To address the challenge of ensuring accurate and reliable diagnostic outcomes for the complex nonlinear systems found in nuclear power plants, a single fault diagnosis method proves inadequate. This paper proposes a comprehensive fault diagnosis strategy that integrates expert knowledge with data-driven methodologies. Initially, it synthesizes knowledge regarding typical faults within the primary coolant system and their associated abnormal indicators through a thorough analysis. Subsequently, this scattered fault knowledge is systematically organized into a coherent fault knowledge graph. The Bayesian Network is then employed to articulate both the fault knowledge and the diagnostic process. Given the difficulty of distinguishing between similar faults exhibiting analogous signals within the Bayesian framework, a data-driven approach is utilized to refine diagnostics based on insights derived from the Bayesian network analysis. This improved method achieves a fault diagnosis accuracy rate of 98.11%. Finally, the diagnostic results are input as evidence into the Bayesian network to infer the abnormal symptoms that would be caused by the fault pattern and to compare them with the actual changes in parameters in the fault sample data. This not only validates the diagnostic results but also enhances the interpretability of the model. The findings of this study contribute to improving the interpretability and accuracy of intelligent fault diagnosis technology.