In recent years, there has been a significant rise in the popularity of time-series based machine learning. While historically image recognition and natural language processing have been at the forefront of machine learning, many of the successes achieved in these fields can also be applied to the time-series domain. This is particularly relevant in the context of Industry 4.0, where the demand for connectivity and the collection of sensor data has increased the need for intelligent methods to extract valuable insights from abundantly collected time series data.
However, industrial process data often present themselves as multi-phase and multivariate time series. These form of time series poses unique challenges for conventional machine learning tasks such as classification and segmentation. Therefore, this dissertation aims at advancing machine learning methods specifically designed for multi-phase multivariate time series. This is done on the example of the publicly available “Hydraulic End-of-Line Testing”-dataset, in which each sample represents the end-of-line test cycle of a freshly manufactured product.
Functional end-of-line testing is a powerful but costly approach for industrial quality assurance, with one of the main cost factors being the often excessive and unnecessarily complex design of industrial testing cycles. Consequently, at its core, the dissertation proposes a three-step machine learning-based approach for optimizing end-of-line test cycles. In order to enable this approach, the thesis explores and proposes new methods for time series segmentation, transfer learning, time series classification, and time series redundancy detection.