Vibration-based structural health monitoring (SHM) seeks to detect, quantify, locate, and prognosticate damage by processing vibration signals measured while the structure is operational. The basic premise of vibration-based SHM is that damage will affect the stiffness, mass or energy dissipation properties of the structure and in turn alter its measured dynamic characteristics. In order to make SHM a practical technology it is necessary to perform damage assessment using only a minimum number of permanently installed sensors. Deducing damage at unmeasured regions of the structural domain requires solving an inverse problem that is underdetermined and(or) ill-conditioned. In addition, the effects of local damage on global vibration response may be overshadowed by the effects of modelling error, environmental changes, sensor noise, and unmeasured excitation. These theoretical and practical challenges render the damage identification inverse problem ill-posed, and in some cases unsolvable with conventional inverse methods. This dissertation proposes and tests a novel interpretation of the damage identification inverse problem. Since damage is inherently local and strictly reduces stiffness and(or) mass, the underdetermined inverse problem can be made uniquely solvable by either imposing sparsity or non-negativity on the solution space. The goal of this research is to leverage this concept in order to prove that damage identification can be performed in practical applications using significantly less measurements than conventional inverse methods require. This dissertation investigates two sparsity inducing methods, L1-norm optimization and the non-negative least squares, in their application to identifying damage from eigenvalues, a minimal sensor-based feature that results in an underdetermined inverse problem. This work presents necessary conditions for solution uniqueness and a method to quantify the bounds on the non-unique solution space. The proposed methods are investigated using a wide range of numerical simulations and validated using a four-story lab-scale frame and a full-scale 17 m long aluminum truss. The findings of this study suggest that leveraging the attributes of both L1-norm optimization and non-negative constrained least squares can provide significant improvement over their standalone applications and over other existing methods of damage detection.

Iterative Krylov subspace methods have proven to be efficient tools for solving linear systems of equations. In the context of ill-posed inverse problems, they tend to exhibit semiconvergence behavior making it difficult detect ``inverted noise" and stop iterations before solutions become contaminated. Regularization methods such as spectral filtering methods use the singular value decomposition (SVD) and are effective at filtering inverted noise from solutions, but are computationally prohibitive on large problems. Hybrid methods apply regularization techniques to the smaller ``projected problem" that is inherent to iterative Krylov methods at each iteration, thereby overcoming the semiconvergence behavior. Commonly, the Golub-Kahan bidiagonalization is used to construct a set of orthonormal basis vectors that span the Krylov subspaces from which solutions will be chosen, but seeking a solution in the orthonormal basis generated by the Arnoldi process (which is fundamental to the popular iterative method GMRES) has been of renewed interest recently. We discuss some of the positive and negative aspects of each process and use example problems to examine some qualities of the bases they produce. Computing optimal solutions in a given basis gives some insight into the performance of the corresponding iterative methods and how hybrid methods can contribute. / Master of Science

Magnetic induction tomography (MIT) is a non-invasive, soft field imaging modality that has the potential to map the electrical conductivity (σ) distribution inside an object under investigation. In MIT, a number of exciter and receiver coils are distributed around the periphery of the object. A primary magnetic field is emitted by each exciter, and interacts with the object. This induces eddy currents in the object, which in turn create a secondary field. This latter is coupled to the receiver coils and voltages are induced. An image reconstruction algorithm is then used to infer the conductivity map of the object. In this thesis, the application of MIT for volumetric imaging of objects with low conductivity materials (< 5 Sm-1) and dimensions < 1 m is investigated. In particular, two low conductivity applications are approached: imaging cerebral stroke and imaging the saline water in multiphase flows. In low conductivity applications, the measured signals are small and the spatial sensitivity is critically compromised making the associated inverse problem severely non-linear and ill-posed.The main contribution from this study is to investigate three non-linear optimisation techniques for solving the MIT inverse problem. The first two methods, namely regularised Levenberg Marquardt method and trust region Powell's Dog Leg method, employ damping and trust region strategies respectively. The third method is a modification of the Gauss Newton method and utilises a damping regularisation technique. An optimisation in the convergence and stability of the inverse solution was observed with these methods compared to standard Gauss Newton method. For such non linear treatment, re-evaluation of the forward problem is also required. The forward problem is solved numerically using the impedance method and a weakly coupled field approximation is employed to reduce the computation time and memory requirements. For treating the ill-posedness, different regularisation methods are investigated. Results show that the subspace regularisation technique is suitable for absolute imaging of the stroke in a real head model with synthetic data. Tikhonov based smoothing and edge preserving regularisation methods also produced successful results from simulations of oil/water. However, in a practical setup, still large geometrical and positioning noise causes a major problem and only difference imaging was viable to achieve a reasonable reconstruction.

