Fakultät für Elektrotechnik und Informatik
Permanent URI for this collection
Browse
Recent Submissions
Now showing 1 - 5 of 1259
- ItemComparison and analysis of event-triggered state estimation methods for nonlinear systems(Hannover : Gottfried Wilhelm Leibniz Universität, Institut für Regelungstechnik, 2025) Ji, JiaxinEvent-triggered state estimation has gained significant attention in resource-constrained environments such as network control system due to its ability to save communication and energy resources. For example, it has been observed that event-triggered Kalman filtering-based approaches can achieve resource efficiency in general nonlinear systems [1]. Also the idea of Moving Horizon Estimation (MHE) with its application in event-triggered settings is known to address state estimation tasks effectively [2]. The presented work investigates and compares different such event-triggered state estimation techniques for nonlinear systems, focusing on their performance under different mechanisms and triggering rules. In this work, we introduce the principles of event-triggered state estimation and focus on two types of triggering mechanisms: innovation-based and send-on-delta, combined with stochastic and deterministic triggering rules. We compare the common state estimation methods, including the Kalman filtering-based approaches: Extended Kalman Filter (EKF), Unscented Kalman Filter (UKF) and Cubature Kalman Filter (CKF), as well as optimization-based MHE. Numerical results are presented for four nonlinear systems, including a two-link robot arm. Due to the limitations of event-triggered MHE for nonlinear systems, a linear system-based design is used for comparison. The results demonstrate that CKF achieves the best performance in highly nonlinear systems, while EKF and MHE are more effective in linear or weakly nonlinear settings. Notably, an innovation-based mechanism consistently offers more appropriate triggering sample time, while a deterministic rule enhances the estimation performance. Despite its accuracy, MHE incurs the highest computational costs, followed by CKF, UKF, and EKF. 1. Marzieh Kooshkbaghi, Horacio J Marquez, and Wilsun Xu. “Event-triggered approach to dynamic state estimation of a synchronous machine using cubature Kalman filter”. In: IEEE Transactions on Control Systems Technology 28.5 (2019), pp. 2013–2020. 2. Xunyuan Yin and Jinfeng Liu. “Event-triggered state estimation of linear systems using moving horizon estimation”. In: IEEE Transactions on Control Systems Technology 29.2 (2020), pp. 901–909.
- ItemLearning-based multi-model fitting for 3D Scene understanding(Institutionelles Repositorium der Leibniz Universität Hannover, 2025) Kluger, FlorianThis thesis approaches the problem of robustly fitting multiple geometric models to data in the presence of noise and outliers. Geometric models such as vanishing points, homographies, fundamental matrices and geometric primitives provide information about the 3D structure of data. Extracting this information from images is crucial for attaining a spatial understanding of a scene observed by a machine. With multiple models present in the data, estimating the parameters of each model becomes intractable as the belonging between data points and models is unknown beforehand. Previous works thus detect models via handcrafted heuristics based on random sampling of data points. Since existing datasets for multi-model fitting only contain few scenes, they do not allow for an evaluation of these methods in diverse conditions. We therefore introduce four new datasets for the tasks of vanishing point, homography and fundamental matrix estimation. They contain either real-world or synthetically generated images of indoor or outdoor scenes, offering a wide variety of environments. Three of these datasets contain a significantly larger number of scenes than existing ones and thus enable training of methods for robust multi-model fitting which utilise deep learning. Using these datasets, we present the first learning-based methods for robust multi-model fitting. Extending recent advances in robust model fitting to multi-model problems, both methods guide RANSAC-based model estimators via a neural network. CONSAC detects geometric models sequentially by computing sample weights for each data point, and updating them conditioned on previously found models at each iteration. It predicts accurate model parameters for vanishing point and homography estimation, and our experiments show that it is more robust to outliers than its competitors. PARSAC significantly accelerates this technique by processing all models independently. By computing multiple sample and inlier weights for each data point, the neural network of PARSAC provides a soft segmentation of the data into individual models in a single forward pass. Combined with a weighted inlier counting technique, this allows PARSAC to estimate models in parallel, thereby reducing computation times by up to two orders of magnitude in practise. As our experiments show, PARSAC is currently the fastest robust multi-model fitting approach with state-of-the-art accuracy for vanishing points, homographies and fundamental matrices. Based on CONSAC, we introduce a method for primitive-based 3D scene decomposition. The task of fitting a set of cuboids to a depth map that abstract its 3D shape is formulated as a robust multi-model fitting problem. This method includes two different approaches for estimating cuboid parameters from a minimal set of points, based on iterative numerical optimisation and neural network regression, respectively. A novel occlusion-aware inlier counting technique avoids the selection of overly large and occluding cuboids which do not fit the scene well. Compared to prior work, our method yields more sensible scene decompositions.
