Promotionszentrum Angewandte Informatik
Refine
Document Type
- Doctoral Thesis (19)
- Article (1)
Has Fulltext
- yes (20)
Keywords
- Machine Learning (2)
- visual analytics (2)
- 3D-Druck (1)
- 3D-Rekonstruktion (1)
- 5G (1)
- Algorithms (1)
- Augmented Reality (1)
- AutoML (1)
- Automotive Ethernet (1)
- Automotive Security (1)
Institute
Heterogeneity in imaging protocols, patient populations, and institutional data governance remains a central barrier to the clinical deployment of breast imaging AI systems. This cumulative dissertation addresses this challenge through a unified research program that integrates (i) interpretable Magnetic Resonance Imaging (MRI) radiomics for molecular subtype classification, (ii) multimodal volumetric deep learning for therapy-response prediction, and (iii) federated learning for multi-institutional collaboration under both moderate and extreme non-Independent and non-Identically Distributed (non-IID) conditions.
Paper I (“Leveraging MRI Radiomics and Machine Learning for Accurate Differentiation of Triple-Negative Breast Cancer Subtype”) established a rigorous Image Biomarker Standardisation Initiative (IBSI)-compliant radiomics pipeline for differentiating Triple-Negative Breast Cancer (TNBC) from other subtypes. Using multi-sequence breast MRI, the approach achieved high discriminative performance (Area Under the Receiver Operating Characteristic Curve (AUROC) 0.93) while preserving interpretability, forming a reproducible baseline for subsequent deep learning studies.
Paper II (“Predicting Pathological Complete Response in Breast Cancer Using a Dual 3D ResNet–Transformer Architecture with Multimodal Data Fusion”) introduced a dual-stream 3D deep learning architecture that fuses pre-treatment Dynamic Contrast-
Enhanced MRI (DCE) with structured clinical variables using transformer-based attention. The multimodal model improved the prediction of Pathological Complete Response (pCR) compared to the imaging-only and clinical-only baselines. Furthermore, the core imagingonly architecture was successfully generalized to an independent external clinical cohort, demonstrating the robust predictive capacity of volumetric modeling in precision oncology.
Paper III (“Explainable and Fair Federated Learning with XGBoost for Predicting Pathological Complete Response in Breast Cancer using Multi-Center DCE-MRI Data”) examined Federated Learning (FL) on a multi-center MRI dataset to address the datasharing constraints. The results show that FL maintains the balanced accuracy of a centralized model, improves subgroup fairness, and outperforms single-site training, highlighting its suitability for privacy-preserving clinical collaboration.
Paper IV (“Performance of federated versus centralized learning for mammography classification across film–digital domain shift”) extended this analysis to a strong crossdomain setting involving scanned films versus full-field digital mammography. While FL preserves performance in the dominant digital domain, it shows precision degradation in film images, revealing structural domain-shift limitations and motivating domain-aware personalization and site-specific calibration strategies.
Overall, these studies consolidate an evidence-based view of how interpretability,
multimodal modeling, and federated learning interact under real-world heterogeneity in breast imaging, and highlight practical design trade-offs for translation to clinical settings.
Purpose
We present a comparative study of segmentation methods for high-power laser applications, focusing on two specific challenges: detection of microscopic surface damage on optical components and detection of radiochromic films for reconstructing high-dimensional particle phase space distributions.
Methods
Both applications involve complex morphological variations and non-homogeneous contrast conditions, requiring robust and scalable analysis methods. We evaluate two conventional algorithms and two deep learning-based instance segmentation models, including YOLOv8n-seg and a Detectron2-based Mask R-CNN implementation. All models are evaluated on real datasets that reflect the experimental complexities. We focus particular attention to the accuracy of contour detection, using geometric evaluation metrics such as radial contour comparison, Hausdorff distance, Chamfer distance, as well as intersection-over-union, and analysing runtime performance.
Results
Our results indicate that the YOLOv8n-seg model outperforms the conventional surface damage segmentation method in accuracy, but with 12 times higher computational requirements. In contrast, for radiochromic films analysing YOLOv8n-seg achieves both higher accuracy and faster evaluation. In comparison to YOLOv8n-seg model, Detectron2-based Mask R-CNN implementation lags in both segmentation performance and runtime.
