Damage Detection Using Supervised Machine Learning Algorithms for Real-World Engineering Structures

Vibration-based structural health monitoring represents an efficient way to evaluate structural integrity and the presence of damage at an early stage. These methods usually assume that damage manifests itself as a deviation in the modal properties of the structure with respect to its normal conditions. Traditional procedures for modal parameters estimation require the use of a dense sensor arrangement and complex logic techniques, thus making them not particularly suitable for the case of large engineering structures, where the need for cost-effective monitoring solutions is of utmost importance because of the large number of substructures to be monitored. This paper proposes the use of simple statistical and spectral features as a mean to characterize accelerations signals. Starting from this set of features, the principal component analysis (PCA) is first used to reduce data dimensionality still preserving the relevant information about the structural conditions, then a k-Nearest Neighbors (k-NN) procedure is adopted as a supervised machine learning method to classify different types of damage. The procedure is validated using the experimental data from the permanent monitoring system of the G. Meazza stadium grandstands of, where one accelerometer per stand is installed to get vibration data during the main events. Four grandstands located on the same ring and having the same nominal geometry are considered. The leading idea is to reproduce different scenarios where, due to the impossibility of imposing realistic damages, one grandstand is assumed to be the safe structure, while the others represent a proxy for small structural changes to be identified