Flatness Prediction of Cold Rolled Strip Based on Deep Neural Network with Improved Activation Function

Area of Science:

• Materials Science
• Computer Science
• Industrial Engineering

Background:

• Cold rolled strip flatness is a critical quality indicator in industrial manufacturing.
• Existing methods for strip shape prediction face limitations in accuracy and efficiency.

Purpose of the Study:

• To develop an improved deep neural network (DNN) for accurate strip shape prediction.
• To enhance the prediction accuracy of cold rolled strip flatness in steel production.

Main Methods:

• Modeling strip production data using an improved deep neural network (DNN).
• Analysis of activation function properties and non-convex optimization in deep networks.
• Proposal and evaluation of an improved Swish activation function with batch normalization.

Main Results:

• The improved DNN achieved higher prediction accuracy for cold rolled strip flatness compared to standard DNNs.
• The mean square error for flatness prediction was reduced by 35% using the improved DNN.
• The improved Swish activation function demonstrated lower loss than other activation functions on the MNIST dataset.

Conclusions:

• The improved DNN model meets and exceeds industrial requirements for strip flatness prediction.
• This approach offers significant potential for reducing scrap yield and production costs in the steel industry.
• The developed model provides valuable guidance for industrial production processes.