Noise-robust multi-view graph neural network for fault diagnosis of rotating machinery

Modern large-scale equipment is deployed with multiple sensors to monitor the operating state in real time, thus imposing higher requirements on intelligent fault diagnosis methods. However, current deep learning-based methods for multi-sensor information fusion often rely on features extracted from a single domain, which are incompetent to characterize the diversity and complexity of multi-sensor signals. To fully exploit the potential of multi-domain features, a Multi-view Graph Neural Network (MvGNN) combining time domain (TD) and frequency domain (FD) features is proposed for the fault diagnosis of a multi-sensor rotating machine system. Firstly, the normalized TD signals are modeled as graph-structured data in terms of the k nearest neighbor (kNN) k NN) algorithm. The nodes' initial features are transformed into two different feature spaces using Convolutional Neural Network (CNN) and Fast Fourier Transform (FFT) to form multi-view (TD and FD) graphs. Subsequently, single-view graphs are learned by independent graph convolution blocks separately to aggregate the multi-sensor information. Lastly, a view-attention block is designed to compute the unified representation of the multi-view graph, which is subsequently input into a classifier to diagnose the health state. To verify the capabilities of MvGNN, two case studies are performed on public datasets. Experimental results show that the proposed method has satisfactory diagnostic accuracy and surpasses comparative methods. In addition, the abundant information contained in the multi-view graph endows the proposed method with stronger robustness in a noisy environment.