Session I: Digital Twins for City Infrastructure (2:00 PM - 4:00 PM, Tues, Sept. 17, 2019)
Abstract: Digital Twin (DT) has been developed and applied in manufacture industry for more than a decade but DT is relatively new for civil infrastructure engineering. This talk will briefly examine the original DT framework with the technical requirements for infrastructure DT and report on the applied research in developing AI-based methods and tools for constructing high-fidelity digital twins, which are applied to infrastructure inspection, performance modeling and near real-time operation management to improve infrastructure resilience.
communities.bentley.com/.../274595
Abstract: Reality capture and VR/AR technologies are transforming ways of managing cities with the end goal of creating smart and resilient communities. In this talk, we will present our research work in large-scale urban mapping and modeling in the context of disaster resilience and smart cities. In particular, we will share our findings on how to use AI, cloud computing, and BIM/GIS technologies to create digital twins that can be used to assess disaster impacts, enhance urban mobility, and simulate facility operations.
communities.bentley.com/.../274596
Abstract: It is challenging for travelling public, especially the underserved populations including those with visual impairment, Autism Spectrum Disorder (ASD), to easily navigate through large and complex facilities such as transportation hubs, sports arenas and exhibition halls. Supported by both the National Science Foundation and Bentley, our joint-team of CUNY, Rutgers and Lighthouse Guild aims at bettering the location-aware services by developing and applying smart digital twins of transportation hubs. Using Port Authority Bus Terminal (PABT) in New York City as example, our goal is to construct a PABT digital twin to monitor and analyze both the spatial semantics of and the activity (crowd, vehicles, etc.) statistics within the facilities. While huge amounts of multimodal data have been collected via various sensors, learning semantics and activities typically require the annotations of large amounts of the data, which is a daunting and expensive task. Our talks will present the research progress achieved this year for learning scene semantics and crowd statistics, from 3D or 2D data respectively, with minimal annotations, using deep learning approaches.
communities.bentley.com/.../274597
(1) Unsupervised Feature Learning for Point Cloud Understanding by Contrasting and Clustering with Graph Convolutional Neural Network. Ms Ling Zhang, Master of Computer Engineering, Grove School of Engineering, The City College, The City University of New York (Advisor: Professor Zhigang Zhu)
Abstract: To alleviate the cost of collecting and annotating large-scale point cloud datasets, we propose an unsupervised learning approach to learn features from unlabeled point cloud "3D object" dataset by using part contrasting and object clustering with deep graph neural networks (GNNs). The contrasting learning step forces a ContrastNet to learn high-level semantic features of objects but likely ignores low-level features, while the ClusterNet in the clustering step improves the quality of learned features by being trained to discover objects that belong to the same semantic categories by using cluster IDs. We have conducted extensive experiments to evaluate the proposed framework on point cloud classification tasks.
communities.bentley.com/.../274599
(2). Learning to Monitor Crowds with Minimal Data Using Semi-supervised GANs. Mr. Gregory Olmschenk, PhD Candidate, Department of Computer Science, The Graduate Center, The City University of New York (Advisor: Professor Zhigang Zhu)
Abstract: In this work, we demonstrate how state-of-the-art deep learning methods for crowd analysis (people counting) can be trained using minimal annotated data. Our semi-supervised approach with generative adversarial networks (GANs) can also be applied more generally to a wide variety of learned tasks. For the purpose of crowd analysis, our method allows any facility to use the latest network architectures with minimal data annotation specific to their site.
communities.bentley.com/.../274598
Session 2: Digital Twins for Improving Infrastructure Health (2:00 PM – 3:30 PM, Thurs Sept. 19, 2019)
Abstract: This presentation will provide an overview of recent projects involving calibration of the digital twin of a long span cable stayed bridge, and the automated damage detection of a typical steel girder bridge using its digital twin. A calibrated digital twin can be used to supplement the biennial inspections of the bridge, yield more accurate load ratings, enhance permit vehicle evaluation, verify integrity following a rare or extreme event, or automate the process of damage identification and localization. Both real-world case studies illustrate the effective ways in which bridge digital twins can be used to more effectively manage and maintain our bridge infrastructure.
communities.bentley.com/.../274600
2. Use of Computer Vision for Traffic Load Estimation and Bridge SHM, Maria Feng, Renwick Professor of Civil Engineering, Columbia University, New York City.
