Abstract: In order to improve classification accuracy of traditional multinomial-Dirichlet model in sparse data scenario, a hierarchical Bayesian network parameter estimation method was proposed based on variational inference. Introducing a hyper-prior into the traditional multinomial-Dirichlet model, the hierarchical multinomial-Dirichlet model was constructed to estimate the conditional distribution in Bayesian networks. Analyzing the prior dependency structure of hierarchical multinomial-Dirichlet model, a fast and accurate self-organizing variational reasoning algorithm was developed. Compared with the traditional classification model, the hierarchical multinomial-Dirichlet model proposed in this paper shows a significant performance improvement in dealing with the fault classification problem of liquid rocket engines with small data sets.