A Statistical Approach to Predict Operating System Failures Based on Multiple Failures Association
Empirical studies have shown robust evidence of OS failure patterns characterized by multiple combinations of failure events composed of the same or different failure types. In this paper, we present a statistical approach to predict OS failures based on multiple failures association. Once we identify systematic failure associations in field data, then we compute the probability of a given failure to occur within a time interval upon the occurrence of a particular pattern of preceding failures. Due to the nature of the failure data, in which the failure types must be handled as categorical variables, we used the logistic regression method to tackle our research problem-especially its variant multinomial logistic regression. Our approach was able to predict OS failures with good to high accuracy (81% to 95%). The resulting regression models proved robust enough to deal with different prediction time intervals with no degrading effect on their accuracy.
