To avoid unpredictable losses because of network failure, the reliability of the network needs to be evaluated in some application scenarios. This paper start the network failure prediction research upon 14 months' network alarm logs we collected. The logs are of one Metropolitan area network. The research method is shown as below: firstly, construct features to represent network characteristics by the means of the feature construction method which is based on two levels time windows; secondly, select optimal parameter combination to create the feature files through multiple experiments; thirdly, design and build adaptive failure prediction model according to classification learning methods. Numbers of experiments show that accuracy of predicting whether the network failure takes place in 6 hours is up to 70%, is better than the prediction result of Weibull distribution model obviously; the results of classification prediction for network equipment failure are slightly better than the prediction method on the basis of Weibull distribution. Preliminary research results show that most network failures can be predicted through analyzing previous network running logs and the method proposed in this paper is verified to be with good prediction effect. This method can detect failures in practical application on early stage and reduce unnecessary economic losses.