A Graph Convolutional Shrinkage Network-based Fault Diagnosis Method for Industrial Process

With the increasing demand of safety monitoring for industrial process, fault diagnosis is facing greater challenges. In the paper, a graph convolutional shrinkage network is proposed for fault diagnosis. First, considering the correlation between process variables, Pearson correlation coefficient is used to convert the time series to graph structure with nodes and edges. Second, a graph convolutional shrinkage layer with anti-noise ability is designed, in which Chebyshev kernel is used as convolution kernel to extract the feature information of graph structure. In addition, a soft thresholding shrinkage layer is added to help adaptively determine the filtering thresholds of graphic feature information. Third, the graph classification for fault diagnosis is performed using the readout layer, fully connected layer and softmax layer. Finally, the proposed method's effectiveness and robustness are evaluated using TE process data. The experimental results demonstrate that the proposed fault diagnostic approach is reliable and capable of correctly identifying the fault states in noisy situations.