Skip to Content

Research


Research focusExtended adaptation in data streamsOnline-adaptive cluster analysisCluster-based anomaly and drift detectionOnline-adaptive controlMLPro – Open-source framework and reference architectureAll publications

Research focus

We conduct research in the field of Online Machine Learning (OML). This branch of artificial intelligence focuses on methods and systems that continuously adapt to changing conditions during operation. This key capability distinguishes them from widely deployed AI systems that are trained on large datasets and do not learn further once deployed.

OML methods are applied in the real-time analysis of data streams, for example in sensor networks. A typical application is the detection of anomalies and drift in incoming data, which in industrial contexts can be used for the early identification and mitigation of technical faults (predictive maintenance). OML also enables early detection capabilities in the analysis of daily financial data and in business analytics and business intelligence contexts. In addition, OML methods are applied in adaptive control and adaptive simulation.

blueAIC contributes to the advancement of this field through its own research and peer-reviewed publications. Our agenda focuses on the following areas:

  • Novel adaptation mechanisms for data stream applications
  • Efficient online-adaptive algorithms
  • Online-adaptive control
  • Online-adaptive simulation

For the structured and production-oriented implementation of selected methods, we use the open-source framework MLPro.

Previous contributions were developed at the South Westphalia University of Applied Sciences (Germany), Department of Electrical Engineering, in the research group of Automation Technology and Learning Systems.


Top


Extended adaptation in data streams

This research line addresses novel adaptation mechanisms in data stream environments that go beyond purely forward-directed incremental model updates. The focus lies on OML applications understood as structured stream workflows, composed of adaptive stream tasks and their interaction through cooperative adaptation mechanisms:

  • Event-oriented adaptation

  • Cross-task adaptation cascades

  • Reverse adaptation


Publications
Authors

DetlefArend, Laxmikant ShrikantBaheti, SteveYuwono, Syamraj Purushamparambil SatheeshKumar, AndreasSchwung

Journal

Machine Learning with Applications, Vol. 21, September 2025, Article 100715

Abstract

In this paper, we present version 2.0 of the open-source middleware MLPro for applied machine learning in Python. Notably, it introduces the new sub-framework MLPro-OA for online machine learning, focusing on standards and templates for classic and online-adaptive data stream processing (DSP/OADSP). As part of this, we provide three novel adaptation mechanisms:The first, event-oriented adaptation, enables localized, event-driven parameter updates within individual tasks. The second, cascaded adaptation, allows adaptation events to propagate across multiple dependent tasks, creating task-spanning adjustment cascades decoupled from the forward-facing DSP. The third, reverse adaptation, allows tasks to revise prior adjustments by explicitly processing obsolete instances discarded from a preceding sliding window. Furthermore, we provide insights into the underlying design criteria of MLPro-OA, which were developed through extensive requirements engineering. In the practical part of this work, we demonstrate the essential functionalities of MLPro-OA using reproducible examples.

DOI

10.1016/j.mlwa.2025.100715

Supplementary materials
Authors

Detlef Arend, Andreas Schwung

Conference

IEEE, 2025 IEEE 8th International Conference on Industrial Cyber-Physical Systems (ICPS)

Abstract

There are essentially two global trends that are currently heating up the AI sector. An increasing number of large language models (LLM) creates enormous resource requirements due to extensive data storage and complex training runs. Furthermore, advancing digitalization leads to an increased volume of data and the associated need for real-time analysis. In both areas, it is becoming apparent that the established AI models are reaching their limits despite all optimizations. This is where online machine learning comes in, with new models optimized to process incoming data in real-time and continuously adapt to it. This brings special requirements for the algorithms and their orchestration in extensive stream workflows. In this article, we focus on the latter aspect and highlight the two concepts of cascaded and reverse adaptation, which enable significant functional improvements individually and in combination. A corresponding extension of our open-source framework MLPro also enables us to demonstrate the adaptation mechanisms mentioned practically.

DOI

10.1109/ICPS65515.2025.11087907

Supplementary materials

GitHub - fhswf/paper-da-ieee-icps-2025: Paper IEEE ICPS 2025 - Arend/Schwung - "Online-adaptive data stream processing via cascaded and reverse adaptation"


Top


Online-adaptive cluster analysis

A current focus lies on the development of a framework for online-adaptive cluster analysis. The architecture supports a variable number of clusters as well as an extended statistical data representation for each cluster. The design emphasizes efficient update mechanisms and the replaceability and extensibility of core model components.

