Refine
Document Type
- Doctoral thesis (5)
- Scientific article (1)
- Bachelor thesis (1)
- Conference proceeding (1)
Has Fulltext
- yes (8)
Is part of the Bibliography
- no (8)
Keywords
- Machine learning (8) (remove)
Physics-informed neural networks (PINN) are machine-learning methods that have been proved to be very successful and effective for solving governing equations of fluid flow. In this work we develop a robust and efficient model within this framework and apply it to a series of two-dimensional three-component (2D3C) stereo particle-image velocimetry datasets, to reconstruct the mean velocity field and correct measurements errors in the data. Within this framework, the PINNsbased model solves the Reynolds-averaged-Navier-Stokes (RANS) equations for zeropressure-gradient turbulent boundary layer (ZPGTBL) without a prior assumption and only taking the data at the PIV domain boundaries. The TBL data has different flow conditions upstream of the measurement location due to the effect of an applied flow control via uniform blowing. The developed PINN model is very robust, adaptable and independent of the upstream flow conditions due to different rates of wall-normal blowing while predicting the mean velocity quantities simultaneously. Hence, this approach enables improving the mean-flow quantities by reducing errors in the PIV data. For comparison, a similar analysis has been applied to numerical data obtained from a spatially-developing ZPGTBL and an adverse-pressure-gradient (APG) TBL over a NACA4412 airfoil geometry. The PINNs-predicted results have less than 1% error in the streamwise velocity and are in excellent agreement with the reference data. This shows that PINNs has potential applicability to shear-driven turbulent flows with different flow histories, which includes experiments and numerical simulations for predicting high-fidelity data.
The creation of adequate simulation models for complex assemblies is an extensive process that requires a lot of experience, and on the other hand involves a multitude of manual, tedious tasks. These are significant obstacles for improving the process performance and capabilities. The objective of this research is to develop methods which digitally imitate the way of thoughts of the engineer in the design process towards a digital system understanding and which support the automation of the involved manual workflow.
This thesis presents a strategy to translate engineering reasoning and actions to an equivalent in the computer domain. A cardinal step is to gain understanding of system arrangements, boundary conditions and its components. Based on this evaluation, the identification of assembly parts is forming the foundation for optimized process chains for the transfer to the analysis environment. Model complexity relates to computational effort, which in turn affects model capabilities and manageability. To achieve a satisfactory compromise of model quality and complexity, this transfer process is strongly dependent on the visual analysis, reasoning and manual implementation of skilled engineers.
The principle of translating engineering logics is pursued from the assembly system to its smallest parts. Component segmentation methods allow subdividing regions of interest into substructures which are assigned with a feature vector. This vector comprises metrics describing the substructures with regard to specific aspects and is the key decision point for subsequent steps as idealization, suitable Finite-Element modeling and ultimately building an analysis model. The created system database is continuously maintained and supports these process chains as well as the final setup of the assembly simulation model.
An automated workflow like this implies advantages for efficiency, but also creates opportunities for further use cases. This workflow has been exploited for generating a training data set from the different simulation variants as a basis to a knowledge representation imitating engineering experience. An algorithm from the graph neural network field is applied to this data set as a conceptual approach. The intention pursued in this concept is to model the learning progress about estimating the influence of modelling decisions on simulation results and quality.
This research proposes a holistic strategy and describes methods to achieve the objectives of decreasing manual effort, introducing an automated and geometry-based process and digitally replicating engineering experience by introducing a knowledge database.
