20005 research outputs found
Sort by
Structural health monitoring of composites by combining machine learning and synthetic evaluation methods with vibro-acoustic modulations
No full textThe vibro-acoustic modulation (VAM) method is known for its high sensitivity to detect even small damages and is applied non-destructively. VAM uses a high-frequency ultrasonic carrier wave, that is modulated in the specimen by a high-amplitude pump wave of significantly lower frequency. The method was introduced in the 1990s, and the literature today includes over 200 journal publications in which VAM is applied to many laboratory-based applications. However, no industrial application has yet been proclaimed. Since transitioning from the laboratory to the industry is one of the biggest hurdles, this thesis aims to address and overcome the limitations to enable the next step toward industrial applications.
The requirements on the method depend on whether it will be used in the context of non-destructive testing (NDT) or the structural health monitoring (SHM). VAM, as traditionally described in the literature, has no defined baseline of the measurement. Hence, measurements of different specimens can not be reliably compared. In SHM applications, the first measurement is usually assumed to be pristine and is utilised as this baseline. Despite current investigations into estimating a baseline, it remains a major challenge (due to complex signal modulation and dependencies on the material, preexisting damage, \textit{etc.}). To overcome this issue---specifically for NDT applications---this thesis proposes a data-driven approach that is validated with two applications.
First, the adhesive bonding of fibre composite structures is investigated. Single-lap shear specimens with so-called weak bonds and kissing bonds were prepared by inserting a non-stick film, or by contaminating the bond line with a release agent. It is shown that a data-driven evaluation can accurately differentiate between undamaged and damaged specimens in specific frequency ranges, even though the differences based on the traditional evaluation methods are minimal. The trained neural networks are evaluated to recursively generate information on the significance of the input values, leading to a deeper understanding of the VAM method and its mechanisms. The applicability of this data-driven evaluation is confirmed by testing welded steel specimens, where half of them contained crystallisation defects that resulted from false welding parameters.
When VAM is applied as an SHM method, it is advantageous to leverage the ambient vibrations as the modulating pump wave. Ideally, SHM systems are built with low-power and energy-harvesting devices to reduce installation and maintenance costs. The challenge results from variations of the ambient vibration due to changes of environmental influences. Thus, the traditional VAM measurements would neither be consistent nor comparable between the measurements over time.
This challenge is overcome by the proposed synthetic computation of the VAM signal. This synthetic computation minimises the dependency on the ambient vibration so that the VAM measurement can be performed on complex components without interference from environmental changes. The presented approach significantly reduces the requirements on the sensor nodes in terms of sampling rate, measured data points, data size, and energy required to drive the high-frequency emitter. Thus, the usage of self-sustaining energy-harvesting sensor nodes comes into reach. Furthermore, the synthetic method can be implemented into most existing SHM systems that contain acoustic emissions or guided wave measurements with minimal adaptations. Finally, the viability of the synthetic VAM method is demonstrated on larger and more complex structures.Die Vibroakustische Modulation (VAM) ist eine zerstörungsfreie Prüfmethode, die sich durch eine hohe Empfindlichkeit für die Detektion von Schäden auszeichnet. Erstmals wurde die Methode in den 1990er Jahren publiziert. Mittlerweile befassen sich bereits über 200 Veröffentlichungen mit VAM, welche sich jedoch auf kleine Strukturen beschränken, die hauptsächlich under idealen Laborbedingungen getestet wurden.
Eine industrielle Anwendung wurde noch nicht publiziert.
Die Anforderungen an die Methode variieren je nach Anwendung im Rahmen der zerstörungsfreie Prüfung (NDT) oder bei der Überwachung des strukturellen Zustands (SHM).
Die größte Herausfoderung liegt in der genauen Zuordung von einem Modulationswerts zu einem Schadenszustands, mit dem die Messungen an verschiedenen Proben verglichen werden können.
Bei SHM-Anwendungen wird dies üblicherweise umgangen, indem der initiale Zustand als fehlerfrei angenommen wird, und damit als Vergleichswert dient.
