Neural Network Models for Industrial Batch Annealing Operation

Abstract

Modern industrial batch annealing operations are highly automated, where the process cycles of individual stacks are designed with a thermal model. The inability of thermal model in capturing annealing kinetics and the associated process inefficiency was recently illustrated. Subsequently, this limitation was overcome through an integrated batch annealing model, with prediction capability extended to annealing kinetics, microstructure, and mechanical properties. However, the integrated model requires significant processing time in deterministically designing the industrial process cycle for a battery of furnaces. Neural network models are an attractive alternative to the deterministic models due to the accurate online data measurement of modern operations and the high processing speed of neural network in prediction mode. In the present work, based on the industrial data, a neural network model has been developed for correlating the coil dimensions and annealing parameters to the mechanical properties. The noise associated with the industrial data has been characterized and the associated complexity in developing the neural network model has been studied. Of the various neural network algorithms examined in the present work, the stochastic noise addition was observed to be an effective method in enhancing the neural network model performance.

Keywords:

• Back propagation
• Batch annealing
• Cold rolling
• Data classification
• Industrial data
• Mechanical properties
• Model
• Model validation
• Multilayer perceptron
• Neural network
• Noise characterization
• Noisy data
• Stochastic noise
• Steel
• Weight decay

ACKNOWLEDGMENTS

The authors would like to thank Professor Mathai Joseph, Executive Director, Tata Research Development and Design Centre (TRDDC) for approving and supporting this work. The authors are also thankful to the managements of Tata Steel and Tata Research Development and Design Centre for approving the publication of this work.

Notes

*Amplitude of noise vector in terms of percentage of the range of the output variables.

^#Percentage of test data sets within 10% error range.