On the use of multivariate autoregressive models for vibration-based damage detection and localization

This paper proposes a novel method suitable for vibration-based damage identification of civil structures and infrastructures under ambient excitation. The damage-sensitive feature employed in the presented algorithm consists of a vector of multivariate autoregressive parameters estimated from the vibration responses collected at different locations of the analyzed structure. Outlier analysis and statistical pattern recognition are exploited for damage detection and localization. In particular, the Mahalanobis distance between a set of reference (i.e., “healthy”) and inspection parameters is evaluated. A threshold is then selected to determine whether the inspection vectors refer to damaged or undamaged conditions. The effectiveness of the proposed approach is proved using numerical simulations and experimental data from a benchmark test. The analysis results show that the largest values of Mahalanobis distance can be found in the proximity of those sensors closest to the damaged elements. Thus, the Mahalanobis distance applied to vectors of multivariate autoregressive parameters has proven to be a robust indicator for damage detection and localization.