Search

Confounding variability caused by environmental and/or operational conditions is a big challenge in the structural health monitoring (SHM) of large-scale civil structures. The elimination of such variability is of paramount importance in avoiding economic and human losses. Machine learning-aided data normalization provides a good solution to this challenge. Despite proper studies on data normalization using structural responses/features acquired from contact-based sensors, this issue has not been explored properly via new features, such as displacement responses from remote sensing products, including synthetic aperture radar (SAR) images. Hence, the main aim of this work was to eliminate environmental variability, particularly thermal effects, from different and limited structural displacements retrieved from a few SAR images related to long-term health monitoring programs of long-span bridges. For this purpose, we conducted a comprehensive comparative study to investigate two supervised and two unsupervised data normalization algorithms. The supervised algorithms were based on Gaussian process regression (GPR) and support vector regression (SVR), for which temperature records acquired from contact temperature sensors and structural displacements retrieved from spaceborne remote sensors produce univariate predictor (input) and response (output) data for the regression problem. For the unsupervised algorithms, this paper employed principal component analysis (PCA) and proposed a deep autoencoder (DAE), both of which conform with unsupervised reconstruction-based data normalization. In contrast to the GPR- and SVR-based data normalization algorithms, both the PCA and DAE methods only consider the SAR-based displacement (output) data without any requirement of the environmental and/or operational (input) data. Limited displacement sets of long-span bridges from a few SAR images of Sentinel-1A, related to long-term SHM programs, were considered to assess the aforementioned techniques. Results demonstrate that the proposed DAE-aided data normalization is the best approach to remove thermal effects and other unmeasured environmental and/or operational variability. [ABSTRACT FROM AUTHOR]

The development of satellite sensors and interferometry synthetic aperture radar (InSAR) technology has enabled the exploitation of their benefits for long-term structural health monitoring (SHM). However, some restrictions cause this process to provide a small number of images leading to the problem of small data for SAR-based SHM. Conversely, the major challenge of the long-term monitoring of civil structures pertains to variations in their inherent properties by environmental and/or operational variability. This article aims to propose new hybrid unsupervised learning methods for addressing these challenges. The methods in this work contain three main parts: (i) data augmentation by the Markov Chain Monte Carlo algorithm, (ii) feature normalization, and (iii) decision making via Mahalanobis-squared distance. The first method presented in this work develops an artificial neural network-based feature normalization by proposing an iterative hyperparameter selection of hidden neurons of the network. The second method is a novel unsupervised teacher–student learning by combining an undercomplete deep neural network and an overcomplete single-layer neural network. A small set of long-term displacement samples extracted from a few SAR images of TerraSAR-X is applied to validate the proposed methods. The results show that the methods can effectively deal with the major challenges in the SAR-based SHM applications. [ABSTRACT FROM AUTHOR]

• Suggesting a non-parametric method for early damage alerts using limited vibration data. • Focusing on unsupervised learning without relying on labeled data. • Introducing a new non-parametric clustering method with automated cluster selection. • Developing a novel density-based anomaly detector for damage detection without hyperparameter tuning. • Validating the method effectiveness on real-world bridges. Machine learning-assisted vibration monitoring is an intelligent, automated, and popular strategy for evaluating civil structures and damage alarming. However, implementing this strategy under a short-term monitoring program may encounter challenges such as limited vibration data, profound environmental and operational variations, and the limitations of state-of-the-art solutions under these conditions. The main purpose of this paper is to propose a novel machine learning technique in terms of unsupervised learning for damage alarming with limited vibration data. The crux of this technique lies in two fully non-parametric parts of data partitioning and anomaly detection. Initially, a non-parametric clustering approach with a novel procedure is presented to divide limited vibration data into clusters. Subsequently, a new density-based anomaly detector is developed to prepare indicators for damage alarming. Limited eigenfrequencies of full-scale bridge structures are used to validate the proposed solution. Results can substantiate its effectiveness and practicability in short-term monitoring programs. [ABSTRACT FROM AUTHOR]

• Proposing a novel unsupervised meta-learning method for long-term SHM with large data and profound environmental effects. • Proposing a locally unsupervised feature selection approach called nearest cluster selection. • Proposing a locally robust Mahalanobis distance via local metric learning and robust statistics for anomaly detection. • Accurate warning of damage occurrence under severe environmental variability caused by freezing weather. Long-term monitoring brings an important benefit for health monitoring of civil structures due to covering all possible unpredictable variations in measured vibration data and providing relatively adequate training samples for unsupervised learning algorithms. Despite such merits, this process may encounter large data with missing values and also yield erroneous results caused by severe environmental changes, particularly those emerge as sharp increases in modal frequencies during freezing weather. To address these challenges, this article proposes a novel unsupervised meta-learning method that entails four steps of an initial data analysis, data segmentation, subspace searching by a novel approach called nearest cluster selection, and anomaly detection. The first step intends to initially analyze measured data/features for cleaning missing samples. Next, the second step exploits spectral clustering to divide clean data into some segments. In the third step, the proposed nearest cluster selection is utilized to measure dissimilarities between the segments by a distance metric and select a cluster with the minimum distance as the representative of the main segment. Finally, a locally robust Mahalanobis-squared distance is applied by merging the concepts of robust statistics and local metric learning for online anomaly detection. The key innovations of this research contain developing a new unsupervised learning strategy alongside a locally robust distance and proposing the idea of nearest cluster selection. Long-term modal frequencies of full-scale concrete and steel bridges are used to verify the proposed method. Results demonstrate that this method succeeds in mitigating severe environmental effects and accurately detecting damage. [ABSTRACT FROM AUTHOR]