Pseudo-sample enhanced fault diagnosis method for unmanned underwater vehicles based on graph adversarial modeling and metric constraints

Thruster failures pose a threat to the navigation safety of unmanned underwater vehicles (UUVs) in complex marine environments. However, due to the influence of test conditions and fault randomness, the inter-class imbalance leads to the difficulty in diagnosis. To address this issue, a pseudo-sample enhanced fault diagnosis method is proposed by introducing metric constraints into a graph adversarial model. Firstly, this method constructs a feature-consistent pseudo-sample generation model based on graph adversarial modeling and improves the modeling ability of the model for complex association features between samples by mining the potential topology in the data. Based on this, a joint optimization strategy combining adversarial loss and feature space metric loss is designed, which makes the generated pseudo-samples closer to the real data and improves the model robustness under imbalanced data. Experimental tests using the vibration data of real UUV thrusters, along with comparisons against multiple baseline models, are conducted. The results show that the proposed method achieves superior diagnostic performance and stability across different operating conditions.