Conclusion
These results highlight the potential of YOLOv8n-seg model in addressing specific data-related challenges in modern laser diagnostics and support their role in the development of next-generation automated analysis systems.
Während der Durchführung von Fertigungsprozessen und Bauprojekten treten oftmals Schwierigkeiten auf, beispielsweise durch Fehlplanungen oder nachträgliche Planungsänderungen. Im Baubereich gehören dazu etwa veränderte Rohrführungen, nachträgliche Wanddurchbrüche u. ä., was eine zeitnahe Kommunikation zwischen den verantwortlichen Architekten, Bauleitern und Bauherren erfordert. Dennoch werden gerade bei Bau- oder Umbaumaßnahmen trotz der Dokumentation des Baufortschritts Fehler oder Abweichungen vom ursprünglichen Plan oft erst dann festgestellt, wenn sich daraus später Probleme ergeben – prominente Beispiele, wie den Berliner Flughafen, gibt es hier reichlich. Eine Ursache liegt darin, dass die Dokumentation üblicherweise lediglich durch eine Ansammlung von Plänen und Bildern, ggfs. verknüpft mit deren Aufnahmedatum, geschieht. Auch die Bauüberwachung erfolgt i. d. R. manuell und häufig nur in Form von Baubesichtigungen.
Die rechnergestützte Verwaltung der korrespondierenden Daten (2D-Baupläne, 3D-Architekturmodelle, 2D-/3D-Aufnahmen u. ä.) ermöglicht es hierbei, die gesamten Kommunikations- und Koordinationsprozesse während der Bauphase zu beschleunigen und zu vereinfachen. Dazu müssen zunächst allen Stakeholdern identische, aktuelle Daten vorliegen, was mittels einer kollaborativen Management-Anwendung möglich ist. Dadurch lassen sich etwaige Fehler unmittelbar identifizieren und ggf. direkt verhindern. Ein während der Bauphase automatisierter Abgleich der Aufnahmen mit den Originalplänen erlaubt außerdem eine Echtzeit-Überwachung und Fehlerminimierung. Des Weiteren können auf diese Weise spätere Gebäudewartungen über Augmented-Reality-Techniken unterstützt und deutlich vereinfacht werden. Ein Beispiel hierfür ist die Überblendung der realen Umgebung mit den während der Baumaßnahmen bereits adaptierten originalen Plänen, um so etwa interaktiv anzeigen zu können, wo Rohre oder Kabel tatsächlich verlaufen. Ähnliches gilt für die Überwachung von Fertigungsprozessen, wie etwa beim 3D-Druck von Bauteilen. Auch bei diesem Anwendungsfall kann eine frühzeitige und möglichst automatisierte zerstörungsfreie Abweichungserkennung und Fehlerkorrektur Folgekosten deutlich reduzieren.
Ein automatisierter Abgleich zwischen den ursprünglichen Planungsdaten und einer digitalen Beschreibung des realen Ist-Zustands inklusive einer Rückführung ermittelter Differenzen in die Planungsdaten ermöglicht es, nachträglich gewünschte oder notwendig gewordene, ebenso wie fehlerhafte, Abweichungen vom Bau- oder Fertigungsplan angemessen präsentieren und dokumentieren zu können, um somit z. B. allen Beteiligten zu ermöglichen, fundierte Entscheidungen über das weitere Vorgehen treffen zu können. Als Planungsdaten werden die Ergebnisse des Konstruktionsprozesses angesehen, die i. d. R. aus CAD-Daten bestehen. Diese beschreiben den ursprünglichen Soll-Zustand und bilden beispielsweise die Grundlage für einen Bau- oder Produktionsprozess ab. Die digitale Beschreibung des realen Ist-Zustands in Form von aktuellen Messdaten kann aus unterschiedlichen Quellen stammen, wie z. B. durch den Einsatz von 3D-Scannern. Die 3D-Modelle der ursprünglichen Planungsdaten sowie die der 3D-rekonstruierten Ist-Daten unterscheiden sich typischerweise sowohl in ihrer Geometrie und Topologie als auch in ihrem generellen Aufbau. Neben der Zuordnung und Anzahl enthaltener Dreiecke etc. können die 3D-Modelle der Planungs- bzw. Rekonstruktionsdaten jeweils aus einem einzigen Dreiecksnetz oder aber aus mehreren Dreiecksnetzen bestehen. Sollen einzelne Bauteile überprüft werden, bestehen diese Ist-Daten i.d.R. aus einem einzigen Dreiecksnetz mit unterschiedlicher Topologie. Ein einfacher geometrischer Vergleich zum Ermitteln der Differenzen zwischen Planungs- bzw. Rekonstruktionsdaten ist deshalb nicht ohne weiteres möglich. Im Anwendungsfall \textit{Bauüberwachung} bedeutet dies zudem, dass Objekte von Interesse jeweils einzelne Teilkomponenten darstellen können, wie z. B. Türen oder Fenster in einem Raum, welche für einen Vergleich zunächst aus den Daten extrahiert werden müssen.
