Data-driven methods for threshold determination in time-series based damage detection

Structural vibration monitoring has received a lot of attention from the research community in the past few years. The objective is to create automatic structural assessment techniques that can be realized through programmed vibration analysis. Till now many vibration-based damage features have been proposed, yet to truly automate the damage identification process, reliable damage threshold construction techniques are also need. In this paper, two data-driven methods based on resampling and nearest neighbor rule are applied for threshold construction for damage features from autoregression (AR) analysis of vibration signals. Both threshold calculation techniques are rooted in empirical feature probability estimation. The proposed thresholds are then tested on features extracted acceleration measurements collected from a 5 DOF test specimen. The resampling method is applied to Mahalanobis distance of AR model coefficients, while the nearest neighbor rule is used on a combination of coefficient distance feature and the residual autocorrelation feature. Both methods perform well in this case study.