Bayesian-optimized unsupervised learning approach for structural damage detection

Structural health monitoring (SHM) is developing rapidly to fulfill the world's need for resilient and sustainable communities. Due to the current advancements in machine learning and data science, data-driven SHM is an attractive solution for real-time damage detection compared to the traditional nondestructive evaluation techniques. However, most widely available data-driven SHM methods rely on fully or partially simulated data to train the statistical model, and thus require a number of predefined assumptions and parameters, or are not adapted for post-extreme events damage diagnosis. In this study, we propose a density-based unsupervised learning approach for structural damage detection and localization. This approach leverages cumulative intensity measures for damage-sensitive feature extraction for the first time in an unsupervised learning approach. Furthermore, a statistical model construction process is proposed based on kernel density maximum entropy (KDME) and Bayesian optimization. The framework is evaluated in three case studies. The first two involve a numerical three-story building and a numerical nine-story asymmetrical building that are both subjected to 100 ground motion excitations while considering environmental variations. The proposed framework is able to detect and localize damage in those case studies with an average accuracy of 92%. The third case study, which contains 44 shake-table tests of a three-story frame structure with masonry infill, is used to experimentally validate the proposed framework in damage detection. The three case studies demonstrate the potential and robustness of the proposed Bayesian-optimized, multivariate KDME novelty detection framework for detecting and localizing structural damage, especially after extreme events.