Sei (X(t), t>= -r) ein stationärer stochastischer Prozess, der die affine stochastische Differentialgleichung mit Gedächtnis dX(t)=L(X(t+s))dt+sigma dW(t), t>= 0, löst, wobei sigma>0, (W(t), t>=0) eine Standard-Brownsche Bewegung und L ein stetiges lineares Funktional auf dem Raum der stetigen Funktionen auf [-r,0], dargestellt durch ein endliches signiertes Maß a, bezeichnet. Wir nehmen an, dass eine Trajektorie (X(t), -r 0, konvergiert. Diese Rate ist schlechter als in vielen klassischen Fällen. Wir beweisen jedoch eine untere Schranke, die zeigt, dass keine Schätzung eine bessere Rate im Minimax-Sinn aufweisen kann. Für zeit-diskrete Beobachtungen von maximalem Abstand Delta konvergiert die Galerkin-Schätzung immer noch mit obiger Rate, sofern Delta is in etwa von der Ordnung T^(-1/2). Hingegen wird bewiesen, dass für festes Delta unabhängig von T die Rate sich signifikant verschlechtern muss, indem eine untere Schranke von T^(-s/(2s+6)) gezeigt wird. Außerdem wird eine adaptive Schätzung basierend auf Wavelet-Thresholding-Techniken für das assoziierte schlechtgestellte Problem konstruiert. Diese nichtlineare Schätzung erreicht die obige Minimax-Rate sogar für die allgemeinere Klasse der Besovräume B^s_(p,infinity) mit p>max(6/(2s+3),1). Die Restriktion p>=max(6/(2s+3),1) muss für jede Schätzung gelten und ist damit inhärent mit dem Schätzproblem verknüpft. Schließlich wird ein Hypothesentest mit nichtparametrischer Alternative vorgestellt, der zum Beispiel für das Testen auf Gedächtnis verwendet werden kann. Dieser Test ist anwendbar für eine L^2-Trennungsrate zwischen Hypothese und Alternative der Ordnung T^(-s/(2s+2.5)). Diese Rate ist wiederum beweisbar optimal für jede mögliche Teststatistik. Für die Beweise müssen die Parameterabhängigkeit der stationären Lösungen sowie die Abbildungseigenschaften der assoziierten Kovarianzoperatoren detailliert bestimmt werden. Weitere Resultate von allgemeinem Interessen beziehen sich auf die Mischungseigenschaft der stationären Lösung, eine Fallstudie zu exponentiellen Gewichtsfunktionen sowie der Approximation des stationären Prozesses durch autoregressive Prozesse in diskreter Zeit. / Let (X(t), t>= -r) be a stationary stochastic process solving the affine stochastic delay differential equation dX(t)=L(X(t+s))dt+sigma dW(t), t>= 0, with sigma>0, (W(t), t>=0) a standard one-dimensional Brownian motion and with a continuous linear functional L on the space of continuous functions on [-r,0], represented by a finite signed measure a. Assume that a trajectory (X(t), -r 0. This rate is worse than those obtained in many classical cases. However, we prove a lower bound, stating that no estimator can attain a better rate of convergence in a minimax sense. For discrete time observations of maximal distance Delta, the Galerkin estimator still attains the above asymptotic rate if Delta is roughly of order T^(-1/2). In contrast, we prove that for observation intervals Delta, with Delta independent of T, the rate must deteriorate significantly by providing the rate estimate T^(-s/(2s+6)) from below. Furthermore, we construct an adaptive estimator by applying wavelet thresholding techniques to the corresponding ill-posed inverse problem. This nonlinear estimator attains the above minimax rate even for more general classes of Besov spaces B^s_(p,infinity) with p>max(6/(2s+3),1). The restriction p >= 6/(2s+3) is shown to hold for any estimator, hence to be inherently associated with the estimation problem. Finally, a hypothesis test with a nonparametric alternative is constructed that could for instance serve to decide whether a trajectory has been generated by a stationary process with or without time delay. The test works for an L^2-separation rate between hypothesis and alternative of order T^(-s/(2s+2.5)). This rate is again shown to be optimal among all conceivable tests. For the proofs, the parameter dependence of the stationary solutions has to be studied in detail and the mapping properties of the associated covariance operators have to be determined exactly. Other results of general interest concern the mixing properties of the stationary solution, a case study for exponential weight functions and the approximation of the stationary process by discrete time autoregressive processes.

