A Multi-stage Machine Learning Methodology for Health Monitoring of Largely Unobserved Structures Under Varying Environmental Conditions

Structural Health Monitoring (SHM) via data-driven techniques can be based upon vibrations acquired by sensor networks. However, technical and economic reasons may prevent the deployment of pervasive sensor networks over civil structures, thus limiting their reliability in terms of damage detection. Moreover, the effects of environmental (and operational) variability may lead to false alarms. To address these challenges, a multi-stage machine learning (ML) method is here proposed by exploiting autoregressive (AR) spectra as damage-sensitive features. The proposed method is framed as follows: (i) computing the distances between different sets of the AR spectra via the log-spectral distance (LSD), providing also the training and test datasets; (ii) removing the potential environmental variability by an auto-associative artificial neural network (AANN), to set normalized training and test datasets; (iii) running a statistical analysis via the Mahalanobis-squared distance (MSD) for early damage detection. The effectiveness of the proposed approach is assessed in the case of limited vibration data for the laboratory truss structure known as the Wooden Bridge. Comparative studies show that the AR spectrum is a reliable feature, sensitive to damage even in the presence of a limited number of sensors in the network; additionally, the multi-stage ML methodology succeeds in early detecting damage under environmental variability.