Improving SHM multi-structure adaptability via reduced a priori information
Rodrigo Sarlo, Alan Smith, and Mohammad-Hesam Soleimani-Babakamali
One of the current challenges in structural health monitoring (SHM) is the design of methods that are flexible enough to scale to a variety of as-built structures. Current methods have limited scalability partly because they tend to rely on significant a priori information, e.g., an assumed model/geometry of the structure, training examples of specific damage cases, etc. Not only is such information sometimes difficult to obtain, it also results in a number of specific challenges that will be postulated as part of this talk. These include: 1) deviations of the as-built behavior from the assumed “as-planned” behavior and 2) limits on an SHM algorithm’s ability to deal with new or evolving damage or environmental scenarios. Based on these challenges, the talk will introduce a set of on-line monitoring methodologies which aim to reduce reliance on a priori information while accounting for the numerous variations present in the true built environment. The first challenge is addressed through the conversion of dense reality capture date (e.g., from LiDAR) to workable structural models, enabling the analysis of as-built geometries that evolve over the structure’s lifetime. The second challenge is addressed via an unsupervised SHM framework which exploits simple, high-dimensional features in combination with generative adversarial networks (GANs). Synthetic data from the GANs is analyzed using system reliability principles in order to adaptively define damage detection thresholds completely online. Both frameworks will be demonstrated on a variety of full-scale experimental structures to support their scalability potential.