Informatik
Refine
Document Type
- Master's Thesis (16)
Language
- English (16) (remove)
Has Fulltext
- yes (16)
Is part of the Bibliography
- no (16)
Keywords
- ROS (2)
- Automated optimisation (1)
- Blockchain (1)
- CNN (1)
- Circular Economy (1)
- Compositional data (1)
- Computer Vision (1)
- DLT (1)
- Data analysis (1)
- Distributed Ledger Technology (1)
This thesis evaluates the feasibility of conducting visual inspection tests on power industry constructions using object detection techniques. The introduction provides an overview of this field’s state-of-the-art technologies and approaches. For the implementation, a case study is then conducted, which is done in collaboration with the partner company OMICRON Electronics GmbH, focusing on power transformers as an example. The objective is to develop an inspection test using photographs to identify power transformers and their subcomponents and detect existing rust spots and oil leaks within these components. Three object detection models are trained: one for power transformers and sub-components, one for rust detection, and one for oil leak detection. The training process utilizes the implementation of the YOLOv5 algorithm on a Linux-based workstation with an NVIDIA Quadro RTX 4000 GPU. The power transformer model is trained on a dataset provided by the partner company, while open-source datasets are used for rust and oil leak detection. The study highlights the need for a more powerful GPU to enhance training experiments and utilizes an Azure DevOps Pipeline to optimize the workflow. The performance of the power transformer detection model is satisfactory but influenced by image angles and an imbalance of certain sub-components in the dataset. Multi-angle video footage is a proposed solution for the inspection test to address this limitation and increase the size of the dataset, focusing on reducing the imbalance. The models trained on open-source datasets demonstrate the potential for rust and oil leak detection but lack accuracy due to their generic nature. Therefore, the datasets must be adjusted with case-specific data to achieve the desired accuracy for reliable visual inspection tests. The results of the case study have been well-received by the partner company’s management, indicating future development opportunities. This case study will likely be a foundation for implementing visual inspection tests as a product.
Lack of transparency and traceability of products and their raw materials means that most products can only be thrown away or not properly recycled due to a lack of relevant data. This conflicts with the circular economy principles, which are demanded by several initiatives, including the European Union. The aim of this master thesis is to analyze this conflict and to propose a technical solution based on Distributed Ledger Technology that enables transparency and traceability of products and their materials. Therefore, the thesis addresses two central research questions: 1. How can traceability and transparency be enabled by integrating a DLT solution? 2. How would a prototype with the integration of smart contracts and DLT look like? To answer these questions, a blockchain solution is implemented using Hyperledger Fabric. The solution uses the immutability and decentralized nature of DLT to record and track the movement of products and their materials throughout their life cycle in the Circular Economy. Furthermore, with private data collections, confidentiality, and privacy are granted while ensuring transparency. The thesis contributes to the Circular Economy field by exploring the principles, models, and challenges of the Circular Economy and the circularity goals of a Digital Product Passport to develop a suitable technical solution. The chosen blockchain framework, Hyperledger Fabric, is presented, and its key components and features are highlighted. The thesis also delves into the design decisions and considerations behind the Digital Product Passport platform, explaining the architecture and transaction flow together with the prototype implementation and demonstration to showcase the functionality of the solution. Results and analysis provide insights into the challenges of the Circular Economy, sustainable resource management, and the Digital Product Passport, resulting in recommendations for future improvements and enhancements. Overall, this thesis offers a practical solution utilizing DLT to enable transparency and traceability in the Circular Economy, contributing to the realization of sustainable and efficient resource management practices to ultimately contribute to the set Circular Economy initiatives.
Programmable Logic Controller (PLC) modules are used in industrial settings to control and monitor various manufacturing processes. Detecting these modules can be helpful during installation and maintenance. However, the limited availability of real annotated images to train an object detector poses a challenge. This thesis aims to research object detection of these modules on real images by using synthetic data during training. The synthetic images are generated from CAD models and improved with Generative Adversarial Networks (GANs). The CAD models are rendered in different scenes, and perfectly annotated images are automatically saved. A technique called domain randomization is applied during rendering. It renders the modules in different poses with constantly changing backgrounds. As the CAD models do not visually resemble the real modules, it is necessary to improve the synthetic images. This project researches StarGAN and CycleGAN for the task of image-to-image translation. A GAN is trained with real and synthetic images and can then translate between these domains. YOLOv8 and Faster R-CNN are tested for object detection. The best mean Average Precision (mAP) is achieved when training with a synthetic dataset where 50% of the images were improved with StarGAN. When trained with YOLOv8 and evaluated on a real dataset, it achieves a mAP of 84.4%. Overall, the accuracy depends on the quality of the CAD models. Using a GAN improves the detection rate for all modules, but especially for unrealistic CAD models.
