OPUS Online Publikationen der Universität Stuttgart
Not a member yet
17148 research outputs found
Sort by
A conversion of the geoid to the quasigeoid at the Hong Kong territories
No full textA levelling network was readjusted and a new geoid model compiled within the framework of geodetic vertical datum modernization at the Hong Kong territories. To accomplish all project objectives, the quasigeoid model has to be determined too. A quasigeoid model can be obtained from existing geoid model by applying the geoid-to-quasigeoid separation. The geoid-to-quasigeoid separation was traditionally computed as a function of the simple planar Bouguer gravity anomaly, while disregarding terrain geometry, topographic density variations, and vertical gravity changes due to mass density heterogeneities below the geoid surface. We applied this approximate method because orthometric heights of levelling benchmarks in Hong Kong were determined only approximately according to Helmert’s theory of orthometric heights. Considering a further improvement of the accuracy of orthometric heights by applying advanced numerical procedures, we determined the geoid-to-quasigeoid separation by applying an accurate method. The comparison of the accurately and approximately computed values of the geoid-to-quasigeoid separation revealed significant differences between them. The approximate values are all negative and reach -2.8 cm, whereas values from the accurate method vary between -4.1 and + 0.2 cm. In addition, we assessed the effect of anomalous topographic density on the geoid-to-quasigeoid separation by employing a newly developed digital rock density model. According to our estimates the effect of anomalous topographic density reaches a maximum value of 1.6 cm, reflecting a predominant presence of light volcanic rocks and sedimentary deposits at the Hong Kong territories. Our numerical findings indicate that the conversion between geoid and quasigeoid models should be done accurately, even in regions with a moderately elevated topography.Hong Kong Polytechnic Universit
Development of a deployment platform for ONNX models
No full textArtificial Intelligence models and specifically Machine Learning models are experiencing increasing adoption in various fields and domains. Consequentially the demand for efficient deployment solutions is becoming urgent. Ensuring seamless model management, reliable deployment and fast inference remains a key challenge.
This work presents a solution to the mentioned problem. The solution is a platform for ONNX model deployment, providing a streamlined approach to model versioning, metadata management, and inference execution. To enable the efficient model file storage and their associated metadata, the platform leverages MongoDB alongside GridFS.
Additionally the platform manages model versioning, where each model version is stored as a seperate entry, enabling multiple versions of a model to exist without having to delete previous versions. Deployment and inference are tested using performance metrics, like resource utilization and speed. Usability and robustness though are evaluated through structured test cases and user feedback. All in all the goal is to develop a prototype of the platform quickly utilizing rapid prototyping, while iteratively evaluating it with the help of design science.
Deployment results indicate efficient resource utilization and rapid inference, with challenges in scalability, especially for large models. Usability testing confirms an intuitive interface, ease of use and general user satisfaction. Robustness testing shows that the platform handles unexpected scenarios effectively without failures, while remaining operable and avoiding complete crashes.
Finally the platform successfully addresses ONNX model deployment challenges, while maintaining ease of use, even for non-technical users. Future enhancements could include enhanced model versioning, inference optimizations and integration with external platforms
A sphingolipid rheostat controls apoptosis versus apical cell extrusion as alternative tumour-suppressive mechanisms
No full textEvasion of cell death is a hallmark of cancer, and consequently the induction of cell death is a common strategy in cancer treatment. However, the molecular mechanisms regulating different types of cell death are poorly understood. We have formerly shown that in the epidermis of hypomorphic zebrafish hai1a mutant embryos, pre-neoplastic transformations of keratinocytes caused by unrestrained activity of the type II transmembrane serine protease Matriptase-1 heal spontaneously. This healing is driven by Matriptase-dependent increased sphingosine kinase (SphK) activity and sphingosine-1-phosphate (S1P)-mediated keratinocyte loss via apical cell extrusion. In contrast, amorphic hai1afr26 mutants with even higher Matriptase-1 and SphK activity die within a few days. Here we show that this lethality is not due to epidermal carcinogenesis, but to aberrant tp53-independent apoptosis of keratinocytes caused by increased levels of pro-apoptotic C16 