Um diese Herausforderungen bei NDT-Anwendungen zu überwinden, wird in dieser Arbeit ein datengetriebener Ansatz vorgeschlagen, der an zwei Anwendungsfällen validiert wird.
Die Detektion von Adhäsionsdefekten in Faserverbunden ist relevant für die Fertigung von Faser-Verbunden.
Überlappklebungen wurden unter idealen Bedingungen und mit eingebrachten Defekten hergestellt. Dazu wurde entweder eine PTFE-Folie eingelegt oder die Klebestelle mit einem Trennmittel kontaminiert.
Beide Defekte verringern die erreichte Scherfestigkeit der Proben signifikant.
Es wurde gezeigt, dass durch eine datengetriebene Auswertung von VAM eine präzise Klassifikation der Proben in die einzelnen Herztellungsarten erfolgen kann, trotz nur minimaler Unterschiede in der Auswertung der traditionell verwendeten Schadens-Indezes von VAM. Zusätzlich wurden die trainierten neuronalen Netze rekursiv evaluiert, um Informationen über die Bedeutung der Seitenbänder (Eingangswerte) zu generieren, was zu einem tiefgehenderen Verständnis der VAM-Methode und ihrer Mechanismen führt.
Bei der zusätzlichen Überprüfung von geschweißten Stahlproben, bei denen ein Kristallisationsfehler in der Schweißnaht eingebracht wurde, konnte die Relevanz der datengetriebenen Auswertung bestätigt werden.
Die Anwendung von VAM als SHM-System bringt zusätzliche Herausforderungen mit sich. Idealerweise werden SHM-Systeme mit energieeffizienten Sensorknoten gebaut, um die Installations- und Wartungskosten zu minimieren. Bei der Anwendung von VAM als SHM-Methode empfiehlt es sich, die bereits vorhandenen Vibrationen als modulierende niederfrequente Schwingung zu nutzen. Die Herausforderung dabei liegt in der Variation dieser Schwingungen aufgrund wechselnder Umwelteinflüsse. Daher sind konventionelle VAM-Messungen an realen Strukturen weder konsistent noch über die Zeit vergleichbar.
Diese Herausforderung konnte durch eine synthetische Berechnung des VAM-Signals umgangen werden.
Mit dieser Methode kann die Abhängigkeit von Umgebungsschwingungen minimiert werden, sodass VAM-Messungen an komplexen Bauteilen ohne Beeinträchtigung durch Umgebungsänderungen durchgeführt werden können.
Der vorgestellte Ansatz reduziert deutlich die Anforderungen an Sensorknoten in Bezug auf Abtastrate, gemessene Datenpunkte, Datengröße und die für den Betrieb des Ultraschallsenders erforderliche Energie.
Dadurch könnten auch autarke Sensorknoten eingesetzt werden.
Zudem ist VAM in die meisten bestehenden SHM-Systeme integrierbar, bei denen die akustischen Emissionen oder geführte Wellenmessungen gemessen werden.
Die Anwendbarkeit der traditionellen sowie der synthetischen VAM wurde auch an größeren und komplexeren Strukturen demonstriert, wobei insbesondere die synthetische Variante vielversprechende Ergebnisse erzielte
Process Understanding of transamination reaction in chiral pharmaceutical intermediate production catalyzed by an engineered amine transaminase
No full textChiral amines are key building blocks for the synthesis of many active pharmaceutical ingredients (APIs). Biocatalytic routes offer significant advantages to provide sustainable access to such motifs on commercial scale, with sacubitril valsartan sodium hydrate as a recent example. In this study a deeper mechanistic and kinetic understanding of the central biocatalytic step in the synthesis of sacubitril valsartan sodium hydrate, applying the evolved transaminase CDX-043, was gained. The equilibrium of the transamination reaction was investigated in detail, and two kinetic models (ping-pong two-substrate kinetics and Michaelis–Menten double substrate kinetics) were established, considering substrate and product inhibition. The determined equilibrium constant indicates that the equilibrium lies strongly on the product side. The results of the kinetic studies demonstrate that the transaminase reaction is in conformity with the Michaelis–Menten double substrate kinetic model. Product inhibition was found to be more severe than substrate inhibition. The application of a plug flow reactor (PFR) was shown to be the preferred reactor setup to reduce the occurring inhibition
Geometric learning of latent parameters with Helmholtz Machines
No full textIn this thesis, we use concepts from Information Geometry (IG), such as Natural Gradient Descent (NG), to improve the training of a Helmholtz Machine (HM) through the design and implementation of a novel algorithm called the Natural Reweighted Wake-Sleep (NRWS).
