Research On Prediction Of Steel Rolling Performance Based On Deep Learning

In recent years,China's iron and steel industry has developed rapidly.,and the scale of iron and steel output is in the forefront of the world.Because steel is widely used in various fields of society,such as bridges,tall buildings,ships,machinery,railways and aviation,the quality of steel will directly affect the field of people's livelihood and social public safety.The mechanical properties of steel are mainly reflected in the tensile strength of the steel.In recent years,deep learning has developed rapidly.With its powerful automatic feature extraction capabilities,it can quickly analyze large amounts of data and has strong data fitting capabilities.Therefore,it has been widely used in data feature extraction and prediction,and has achieved good results.In order to improve the prediction accuracy of steel tensile strength,this paper combines deep convolutional network(CNN)and long-short-term memory network(LSTM),uses evolutionary algorithm to optimize network parameters,and applies the model to hot-rolled steel performance prediction.The main content of the paper is as follows:(1)The status quo of research on the prediction of rolling properties is analyzed,and the rolling process and related components are introduced.Aiming at the characteristics of large rolling data,low signal-to-noise ratio,and uneven distribution,the minimum and maximum normalization and outlier detection methods were used to preprocess the sample data.(2)The related theories and basic knowledge of deep learning are introduced,and the convolutional neural network is applied to the feature extraction of hot rolled steel data in this paper.By virtue of its local receptive field mechanism,the model combines low-level features to form a more abstract high-level representation.Discovering the distributed feature representation of the data can not only reduce the difficulty of extracting features from steel rolling data,but also fully mine high-dimensional complex data.Make up for a series of traditional problems such as poor generalization ability and insufficient training.(3)In view of the fact that the rolling data not only has spatial characteristics,but also has certain time-series correlation,it is proposed to combine the long and short memory network with the convolutional neural network,use the convolutional neural network for feature extraction,and then send the extracted features to the LSTM network for sequence prediction.Aiming at the problems that improper network parameter setting may bring insufficient accuracy or too long processing time,the evolutionary algorithm is introduced to optimize the CNN-LSTM model to build a deep learning model with good performance.