Optimal-adaptive online control of non-stationary manufacturing processes

Motivation and Goals:

Aim of the project is the research, development and evaluation of adaptive-optimal control methods for manufacturing processes in metal processing. By means of the methods Initial, offline trained, control policies have to be adopted to the physical process during processing time to:

Adopt process behavior which is not part of the simulated environment
Adopt non-stationary process behavior

Research

Results

Development and evaluation of methods for optimal adaptive control of manufacturing processes based on reinforcement learning
Development and evaluation of adaption strategies for minimizing process waste under explorative adoption, inter alia by detecting systematic non-stationary process behavior.

Implemented versions of the control agent: Approximate online reinforcement learning for manufacturing process control based on recurrent artificial neural networks for partial observable process environments and methods for save and efficient process exploration.
Evaluation results for the control agents based on one or more exemplary applications in metal processing.

Application example: Deep Drawing