Statistical Methods for Automated Detection of Extreme Events

摘要

In several previous reports, the authors have considered the problem of outlierdetection. In this report, several issues are addressed which are necessary to make our previous results broadly applicable in automated operational environments. This report basically consists of Appendix 1 and Appendix 2. The first appendix extends previous outlier detection methodology using the generalized likelihood function in such a manner that ground truth is no longer required. If the data vector contains several features, say for example 10, and several stations, say 8, observe the event, then the generalized likelihood ratio approach would require an estimate of an 80 x 80 covariance matrix, which will most likely not be feasible. In the second appendix, this problem is solved by compressing.the feature vector in such a way that the detection probabilities are still optimal and estimation of a large covariance matrix is no longer required. For instance, in the example suggested in the above, the 80 x 80 matrix would be replaced by an 8 x 8 matrix. Simulation studies show no loss in this method over the previous one.