In practical engineering, there are few samples of arrester failures, and it is difficult for intelligent algorithms such as neural networks to make accurate judgments. To this end, a fault diagnosis method of arrester based on Bayesian network was proposed. Firstly, the principal component analysis was used to extract 21 characteristic parameters that affect the operation of arrester. And then the extracted characteristic parameters was chosen to establish two-layer information architecture defect diagnosis model. The classification probability of different categories was calculated according to the existing real-time data. If the first classification result indicated that the arrester is abnormal, new detection evidence was added for the second diagnosis. Finally, 6 arresters under the same voltage level in a certain area were selected to analyze and verify the validity and correctness of the proposed method.