Bridge collapse prediction by small displacement data from satellite images under long-term monitoring

Long-term structural health monitoring (SHM) is an effective approach to continuously assess the health and safety of civil structures. In most vibration-based SHM applications based on contact-based sensors, a long-term SHM strategy contains a large volume of measured data with high-dimensional data samples. Therefore, the course of dimensionality and Big Data is one of the major challenges in SHM. Thanks to developments of satellite sensors and synthetic aperture radar (SAR), it has been possible to exploit their benefits for SHM of civil structures via satellite images as remote sensing. The great advantage of the SAR-based SHM strategy is the possibility of obtaining prior information of civil structures regarding their normal conditions. However, some restrictions cause this process provides a small number of images leading to the problem of small data for the SAR-based SHM strategy. On the other hand, the major challenge in most of the long- term SHM projects is related to variations in their inherent properties resulting from environmental and/or operational variability (EOV). This issue is challenging and will be worse in the SAR-based SHM application when a small number of data is present. Therefore, the main objective of this study is to propose effective and efficient multi-stage unsupervised learning methods for addressing the aforementioned challenges. The proposed method contains three main parts including: (i) data augmentation by a Markov Chain Monte Carlo (MCMC) algorithm, (ii) feature normalization for removing the EOV conditions, and (iii) decision-making for pre-collapse prediction via Mahalanobis-squared distance (MSD). The major innovation of this method concentrates on feature normalization, which is based on the concepts of machine learning, unsupervised learning, artificial neural network (ANN). In the proposed method called MCMC-ANN-MSD, an effective iterative hyperparameter selection is proposed to choose the neuron sizes of the hidden layers with the focus on dealing with the EOV effects. A small set of displacement samples extracted from a few satellite images of TerraSar-X regarding a long-term monitoring scheme of the Tadcaster Bridge in United Kingdom is applied to validate the proposed method. Results show that the proposed multi-stage unsupervised learning methods effectively deal with the major challenges in the SAR-based SHM and provide practical tools for real applications.