MVGNet: Multiview Graph Network With Interactive Shared Fusion for Fault Diagnosis of Wind Turbines

Abstract

Intelligent fault diagnosis of wind turbines (WTs) has gained considerable attention in recent years, and deep learning-based methods have shown excellent performance in modeling multivariate supervisory control and data acquisition (SCADA) data. However, these methods are performed in Euclidean space and cannot fully exploit the complex coupling relationship and potential fault propagation between different sensors of WTs. To address this issue, we transform SCADA data into graph structure data with sensors (as nodes) and topological connections (as edges) between sensors to represent complex interactive information. Specifically, we propose a multiview graph network with an interactive shared fusion model (MVGNet) for fault diagnosis of WTs. First, a multiview graph construction framework considering multiple measures is constructed to fully explore the fault-related features. In particular, we propose a directed causal graph for SCADA data modeling for the first time, which is beneficial for mining the causal relationships among multiple variables of WTs and providing complementary information with the correlation graph to enrich the feature representations. Then, a multihead graph attention network (GAT) with residual connection is designed to perform multiview graph representation learning. Furthermore, to take full advantage of the complementarity and consistency of multiview graphs, we propose a multigraph interactive shared fusion (MISF) module, which can adaptively perform shared learning among multiple views, provide more interactive operations, and effectively fuse multigraph features. Finally, the global graph embedding is generated for the fault diagnosis task. Experimental results on four real-world WTs datasets show that MVGNet has reliable and superior diagnosis performance compared with state-of-the-art methods.

Topics & Concepts