Time series analysis and feature extraction techniques for structural health monitoring applications

Recently, advances in sensing and sensing methodologies have led to the deployment of multiple sensor arrays on structures for structural health monitoring (SHM) applications. Appropriate feature extraction, detection, and classification methods based on measurements obtained from these sensor networks are vital to the SHM paradigm. This dissertation focuses on a multi-input/multi-output approach to novel data processing procedures to produce detailed information about the integrity of a structure in near real-time. The studies employ nonlinear time series analysis techniques to extract three different types of features for damage diagnostics: namely, nonlinear prediction error, transfer entropy, and the generalized interdependence. These features form reliable measures of generalized correlations between multiple measurements to capture aspects of the dynamics related to the presence of damage. Several analyses are conducted on each of these features. Specifically, variations of nonlinear prediction error are introduced, analyzed, and validated, including the use of a stochastic excitation to augment generality, introduction of local state-space models for sensitivity enhancement, and the employment of comparisons between multiple measurements for localization capability. A modification and enhancement to transfer entropy is created and validated for improved sensitivity. In addition, a thorough analysis of the effects of variability to transfer entropy estimation is made. The generalized interdependence is introduced into the literature and validated as an effective measure of damage presence, extent, and location. These features are validated on a multi-degree-of-freedom dynamic oscillator and several different frame experiments. The evaluated features are then fed into four different classification schemes to obtain a concurrent set of outputs that categorize the integrity of the structure, e.g. the presence, extent, location, and type of damage, taking advantage of the capabilities of these features to extract damage-related information. First, a multivariate outlier and localization technique is established under an unsupervised learning assumption. Next, parallel and serial linear discriminant analysis-based classification algorithms are analyzed in a supervised learning paradigm. Finally, a back-propagation neural network is employed as an additional type of damage classifier. For both a simulated structural model and a tested frame experiment, these methods are shown to correctly classify the structure among a variety of different categories 85 to 99% of the time, depending on noise level