The Research On Soft Measurement Of Free Calcium Content In Cement Clinker Based On Long Short-term Memory Network And Attention Mechanism

The content of free calcium oxide f-Ca O in cement clinker is the main index to evaluate the quality of cement clinker.Real-time control of f-Ca O content is the key to ensuring scientific production of cement and improving cement quality.Aiming at the characteristics of time series coupling,dynamic nonlinearity and limited labeled data in cement clinker production,this paper uses data-driven soft measurement technology to study the f-Ca O content in cement clinker,and proposes a soft measurement method based on long shortterm memory network(LSTM)and attention mechanism to monitor f-Ca O content online in real time.The specific research work is as follows:Firstly,this paper introduces the new dry cement production process and the f-Ca O generation mechanism of cement clinker,analyzes the characteristics of cement production process,studies the influencing factors affecting the f-Ca O content of burned mature material in cement production process,selects relevant variables as input to the model through process analysis,and preprocesses its historical data.Secondly,aiming at the problems of dynamic nonlinearity and limited label samples in the firing process of cement clinker,a soft measurement model(LSTM-Attention)based on long short-term memory network and self-attention mechanism is proposed,which can fully explore the dynamic time series characteristics of input samples.The self-attention mechanism improves the ability of the original network structure to suppress redundant features by assigning the secondary weight of the hidden layer feature vector.Experiments show that the soft measurement model based on LSTM-Attention network structure can be used for real-time estimation of f-Ca O content of cement clinker.Then,aiming at the characteristics of timing coupling,limited label samples and timevarying delay in the firing process of cement clinker,a soft measurement model based on multi-task attention and Res-Bi LSTMs is proposed,which mainly introduces multi-task attention combining data reconstruction attention and quality supervision attention mechanism.In addition to considering the temporal correlation between the time steps of the input information,the model further extracts the correlation coupling relationship of the input information under the action of quality supervision.Finally,through experimental comparison,the proposed model has better measurement results under the condition of limited label samples.Finally,in order to further extract the coupling characteristics between the process variables and equipment existing in the firing process,a TPA-LSTM soft measurement model based on the multivariate temporal attention mechanism is proposed.CNNs are used to extract time pattern features with time-invariant characteristics,that is,time series data is transformed into "frequency domain" information.At the same time,a temporal mode attention mechanism is proposed,which aims to select relevant variables for weighting to achieve multivariate time series prediction.Finally,the effectiveness of this model is verified by experiments.