Continuous Health Assessment of Bridges under Sudden Environmental Variability by Local Unsupervised Learning

Continuous health monitoring of civil engineering structures is an important process for ensuring their safety. However, sudden environmental variability makes this process erroneous and unreliable. To address this challenge, we propose a novel unsupervised learning method based on double data clustering. The central core of this method is to perform a data segmentation/clustering process in two levels by using a new clustering technique called local density peak clustering under minimum spanning tree (LDPC-MST). The main goal is to extract the most relevant information insensitive to environmental variations. In the first level of the double clustering algorithm, the LDPC-MST divides all available data points into main clusters. Subsequently, this approach is implemented to find subclusters within each main cluster and attempt to select one of them as the representative set, which contains the most relevant features. Using the representative subclusters of all main clusters, a damage detection indicator based on the Mahalanobis-squared distance is defined to detect any abnormal change caused by damage. The main innovation of this research is to develop a novel locally unsupervised learning method by using the process of double clustering and LDPC-MST. To validate this method, the natural frequencies of a concrete box-girder bridge and a steel arch bridge under strong environmental variations are incorporated. Several comparative analyses are also performed to indicate the superiority of this method over some well-known techniques. Results show that the proposed method can effectively warn the occurrence of damage with smaller rates of false positive, false negative, and total errors in comparison with state-of-the-art techniques.