Abstract: To address the problems of limited fish disease case data, heterogeneous feature descriptions, and incomplete information in aquaculture scenarios, this study proposes a knowledge-driven lightweight diagnostic method. Authoritative literature and monographs on freshwater fish diseases in China were systematically reviewed, and the acquired disease-related information was subjected to multi-granularity semantic disambiguation and standardized processing. An expert knowledge base was constructed, covering 4 disease type, 83 diseases, 9 symptom part, and 61 typical symptom. Boolean encoding was used to transform expert knowledge into feature vectors, and a Naive Bayes diagnostic model was developed to enable small-sample probabilistic reasoning. To evaluate the practical application performance of the model, an interference validation strategy based on combinatorial subset enumeration was designed to simulate symptom omission and false-reporting scenarios in real diagnostic processes. The results showed that under complete-information conditions, the proposed model achieved accuracy, F1-score, and other core metrics above 0.98; under missing or noisy conditions, its performance remained above 0.85 and outperformed the comparative models. On a Raspberry Pi 5 platform, the training and validation times of the model were both at the millisecond level, with an average central processing unit (CPU) utilization of 4.8%. These results indicate that the proposed method has good diagnostic stability and anti-interference ability under small-sample, symptom-omission, and false-reporting conditions. It also has low computational overhead and is suitable for deployment on low-computing-power terminals, providing technical support for auxiliary fish disease diagnosis in grassroots aquaculture scenarios.