Data-Driven Adaptive Nearest-Neighbor Reconstruction for Early-Stage Damage Detection in Structures Subjected to Changing Environments

The utilization of intelligent damage detection techniques that rely on modal frequencies has resulted in remarkable progress in civil structural health monitoring (SHM). However, changing environments often induce nonlinear and non-Gaussian modal variabilities with high dimensionality, obscuring damage intensities. Under such circumstances, traditional SHM methods erroneously classify structures as undamaged. This paper presents an adaptive nearest-neighbor reconstruction difference model to enhance the detectability of early-stage structural damage by alleviating environmental effects. First, nonlinear slow features of high-dimensional frequency data are extracted via exponential slow feature analysis under multisource heterogeneous environmental conditions. A neighborhood set of slow features is then automatically constructed using the k-nearest-neighbor algorithm, followed by the computation of the reconstruction and differential values based on the mean and standard deviation of the nearest neighbors. A dynamic linear update model designed to implicitly remove non-Gaussian frequency variability induced by seasonal temperature is established. Robust Mahalanobis squared distance indicators and thresholds are subsequently defined to capture slight frequency variability. The validity of the proposed early damage detection methodology is verified using the Z24 Bridge. Results indicate that this methodology that requires no additional hyperparameter tuning or prior environmental measurements can effectively eliminate environmental effects by transforming nonlinear and non-Gaussian modal variabilities into linear correlations and unimodal distributions, thereby providing a Type I error rate of 0.35% and a Type II error rate of 0%in early-stage damage detection, outperforming traditional methods.