Multi-Task Learning for Predictive Modeling and Optimization of Manufacturing Processes: A Stretch-Reducing Mill Application

Stretch-reducing mills are complex multi-stand systems in which the final tube thickness depends on the interaction between inter-stand tensions, angular speeds, material deformation, and process conditions. In this study, a multi-task neural network is proposed to model and optimize the operation of a six-stand stretch-reducing mill. A structured pipeline was used to generate a feasible dataset to train and test the model. The proposed architecture employs the inter-stand tensions in input, utilizes a shared network to get the general relationships between variables, and predicts process status, tube thickness, angular speed, and the neutral radius of each stand with dedicated task-specific heads. The model achieved high performance across all the tasks on training, validation, and testing sets. The final tube thickness predictions get a very low error, showing the ability of the proposed network to gather hidden relationships throughout the process. The trained model was integrated with the differential evolution optimization algorithm to obtain the feasible inter-stand distribution and stand angular speeds to achieve the desired final tube thickness while respecting the feasibility constraint set. The results demonstrate that the proposed framework can provide support for the decision-making process optimization, representing an effective tool for intelligent modeling in stretch-reducing mill monitoring and control.