A DYNAMIC MULTI-SOURCE INFORMATION FUSION FRAMEWORK FOR MECHANICAL FAULT DIAGNOSIS

Deep learning methods for fault diagnosis play a critical role in the monitoring and detecting operating conditions of the mechanical equipment. However, the most existing studies are accompanied by single-source sensor and data features, which make them unsuitable for the complex and harsh real-world factory environments. In this paper, we propose a dynamic feature fusion framework based on graph convolutional networks under multi-source information named DF2GCNs. First, we extract multi-source features from raw multi-source signals through the convolutional neural networks. Then, we design a dynamic weighted factor to achieve more effective fusion between the extracted features. After that, the fused features are input to a graph construction module to construct graph topology. Moreover, we utilize the graph convolutional networks to not only extract latent features but also mine the relationships between multiple features to enrich fault-related representations. Finally, Comprehensive experiments on a multi-source test bench demonstrate DF2GCNs outperforms the state-of-the-art (SOTA) methods, striking a good trade-off between diagnostic performance and robustness.