Repository der Technischen Hochschule Ingolstadt
Not a member yet
4170 research outputs found
Sort by
Comparison of Heat Pump Modelling Approaches for Predictive Controllers to Optimize the Efficiency of Modulating Air-Source Heat Pumps: A Case Study from South Germany
No full textEntwicklung eines reflexionsmindernden Bildgebungsverfahrens mit homogener Ausleuchtung und konvolutionalen neuronalen Netzen zur Verschleißerkennung und Klassifizierung komplexer, spiegelnder Zerspanungswerkzeuge
No full textIm Zuge der Digitalisierung stehen Unternehmen, insbesondere Klein- und Mittelständische Unternehmen (KMUs), bei der Etablierung von Künstliche Intelligenz (KI) in industriellen Fertigungsprozessen zunehmend vor großen Herausforderungen. Im letzten Jahrhundert hat die klassische Automatisierung in der Fertigung maßgeblich zur Produktivitätssteigerung und damit auch zur Wettbewerbsfähigkeit von Unternehmen beigetragen. Die Entwicklung und Etablierung von KIgestützter Automatisierung in Unternehmen stellen die nächste Stufe der Produktivitätssteigerung dar und werden die bisherige regelbasierte, klassische Automatisierung zunehmend ablösen. Der Bedarf nach KI-gestützter Automatisierung in der Zerspanungsindustrie, insbesondere in den Bereichen der Werkzeugfertigung und -instandsetzung, gewinnt zunehmend an Bedeutung. Das liegt an den steigenden Kundenanforderungen und am demografischen Wandel, aber auch am Fachkräftemangel, an kostenintensiven Messverfahren sowie an zeit- und ressourcenaufwendigen Schulungen, die zur Bedienung komplexer Messverfahren relevant sind. Das stellt eine Gefahr für die Wettbewerbsfähigkeit der KMUs dar. Klassische Messverfahren wie die Fokusvariation oder triangulationsbasierte Messtechniken liefern zwar eine hohe Genauigkeit, sind jedoch mit erheblichem Zeit- und Kostenaufwand verbunden. Daher eignen sie sich nur bedingt für den Serieneinsatz in der Qualitätskontrolle, insbesondere bei der Vermessung von Schneidwerkzeugen im Rahmen stichprobenbasierter Prüfungen.
Um diese Herausforderungen entgegenzuwirken, wurde im Rahmen dieser Dissertation ein vollständiges KI-gestütztes System zur automatisierten Werkzeuginspektion und -Klassifizierung entwickelt und fundierte Convolutional Neural Networks (CNNs)-basierte Forschung betrieben. Eine der zentralen Kerninnovationen dieser Arbeit liegt in der Entwicklung einer werkzeugspezifischen Lichtquelle, die gezielt auf die optischen Eigenschaften der zylindrischen Schneidwerkzeuge wie Schaftfräser abgestimmt ist und neue Möglichkeiten in der bildbasierten Qualitätskontrolle von glänzenden Oberflächen eröffnet. Die dabei erzeugten hochauflösenden Aufnahmen zeichnen sich durch eine hohe Tiefenschärfe, Farbtreue und gleichmäßige Ausleuchtung aus und ermöglichen eine zuverlässige Verschleißanalyse. Um die Funktionalität dieses neuen Bildgebungsverfahrens zu demonstrieren, wurde zunächst mit regelbasierten Verfahren wie adaptiver Schwellenwertbildung ein Verschleißdetektionsalgorithmus entwickelt und in Gegenüberstellung mit standardisierten Bildaufnahmen aus der Fertigung erforscht. Die ersten vielversprechenden Ergebnisse führten zur Anwendung bestehender Architekturen wie Mask Region-Based Convolutional Neural Network (Mask R-CNN). Diese wurden für die Lokalisierung und Detektion von Verschleiß anhand weniger Bilder genutzt. Außerdem wurde ein CNN-basiertes Verfahren für die semantische Segmentierung entwickelt, das in der Lage ist, normalen und abnormalen Verschleiß an hochreflektierenden Schaftfräsern mit komplexen Freiformflächen zu detektieren und mit hoher Genauigkeit trennscharf voneinander zu unterscheiden.
Bei der semantischen Segmentierung wurden unterschiedliche Ansätze für die Modelloptimierung verfolgt, darunter Variationen der Ankerboxen und des Intersection over Union (IoU)-Schwellenwerts bei Mask R-CNN, um die bestmögliche Modellperformance zu erhalten. Weitere CNN-basierte Ansätze folgten. Hierbei wurden bewährte Architekturen wie U-Net Architecture (UNet) für die Verschleißdetektion modifiziert, aber auch eine kompakte, dimensionsreduzierte Form entwickelt, um normalen von abnormalem Verschleiß mit hoher Genauigkeit zu detektieren und Aussagen über das Verschleißverhalten zu treffen. Dabei wurden unterschiedliche Regularisierungstechniken und Hyperparameteroptimierung angewandt, um die beste Modellleistung an zwei unterschiedlichen hochreflektierenden Werkzeugtypen zu erreichen.
