Deep Reinforcement Learning Driven Control and Optimization for Cyber Physical System in Smart Manufacturing

The development and use of industry solutions intended to increase the autonomy and flexibility of production systems. The objective is to enable systems to autonomously reconfigure, resulting in enhanced adaptability and efficiency in manufacturing procedures. Achieving such goals requires Cyber-Physical Industrial Systems (CPPSs), which integrate digital technologies with industrial processes. A digital twin (DT) serves as the core component of the ‘cyber’ layer in their design, which is built on both physical and cybernetic features. Nevertheless, to respond swiftly to changes in the manufacturing systems environment, the cyber layer’s answers need sophisticated analytical modules to be added to their virtual equivalent. In this study, the researcher provides a strategy for developing CPPS-specific digital twin production systems using a combination of discrete simulation and deep reinforcement learning (DRL) methods. The digital twin serves as a medium of communication between the reinforcement learning agents and their environment to devise a plan for distributing production resources among certain activities. The methods used for this study are Proximal Policy Optimization and Asynchronous Advantage Actor-Critic.