Das Ziel dieser Arbeit liegt daher in der automatisierten Erkennung von Differenzen zwischen den ursprünglichen Planungsdaten und einer 3D-Rekonstruktion des realen Ist-Zustands sowie in der Rückführung der ermittelten Differenzen in die ursprünglichen Planungsdaten. Zur Erreichung des Ziels sind vier Forschungsfragen zu beantworten: Was müssen (modellbasierte) 3D-Aufnahmeverfahren leisten, um eine Detektion von Differenzen zu ermöglichen? Wie können die akquirierten 3D-Messdaten sinnvoll mit den dazu korrespondierenden 3D-Planungsdaten verglichen werden? Wie können ermittelte Differenzen automatisiert in die 3D-Planungsdaten (Soll-Daten) zurückgeführt werden? Wie können die ermittelten Differenzen benutzerorientiert visualisiert werden? Die Bearbeitung dieser Forschungsfragen erfolgt auf der Grundlage der vorgestellten Anwendungsfälle aus der Bauindustrie und der Bauteilfertigung.
Um eine für das Detektieren von Differenzen hinreichend gute 3D-Rekonstruk-tion zu gewährleisten, müssen 3D-Aufnahmeverfahren neben einem ausreichenden Auflösungsvermögen im Verhältnis zur Objektgröße einen möglichst hohen Grad an Robustheit gegenüber Einsatzbeschränkungen aufweisen, wie z. B. variable Lichtverhältnisse, transparente oder stark reflektierende Oberflächen. Es wird dargelegt, dass die Anwendung den Benutzer während der Datenakquise unterstützen sollte (z. B. Freistellung der Zielobjekte, Visualisierung des Scanfortschritts), da sich dies signifikant auf die Qualität der erzielten Messdaten auswirkt.
Für den Vergleich der Planungs- und Messdaten werden zwei verschiedene Verfahren entwickelt. Diese unterscheiden sich darin, ob die Objekte von Interesse für einen Vergleich Teilkomponenten repräsentieren, wie Fenster oder Türen, welche für einen Vergleich zunächst aus den Planungs- als auch Messdaten extrahiert werden, oder ob die zu vergleichenden Datensätze in ihrer Gänze verglichen werden. Das objektspezifische Multi-Objekt-Verfahren für den Anwendungsfall Bauüberwachung, bei dem ein Vergleich extrahierter Teilkomponenten stattfindet, wurde durch eine Punktwolkenanalyse der zu vergleichenden Datensätze umgesetzt und bezieht eine Rückführung der hierdurch detektierten Differenzen in die Geometrie der 3D-Planungsdaten ein. Das generische Gesamt-Objekt-Verfahren für den Anwendungsfall Bauteilfertigung vergleicht die Planungs- und Messdaten gänzlich und ermöglicht das Detektieren von Differenzen durch eine Vereinheitlichung der zu vergleichenden Datensätze auf Basis einer hochaufgelösten Voxelisierung und ermittelt einen Ähnlichkeitskoeffizienten anhand der daraus resultierenden Voxelmengen, welche entweder die voxelisierten Polygonnetzoberflächen oder Volumen abbilden.
Um die Kommunikation zwischen den Beteiligten zu verbessern und Entscheidungsfindungsprozesse zu optimieren, wird gezeigt, wie die ermittelten Differenzen (nachträglich gewünschte Abweichungen sowie echte Fehler) nutzer- und anwendungsfallorientiert visualisiert werden können. Es wird dargelegt, dass eine Überlagerung der aktualisierten und ursprünglichen Planungsdaten sowie eine Separation der Differenzen das Hervorheben ermittelter Differenzen unterstützt. Dies trifft ebenso auf die multicodale Ergebnispräsentation zu, bei der die ermittelten Differenzen mittels verschiedener Codalitäten bereitgestellt werden (z. B. tabellarische Präsentation und interaktives 3D-Echtzeit-Rendering der Differenzen, Voxel- und Polygonnetz-basierte Differenzvisualisierung).
