Failure Prediction Of On-line Systems Based On Log Files

Despite the development of software testing technology is relatively mature, software failure and many other problems at runtimes have been still to occur after testing of software systems. Due to various uncertainties at online software system’s internal and external environment lead to system failure, such failures are difficult to accurately predict in a way. Failure prediction of the online system is to predict the behavior of software system in the operating state accurately in advance, in order to achieve an effective risk reduction of system failure and avoid major losses due to system failures caused. Thus, finding a way to predict online system failure behavior effectively still has practical significance and high research value.We propose a method of online failure prediction based on log files. Firstly, we extract variable data which related with system failure from the system log. We analyze these data and generate fuzzy rules combined with system requirements document. And then, a set of predictive values were obtained by time series prediction using variable data with the ARIMA model. Finally, time series prediction of the independent variable values gets the predicted fuzzy dependent variable values using fuzzy reasoning. According to the error time series prediction of the independent variable values and the predicted fuzzy dependent variable, we can evaluate the system failure behavior. In this paper, failure behavior of the system can be predicted by prediction time series and fuzzy reasoning. And then we evaluate and analyze the results of the final prediction according to the corresponding index. Experimental results show that the proposed method can effectively predict the behavior of the online system failure occurs, and be able to reach a satisfactory prediction.We have contributions as following three aspects:1) We propose a framework of online failure prediction based on time series prediction and fuzzy inference. First we obtain predicted values of all variables through time-series forecasting, then get the fuzzy reasoning predicted values according to the relationship among the variables. Finally we assess the failure behavior of the system by comparing two predicted values.2) We propose a method to generate fuzzy rules by combining bool variables with fuzzy variables. For the situations where both bool data and numerical data exist which significantly influencethe failure of the system, our method avoids losing a lot of useful information.3) The above methods are verified on CASCO rail system. Basing on the log data of the system, we verify the effectiveness of proposed framework. Meanwhile, we verify the feasibility of generating fuzzy rules which combines bool variable and numerical variables.