Improving system identification using clustering

Abstract

System identification involves identification of a behavioral model that best explains the measured behavior of a structure. This research uses a strategy of generation and iterative filtering of multiple candidate models for system identification. The task of model filtering is supported by measurement-interpretation cycles. During each cycle, the location for subsequent measurement is chosen using the predictions of current candidate models. In this paper, data mining techniques are proposed to support such measurement-interpretation cycles. Candidate models, representing possible states of a structure, are clustered using a technique that combines principal component analysis and K -means clustering. Representative models of each cluster are used to place sensors for subsequent measurement on the basis of the entropy of their predictions. Results show that clustering is necessary to identify the different groups of candidate models. The entropy of predictions is found to be a valid stopping criterion for iterative sensor addition. Clustering helps classify models and, thus, it provides useful support to engineers for further decision making.