In response to the challenges of data imbalance and feature extraction encountered in fault diagnosis of plunger-type water injection pump, a fault diagnosis algorithm that integrates metric learning and sampling adaptation mechanism is proposed. Initially, a feature extractor based on Laplace wavelet convolution and multi-scale convolutional attention mechanism is designed to capture key fault characteristics at various scales and adaptively assign feature weights. Subsequently, a metric learning module based on dual center loss is introduced to enforce constraints in the high-dimensional embedding space, facilitating the clustering of same-class samples and increasing the classification capability. Finally, a class-balanced batch sampling mechanism is adapted to enhance the model, overcoming the limitations imposed by data imbalance and further improve the performance of the model. The proposed method is validated using an imbalanced dataset of water injection pump collected from industrial sites, achieving a diagnostic accuracy up to 97.11%.