Semi-Supervised Learning of Multiscale Graph Neural Networks for Industrial Process Fault Diagnosis

The scarcity of labeled data is a critical challenge in industrial process multi-scale modeling, as learning reliable models from limited labeled data and large-scale unlabeled data is crucial for fault diagnosis tasks. To solve the issue effectively, this paper proposes a novel semi-supervised multi-scale graph neural network (SsMsGNN) based on the framework of graph convolution autoencoder networks (GCAN). The SsMsGNN mainly comprises a multiple scale graph structure learning module, a multi-scale feature extraction network, and a multi-scale feature fusion network. The multi-scale feature extraction network captures multi-scale inherent features from raw data through different pathways. Here, two variant multi-scale feature extraction network architectures are presented for GCAN, i.e. the multi-scale pyramid feature (MsPyF) and the multi-scale parallel feature (MsPaF) networks, each employing two optimization strategies (frozen and cross) to improve model performance. Meanwhile, the multiple scale graph structure learning module generates scale-specific adjacency matrices to represent topological relationships between process variables. Finally, the multi-scale feature fusion network aggregates representations from multiple layers to establish the fault classification network. Moreover, the semi-supervised learning mechanism of the GCAN can effectively leverage both labeled and unlabeled data for the proposed SsMsGNNs. Case studies on the Tennessee Eastman and the Three-Phase Flow processes demonstrate that the proposed SsMsGNNs outperform existing methods, showing significant improvements in fault diagnosis performance, especially when incorporating numerous unlabeled data. Codes of the proposed models are available at https://github.com/seajq/SsMsGNN.