In order to quantify the actual impact of maintenance operation on the transformer failure rate, this paper establishes a prediction model of transformer internal latent failure rate based on real-time operating conditions by deeply mining the oil chromatographic monitoring data on a short-time scale. This paper firstly analyzes the modeling process of the traditional Markov failure rate model , points out that the independent variable state duration in the model cannot respond to the iinfluence of maintenance operation, and proposes to modify the state duration parameters. The specific steps are to extract the key dissolved gases based on R clustering analysis-principal component analysis ( RCA-PCA). And then, the mapping relationship between the key dissolved gas contents and the modified state duration is determined by the radial basis function ( RBF) neural network. Finally, the analytical expression of latent failure rate inside transformer with the modified state duration is derived. Results show that the proposed model has the ability to characterize the impact of maintenance operations and the actual failure rate compared with the traditional Markov model.