A scientific publication on this topic is currently in preparation. Further details will be provided after publication.


Top


Cluster-based anomaly and drift detection

Building on our research in online-adaptive cluster analysis, we investigate efficient algorithms for the detection of anomalies and drift in evolving cluster sets. These approaches enable a wide range of practical applications for the automated early detection of changes and irregularities in numerical data streams.

A scientific publication on this topic is currently in preparation. Further details will be provided after publication.


Top


Online-adaptive control

In addition to the analysis of continuous data streams, we investigate online-adaptive control methods at the intersection of classical control engineering and reinforcement learning.

Learning models are directly integrated into control loops in order to adapt and optimize controllers under changing operating conditions in a data-driven manner.

For methodological implementation, a reusable sub-framework for online-adaptive control and learning procedures has been implemented within MLPro. It enables the structured combination of classical control algorithms with learning components. In particular, it includes a PID controller whose parameters are adapted to evolving control plants using reinforcement learning.


Publications
Authors

Detlef Arend, Amerik Toni Singh Padda, Andreas Schwung, Dorothea Schwung

Conference

IEEE, 2025 11th International Conference on Control, Decision and Information Technologies (CoDIT)

Abstract

This paper presents the novel RLPID architecture for online-adaptive control, which combines classical proportional-integral-derivative (PID) control with reinforcement learning (RL). This hybrid approach enables dynamic online adjustment of PID parameters during control operation. Specifically, we propose a multi-objective reward structure that integrates established control criteria and analyze suitable configurations for different system dynamics. The RLPID controller has been implemented within the open-source middleware MLPro, where it is embedded in newly developed sub-frameworks for classical and online-adaptive control. Owing to its hybrid nature, the architecture can be used both in traditional control loops and within the Markov decision process of RL. Its effectiveness and practical applicability are demonstrated in a cascade control scenario.

DOI

10.1109/CoDIT66093.2025.11321229

Supplementary materials

GitHub - fhswf/paper-da-ieee-codit-2025: Paper IEEE CoDIT 2025 - Arend/Singh Padda/Schwung - "Online-adaptive PID control using Reinforcement Learning"


Top


MLPro – Open-source framework and reference architecture

MLPro is an open-source engineering middleware that has been developed since 2021 within the academic environment of the South Westphalia University of Applied Sciences (Germany) in collaboration with additional researchers and students.

The architecture is modular and comprises several sub-frameworks, including Online Machine Learning (MLPro-OA), Reinforcement Learning (MLPro-RL), Game Theory (MLPro-GT), and Supervised Learning (MLPro-SL). MLPro is therefore not a pure OML framework, but an integrative software foundation for diverse learning and decision-making approaches.

It provides a consistent and extensible structure in which selected adaptation, analysis, learning, and control methods can be implemented, combined, and further developed.

MLPro serves both as a technical reference architecture and as a sustainable implementation basis for published and future work. Accordingly, experiments and references to MLPro are included throughout our publications. The framework itself has been described in detail in several peer-reviewed papers.


Publications
Authors

DetlefArend, Laxmikant ShrikantBaheti, SteveYuwono, Syamraj Purushamparambil SatheeshKumar, AndreasSchwung

Journal

Machine Learning with Applications, Vol. 21, September 2025, Article 100715

Abstract

In this paper, we present version 2.0 of the open-source middleware MLPro for applied machine learning in Python. Notably, it introduces the new sub-framework MLPro-OA for online machine learning, focusing on standards and templates for classic and online-adaptive data stream processing (DSP/OADSP). As part of this, we provide three novel adaptation mechanisms:The first, event-oriented adaptation, enables localized, event-driven parameter updates within individual tasks. The second, cascaded adaptation, allows adaptation events to propagate across multiple dependent tasks, creating task-spanning adjustment cascades decoupled from the forward-facing DSP. The third, reverse adaptation, allows tasks to revise prior adjustments by explicitly processing obsolete instances discarded from a preceding sliding window. Furthermore, we provide insights into the underlying design criteria of MLPro-OA, which were developed through extensive requirements engineering. In the practical part of this work, we demonstrate the essential functionalities of MLPro-OA using reproducible examples.

