Soft Sensor Methods Based On Data And Mechanism Hybrid Drive For Complex Industrial Process

In complex industrial process,in order to obtain the key information of the controlled object in time,it is necessary to check and measure some parameters qualitatively or quantitatively.However,some key information of process variables cannot be obtained due to the complex industrial system structure and variable characteristics as well as cost and technology factors.Soft sensor technology is widely used to solve the mentioned problems,At present,most of the researches on complex industrial process are devoted to modeling method while neglecting the influence of input characteristics on target variables.Moreover,due to the difficult characteristics of complex industrial modeling,there are relatively few researches on hybrid drive soft sensor methods.This paper conducts research on the processing of input data features before modeling of complex industrial processes and the establishment of hybrid models.Taking the air preheater rotor as the research object of complex industrial process,the research content are applied to the prediction of the air preheater rotor deformation.The specific research of this paper are as follows:(1)Before carrying out hybrid model,the data driven and mechanism driven model should be established.Firstly,the thermal and stress coupling analysis of the LAP 14948/2400 air preheater is carried out by WORKBENCH19.0 software.It is known that the shape of rotor thermal deformation is "mushroom".Then,the thermal strain caused by thermal stress is mathematically described according to simplified mechanisms and principles such as the law of conservation of energy and boundary conditions,and the thermal deformation of the rotor based on the mechanism model is established.(2)After getting the industrial field data of the air preheater,the mechanism model analysis method is used to screen out the auxiliary variables related to the thermal deformation of the air preheater rotor,and then the Spearman correlation coefficient method is used for correlation analysis.The auxiliary variables closely related to the rotor deformation are obtained as reliable input.Affected by the accuracy of measuring instruments and poor industrial site environment,the data collected in industrial field inevitably have deviations,and the accuracy of the input data has great impact on the results of data-driven soft sensor modeling.So,before data-driven modeling,this article denoising the input data.Then,after normalizing the denoising data,mutually exclusive training samples and test samples are selected for data-driven modeling.(3)The data sets selected in(2)are modeled by BP neural network,Support vector regression(SVR)and SVR based on Particle Swarm Optimization(PSO)respectively,and then the deformation of the rotor is predicted.Then,use the wavelet denoising data combined with SVR to establish a data-driven soft sensor method to predict the deformation of the rotor.The results show that the prediction effect of data-driven modeling based on denoised data is better than the modeling method without denoising,and the complexity of the model is reduced.(4)The data-driven model of combining wavelet and SVR established in(3)and the mechanism driven model established in(1)are combined in parallel,series and hybrid mode respectively,and the hybrid models based on error,superposition and weight factor are established.These three hybrid models are used to predict the thermal deformation of rotor.Experimental results show that the modeling method based on weight factor is better than the other two hybrid modeling methods and also better than the single model,but the performance of the error-based hybrid modeling method is not as good as the single model.There is no fixed standard for the establishment of hybrid-driven model,and the effect of hybrid model is not necessarily better than a single model.It must be combined with specific industrial process analysis.This paper carries out soft-sensing modeling and analysis on the single models and hybrid models,then applies them to complex industrial processes to verify.It shows that the data driven modeling method based on the combination of wavelet and SVR proposed in this paper more accurate than those data-driven modeling methods without denoising.At the same time,three hybrid drive models are compared to find the most accurate to predicting the rotor deformation.It provides more accurate deformation for the research of air preheater air leakage technology.