Abstract: This talk will focus on the use of computer vision to monitor both vehicle loading and bridge response to create a near real-time digital twin model that reflects the current structural conditions of the bridges. To construct such a digital twin requires for identifying bridge model parameters, which are often obtained by performing system identification based on the system output (bridge vibration responses) because the system input (traffic excitation) is difficult to measure. To facilitate the identification, traffic excitation is commonly modeled as spatially uncorrelated white noise, but this assumption is invalid for majority of highway bridges. This talk presents two novel computer vision-based methods for estimating vehicle loads. The first method is stochastic modeling of traffic excitation on bridges based on vehicles’ arrival time and speeds measured by computer vision. The second method is a direct measurement of vehicle loads based on their tire deformations captured by computer vision. The talk further demonstrates the integration of the traffic excitation input with bridge response output for Structural Health Monitoring (SHM).
communities.bentley.com/.../274601
3, Infrastructure Digital Twins of Power Distribution System for Resilient Community Wei Zhang, Associate Professor, Department of Civil and Environmental Engineering, University of Connecticut.
Abstract: Power distribution system usually experiences severe damages during extreme weather events. Power outages that lasted for several days or weeks could significantly affect communities. To better understand the in-service conditions and the history of power distributions systems including poles and wires, infrastructure digital twins are developed using point cloud data and photogrammetry to first construct the 3D reality model, which is then updated with the operational data and converted into a finite element model by assigning material properties and types of connection between different elements. Thus, the high-fidelity digital twins for the power distribution system are established and can be applied to assess different operation and maintenance options, such as tree trimming and replacement of poles, for various weather events to achieve the reliable decisions for resilience enhancement.
communities.bentley.com/.../274602
Abstract: Smart cities can sense, understand their environment and the people. Direct sensing of desired events is difficult and often impossible due to deployment difficulties, lack of available sensors, and cost of maintenance. This talk will examine utilizing existing both mobile and fixed infrastructure sensing to infers information of the environment and individuals in the buildings. Physical information is used to guide the learning models to ensure the models are explainable and can be trained with less amount of data required. The talk explores two examples. 1) micro drones to enter and sense the building response from inside the building. 2) Infrastructure-based sensors to infer occupant location and identity.
communities.bentley.com/.../274604
2. Automated Decision Support for Flood Risk Mitigation Using Image-based Deep Learning, Mohammad Jahanshahi, Assistant Professor of Civil Engineering, Purdue University
Abstract: Floods are the most common and most damaging natural disaster worldwide, both in terms of economic losses and human casualty. Assessing flood risk is important but challenging. In this study, an interactive decision support system is developed to tackle the grand challenge of increasing flood risks. This study proposes a vision-based approach using deep learning that can collect comprehensive data effectively and efficiently without human-involved street surveys. With the proposed feature fusion network, the proposed approach can predict foundation type and height, building type and building stories as well as square footage. Individual property owners can use the tool to decide what measures to take to protect their assets, and local planners and policy makers will be able to develop “adaptation timelines” that prioritize infrastructure projects using estimates of current and future risk.
communities.bentley.com/.../274603
Abstract: Classifying and detecting seismic-vulnerable buildings, such as soft story buildings, is imperative for structural retrofit to reduce seismic losses and improve city resilience. This talk will present the application of deep learning with iModel.js and 3D models for detecting and visualizing the classified buildings. The approach aims at improving the accuracy, efficiency and visualization of the building classification. The trained model has been tested on several cities in California and the results are compared with those previously obtained by image-based deep learning approach.