Non-parametric empirical machine learning for short-term and long-term structural health monitoring

Abstract

Early damage detection is an initial step of structural health monitoring. Thanks to recent advances in sensing technology, the application of data-driven methods based on the concept of machine learning has significantly increased among civil engineers and researchers. On this basis, this article proposes a novel non-parametric anomaly detection method in an unsupervised learning manner via the theory of empirical machine learning. The main objective of this method is to define a new damage index by using some empirical measure and the concept of minimum distance value. For this reason, an empirical local density is initially computed for each feature and then multiplied by the minimum distance of that feature to derive a new damage index for decision-making. The minimum distance is obtained by calculating the distances between each feature and training samples and finding the minimum quantity. The major contributions of this research contain developing a novel non-parametric algorithm for decision-making under high-dimensional and low-dimensional features and proposing a new damage index. To detect early damage, a threshold boundary is computed by using the extreme value theory, generalized Pareto distribution, and peak-over-threshold approach. Dynamic and statistical features of two full-scale bridges are used to verify the effectiveness and reliability of the proposed non-parametric anomaly detection. In order to further demonstrate its accuracy and proper performance, it is compared with some classical and recently published anomaly detection techniques. Results show that the proposed non-parametric method can effectively discriminate a damaged state from its undamaged condition with high damage detectability and inconsiderable false positive and false negative errors. This method also outperforms the anomaly detection techniques considered in the comparative studies.