series signals directly, and the other utilizes short-time Fourier transform (STFT) time-frequency representations. To boost the waveform diversity, we employed signal processing augmentation, domain randomization, and class-specific generative adversarial networks (GAN). Importantly, the time-series dataset is publicly available through IEEE DataPort.

We've compared several neural network architectures under the same training conditions: 1D and 2D convolutional neural networks, long short-term memory networks, CNN-LSTM hybrid models, ResNet, MobileNet, and VGG. It turns out that a compact ResNet provides the best trade-off between accuracy and complexity; it achieved a 96.94% software test accuracy using just 175,685 parameters. Furthermore, after 8-bit quantization and deploying it on a Xilinx Zynq UltraScale+ MPSoC ZCU102 platform, the model still achieved a remarkable 95.87% accuracy, with a measured mean neural-network accelerator latency of only 6.90 ms per input record. My findings establish a solid simulation-based and accelerator-level feasibility for embedded edge AI in aircraft electrical health monitoring, which inspires further end-to-end data acquisition and experimental validation.

多电飞机需要对高频电网进行快速且可靠的监测，但现有的绝大多数电能质量扰动与故障诊断方法都是针对传统的 50 或 60 Hz 电网开发的。本文提出了一种硬件感知的深度学习框架，用于 400 Hz 航空电力系统中电气故障与电能质量扰动的多分类检测。基于受波音 787 电气架构启发的高保真仿真模型，生成了涵盖正常、扰动、开关、开路及短路共 21 种工况的电压与电流波形。由一维时间序列信号和短时傅里叶变换时频表示分别构建了两个数据集，每个数据集均包含 73,500 个样本。通过信号处理增强、域随机化以及特定类别的生成对抗网络提升了波形多样性，且该时间序列数据集已在 IEEE DataPort 上公开发布。我们在统一的训练条件下对比了一维与二维卷积神经网络、长短期记忆网络、CNN-LSTM 混合模型、ResNet、MobileNet 以及 VGG 模型。其中，轻量化 ResNet 在精度与复杂度之间取得了最佳平衡，仅凭 175,685 个参数便实现了 96.94% 的软件测试准确率。经 8 位量化并部署于 Xilinx Zynq UltraScale+ MPSoC ZCU102 平台后，该模型达到了 95.87% 的准确率，实测神经网络加速器处理每条输入记录的平均延迟为 6.90 ms。研究结果在仿真与加速器硬件层面确立了嵌入式边缘 AI 应用于飞机电气健康监测的可行性，并为未来开展端到端数据采集与实验验证提供了动力。