Dual graph driven-consistent representation learning method for semi-supervised fault diagnosis of rotating machinery

Abstract

Graph neural network (GNN) is an effective tool for semi-supervised fault diagnosis of rotating machinery. However, existing GNN based-semi-supervised methods only rely on single graph structure to learn feature representation under limited labeled samples, while the information of different topology graph structures cannot be directly fused due to the large difference of feature extracting, leading to insufficient node relationships and label information mining. Besides, static or limited dynamic feature extraction of neighbor nodes will hinder the expressiveness of semi-supervised GNN models. To overcome these limitations, a dual graph driven-consistent representation learning method (DGDCRL) is proposed in this paper. First, a dual graph structure with two different topology graphs is conducted using graph label passing method, in which limited labeled sample information are fully leveraged and richer topology structure information among nodes can be captured. Second, a consistent representation learning method with gated-dynamic enhanced graph attention module (GDEGAT) is proposed to extract the common embeddings from two topology graphs, where a DEGAT layer is developed to aggregate neighbor information more dynamically and expressively. Besides, to enhance the alignment between the embeddings of the same nodes across two topology graphs, we design a consistent representation loss. Two datasets are used to validate the performance of the proposed method, indicating that the proposed DGDCRL method with GDEGAT module can achieve the effective diagnosis results of rotating machinery under both constant and variable speed conditions, and the DGDCRL method can effectively enhance the semi-supervised diagnostic ability of baseline GNNs under low labeled rates.