The defect prediction is one of the important tasks to preserve an assurance of software quality in the software engineering. In previous works of the defect predicion, two issues are identified. First, there is an issue of a heterogeneous metrics set. Many researchers use a su- pervised learning approach as to generate a defect prediction model. Then, they have to collect a training dataset that has same metrics with an objective dataset. This reduces the amount of available data, and thus should be improved. Second, it is difficult to choose the fittest model since there is an issue of differences with accuracy of a model between datasets. Various so- lutions have been reported regarding the issue of a heterogeneous metrics set. For instance, Some researchers apply an unsupervised learning method to an object dataset since unsuper- vised learning methods have an advantage that they do not need a training dataset. However, less research has focused the second issue. In this paper, we propose an unsupervised learning approach using Restricted Boltzmann Machine as preprocessing of metrics to solve the second issue. We conduct experiments on three empirical datasets. These results show that differences between five unsupervised learning methods are reduced, and all of them belong to the group which has the best AUC values. Furthermore, we confirm that unsupervised learning methods with Restricted Boltzmann Machine as preprocessing of metrics are effective on the source code complexity metrics.