ceramides, sphingolipid counterparts to S1P within the sphingolipid rheostat, which severely compromises the epidermal barrier. Mathematical modelling of sphingolipid rheostat homeostasis, combined with in vivo manipulations of components of the rheostat or the ceramide de novo synthesis pathway, indicate that this unexpected overproduction of ceramides is caused by a negative feedback loop sensing ceramide levels and controlling ceramide replenishment via de novo synthesis. Therefore, despite their initial decrease due to increased conversion to S1P, ceramides eventually reach cell death-inducing levels, making transformed pre-neoplastic keratinocytes die even before they are extruded, thereby abrogating the normally barrier-preserving mode of apical live cell extrusion. Our results offer an in vivo perspective of the dynamics of sphingolipid homeostasis and its relevance for epithelial cell survival versus cell death, linking apical cell extrusion and apoptosis. Implications for human carcinomas and their treatments are discussed.Deutsche Forschungsgemeinschaf
Methodik zur Orchestration kooperativer Smart Services in der Smart Factory
No full textUm den Herausforderungen des 21. Jahrhunderts zu begegnen, müssen sich produzierende Unternehmen an die externe Komplexität des Marktes anpassen. Ein Enabler für die notwendigen Anpassungen ist die Digitalisierung. Im industriellen Umfeld werden die Potenziale der Digitalisierung im Rahmen von Industrie 4.0 verortet. Industrie 4.0 wird in drei Dimensionen unterteilt: Smart Factory, Smart Products und Smart Services. In der Smart Factory wollen Fabrikbetreiber ihre Prozesse kontinuierlich optimieren. Mithilfe von Smart Services können Optimierungspotenziale identifiziert und realisiert werden. In der Praxis erfolgt die Umsetzung von Smart Services zunehmend in Wertschöpfungsnetzwerken mit mehreren rechtlich selbstständigen Akteuren. Mangelndes Vertrauen, Kooperationsbereitschaft, Skalierbarkeit, Gewinnverteilung und ungünstige Rahmenbedingungen erschweren die Umsetzung von Smart Services. Die nachhaltige Kooperation der beteiligten Akteure ist ein kritischer Erfolgsfaktor.
In dieser Arbeit wird eine Methodik vorgestellt, die erstmalig die Kostenersparnis aus Smart Services transparent auf alle Akteure verteilt, sodass die Kooperation für alle Beteiligten vorteilhaft ist. Die Methodik koordiniert die Zusammenarbeit der Akteure im Wertschöpfungsnetzwerk der Smart Factory. Maschinen- und Komponentenlieferanten werden in die Pflicht genommen, die für Smart Services notwendigen Daten aus ihren Betriebsmitteln und Fertigungsprozessen in der erforderlichen Datenqualität zur Verfügung zu stellen. Der Serviceanbieter wird mithilfe der Daten befähigt, Optimierungspotenziale für den Fabrikbetreiber zu realisieren. Um die Zusammenarbeit zwischen den Akteuren langfristig zu sichern, beteiligt der Serviceanbieter die Datenanbieter auf Grundlage der zur Verfügung gestellten Daten am Umsatz.
Die entwickelte Methodik wird auf ein Fallbeispiel angewendet. Die vorgeschlagenen Werkzeuge und Vorgehensweisen sind mit Akteuren dieses Netzwerks validiert. Die Erfassung und Verteilung von Kostenersparnissen führt nach Meinung der elf Validierungsteilnehmer zu einer höheren Kooperationsbereitschaft zwischen den Akteuren.To meet the challenges of the 21st century, manufacturing companies must adapt to the market’s external complexity. One enabler for the necessary adjustments is digitization. In the industrial environment, the potential of digitization is located within Industry 4.0. It is divided into three dimensions: Smart Factory, Smart Products and Smart Services. In the Smart Factory, factory operators want to continuously optimize their processes. Optimization potential can be identified and realized with the help of smart services. The service provider cannot fulfill this task without exchanging data with other players. Furthermore, lack of trust, willingness to cooperate, scalability, profit distribution and unfavorable framework conditions make the implementation of Smart Services difficult. Although there are established methods for identifying and developing smart services, companies have only limited knowledge and a low level of cross-company implementation and deployment. In practice, smart services are increasingly implemented via business ecosystems. The sustainable cooperation of the participating companies is a critical success factor.
In this work, a methodology is presented that transparently distributes the cost savings from Smart Services, so that the cooperation is beneficial for all parties involved. The methodology coordinates the actor’s cooperation in the smart factory’s value network. Machine and component suppliers will be asked to provide the necessary data from their operating equipment and manufacturing processes in the required data quality for Smart Services. With the help of the data the service provider can realize optimization potentials of the factory operator. The service provider involves the stakeholders in revenue based on the data provided.
