Simulation-based validation of multivariate time-series anomaly detection algorithms for aircraft edge systems

This study presents a systematic approach for selecting and evaluating anomaly detection algorithms for the development of an edge system for real-time aircraft data analysis. Using the HAI 21.03 dataset, a highly imbalanced dataset in which anomalies account for less than 3% of the total data, the performance of a prediction-based Long Short-Term Memory Prediction (LSTM Pred) model and reconstruction-based Long Short-Term Memory Autoencoder (LSTM AE), Convolutional Neural Network–Long Short-Term Memory Autoencoder (CNN–LSTM AE), Variational Autoencoder (VAE), and Transformer-based anomaly detection (TransAD) models is compared. Based on the analysis of real-time processing capability and the F1-score, the LSTM AE model is selected for further evaluation. Normal and abnormal flight operation simulation data are generated using the commercial flight simulation software X-Plane. When abnormal data are input into the selected LSTM AE model, increased Mean Squared Error (MSE) values are observed near the injected fault events. These results suggest that LSTM AE is a promising candidate for real-time anomaly detection in aircraft edge systems under the experimental conditions considered in this study.