Aufbauend auf den erarbeiteten Antworten zu den genannten Forschungsfragen wurde eine Difference Detection Tool Pipeline entwickelt, welche die wesentlichen Funktionalitäten zur Erfassung und Manipulation von Planungs- und Rekonstruktionsdaten integriert und diese für eine kontextbezogene Detektion von Differenzen (inklusive der Rückführung selbiger in die dadurch aktualisierten Planungsdaten) handhabbar bereitstellt. Beim Durchlaufen der Pipeline wird dabei jeweils noch zwischen einer Multi-Objekt-Differenz-Detektion (bei der wie im Anwendungsfall Bauüberwachung zunächst Teilkomponenten zu extrahieren sind), sowie einer Gesamt-Objekt-Differenz-Detektion unterschieden (bei der wie im Anwendungsfall Bauteilfertigung die Planungsdaten aus einer Einzelkomponente bestehen). Der Nutzen der entwickelten Difference Detection Tool Pipeline wird schließlich in verschiedenen Anwendungsfeldern exemplarisch aufgezeigt und evaluiert.
Disinformation, misinformation, and harmful online content have increasingly shaped public discourse, political decision-making, and societal trust. Through the expanding role of social networks and digital news platforms, users are confronted with large volumes of heterogeneous information, making it difficult to distinguish reliable content from deceptive or manipulative material. Existing machine learning approaches often focus on binary classification without addressing the linguistic, narrative, and contextual structures that contribute to misleading content. In addition, related forms of harmful communication - such as hate speech, toxic language, propaganda, biased reporting, and extremist narratives - are closely intertwined with disinformation, yet are rarely analyzed together within a unified research setting. To address these challenges, this dissertation presents DisDETECT, a system for the semi-automatic detection, analysis, and contextualization of disinformation in German-language online media. Two datasets were created for this purpose: one covering short texts and one consisting of long-form articles, both including disinformation, narrative structures, propaganda cues, hate and toxic language, and other stylistic or communicative phenomena. Based on these resources, an information extraction pipeline identifies entities, events, topics, claims, narratives, framing techniques, and indicators of bias at both document and corpus level. The extracted elements are integrated into a semantically enriched knowledge graph that organizes structured triples together with embedding-based representations for subsequent analytical steps and knowledge infusion. The methodological component of the thesis introduces a hybrid classification pipeline that combines transformer-based language models, linguistic features, and knowledge graph embeddings. Nineteen models were developed and evaluated across binary, multiclass, and multilabel tasks relevant to disinformation and harmful communication. Transfer learning was applied using multiple external datasets for pre-training and fine-tuning. The outputs of these models form the basis of an assessment mechanism that assigns degrees of reliability to textual content. To support interpretability, a rule-based explanation layer and large language model components generate textual explanations of model decisions, which are presented within the interactive visual analytics interface of the DisDETECT system. The results of this work show that the combination of information extraction, transfer learning, knowledge infusion, and hybrid classification methods supports detailed analysis of disinformation and related communication styles. A user evaluation of the DisDETECT system provides insights into how experts interact with the interface and interpret the presented results. By integrating linguistic, narrative, and contextual perspectives into a single analytical environment, this dissertation contributes to a broader understanding of how harmful content emerges, overlaps, and propagates across online media.
Modern cars are highly connected and complex IT systems, featuring various services such as backend communication, Vehicle to Everything (V2X) communication, app-control of car features, and seamless authentication of charging processes for electric vehicles. This poses new challenges to cybersecurity as many previous attacks have demonstrated. Because of the broad attack surface, it is important to not only consider the security of the car’s external interfaces but also secure the internal communication to hamper lateral movement. However, designing secure communication protocols is prone to errors. Despite this, many common automotive protocol standards have not been thoroughly analyzed. A powerful method for such an analysis is formal verification, which defines and proves the security properties provided by a protocol. This can be further supported by automated tools such as the Tamarin prover.
