Novel Sensor Clustering–Based Approach for Simultaneous Detection of Stiffness and Mass Changes Using Output-Only Data

Abstract

This paper presents a novel sensor clustering-based time series approach for anomaly detection. The basic idea of this approach is that localized change in the properties of a structure may affect the relationship between the accelerations around the position where the damage occurs. Therefore, for both healthy and damaged (or unknown state) structures, autoregressive moving average models with eXogenous inputs (ARMAX) are created for different clusters using the data from the sensors in these clusters. The difference of the ARMAX model coefficients are employed as damage features (DFs) to determine the existence, location, and severity of the damage. To verify this approach, it is first applied to a 4-DOF mass spring system and then to the shear type IASC-ASCE numerical benchmark problem. It is shown that the approach performs successfully for different damage patterns. It is also demonstrated that the approach can not only accurately determine the location and severity of the damage, but can also distinguish between changes in stiffness and mass.