Fault Prediction for Software System in Industrial Internet: A Deep Learning Algorithm via Effective Dimension Reduction

Abstract

In recent years, as information technology develops, Industrial Internet has become a hot issue in international industry. Because of the use of networked software in Industrial Internet, many machines are software-intensive. But software is a product of human brain thinking activities, with the increase of its scale and complexity, there will inevitably be some software defects caused by human errors in the process of design and development. Nowadays, software fault diagnosis mostly relies on personal experience and lacks effective technology and methods, which seriously affects the ability of software-intensive systems. This paper mainly designed a software system fault prediction model, which can be used by software-intensive system users of Industrial Internet to quickly predict whether software failures occur or not. Different fault prediction methods based on deep learning are introduced. We propose a software fault prediction model based on locally linear embedding (LLE) algorithm and long short-term memory (LSTM) algorithm to train the model. Original data sets are from MDP dataset of NASA. We process original datasets by using LLE algorithm to reduce dimensions of datasets. After processed datasets were trained by LSTM algorithm, the prediction model can be obtained. Compared with single LSTM and principal components analysis-long short-term memory algorithm (PCA-LSTM), the results show that locally linear embedding-long short-term memory algorithm (LLE-LSTM) algorithm has a better performance than other existing algorithms in terms of prediction accuracy, precision, recall, f-measure.