KI-STAR

Improving the raw material and energy efficiency of processing plants in the primary raw materials industry Through AI-based (semi-)autonomous control systems[

Motivation

Rock crushing accounts for approximately 3% of global electricity consumption. This makes it one of the most energy-intensive process steps in the extraction and processing of mineral raw materials. At the same time, during the processing of aggregates, approximately 10% of the raw material is discarded as unusable waste due to both natural as well as process-related quality variations. Although the machines in a processing plant partially record data continuously, this data has so far been evaluated or utilized only on a machine-specific basis, rather than in the context of the overall process. The reason for this is that the control systems used are typically based on static operating conditions of the plant and can therefore only respond to a limited extent to the dynamically varying properties of the feed material. Experienced plant operators therefore frequently intervene manually in the process control to achieve the required product quality more efficiently. Given the difficulty of transferring this expertise and the growing shortage of skilled workers, this approach is reaching its limits and calls for digital solutions.

Approach

In the KI-STAR project, the Institute of Advanced Mineral Processing Technologies (AMR) – as the coordinator – along with the Fraunhofer Institute for Production Technology (IPT), Willy Dohmen GmbH & Co. KG, eurogard GmbH, and Optocycle GmbH, in collaboration with the MRE, are developing an intelligent control system for processing plants. For this purpose, the processing path of an aggregate plant is systematically examining as a reference process. In this process, all machines and their control variables, as well as relevant process parameters, quality-influencing disturbances, and application scenarios for (semi-)autonomous control, are derived. This results in five closely interlinked components. First, the MRE develops a modular, dynamic simulation model of the reference plant to digitally describe the behavior and interaction of the machines without interfering with the actual process. A comprehensive sensor system is also being developed for the reference plant to measure product quality parameters, such as particle size and shape. An interoperable data platform consolidates these measured values with machine data and external data – such as power consumption, weather conditions, or demand. Based on the simulation model and this data, an AI agent uses reinforcement learning to learn the effects of individual control settings and derives control recommendations to optimize the energy consumption, throughput, and product quality of the entire plant. In the final step, the practical feasibility of the AI-based control system will be validated using the reference plant and its applicability to larger and more complex processing plants will be evaluated.

Project Objectives and Research Questions

Investigation into how a dynamic simulation model of a processing plant, based on real material flows, can be established as a training environment for an AI agent.
Development of an AI agent that controls processing plants (partially) autonomously depending on the optimization goal (overall plant effectiveness, energy consumption, or product quality).
Validation of the overall system – comprising the data platform, sensor system, simulation, and AI agent – at an aggregate plant (TRL 6) and assessment of its transferability to more complex processing plants.

Work Packages