Improving feature extraction via time series modeling for structural health monitoring based on unsupervised learning methods

Feature extraction by time series modeling based on statistical pattern recognition is a powerful approach to structural health monitoring. Determination of an adequate order and identification of an appropriate model play prominent roles in extracting sensitive features to damage from time series representations. Early damage detection under statistical decision-making via high-dimensional features is another significant issue. The main objectives of this article are to improve a residual-based feature extraction method by time series modeling and propose a multivariate data visualization approach to early damage detection. A simple graphical tool based on Box-Jenkins methodology is presented to identify the most compatible time series model with vibration time-domain measurements. Furthermore, k-Means and Gaussian mixture model clustering techniques are utilized to examine the performance of the residuals of the identified model in damage detection. A numerical concrete beam and an experimental benchmark model are applied to verify the improved and proposed methods along with comparative analyses. Results will show that these approaches are successful in obtaining a sufficient order superior to a state-of-the-art order determination technique, generating uncorrelated residuals, extracting sensitive features to damage, and accurately detecting early damage by high-dimensional data.