Search

Vibration-based Structural Health Monitoring (SHM) is classically approached from two different directions; both involve the acquisition and processing of vibration signals. The first and most popular strategy, which is also followed in the present thesis, relies entirely on the measurements. In contrast, the second approach employs physical models such as finite element (FE) models that are designed based on mechanical principles. In times in which the real-time processing of digital twins for engineering structures becomes more and more realistic, model-based approaches for vibration-based SHM receive increasing attention. Data-driven strategies are still primarily used in vibration-based SHM, and they will remain appealing in situations where precise physical modeling appears cumbersome. Hence, the need for efficient, robust, and reliable data-driven techniques concerning all stages and hurdles of SHM that can prove themselves in practice will never vanish. In this regard, after over 25 years of research, the number of real-life validation studies is still surprisingly low. As for all SHM strategies, the difficulty concerning damage analysis increases with higher levels of realization. Beginning with the goal of detecting damage, SHM finally seeks to predict the remaining lifetime of a structure. The intermediate steps comprise the localization, classification, and assessment of damage. Without the existence of adequately calibrated physics-based models, the successful implementation of methods tackling the objectives beyond damage localization in an unsupervised data-driven scheme is questionable. The term ‘unsupervised’ refers to the fact that knowledge about the manifestation of damage is not available. Especially in civil engineering, this situation pertains in general and is considered in the present thesis. In data-driven SHM, where the area of structural alterations is narrowed down to adjacent sensors, damage localization suffers from the coarse spatial resolution of parsimonious data acquisition systems. Classical modal approaches that hold potential for damage localization require a dense sensor network or significant damage. Originating from the field of fault detection and isolation, estimator- and filter-based methods have proven to be applicable for damage identification of mechanical and civil engineering structures. Notably, they feature an enormous sensitivity towards structural changes when properly designed. Although it remains advantageous for the sake of precise damage localization, these tools such as Kalman or H-infinity filters do not exhibit the inherent demand for a dense sensor network. Consequently, they promise to be viable techniques for the application in vibration-based SHM. A central challenge of this discipline is the discrimination between the natural variability of the structure’s dynamics and the one caused by damage. The former results from varying environmental and operational conditions (EOCs). Especially highly sensitive methods for damage identification are affected by these natural changes, and thus, rely on an efficient data normalization strategy, which can prove itself in practice. In light of these challenges, this thesis provides a real-life validation for the application of quadratic estimators in data-driven vibration-based SHM. To this end, an elaborate technique for estimator-based damage localization is adapted and included in an SHM framework comprising the necessary steps of data normalization and statistical testing. The damage analysis methodology was originally designed for H-infinity filters, which seem well-suited for use in SHM, as they do not assume specific properties of the excitation acting on the structure nor of the involved disturbances. However, previous studies have shown that, in some cases, the filter performance required to achieve high levels of sensitivity towards localized damage cannot be obtained. This issue can be circumvented by employing well-tuned Kalman filters. Therefore, a novel approach for noise covariance estimation is established at first. The associated estimation scheme constitutes a parametric extension of the popular autocovariance least-squares (ALS) technique. The effectiveness of this estimation technique in the context of Kalman filter-based damage localization is studied first using simulations and laboratory experiments. The second part is dedicated to the problem of handling EOCs. This body of work proposes an identification scheme for linear parameter-varying systems based on the interpolation of linear time-invariant systems for different operating points. A simulation study demonstrates the applicability for the purpose of data normalization. Finally, real-life validation of the proposed methods for SHM is conducted. Therefore, a steel lattice mast located outdoors functions as the test object. It is naturally affected by ambient sources of excitation, variability, and uncertainty. The mast, explicitly designed for this validation purpose, is equipped with reversible damage mechanisms that may be activated or removed to reduce the stiffness at multiple locations of the structure. The investigations conducted in this part of the thesis demonstrate proper damage detection of all considered damages as well as localization for the highest degree of severity. These promising results suggest the applicability of the presented methods for Kalman filter tuning, damage localization, and data-normalization in the context of vibration-based SHM.

The synthesis of organic compounds, which is central to many areas such as drug discovery, material synthesis and biomolecular chemistry, requires chemists to have years of knowledge and experience. The development of technologies with the potential to learn and support experts in the design of synthetic routes is a half-century-old challenge with an interesting revival in the last decade. In fact, the renewed interest in artificial intelligence (AI), driven mainly by data availability, is profoundly changing the landscape of computer-aided chemical reaction prediction and retrosynthetic analysis. In this article, we briefly review different approaches to predict forward reactions and retrosynthesis, with a strong focus on data-driven ones. While data-driven technologies still need to demonstrate their full potential compared to expert rule-based systems in synthetic chemistry, the acceleration experienced in the last decade is a convincing sign that where we use software today, there will be AI tomorrow. This revolution will help and empower bench chemists, driving the transformation of chemistry towards a high-tech business over the next decades.

