Automatic Density-Based Clustering for Operational Modal Analysis

Abstract

Estimation of modal parameters from ambient response measurements is a central task in structural health monitoring and rapid condition assessment of structures after natural disasters or other damaging events. This task has traditionally required considerable interaction from the user. Automation of this task enables online assessment of the integrity of structures, increases the accuracy of results by removing user error, and reduces analysis time and associated costs. This paper proposes an unsupervised approach for automatic extraction of modal parameters from measured vibration data. A novel heuristic to automate an existing clustering algorithm called density-based spatial clustering of application with noise was introduced and validated. This heuristic uses a histogram as a nonparametric density estimator and is applicable to data sets containing arbitrarily shaped clusters. The automated clustering procedure can be used with any output-only system identification method that produces modal estimates over a range of model orders. An application was presented using numerical simulations of a 5-story shear frame model under ambient excitations. System identification was performed using covariance-based stochastic subspace identification, and modal estimates were obtained using the proposed approach. The modal estimation process was repeated using 200 independent realizations of structural responses. The accuracy of the predictions was investigated by comparing the predicted modal parameters to the theoretical values from eigenvalue analysis. The results demonstrate the promise of the proposed approach. Validation of the proposed method on real structures will be addressed in future studies.