First, we prove that for any Directed Acyclic Graph (DAG) the associated Fisher Information Matrix (FIM), which describes the geometry of the statistical manifold, has a fine-grained block-diagonal structure that is efficient to invert. By exploiting the fact that the HM is composed of two DAG networks, we adapt its training algorithm into the NRWS implementing NG.
The NRWS not only achieves better performance in the minimum of the optimization loss compared to other training methods, such as the Reweighted Wake-Sleep (RWS) and Bidirectional Helmholtz Machine but also outperforms them in both epochs and wall-clock time. In particular, we present how the NRWS achieves state-of-the-art performance on standard benchmark datasets (MNIST, FashionMNIST, and Toronto Face Dataset) based on the importance sampling estimation of the log-likelihood of the HM.
By adapting Accelerated Gradients (AG) methods to operate within the geometry defined by the FIM of the HM, we further improve the performance of the NRWS. Using first-order AG methods, such as Momentum and Nesterov Momentum, improves the convergence rate of the NRWS without any computational overhead. Additionally, we develop a regularizer method based on the Maximum Entropy Principle, named the Entropy Regularizer (ER), which we show further improves the NRWS by reaching lower optimization loss and narrowing the generalization gap of the algorithm without extra time penalty, which can also be applied to non-geometric training methods. Conveniently, the NRWS framework is compatible with continuous random variables; hence, we show how the FIM can be derived for normally distributed hidden variables.
Finally, we explore the possibilities of using HMs with Convolutional Neural Networks (CNNs) by computing the FIM for such network topologies and showing that the resulting matrix also has a finely-grained block-diagonal structure. We finish by presenting a hypothesis on the difficulties of using CNNs with HMs and NRWS. We make significant contributions to the field of IG and HM, with numerous findings that could be further explored or reused in other research fields. Our results can represent a starting point for future research on improving training algorithms for neural networks and deep learning models using geometric methods, such as the NG.In dieser Dissertation verwenden wir Konzepte der Informationstheorie (IG), wie das natürliche Gradientenverfahren (NG), um das Training der Helmholtz-Maschine (HM) durch die Entwicklung und Implementierung eines neuartigen Algorithmus, des sogenannten Natural Reweighted Wake-Sleep (NRWS), zu verbessern.
Zunächst beweisen wir, dass die zu einem beliebigen gerichteten azyklischen Graphen (DAG) assoziierte Fisher-Informationsmatrix (FIM), die die Geometrie der statistischen Mannigfaltigkeit beschreibt, eine fein abgestufte blockdiagonale Struktur aufweist, die sich effizient invertieren lässt. Wir nutzen diese Beobachtung zusammen mit der Tatsache, dass die HM aus zwei DAG-Netzwerken besteht, um deren Trainingsalgorithmus, das Wake-Sleep Algoritmus (WS), in das NRWS umzuwandeln, das NG implementiert.
Das NRWS erreicht nicht nur eine bessere Performance beim Minimieren des Optimierungsverlustes im Vergleich zu anderen Trainingsmethoden, wie dem Reweighted Wake-Sleep (RWS) und der Bidirectional Helmholtz-Maschine, sondern übertrifft diese auch hinsichtlich der benötigten Epochen und der Laufzeit. Insbesondere zeigen wir, wie das NRWS auf Standard-Benchmark-Datensätzen (MNIST, FashionMNIST und Toronto Face Dataset) eine Spitzenleistung erreicht, basierend auf der Importance-Sampling-Schätzung der Log-Likelihood der HM.
