Automating Industrial Event Stream Analytics: Methods, Models, and Tools

Industrial event streams are an important cornerstone of Industrial Internet of Things (IIoT) applications. For instance, in the manufacturing domain, such streams are typically produced by distributed industrial assets at high frequency on the shop floor. To add business value and extract the full potential of the data (e.g. through predictive quality assessment or maintenance), industrial event stream analytics is an essential building block. One major challenge is the distribution of required technical and domain knowledge across several roles, which makes the realization of analytics projects time-consuming and error-prone. For instance, accessing industrial data sources requires a high level of technical skills due to a large heterogeneity of protocols and formats. To reduce the technical overhead of current approaches, several problems must be addressed. The goal is to enable so-called "citizen technologists" to evaluate event streams through a self-service approach. This requires new methods and models that cover the entire data analytics cycle.
In this thesis, the research question is answered, how citizen technologists can be facilitated to independently perform industrial event stream analytics. ... mehrThe first step is to investigate how the technical complexity of modeling and connecting industrial data sources can be reduced. Subsequently, it is analyzed how the event streams can be automatically adapted (directly at the edge), to meet the requirements of data consumers and the infrastructure. Finally, this thesis examines how machine learning models for industrial event streams can be trained in an automated way to evaluate previously integrated data. The main research contributions of this work are:
1. A semantics-based adapter model to describe industrial data sources and to automatically generate adapter instances on edge nodes.
2. An extension for publish-subscribe systems that dynamically reduces event streams while considering requirements of downstream algorithms.
3. A novel AutoML approach to enable citizen data scientists to train and deploy supervised ML models for industrial event streams.
The developed approaches are fully implemented in various high-quality software artifacts. These have been integrated into a large open-source project, which enables rapid adoption of the novel concepts into real-world environments. For the evaluation, two user studies to investigate the usability, as well as performance and accuracy tests of the individual components were performed.