Neuroadaptive technology (NAT) utilizes real-time measures of neurophysiological activity within a closed control loop to create intelligent software adaptation. Measures of electrocortical and neurovascular brain activity are quantified to provide a dynamic representation of the psychological state of the user, with respect to cognitions, emotions and motivation. As such, NAT can access unique aspects of human information processing, and human intelligence, which can subsequently be used to enable more versatile and more human-like forms of machine intelligence. Current trends in different scientific fields indicate an increased interest in integrating context-sensitive information from the human brain into Artificial Intelligence. NAT'22, the Neuroadaptive Technology Conference 2022, was intended to bring scientists interested in Physiological Computing, Applied Neurosciences and Passive Brain-Computer Interfaces together with experts from the fields of Artificial Intelligence, Machine Learning and Intelligent Systems. The main goals of the conference were an exchange of research questions and findings from these fields and the identification of common goals and joint ventures in the domain of Neuroadaptive Technology, including: real-time signal processing, unsupervised vs. supervised ML, designing neuroadaptive interaction, explainable AI (XAI), neuroadaptive applications, hybrid AI systems (DL + symbolic AI) for applied neurosciences, ethics of neurotechnology in real world (responsibility for action, security), cloud-based solutions for data management and more.
NAT'22 was held in Lübbenau, near Berlin, and organised by the Society for Neuroadaptive Technology. These Proceedings contain the abstracts of six keynote lectures and a total of 39 submissions in the categories of Brain-Computer Interface & Applications, Ethics & Perspectives, Artificial Intelligence & Machine Learning, and a poster session.
Surface variations are an unavoidable byproduct of any manufacturing process and may lead to deviating part performance and even elevated part rejection rates. Because traditional computer aided-design approaches are aimed towards production of idealized, nominal geometric shapes, the wide geometric-and statistical-variability typical for any manufacturing process remains unrepresented and is frequently ignored during design. Thus, the present work aims at a more realistic design approach and, therefore, develops a collection of computer-aided design strategies for accurate representation, statistical analysis and prospective estimation of surface deviations with validation examples on aero engine turbine blades.
The CAD representation of real manufactured surfaces requires the ability to accurately recreate complex geometric shapes. This is achieved by automated re-parametrization of any CAD face of interest as B-spline surface with a rather dense control point grid. Face matching to scanned manufactured samples is then performed by calculating control point displacements, which successfully deliver surface representation errors below typical measurement uncertainties on multiple matching examples from turbine shank and hot-gas faces. Since inference of performance variability due to manufacturing is usually limited by the amount of scanned manufactured parts, a probabilistic model is formulated based on singular-value decomposition of control point displacements and identification of dominant manufacturing modes. This allows generation of an infinite set of synthetic deviating surfaces faithful to experimental deviation patterns.
Nominal geometric features may significantly differ between design iterations and manufacturing modes may not necessarily be transferable between different designs. Thus, deviation estimation may remain infeasible before manufacturing. To enable deviation estimation during the design phase, the present work proposes a machine learning strategy to identify deviation patterns explained by nominal geometric properties-such as relative position and local orientation-and use them for deviation estimation on new designs. This strategy is able to predict realistic stress variability induced by shank deviations of a turbine blade design using only surface deviation information from three given designs, which encourages machine-learning approaches as valuable tool for geometric deviation estimation as part of robust design.
Within the last decades, the number of social networks is growing fast. The competition of retaining the customers to grow their platform and increase their profitability is rising. That is why companies need to detect possible churners to retain these. The problem of predicting the users’ lifetime, churning users, and the reasons for churning can be tackled by using machine learning.
The goal of this bachelor thesis is to build machine learning models to predict user churn and the user lifetime within the social network Jodel, a location-based anonymous messaging application for Android and iOS.
To get the best possible prediction results, we have started with extensive literature research, whose approaches we have tested and added to a machine learning pipeline to build predictive models. With these models, we have investigated the performance after different observation time windows and have finally compared the strongest models to detect similarities and understand the insights to learn their behaviour.
The results of this thesis are machine learning models for a selected representative set of communities varying in size within the Kingdom of Saudi Arabia and a country model leveraging all data. These models are used for a regression task by predicting the lifetime of a user and a multi-label classification of a user into six different churn classes. Additionally, we have also given models for a binary classification, where the model will predict if the user will churn within a given time or not. These models have shown general strong predictive power, which is shrinking when limiting the observation time window. Especially the binary classification yielded high accuracy of over 99%.
