Learning-based Edge Computing Architecture for Regional Scheduling in Manufacturing System

This paper proposes a novel edge-computing based structure to support learning-based decision-making in industry manufacturing field. This structure consists of four functional layers, respectively realizing model establishment, task allocation and task processing work. In order to take full advantage of the distributed computing resources at the edge, the manufacturing computing task can be further decomposed into several sub-tasks, separating the complex computing problem with large problem size into regional scheduling ones with much smaller problem size. All the sub-tasks are allocated to the edges, accomplished by the algorithm deployed on computing devices of region-related edge node, which contributes to faster data-processing and problem-solving speed. A simulation test has been performed in which a multi-AGV scheduling problem was solved according to a distributed reinforcement learning method configured in such edge computing architecture. The objective of each edge node is to acquire AGV schedule of related region that minimizes the makespan. Simulation results demonstrate that this distributed edge computing system can be enabled to learn satisfying solution and converge much faster when it is compared with conventional method applied in centralized architecture.