Research On Prediction And Model Optimization Of Cement Specific Surface Area Based On Spatio-temporal Convolutional Network

The specific surface area of cement is one of the important indicators of cement quality inspection in the cement production process.Realizing its accurate prediction is of great significance to cement production scheduling,energy saving and consumption reduction,and improvement of cement product quality.However,due to the non-linearity,uncertainty,multiple interferences,dynamic changes,time-varying delays and incomplete information in the cement processing industry,it is difficult for us to establish an accurate soft-sensing model for cement quality inspection.In response to the above problems,this paper proposes to establish a spatio-temporal graph decoupling convolution network(STG-DCNN)based on the theory of multi-scale feature extraction.Aiming at the problem of difficulty in determining the optimal structure and the problem of different feature depths that need to be extracted for features of different dimensions,a STG-DCNN deep self-healing model based on Improved Two-population Differential Evolution Algorithm(ITDE)is further proposed.It realizes the multi-scale feature extraction of temporal and spatial dimensional features in the cement grinding process while realizing the automatic optimization of the model parameters.Finally,the effective prediction of the specific surface area of the cement product is realized.The specific research work is as follows:First of all,combined with the production process of the cement grinding system,the existing typical combined grinding system,double loop cement grinding system,and semi-final cement grinding system are specifically analyzed and compared,and the key variables that affect the specific surface area of the cement product are studied.Based on the detection of the specific surface area of the cement product,the key scientific problems to be solved in this paper are introduced,and the technical route and specific solutions are given.Secondly,based on the theory of multi-scale input and multi-scale feature fusion,a multi-scale spatiotemporal feature extraction network based on convolutional neural network is established for the time-varying time delay,spatial variable coupling characteristics,nonlinearity and uncertainty characteristics of the cement grinding process.The model uses sliding window technology to map the time series data containing time-varying delay information obtained by gray correlation analysis and the spatial sequence data containing variable coupling information obtained by permutation entropy theory to the input layer of the prediction model.Then,a dual-channel convolutional neural network is used to learn and mine the change law between the features contained in the input data and the specific surface area of the cement product.This model can better mine the cement grinding data and realize the effective prediction of the specific surface area of the cement product.Finally,it is difficult to determine the feature depths of different dimensions and the optimal parameters of the deep learning model under complex working conditions are difficult to select.Taking the global reverse training error of the STG-DCNN model as the objective function,an STG-DCNN deep self-healing model based on the improved dual-group differential evolution algorithm is established.This model designs an adaptive mutation scaling factor on the basis of the dual-group differential evolution algorithm to improve the scientific nature of the mutation operator,thereby further obtaining better global search capabilities and increasing the convergence speed.The established ITDE-STG-DCNN cement specific surface area self-healing model gets rid of the intervention of manual experience in the selection of model structural parameters.In the experiment,the actual data in the cement production process was used for verification and analysis,and STG-DCNN was compared and verified with the traditional CNN modeling method,SVM,XGBoost and Random Forest.The ITDE-STG-DCNN and the STG-DCNN target model designed based on the basic differential evolution algorithm,the dual population differential evolution algorithm and the genetic algorithm are compared and analyzed based on the STG-DCNN under the optimal artificial strategy.The results show that the method proposed in this paper obtains greater gains and strong generalization ability than the proposed comparison scheme.Under the requirements of actual industrial applications,it realizes the effective prediction of the specific surface area of the cement product and the automatic optimization of the model.