Systems are constantly increasing in complexity. This poses challenges to managing and using system knowledge. The Systems Modeling Language (SysML) is a modeling language specifically for systems, while Machine Learning (ML) is a tool to tackle complex problems. Currently, no bridge between systems modelled in SysML and ML regarding said systems has been proposed in literature. This thesis presents an approach that uses Model-driven Software Engineering (MDSE) and Template-based Code Generation (TBCG) to generate a ML IPython Notebook (IPYNB) from a SysML model. A mapping configuration using JavaScript Object Notation (JSON) allows the definition of mappings between SysML elements and template variables, enabling configuration and user-supplied templates. To test the approach, a SysML model describing ML to predict the weather based on data is created. Python ML templates are supplied and template variables mapped with the JSON mapping configuration are proposed in the thesis. The outcome is an executable IPYNB that contains all information from the SysML model and follows the modelled workflow. The findings of the work show that model-driven ML using SysML as a modeling language is beneficial due to the representation of ML knowledge in a general-purpose modeling language and the reusability of SysML model elements. It further shows that TBCG and a mapping configuration allow for more flexible code generation without changing the source implementation.
Offline speech to text engine for delimited context in combination with an offline speech assistant
(2022)
The inatura museum in Dornbirn had planned an interactive speech assistant-like exhibit. The concept was that visitors could ask the exhibit several questions that they would like to ask a flower. Solution requirements regarding the functionalities were formulated, such as the capacity to run offline because of privacy reasons. Due to the similarity of the exhibit, open-source offline Speech To Text (STT) engines and speech assistants were examined. Proprietary cloud-based STT engines associated with the corresponding speech assistants were also researched. The aim behind this was to evaluate the hypothesis of whether an open-source offline STT engine can compete with a proprietary cloud-based STT engine. Additionally, a suitable STT engine or speech assistant would need to be evaluated. Furthermore, analysis regarding the adaption possibilities of the STT models took place. After the technical analysis, the decision in favour of the STT engines called "Vosk" was made. This analysis was followed by attempts to adapt the model of Vosk. Vosk was compared to proprietary cloud-based Google Cloud Speech to Text to evaluate the hypothesis. The comparison resulted in not much of a significant difference between Vosk and Google Cloud Speech to Text. Due to this result, a recommendation to use Vosk for the exhibit was given. Due to the lack of intent parsing functionality, two algorithms called "text matching algorithm" and "text and keyword matching algorithm" were implemented and tested. This test proved that the text and keyword matching algorithm performed better, with an average success rate of 83.93 %. Consequently, this algorithm was recommended for the intent parsing of the exhibit. In the end, potential adaption possibilities for the algorithms were given, such as using a different string matching library. Some improvements regarding the exhibit were also presented.
The demand for managing data across multiple domains for product creation is steadily increasing. Model-Driven Systems Engineering (MDSE) is a solution for this problem. With MDSE, domain-specific data is formalized inside a model with a custom language, for example, the Unified Modelling Language (UML). These models can be created with custom editors, and specialized domains can be integrated with extensions to UML, e.g., the Systems Modeling Language (SysML). The most dominant editor in the open-source sector is Eclipse Papyrus SysML 1.6 (Papyrus), an editor to create SysML diagrams for MDSE.
In the pursuit of creating a model and diagrams, the editor does not support the user appropriately or even hinders them. Therefore, paradigms from the diagram modelling and Human Computer Interaction (HCI) domains, as well as perceptual and design theory, are applied to create an editor prototype from scratch. The changes fall into the categories of hierarchy, aid in the diagram composition, and navigation. The prototype is compared with Papyrus in a user test to determine if the changes have the effect of improving usability.
The study involved 10 participants with different knowledge levels of UML, ranging from beginners to experts. Each participant was tested on a navigation and modelling task in both the newly created editor, named Modelling Studio, and Papyrus. The study was evaluated through a questionnaire and analysis of the diagrams produced by the tasks.
