Research on piston aircraft engine fault diagnosis based on multisource information fusion by graph convolutional network

Abstract

Vibration signals have been widely used in engine fault diagnosis due to their excellent ability to characterize faults. However, they suffer from disadvantages such as susceptibility to noise interference and difficulty in extracting weak fault features. To address this, this article introduces acoustic emission (AE) signals to complement vibration signals with complementary high- and low-frequency characteristics, thereby enabling effective diagnosis of subtle engine faults. Nevertheless, traditional single-level fusion models still face significant challenges in comprehensively extracting faint fault features and deeply exploiting information complementarity. To overcome these limitations, this article innovatively proposes a multisource multilevel information fusion model based on graph convolutional networks (GCNs). The framework utilizes GCNs to compensate for traditional deep learning algorithms’ inability to capture spatial relationships among nodes; introduces an attention-based multiscale feature cross-fusion module to achieve cross-fusion of multisource features across different scales; and integrates a multilevel fusion strategy with Dempster–Shafer evidence theory to accomplish decision-level fusion, enabling multisource information integration at multiple scales and deep levels. The effectiveness of the proposed method in diagnosing aircraft engine faults is verified through the establishment of experiments, and excellent performance is achieved in the remaining working conditions, which prove that the method has a certain degree of universality.