Dynamic decision-making for knowledge-enabled distributed resource configuration in cloud manufacturing considering stochastic order arrival

Abstract

The emergence of COVID-19 caused the stagnation of production activities and promoted the changing market demand. These uncertainties not only brought great challenges to the manufacturing approaches led by a single enterprise, but also threatened the stability of inherent supply chain. To maintain market competitiveness, an efficient distributed manufacturing resource allocation method is urgently needed by manufacturers. Cloud manufacturing (CMfg) is an advanced service-oriented manufacturing paradigm that breaks physical space constraints to integrate distributed resources across enterprises, and provides on-demand configuration of manufacturing services for personalized consumer needs in real-time. The focus of this paper is to achieve dynamic configuration of distributed resources in CMfg considering stochastic order arrival, while reducing overall completion time and improving resource utilization. First, a dynamic knowledge graph for distributed resources is constructed, and its definition and construction methods are introduced. Secondly, semantic matching between massive optional manufacturing resources and multiple types of subtasks is achieved through knowledge extraction, thereby obtaining a candidate set of manufacturing resources that meet basic requirements for each subtask. An artificial intelligence (AI) scheduler based on deep reinforcement learning is developed, and order urgency is incorporated into the design of state observation vectors. AI scheduler can generate optimal decision results based on environmental observations, select high-quality manufacturing service compositions over candidate sets, and ultimately achieve efficient distributed resources configuration. Finally, Dueling DQN-based training method is put forward to optimize AI scheduler, enabling adaptable decision-making performance in dynamic environment. In the experiment, a simulation environment with 18 different settings is designed that considers stochastic processing time, random order compositions and various order arrival patterns. The proposed graph-based matching method, scheduling policy learning method and dynamic decision-making method are tested in the simulation environment. The experiment results demonstrate that the cognitive and AI joint driven distributed manufacturing resource configuration method is superior to traditional methods in terms of policy learning speed and scheduling solution quality. © 2023 Elsevier Ltd