The developed methodology is applied to a case study with a real business ecosystem. The proposed tools and procedures are validated with the actors of this network. According to the eleven participants, the recording and distribution of cost savings led to a higher willingness to cooperate between the actors of this value network
Accelerating segment anything models via token merging : a comparative study and a spectrum preservation-based approach
No full textThe Segment Anything Model (SAM) has emerged as a significant advancement in image segmentation, demonstrating exceptional generalization across diverse datasets with minimal task-specific tuning. However, its computational demands, inherited from Vision Transformers (ViTs), pose considerable challenges for deployment in resource-constrained environments. This thesis addresses these challenges by integrating token merging strategies, which have proven effective in enhancing the efficiency of ViTs without additional training. Specifically, we conduct a comprehensive analysis of SAM’s architecture and adapt existing token merging techniques to reduce computational overhead while maintaining high segmentation accuracy. We propose an architecture for SAM that incorporates these strategies and evaluate its performance and computational efficiency across various datasets, showing that our approach effectively accelerates SAM’s inference speed while preserving segmentation quality. Furthermore, we propose GradToMe based on PiToMe, an innovative method that leverages gradient approximation and grid-based sampling to combine similar tokens. This approach emphasizes spectrum preservation to retain critical information during the token reduction process, thereby improving the effectiveness of token merging and further saving computational costs. Consequently, our results demonstrate that this approach enhances the feasibility of deploying SAM in real-time applications, making it more suitable for use in resource-limited environments without compromising performance. Code is available at: https://github.com/xxjsw/tome_sam.Das Segment Anything Model (SAM) hat sich als ein bedeutender Fortschritt in der Bildsegmentierung etabliert und zeigt außergewöhnliche Generalisierungsfähigkeiten über verschiedene Datensätze hinweg, bei minimaler aufgabenspezifischer Feinabstimmung. Allerdings stellen die hohen rechnerischen Anforderungen, die vom Vision Transformer (ViT) übernommen wurden, erhebliche Herausforderungen für den Einsatz in ressourcenbeschränkten Umgebungen dar. Diese Thesis geht diese Herausforderungen an, indem sie Token-Merging-Strategien integriert, die sich als effektiv erwiesen haben, um die Effizienz von ViTs ohne zusätzliche Trainingsphase zu verbessern. Insbesondere führen wir eine umfassende Analyse der SAM-Architektur durch und passen bestehende Token-Merging-Techniken an, um den Rechenaufwand zu verringern, ohne die Segmentierungsgenauigkeit zu beeinträchtigen. Wir schlagen eine Architektur für SAM vor, die diese Strategien integriert, und evaluieren ihre Leistung sowie ihre rechnerische Effizienz über verschiedene Datensätze hinweg. Dabei zeigen wir, dass unser Ansatz die Inferenzgeschwindigkeit von SAM effektiv beschleunigt, während die Segmentierungsgenauigkeit erhalten bleibt. Darüber hinaus schlagen wir mit GradToMe, basierend auf PiToMe, eine innovative Methode vor, die Gradientenapproximation und gitterbasierte Stichproben nutzt, um ähnliche Tokens zu identifizieren. Diese Methode legt besonderen Wert auf die Erhaltung des Spektrums, um sicherzustellen, dass während des Token-Merging-Prozesses kritische Informationen erhalten bleiben, was den Token-Merging-Prozess optimiert und die Inferenzgeschwindigkeit weiter steigert. Unsere Ergebnisse zeigen, dass dieser Ansatz die Machbarkeit des Einsatzes von SAM in Echtzeitanwendungen verbessert, wodurch es besser für den Einsatz in ressourcenbegrenzten Umgebungen geeignet ist, ohne die Leistung zu beeinträchtigen. Der Code ist verfügbar unter folgendem Link: https://github.com/xxjsw/tome_sam
Endogenous estrogen metabolites as oxidative stress mediators and endometrial cancer biomarkers
No full textBackground. Endometrial cancer is the most common gynecologic malignancy found in developed countries. Because therapy can be curative at first, early detection and diagnosis are crucial for successful treatment. Early diagnosis allows patients to avoid radical therapies and offers conservative management options. There are currently no proven biomarkers that predict the risk of disease occurrence, enable early identification or support prognostic evaluation. Consequently, there is increasing interest in discovering sensitive and specific biomarkers for the detection of endometrial cancer using noninvasive approaches.
Content. Hormonal imbalance caused by unopposed estrogen affects the expression of genes involved in cell proliferation and apoptosis, which can lead to uncontrolled cell growth and carcinogenesis. In addition, due to their ability to cause oxidative stress, estradiol metabolites have both carcinogenic and anticarcinogenic properties. Catechol estrogens are converted to reactive quinones, resulting in oxidative DNA damage that can initiate the carcinogenic process. The molecular anticancer mechanisms are still not fully understood, but it has been established that some estradiol metabolites generate reactive oxygen species and reactive nitrogen species, resulting in nitro-oxidative stress that causes cancer cell cycle arrest or cell death. Therefore, identifying biomarkers that reflect this hormonal imbalance and the presence of endometrial cancer in minimally invasive or noninvasive samples such as blood or urine could significantly improve early detection and treatment outcomes.