The best models have been used for predicting user churn within other communities to detect communities with possible similar behaviour. These similarities then have been determined by features’ importance, where the most important features have got fed back into empirics. This has shown statistically significant differences between user groups with a different active time but as of today no clear trends were visible that had led us to define the communities’ behaviours.
Since the competition of social networks is still growing, the retaining of users will stay a core marketing strategy, which will need to be tackled by machine learning and artificial intelligence. The created models could be useful for predicting churning users within the platform Jodel to detect these customers that will churn within a given time.
Researches did not focus much on anonymous and location-based messaging. That is why the results of this thesis on the anonymous messaging application Jodel opens a variety of possible tasks for the future in this context.
Motivated by computing functionals of high-dimensional, potentially metastable diffusion processes, this thesis studies robustness issues appearing in the numerical approximation of expectation values and their gradients. A major challenge being high variances of corresponding estimators, we investigate importance sampling of stochastic processes for improving statistical properties and provide novel nonasymptotic bounds on the relative error of corresponding estimators depending on deviations from optimality. Numerical strategies that aim to come close to those optimal sampling strategies can be encompassed in the framework of path space measures, and minimizing suitable divergences between those measures suggests a variational formulation that can be addressed in the spirit of machine learning. A key observation is that while several natural choices of divergences have the same unique minimizer, their finite sample properties differ vastly. We provide the novel log-variance divergence, which turns out to have favorable robustness properties that we investigate theoretically and apply in the context of path space measures as well as in the context of densities, for instance offering promising applications in Bayesian variational inference.
Aiming for optimal importance sampling of diffusions is (more or less) equivalent to solving Hamilton-Jacobi- Bellman PDEs and it turns out that our numerical methods can be equally applied for the approximation of rather general high-dimensional semi-linear PDEs. Motivated by stochastic representations of elliptic and parabolic boundary value problems we refine variational methods based on backward SDEs and provide the novel diffusion loss, which can be related to other state-of-the-art attempts, while offering certain numerical advantages.
Machine-Learning-Verfahren in der Produktlinienoptimierung - Simulationsrechnungen und Robustheit
(2021)
Die Produktlinienoptimierung auf Basis der Conjointanalyse ist eines der wichtigsten Probleme der Marketingforschung. Das Ziel der Produktlinienoptimierung ist das Identifizieren einer optimalen Produktlinie bzgl. des Gewinns und Marktanteils. Für die Modellierung der Auswahlwahrscheinlichkeiten der Produkte unter Berücksichtigung der Konsumentenpräferenzen werden deterministische und probabilistische Entscheidungsregeln verwendet. Darauf aufbauend werden für jede Entscheidungsregel jeweils eine Gewinn- und eine Marktanteilsmaximierungszielfunktion aus der Literatur ausgewählt. Als Kombinatorisches Optimierungsproblem ist die Produktlinienoptimierung NP-hard, wodurch die Berechnung der Optima schon bei mittelgroßen Problemgrößen nicht mehr möglich ist. Aus diesem Grund sind Verfahren aus dem Machine-Learning nötig, um die Optima in angemessener Rechenzeit zu approximieren. Die verwendeten Machine-Learning-Verfahren sind Genetische Algorithmen, das Min-Max Ant-System, Particle-Swarm-Optimization und Simulated Annealing, deren Parameter in Abhängigkeit der Problemgröße adaptiv definiert werden. Durch einen neuartigen Clusteransatz können Genetische Algorithmen das deterministische Produktlinienoptimierungsproblem im Sinn der Zielfunktionswertqualität am besten lösen. Das Min-Max Ant-System mit lokaler Suche ist am besten für die deterministische Gewinnmaximierung geeignet. Für die Zielfunktionen mit probabilistischer Entscheidungsregel zeigt sich das Min-Max-System ohne lokale Suche als überlegen. Für die Untersuchung der Robustheit der hierarchischen Bayes-Schätzung wird ein realer Datensatz verwendet. Die Robustheit wird maßgeblich durch die Wahl der Entscheidungsregel beeinflusst. Die Zielfunktionen mit probabilistischer Entscheidungsregel reagieren robust auf die hierarchische Bayes-Schätzung, wohingegen die Zielfunktionen mit deterministischer Entscheidungsregel sehr sensitiv auf sich verändernde Konsumentenpräferenzen reagieren.
