Strip Exit Thickness Prediction Model Based On Improved Activation Function RNN

The exit thickness accuracy is one of the most important quality indicators for measuring the continuous strip rolling products.Therefore,the issue of improving the thickness accuracy of the strip exit is a hot topic for many scholars,At present,the neural network method has been applied to this field.A deep learning method is used to improve the prediction accuracy of the strip exit thickness.1.Firstly,the differences between the BP neural network,DNN model and RNN model in terms of network structure and training methods were introduced.Secondly,the rolling data was normalized,all variables were mapped to the interval[0,1],finally three models were implemented by using Python.Experimental results showed that,the MSE of using BP neural network,DNN model and RNN model for predicting the exit thickness are 0.014,0.011,and 0.012,respectively.2.Firstly,the gradient acceleration activation function GAAF was introduced,there are an infinite number of discontinuities in the function.There is a non-conductible situation in the backpropagation process,the shape function was only empirically given according to the experimental results.Aiming at the above problems of GAAF,two improvements were made,The smooth GAAF function was proposed and the general definition of the shape function was given.Finally,the improved algorithm was tested using the commodity image classification public dataset Fashion-MNIST.Experimental results showed that the improved algorithm could improve the prediction accuracy of the data set by 0.6%.3.Firstly,in order to make the original rolling data applicable to the recurrent neural network,the strip thickness data was serialized.Secondly,according to the production process of the cold strip,the improved GAAF algorithm was introduced to establish a RNN model.Finally,the experimental results showed that the RNN using the improved activation function could make the MSE of the exit thickness reach 0.009,and the prediction accuracy was improved by 0.003 compared with the RNN before the improvement.