【作者】 邵开丽；

【导师】 王命延；

【作者基本信息】 南昌大学 ， 计算机软件与理论， 2007， 硕士

【摘要】 在计算机技术飞速发展的今天，人们对计算机系统的依赖性越来越高。而在造成计算机系统错误的因素中，软件占了绝大部分。随着软件体系规模的日益增大及其复杂性的日益增强，软件的可靠性问题更为突出。因此，如何保证软件的质量，有效地对软件产品特性进行度量和预测，对开发期间的现状进行控制，设计并开发可靠的软件己成为当务之急。而软件可靠性预测能够使开发和测试的相关人员对软件的可靠性有一个大致的了解，所以对于软件如何进一步开发、测试和质量的控制都具有十分重要的意义。神经网络自开创以来一直深受许多学者的重视，并广泛运用于各种领域，取得了辉煌的成就。预测是神经网络的又一个重要应用领域，这是因为神经网络具有优良的非线性特性，特别适用于高度非线性系统的处理。所以基于神经网络的智能预测是解决非线性预测问题的有效方法，为预测理论开辟了新的广阔发展空间。本文概述了国内外关于软件可靠性预测方法的研究现状，重点研究和分析了传统的软件可靠性模型和BP神经网络的结构和特点，指出了它们的不足之处。并结合前人运用神经网络进行软件可靠性预测的相关理论和成果，针对它们的不足之处进行了改进。特别是BP神经网络具有的收敛速度慢和易陷入局部极小值等问题，主要从算法和网络结构两方面着手进行了改进。在启发式改进算法方面通过增加动量项和可变的学习速率；数值优化方面主要使用Levenberg-Marquardt算法训练网络。而网络结构主要针对初始权值的选取和隐含层节点数的确定进行了优化设计；同时结合区间探测法、逐步搜索法、数据归一化和交叉验证等数学工具，提出了基于神经网络的软件可靠性预测模型。最后，通过运用MATLAB仿真工具对各种模型进行了数值仿真分析，证实了新模型同传统模型相比预测精度更高，泛化能力更强和良好的一致性。为了进一步验证这一模型的实用性，特将其应用到某手机软件系统以估测此软件的可靠性。更多还原

【Abstract】 With the rapid development of computer technology today, we depend more andmore on computer systems. Software is the primary factor among all that result incomputer system errors. The increasingly extended scale and complexity of softwaresystem make its reliability issue much more important. So how to secure the qualityof software as well as design and develop reliable software is our urgent task. As theprediction of software reliability enables the developers and testers to get generalideas about the software reliability before testing and using it, it is important forfurther development, testing and quality control of software.Neural Network (NN) has been deeply appreciated by many scholars. Now it isused with great success in many fields. Prediction is one of the important applicationfields of NN. Most of the general predicting methods are based on linear analysis,when it comes to non-linear they met many difficulties. Therefore NN is competentfor non-linear proceeding for its excellent non-linear character. Predicting methodsbased on NN extend the space of predicting research.This paper summarizes the status of software reliability prediction methods inhome and abroad. Through research and analysis of traditional software reliabilitymodels and neural network structure and characteristics we pointed out theirdeficiencies. The low speed of learning is a main problem for the BP neural networksin fault diagnosis systems. On the basis of others theory and results of softwarereliability prediction on neural network, the momentum term and variable studyingrate are increased to the standard BP algorithm, and LM algorithm is used to train theneural network because the algorithm can solve problem of velocity effectively, anddesigning Optimally of the network structure and numbers of the hidden layer nodes.We come up with the better software reliability prediction model based on ANN withthe help of many mathematic tools such as the Zone-Probing method, the Searchmethod gradually, data scaling, cross validation and so on.At last, Simulation shows that compared with the classic models, the new modelhas better prediction precision, better generalization ability and lower dependence on the number of samples.for the further verification of this model we apply it to somecellular phone software to predict the software reliability of it.更多还原