Freezing weather can introduce challenges in long-term structural health monitoring of civil structures, particularly bridges. A noticeable impact of freezing temperature is the emergence of sudden and sharp increases in structural modal frequencies, causing false alarm and mis-detection errors in change detection of civil structures. This paper proposes an innovative unsupervised data normalization method to mitigate freezing effects. The proposed method integrates locally robust principal component analysis (LRPCA) with Gaussian density distance (GDD) clustering, called GDD-LRPCA, which automatically determines the number of clusters. Initially, a training set of original modal frequencies is partitioned via the GDD clustering. Subsequently, an individual LRPCA model is fitted to each partition to extract new normalized modal frequencies insensitive to freezing effects. The groundbreaking nature of this research relies on developing an integrated unsupervised data normalizer by leveraging advanced machine learning algorithms such as local learning, robust learning, and hybrid unsupervised learning. The major advantage of the proposed method is its non-parametric nature obviating any supplementary technique for hyperparameter optimization. The validity of this method is benchmarked by real-world bridge structures along with several comparative analyses. Results demonstrate that GDDLRPCA effectively removes the freezing effects from structural modal frequencies and outperforms its counterparts in unsupervised data normalization.
Loading....