Online Failure Prediction For Software Systems

With the increase of requirements for computer software system, it has grown in complexityup to a point where its failure behavior is in part unpredictable. Accurate failure predictions canhelp in mitigating the impact of computer failures. However, the problems with many softwarefailure prediction methods are still far from providing satisfactory accurately. Of course,providing accurate predictions with sufficient lead of time is challenging.In this paper, a methodology is proposed to predict failure of software during runtime withneural network and to improve prediction accuracy. The core idea of our method is as follows:first of all, we use neural network to train the value of the variables into whole software systems,and to predict their values in the next cycle; then according to the influence of the relationshipbetween variables, to build function invoke diagrams, and to obtain the nonlinear functionrelationship between the variables by it; finally, we compare the predictive value of the variableswhether meet this nonlinear relationship with a certain range of error, if meet, it was consideredto be normal, otherwise, argue that failure occurred. In order to verify the feasibility of ourmethod, we mainly made two experiments: online failure prediction for failure time serial andcasco railway transportation system. In order to examine if changes in network architecture andthreshold can have positive effects on the results of online failure prediction, we from those twoaspects to improve our method, and through evaluation metric which are precision, recall,accuracy and F-Measure to evaluate and analysis the results of experiments. The results shownthat our proposed method can more accurately predict the failure occurred in the runtime of thesoftware system, and the results of experiments are very satisfactory.In this thesis, we have carried on the following three aspects of work:1) We propose to build a neural network model by variables influence diagram;2) We propose use the ELM learning algorithm to train the neural network, so as to avoid theproblem that the local optimization of neural network;3) We propose a parallel prediction method by use three neural networks, so as to improve theaccuracy of online failure prediction results.