Dipl.-Ing. Lucas Weber

This follow-up research project builds on previous work that demonstrated the value of change point analysis (CPA) for knowledge discovery in heterogeneous industrial time series data. The project extends existing methods to support large-scale analysis of complex thermodynamic systems, with a particular focus on power plant data characterized by high dimensionality, heterogeneity, and varying data quality. Beyond detecting individual change points, the developed methods enable higher-level analyses such as sequence-based pattern discovery and the identification of related or decoupled changes across hundreds of time series.

A central objective of the project is to transition the previously developed prototypical algorithms into a scalable, platform-based environment close to the data sources (cloud deployment). In collaboration with Siemens Energy platform teams, the project aims to improve scalability, availability, and iteration speed, enabling broader evaluation across diverse real-world use cases. This setup allows domain experts to apply the methods without direct access to source code and supports faster feedback cycles for algorithmic refinement.

The research further focuses on systematic evaluation of the algorithms on additional industrial use cases, expert-guided parameter selection, and adaptation of methods to handle edge cases and varying data quality. By strengthening the connection between algorithmic development, scalable infrastructure, and expert-driven evaluation, the project advances the practical applicability of change point analysis as a decision-support tool for industrial time series analysis.

This research project investigates the use of unsupervised change point analysis (CPA) methods for knowledge discovery in large-scale industrial time series datasets. Change point analysis offers a promising tool to mine large time series datasets by identifying abrupt and unexpected changes in time series data without requiring supervision.

The project systematically evaluates the applicability, feasibility, and parameterization of state-of-the-art CPA algorithms for real industrial data. Key activities include a comprehensive literature review, detailed data profiling and preparation, conceptual development of an evaluation framework, and quantitative comparison of multiple algorithms using representative use cases provided by industry partners. We extend the typical one-dimensional use-case, where CPA is applied to isolated signals, by comparing the extracted events of multiple signals to find relationships and to mine unexpected, isolated events.

In addition, a visualization prototype will be developed to support interpretation of detected changes and to demonstrate how results can be integrated into engineering workflows and monitoring processes.

By assessing whether detected change points correspond to meaningful anomalies/events in combined cycle power plants (steam generators specifically), the project aims to advance unsupervised data mining methods for complex industrial systems and provide practical guidance for their deployment in power plant monitoring.

Modern complex systems, such as power plants or other industrial structures, combined with the rise of IoT and Industry 4.0, produce thousands of time series measuring different aspects within these systems. As time series measure the state of these complex systems, the correct identification and integration of these time series are key to enabling advanced analytics and further optimization. As acquiring contextual information about each time series and their relations is currently a time-consuming and error-prone manual process, techniques to support or even automate this process are in high demand. While there are different available metadata formats, such as Brick, this metadata often is not available for all data sources and is not commonly used for all systems. Integrating time series at scale requires efficient algorithms and robust concepts that can deal with the heterogeneity and high volume of time series from different domains.

Additional Applications and Outcomes:

Changepoynt Python Package

Changepoint correlation heavily relies on suitable changepoint detection algorithms, many of which were implemented from research papers within a pip-installable package “changepoynt” (https://changepoynt.de). Changepoint detection, a critical task in time series analysis, identifies abrupt shifts or transitions in data patterns, offering insights into underlying phenomena. Developed with flexibility and scalability in mind, “changepoynt” integrates a range of state-of-the-art methods for changepoint detection, empowering researchers across domains to efficiently analyze and interpret their data.

CATCH: Contextual Anomaly Tracking with Changepoint Detection

Together with a research partner from the industry, we basically use the inverse of our idea of relationship discovery to detect contextual anomalies. The hypothesis of the project states that signals, which should have relations (e.g. Input-Output measurements of a dynamical system), behave anomalously if they stop showing simultaneous changes. In contrast to classical anomaly detection methods, change point anomaly (the comparison of multiple changepoint signals) is mainly targeted at contextual anomalies, where two signals are measuring the same component, and consequentially should change at similar times when the plant changes operational status. In case the signals change separately, a contextual anomaly occurs.  While the methods are available in theory, the project is necessary to test the applicability, feasibility, and correct parametrization of the methods for selected use cases. A demonstrator for a two-dimensional case can be found under https://anomaly.changescore.de/ and for the multi-dimensional case under https://heatmap.changescore.de/.