Fault Diagnosis of Rolling Bearing Based on WHVG and GCN

In recent years, emerging intelligent algorithms have achieved great success in the domain of fault diagnosis due to effective feature extraction and powerful learning ability. However, the current models can only handle the data in Euclidean space, ignoring latent structure relationships of the signal, which can provide additional helpful information to distinguish diverse fault patterns. To address this issue, a graph convolution network (GCN) incorporating the weighted horizontal visibility graph (WHVG) is proposed for bearing faults diagnosis. The WHVG is utilized to transform time series to graph data from a geometric perspective. Edges are weighted by the difference between the sampling indexes to weaken the influence of remote nodes that are considered as noise. Furthermore, the graph isomorphism network (GIN) is improved as GIN+ to learn the graph representation and perform fault classification. Finally, the validity of WHVG and GIN+ is testified by three real-world bearing datasets. Meanwhile, the GIN+ model is compared with other machine learning models, multilayer perceptron (MLP), long short-term memory (LSTM), and two GCN models. The experimental results show that GIN+ boosts the performance and the internal structure relationships of the data contribute to the bearing faults diagnosis.

Automated summary

Emerging intelligent algorithms have achieved great success in fault diagnosis, but current models struggle with capturing all structure relationships in the data. To address this, a graph convolution network (GCN) incorporating the weighted horizontal visibility graph (WHVG) is proposed, improving performance and benefiting from the internal structure relationships of the data in bearing faults diagnosis.