Durch die Anpassung beschleunigter Gradientenverfahren (AG) an die Geometrie, die durch die FIM der HM definiert wird, verbessern wir die Leistung des NRWS weiter. Der Einsatz von AG-Methoden erster Ordnung, wie Momentum und Nesterov-Momentum, beschleunigt die Konvergenzrate des NRWS ohne zusätzlichen Rechenaufwand. Darüber hinaus entwickeln wir eine Regularisierungsmethode, die auf dem Prinzip der maximalen Entropie basiert, den sogenannten Entropieregularisator (ER). Dieser verbessert das NRWS zusätzlich, indem er niedrigere Optimierungsverluste erreicht und die Generalisierungslücke des Algorithmus ohne zusätzlichen Zeitaufwand verringert. Diese Methode kann auch auf nicht-geometrische Trainingsmethoden wie das RWS angewendet werden.
Praktischerweise ist das gesamte NRWS-Framework mit kontinuierlichen Zufallsvariablen kompatibel, sodass wir zeigen, wie die FIM für normalverteilte verborgene Variablen abgeleitet werden kann. Schließlich untersuchen wir die Möglichkeit, HMs mit Convolutional Neural Networks (CNNs) zu verwenden, indem wir die FIM für solche Netzwerktopologien berechnen und zeigen, dass die resultierende Matrix ebenfalls eine fein abgestufte blockdiagonale Struktur aufweist. Wir schließen mit einer Hypothese zu den Schwierigkeiten, CNNs mit HMs und NRWS zu kombinieren.
Wir leisten bedeutende Beiträge auf dem Gebiet der IG und HM mit zahlreichen Erkenntnissen, die weiter erforscht oder in anderen Forschungsfeldern wiederverwendet werden könnten. Unsere Ergebnisse können einen Ausgangspunkt für zukünftige Forschungen mit dem Ziel Trainingsalgorithmen für neuronale Netzwerke und Deep-Learning-Modelle mithilfe geometrischer Methoden wie NG zu verbessern darstellen
Noisy group testing via spatial coupling
No full textWe study the problem of identifying a small number k~nθ, 0<θ <1, of infected individuals within a large population of size n by testing groups of individuals simultaneously. All tests are conducted concurrently. The goal is to minimise the total number of tests required. In this paper, we make the (realistic) assumption that tests are noisy, that is, that a group that contains an infected individual may return a negative test result or one that does not contain an infected individual may return a positive test result with a certain probability. The noise need not be symmetric.We develop an algorithm called SPARC that correctly identifies the set of infected individuals up to o(k) errors with high probability with the asymptotically minimum number of tests. Additionally, we develop an algorithm called SPEX that exactly identifies the set of infected individuals w.h.p. with a number of tests that match the information-theoretic lower bound for the constant column design, a powerful and well-studied test design
The role of creativity and innovation in the quality of our lives, the planet and science
No full textThis editorial is the first in a new set-up, in which we, as the editorial team of Creativity and Innovation Management, share with our community and wider readership our experiences, views and perspectives, expectations or questions we may have. The intention is to publish such an editorial every half year. The topic of this one: creativity, innovation and quality
Material flow control in make-to-stock production systems : an assessment of order generation, order release and production authorization by simulation
No full textMaterial flow control (MFC) is a key element of production planning and control. The literature typically categorizes different MFC methods according to how MFC is realized. This distinction overlooks that MFC decisions can be subdivided into three independent tasks that are executed as orders progress through the system: (i) order generation, (ii) order release, and (iii) production authorization. MFC methods are typically designed for only one of these three tasks, which leaves a large part of the order flow uncontrolled. This study therefore not only provides a new categorization of MFC methods, but also argues for the simultaneous application (or the combining) of three different MFC methods for order generation, order release, and production authorization. To support this argument, the performance effects of an integrated MFC approach are evaluated. Findings show that each individual MFC method impacts different performance metrics, which can be explained by the presence of a hierarchy of workloads, where each workload level constrains the succeeding hierarchical level. Each MFC method has a main impact on a different workload. This has important implications for the design of MFC methods and extends recent literature on hierarchical production planning and control systems
Hemodynamics in arterial bypass graft anastomoses with varying cuff sizes and proximal flow paths: a fluid–structure interaction study
