Conditional Enhanced Variational Autoencoder-Heterogeneous Graph Attention Neural Network: A Novel Fault Diagnosis Method for Electric Rudders Based on Heterogeneous Information

Abstract

In machine fault diagnosis, despite the wealth of information multi-sensor data provide for constructing high-quality graphs, existing graph data-driven diagnostic methods face challenges posed by handling these heterogeneous multi-sensor data. To address this issue, we propose CEVAE-HGANN, an innovative model for fault diagnosis based on the electric rudder, which can process heterogeneous data efficiently. Initially, we facilitate interaction between conditional information and the original features, followed by dimensional reduction via a conditional enhanced variational autoencoder, thereby achieving a more robust state representation. Subsequently, we define two meta-paths and employ both the Euclidean distance and Pearson coefficient in crafting an effective adjacency matrix to delineate the relationships among edges within the graph, thereby effectively representing the complex interrelations among these subsystems. Ultimately, we incorporate heterogeneous graph attention neural networks for classification, which emphasizes the connections among different subsystems, moving beyond the reliance on node-level fault identification and effectively capturing the complex interactions between subsystems. The experimental outcomes substantiate the superiority of the electric rudder-based CEVAE-HGANN model fault diagnosis.