A Unified Distributed Spatiotemporal Signal Processing Framework for Structural Health Monitoring
Abstract:
Conventionally, structural health monitoring (SHM) routines are time-based and are carried out offline. So they cannot provide real-time situational awareness. More recently, online health monitoring systems start to appear thanks to the fast developing technologies in sensing, hardware miniature, embedded computing and communication. While sensing technologies could be dramatically different for various structures, the fundamental idea in detecting anomalies in a structure is very much alike: sensor measurements or features extracted from these measurements are compared with the predictions from analytical/numerical structural models or human experience. Any deviations indicate that some anomalies may exist in the structure. In this project, we aim at developing a unified framework that couples spatiotemporal sensing data with physics-based and data-driven models for SHM. Pure physics model based methods or pure data-driven approaches have their respective limitations. More sophisticated statistical modeling that can accommodate all available domain knowledge while take advantage of the huge amount of data collected by modern sensing systems is preferred. Sensors are usually deployed in a highly dense manner and observations are collected at rapid rates. Therefore, nearby sensors are likely to have similar measurements due to the spatial correlation. Their observations within a short period of time are temporally correlated. This makes the information contained in the spatial configuration and temporal coupling very precious for data processing. In this project, a dynamic Markov random fields (DMRF) based framework is proposed unifying both temporal and spatial information for statistical inference. Unlike FEM based models which provide precise values of all elements given boundary conditions, DMRF models the relations between elements in a structure. The neighborhood system, form of potential functions and parameters can be either manually constructed based on domain knowledge or learned from large training data, or combine both. Critical issues in wireless sensor networks such as communication costs are studied/optimized and the solution is integrated in the proposed framework. The developed framework finds applications in various areas. Cup anemometers are commonly used for wind speed measurement in the wind industry. Anemometer malfunctions lead to excessive errors in measurement and directly influence the wind energy development for a proposed wind farm site. Since the accuracy of anemometers can be severely affected by the environmental factors such as icing and the tubular tower itself, in order to distinguish the cause due to anemometer failures from these factors, our methodologies utilize the spatial relations (multiple anemometers at different locations) and temporal variations under the influence of environmental factors. In another application, the knowledge of the existing pavement condition is very important to the success of any rehabilitation project. The performance of a pavement depends on many factors such as the structural adequacy, the properties of the materials used, traffic loading, climatic conditions and the construction methods. Accurate prediction of the remaining service life (RSL) of pavements has always been a difficult problem. We approach this problem from several aspects. The procedures utilizing non-destructive testing methods, such as the falling weight deflectometer deflection studies, often result in relatively better quantitative assessment of the pavement condition. The spatial relations of deflection measurements at different locations are examined and a technique based on the similarity search method is adopted in predicting the RSL of pavements.
Two Ph.D. students (graduated in April and July of 2013 respectively) and two M.S. students (graduated in May and December of 2012 respectively) were supported by this project.
Submitted by Qi Cheng
on
Abstract:
Conventionally, structural health monitoring (SHM) routines are time-based and are carried out offline. So they cannot provide real-time situational awareness. More recently, online health monitoring systems start to appear thanks to the fast developing technologies in sensing, hardware miniature, embedded computing and communication. While sensing technologies could be dramatically different for various structures, the fundamental idea in detecting anomalies in a structure is very much alike: sensor measurements or features extracted from these measurements are compared with the predictions from analytical/numerical structural models or human experience. Any deviations indicate that some anomalies may exist in the structure. In this project, we aim at developing a unified framework that couples spatiotemporal sensing data with physics-based and data-driven models for SHM. Pure physics model based methods or pure data-driven approaches have their respective limitations. More sophisticated statistical modeling that can accommodate all available domain knowledge while take advantage of the huge amount of data collected by modern sensing systems is preferred. Sensors are usually deployed in a highly dense manner and observations are collected at rapid rates. Therefore, nearby sensors are likely to have similar measurements due to the spatial correlation. Their observations within a short period of time are temporally correlated. This makes the information contained in the spatial configuration and temporal coupling very precious for data processing. In this project, a dynamic Markov random fields (DMRF) based framework is proposed unifying both temporal and spatial information for statistical inference. Unlike FEM based models which provide precise values of all elements given boundary conditions, DMRF models the relations between elements in a structure. The neighborhood system, form of potential functions and parameters can be either manually constructed based on domain knowledge or learned from large training data, or combine both. Critical issues in wireless sensor networks such as communication costs are studied/optimized and the solution is integrated in the proposed framework. The developed framework finds applications in various areas. Cup anemometers are commonly used for wind speed measurement in the wind industry. Anemometer malfunctions lead to excessive errors in measurement and directly influence the wind energy development for a proposed wind farm site. Since the accuracy of anemometers can be severely affected by the environmental factors such as icing and the tubular tower itself, in order to distinguish the cause due to anemometer failures from these factors, our methodologies utilize the spatial relations (multiple anemometers at different locations) and temporal variations under the influence of environmental factors. In another application, the knowledge of the existing pavement condition is very important to the success of any rehabilitation project. The performance of a pavement depends on many factors such as the structural adequacy, the properties of the materials used, traffic loading, climatic conditions and the construction methods. Accurate prediction of the remaining service life (RSL) of pavements has always been a difficult problem. We approach this problem from several aspects. The procedures utilizing non-destructive testing methods, such as the falling weight deflectometer deflection studies, often result in relatively better quantitative assessment of the pavement condition. The spatial relations of deflection measurements at different locations are examined and a technique based on the similarity search method is adopted in predicting the RSL of pavements.
Two Ph.D. students (graduated in April and July of 2013 respectively) and two M.S. students (graduated in May and December of 2012 respectively) were supported by this project.