Two-layer fault diagnosis model of aircraft based on LSTM

Different component malfunctions of the aircraft may cause fault characteristics crossing easily and similar trajectory deviations, making traditional single-type diagnostic approaches to risk misidentification. To accurately ascertain the fault status based on sequential data such as flight trajectories and attitudes, this paper introduces a two-tiered fault diagnosis model leveraging long short-term memory (LSTM) neural networks. The first layer of the model discerns the fault mechanism, while the second layer pinpoints the specific fault type. Datasets encompassing four distinct actuator faults and four varying degrees of thrust drop faults (20 %, 40 %, 60 %, and 80 %,) are constructed for model training and validation. Simulation results show that employing a two-layer LSTM network enhances both training convergence and testing precision, achieving a diagnostic accuracy of 99.38 %, surpassing the 96.74 % accuracy of a single-layer network. Furthermore, the LSTM network's diagnostic accuracy, utilizing sequence information, significantly outperforms that of the Least Squares Classifier and the BP neural network, which rely on individual state information. Thus, the efficacy of proposed two-tiered LSTM-based fault diagnosis model in improving diagnostic accuracy and reliability in aircraft systems has been validated.