Metric learning-based two-stage imbalanced fault diagnosis model for water injection pump

To address the severe data imbalance and the difficulty of feature extraction in fault diagnosis of plunger‐type injection pumps, we propose a fault diagnosis algorithm that integrates metric learning with a sampling‐adaptation mechanism. First, we design a feature extractor based on Laplacian wavelet convolution and a multi‐scale attention mechanism to capture critical fault features and adaptively assign feature weights. Second, we introduce a metric‐learning module driven by dual-center loss, imposing inter-class and intra-class distance constraints in a high-dimensional embedding space to promote clustering of samples from the same fault class and thus enhance fault recognition capability. Finally, we adopt a class-balanced batch sampling mechanism to adaptively strengthen the model, overcoming the performance limitations imposed by data imbalance and further improving overall accuracy. Validated on a real, imbalanced injection-pump dataset, the proposed method achieves a diagnostic accuracy of 97.11%.