DOI

10.1016/j.mlwa.2025.100715

Supplementary materials
Authors

Steve Yuwono, Detlef Arend, Andreas Schwung

Journal

SoftwareX, Vol. 28, December 2024, Article 101956

Abstract

Game theory, a fundamental aspect of mathematical economics and strategic decision-making, has been increasingly applied to various fields, including economics, biology, computer science, and engineering. Despite its growing importance, there is a significant lack of flexible and user-friendly tools for standardized modeling of them, particularly for real-world applications. Hence, we developed MLPro-GT as part of our open-source MLPro project, which offers modular and standardized yet flexible components, extensive documentation, and a variety of examples. MLPro-GT allows researchers and practitioners to easily incorporate game theory into their applications while lowering the entry barrier for students. This makes individual work more reproducible, shareable, and reusable.

DOI

10.1016/j.softx.2024.101956

Authors

Detlef Arend,  Mochammad Rizky Diprasetya, Steve Yuwono, Andreas Schwung

Journal

Software Impacts, Vol. 14, December 2022, Article 100421

Abstract

In recent years, many powerful software packages have been released on various aspects of machine learning (ML). However, there is still a lack of holistic development environments for the standardized creation of ML applications. The current practice is that researchers, developers, engineers and students have to piece together functionalities from several packages in their own applications. This prompted us to develop the integrative middleware framework MLPro that embeds flexible and recombinable ML models into standardized processes for training and real operations. In addition, it integrates numerous common open source frameworks and thus standardizes their use. A meticulously designed architecture combined with a powerful foundation of overarching basic functionalities ensures maximum recombinability and extensibility. In the first version of MLPro, we provide sub-frameworks for reinforcement learning (RL) and game theory (GT).

DOI

10.1016/j.simpa.2022.100421

Authors

Detlef Arend,  Steve Yuwono, Mochammad Rizky Diprasetya, Andreas Schwung

Journal

Machine Learning with Applications, Vol. 9, September 2022, Article 100341

Abstract

Nowadays there are numerous powerful software packages available for most areas of machine learning (ML). These can be roughly divided into frameworks that solve detailed aspects of ML and those that pursue holistic approaches for one or two learning paradigms. For the implementation of own ML applications, several packages often have to be involved and integrated through individual coding. The latter aspect in particular makes it difficult for newcomers to get started. It also makes a comparison with other works difficult, if not impossible. Especially in the area of reinforcement learning (RL), there is a lack of frameworks that fully implement the current concepts up to multi-agents (MARL) and model-based agents (MBRL). For the related field of game theory (GT), there are hardly any packages available that aim to solve real-world applications. Here we would like to make a contribution and propose the new framework MLPro, which is designed for the holistic realization of hybrid ML applications across all learning paradigms. This is made possible by an additional base layer in which the fundamentals of ML (interaction, adaptation, training, hyperparameter optimization) are defined on an abstract level. In contrast, concrete learning paradigms are implemented in higher sub-frameworks that build on the conventions of this additional base layer. This ensures a high degree of standardization and functional recombinability. Proven concepts and algorithms of existing frameworks can still be used. The first version of MLPro includes sub-frameworks for RL and cooperative GT.

DOI

10.1016/j.mlwa.2022.100341

Supplementary materials

GitHub - fhswf/MLPro: MLPro - The Integrative Middleware Framework for Standardized Machine Learning in Python


Learn more


Top



All publications

Authors

DetlefArend, Laxmikant ShrikantBaheti, SteveYuwono, Syamraj Purushamparambil SatheeshKumar, AndreasSchwung

Journal

Machine Learning with Applications, Vol. 21, September 2025, Article 100715

Abstract

In this paper, we present version 2.0 of the open-source middleware MLPro for applied machine learning in Python. Notably, it introduces the new sub-framework MLPro-OA for online machine learning, focusing on standards and templates for classic and online-adaptive data stream processing (DSP/OADSP). As part of this, we provide three novel adaptation mechanisms:The first, event-oriented adaptation, enables localized, event-driven parameter updates within individual tasks. The second, cascaded adaptation, allows adaptation events to propagate across multiple dependent tasks, creating task-spanning adjustment cascades decoupled from the forward-facing DSP. The third, reverse adaptation, allows tasks to revise prior adjustments by explicitly processing obsolete instances discarded from a preceding sliding window. Furthermore, we provide insights into the underlying design criteria of MLPro-OA, which were developed through extensive requirements engineering. In the practical part of this work, we demonstrate the essential functionalities of MLPro-OA using reproducible examples.

