Research And Application Of Software Defect Prediction Based On Deep Learning

In order to improve the quality and reliability of big-data modeling platform system for commercial aircraft intelligent manufacturing,utilizing reasonably project resources,reducing test costs and improving research and development efficiency,software defect prediction methods based on deep learning were researched and implemented.The main contents were as follows:First,a software defect prediction method based on stacked denoising sparse auto-encoder was proposed.Auto-encoder based on deep learning theory was applied to feature extraction of software defect prediction,and improved the loss cost function and sparsity constraint method.At the same time,in order to eliminate the influence of noise on the original data,a software defect prediction method based on stacked denoising sparse auto-encoder was proposed.The method can automatically learn features from the original defect data.By setting different number of hidden layers and sparsity constraint parameters and noise ratio,the required features can be extracted directly and efficiently from the original data,and then combined with Logistic regression classification to classified and predicted for the extracted features.Experimental results using the Eclipse defect datasets show that the proposed method has better prediction performance than the software defect prediction methods based on traditional feature extraction.Second,a software defect prediction method based on deep stacking forest was proposed.Deep neural networks require large-scale train data dur ing training,so the work of small-scale data will not reach ideal results due to training under-fitting.Moreover,deep neural networks have too many hyperparameters and the parameters are complicated and difficult to adjust.In order to solve these problems,deep forest was applied to software defect prediction,and improved the feature vector generation method and feature transformation method and layer-by-layer learning method.A software defect prediction method based on deep stacking forest was proposed.The method firstly use random sampling method to transform the original features,and then use stacking forests to layer-by-layer representation learning for the transformational features.Experimental results show that the proposed method can significantly improve the prediction performance and time efficiency with the software defect prediction method based on deep forest.Finally,the proposed software defect prediction methods were used to predict defects of big-data modeling platform system.Firstly extracting program modules of the software system,designing metrics and measuring,the feature datasets were obtained.Then part of the program modules are randomly selected for manual testing and tagging.The software defect prediction models were built by the defect datasets of tagged program modules to predict the defect tendencies of the non-tagged program modules.Finally,by comparing the prediction results with the actual test results,it was verified the effectiveness of the proposed software defect prediction methods.