Early Damage Detection for Partially Observed Structures with an Autoregressive Spectrum and Distance-Based Methodology

Vibration-based Structural Health Monitoring (SHM) methods often rely upon vibration responses measured with a pervasive network of sensors. In some cases, it does not look possible for technical and economic reasons to equip civil structures with a distributed sensing system. Hence, the amount of information to handle for damage detection may be seriously affected by environmental and/or operational variability, leading to false detection results. To address this challenge, we present a parametric spectral method based on AutoRegressive (AR) modeling to set the damage-sensitive structural features. The spectra of the AR models associated with the normal and damaged conditions are collected into two matrices, to provide individual multivariate feature datasets in the frequency domain. By vectorising the matrices, two series of feature samples relevant to the normal and damaged conditions are obtained. To detect damage, the Logspectral distance method is adopted to measure the similarity between the two aforementioned feature vectors. The effectiveness and accuracy of the proposed approach are assessed through limited vibration data relevant to the IASC-ASCE benchmark problem. Results show that the AR spectrum stands as a reliable and sensitive feature for partially observed structures, hence in the case of limited sensor locations; additionally, the presented distance methodology succeeds in detecting early damage.