Machine learning has become an influential and useful tool for many civil engineering applications, particularly structural health monitoring (SHM). For this reason, this article aims to propose a novel machine learning method in terms of unsupervised information-based anomaly detection for SHM under long-term and short-term monitoring. The crux of this method is to define a new anomaly score or information content by using the concepts of local density, unsupervised feature selection via a one-class nearest neighbor rule, local cutoff distance, and minimum distance value. A non-parametric approach is then proposed to choose adequate nearest neighbors of each feature and determine its local cutoff distance and local density. To detect damage through the concept of anomaly detection, it is necessary to compare the anomaly scores with an alarming threshold. Accordingly, a new probabilistic method under semi-parametric extreme value (SEV) theory is proposed to estimate a quantile from some extreme samples and use this quantile value as a threshold limit. Due to the importance of selecting adequate extreme samples, this process is carried out by a two-stage iterative approach. The major contributions of this article include: (i) proposing a novel unsupervised learning method in a non-parametric fashion, (ii) defining a new anomaly score for SHM applications, (iii) advancing an unsupervised nearest neighbor selection, and (iv) developing a new probabilistic threshold estimation based on the SEV theory without any model selection, modeling procedure, and parameter estimation. Dynamic and statistical features extracted from vibration data of two full-scale bridges are applied to validate the proposed methods along with several comparative studies. Results demonstrate that the methods presented here are effective and reliable tools for accurate SHM via high- and low-dimensional features with superiority over some well-known techniques.