Small AI Models Could Guard Aircraft Power Systems Against Hidden Electrical Faults

Checking Whether Synthetic Signals Match Real Aircraft Electrical Data

Synthetic aircraft electrical signals were checked against real aircraft power-system records in three ways. The first test looked at frequency content. Each signal was turned into a power spectrum, which shows how much energy sits at each frequency. Generated signals had very small spectral errors compared with real ones. Mean squared errors were on the order of 10 to the minus ninth across all classes. That is tiny. It means the synthetic data kept the main electrical fingerprints of aircraft power systems. These include 400 Hz operation, harmonic components, flicker-related sidebands, and broadband transient energy. A model trained on such signals would see the same frequency patterns it expects from real flight power lines.

The second test compared waveform shape over time. A method called dynamic time warping measured distance between signals. It allowed small stretching or shifting without counting that as a big mismatch. Generated records sat close to real records. In many cases, they were nearly as close as real records were to other real records. Disturbances with abrupt changes showed larger distances. Voltage interruptions and fault events are examples. Smoother disturbances stayed closer together. The small extra distance between synthetic and real signals was not treated as distortion. It looked like controlled variation of the same physical event.

The third test placed real and synthetic samples in a compressed feature space. This is a mathematical map that groups similar electrical conditions together. Real and generated samples landed in the same class clusters. They did not split into separate blobs. There was no clear sign of mode collapse, where a generator produces only a few repeated patterns. Instead, synthetic samples spread through each cluster. That added variety inside normal operation, power-quality disturbances, and fault conditions. The result suggests the generator learned the underlying structure of aircraft electrical data. It did not invent strange outliers.

Comparing Fault Detectors on Aircraft Power Disturbances

Several deep-learning detectors were tested on simulated aircraft power disturbances. All of them cleared a high bar. Each architecture reached more than 94% test accuracy. That result shows the classification task is solvable by several model families. The strongest detector was a one-dimensional convolutional network. It scored 97.29% on held-out test signals. Its precision, recall, and F1 score were 97.32%, 97.29%, and 97.30%. Those numbers mean it rarely mislabeled a disturbance type. The cost was size. That network used 8.34 million parameters. A compact residual network came close behind. It reached about 96.94% accuracy with only 175,685 parameters. The gap was just 0.35 percentage points. That small loss is easy to accept when the model is far smaller. Larger standard networks did not pay for their size. VGG16 used about 138 million parameters and reached 96.55%. ResNet18 used about 11.3 million parameters and reached 96.41%. A lightweight VGG19 used about 20.3 million parameters and reached 96.48%. ResNet50 used about 23.5 million parameters but fell to 95.88%. The pattern was clear. More parameters did not bring proportional accuracy gains. Smaller convolutional models also performed well. A plain two-dimensional CNN with about 546,000 parameters reached 96.27%. A denoising-based two-dimensional CNN with 177,334 parameters reached 96.33%. These results show that the useful fault patterns can be captured without a very large network. Sequence models did not help as much as expected. The one-dimensional CNN-LSTM had the lowest accuracy at 94.57%, despite about 480,000 parameters. The two-dimensional CNN-LSTM reached 95.68% with about 756,000 parameters. Adding memory for time sequences did not improve classification on these short electrical records. Convolutional layers alone found the important shapes in the signals. The compact residual network also had balanced class performance. Precision stayed above 92% and recall above 91% across all 21 classes. Most classes exceeded 95% recall. Its false-negative rate ranged from 0.29% to 8.29%. The weakest class had precision of 92.91%, recall of 91.71%, and F1 of 92.31%. Stronger classes missed fewer than about 0.6% of real events. In safety-critical monitoring, that worst-case miss rate matters more than average accuracy. A missed fault can be worse than a false alarm.

Stress Testing the Detector With Added Electrical Noise

To check how well the detector held up, extra electrical noise was added to the test signals. The noise was white Gaussian noise, a common way to imitate random interference in an aircraft power system. It was applied at several signal-to-noise ratios, from 20 dB down to 0 dB. On that scale, lower numbers mean more noise. At 0 dB, the added noise had the same power as the signal itself.

The detector’s accuracy fell slowly as the noise grew. At 20 dB, it reached 96.93%. At 12.5 dB, it was still 96.41%. At 10 dB, it held at 96.04%. At 7.5 dB, it dropped to 95.25%. At 5 dB, it fell to 92.74%. At 2.5 dB, it reached 86.55%. At 0 dB, it was 71.26%.

These numbers matter because aircraft electrical systems can be noisy. The result shows the detector remains useful under moderate interference. It keeps above 96% accuracy at 10 dB and above 92% at 5 dB. Even at severe 0 dB noise, most samples were still classified correctly. But the drop is large enough to matter in real use. A deployed system would likely need good sensors, filtering, or confidence checks before acting on a warning.

