Alarm systems that warn operators when abnormal conditions arise are important for the safe operation of chemical plants. Nuisance alarms, such as alarm floods or repeating alarms, cause operators to misinterpret plant conditions, which may result in accidents. Event correlation analysis detects nuisance alarms buried in the noise of a plant's event log data by detecting similarities among their occurrences. However, this method fails to detect the similarity between two physically related events when the variance in time lag between them is large. We propose a new method of event correlation analysis in which the time window size is expanded gradually according to evaluation results of similarities. The usefulness of the proposed method is demonstrated through a case study of an azeotropic distillation column at a chemical plant.