In this work, we formally analyze the security of automotive communication protocol standards for in-vehicle and external communication and propose security extensions or mitigations for the issues we identify. For in-vehicle communication, we analyze AUTOSAR SecOC, an authentication protocol for CAN and automotive Ethernet communication, and the communication middlewares SOME/IP and DDS. We identify multiple attacks on SOME/IP and propose two extensions to address its missing security mechanism, one based on asymmetric cryptography and digital certificates and one based on symmetric cryptography utilizing a central authentication server. In our analysis of DDS, we identify a vulnerability in the encryption algorithm that can enable replay attacks in some circumstances. Furthermore, we systematically compare different approaches to secure automotive Ethernet-based in-vehicle communication, also considering MACsec, IPsec, and TLS. Finally, we introduce a new dataset for in-vehicle communication that highlights the limitations of the current state-of-the-art and is a useful resource for developing vehicular IDSs.
For external communication, we focus on two important use cases: Vehicle diagnostics and Electric Vehicle (EV) charging. We formally analyze the authentication mechanisms in the diagnostics protocol UDS, pointing out insecure configurations and identifying two novel vulnerabilities. We propose and formally verify mitigations for these vulnerabilities and practically evaluate a resulting attack. For the second use case, we propose two extensions to the ISO 15118 protocol for EV charging. Our first extension implements missing privacy properties, while our second extension augments the protocol standard with security against quantum attackers and a mechanism for cryptographic agility. Both extensions are formally verified and practically evaluated. Finally, we propose a novel concept for a privacy-preserving EV charging protocol.
This thesis demonstrates how formal verification can improve the security of automotive communication protocol standards and, thus, should become an integral part of the development of new security protocols.
Magnetic Particle Imaging (MPI) is a recently developed imaging technique that allows for both high spatial and temporal resolution. Compared with other classical modalities there is no exposition to ionizing radiation. These features make it a promising modality in medical applications. However, the corresponding imaging task constitutes a severely ill-posed inverse problem which requires regularization techniques to produce acceptable results. Currently used reconstruction methods are based on a time-consuming and memory-intensive calibration process. This calibration process is at the core of the measurement-based approach. To avoid the calibration, model-based approaches are of interest in MPI. In this thesis we outline the main contributions obtained both to the model-based approach and to the measurement-based approach. The main contribution in measurement-based approach is the employment of a Plug-and-Play reconstruction algorithm that uses a pretrained deep-learning-based Gaussian denoiser in a zero-shot fashion, with additional L1-prior, to achieve fast and competitive reconstructions on the OpenMPI dataset. Concerning the model-based approach, we consider a two-stage algorithm based on a reconstruction formula for Field-Free Point (FFP) MPI. This algorithm consists of two stages: the Core Stage and the Deconvolution Stage. We provide a variational formulation of the Core Stage and improve the Deconvolution Stage with TV-like regularization and with a Nonnegative Fused LASSO algorithm, for which we also prove convergence. The two-stage algorithm is capable to deal with scans that are not dependent on the scanning trajectory. We show how this property can be leveraged in multi-patch MPI in a simulated scenario.
As a further contribution to model-based MPI, we have also provided reconstruction formulae for the case of 3D Field-Free Line (FFL) MPI. Such formulae were missing in the literature at the beginning of this project. We show the applicability of the 3D FFL reconstruction formulae in simulated scenarios.
Finally, we further develop our methodology to apply the two-stage model-based algorithm to real MPI data. In particular, we show the first reconstruction with our algorithm on real 2D MPI data collected by a Bruker's scanner. The results are interesting in view of the fact that these are the first reconstructions obtained on real multi-dimensional (2D) MPI with a method that does not depend on the specific choice of the scanning trajectory. To highlight the flexibility of our method, we additionally display reconstructions from data obtained with an MPI scanner that does not use Lissajous trajectories. The results obtained demonstrate that the methods developed in this work are capable of competitive reconstruction quality, while offering flexibility for future general-purpose model-based MPI reconstructions.
