Failure boundaries are always far away from the center of the design space of variables for low failure probability problems. In order to make full use of samples, a partitioning method of experimental levels is proposed for improving general design of experiments in this paper. The method is implemented by non-uniformly partitioning experimental levels according to the probability density function of variables, which leads to a wide interval of adjacent experimental levels at the high probability density value, while a narrow one at the low. The translational propagation latin hypercube design was improved by using the non-uniformly partitioning experimental levels method. To validate the practicability and effectiveness of the proposed method, two numerical examples are presented and the results show that the improved translational propagation latin hypercube design is more effective than the previous.