Health Check for Neural Networks

Steady growth in the variety of steel grades has created the need for self-learning statistical models – like neural networks – in order to achieve high setup accuracy and tight product tolerances, as well as to apply to new materials. The performance of neural networks depends in turn on the quality and quantity of training data available. Erroneous data lead to poor modeling behavior, weakening pre-calculation accuracy – and the performance of the mill declines. Improvements to the pre-calculations of thickness, width, strip temperature, profile, and flatness using neural networks results in higher quality and more exact strip dimensions.