Triplet metric driven multi-head GNN augmented with decoupling adversarial learning for intelligent fault diagnosis of machines under varying working condition

Effective application of intelligent diagnosis methods requires that training and testing data to follow a consistent probability distribution. However, changes in working condition lead to inevitable changes in probability distribution of collected signals, which makes the intelligent models unable to accurately classify unlabeled data obtained from different working conditions. To solve this problem, a triplet metric driven multi-head graph neural network (GNN) augmented with decoupling adversarial learning is proposed. First, we creatively use square L2 distance and triplet loss to train and measure similarity between vibration signals, and convert multiple one-dimensional signals into a sample in graph structures. This change adds correlation information between samples to dataset, thereby improving training effect of model. Then, to comprehensively analyze information contained in the vertices and edges of graphs, a graph neural network augmented with multi-head attention mechanism is constructed. Finally, in response to the complex situation of the target domain containing multiple unknown conditions, a multi-domain decoupling adversarial learning strategy is proposed to achieve the fine-grained adaptation of distinguishable structures in target domain. Three experiments are conducted to compare with six well-established models, and average accuracy of 96.67 %, 86.85 %, 100 % are achieved by proposed method, which shows significant superiority of the proposed framework.

Automated summary

The proposed method improves the accuracy of intelligent diagnosis using triplet loss and graph neural network structure, adapting to different working conditions in the target domain through decoupling adversarial learning, significantly enhancing the model's performance.