Data-driven structural diagnosis and conditional assessment: from shallow to deep learning

Abstract

As compared to conventional physics-based techniques, advances in sensors and computing technologies have been promoting data-enabled structural diagnosis and conditional assessment using machine learning techniques in structural health monitoring (SHM). Machine learning helps civil engineers to extract valuable information from large amount of data to make time-sensitive decision. The application of different machine learning techniques to large-scale civil structures is, however, still impeded by challenges. In this study, we use representative supervised support vector machine (shallow learning) and deep Bayesian deep belief network (deep learning) to demonstrate their merits and limitations in structural diagnosis and conditional assessment. A benchmark in the literature is used for the demonstration. The results showed that the shallow learning highly relies on the hand-crafted features, while optimization of kernels is another challenge during learning process. The deep learning could promote the learning accuracy without kernel design. Although the noise could lead to difficulty in data mining, the comparison demonstrated that the deep learning has less sensitivity to the impacts of noise interference than those of shallow learning.