Summary. This review analyzes the role of estrogen metabolites as potential biomarkers for the early detection and monitoring of endometrial cancer.Projekt DEA
Biocatalytic stereocontrolled head-to-tail cyclizations of unbiased terpenes as a tool in chemoenzymatic synthesis
No full textTerpene synthesis stands at the forefront of modern synthetic chemistry and represents the state-of-the-art in the chemist’s toolbox. Notwithstanding, these endeavors are inherently tied to the current availability of natural cyclic building blocks. Addressing this limitation, the stereocontrolled cyclization of abundant unbiased linear terpenes emerges as a valuable tool, which is still difficult to achieve with chemical catalysts. In this study, we showcase the remarkable capabilities of squalene-hopene cyclases (SHCs) in the chemoenzymatic synthesis of head-to-tail-fused terpenes. By combining engineered SHCs and a practical reaction setup, we generate ten chiral scaffolds with >99% ee and de , at up to decagram scale. Our mechanistic insights suggest how cyclodextrin encapsulation of terpenes may influence the performance of the membrane-bound enzyme. Moreover, we transform the chiral templates to valuable (mero)-terpenes using interdisciplinary synthetic methods, including a catalytic ring-contraction of enol-ethers facilitated by cooperative iodine/lipase catalysis.Deutsche Forschungsgemeinschaf
Machine learning-based metabolic rate estimation from wearable sensors
No full textAdaptive devices such as exoskeletons and prostheses can enhance human physical capabilities or replace the functionality of missing body parts. However, adjusting these devices for the specific needs of an individual remains a time-consuming and costly procedure. A key objective in optimizing these devices is minimizing the user’s energy expenditure (EE), a metric closely related to metabolic cost. Traditional methods for estimating metabolic cost, such as indirect calorimetry, are performed in controlled environments, limiting real-world applicability. This study aims to bridge this gap by exploring the use of traditional machine learning (ML) methods to estimate metabolic cost in real-time environments, utilizing wearable sensors integrated into adaptive devices. Using the dataset from Ingraham et al. (2019), which includes data from ten healthy subjects performing various exercises, the study investigates how different sensor combinations impact prediction accuracy.
This thesis evaluated multiple ML models, including Random Forest (RF), Support Vector Machines (SVM), Linear Regression (LR), Decision Trees (DT), and Multilayer Perceptrons (MLP), within two cross-validation methods: Leave-One-Subject-Out (LOSO) and Leave-One-Time-Out (LOTO).
Key findings from this evaluation include: In the LOSO setting, RF outperformed other models, achieving the lowest RMSE in several sensor regions, including Hexoskin, EMG Pants, and Best Combination, with the ’Best Combination’ region showing the best results. In contrast, MLP performed well in the LOTO setting, with its strongest performance observed in the ’Best Combination’ region. SVM demonstrated robust performance when all sensor data was combined, emphasizing the potential of multimodal sensor fusion. Hyperparameter tuning and sensor feature selection were crucial factors in optimizing model performance, particularly for more complex models like RF and MLP. The results suggest that while traditional ML methods can estimate EE effectively, challenges remain in refining preprocessing techniques, tuning hyperparameters, and optimizing sensor combinations. This thesis outlines the importance of model selection, sensor fusion, and parameter optimization in developing more accurate and real-time energy expenditure prediction systems for wearable technologies.Adaptive Geräte wie Exoskelette und Prothesen können die körperlichen Fähigkeiten des Menschen verbessern oder die Funktionalität fehlender Körperteile ersetzen. Die Anpassung dieser Geräte an die Bedürfnisse einer Person, ist jedoch nach wie vor ein zeit- und kostenaufwändiges Verfahren. Ein wichtiges Ziel bei der Optimierung dieser Geräte ist die Minimierung des Energieaufwands des Benutzers, der eng mit den mit den Stoffwechselkosten zusammenhängt. Herkömmliche Methoden zur Schätzung der Stoffwechselkosten, wie die indirekte Kalorimetrie, werden in kontrollierten Umgebungen durchgeführt, was die Anwendbarkeit in der realen Welt einschränkt. Diese Thesis zielt darauf ab diese Lücke zu schließen, indem sie die Anwendung traditioneller Machine Learning Methoden zur Schätzung der Stoffwechselkosten in Echtzeitumgebungen unter Verwendung von tragbaren Sensoren, die in adaptive Geräte integriert sind, untersucht. Unter Verwendung des Datensatzes von Ingraham et al. (2019), der Daten von zehn gesunden Probanden enthält, die verschiedene Übungen durchführten, untersucht die Thesis, wie sich verschiedene Sensorkombinationen auf die Vorhersagegenauigkeit auswirken.
