Adaptive data-driven predictive control using behavioral approach for autonomous powder compaction

  • Contact:
    Professor Dr.-Ing. Naim Bajcinca

    Technische Universität Kaiserslautern

    Fachbereich Maschinenbau und Verfahrenstechnik

    Lehrstuhl für Mechatronik in Maschinenbau und Fahrzeugtechnik



    Professor Dr. Markus Thommes

    Technische Universität Dortmund

    Fakultät Bio- und Chemieingenieurwesen

    Lehrstuhl für Feststoffverfahrenstechnik



Powder compaction is a common dry granulation method to transfer powder material in compacts such as tablets. Industrial processes are mainly performed on rotary tablet presses, where multiple sub-processes are included. Feeding and blending of the powder component are subsequentially followed by the die filling with powder, powder compression and the ejection of the compact from the die. Hereby the temporal scales range from minutes (feeding, blending) to seconds (filling) and milliseconds (compression, ejection), but these processes are coupled with respect to the material flow. This leads to a complex control task, which aims for achieving the desired compact active content (dose) and radial fracture force (hardness) related to the product properties in terms of weight, active weight fraction, porosity and degree of lubrication. Established strategies for these control problems are currently based on human operator intervention, which is in general prone and predisposed for slow action and reaction. Therefore, the aim of this project is to replace human intervention by an autonomous control system for powder compaction. A key element for autonomous powder compaction is a process monitoring system capable of characterizing product quality deviations. To this end, a novel sensor system will be developed, which combines different methods (UV-Vis spectroscopy, NIR spectroscopy and machine data) and sensor types (direct, hybrid and soft). Thereby all sensors will be designed for in-situ determination to obtain a real-time feedback on the process state. The experimental data will be used to develop mathematical models for powder compaction. Different model classes including linear autoregressive models with exogenous inputs (ARX) and nonlinear autoregressive models (NARX) have to be considered due to the described complexity of the process. Crucial steps during model development are the raw data preprocessing, modelling of the process steps individually and connection of these. The obtained data-driven model is used to develop offline and online predictive control policies (DPC) algorithms with respect to the behavior theory. The theoretical basement is the formulation of optimal control problems (OCPs), while the aim is to optimize the process steps in a closed-loop powder compaction. In conclusion, we will develop and implement a control algorithm capable to autonomously adjust the product quality in terms of compact dose and hardness in powder compaction. Hereby autonomy implies an online self-adaption of the parameter set points for the different process steps and phases. During the start-up phase, the waste is minimized, while during manufacturing the production rate is maximized and process disturbances such as feed rate fluctuations are balanced. Thereby the product quality and process efficiency are enhanced in comparison to manual process management.