NCSU Institutional Repository >
NC State Theses and Dissertations >
Please use this identifier to cite or link to this item:
|Title: ||Structural Health Monitoring and Detection of Progressive and Existing Damage using Artificial Neural Networks-Based System Identification|
|Authors: ||Saadat, Soheil|
|Advisors: ||Dr. Mohammad N. Noori, Committee Chair|
Dr. Gregory D. Buckner, Committee Co-Chair
Dr. Tadatoshi Furukawa, Committee Member
Dr. Edward Grant, Committee Member
Dr. Melur K. Ramasubramanian, Committee Member
|Keywords: ||system identification|
artificial neural networks
health monitoring and damage detection
|Issue Date: ||10-Mar-2003|
|Discipline: ||Electrical Engineering|
|Abstract: ||In recent decades, the growing number of civil and aerospace structures has accelerated the development of damage detection and health monitoring approaches. Many are based upon non-destructive and non-invasive sensing and analysis of structural characteristics, and most use structural response information to identify the existence, location, and time of damage. Model based techniques such as parametric and non-parametric system identification seek to identify changes in the parameters of a dynamic structural model. Restoring forces in real structures can exhibit highly non-linear characteristics, thus accurate non-linear system identification is critical. Parametric system identification approaches are commonly used, but these require a priori assumptions about restoring force characteristics. Non-parametric approaches do not require such information, but they typically lack direct associations between the model and the structural dynamics, providing limited utility for accurate health monitoring and damage detection.
This dissertation presents a novel 'Intelligent Parameter Varying' (IPV) health monitoring and damage detection technique that accurately detects the existence, location, and time of damage occurrence without any assumptions about the constitutive nature of structural non-linearities. This technique combines the advantages of parametric techniques with the non-parametric capabilities of artificial neural networks by incorporating artificial neural networks into a traditional parametric model. This hybrid approach benefits from the effectiveness of traditional modeling approaches and from the adaptation and learning capabilities of artificial neural networks. The generality of this IPV approach makes it suitable to a wide range of dynamic systems, including those with non-linear and time-varying characteristics.
This IPV technique is demonstrated using a lumped-mass structural model with an embedded array of artificial neural networks. These networks identify the non-linear and time-varying storing forces that would be difficult or impossible to model using traditional modeling techniques. This approach preserves direct associations between the model and the underlying system dynamics, making it ideally suited for health monitoring. Backpropagation of error is used to identify the 'optimal' network parameters from recorded acceleration responses.
Chapter 1 presents an introduction to commonly used health monitoring and damage detection strategies, discusses their advantages and shortcomings, and identifies the building blocks of an effective health monitoring and damage detection strategy. Chapter 2 presents the principles of modeling and system identification. Different modeling and optimization techniques are introduced and their relevance to health monitoring and damage detection are identified. Chapter 3 introduces artificial neural networks, in particular Radial Basis Function Networks (RBFNs), for function approximation as related to the development of the IPV technique. Chapter 4 presents the development and implementation of the IPV technique. It includes development of: (1) a computational model of a typical three-story, base-excited structure, (2) computational models for elastic, elasto-plastic, and hysteretic restoring forces, (3) structural damage mechanisms, (4) structural response simulations to synthetic and recorded ground excitations, and (5) the IPV technique implementation. Chapter 5 is devoted to studying the effects of changes in artificial neural network parameters on IPV accuracy and performance. Chapter 6 is devoted to studying the effects of measurement noise on IPV accuracy. Chapter 7 identifies the main advantages of IPV over other techniques and provides future research directions.|
|Appears in Collections:||Dissertations|
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.