Self-learning Processes in Smart Factories: Deep Reinforcement Learning for Process Control of Robot Brine Injection

Abstract The goal of this paper is to investigate the application of adaptive learning algorithms, which enables industrial robots to cope with natural variations exhibited in a brine injection process related to the production of bacon. Due to the variations in bacon meat, the traditional needle-based brine injection process is not capable of injecting the correct amount of brine, leading to either ruined or unflavored bacon. In the presented work a Deep Deterministic Policy Gradient (DDPG) reinforcement learning algorithm is introduced in the injection process to improve process control. To accelerate training of the reinforcement learning algorithm, a simulation environment of the brine absorption is generated based on 64 conducted experiments. The simulation environment estimates the amount of absorbed brine given injection pressure and injection time. Tests are run in the simulation where the starting mass is generated from a normal distribution with mean 80.5g, and a standard deviation of 4.8 g and 20.0 g respectively. With a target of 15 % mass increase, the agent can produce an average mass increase of 14.9 % for the first test and 14.6 % for the second test. This indicates that the model can successfully adapt to a high variety input, thereby showing potential for process control in brine injection, coping with natural variation in meat structure.