Detection of Abnormal Conditions in Electro-Mechanical Actuators by Physics-Informed Long Short-term Memory Networks

Electro-Mechanical Actuators (EMAs) are projected to revolutionize the flight control actuator paradigm, potentially replacing hydraulic-powered systems in the future. Consequently, the functioning of EMAs is destined to become critical for the safe and reliable operation of aircraft. Abnormal conditions of the mechanical components of EMAs can lead to their failure. The objective of this work is to develop a method for the early detection of abnormal conditions of the components of EMAs. The proposed method is based on a signal reconstruction model that estimates the motor position of EMA as expected in normal conditions of its components. Then, the presence of an abnormal condition is identified when the difference between the motor position and its reconstructed position in normal conditions exceeds a preset threshold. The signal reconstruction model employs a Physics-Informed Long Short-Term Memory network (PILSTM), whose architecture combines a physics-informed cell for the solution of the differential equations governing the EMA operation, and a data-driven Long Short-Term Memory (LSTM) cell which receives in input the output of the physics-informed cell and reconstructs the position expected in normal conditions. The proposed method is applied to data simulated by a high-fidelity model of EMAs. The results show that PILSTM is able to provide accurate, physics-consistent estimates of the motor position of EMA in normal conditions and the missed and false detection alarms are lower than those of other state-of-the-art methods.