- ItemImpedance Characteristics of Monolayer and Bilayer Graphene Films with Biofilm Formation and Growth(Basel : MDPI, 2022) Nakagawa, Ryoichi; Saito, Kai; Kanematsu, Hideyuki; Miura, Hidekazu; Ishihara, Masatou; Barry, Dana M.; Kogo, Takeshi; Ogawa, Akiko; Hirai, Nobumitsu; Hagio, Takeshi; Ichino, Ryoichi; Ban, Masahito; Yoshitake, Michiko; Zimmermann, StefanBiofilms are the result of bacterial activity. When the number of bacteria (attached to materials’ surfaces) reaches a certain threshold value, then the bacteria simultaneously excrete organic polymers (EPS: extracellular polymeric substances). These sticky polymers encase and protect the bacteria. They are called biofilms and contain about 80% water. Other components of biofilm include polymeric carbon compounds such as polysaccharides and bacteria. It is well-known that biofilms cause various medical and hygiene problems. Therefore, it is important to have a sensor that can detect biofilms to solve such problems. Graphene is a single-atom-thick sheet in which carbon atoms are connected in a hexagonal shape like a honeycomb. Carbon compounds generally bond easily to graphene. Therefore, it is highly possible that graphene could serve as a sensor to monitor biofilm formation and growth. In our previous study, monolayer graphene was prepared on a glass substrate by the chemical vapor deposition (CVD) method. Its biofilm forming ability was compared with that of graphite. As a result, the CVD graphene film had the higher sensitivity for biofilm formation. However, the monolayer graphene has a mechanical disadvantage when used as a biofilm sensor. Therefore, for this new research project, we prepared bilayer graphene with high mechanical strength by using the CVD process on copper substrates. For these specimens, we measured the capacitance component of the specimens’ impedance. In addition, we have included a discussion about the possibility of applying them as future sensors for monitoring biofilm formation and growth.
- ItemData-driven methods for spatio-temporal predicition(Institutionelles Repositorium der Leibniz Universität Hannover, 2025) Sao, AshutoshModerne Städte stehen vor großen Herausforderungen wie Verkehrsstaus und Umweltverschmutzung, die in erster Linie auf eine unzureichende Planung zurückzuführen sind, die sich häufig auf traditionelle Methoden wie manuelle Erhebungen und statische Datenanalysen stützt. Diese Ansätze werden der Komplexität städtischer Systeme nur schwer gerecht und führen zur Überfüllung von Straßen und Orten, ungleicher Ressourcenverteilung und eskalierenden Umweltschäden. Die Ineffektivität dieser traditionellen Methoden ist vor allem darauf zurückzuführen, dass es an großen Datenmengen und begrenzten Rechenressourcen für die Verarbeitung komplexer städtischer Dynamiken mangelt. Jüngste Fortschritte in der GPS-Technologie, die inzwischen in alltägliche Geräte wie Fahrzeuge und Smartphones integriert ist, haben jedoch in Kombination mit leistungsstarken Rechenressourcen wie GPUs und TPUs die Erfassung und Analyse umfangreicher räumlich-zeitlicher Daten verändert. Diese Innovationen ermöglichen Prognosemodelle zur Vorhersage von Verkehrsmustern, Ressourcenbedarf und Umwelttrends, die Stadtplanern umfassende Einblicke bieten. Folglich unterstreichen sie das Potenzial datengesteuerter Methoden für raumzeitliche Vorhersagen zur Optimierung der städtischen Infrastruktur, zur Verbesserung des Ressourcenmanagements und zur Bewältigung der dynamischen Anforderungen moderner Städte. Trotz dieser Fortschritte gibt es im Bereich der datengesteuerten raum-zeitlichen Vorhersage noch einige Herausforderungen. Erstens erfordert die effektive Bewältigung verschiedener städtischer Vorhersageaufgaben die Identifizierung und Erfassung gemeinsamer und aufgabenspezifischer Beziehungen innerhalb raum-zeitlicher Daten. Zweitens erfordern spezifische Anwendungen, wie die Vorhersage der Belegung von Ladestationen für Elektrofahrzeuge, aufgrund einzigartiger Datenmerkmale wie schiefe Verteilungen und begrenzte Verfügbarkeit maßgeschneiderte Ansätze. Schließlich bleibt die Entwicklung präziser Vorhersagemodelle für datenarme Regionen eine kritische Herausforderung, da nicht genügend Trainingsdaten zur Verfügung stehen. Um diese Herausforderungen zu bewältigen, werden in dieser Arbeit neuartige, auf Deep Learning basierende Lösungen vorgeschlagen. Zur Bewältigung verschiedener städtischer Vorhersageaufgaben stellen wir ST-SampleNet vor, ein Transformer-basiertes raum-zeitliches Modell, das gemeinsame raum-zeitliche Beziehungen effektiv erfasst. Als Beispiel für eine domänenspezifische Herausforderung stellen wir DFDS vor, ein spezialisiertes Modell für die Vorhersage der Belegung von EV-Ladestationen, das dynamische und statische Informationen integriert, um Datenschieflage und begrenzte Verfügbarkeit zu berücksichtigen. Für datenarme Szenarien präsentiert diese Arbeit MetaCitta, ein Meta-Learning-Framework, das das Wissen aus datenreichen Städten nutzt, um Vorhersagen in Regionen mit begrenzten Daten zu verbessern und so eine robuste und übertragbare Modellierung in verschiedenen Kontexten zu gewährleisten. Insgesamt werden in dieser Arbeit neuartige Methoden zur Nutzung raum-zeitlicher Daten für die Bewältigung kritischer städtischer Herausforderungen vorgestellt. Durch die Präsentation innovativer Lösungen für eine nachhaltige, datengesteuerte Stadtentwicklung wird eine Grundlage für intelligentere Städte mit optimierter Infrastruktur, effizientem Ressourcenmanagement und einer verbesserten Lebensqualität für die Bewohner geschaffen.
- ItemSolutions of 3×3 Systems of Linear Equations Presented on a Golden Plate(Genève : CERN, 2025-07) Lamjahdi, Mohamed El MamiThis paper presents a direct solution method for 3×3 systems of linear equations. The derivation is based on a sketch of the Reverse-2X Method introduced in [LAM25], as well as the Reverse Gauss-Jordan Algorithm. Owing to the requirement of only a single mathematical operation to determine the third variable, the proposed solution outperforms all state-of-the-art methods, including the 4X Method described in [LAM25].