Soft Sensor Modeling For Raw Material Decomposition Rate In Cement Process Using Improved Echo State Network

Pre-decomposition technology is the key to the new dry process cement production technology.The raw material decomposition rate is an important indicator in the cement production control process.It is important to ensure the accuracy and real-time nature of the decomposition rate.At present,industrial sites use manual sampling and laboratory analysis to obtain the decomposition rate.The obtained data has a long lag and cannot reflect the real-time operating conditions,which cannot meet the needs of industrial real-time control.Data-driven soft measurement modeling method uses the information of related process variables in the system to establish the mapping relationship of the dominant variables.Using the process variable information in the pre-decomposition system to establish a soft-sensor model of the raw material decomposition rate,it can provide accurate and real-time decomposition rate information for the control process and ensure cement quality.This paper analyzes the characteristics of the process and variable data of the pre-decomposition system,establishes a dynamic soft-sensor model of correlation reasoning based on fast sampling rate input data and slow sampling rate output data.Data to meet the needs of industrial control systems.Considering the characteristics of dynamic multivariate time series data,this paper proposes to establish a soft-sensing model based on an improved echo state network,which is applied to the prediction of raw material decomposition rate.main tasks as follows:(1)The data used to build the soft measurement model includes fast sampling rate input data and slow sampling rate output data.It has the characteristics of multivariable,nonlinear,and strong coupling.Therefore,it is necessary to choose a network with good nonlinear mapping ability to learn To achieve a more accurate prediction effect.This paper proposes an ESN incremental random learning algorithm based on orthogonal least squares(OLS).The learning model uses the OLS algorithm to set a supervision mechanism to gradually increase the nodes of the random static non-linear layer,and sets an adaptively adjusted threshold to control the speed of the model’s error decline.In this paper,a comparative experiment is designed to observe the differences in prediction results of the four models(φ-ESN,SCN-φESN,OLS-φ-ESN,and ESN-RBF)in the two sets of benchmark nonlinear time series tasks.The experimental results show that the incremental random learning model based on the adaptive threshold OLS algorithm is improved in prediction performance and the training model is more compact than the incremental random learning model based on orthogonal matching pursuit(OMP)algorithm..Based on the experimental results,the differences between the model using the incremental deterministic learning algorithm ESN-RBF and the model of the incremental random learning algorithm are further analyzed.In this paper,we explore the differences between the predictive ability and generalization ability of OLS-(p-ESN under the action of four activation functions.The resulting model is more compact.(2)The above experimental research process found that the structural parameter spectral radius of the reserve pool has a significant effect on the prediction capability of the φ-ESN network,but the effect of improving the network prediction capability only by adjusting the spectral radius is limited.This paper studies experimentally how the reserve pool neuron adaptively adjusts its internal weights and biases through IP learning when the expected distribution of neuron output is a Gaussian distribution with a given constraint,so that the output converges to the expected output distribution.The experimental results show that the prediction performance of(p-ESN based on the IP learning rules is better than the best performance obtained by adjusting the spectral radius.In addition,the variance of the expected distribution has a significant impact on the final prediction performance.(3)In two sets of benchmark data prediction tasks,OLS-φ-ESN exhibits excellent prediction results.Therefore,in order to predict the decomposition rate of raw materials,this paper uses OLS-φ-ESN to build a soft measurement model.First,the fast-rate sampled input data is anomalous,filtered,and normalized;then the PCA method is used to reduce the dimension,and the low-dimensional expression of the input data is sent to the reserve pool of the OLS-φ-ESN model for dynamic spatial-temporal feature extraction,Get the reserve pool state time series;then extract the reserve pool state vector with the slow-rate sampling and output mentor samples at the corresponding time as the input vector of the static random nonlinear layer;Finally,linear regression is performed through the output layer.The established OLS-cp-ESN model is compared with the φ-ESN model.The experimental results show that the prediction performance of the two models is close,and the prediction performance of the OLSφ-ESN model in the benchmark data prediction task.Obviously better than the φ-ESN model.Further analysis shows that the random algorithm and the incremental random algorithm have not obvious advantages in complex modeling tasks such as high-dimensional small sample soft decomposition of the decomposition furnace.The model performance can be further improved by expanding the sample set size.