The test was stricter than ordinary training conditions. Extra noise was added during inference, on top of noise already used while learning. That made it a stress test beyond what the model had seen. The gradual decline also suggests the detector is not relying on fragile amplitude details. It appears to use stable time and frequency patterns in the signals. Those patterns survive better when random interference is present.

Measuring How a Compact Model Runs on Aerospace Hardware

A compact fault classifier was placed on a Xilinx Zynq UltraScale+ board to test how it behaves in aerospace hardware. The model works on short-time Fourier transform images, not raw voltage traces. That means an aircraft system would first turn electrical signals into time-frequency pictures before classification. The board received batches of 1,000 inputs from a host computer over Ethernet. Timing was measured many times and averaged to get one number per input. A separate test removed round-trip communication delay. The resulting 6.90 milliseconds is only the chip-side work: moving data inside the system-on-chip and running the network. That speed equals about 145 classifications per second. For an onboard monitor, that rate can support continuous checking without long gaps between tests.

The hardware design kept memory traffic organized. Input signals, weights, and outputs lived in external DDR memory. A scheduler sent work to separate blocks: one for convolution layers, one for dense layers, and one for small math operations such as addition and resizing. Arbitrators controlled reads and writes through AXI4 interfaces. A profiler recorded start and stop times so engineers could see where time was spent. After synthesis, the processor used 391 of 2,520 DSP slices. That is about 15.5% of the arithmetic capacity on that chip. The remaining space matters in avionics because other signal-processing tasks may share the same hardware.

The fixed-point version used 8-bit weights and activations. Its accuracy fell by 1.07 percentage points compared with the floating-point software model. That small drop shows the network still separates fault patterns after conversion to lower precision. The average latency of 6.90 milliseconds is not the whole story for real-time use. A worst-case per-input time of 14.40 milliseconds was observed across all classes. It occurred for class 11. Bounded timing helps avionics teams plan predictable schedules, because safety-critical systems need to know when a result will arrive.

The measured latency covers only the neural-network core on the FPGA. It does not include sensing, isolation, data acquisition, or the short-time Fourier transform front end. A complete onboard system would add those pieces before classification. Representative sensor and converter parts can operate in microseconds, but that full pipeline was not experimentally verified here. The hardware result therefore shows a fast, compact inference engine with enough chip headroom for future integration.

A Small Watchdog for Aircraft Power

A compact detector now shows that aircraft electrical faults can be recognized on board. It does not need a large data center to do it. The key result is simple in shape and hard to beat in practice. A small residual network reached about 97 percent accuracy on simulated aircraft power faults. It used only 175,685 parameters. That is far smaller than many standard deep-learning models.

Large networks did not win by much. A one-dimensional convolutional model reached 97.29 percent, but it needed 8.34 million parameters. The extra accuracy was small compared with the added cost. For an aircraft, that difference can mean less heat, less weight, and more room for other safety tasks.

On a Xilinx Zynq board, the same compact model ran with an average latency of 6.90 milliseconds per classification. The worst case was 14.40 milliseconds. That bounded timing helps with predictable scheduling in safety-critical avionics. The processor used only 391 of 2,520 DSP slices on the chip. That leaves space for other signal-processing work.

The detector also survived added electrical noise better than a fragile model would. Accuracy stayed above 96 percent at moderate interference and above 92 percent when the noise was heavier. At extreme noise, performance fell sharply, but it did not collapse all at once. That suggests the model learned useful patterns in time and frequency.

The biggest obstacle is that the evidence still comes from simulation. Training and test samples were related variants of the same simulated aircraft power environment. That can make accuracy look stronger than it would be on truly independent flight data. For safety-critical monitoring, the worst-case miss rate matters more than average accuracy. Some classes had false-negative rates near 8 percent. A missed fault can be worse than a false alarm.

The measured latency also covers only the neural-network core. It does not include current sensing, voltage isolation, data acquisition, or the short-time Fourier transform front end. Those parts are fast in principle, but they have not been validated as one integrated onboard pipeline.

The opportunity is real. A compact model can sit close to the power system and watch for faults in near real time. It could support health monitoring, maintenance planning, and protection logic on future more-electric aircraft. Synthetic signal generation may help fill gaps, but it cannot replace real flight evidence. The next step is to move from simulated envelopes to a ground-based aircraft power-system test bed. Then the system must be checked against independently generated operating conditions and real flight data.

If a watchdog can run small, fast, and mostly accurate on board, how much proof of independence should it earn before it is trusted with an aircraft’s electrical health?