DOI

10.1016/j.mlwa.2025.100715

Supplementary materials
Authors

Detlef Arend, Amerik Toni Singh Padda, Andreas Schwung, Dorothea Schwung

Conference

IEEE, 2025 11th International Conference on Control, Decision and Information Technologies (CoDIT)

Abstract

This paper presents the novel RLPID architecture for online-adaptive control, which combines classical proportional-integral-derivative (PID) control with reinforcement learning (RL). This hybrid approach enables dynamic online adjustment of PID parameters during control operation. Specifically, we propose a multi-objective reward structure that integrates established control criteria and analyze suitable configurations for different system dynamics. The RLPID controller has been implemented within the open-source middleware MLPro, where it is embedded in newly developed sub-frameworks for classical and online-adaptive control. Owing to its hybrid nature, the architecture can be used both in traditional control loops and within the Markov decision process of RL. Its effectiveness and practical applicability are demonstrated in a cascade control scenario.

DOI

10.1109/CoDIT66093.2025.11321229

Supplementary materials

GitHub - fhswf/paper-da-ieee-codit-2025: Paper IEEE CoDIT 2025 - Arend/Singh Padda/Schwung - "Online-adaptive PID control using Reinforcement Learning"

Authors

Detlef Arend, Andreas Schwung

Conference

IEEE, 2025 IEEE 8th International Conference on Industrial Cyber-Physical Systems (ICPS)

Abstract

There are essentially two global trends that are currently heating up the AI sector. An increasing number of large language models (LLM) creates enormous resource requirements due to extensive data storage and complex training runs. Furthermore, advancing digitalization leads to an increased volume of data and the associated need for real-time analysis. In both areas, it is becoming apparent that the established AI models are reaching their limits despite all optimizations. This is where online machine learning comes in, with new models optimized to process incoming data in real-time and continuously adapt to it. This brings special requirements for the algorithms and their orchestration in extensive stream workflows. In this article, we focus on the latter aspect and highlight the two concepts of cascaded and reverse adaptation, which enable significant functional improvements individually and in combination. A corresponding extension of our open-source framework MLPro also enables us to demonstrate the adaptation mechanisms mentioned practically.

DOI

10.1109/ICPS65515.2025.11087907

Supplementary materials

GitHub - fhswf/paper-da-ieee-icps-2025: Paper IEEE ICPS 2025 - Arend/Schwung - "Online-adaptive data stream processing via cascaded and reverse adaptation"

Authors

Steve Yuwono, Detlef Arend, Andreas Schwung

Journal

SoftwareX, Vol. 28, December 2024, Article 101956

Abstract

Game theory, a fundamental aspect of mathematical economics and strategic decision-making, has been increasingly applied to various fields, including economics, biology, computer science, and engineering. Despite its growing importance, there is a significant lack of flexible and user-friendly tools for standardized modeling of them, particularly for real-world applications. Hence, we developed MLPro-GT as part of our open-source MLPro project, which offers modular and standardized yet flexible components, extensive documentation, and a variety of examples. MLPro-GT allows researchers and practitioners to easily incorporate game theory into their applications while lowering the entry barrier for students. This makes individual work more reproducible, shareable, and reusable.

DOI

10.1016/j.softx.2024.101956

Authors

Steve Yuwono, Marlon Löppenberg, Detlef Arend, Mochammad Rizky Diprasetya, Andreas Schwung

Journal

Software Impacts, Vol. 16, Mai 2023, Article 100509

Abstract

Although the integration of machine learning into production systems has demonstrated significant potential, its real-life applications remain challenging. It is often necessary to rely on digital representations of the actual systems to design and test machine learning algorithms for controlling the systems because conducting such processes directly in real systems is expensive and risky. Hence, in this paper, we present a standardized, multi-purpose and flexible production systems simulator in Python for scientific and industrial activities, namely MLPro-MPPS, that is integrated with the MLPro package. Consequently, this allows the simulations by MLPro-MPPS to be compatible with machine learning tasks.

