Prädiktiv-reaktives Scheduling zur Steigerung der Robustheit in der Matrix-Produktion

Due to the increasing individualization of products, manufacturing companies are offering more and more variants with decreased quantities per variant. In addition, customer demand is becoming more volatile and difficult to predict. The main challenge is to eco-nomically produce a fluctuating mix of variants with fluctuating total quantities. Matrix-Production systems aiming for a production in batch size 1 decoupled from a takt are therefore a current object of research. In addition to the design of these systems, an increasingly important role is filled by production planning and control, since the material flows in such production systems are highly complex.

The state of research is characterized by a multitude of predictive-reactive methods for scheduling even in complex production systems. However, there is no approach that specifically considers robustness in predictive planning in order to enable reactive rescheduling to maintain the desired logistical performance despite unforeseen disruptions.
Therefore, a method for predictive-reactive product control of matrix-structured produc-tion systems was developed in this thesis, which allows the determination of an optimal degree of robustness in predictive robust scheduling and thus enables an optimal mix of prevention and reaction in production control. ... mehrThe method consists of three parts: First, in predictive robust scheduling, a schedule is generated on the basis of the pro-duction program, in which a desired extent of slip times between processing steps is then inserted. The robust schedules are then carried out in a discrete-event simulation. In the event of longer disturbances, a rescheduling corridor is determined secondly, which indicates which processing steps of which orders must be rescheduled depending on the duration of the disturbance and the underlying schedule. The rescheduling corridors are then rescheduled thirdly in reactive rescheduling and the results are transferred to the discrete event simulation for reintegration. Reactive rescheduling uses reinforcement learning based on a decentralized Markov process to learn optimal selection strategies for orders depending on the station. The method was tested in an application for a concept of a flexible body-in-white production with a partner from the automotive industry.
The developed method contributes to the understanding of the concept of robustness as well as to the application possibilities and limits of reinforcement learning in production control. To the author’s knowledge, the work is the first approach to integrate robustness considerations directly into predictive-reactive scheduling approaches in order to improve the logistical performance.