Intelligent monitoring of flight control system for complex operating conditions: the LSTM-attention method

Abstract

The Flight Control System (FCS) is a mission-critical component of Unmanned Aerial Vehicle (UAV), and its operational condition directly affects flight safety. In view of the shortcomings of the existing monitoring methods for UAV’s FCS, such as insufficient utilization of time information and lack of interpretability, this paper proposes a novel framework that integrates Long Short-Term Memory networks (LSTM) with the attention mechanism. This method uses a sliding window to construct time-series samples from flight parameters, extracts time-series features through two layers of LSTM networks, and employs an additive attention mechanism to enable the model to focus on key historical moments for predicting the FCS state. Based on the prediction residuals, adaptive thresholds and multi-point decision-making strategies are adopted to output three-level early warnings (normal, concerned, and alarm). Based on the attention weights, conducting interpretability analysis is helpful for achieving fault location. Experiments based on real UAV flight data show that the LSTM-Attention model outperforms BPNN, SVM and standard LSTM in terms of mean square error (MSE) and mean absolute error (MAE). In anomaly detection, the average F1 score reaches 94.9%, and the inference time is only in the millisecond range, meeting the real-time requirements.9%, and the inference time is only in the millisecond range, meeting the real-time requirements.