Real-Time Aerospace Power Monitoring with Quantized ResNet

A compact ResNet model enables real-time fault detection in 400 Hz aerospace electrical systems, achieving 95.87 percent accuracy on a Xilinx Zynq UltraScale Plus MPSoC. This framework provides an alternative to traditional grid-monitoring tools designed for 50 or 60 Hz terrestrial grids, which lack the temporal resolution to capture transient electrical faults in high-frequency, variable-load aircraft environments. For engineers, this confirms that hardware-aware model design and 8-bit quantization can bring high-fidelity fault diagnosis directly to the edge, replacing reliance on heavy, non-embedded hardware.

Detecting Faults in High-Frequency Waveforms

Power systems in aircraft operate at 400 Hz to improve power density and reduce weight. Standard diagnostic tools rely on steady-state analysis, but these struggle with the transient anomalies inherent in high-frequency networks. To overcome this, the researchers trained a ResNet model on spectrogram representations generated by short-time Fourier transforms. The short-time Fourier transform takes the 1D voltage signal over time and decomposes it into a 2D time-frequency representation, or spectrogram. This process creates an image-like structure where the horizontal axis represents time and the vertical axis represents frequency, allowing the 2D convolution filters in the ResNet to treat harmonic distortions and spectral shifts as spatial features to be classified. Instead of simply monitoring voltage levels, the model learns to identify specific patterns that distinguish a fault from a normal switching event. A standard switching event might cause a momentary voltage spike, but a catastrophic failure like a short-circuit introduces a distinct, wide-spectrum energy shift that the ResNet architecture is uniquely tuned to isolate.

Quantization and Hardware Throughput

Deploying models onto an embedded Xilinx Zynq UltraScale Plus MPSoC ZCU102 requires balancing predictive performance with finite silicon resources. The researchers applied 8-bit quantization to the model weights. By converting 32-bit floating-point parameters to 8-bit integers, the system reduces the required memory bandwidth. Crucially, the Zynq MPSoC contains dedicated DSP slices and logic fabric that can perform far more parallel 8-bit integer arithmetic operations per clock cycle than 32-bit float operations. This higher throughput is the driver for achieving a 6.90 ms inference latency on the embedded fabric. Comparing the software-based model at 96.94 percent accuracy to the quantized version at 95.87 percent shows a performance delta of only 1.07 percentage points, proving that the model remains robust despite the reduction in numerical precision.

Training for Real-World Diversity

The training process relied on synthetic data, which is necessary because waiting for years of rare, catastrophic flight-test failure data is impractical. To bridge the gap between simulation and the actual cockpit, the researchers used generative adversarial networks and domain randomization. These techniques inject artificial signal variations, such as jitter, white noise, and amplitude fluctuations, into the simulation data. By forcing the model to ignore this noise during training, it becomes resilient to the chaotic electrical environment found in flight hardware.

This approach shifts the burden of model development away from physical data collection. However, a significant gap remains between synthetic performance and production reliability. Simulated environments often fail to model specific, non-linear factors like electromagnetic interference from other onboard systems and long-term sensor drift, both of which degrade signal quality in ways synthetic noise profiles cannot fully replicate. Future validation must move from simulation-derived weights to experimental data acquisition in physical aircraft electrical systems to determine how much the accuracy drops when exposed to these unmodeled, real-world variables. Until hardware-in-the-loop testing on a test stand or flying aircraft is completed, the model's resilience to actual airframe EMI remains a known unknown.