Beschreibung
Comprehensive, accurately labeled sensor datasets are an essential prerequisite for training supervised machine learning models used for tasks such as quality control, predictive maintenance, and defect detection in the manufacturing industry. However, the provision of such datasets still poses two specific challenges: first, performing exploratory data analysis (EDA) to provide data scientists with the necessary knowledge in the domain context to label the data, and second, a lack of visual interactive labeling (VIAL) approaches to efficiently annotate large volumes of industrial sensor data with accurate labels. This dissertation proposes the innovative VIEDAL process, integrating guidance systems for both EDA and VIAL tasks. Drawing from real-world use cases, this thesis presents a detailed system design to support each task, addressing feasibility and usefulness through a comprehensive design study. The EDA guidance system records domain expert interactions to generate guided sessions for novices, while the VIAL guidance system incorporates unsupervised and active learning approaches to streamline dataset annotation. Through user studies, the effectiveness of the proposed systems is evaluated, demonstrating reproducibility of expert key insights through generated EDA sessions and faster creation high quality labeled datasets. Additionally, this work discusses approaches for transferring recorded EDA sessions and VIAL models between use cases to streamline future guidance system implementations. The results of this thesis provide a foundational for further research to expedite the creation of labeled sensor datasets, thereby facilitating faster development and integration of machine learning models for enhancing production processes in the manufacturing industry.
