Metal oxide arrester with electric charge runs for a long time, and the latent defects cannot be identified by regular maintenance or online monitoring. Therefore, this paper proposes a data mining method to assess the operating condition of the arrester with the typical operating parameters. First, the characteristic parameters from the lightning detection, online monitoring, on-site inspection, and pre-operation information of arresters are selected and formed a defect feature quantity database. Then, the semi-trapezoidal model is used to normalize the quantitative parameters, and the natural language processing technology is introduced to normalize the qualitative parameters. Besides, a data fusion method based on random forest optimization is proposed. Finally, all of the arrester data for a substation are adopted for analysis. The example shows that the assessment accuracy of the proposed model is 93.12%, and it has better generalization ability than the decision tree model and the support vector machine model.