Interpretable Multi-Step Production Optimization Utilizing IoT Sensor Data

Abstract

In an industrial manufacturing process, such as petroleum, chemical, and food processing, with the deployment of thousands of sensors in the plants, we have the chance to provide real-time onsite management for the processes. Beyond the real-time status update, utilizing vast IoT data and creating machine learning and optimization models provide us with intelligent business recommendations. Those are used by the site engineers and managers to make real-time decisions in a situation with multiple conflicting operational and business goals. Those goals include maximizing financial gain, minimizing costs, limiting the usage of certain raw materials or additives, decreasing environmental impact, and more. When formalizing these decision-making tasks, often there is no prior knowledge of compromise between the conflicting goals. That poses a challenge to generate a proper objective function. In this paper, we create a Multi-Step optimization process to address this uncertainty of selecting proper objectives and their preferences. Instead of using an explicit trade-off to create a single weighted objective function (as a traditional approach) and rely on a single attempt to find the optimal solution, we decompose this problem into multiple steps. In each step, we optimize only one objective from one KPI with an exact semantic meaning. We demonstrate the usability of the approach using a practical application from an oil sands processing facility, provide modeling results focusing on the response to business priorities, performance, and interpretability. The multi-step approach presents the convergence of the target goal with an outcome KPI with comparison for each step to illustrate the enhanced interpretability.