Visual Interactive Exploration and Labeling of Large Volumes of Industrial Time Series Data

Tristan Langer, Viktor Welbers, Yannik Hahn, Mark Wönkhaus, Richard Meyes, Tobias Meisen

Abstract: In recent years, supervised machine learning models have become increasingly important for the advancing digitalization of the manufacturing industry. Reports from research and application show potentials in the use for application scenarios, such as predictive quality or predictive maintenance, that promise flexibility and savings. However, such data-based learning methods require a large training sets of accurately labeled sensor data that represents the manufacturing process in the digital world and allow model to learn corresponding behavioral patterns. Nevertheless, the creation of these data sets cannot be fully automated and requires the knowledge of process experts to interpret the sensor curves. Consequently, the creation of such a data set is time-consuming and expensive for the companies. Existing solutions do not meet the needs of the manufacturing industry as they cannot visualize large data sets, do not support all common sensor data forms and offer little support for efficient labeling of large data volumes. In this paper, we build on our previously presented visual interactive labeling tool Gideon-TS that is designed for handling large data sets of industrial sensor data in multiple modalities (univariate, multivariate, segments or whole time series, with and without timestamps). Gideon-TS also features an approach for semi-automatic labeling that reduces the time needed to label large volumes of data. Based on the requirements of a new use case, we extend the capabilities of our tool by improving the aggregation functionality for visualizing large data queries and by adding support for small time units. We also improve our labeling support system with an active learning component to further accelerate the labeling process. We evaluate the extended version of Gideon-TS on two industrial exemplary use cases by conducting performance tests and by performing a user study to show that our tool is suitable for labeling large volumes of industrial sensor data and significantly reduces labeling time compared to traditional labeling methods.