Unsupervised Graph Contrastive Learning-based Anomaly Detection Method for Rotating Machinery

By merits of strong ability in exploiting the degradation information of machinery, deep learning (DL)-based methods have achieved great success in rotating machinery fault diagnosis. Among them, supervised DL-based diagnosis methods usually need to use sufficient labeled samples to train a superior DL model. However, labeled samples in real industry are often difficult to obtain, failing in providing sufficient labeled samples for model training. Therefore, an unsupervised graph contrastive learning-based anomaly detection method for rotating machinery is proposed. An improved short-time Fourier transform-based graph is conducted, where frequency is regarded as the node and time-domain feature is the node feature. On the basis, the Pearson correlation coefficient is calculated to determine the edge connection. The graphs with similar energies are selected as positive pairs, and the graphs constructed from high-frequency components obtained by empirical mode decomposition are selected as negative pairs. Then the graph convolutional network is adopted to learn the low dimensional representation of graph features, and a designed contrastive loss function is used to maximize the similarity between positive pairs and enlarge the difference between negative pairs. Finally, division of normal samples and abnormal samples is realized by the density-based spatial clustering of applications with noise. Comparative experimental results show that the proposed method achieves a competitive result.