Uncertainty-Aware LSTM Based Dynamic Flight Fault Detection for UAV Actuator

Accurate fault detection for unmanned aerial vehicle (UAV) actuators is essential for ensuring flight safety and mission completion. Without the requirement of modeling complex physical mechanism, data-driven actuator fault detection approaches have attracted much attention. Among them, the long short-term memory (LSTM) approach has shown superior performance due to its capability of modeling complex spatial-temporal features. However, the modeling uncertainty of LSTM is actually changeable under different flight conditions, which has not been well considered in the existing researches. In this article, a novel uncertainty-aware LSTM (UA-LSTM) based dynamic flight condition fault detection approach for UAV actuator is proposed. A prediction-based fault detection model is first set up based on LSTM, utilizing the information from both the flight action modes and the actuator effect. Its inputs are specifically selected by both physical mechanism and data correlation analysis. Furthermore, time-series features indicating the prediction uncertainty of the model are constructed on the selected inputs to characterize dynamic flight conditions more accurately. Then, an adaptive threshold estimation space is set up based on an enhanced distribution-based condition clustering approach. Fault detection thresholds for various flight conditions are obtained and are smoothed to reduce the influence of disturbances. Finally, the fault detection model with stepwise adaptive detection threshold is acquired. Experimental results on both simulation and real flight data illustrate that the proposed approach is superior for actuator fault detection under dynamic flight conditions.