Die vorliegende Arbeit zeigt eine datengest��tzte Methode zur Analyse des mehrk��rperdynamischen Verhaltens von Eisenbahnbr��cken auf. Dabei wird insbesondere auf die Tragwerksantwort infolge mehrk��rperdynamischer Berechnungen eingegangen. Dies ist f��r die Ingenieurpraxis relevant, da sich aus mehrk��rperdynamischen Berechnungen in vielen F��llen eine Reduktion der Tragwerksantwort (im Vergleich zu dynamischen Berechnungen mittels Eink��rpermodell) ergibt. Dadurch k��nnten zuk��nftige Br��ckentragwerke wirtschaftlicher werden. Hierzu wird die Tragwerksantwort von 1000 fiktiven Einfeldtr��gerbr��cken mit St��tzweiten bis 40 m bei der ��berfahrt von konkreten Zugkonfigurationen im Hochgeschwindigkeitsverkehr mit zwei unterschiedlichen Berechnungsmodellen numerisch untersucht. Dazu werden zun��chst Br��ckendaten, die tats��chlich existierende Br��ckentragwerke beschreiben, ausgewertet. Unter der Anwendung stochastischer Methoden l��sst sich anhand dieser Br��ckendaten ein Parameterfeld aus Br��ckenparametern erzeugen, das f��r dynamische Berechnungen herangezogen werden kann. Dabei entspricht jeder Punkt des Parameterfeldes einem fiktiven Br��ckentragwerk. In einem weiteren Schritt werden dynamische Berechnungen f��r alle so erzeugten, fiktiven Br��cken durchgef��hrt. Als Berechnungsmodelle kommen dabei ein Eink��rpermodell, bei dem ��berfahrende Z��ge durch Einzellasten ersetzt werden, sowie ein Mehrk��rpermodell, bei dem ��berfahrende Z��ge durch Mehrmassenschwinger ersetzt werden, zur Anwendung. Danach wird die Differenz der Tragwerksantwort aus den beiden Berechnungsmodellen in Abh��ngigkeit der Parameter des Parameterfeldes ausgewertet. Zwischen den Berechnungsergebnissen wird linear interpoliert, wodurch die funktionelle Abh��ngigkeit oben erw��hnter Differenz von den Br��ckenparametern numerisch beschrieben wird. Zudem werden dynamische Berechnungen an zwei simplen Feder-D��mpfer-Modellen, denen die Parameter des Parameterfeldes zugeordnet werden, vorgenommen. F��r die beiden Feder-D��mpfer-Modelle wird ebenfalls die Differenz der Tragwerksantwort berechnet und diese mit der Differenz der Tragwerksantwort aus den ersten beiden Berechnungsmodellen verglichen. Es wird untersucht, ob eine Korrelation zwischen den beiden Differenzen der Tragwerksantwort vorliegt. Durch obige Vorgangsweise ergibt sich f��r die betrachteten Zugkonfigurationen ein Zusammenhang zwischen der Differenz der Tragwerksantwort einerseits und Br��ckenparametern andererseits. Der beschriebene Zusammenhang kann je nach zugrunde gelegtem Berechnungsaufwand entweder zur Absch��tzung oder zur Bestimmung der ��nderung der Tragwerksantwort zufolge mehrk��rperdynamischer Berechnungen (im Vergleich zu dynamischen Berechnungen mittels Eink��rpermodell) herangezogen werden. Au��erdem zeigen die Berechnungsergebnisse, das simple Feder-D��mpfer-Modelle unter bestimmten Voraussetzungen f��r eine grobe Absch��tzung des gefundenen Zusammenhanges herangezogen werden k��nnen., This diploma thesis analyses the multi-body-dynamical behaviour of railway bridges which is characterised in terms of dynamical bridge responses. Predicting this behaviour is astute to engineering design, since multi-body-dynamical analyses of railway bridges often result in a reduction of dynamical bridge responses (compared to the responses obtained by single-bodydynamical analyses). Thus, the design efficiency of railway bridges can be increased. The study of multi-body-dynamical behaviour of railway bridges was conducted by using 1000 virtual simply-supported railway bridges with spans of up to 40 m subjected to high-speed rail traffic. Using two different mechanical models, all bridges were analysed during the passage of specified train configurations. First, a data set of exisitng railway bridges was statistically analysed. Using stochastical methods, a parameter field for bridge parameters could be obtained from the data set. Each point of the parameter field represented a virtual bridge. Second, single-body as well as multibody dynamic analyses were conducted for every virtual bridge. For the single-body dynamic analyses, trains were replaced by point loads acting on the bridge girder. For the multi-body dynamic analyses, trains were replaced by multiple degree of freedom systems. The differences in bridge responses obtained from the two mechanical models was computed as a function of the bridge parameters. Using linear interpolation, a numerical function describing the differences in bridge responses was obtained. In addition, bridge parameters were assigned to two common spring-damper-models. The two spring-damper-models subjected to dynamic loads were analysed using dynamic analysis. The differences in dynamic responses of the two models was computed and compared to the differences in dynamic bridge responses obtained before. The correlation of the differences was also examined. With the help of results obtained using the methods described above, a relation between bridge parameters and the differences in dynamic bridge responses between the single-body and the multi-body model was derived. Depending on the computational effort made, this relation can be used to either estimate or to predict the differences in dynamic bridge responses between the single-body and the multi-body model. It is concluded that the results obtained using the spring-damper-models allow for a gross estimate of the differences in dynamic bridge responses between the single-body and the multi-body model.

This paper reports the results of our experiments aimed at the automatic optimization of the number of parameters in the semi-continuous phonetically tied HMM based speech recognition system that is part of the speech-to-speech translation system JANUS-2. We propose different algorithms devised to determine the optimal number of model parameters. In recognition experiments performed on a spontaneous human-to-human dialog database, we show that automatic optimization of the acoustic modeling parameter size with the proposed algorithm improves the recognition performance without increasing the required amount of computing power and memory.