Study Of Strip Flatness Model Based On Improved Activation Function DNN

As the industry's requirements for the quality of cold-rolled strips continue to increase,the quality of the flatness has become one of the most important indicators for measuring the quality of cold-rolled strip.In this paper,the neural network model is applied to the strip production data of a 1250mm cold-rolled strip mill in order to improve the accuracy of the flatness prediction.And we make improvements to the deep neural network,the prediction accuracy of the improved model is further improved.The specific work is as follows:1.Traditional neural network modeling.Firstly,preprocess the strip production data,delete null values,irrelevant columns and noise data,select input variables and output variables.The BP neural network and the deep neural network are implemented in Python,and then use them to model the processed data respectively.After training,the mean square error of the BP network model is 8.329,and the mean square error of the deep neural network model is 3.731.The prediction accuracy of the deep neural network model is higher than that of the BP neural network model.2.Make improvements to the deep learning algorithm.Firstly,the activation function commonly used in neural networks is analyzed.We propose that in the deep network,the monotonicity of the activation function is independent of the convexity of the loss function,and we prove that the loss of the deep network is not convex whether the activation function is monotonous or not.And then the non-convex optimization problem of the loss function is discussed.We extend the discussion results of the non-convex optimization of the deep linear network to the deep nonlinear network,prove that under the condition that the hypothesis and the proposed theorem are satisfied,the critical point of the network is the global minimum point.Finally,the Swish activation function is improved by the batch normalization method.The improved function is tested on the data set Mnist containing 70,000 handwritten digits,the loss is 0.1088 on the test set,which is lower than other activation functions,and the improved Swish function is better than the Swish function.3.Improved deep learning algorithm modeling.The preprocessed strip production data is modeled by a deep neural network with improved activation function.After training,the model's mean square error is 1.305,The prediction accuracy of the model is further improved compared with the traditional deep neural network model.The improved model is used as the final model with flatness prediction,and the model's accuracy meets industrial requirements.