Long-term health monitoring of concrete and steel bridges under large and missing data by unsupervised meta learning

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 se-lection 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 fre-quencies 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.

Automated summary

Long-term monitoring is crucial for health monitoring of civil structures as it covers all possible unpredictable variations in vibration data and provides adequate training samples for unsupervised learning. However, it may encounter large data with missing values and erroneous results due to environmental changes. To address these challenges, this article proposes a novel unsupervised meta-learning method involving four steps of data analysis, segmentation, subspace searching, and anomaly detection.