Abstract: In the process of oilfield production, the stable operation of pumping wells is crucial for improving production efficiency and economic benefits. However, most of the existing fault diagnosis techniques rely on a single data source (e.g., schematic data or production parameters) for model training, and the diagnostic accuracy is seriously insufficient, or even the diagnostic failure occurs when faced with complex working conditions such as rod breakout and pump leakage. In this study, a CNN-XGB fault diagnosis model for multi-source data fusion is proposed. The model combines the convolutional neural network (CNN) and extreme gradient boosting (XGB) algorithms to extract the image features of the pump power diagram and the well production parameter features, respectively, to capture the feature information reflecting different working conditions from multiple angles. By integrating these features and inputting them into a multilayer perceptron (MLP), the model is able to achieve more accurate classification results, which significantly improves the specificity recognition capability. Experimental results show that the fusion model achieves more than 95% diagnostic accuracy and recall under six typical working conditions, demonstrating higher diagnostic accuracy and robustness compared to the traditional CNN and XGB models. This method effectively solves the limitation of a single data source in fault diagnosis, provides an innovative technical means for intelligent diagnosis of oilfield pumping well conditions, and has important practical application value and innovative significance.