Reinforcement Learning-based Operational Decision-Making in the Process Industry Using Multi-View Data

Owing to the frequent fluctuations encountered in raw material characteristics and operational conditions in the process industry, traditional data-driven approaches prove inadequate in adapting the adjustment of operational variables. Furthermore, the potential of multi-view data, including images, audio, and sensor data, remains underexploited in industrial processes. This study proposes an operational decision-making method based on feedstock-guided multi-view actor-critic (FMAC-ODM) using multi-view data to address these issues. This method utilizes the idea of reinforcement learning (RL) for enhanced decision-making. First, the problem of optimizing operational variables is reformulated into a continuous RL problem to acquire an improved decision-making policy aligned with the current operational conditions. Subsequently, the inclusion of feedstock properties in the state space is implemented to provide essential guidance for the decision-making process. Finally, in pursuit of a comprehensive understanding and bolstering the precision of the decision-making strategy, multi-view data sourced from the industrial site is harnessed as a surrogate for human observation. The effectiveness of the proposed decision-making method is substantiated through its practical application in the industrial flotation process.