Beschleunigung der Verdichterkennfeldberechnung mithilfe von Methoden des maschinellen Lernens
(2021)
In der heutigen Triebwerksentwicklung ist die Verwendung komplexer und zeitaufwändiger numerischer Strömungssimulationsverfahren (3D-CFD) unerlässlich. Dies gilt auch und insbesondere für den Bereich der Verdichterkennfeldberechnung, welcher viele zeitintensive 3D-CFD Berechnungen benötigt. Dabei sind zur qualitativen Beurteilung eines Verdichterentwurfs sowohl Betriebspunkte wie, Reiseflug, Start und Landung, hinreichend genau abzubilden, als auch die kritischen, den Verdichterarbeitsbereich limitierenden Betriebsgrenzen Pumpen und Sperren zu detektieren. Bisherige Arbeiten zur automatisierten Verdichterkennfeldberechnung basieren auf strukturierten Berechnungen von verschiedenen Drehzahllinien, auf welchen jeweils isoliert Pump- und Sperrgrenze gesucht werden. Durch die Beschränkung auf einzelne Drehzahlen wird jedoch nicht der gesamte Charakter des Kennfeldes erfasst, so dass unbekannte Betriebsbereiche aus linearer Interpolation abgeleitet werden müssen. Ein zusätzlicher Nachteil solcher auf einzelne Drehzahllinien fixierten Methoden ist ihre geringe Parallelisierbarkeit.
Der Fokus dieser Arbeit liegt daher auf der Entwicklung eines effizienten Verfahrens zur Erfassung des gesamten Verdichterkennfeldes. Die zwei wesentlichen Anforderungen an das Verfahren sind erstens die Reduktion der Anzahl der notwendigen CFD-Berechnungen zur hinreichend genauen Beschreibung des Verdichterkennfeldes sowie zweitens die Beschleunigung jeder einzelnen 3D-CFD-Berechnung. Zu diesem Zweck wird zur Kennfeldberechnung eine Strategie vorgeschlagen, welche sich von der üblichen strukturierten Berechnung einzelner Drehzahllinien löst und stattdessen mit unstrukturierten, zufällig bestimmte Stützstellen arbeitet. Dabei wird ein zweiphasiges Verfahren entwickelt, bei dem zunächst die Pump- und Sperrlinien in ihrer Gesamtheit mit einer iterativen, hoch parallelisierbaren, auf Support-Vector-Machine beruhenden Strategie bestimmt werden. Als nächster Schritt wird mit Methoden der statistischen Versuchsplanung eine ausreichende Dichte von Stützstellen innerhalb der Betriebsgrenzen des Verdichters generiert. Abschließend werden auf Basis aller verwendeten Stützstellen Antwortflächen für Verdichterdruckverhältnis, Wirkungsgrad und Eintrittsmassenstrom aufgebaut.
Zur Reduktion der Rechenzeit jeder einzelnen 3D-CFD Rechnung werden unterschiedliche Methoden zur Erzeugung von Startlösungen betrachtet. In diesem Rahmen werden Initialisierungsansätze aus reduzierten Strömungsmodellen und aus der Superposition von bereits bekannten Strömungslösungen auf Basis der Methode der Proper-Orthogonal-Decomposition (POD) untersucht.
Als Validierung wird abschließend das entwickelte Verfahren zur Kennfeldberechnung in Kombination mit dem POD-Initialisierungsansatz erfolgreich auf die Analyse eines 4.5- stufigen Forschungsverdichters angewendet.