The findings are that Modelling Studio’s changes to the hierarchical elements improved their rating. Furthermore, aid for diagram composition could be reinforced by changes to the alignment helper tool and adjustments to the default arrow behaviour of a diagram. Lastly, model navigation adjustments improve a link’s visibility and rating of a specialized link (best practice). The introduction of breadcrumbs had limited success in bettering navigation usability. The prototype deployed a broad spectrum of changes that found improvement already, which can, however, be further improved and tested more thoroughly.
Erosion due to cavitation is a common problem for any kind of water turbine. Most of the currently used techniques to detect cavitation are using an Acoustic Emission (AE) sensor and highspeed cameras during operation. For the pelton wheel which is subject of this thesis it is impossible to take pictures during operation, because of the splashing water and the mist. Therefore this thesis aims to explore possibilities in detecting erosion on the buckets of the pelton wheel on images taken during manual inspections. Since the provided images are snapshots taken with a mobile phone camera without a tripod, a lot of effort was invested in the preprocessing of the images. For the main task, the classification of the erosion, two methods were evaluated: Local Binary Patterns (LBP) + kN-earest neighbor classification and the classification with a Convolutional Neural Network (CNN). The given 2405 images, contained 4810 buckets on which the erosion was graded from zero to four. This means the baseline for the classification accuracy is 20%. LBP + kNearest neighbor classification scored 32.03%. The chosen CNN model, a light version of the Xception architecture outperformed the LBP + kNearest classification with 58,29%. The biggest issue found during research is the variance of the erosion grading by the maintainance personnel. Reasons for this are: no objective grading critera like the area of erosion in mm2, classification by different employees, a shift in grading from overall bucket condition to erosion from cavitation and too many classes for grading. The mentioned reasons were confirmed by the manual classification experiment were an IllwerkeVKW employee had to perform the grading on images of the dataset. The contestants accuracy score was 36% for this task. The result of 58,29% classification accuracy indicates that an automated grading of erosion by cavitation is feasible.
Skiing is one of the most popular winter sports in the world and especially in the alps. As the skiers enjoy their time on the slopes the most annoying thing that could happen is long waiting times at a lift. Unfortunately, because of climate changes, this happens more regularly because smaller skiing areas at lower altitudes have to close and the number of good skiing days decreases as well. This leads to a increase in the number of skiers in the skiing areas which inevitably leads to longer waiting times and dissatisfied skiers. To prevent this from happening, the carriers of the skiing areas have to manage the skiers flow and distribution and what better way to analyse the current situation and possible changes then by simulating the whole area. A simulation has the advantage of being flexible with regards to time as well as configuration. Be it simulating a skiing day and look into detail of the behaviour of a single skier and how it moves in the area by simulating in real time or setting the focus to the whole area and find out when and where queues are forming throughout the whole day by speeding up the time and simulate the day in only seconds, everything is possible. Even simulating a scenario where some part of the area is closed and the skiers cannot take specific lifts due to some technical error or some slopes because of to less snow. By simulating and analysing all these scenarios not only does the experts of the skiing area gain valuable statistical information about the area but can also simulate changes to the system like a crowd fl ow control or an increase or decrease in capacity of a lift. The simulation built in context with this work for the skiing area of Mellau shows all those applications but can also be used as a basis for further improvements of the skiing area or be expanded to other areas like Damüls. The simulation was implemented using the Anylogic simulation environment and the statistical evaluation was also performed in this program.
With the rise of people wearing smartwatches and the ever-lasting issue of stress, there has been an interest in detecting stress with wearables in real-time. This allows for interventions that take place exactly when stress occurs. However, many situations require all of our attention, making them unsuitable for any interventions. Additionally, many approaches currently do not factor in this aspect, running the risk of offering users undesirable interventions.
This thesis examines how contextual user information can be incorporated into a stress intervention system to reduce undesirable intervention timings. The system is split into detecting stress using heart rate variability (HRV) metrics obtained from a photoplethysmography (PPG) signal, and inferring user context from available sensor data. It is evaluated with a simulation-based approach using daily schedules of created personas and randomly sampled stressors during daily life.
The results obtained indicate the benefit of adding contextual user information to a stress intervention system. Depending on the busyness of the schedule, it can greatly decrease the number of received interventions. However, as these findings are attained without performing a user testing, it is unclear how they compare to results from real-world usage.