报告简介:

Structural Health Monitoring (SHM) and rapid damage assessment after natural and man-made hazards are important focuses within Civil Engineering in recent years. Meanwhile, Artificial Intelligence (AI) and Machine Learning (ML) technologies are developing rapidly, especially in the applications of Deep Learning (DL) for Computer Vision (CV). In this data explosion epoch, structural response records (time series), videos, and images as the data media are more ubiquitous and available than ever, leading to their more important roles in real-time analysis. There are several real project applications offering insights into the implementation of the state-of-the-art ML/DL technologies into SHM and post-disaster reconnaissance efforts.

In time/frequency domains, one highway bridge system was investigated. Based on nonlinear time history analysis simulations of the bridge system, a data-driven damage detection approach was explored for bridge columns, the most critical components of bridge systems. Damage feature selection was demonstrated through unsupervised learning. Subsequently, a Support Vector Machine (SVM) was applied for several classification problems of engineering interests. Beyond a bridge system in transportation infrastructure, a one 3-story braced steel frame structure was also studied for damage state detection and localization under different hazard levels. The floor accelerations from undamaged/damaged states were utilized as input data, the AutoRegressive Integrated Moving Average (ARIMA) modeling as a time series method was applied as a feature extractor, and finally Random forest/Decision Tree algorithms were used as classifiers for detection and localization.

In vision-based SHM area, the state-of-the-art DL techniques were implemented for image-based structural damage recognition using Convolutional Neural Networks (CNNs) with several architectures, i.e. VGG Net and Residual Net (ResNet). Selected from the Structural ImageNet, which contains over 10,000 images from reconnaissance efforts, a small experimental dataset was established with 2,000 manually labeled images based on domain knowledge. Addressing the issue of a small number of labeled images, instead of training CNN from scratch, the Transfer Learning (TL) approach was applied in two different ways with Feature Extractor (FE) and Fine-Tuning (FT). Four recognition experiments were conducted on different configurations of VGG Net and ResNet using both FE and FT. For a better understanding of how CNNs work, Class Activation Map (CAM) and Local Interpretable Model-agnostic Explanations (LIME) were implemented with visualization to explain the model’s performance.