Deep hierarchical sorting networks for fault diagnosis of aero-engines

Abstract

In modern industry, timely health assessments of aero-engines are crucial for ensuring their proper functionality and the safety of aviation operations. However, during the collection of operating data for aero-engines, influential fault features may exhibit hysteresis or even overwhelmed due to transmission delays in some sensors. Furthermore, these features in the data at interval points are difficult to extract using traditional deep neural networks. Moreover, in aero-engine fault diagnosis, the number of normal samples is significantly higher than that of fault samples. As a result, traditional deep neural networks tend to focus on normal samples while fault samples are neglected, increasing the risk of missed diagnoses or misdiagnoses. To address these problems, this paper proposes a parallel convolutional neural network based on hierarchical sorting of state points (FSHSM-PCNN), to improve the synergistic effect between state point data at different hierarchical levels via the hierarchical sorting module, and to efficiently extract fault information via the parallel convolutional neural network. First, the state point data in the original samples is internally sorted along the time dimension by the fault significance-based hierarchical sorting module (FSHSM), and the different levels of state point data obtained after sorting reveal a reinforced synergistic effect. Second, a parallel convolutional neural network is developed to extract temporal status features and reinforced synergistic features, and the fused information is used for fault diagnosis. Finally, the performance of the proposed FSHSM-PCNN is evaluated using actual monitoring data from aero-engines. The experimental results show that the proposed method is effective in extracting fault features from the monitoring data. Compared to other methods in the ablation study, the proposed method improves average performance in aero-engine fault diagnosis by 12.46 %, 7.07 %, and 12.62 %, respectively. In diagnosis tasks with imbalanced datasets, its accuracy exceeds that of other methods by at least 5.01 %.