The Research On Soft Sensing Of Free Calcium Content In Cement Clinker Based On Attention Mechanism And Convolutional Neural Network

The content of free calcium oxide(f-Ca O)in cement clinker is an important index for cement quality.At present,the monitoring of f-Ca O content mainly depends on the off-line manual detection with long sampling interval(typically one hour).The off-line detection has an obvious time lag in guiding real-time quality control and cement production optimization.Therefore,based on the data-driven soft sensing mehod,this paper puts forward a soft sensing method of free calcium content in cement clinker based on attention mechanism and CNN to realize the real-time online cement quality monitoring.The specific research work is as follows:Firstly,the paper introduces the new dry cement production process and the formation mechanism of f-Ca O in cement clinker during the cement calcination.The strong coupling of cement firing process and the time consumption of each production stage are researched deeply.The difficulties in the f-Ca O content modeling process are analyzes and the corresponding solutions are proposed.The variables related to the cement clinker f-Ca O content are selected as input variables and the related historical data are preprocessed.Secondly,aiming at the characteristics of dynamic nonlinearity,strong coupling,time-varying delay and uncertainty in cement firing process,a soft sensing model based on multivariate time series analysis and CNN(MVTS-CNN)is established.The model first uses the multivariable time series analysis method to process the input variable time series,then applies the multivariable time series feature extraction method to extract the features of the graph structured multivariable time series.The experimental results show that the MVTS-CNN model has high prediction accuracy and generalization ability,and can be used for real-time online f-Ca O content monitoring.Then,according to the characteristics of multiple working conditions in the process of cement firing,a soft sensing method based on single attention mechanism and MVTS-CNN(AMTS-CNN)is proposed.From the aspects of data redundancy and condition features under a certain produce condition,this paper introduces time-series attention mechanism and channel attention mechanism into MVTS-CNN model separately and gets corresponding AMTS-CNN model.The experimental results show that the AMTS-CNN model based on channel attention mechanism has better prediction effect,and compared with MVTS-CNN,AMTS-CNN has better performance and better adaptability in different produce conditions.Finally,in order to enhance the model performance in the dynamic and multiple produce condition,a soft sensing method based on dual attention mechanism and MVTS-CNN(DATS-CNN)is proposed.Serial attention mechanism and parallel attention mechanism are constructed respectively based on the position difference between time-series attention and channel attention,and introduced into the MVTS-CNN;Meanwhile,the corresponding DATS-CNN models are constructed according to the position of time-series attention and channel attention in serial attention mechanism and the feature fusion method in parallel attention mechanism.The experimental results show that compared with AMTS-CNN,DATS-CNN model has better prediction effect in multiple dynamic produce process.