Buildings, bridges, power grids, and other infrastructure are essential parts of the urban system. It is critical to diagnose their health conditions continuously. The availability of low-cost sensing, wireless communications, and data processing technologies has made it possible to deploy large-scale sensor networks to monitor urban infrastructure. The data from these networks offer new opportunities to integrate spatial and temporal information for infrastructure health monitoring. In the first part of this talk, we design statistical and machine learning based methods to recover the underlying infrastructure network and its connectivity by utilizing probabilistic graphical to model infrastructure damage and sensor measurements. We illustrate the approach by demonstrating a provably correct method to recover the power distribution network topology. The proposed method is validated using 110,000 customers’ data and over 30 distribution grids with sizes varying from 113 to 3000 buses.
In the second part of this talk, we focus on designing algorithms for detection and localization of infrastructure damage. Two practical problems of damage identification are 1) the post-damage information is usually unknown and 2) multiple types or locations of damage may present simultaneously. To tackle these challenges, we propose utilizing graphical models to create algorithms that can learn the post-damage distributions and cooperatively detect multiple faults leading to localizing damage in space and time. A cooperative Bayesian damage detection and localization method is shown to isolate the damage location in an optimal way and can be even computed in a distributed manner. The methods are tested utilizing various datasets collected from experiments with buildings subject to different ground motions in shake table tests. In infrastructure systems, knowledge transfer is possible because each system is governed by physical laws, which are many mathematical similarities though in different domains. We demonstrate that, with simple modifications, the damage detection method for buildings can be applied to detect distribution grid power outages as well.
At last, we demonstrate that presented theoretical works are integrated into a wireless infrastructure health monitoring system, SnowFort. SnowFort is an open source wireless sensor network and data analytics system designed for infrastructure and environmental monitoring. With optimized communication architecture, the wireless sensor network can easily scale up to hundreds of sensors and each sensor can operate between 125 and 244 days with two AA batteries. Utilizing cloud computing technology, SnowFort supports real-time infrastructure system health diagnosis and data visualization. We demonstrate how to use SnowFort to monitor a steel structure health conditions in shake table experiments.
Yizheng Liao is a visiting scholar in the Department of Civil and Environmental Engineering (CEE) at Stanford University. He received the Ph.D. degree in CEE in 2018 and the M.S. degree in Electrical Engineering in 2013, both at Stanford University. Prior to Stanford University, he received the M.S. and B.S. degrees in Electrical and Computer Engineering at Worcester Polytechnic Institute in 2011. He is the receipts of the best student paper gold award of 2016 IEEE International Conference on Probability Methods Applied in Power Systems and the best paper award of 2016 ASCE Engineering Mechanics Institute Conference. He also received the Charles H. Leavell Fellowship and Gerald L. Pearson Memorial Fellowship.