In dieser Arbeit wurden mehrere ML-Modelle bewertet, darunter Random Forests (RF), Support Vector Machines (SVM), Linear Regression (LR), Decision Trees (DT), Multi-layer Perceptrons (MLP), und AdaBoost (ADA) mit zwei Cross-Validation-Verfahren: Leave-One-Subject-Out (LOSO) und Leave-One-Task-Out (LOTO). Die wichtigsten Ergebnisse dieser Thesis sind: Mit LOSO Cross Validation schnitt Random Forests (RF) besser ab als andere Modelle, mit dem niedrigsten root mean square error (RMSE) in mehreren Sensorregionen, einschließlich Hexoskin, EMG Pants und Best Combination, wobei die Region Best Combination die besten Ergebnisse zeigte. Im Gegensatz dazu schnitt Multi-layer Perceptrons (MLP) in der LOTO-Validierung gut ab, wobei die stärkste Leistung in der Region Beste Kombination beobachtet wurde. SVM erreichte eine gute Leistung, wenn alle Sensordaten kombiniert wurden, was das Potenzial der Kombination verschiedener Sensortypen unterstreicht. Das Tuning der Hyperparameter und die Auswahl der Sensoren waren entscheidende Faktoren für die Optimierung der Modellleistung, insbesondere bei komplexeren Modelle wie RF und MLP. Die Ergebnisse deuten darauf hin, dass traditionelle ML-Methoden zwar energy expenditure (EE) effektiv schätzen können, es bleiben aber Herausforderungen in der Verfeinerung der Preprossesing-Techniken, des Tunings von Hyperparametern und der Optimierung der Sensorauswahl. Diese Herausforderungen zu bewältigen ist ein wichtiger Teil der Entwicklung genauerer Systeme zu Berechnung von Energieverbrauch in Echtzeit für tragbare Technologien
Analyse verschiedener Vorkonditionierer für Kernel-Matrizen basierend auf der PLSSVM Bibliothek mit Hilfe von SYCL
No full textPLSSVM is a library that enables the efficient training and execution of Support Vector Machines, which can be used to classify data. It does so by utilizing various high performance computing frameworks to construct and solve a system of linear equations. The conjugate gradient algorithm is used to iteratively solve this linear system. Large datasets with many features, resulting in ill-conditioned kernel matrices have a negative impact on the convergence of the CG method.
To remedy this problem, the goal of this thesis is to analyze different preconditioners in the context of the preconditioned conjugate gradient algorithm, in order to reduce the condition number of the linear system, leading to better convergence and higher stability in regards to different hyperparameter sets.
To achieve this goal three different preconditioners were implemented with SYCL and tested, showing that the usage of a preconditioners can indeed help to improve the mentioned aspects, resulting in fewer iterations (up to 78%) to converge and enabling the usage of hyperparameter combinations that were not possible before.PLSSVM ist eine Bibliothek, die das effiziente Training und die Ausführung von Support Vector Machinen ermöglicht, welche zur Klassifizierung von Daten eingesetzt werden können. Für das Training wird ein Gleichungssystem aufgestellt, welches mit Hilfe von verschiedenen hochperformanten Frameworks effizient gelöst werden kann. Die Lösung dieses Gleichungssystems erfolgt iterativ durch das Verfahren der konjugierten Gradienten. Aufgrund von großen Datensätzen mit vielen Features, kommt es dabei aktuell immer wieder zu Instabilitäten, gerade in Hinblick auf die Kernelmatrizen, welche eine hohe Konditionszahl aufweisen.
Um dieses Problem zu lösen, beschäftigt sich die Masterarbeit mit der Analyse verschiedener Vorkonditionierer für das Verfahren der konjugierten Gradienten. Diese haben das Ziel die Konditionszahl der Gleichungssystems zu verbessern und so die Konvergenz und Stabilität des Verfahrens zu gewährleisten, gerade auch im Bezug auf unterschiedliche Kombinationen von Hyperparametern.
Im Rahmen der Masterarbeit wurden drei verschiedene Vorkonditionierer mit Hilfe von SYCL implementiert und getestet. Die Ergbenisse zeigen hierbei, dass der Einsatz eines Vorkonditionieres die angesprochenen Ziele erreichen kann, indem die Anzahl der Iterationen zur Konvergenz bis zu 78% verringert wird und der Einsatz von vorher nicht möglichen Hyperparameterkombinationen nun möglich ist