The event correlation analysis is a knowledge extraction method that detects statistical similarities among the discrete events of alarms and operations. The method uses operation data from a plant to quantify the degree of similarity between events separated by a time-lag by evaluating the cross correlation function. If high similarity between two events is not detected, the time window size is doubled, and the log data of two events are reconverted into sequential binary data using the new time window size. The expansion of the time window size and recalculation of the similarity continues until either high similarity is detected or the time window size becomes larger than the maximum pre-determined size. We applied the method to the plant operation data of an ethylene plant. The results showed that it was able to correctly identify similarities between two physically related events, even when the conventional method using a constant time window size failed due to the large variance in time lag. Using the method, we can effectively identify nuisance alarms and routine operations within a large amount of event data.