Authors:
Tristan Langer
;
Viktor Welbers
and
Tobias Meisen
Affiliation:
Chair of Technologies and Management of Digital Transformation, University of Wuppertal, Lise-Meitner-Straße 27-31, 42119 Wuppertal, Germany
Keyword(s):
Labeling, Time Series Analysis, Sensor Data, Visual Analytics.
Abstract:
Modern digitization in industrial production requires the acquisition of process data that is subsequently used in analysis and optimization scenarios. For this purpose, the use of machine learning methods has become more and more established in recent years. However, training advanced machine learning models from scratch requires a lot of labeled data. The creation of such labeled data is a major challenge for many companies, as the generation process cannot be fully automated and is therefore very time-consuming and expensive. Thus, the need for corresponding software tools to label complex data streams, such as sensor data, is steadily increasing. Existing contributions are not designed for handling large datasets and forms common for industrial applications, and offer little support for the labeling of large data volumes. For this reason, we introduce Gideon-TS — an interactive labeling tool for sensor data that is tailored to the needs of industrial use. Gideon-TS can integrate
time series datasets in multiple modalities (univariate, multivariate, samples, with and without timestamp) and remains performant even with large datasets. We also present an approach to semi-automatic labeling that reduces the time needed to label large volumes of data. We evaluated Gideon-TS on an industrial exemplary use case by conducting performance tests and a user study to show that it is suitable for labeling large datasets and significantly reduces labeling time compared to traditional labeling methods.
(More)