The rapid and accurate diagnosis of accidents is crucial for the operation and maintenance of reactors. However, traditional accident diagnosis methods based on knowledge or pure data-driven approaches often struggle to strike a balance among computational efficiency, accuracy, and interpretability. To address this, this study proposes a hybrid-driven accident diagnosis model that combines knowledge and data. First, for the Loss of Coolant Accident (LOCA), Station Blackout (SBO), Main Feedwater Line Break (MFLB), and Steam Generator Tube Rupture (SGTR) accidents, preliminary diagnosis models are constructed based on four algorithms: eXtreme Gradient Boosting (XGBoost), Support Vector Machine (SVM), Back-Propagation Neural Network (BPNN), and Long Short-Term Memory Neural Network (LSTM), respectively. Then, by integrating knowledge, the confusion matrix is used to locate the misdiagnosed operating conditions and accident types. Intensive sampling is carried out for the misdiagnosed intervals to generate a high-quality dataset. Based on this dataset and the misdiagnosed operating conditions, a re-diagnosis model is further trained to calibrate the misdiagnosed results, ultimately forming an accurate diagnosis model. In addition, the SHAP algorithm is used to analyze the importance of feature parameters in the diagnosis model, monitoring the contribution of each parameter to the diagnosis results. The results show that the hybrid-driven accident diagnosis model proposed in this paper can effectively improve the diagnostic accuracy of pure data-driven models, and has high interpretability and practicality.