RPA-LSTM: Reversible normalization and prediction-assisted LSTM for aero-engine intake-system ECU anomaly detection

Reliable anomaly detection in aero-engine monitoring systems is critical for flight safety, maintenance planning, and operational reliability under variable operating conditions. For turbocharged aviation piston engines, intake-system faults induced by turbocharger wastegate actuator degradation frequently occur during takeoff, climb and other high-load phases, creating an urgent demand for robust time-series diagnosis methods. This paper proposes a Reversible Normalization and Prediction-Assisted LSTM (RPA-LSTM) framework for general aero-engine ECU intake anomaly detection, and real-world AE300 airborne ECU data are adopted as the verification dataset. The model takes intake-manifold pressure and power-lever position as input sequences. Reversible Instance Normalization (RevIN) is used to alleviate instance-wise distribution shifts, while an auxiliary prediction branch is introduced to strengthen temporal representation learning. A shared LSTM encoder captures sequential dependencies, and only the reconstruction branch is retained for lightweight inference. Experiments on 1.2 million in-service AE300 ECU samples demonstrate that the proposed method achieves an accuracy of 0.9992, a precision of 0.9295, a recall of 0.9791, and an F1-score of 0.9537. Compared with seven mainstream time-series baselines, it achieves the highest recall with significantly lower computational cost than Transformer, demonstrating a favorable trade-off between detection sensitivity and deployment efficiency. Ablation results verify that both RevIN and the prediction-assisted branch bring stable performance gains, while an extra frequency branch cannot improve detection accuracy under the two-variable input setting. Supplementary experiments including Monte Carlo random subsampling cross-validation, noise robustness tests, multi-flight-phase adaptability analysis, detection delay and mean time between false alarms (MTBFA) measurement, as well as anomaly ratio threshold sensitivity analysis are conducted in this paper. The results fully prove the generalization, anti-interference ability and practical deployment value of the lightweight RPA-LSTM framework for aero-engine ECU monitoring