Environmental variability is a major challenging issue in bridge health monitoring because bridges are more prone to such variability than other civil structures. To deal with this challenge, this article proposes a new machine-learning method for early damage detection under environmental variability by means of the k-medoids clustering, a new damage indicator, and an innovative approach for selecting a proper cluster number. Estimation of a reliable alarming threshold is another important challenge for early damage detection via most of the machine-learning methods. On this basis, a novel probabilistic approach using the theory of extreme value and a goodness-of-fit measure is proposed to estimate an alarming threshold. The major contributions of this article include proposing a new damage indicator suitable for decision making by clustering-based algorithms, an innovative cluster selection algorithm for dealing with the problem of environmental variability and increasing damage detectability, and a novel probabilistic method for threshold estimation. Modal-based features of the well-known Z24 Bridge are considered to verify the accuracy and effectiveness of the proposed approaches along with several comparative studies. Results show that the methods presented here are highly able to detect early damage even under strong environmental variations and estimate a reliable threshold.

环境可变性是桥梁健康监测中的一个主要挑战性问题，因为桥梁比其他民用建筑更容易出现这种可变性。为了应对这一挑战，本文提出了一种新的机器学习方法，可以利用k-medoids聚类，新的损坏指标以及用于选择合适聚类编号的创新方法。通过大多数机器学习方法来估计可靠的警报阈值是早期发现损坏的另一个重要挑战。在此基础上，提出了一种利用极值理论和拟合优度度量的新概率方法来估计警报阈值。本文的主要贡献包括：提出一种适用于基于聚类算法进行决策的新损伤指标，一种用于解决环境变异性和增加损伤可检测性的创新聚类选择算法，以及一种用于阈值估计的新型概率方法。考虑了著名的Z24桥的基于模态的功能，以验证所提出方法的准确性和有效性以及一些比较研究。结果表明，即使在强烈的环境变化下，此处介绍的方法也能够高度检测早期损坏并估算可靠的阈值。