Drilling pumps serve as the heart of drilling operations, its operational status and service life significantly impact drilling efficiency. Timely and accurate diagnosis of drilling pump failures is crucial for ensuring the efficient and safe operation of drilling pumps and achieving predictive maintenance. Manual inspections suffer from poor timeliness, difficulty in detecting hidden equipment failures, planned maintenance typically fails to account for the dynamic changes in drilling pump conditions while unplanned shutdowns and repairs result in high costs and reduced equipment uptime. This paper selects the mainstream pump types used in petroleum drilling operations, specifically the F-1600 and F-2200 types as the research objects, focusing on the common failures of valve leakage and piston wear in drilling pumps. By comprehensively considering real-time logging data and relevant basic information of drilling pumps, a drilling pump fault diagnosis model based on CNN-LSTM-DNN (CLDNN) is constructed. The model combines convolutional neural networks (CNN), long short-term memory networks (LSTM), and fully connected neural networks (DNN) to effectively captures and memorizes the dependency relationships between key features, and performs multi-scale mining of feature patterns and correlation rules in drilling pump failure sequence data. Through multidimensional evaluation and validation, the model demonstrates excellent performance, achieving a fault diagnosis accuracy of 98.68%, recall rate of 100%. This provides an efficient and reliable data-driven fault diagnosis method for the fields of engineering technology and equipment manufacturing.