Neural Networks Modeling for Software Detected and Corrected Fault Prediction

摘要

Traditional software reliability models describe the software fault detection process (FDP), assuming faults are corrected immediately once detected. With respect to this impractical assumption, these models have been extended for incorporation of the fault correction process (FCP) as well. Extended analytical SRGMs assume FCP as a delayed FDP, which clearly simplifies the relationship between these two processes. On the other hand, data-driven artificial neural network (ANN) models are welll known for its flexibility and assumption relaxation. Extended ANN models for FDP & FCP have been developed. However, different kinds of networks have different performance, and model selection is also a critical issue for ANN models. Accordingly, in this paper a general ANN modeling framework for FDP&FCP is proposed, representing the popular feedforward, recurrent, and radial basis function networks (FFNN, RNN, RBF). Furthermore, a cross-validation scheme is proposed to compare the performance of different networks. Experimental study is conducted with two simulated datasets. Also, the practical application issue of early prediction is discussed.