Unsupervised Condition Monitoring of Structures Using VMD and Isolation Forest

Abstract

In this paper, an unsupervised condition monitoring of civil infrastructures under environmental and operational variations is proposed. To this end, a couple of the lowest natural frequency signals of the structure recorded over a long period of time are studied for damage. Most of the existing techniques are supervised and require baseline information from the healthy state of the structure. In this paper, however, we explore the possibility of performing unsupervised condition monitoring using Isolation Forest (IF) as a well-known unsupervised machine learning approach. We show that a preprocessing of the frequency signals using the Variational Mode Decomposition (VMD) algorithm plays a key role in the success of the IF algorithm in the unsupervised condition monitoring of structures. To investigate the performance of the proposed method, the benchmark problem of the Z24 bridge is studied in this paper. The results show that the proposed methodology stays successful in most of the cases but at a period of very cold temperature where a nonlinear relationship between the frequencies and temperature presents.