Researchers at the Civil Aviation Flight University of China have trained an artificial-intelligence system to spot abnormal behaviour in aircraft-engine sensor records. The model flagged 97.91% of labelled abnormal windows in 1,200,109 electronic-control-unit samples from AE300 engines in service, according to a peer-reviewed Scientific Reports paper published on September 11.
The AE300 is a 168-horsepower turbocharged piston engine with electronic control and a single power lever. Its electronic control unit records both the pilot's power demand and the engine's response. The researchers focused on power-lever position and intake-manifold pressure because degradation of the turbocharger wastegate actuator can disturb that relationship during takeoff, climb and other high-load phases. These phases also produce rapid changes during normal flight, so a fixed pressure limit can confuse an ordinary transition with a fault.
The new system, called Reversible Normalization and Prediction-Assisted Long Short-Term Memory, or RPA-LSTM, processes the two sensor streams in short blocks of consecutive readings. It first rescales each block using its own mean and variance, which reduces the change in data distribution between flight conditions. A shared three-layer LSTM network then learns how the two signals normally develop over time. During training, one output branch reconstructs the input sequence while a second branch predicts subsequent readings. The prediction branch is removed for inference. If the remaining branch cannot reconstruct a new block closely enough, the system marks that block as abnormal.
After removing duplicate records, the team used 951,929 normal samples for training and 248,180 samples for testing. The split followed time order, and the test records contained annotated abnormal points. The model therefore learned normal behaviour from earlier data and was assessed on later records that included faults. Each result was scored at the window level, meaning that the unit being classified was a short run of readings rather than a single measurement.
RPA-LSTM reached 92.95% precision, 97.91% recall and an F1 score of 95.37%. It missed 42 abnormal windows. Under the same test protocol, a conventional deep neural network missed 152, an autoencoder missed 214 and a DLinear forecasting model missed 347. The new model produced slightly more false alarms than two of those baselines, but its precision indicates that 92.95% of the windows it flagged were labelled abnormal.
The component tests help explain the improvement. Removing the normalization stage lowered recall from 97.91% to 94.42%. Removing the auxiliary prediction branch lowered it to 97.01%. Adding a frequency-analysis branch did not improve the result with the two available sensor variables, so the authors left that branch out of the final design.
The dataset is not public because it contains restricted operational engine records, and the experiment covers one engine family using two signals. The paper also reports offline tests for noise, changing flight phases, detection delay and the time between false alarms, but it does not describe an onboard production deployment. A prospective fleet trial would need to show whether the alert rate and detection delay remain acceptable when maintenance teams investigate each warning and compare it with a confirmed engine condition.
Sources
- Linfeng Zhong and colleagues, Scientific Reports, September 11, 2026. DOI: 10.1038/s41598-026-67776-4. Method, AE300 dataset, results and limitations.
- Author manuscript on Research Square. Full model procedure, data split, baseline results and ablation tests.
- Diamond Aircraft, AE300/AE330 E4 Series. Official engine specifications and electronic-control description.
- Lead image: US Navy photograph released by NAVCENT Public Affairs, public domain. The photograph is illustrative and does not show the AE300 study.
.jpg){kind=link}