Health monitoring of civil structures by continuous measurements of vibration data is one of the effective and reliable techniques for ensuring structural safety and serviceability. Envi-ronment changes and the methodology for damage assessment are two important challenges that play critical roles in achieving outstanding results due to the emergence of serious errors. This research intends to propose a new vibration-based method under the idea of locally un-supervised learning for addressing the challenges. The proposed method consists of two steps of data clustering and damage assessment. In the first step, a new clustering algorithm called locally undirected-based graph density peak clustering (LUG-DPC) is presented to split dy-namic features into pre-determined clusters and supply local information. The second step utilizes such local information to estimate local mean vectors and local covariance matrices that make the main elements of the anomaly detector based on the Mahalanobis distance. The major contributions of this paper contain developing an innovative machine learning-aided method and introducing the LUG-DPC for damage assessment Long-term continuous natural frequencies of a full-scale concrete bridge are used to verify the proposed method with some comparisons. Result indicates the proposed method can alleviate the environment effects and obtain reasonable results with inconsiderable errors compared to some well-known tech-niques