DOI

10.1016/j.simpa.2023.100509

Supplementary materials

GitHub - fhswf/MLPro-MPPS: An extension feature of MLPro related to a simulation framework for multi-purpose production systems (MPPS)

Authors

Steve Yuwono, Marlon Löppenberg, Detlef Arend, Mochammad Rizky Diprasetya, Andreas Schwung

Conference

2023 IEEE 2nd Industrial Electronics Society Annual On-Line Conference (ONCON)

Abstract

In the current era of automation, there has been a growing interest in integrating artificial intelligence into manufacturing systems. A digital representation of the systems is often important to develop, test, and implement learning algorithms. However, developing a digital representation that seamlessly integrates with artificial intelligence remains challenging. The majority of existing simulators lack the capability of such integration or are not available as open-source platforms. To overcome these challenges, we propose MLPro-MPPS, a versatile and configurable Python-based production systems simulations framework integrated with the established machine learning framework, namely MLPro. MLPro offers advanced templates for machine learning in both scientific and industrial contexts, where both MLPro and MLPro-MPPS are open-source frameworks. In this paper, we present the design and fundamentals of MLPro-MPPS, which accurately models the behaviour of real hardware. We demonstrate the potential of MLPro-MPPS in facilitating standardized machine learning within production system simulations and conduct two sample applications of production systems within multiple machine learning methods. The results show that MLPro-MPPS offers a clean and reproducible simulation framework.

DOI

10.1109/ONCON60463.2023.10431280

Authors

Detlef Arend,  Mochammad Rizky Diprasetya, Steve Yuwono, Andreas Schwung

Journal

Software Impacts, Vol. 14, December 2022, Article 100421

Abstract

In recent years, many powerful software packages have been released on various aspects of machine learning (ML). However, there is still a lack of holistic development environments for the standardized creation of ML applications. The current practice is that researchers, developers, engineers and students have to piece together functionalities from several packages in their own applications. This prompted us to develop the integrative middleware framework MLPro that embeds flexible and recombinable ML models into standardized processes for training and real operations. In addition, it integrates numerous common open source frameworks and thus standardizes their use. A meticulously designed architecture combined with a powerful foundation of overarching basic functionalities ensures maximum recombinability and extensibility. In the first version of MLPro, we provide sub-frameworks for reinforcement learning (RL) and game theory (GT).

DOI

10.1016/j.simpa.2022.100421

Authors

Detlef Arend,  Steve Yuwono, Mochammad Rizky Diprasetya, Andreas Schwung

Journal

Machine Learning with Applications, Vol. 9, September 2022, Article 100341

Abstract

Nowadays there are numerous powerful software packages available for most areas of machine learning (ML). These can be roughly divided into frameworks that solve detailed aspects of ML and those that pursue holistic approaches for one or two learning paradigms. For the implementation of own ML applications, several packages often have to be involved and integrated through individual coding. The latter aspect in particular makes it difficult for newcomers to get started. It also makes a comparison with other works difficult, if not impossible. Especially in the area of reinforcement learning (RL), there is a lack of frameworks that fully implement the current concepts up to multi-agents (MARL) and model-based agents (MBRL). For the related field of game theory (GT), there are hardly any packages available that aim to solve real-world applications. Here we would like to make a contribution and propose the new framework MLPro, which is designed for the holistic realization of hybrid ML applications across all learning paradigms. This is made possible by an additional base layer in which the fundamentals of ML (interaction, adaptation, training, hyperparameter optimization) are defined on an abstract level. In contrast, concrete learning paradigms are implemented in higher sub-frameworks that build on the conventions of this additional base layer. This ensures a high degree of standardization and functional recombinability. Proven concepts and algorithms of existing frameworks can still be used. The first version of MLPro includes sub-frameworks for RL and cooperative GT.

DOI

10.1016/j.mlwa.2022.100341

Supplementary materials

GitHub - fhswf/MLPro: MLPro - The Integrative Middleware Framework for Standardized Machine Learning in Python

 

Top