No full textThis article investigates the effect of the cuff size of arterial bypass grafts and the flow conditions on the hemodynamics in the anastomosis (connection) to the artery, using numerical simulations. We consider a fluid-structure interaction problem which is solved based on a partitioned scheme. Additionally, we employ computational fluid dynamics to investigate the effect of a rigid wall assumption. The work focuses on clinically relevant hemodynamic quantities associated with the development of intimal hyperplasia. We also include a model for the prediction of hemolysis into the simulation. The results show that even minor changes of the cuff size can result into significant differences in the corresponding quantities of interest. The importance of the inflow path is shown to be lower than that of the cuff size. The usually employed rigid wall assumption is found to be adequate to address wall shear stress oscillations but falls short on predicting maximum and minimum wall shear stress-related quantities of interest
Systematic construction of continuous-time neural networks for linear dynamical systems
No full textDiscovering a suitable neural network architecture for modeling complex dynamical systems poses a formidable challenge, often involving extensive trial and error and navigation
through a high-dimensional hyperparameter space. In this paper, we discuss a systematic approach to constructing neural architectures for modeling a subclass of dynamical systems, namely, linear time-invariant (LTI) systems. We use a variant of continuous-time neural networks in which the output of each neuron evolves continuously as a solution of a first-order or second-order ordinary differential equation. Instead of deriving the network architecture and parameters from data, we propose a gradient-free algorithm to compute sparse architecture and network parameters directly from the given LTI system, leveraging its properties. We bring forth a novel neural architecture paradigm featuring horizontal hidden layers and provide insights into why employing conventional neural architectures with vertical hidden layers may not be favorable. We also provide an upper bound on the numerical errors of our neural networks. Finally, we demonstrate the high accuracy of our constructed networks on three numerical examples
The radial spanning tree in hyperbolic space: degree and edge-length*
No full textConsider a stationary Poisson process η in a d-dimensional hyperbolic space of constant curvature −κ and let the points of η together with a fixed origin o be the vertices of a graph. Connect each point x ∈ η with its radial nearest neighbour, which is the hyperbolically nearest vertex to x that is closer to o than x. This construction gives rise to the hyperbolic radial spanning tree, whose geometric properties are in the focus of this paper. In particular, the degree of the origin is studied. For increasing balls around o as observation windows, expectation and variance asymptotics as well as a quantitative central limit theorem for a class of edge-length functionals are derived. The results are contrasted with those for the Euclidean radial spanning tree
Patterns and thresholds for soil pH across Europe in relation to soil health and degradation
No full textSoil pH indicates the level of acidity or alkalinity in the soil environment, influencing various biogeochemical and physical processes. Additionally, soil pH levels are crucial in determining the bioavailability of elements such as iron, aluminium, and heavy metals which can be harmful. As such, pH is an important soil health and degradation indicator. Although there is a well-established understanding of soil pH at localized levels, the spatial and temporal variations, as well as significant thresholds at national and continental scales, are not sufficiently documented. Here we analyse the European topsoil pH data (LUCAS) in combination with other soil properties from the LUCAS survey, to identify thresholds and spatial patterns of soil pH across Europe in relation to soil health and degradation. At the European scale we found: 1) the water balance, calculated as mean annual precipitation minus potential evapotranspiration (MAP-PET), provides essential context to interpret soil pH; 2) the shift from organic carbon-rich soils to those dominated by inorganic carbon is observed at a pH of about 7.2, however, soil moisture levels may be more critical than pH for the accumulation of soil organic carbon; 3) we identified three distinct clusters within the multivariate regression tree: acidophiles (below pH 5.2), neutrophiles (pH 5.2–6.9) and alkaliphiles (above pH 6.9), while optimum microbial diversity occurred between pH 6 and 7. Earthworm abundance, as reported by the sWorm database, is more nuanced and dependent on land use; 4) risk of degradation by heavy metals cannot be captured by a single pH threshold. Finally, we identify soil pH thresholds that can aid policymakers in identifying regions that may require protection or intervention