Search

ZusammenfassungVirtual Reality (VR) gilt als vielversprechende Technologie, nicht zuletzt deshalb, weil damit komplexe Inhalte vermittelt werden können. Da VR ein hohes Potenzial für Interaktivität und Immersion aufweist, findet die Technologie auch in der Industrie Anwendung, beispielsweise für die realitätsnahe Simulation von Montage- und Instandhaltungsarbeiten. Wissenschaftliche Befunde zur Wirksamkeit des Lernens in industriellen VR-Umgebungen existieren bislang kaum. Es ist daher auch unklar, wie Lerneffekte in virtuellen Umgebungen verbessert werden können. Der vorliegende Artikel untersucht, ob VR-gestütztes Training im Vergleich zu klassischem training-on-the-job (begleitet von einem Tutor) zu einer Steigerung des Lernerfolgs führt. Basierend auf einem VR-Trainingstool zum Erlernen von Montageabläufen wurde eine Laborstudie mit 24 ProbandInnen durchgeführt. Die Ergebnisse zeigen, dass VR-gestütztes Training zwar als weniger anstrengend empfunden wurde, der Lernerfolg jedoch im traditionellen training-on-the-job signifikant besser ausfiel. Auf Basis dieser Erkenntnisse werden die Chancen und Risiken aktueller VR-Trainingssimulationen diskutiert und konkrete Handlungsempfehlungen zur Steigerung des Lerneffekts formuliert.

The identification of appropriate manufacturing process parameters typically relies on rule-based schemes, expertise, and domain knowledge of highly skilled workers. Usually, the parameter settings remain the same for each part in an individual production lot once an acceptable quality is reached. Each part, however, has slightly different properties and part-specific parameter settings have the opportunity to increase quality and reduce scrap. We propose a simple linear regression model to identify process parameters based on experimental data and extend that model with ideas from time series analysis to achieve highly-accurate, part-specific parameter settings in a real-world manufacturing use case. We show the usefulness of exploiting the (autocorrelated) structure of regression residuals to improve the predictive performance of regression models in manufacturing environments.

This paper analyzes the process variable data from injection molding processes to identify the key process variables, which can be predicted by other process variables, which highlights the interdependence among different process variables in various production scenarios. The available data from injection molding machines provide information for the run-time, setup parameters of machines, and measurements of different process variables through sensors. For predictive modeling, we employed different linear regression models with the recursive backward feature selection and SVM regression models using a radial kernel to predict nonlinear process variables. We also applied the linear and SVM regression models for outlier data in the process variables assuming that the upper bound outliers represent the perturbed state of process variables during production. These perturbations are affected by material type, machine type (age and performance), regime changes, or other external effects and subsequently affect the predictability of process variables and production output. Such cases are different compared to the normal or controlled range of process variables. Our analysis shows that the predictability varies for different material types due to the interdependence of the associated process variables. We further highlight that various process variables exhibit nonlinear relationships and cannot be predicted using linear models. We additionally look for the interdependence of process variables used previously by three studies as input features to predict product quality.

The goal of this paper is to test if data partitioning based on domain knowledge improves the performance of unsupervised anomaly detection algorithms for the detection of faulty internal combustion engines. We test three common anomaly detection algorithms to predict defective engines at the end of an assembly line, both with and without data partitioning. The algorithms are trained on high-dimensional vibration data. To evaluate and compare the detection performance of partitioned and unpartitioned approaches, we use a labeled collection of known anomalies. Using domain knowledge to partition the data improves the anomaly detection performance of all tested algorithms. A divide and conquer strategy based on data partitioning, thus, appears to be a viable anomaly detection approach in cases where abundant unlabeled data is available.

Technological change, based on digitalization, is happening fast and companies have to place emphasis on qualifying their employees and improve their infrastructure to be up to date. Learning Factories are adequate facilities for qualifying employees in a safe environment with proper guidance and infrastructure. However, there are three important questions in connection to handling the change in companies. The first is “How is the qualification and infrastructure status quo in a company?” The second is “Which readiness level should be achieved to handle the change appropriate?” The third is “What is the gap and how to close it in terms of qualification and infrastructure?” For answering the first question, an investigation in the company is necessary and a rating has to be conducted. Nevertheless, a standardized maturity model can help to answer this question. The maturity model also supports the answer to the second question. By using the reference table a target level can be defined. The answer to the third question is the delta between the status quo and the target level of the company’s readiness investigation. The gap between the status quo and target is usually indicated by infrastructure, process, organizational and qualification improvements. Furthermore, the gap and hence the necessary improvement is the basis for training in a learning factory. Derived from the gap, the learning factory can prepare a training curriculum and the needed infrastructure to train the staff with the utmost precision.

This paper analyzes production data from injection molding processes to identify key interactions between the process variables from different material categories using the network inference method called "bagging conservative causal core network" (BC3net). This approach is an ensemble method with mutual information that is measured between process variables to select pairs that show significant shared information. We construct networks for different time intervals and aggregate them by calculating the proportion of significant pairs of process variables (weighted edges) for each production process over time. The weighted edges of the aggregated network for each product are used in a machine learning model to optimize the network interval size (interval split) and feature selection, where edge weights are the input features and material categories are the output classification labels. The time intervals are optimized based on the classification accuracy of the machine learning model. Our analysis shows that the aggregated edge features of inferred networks can classify different material categories and identify critical features that represent interdependence in the associated process variables. We further used the "one vs. other" labels for the machine learning models to identify material-specific interactions for each material category. Additionally, we constructed an aggregated network over all samples in which the process variable interactions were steady over time. The resulting network showed modular characteristics where process variables of similar categories were grouped in the same community. publishedVersion