In the model-based diagnosis method of distribution network faults, the minimum hitting set calculation has a great impact on the whole diagnostic process. A new minimum hitting set algorithm suitable for the distribution network topology is proposed to improve the efficiency and accuracy rate of model-based diagnosis. The new fitness function in the algorithm allows the particle iterating towards the minimum hitting set directly. As a result, the search efficiency for the solution space is improved correspondingly. A search strategy of "feature learning" is used to reduce the search for unsolved space. Analyzing the relation between analytic redundancy relations of distribution network components, the theoretical basis and implementation method of algorithm stratification are put forward. The example shows that the improved minimum hitting set algorithm has shorter solution time and higher accuracy. Applying the improved minimum hitting set algorithm to model-based diagnosis, the efficiency and accuracy of fault diagnosis are improved. This work is supported by National Natural Science Foundation of China (No. 51377047).