Machine Learning (ML) is an exponentially growing technological sector fueled by digitalization and an ever-increasing demand for more ML-powered applications. This increasing demand results in a growing need for Artificial Intelligence (AI) experts, who remain in limited supply. However, the big economies worldwide already face a shortage of qualified personnel. Therefore, it is uncertain whether the current supply of AI experts can adequately meet the increasing demand for AI applications. Furthermore, this bottleneck may also affect innovations, as the private sector (startups, corporations) and the public research sector may start to compete for the finite pool of AI experts. Automated Machine Learning (AutoML) is a research domain that may provide a solution to this problem. The use of AutoML solutions, which automates the laborious data science workflow, enables AI experts to increase their efficiency. Furthermore, business domain experts may use AutoML solutions to access ML without requiring AI experts. However, in many cases, employing AutoML solutions requires AutoML, ML, and programming expertise — skills that business domain experts may not necessarily possess. Finally, selecting the best AutoML solution for a use case is not trivial. It requires manual investigation of its properties in technical documentation, source code and experimentation to identify its effectiveness for a use case. This thesis introduces Meta AutoML, a concept that addresses the limitations of individual AutoML solutions and enables both AI experts and business domain experts to generate effective ML models efficiently. Meta AutoML accomplishes this through the automation of the data science workflow using existing AutoML solutions. It introduces a meta-layer which administers the execution of the AutoML solutions and only displays the relevant information to the user. Therefore, the user is no longer required to possess technical expertise in ML, AutoML, and programming. Meta AutoML autonomously performs all the necessary data science workflows using the underlying AutoML solutions to find the most effective ML model. Furthermore, a unified AutoML interface is defined, that would standardize the functionalities and parametrization AutoML solutions offer. In addition, this thesis provides User Experience (UX) recommendations to ensure a good usability in AI systems. Systems which implement the Meta AutoML concept should implement these recommendations to ensure a good usability for both expert groups.
In a benchmark using 40 classification datasets, the Proof-of-Concept (PoC) Meta AutoML system Ontology-based Meta AutoML (OMA-ML) was evaluated, finding that it can generate top results in 29 out of the 40. However, the Meta AutoML process is highly inefficient, as in many cases, the user is only interested in one ML model, the best performing one. In an effort to increase the efficiency of the Meta AutoML process, two AI-based optimization approaches were evaluated. The first is the rule-based training strategy approach, which uses training strategies to optimize the individual data science workflow steps. Through the usage of the top-3 optimization strategy, the Meta AutoML training process can improve its training efficiency by up to 70% without reducing the best ML model prediction performance. The second approach is ML-based optimization, which aims to utilize ML models to predict the optimal training configuration parameter values for a dataset and ML task. Two ML-based approaches were evaluated; the first aimed to predict the best AutoML solutions for a training, and the second the optimal training runtime. However, neither ML-based approach could be used to predict an optimal training configuration. Finally, the usability aspect of OMA-ML was evaluated through a custom 4-stage usability evaluation methodology. Using this usability methodology, OMA-ML was evaluated by User Experience (UX) experts in an expert review, as well as AI and business domain experts through a usability study. Results show that while OMA-ML did not reach the target usability state yet, with each evaluation iteration, its usability improves. The results from the individual contributions show that the novel Meta AutoML can optimize its effectiveness, efficiency, and usability.
Manual assembly remains a crucial aspect of industrial production, especially in non-standardized pre-assembly lines where human adaptability to new processes plays a significant role. Despite the human ability to quickly adjust to new tasks, errors frequently occur, particularly at the beginning and end of a task or during transitions between workstations, often due to unfamiliarity or fatigue. This thesis presents the development of an assistance system designed to enhance the accuracy of manual assembly tasks by recognizing and classifying fine-grained human actions through video data. The system guides workers and alerts them to potential errors, including those that might arise when components are placed correctly but improperly, such as a screw inserted without being tightened. Given the lack of industrial-standard training data, a new dataset, Industrial Hand Assembly Dataset V1 (IHADV1), is introduced, containing eleven assembly classes. The use of skeleton-based methods for feature extraction supports efficient and cost-effective training of a spatiotemporal Transformer network. Initial models demonstrated 87% accuracy with a significant reduction in trainable parameters, highlighting the dataset’s efficacy for capturing detailed assembly movements and the value of spatiotemporal analysis in skeletal data. Further advancements in the methodology, including the use of cross-attention mechanisms at the encoder level, resulted in an accuracy of 99%. The thesis also explores self-supervised learning techniques, such as random masking on non-industrial data, leading to a model capable of achieving over 90% accuracy on the fine tuned classes with a significant reduction in labeled data. Additionally a curriculum-based self-learning approach was developed to enable the model to adapt to evolving industrial environments and integrate new assembly classes, ensuring continuous improvement during operational deployment. The findings suggest promising applications for the assistance system in industrial settings, with potential for scalable and self-sustaining advancements in assembly line efficiency.
