Software Reliability Prediction Based On Depth LSTM Network

As an important means to guarantee the quality of software,how to effectively predict the reliability of software is an urgent problem to be solved in the current software engineering research.So far,many scholars have studied the software reliability prediction problems and put forward many methods,but these methods all have some problems such as low prediction accuracy and insufficient scope of application.As a deep learning model,LSTM can deeply dig out the inherent laws of time series data through the learning of historical data and advanced machine learning based on selective memory,which is suitable for the processing of the data of software failure time.Based on the standard Particle Swarm Optimization(PSO)algorithm,a multi-layer heterogeneous particle swarm optimization algorithm is proposed to optimize the parameters of deep LSTM neural networks with its strong global optimization ability.And the optimized depth LSTM neural network is used for software reliability prediction.The research work of this paper is divided into the following three parts:(1)Aiming at the problems that the standard PSO algorithm is prone to premature convergence and the local optimization ability is poor,a multi-layer heterogeneous particle swarm optimization algorithm(MHPSO)is proposed.In this algorithm,the population structure of particle swarm is set as a hierarchical structure,the concept of attractor is introduced,the velocity update equation of particle is modified,the information exchange ability between particle and particle is enhanced,and the optimization performance of PSO is improved.(2)In view of the low prediction accuracy of the existing software reliability prediction algorithms,a deep LSTM(MHPSO-LSTM)neural network model optimized based on the MHPSO algorithm is proposed.In this model,the global optimization ability of MHPSO is used to optimize the initial weights of deep LSTM networks,which avoids the problem of falling into a local minimum due to the randomization of the initial weights,and improves the performance of software reliability prediction using deep LSTM neural network.(3)The performance of MHPSO and MHPSO-LSTM network models are verified by experiments.For the MHPSO algorithm,the performance of the algorithm is validated from the perspectives of benchmark function optimization,population diversity and algorithm scalability through comparison with PSO and QPSO.For the MHPSO-LSTM network model,the reliability of the software is predicted using the published software defect data sets and the performance of the prediction is verified by comparing with the prediction results of BP,RNN and conventional depth LSTM network models.