| Περίληψη: | This thesis addresses the identification of stochastic systems operating under different conditions, based on data records corresponding to a sample of such operating conditions. This topic is very important, as systems operating under different, though constant conditions at different occasions (time intervals) are often encountered in practice. Typical examples include mechanical, aerospace or civil structures that operate under different environmental conditions (temperature or humidity, for instance) on different occasions (period
of day, and so on). Such different operating conditions may affect the system characteristics, and therefore its dynamics.
Given a set of data records corresponding to distinct operating conditions, it is most desirable to establish a single global model capable of describing the system throughout the entire range of admissible operating conditions. In the present thesis this problem is treated via a novel stochastic Functional Pooling (FP) identification framework which introduces functional dependencies (in terms of the operating condition) in the postulated model structure.
The FP framework offers significant advantages over other methods providing global models by interpolating a set of conventional models (one for each operating condition), as it:
(i) treats data records corresponding to different operating conditions simultaneously, and fully takes cross-dependencies into account thus yielding models with optimal statistical accuracy,
(ii) uses a highly parsimonious representation which provides precise information about the system dynamics at any specified operating condition without resorting to customary interpolation schemes,
(iii) allows for the determination of modeling uncertainty at any specified operating condition via formal interval estimates.
To date, all research efforts on the FP framework have concentrated in identifying univariate (single excitation-single response) stochastic models. The present thesis aims at (i) properly formulating and extending the FP framework to the case of multivariate stochastic systems operating under multiple operating conditions, and (ii) introducing an approach based on multivariate FP modeling and statistical hypothesis testing for damage detection under different operating conditions.
The case of multivariate modeling is more challenging compared to its univariate counterpart as the couplings between the corresponding signals lead to more complicated model structures, whereas their nontrivial parametrization raises issues on model identifiability. The main focus of this thesis is on models of the Functionally Pooled Vector AutoRegressive with eXogenous excitation (FP-VARX) form, and
Vector AutoRegressive Moving Average (FP-VARMA) form. These models may be thought of as generalizations of their conventional VARX/VARMA counterparts with the important distinction being that the model parameters are explicit functions of the operating condition.
Initially, the identification of FP-VARX models is addressed. Least Squares (LS) and conditional Maximum Likelihood (ML) type estimators are formulated, and their consistency along with their asymptotic
normality is established. Conditions ensuring FP-VARX identifiability are postulated, whereas model structure specification is
based upon proper forms of information criteria. The performance characteristics
of the identification approach are assessed via Monte Carlo studies, which also demonstrate the effectiveness of the
proposed framework and its advantages over conventional identification approaches based on VARX modeling.
Subsequently, an experimental study aiming at identifying the temperature effects on the dynamics of a smart composite beam via conventional model and novel global model approaches is presented. The conventional model approaches are based on non-parametric and parametric VARX representations, whereas the global model approaches are based on parametric Constant Coefficient Pooled (CCP) and Functionally Pooled (FP) VARX representations. Although the obtained conventional model and global representations are in rough overall agreement, the latter simultaneously use all available data records and offer improved accuracy and compactness. The CCP-VARX representations provide an ``averaged'' description of the structural dynamics over temperature, whereas their FP-VARX counterparts allow for the explicit, analytical modeling of temperature dependence, and attain improved estimation accuracy.
In addition, the identification of FP-VARMA models is addressed. Two-Stage Least Squares (2SLS) and conditional ML type estimators are formulated, and their consistency and asymptotic normality are established. Furthermore, an effective method for 2SLS model estimation featuring a simplified procedure for obtaining residuals in the first stage is introduced. Conditions ensuring FP-VARMA model identifiability are also postulated. Model structure specification is based upon a novel two-step approach using Canonical Correlation Analysis (CCA) and proper forms of information
criteria, thus avoiding the use of exhaustive search procedures. The performance characteristics of the identification approach are assessed via a Monte Carlo
study, which also demonstrates the effectiveness of the proposed framework over conventional identification approaches based on VARMA modeling.
An approach based on the novel FP models and statistical hypothesis testing for damage detection under different operating conditions is also proposed. It includes two versions: the first version is based upon the obtained modal parameters, whereas the second version is based upon the discrete-time model parameters. In an effort to streamline damage detection, procedures for compressing the information carried by the modal or the discrete-time model parameters via Principal Component Analysis (PCA) are also employed. The effectiveness of the proposed damage detection approach is assessed on a smart composite beam with hundreds of experiments corresponding to different temperatures. In its present form, the approach relies upon response (output-only) vibration data, although excitation-response data may be also
used. FP-VAR modeling is used identify the temperature dependent structural dynamics, whereas a new scheme for model structure selection is introduced which avoids the use of exhaustive search procedures. The experimental results verify the capability of both versions of the approach to infer reliable damage detection under different temperatures. Furthermore, alternative
methods attempting removal of the temperature effects from the damage sensitive features are also employed, allowing for a detailed and concise comparison.
Finally, some special topics on global VARX modeling are treated. The focus is on the identification of the Pooled (P) and Constant Coefficient Pooled (CCP) VARX model classes. Although both model classes are of limited scope, they are useful tools for global model identification. In analogy to the FP-VARX/VARMA model case, the LS and conditional ML type estimators are studied for both model classes, whereas conditions ensuring model identifiability are also postulated. The
relationships interconnecting the P-VARX and CCP-VARX models to the FP-VARX models in terms of compactness and achievable accuracy are studied, whereas their association to the conventional VARX models is also addressed. The effectiveness and performance
characteristics of the novel global modeling approaches are finally assessed via Monte Carlo studies.
|
|---|
