Deep Reinforcement Learning for Optimising a Forging Line
Abstract
The project´s objective is to create a prototype of an AI-based process control system that with high precision can autonomously control Bharat Forge Kilsta´s forging line in such a way that the steel rod temperature, just before forging, falls within the temperature range that optimizes the properties of the forged component. The direct purpose is to reduce scrap, which today is very costly and thus a major challenge for the company.
In order to develop the control algorithm, we are aiming for Deep Reinforcement Learning to find the optimal control policy using a digital twin of the heating process for the forge. This requires large amount of computational resources (GPUs) for model training, hyperparameter tuning and model quality evaluation. We have currently implemented two different control algorithms and we have seen that our approach would need larger network sizes for being able to control it properly. Unfortuantely, this requires more resources than we have available at Karlstads Universitet where we do not have the required GPUs and cannot acquire them in time for the project to succeed. Therefore, we apply for resources here to parallelize the model training and verification.