Zentrale Grenzwertsätze und Konfidenzmengen werden in zwei verschiedenen, nichtparametrischen, inversen Problemen ähnlicher Struktur untersucht, und zwar in der Kalibrierung eines exponentiellen Lévy-Modells und im Dekonvolutionsmodell. Im ersten Modell wird eine Geldanlage durch einen exponentiellen Lévy-Prozess dargestellt, Optionspreise werden beobachtet und das charakteristische Tripel des Lévy-Prozesses wird geschätzt. Wir zeigen, dass die Schätzer fast sicher wohldefiniert sind. Zu diesem Zweck beweisen wir eine obere Schranke für Trefferwahrscheinlichkeiten von gaußschen Zufallsfeldern und wenden diese auf einen Gauß-Prozess aus der Schätzmethode für Lévy-Modelle an. Wir beweisen gemeinsame asymptotische Normalität für die Schätzer von Volatilität, Drift und Intensität und für die punktweisen Schätzer der Sprungdichte. Basierend auf diesen Ergebnissen konstruieren wir Konfidenzintervalle und -mengen für die Schätzer. Wir zeigen, dass sich die Konfidenzintervalle in Simulationen gut verhalten, und wenden sie auf Optionsdaten des DAX an. Im Dekonvolutionsmodell beobachten wir unabhängige, identisch verteilte Zufallsvariablen mit additiven Fehlern und schätzen lineare Funktionale der Dichte der Zufallsvariablen. Wir betrachten Dekonvolutionsmodelle mit gewöhnlich glatten Fehlern. Bei diesen ist die Schlechtgestelltheit des Problems durch die polynomielle Abfallrate der charakteristischen Funktion der Fehler gegeben. Wir beweisen einen gleichmäßigen zentralen Grenzwertsatz für Schätzer von Translationsklassen linearer Funktionale, der die Schätzung der Verteilungsfunktion als Spezialfall enthält. Unsere Ergebnisse gelten in Situationen, in denen eine Wurzel-n-Rate erreicht werden kann, genauer gesagt gelten sie, wenn die Sobolev-Glattheit der Funktionale größer als die Schlechtgestelltheit des Problems ist. / Central limit theorems and confidence sets are studied in two different but related nonparametric inverse problems, namely in the calibration of an exponential Lévy model and in the deconvolution model. In the first set-up, an asset is modeled by an exponential of a Lévy process, option prices are observed and the characteristic triplet of the Lévy process is estimated. We show that the estimators are almost surely well-defined. To this end, we prove an upper bound for hitting probabilities of Gaussian random fields and apply this to a Gaussian process related to the estimation method for Lévy models. We prove joint asymptotic normality for estimators of the volatility, the drift, the intensity and for pointwise estimators of the jump density. Based on these results, we construct confidence intervals and sets for the estimators. We show that the confidence intervals perform well in simulations and apply them to option data of the German DAX index. In the deconvolution model, we observe independent, identically distributed random variables with additive errors and we estimate linear functionals of the density of the random variables. We consider deconvolution models with ordinary smooth errors. Then the ill-posedness of the problem is given by the polynomial decay rate with which the characteristic function of the errors decays. We prove a uniform central limit theorem for the estimators of translation classes of linear functionals, which includes the estimation of the distribution function as a special case. Our results hold in situations, for which a square-root-n-rate can be obtained, more precisely, if the Sobolev smoothness of the functionals is larger than the ill-posedness of the problem.