Anomaly Detection for Aircraft Control System Based on Prediction and Residual Estimation With Dual Temporal Convolutional Networks

With the rapid development of complex aircraft, more accurate and reliable anomaly detection systems are needed to detect minor potential failures, and assist engineers in formulating maintenance strategies in advance. This study presents a dual-temporal convolutional network architecture for anomaly detection in aircraft control systems. The proposed framework introduces two synergistic components: 1) a primary prediction network (P-Network) generating time series prediction, and 2) a residual estimation network (R-Network) estimating the prediction residuals of P-Network. This approach considers an inevitable issue, i.e., the tolerance difference in prediction residuals caused by the uneven distribution of training data. The dynamic threshold mechanism adaptively adjusts the anomaly detection threshold by combining global training residuals and residual estimation. The experimental results based on the aircraft control system data show that the proposed method achieves excellent performance in anomaly detection, especially in detecting minor anomalies, which outperforms several benchmark methods.