Skip to main content

A Platform for Real-time Process Control using Deep Learning and Reinforcement Learning

A digital platform to be implemented particularly at small and medium sized companies where transformation to digitalization and use of the state-of-the-art machine learning benefits are difficult to realize due to lack of skilled personal.

Background

The research is about monitoring predicting and identifying sources of variations occurring and propagating in a chain of manufacturing processes. The research focuses mainly on dimensional and geometric variations. Developing and implementation of deep learning and reinforcement learning models constitute the application aspect of the research.

The state -of-art in the transformation towards digitalization involves developing a model of industrial common practice of the digitalization step is to create digital twins as offline virtual testbeds and require re-computation to respond to changes. The resources to build such platforms are rather established. However, for real time or online application where rapid responses are desired, Deep Learning (DL) and Reinforcement Learning (RL) based systems can provide actions that can respond to changes in real-time, thereby increasing efficiency in resource utilization and productivity.

Aims and objectives

This project aims to create a digital platform (with both frontend and backend) to be implemented particularly at SME’s where transformation to digitalization and use of the state-of-the-art machine learning benefits are difficult to realize due to lack of skilled personal to develop such systems. Using the generated files, RL agents can be trained and saved with no or limited programming skills person can select process objectives/goals and constraints, from which all the necessary RL files are where a semi-skilled. The ability to generate files, train and save models will enable manufacturing organizations for a faster transformation to digital, sustainable, and intelligent organizations.

Project plan

Following the results obtained in earlier efforts the details of the architecture and the interfaces will be done along with the case specification. The output of the case specification will be the basis in developing the Model and RL Experimental Modul which includes process model development and the Reinforcement Learning from a process behavior data. After this how the platform can ease/enable rapid industrial transformation to SME will be demonstrated.

Applied interdisciplinarity

Digitalization and application of machine learning in the manufacturing industry essentially require the close collaboration among different domains due to the complexity of the involved processes. In this project clear collaborations with material science and machine element design is envisaged that will use the results of this work to advance and apply it in the respective domains. As the core elements of the work are data driven, the application of the results from this effort can be further applied in systems where there are processes in need of real time control.

Papers

  • Yacob, F.; Semere D. / A multilayer shallow learning approach to variation prediction and variation source identification in multistage machining processes. Journal of Intelligent Manufacturing 32(2), April 2021.
  • Yacob, F., Semere, D. & Nordgren, E. Anomaly detection in Skin Model Shapes using machine learning classifiers. Int J Adv Manuf Technol 105, 3677–3689 (2019).
  • Hoskins, J. C., & Himmelblau, D. M. Process control via artificial neural networks and reinforcement learning. Computers & chemical engineering, 16(4), 241-251

KTH Collaborations

Duration

September 2021 – December 2023

Project participants