Abstract As the key machine for oil extraction, the water injection pump plays an essential role in ensuring the safe and stable production of oil. However, the water injection pump is prone to failures during operation, leading to considerable losses due to frequent shutdowns. Therefore, it is of great significance to develop an intelligent fault diagnosis method. In this work, a fault diagnosis method with improved center loss-based metric learning is proposed. Firstly, a multi-scale convolution with an attention mechanism is employed to extract key fault features from different temporal-spatial scales with adaptive weighted fusion. Furthermore, an improved center loss is imposed as a constraint on the network-generated embedding space, which enables better clustering of embedding features and consequently a reliable decision boundary for different health states. The proposed method is validated on water injection pump data and outperforms multiple comparative methods, which is able to maintain a better performance even in a highly noisy operating environment.