Taming complexity: Generative doppelgangers for stochastic data trends in complex industrial manufacturing systems

Abstract

The defining characteristics of complex industrial systems are interconnected processes that generate immense amounts of stochastic data, often hindering operational optimization, especially metrics such as Overall Equipment Effectiveness (OEE). To address the limitations of traditional methods and earlier machine learning techniques in capturing this complexity, this paper proposes a novel approach using generative doppelgangers, a Generative Adversarial Network (GAN)-based model, to simulate the operational behavior of these systems. This "behavioral doppelganger" learns intricate relationships within historical operational data from a production facility, enabling proactive what-if analyses for OEE optimization. The proposed framework's ability to replicate the impact of process parameters on availability, quality, and performance, which collectively contribute to OEE, is highlighted. The research validates this approach using real data from an industrial sugar plant, demonstrating its potential to provide valuable insights into system behavior under different operational scenarios for proactive optimization.