Simulation of industrial material flows for virtual commissioning with graph neural networks

考虑到虚拟调试（VC）的要求，没有一种整体的物流模拟方法可以确保大量件货的物流动态的时间确定性计算。随着运动精度的提高，大量单件货物的计算时间也会增加。学习模拟器在预测复杂现象方面比经典模拟器的计算效率高得多。与求解微分方程的模拟器经典方法不同，学习模拟器的方法基于可参数化函数，该函数使用数据和机器学习训练为模型行为。 在自己的初步工作中，它表明，学习模拟器也可以适用于工业物质流的背景下，VC和物质流行为的映射可以表示与图形神经网络（GN）模拟器。这可以大大减少计算时间。对于一般的适用性和普遍性，一个合适的方法自动生成合适的数据，以及一个GN模拟器集成到一个实时仿真环境中，必须研究。本项目的目的是在VC环境下，在GN中进行基于实时计算的高分辨率物质流模拟。这应该创造了在VC仿真中实时仿真高分辨率行为模型的可能性。预先，GN将使用非真实的物理材料流模拟和来自真实的系统的数据在真实的模型行为上进行训练。

Considering the requirements of virtual commissioning (VC), there is no holistic approach to material flow simulation that ensures time-deterministic calculation of the dynamics of material flows for a large number of piece goods. As the accuracy of the movement increases, the computation time for a large number of piece goods increases. Learning simulators can be much more computationally efficient than classical simulators in predicting complex phenomena. In contrast to classical approaches for simulators, which solve differential equations, the approach of learning simulators is based on a parameterizable function that is trained to a model behavior using data and machine learning. In own preliminary work it is shown that a learning simulator can also be suitable for industrial material flow in the context of VC and that the mapping of material flow behavior can be represented with a Graph Neural Network (GN) simulator. This can significantly reduce the computation time. For a general statement of the applicability and generalizability, a suitable methodology for the automatic generation of suitable data as well as an integration of a GN simulator into a real-time simulation environment must be researched. The aim of this project is to make a high-resolution material flow simulation based on a real-time calculation within a GN usable in a VC. This should create the possibility to simulate high-resolution behavior models within a VC simulation in real-time. In advance, the GN will be trained on the real model behavior using non-real-time capable physical material flow simulations and data from a real system.