Embedded Files
Featured Projects
Featured Projects
Classification of soft-story buildings using deep-learning models with 3D point clouds.
Classification of soft-story buildings using deep-learning models with 3D point clouds.
This project originated from the work done by Peng-Yu during his summer internship at Bentley Systems during 2019. An automated workflow was developed to segment point-cloud data at the city of Santa Monica (CA), extract the density features of point clouds, and classify seismically vulnerable buildings through deep neural networks. Transfer Learning techniques were used to optimize the classification performance. The results have been recently submitted to Journal of Computing in Civil Engineering and is being reviewed.
Development of regional seismic assessment toolbox for non-ductile reinforced concrete buildings in Los Angeles.
Development of regional seismic assessment toolbox for non-ductile reinforced concrete buildings in Los Angeles.
An automated workflow for regional seismic assessment of non-ductile reinforced concrete buildings was developed. The proposed workflow includes 5 components: (1) selection of ground motions given the site-specific hazard level; (2) design of pre-1970s concrete buildings based on 1967 Uniform Building Code; (3) generation of detailed finite element models in OpenSees; (4) computation of fragility functions and evaluation of building damages including collapse; (5) estimate corresponding seismic losses.
To automatically generate structure-specific analytical models for 1,500 buildings in Los Angeles, web scrapying technology was used through public data sources. Deep-learning models and transfer learning techniques were conducted to obtain building information from Google street-view images.
Part of the results have been recently submitted to Earthquake Spectra and is being reviewed.
Large scale building inventory using deep convolutional neural networks.
Large scale building inventory using deep convolutional neural networks.
Multiple deep neural networks were used in this project to collect building information from Google street-view images. For example, Faster RCNN with ssd_mobilenet was used to train a floor-detection model. VGG net and Inception net were used to train a classifier of building materials. Through the transfer learning technique, the training and inference can be done in couple of minutes via singe GPU.
Page updated
Google Sites
Report abuse