Abstarct: Nowadays, during the construction and development of software projects, one of the problems that developers mainly struggle with is testing and finding software faults. But this process consumes a lot of time and resources. The concept of software faults prediction by spending less time and resources can greatly help the finding of software faults. Software fault prediction is a process by which faults-prone points in software modules are identified using a prediction model that is built. The software fault prediction model is built by training machine learning techniques, whose sources for construction and training include the fault data set that includes a large number of software features. Software has different levels including modules, classes, and functions. One of the types of software fault prediction model is the prediction fault model in functions. In this research, a new databaxse called BugHunter was used to help predict software faults, which consisted of 15 Java projects at three levels: module, class, and functions. So far, different machine learning techniques such as simple Bayes, logistic regression, random forest, etc. have been used on this data set to discover software faults. The problem that exists in the meantime is that the accuracy of these prediction models is low and the single algorithmic methods in this field have reached their threshold. In this research, we tried to improve this threshold by using the combined method. According to the subject under study, it has been tried to present a model for the software faults predicting in Java language functions by combining random forest, decision tree, k-nearest neighbor, naive bayes and AdaBoost algorithms. During these steps, different configurations for these algorithms have been tested so that we can optimize these algorithms in order to improve the results. To combine the algorithms, two methods of voting and staking algorithms have been used. Finally, the proposed method was tested on the Bughunter databaxse, and according to the results obtained, the F1-score was increased to 93% in the best case.