Smart Forge - Sustainable production through AI controlled forging oven
| Reference number | |
| Coordinator | RISE Research Institutes of Sweden AB - RISE Research Institutes of Sweden |
| Funding from Vinnova | SEK 5 000 000 |
| Project duration | October 2020 - March 2023 |
| Status | Completed |
| Venture | Strategic innovation programme for process industrial IT and automation – PiiA |
| Call | PiiA: Digitalization of industrial value chains, spring 2020 |
Important results from the project
The main goal with Smart Forge was to create a prototype system for automatic control of Bharat Forge Kilstas´s forging line, with the aim of reducing scrap. The approach was to first create a physics-based simulator of the forging line, and then use this for the development and testing of control algorithms. Two algorithms were developed: a linear optimizer, and a reinforcement learning algorithm. The latter has been tested in real production and proved to be able to control the forging line during normal production. The next step is to expand the tests to include more production modes.
Expected long term effects
The project has demonstrated proof-of-concept -- that an AI trained entirely in a simulator can be used to actively control the power input to an induction furnace, to control the temperature of a metal object moving through the furnace. The project has also led to an open source communication platform for remote connection to OPC-UA servers. More work is needed to create a product ready for full deployment, and continued collaboration between RISE, Viking Analytics and Bharat Forge Kilsta is therefore being planned.
Approach and implementation
The project´s results show that the approach is sound, but several challenges were encountered during its course. First and foremost it was about how the communication between the algorithms and the production line should be done, but also how realistic the physics of the simulator has to be and how the training of reinforcement learning AI should be done. Reinforcement learning requires a well-designed reward function, and it turned out to be a challenge to design one for warmholding control. We therefore decided to try to tackle normal production and warmholding recovery first.