UC Berkeley

Civil structures, including bridges, tunnels, and skyscrapers, are becoming susceptible to losing their intended functionality as they deteriorate through the service life. Furthermore, this situation is exemplified in the face of natural hazards and extreme events. Therefore, monitoring and rapid reconnaissance of the condition and health states of such structures is important for effective decision-making towards building more resilient infrastructure systems. Traditionally, such monitoring and reconnaissance efforts require onsite human inspection. However, given the growth of buildings and other infrastructure in urban centers, such an inspection process is infeasible because of limited human resources, financial burden, and time consuming efforts.

In this dissertation, methods to automate the monitoring and reconnaissance processes are proposed. First, methods are introduced to automate the data collection process, where information, that are highly relevant to the health states of structures as well as the entire infrastructure systems, are collected. Second, algorithms are introduced to automate the data processing, where results regarding the health states of structures and infrastructure systems are obtained. Such results are essential for the decision-making process to increase resiliency of the infrastructure systems. The most important technique to automate the above processes is Artificial Intelligence (AI), in particular, Machine Learning (ML) algorithms.

In the case of a single structure, the process of observing its response and determining its health state is called Structural Health Monitoring (SHM). A novel SHM framework utilizing Deep Learning (DL) is proposed. The framework is based on Long Short-Term Memory (LSTM) Encoder-Decoder architecture, which is a variant of the Recurrent Neural Network (RNN), applied to Time Series (TS) data. The TS data is processed through the LSTM network, where the information in the TS data is condensed into a Latent Space Vector (LSV), which is processed through traditional ML algorithms to output the structural health conditions, including the overall health conditions, the locations and severity of damage. To enforce the encoding (i.e., condensation) process of the TS into the LSV without information loss, an Encoder-Decoder architecture is proposed. Moreover, a method for fast prediction of the structural responses, which uses variants of the LSTM network, as well as a novel network called Temporal Convolutional Network (TCN), is proposed, and these models (variants) are compared against each other in terms of the accuracy of predicting the structural response. The proposed models are anticipated to complement/replace the traditional physical simulations for faster prediction of the structural response when immediate results are required, e.g., for rapid decision-making.

On the data collection side of a single structure, the quality of data obtained from the sensor network is critical to the diagnosis (i.e., determination of the health conditions of the structure). If the sensors are not placed on locations that are sensitive enough to the structural damage, the collected data is not useful for the purpose of diagnosis. In this dissertation, an Optimal Sensor Placement (OSP) method is proposed. The causal relationship among the sensor recordings is identified and quantified through Directed Information (DI). In this method, the sensors are added sequentially, i.e., one sensor at a time, until the specified number of sensors (typically based on expert opinion and availability of resources) is satisfied. The new sensor is added at a location where the causal relationship with the existing sensors is the lowest to ensure low redundancy of the information stored in the array of sensors.

For the case of infrastructure on a regional (e.g., city) scale, a method to effectively collect reconnaissance results following an earthquake event is proposed. Social media posts by people near the source of the earthquake, news reports, as well as information from official resources, e.g., United States Geological Survey (USGS), are collected automatically following the earthquake event. Such information is subsequently summarized as a briefing, which provides valuable reference for further detailed reconnaissance (field investigation) and emergency response. The Natural Language Processing (NLP) method is adopted in the formulation of these briefings. Moreover, a practical method to quantify the regional recovery state (a step towards quantifying a metric for the resilience of the affected community) following the earthquake event is proposed. This is based on the number of relevant posts collected from the social media. The recovery is quantified as the averaged recovery states of several key aspects, e.g., water supply to the community, electricity supply to the community, and availability/resumption of the functionality of essential facilities, e.g., medical services by hospitals.