WGCAN: A Weighted Graph Convolutional Adversarial Network for Open-Set Machinery Fault Diagnosis

Recent advancements in unsupervised domain adaptation (UDA)-based rotating machinery fault diagnosis have been notable. However, most intelligent diagnosis algorithms predominantly rely on 1-D convolutional neural networks (CNNs) for feature extraction, neglecting the structural information within the data. Additionally, most existing algorithms operate under a closed-set setting, assuming that the source working condition and target working condition share the same label space. Unfortunately, it is impractical in real-world applications as new faults are likely to emerge under variable working conditions. To address these problems, a novel cross-domain open-set fault diagnosis method named weighted graph convolutional adversarial network (WGCAN) is proposed in this article. First, a CNN-graph convolutional network (CNN-GCN) hybrid architecture is designed for feature extraction from vibration frequency domain signals, where graph convolutional blocks are concatenated after 1-D convolutional blocks and the graph structure is generated by a graph generation layer (GGL). Subsequently, the obtained high-level features are fed into a fault classifier and adversarial learning is employed to establish a boundary between known and unknown faults. In addition, a classifier-based auxiliary weighting module is included to provide source similarity weights for target samples in adversarial training. Experiments on two gearbox datasets demonstrate that the proposed WGCAN can simultaneously identify known and unknown faults under variable working conditions and outperform other diagnosis algorithms.