Es wurden im Rahmen dieser Arbeit auch die State of the Art (SOTA) CNNs-Architekturen für die Klassifizierung von Werkzeugen erforscht. Hierbei wurde die Auswirkung unterschiedlicher Beleuchtungsbedingungen, der Bildqualität sowie der unterschiedlichen Trainingsstrategien untersucht, um die am besten geeignete Architektur und Trainingsstrategie für eine KI-basierte Werkzeugklassifizierung zu identifizieren.In the context of digitalization, companies, especially Small and Medium-Sized Enterprises (SMEs), are increasingly facing major challenges in establishing Artificial Intelligence (AI) in industrial manufacturing processes. In the last century, classic automation in manufacturing contributed significantly to productivity increases and thus also to the competitiveness of companies. Developing and establishing AI-supported automation in companies represents the next stage of increasing productivity and will increasingly replace the existing rule-based, classic automation. Developing and setting up AI-based automation in companies is the next step in boosting productivity and will gradually replace the traditional rule-based automation we’ve seen so far. There’s a growing need for AI-based automation in the machining industry, especially in tool manufacturing and restoration. This is because of increasing customer demands and demographic change, but also because of a shortage of skilled workers, costly measurement methods, and time-consuming and resource-intensive training that’s needed to use complex measurement methods. The classic measurement methods such as focus variation or triangulation-based measurement techniques provide high accuracy, but they are associated with considerable time and cost. Therefore, they are only suitable for limited use in quality control, especially when measuring cutting tools in sample-based tests. The AI-based measurement method is a promising alternative to traditional measurement methods.
To address these challenges, a complete AI-based system for automated tool inspection and classification was developed as part of this dissertation, and in-depth research based on CNNs was carried out. One of the central core innovations of this work lies in the development of a tool-specific light source that is specifically tuned to the optical properties of cylindrical cutting tools such as end mills, opening up new possibilities in image-based quality control of shiny surfaces. High-resolution images generated in this way are characterized by high depth of field, color fidelity, and uniform illumination, enabling reliable wear analysis. To demonstrate the functionality of this new imaging method, a wear detection algorithm was first developed using rule-based methods such as adaptive thresholding and researched in comparison with standardized images from production. The first promising results led to the implementation of existing architectures such as Mask R-CNN. These were used to localize and detect wear based on a small number of images. In addition, a CNN-based method for semantic segmentation was developed that is capable of detecting normal and abnormal wear on highly reflective end mills with complex free-form surfaces and distinguishing between them with high accuracy.
In semantic segmentation, different approaches were pursued for model optimization, including variations of anchor boxes and the IoU threshold in Mask R-CNN, in order to achieve the best possible model performance. Further CNN-based approaches followed. Here, proven architectures such as UNet were modified for wear detection, but a compact, reduced-dimension form was also developed to detect normal and abnormal wear with high accuracy and to make statements about wear behavior. Different regularization techniques and hyperparameter optimization were applied to achieve the best model performance on two different highly reflective tool types.
As part of this work, SOTA CNNs architectures for tool classification were also investigated. The effect of different lighting conditions, image quality, and different training strategies was studied to identify the most suitable architecture and training strategy for AI-based tool classification.
In semantic segmentation, different approaches were pursued for model optimization, including variations of anchor boxes and the intersection-over-union (IoU) threshold in Mask-R-CNN, in order to achieve the best possible model performance. Further CNN-based approaches followed. Here, proven architectures such as U-Net4 were modified for wear detection, but a compact, reduceddimension form was also developed to detect normal and abnormal wear with high accuracy and to make statements about wear behavior. Different regularization techniques and hyperparameter optimization were applied to achieve the best model performance on two different highly reflective tool types.
As part of this work, state-of-the-art (SOTA) convolutional neural network (CNN) architectures for
tool classification were also investigated. The effect of different lighting conditions, image quality,
and different training strategies was studied to identify the most suitable architecture and training
strategy for AI-based tool classification
Die Wesentlichkeitsanalyse in der Nachhaltigkeitsberichterstattung nach CSRD: Eine kritische Würdigung
No full text
Analysis and forecasting of financial risks of German enterprises in the context of unpredictable price dynamics of energy resources
No full textThe relevance of this topic is due to the fact that at present in Germany there is an unpredictable price dynamics of energy resources, changes in interest rates, the dynamics of the euro exchange rate, and is also due to such factors as the instability of the political situation and the changing principles of state influence in the financial
sector.
Purpose of the work: to study the economic impact of financial risk on the German financial system, to explore
methods for measuring, identifying and managing financial risks at the company level.
Research methodology: In the course of the study, such modern methods of studying economic phenomena and
processes as a dialectical approach, system analysis, formal logic, statistical research techniques were used
Integrating cross-modality fusion for joint audio-visual quality assessment
No full textAssessing the quality of audio-visual (AV) content is essential for understanding how users perceive the overall experience of combined audio and video in the modern multimedia applications. Even though there has been a lot of work in assessing the quality of audio and video separately, the integration of these channels has not been well explored especially using advanced deep learning techniques. Current literature tends to use non-industry-standard datasets and simple fusion strategies, thus hindering the real-world relevance and advancement of the field. To address these gaps, this work conducts a series of experiments starting from simple approaches and baseline models, progressively advancing to more sophisticated methods such as integrating cross-modality fusion via cross and self-attention.
Our method incorporates a deep neural network based audio model along with a dedicated video quality model where the features of both modalities are fused using attention mechanisms. We extend the feature set of one modality when needed in order to preserve temporal correspondence and relevance. One of the novelties of the proposed approach is the usage of the Concordance Correlation Coefficient (CCC) loss, which is a measure that has been employed in emotion recognition but not in AV quality modeling before. This loss function increases the stability of our quality estimations since it forces the model to output quality scores that are closer to the actual human ratings.
We also make a unique contribution by using an AV dataset that is closer to the industry practices, which contains high quality audio and video content with realistic distortions. To the best of our knowledge, this is the first work that employs cross-attention for AV feature fusion in an intrusive AV quality assessment setting. The experimental results show that this approach can enhance the prediction performance and yield high Pearson and Spearman correlations as well as low RMSE. This work sets a new baseline for AV quality assessment and demonstrates how cross-modal fusion can be useful in real-life multimedia applications