Most of the existing researches on belief rule-base inference methodology focus on the problem of parameter optimization, while few on the problem that the belief rule-base system cannot operate normally due to the incomplete input information caused by the difficulty of data acquisition. In order to solve this problem, an inference method based on incomplete input is proposed. First, the precondition attributes distribution was obtained from experience or historical data using statistical method. Then, the step sampling method was used to obtain multiple candidate values for the missing precondition attribute, which was combined with other precondition attributes. Each of the inputs was inferred using the belief rule-base inference methodology. Finally, the ER algorithm was used to fuse all the input inference results. In the simulation experiment of automobile engine fault diagnosis, the method of this paper was compared with other methods. Results show that the cumulative inference error of the proposed method was obviously smaller than other methods when there was sufficient historical data. Moreover, through multiple experiments, it was found that the proposed method has good adaptability to different distributions. This method reduces the inference error on the basis of minimizing the computational complexity, and provides a new idea for the belief rule-base inference of incomplete input information.