Refine
Document Type
- Master's Thesis (3)
- Book (1)
- Conference Proceeding (1)
Institute
Language
- English (5) (remove)
Keywords
- Machine learning (5) (remove)
The usage of data gathered for Industry 4.0 and smart factory scenarios continues to be a problem for companies of all sizes. This is often the case because they aim to start with complicated and time-intensive Machine Learning scenarios. This work evaluates the Process Capability Analysis (PCA) as a pragmatic, easy and quick way of leveraging the gathered machine data from the production process. The area of application considered is injection molding. After describing all the required domain knowledge, the paper presents an approach for a continuous analysis of all parts produced. Applying PCA results in multiple key performance indicators that allow for fast and comprehensible process monitoring. The corresponding visualizations provide the quality department with a tool to efficiently choose where and when quality checks need to be performed. The presented case study indicates the benefit of analyzing whole process data instead of considering only selected production samples. The use of machine data enables additional insights to be drawn about process stability and the associated product quality.
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.
The purpose of an energy model is to predict the energy consumption of a real system and to use this information to address challenges such as rising energy costs, emission reduction or variable energy availability. Industrial robots account for an important share of electrical energy consumption in production, which makes the creating of energy models for industrial robots desirable. Currently, energy modeling methods for industrial robots are often based on physical modeling methods. However, due to the increased availability of data and improved computing capabilities, data-driven modeling methods are also increasingly used in areas such as modeling and system identification of dynamic systems. This work investigates the use of current data-driven modeling methods for the creation of energy models focusing on the energy consumption of industrial robots.
For this purpose, a robotic system is excited with various trajectories to obtain meaningful data about the system behavior. This data is used to train different artificial neural network (ANN) structures, where the structures used can be categorized into (i) Long Short Term Memory Neural Network (LSTM) with manual feature engineering, where meaningful features are extracted using deeper insights into the system under consideration, and (ii) LSTM with Convolutional layers for automatic feature extraction. The results show that models with automatic feature extraction are competitive with those using manually extracted features. In addition to the performance comparison, the learned filter kernels were further investigated, whereby similarities between the manually and automatically extracted features could be observed. Finally, to determine the usefulness of the derived models, the best-performing model was selected for demonstrating its performance on a real use case.
Many test drives are carried out in the automotive environment. During these test drives many signals are recorded. The task of the test engineers is to find certain patterns (e.g. an emergency stop) in these long time series. Finding these interesting patterns is currently done with rule based processing. This procedure is very time consuming and requires a test engineer with expertise. In this thesis it is examined if the emerging field of machine learning can be used to support the engineers in this task. Active Learning, a subarea of machine learning, is used to train a classifier during the labeling process. Thereby it proposes similar windows to the already labeled ones. This saves the annotator time for searching or formulating rules for the problem. A data generator is worked out to replace the missing labeled data for tests. The custom performance measure “proportion of seen samples” is developed to make the success measurable. A modular software architecture is designed. With that, several combinations of Time Series Classification algorithms and query strategies are compared on artificial data. The results are verified on real datasets, which are open source available. The best performing, but computational intensive solution is an adapted RandOm Convolutional KErnel Transform (ROCKET). The custom query strategy “certainty sampling” shows the best results for highly imbalanced datasets.