Deep Feature Decomposition Autoencoder: Intelligent Fault Diagnosis for Unknown Faults in Landing Gear Systems

The landing gear system of civil aircraft has the design characteristics of high safety and reliability, which contributes to the scarcity of fault types and samples that can be obtained in practice. The incompleteness of fault labels and samples can not only cause the small sample problem but also lead to the zero-shot problem, meaning that traditional data-driven fault diagnosis models may not apply to identifying unknown faults. Aiming at this problem, a deep feature decomposition autoencoder (DFDAE) method is proposed to realize the unknown fault diagnosis of the aircraft landing gear system. Firstly, a set of parallel semi-supervised asymmetric convolutional autoencoders is constructed, combined with a deep feature decomposition strategy to extract features with interpretable differences from high-dimensional data. Secondly, an adaptive threshold strategy based on the distance difference ratio is proposed. The distance relationship between the test sample and the training set is analyzed to adaptively determine the threshold, avoiding the deviation caused by human subjective experience and realizing the diagnosis of unknown faults. In addition, a silhouette-density clustering evaluation strategy is used to predict the number of unknown fault categories. Finally, the performance of the proposed algorithm is analyzed using a simulation data set of the landing gear system. The results show that DFDAE can realize the unknown fault diagnosis of the aircraft landing gear system with little training in the fault category.