Fault diagnosis based on sequential pattern mining

To find the rules of fault transfer and their mutual influences, sequential pattern mining technology was introduced into analysis of time series in monitoring system. A symbolizing method based on kernel density estimation was proposed. Characteristics of data distribution were used to transform multiple time series of consecutive numerical values into one symbol series so as to obtain suitable mining sequential pattern. After the data restriction, integration, and simplification, a sequence data set which was relevant to faults was formed. A sequential pattern mining algorithm was adopted to process the synthetic data of Tennessee-Eastman procedure. The resulting sequential patterns indicated the main change information of the production of fault. Test demonstrated the feasibility and validity of this method. It helped engineers to understand inherent interactional relationships in complex system in order to make reasonable diagnosis.