Numerical Evaluation and Optimization of the Mechanical Properties of Particle Reinforced Composites
(2025)
In today’s world, composite materials form the bases of many products, ranging from simple’ packaging material to the parts of a wind turbine. Even though computational methods have entered the modern product development cycle, material development has largely stayed unchanged with the use of personal and energy intensive trial and error approaches. This traditional method necessitates the manufacturing of multiple specimens for physical testing, resulting in a long development time while also generating a lot of waste in the process. Numerical methods such as Finite Element Analysis (FEA) in combination with optimization methods can act as an alternative pathway to shrink the development time of a novel composite and reduce the wastage of materials, time and money.
The goal of this work was to develop a numerical method for particle reinforced composites that generates microstructures with targeted effective material properties intended for a specific application. These requirements stated during the development process alongside with a set of constraints are utilized for the optimization algorithm to generate an optimized microstructure, drawing upon a data bank for the individual material phases.
Digital twins of particles encountered during the research are obtained by use of analytical functions such as Spherical Harmonics, Super Ellipsoids or other equations, which are in the following called ‘exact’ particles. Numerical studies based on FEA were conducted on representative volume elements (RVE) of particle reinforced composites to obtain effective elastic properties The results of the FEA calculations for spherical particles were then compared with results obtained with the use of micromechanical models, such as the Mori-Tanaka scheme, Dilute inclusion. Evaluations of the effect of the particle distribution on the elastic properties of the composite were studied, comparing a homogeneous particle distribution with two distinctly different particle cluster distributions. A numerical surrogate model was developed to approximate the effective elastic properties of the composite with ‘exact’ particles. This method is intended to reduce the computational effort such as calculation time and RAM requirements of evaluating the composites effective elastic material properties in comparison to RVE containing the surrogates’ ‘exact’ particle counterpart. This simplification makes it viable to explore different combinations of particle shapes for different matrix materials. Heuristic optimization methods such as Simulated Annealing, Genetic Algorithm and Particle Swarm Optimization (PSO) were explored for finding an optimal material combination to achieve the targeted effective material properties of the composite. For this a function was derived to obtain the optimal material mixture of the composite, which achieves the targeted effective material properties of the compound. The different heuristic methods were compared according to their numerical stability during optimization and the PSO method was chosen. Numerical methods to generate and evaluate conductive particle reinforced polymer matrix composites were explored, which can utilize a material mixture of two particle shapes such as disc-shaped and line-shaped in a polymer matrix. Lastly the application of machine learning methods such as feed forward neural networks were explored to enable a swift quantification of all the possible solutions with regards to different particle and matrix materials and provide a material mixture which achieves the targeted effective elastic properties.
Fingerprints, i.e. ridge and valley patterns on the tip of a human finger, are one of the most important biometric characteristics due to their known uniqueness and persistence properties. Large-scale fingerprint recognition systems are not only used worldwide by law enforcement and forensic agencies, they are also deployed in the mobile market and in nationwide applications. In recent years, contactless fingerprint recognition has become a viable alternative to established contact-based methods. The contactless capturing process avoids distinct problems, e.g. signal of low contrast caused by dirt or humidity and left-over latent fingerprints on the capture surface. Moreover, contactless schemes provide a faster and more hygienic as well as a more convenient capturing process and hence have a higher user acceptance. However, contactless fingerprint recognition introduces new challenges. Environmental influences such as an uncontrolled background and varying illumination and an unconstrained finger positioning are especially a challenge for mobile recognition schemes.
This Thesis contributes to an efficient and secure mobile contactless fingerprint recognition process. The work addresses various vital aspects along the contactless fingerprint recognition pipeline. The mobile, automatic capturing, segmentation and pre-processing of contactless fingerprint samples represents a central focus of this Thesis. Furthermore, contributions to the topics of quality assessment, feature extraction and presentation attack detection are conducted. To enable new research directions, such as training deep learning-based algorithms, a generator for synthetic mobile contactless fingerprint samples is also suggested. The results proposed in this Thesis show improvements on several components of the recognition method which contribute to an increased biometric performance, security and comfort level. Moreover, challenges and limitations are discussed.