We deal with normal sequence data, for example, from real-time process control, which has been analyzed by the Cumulative Sum Test. The purpose of the paper is to give a new procedure for the change detection problem, which detects a change of the mean or the variance, and gives these estimators after the change. First, using Minimum Description Length (MDL) criterion, we propose a procedure which deals with the detection and the estimation consistently. Second, we give a method to adjust an error detection probability and a delay length of detection, which are in trade-off relation. Finally, the new procedure compares favorably with the Cumulative Sum Test.
View full abstract