Research On Soft Sensor Modeling Based On Weighted Gaussian Model Regression

Some quality variables in the industrial process are of great significance for the monitoring,control and optimization of the production process,but the reality is that the measurement of these quality variables is often limited by technical bottlenecks,device cost,and instrument reliability,which is difficult to directly use hardware sensors obtained online in real time,therefore soft-sensing technology is applied.As an alternative to hardware measurement,soft sensor can provide efficient and low-cost predictions for quality variables.Based on the existing research results of soft-sensing technology,oriented to industrial processes with non-linear and time-varying characteristics,considering the time delay information and time order in the process data,this thesis makes improvements based on the Weighted Gaussian model regression method.And the main researches are listed as follows:(1)Aiming at the problems of the industrial data timing mismatch,an online soft-sensor algorithm based on moving grey relational analysis algorithm is proposed.The method firstly uses the moving gray relational analysis to estimate process delay parameter for extracting the process delay information.When the new sample arrives,the model is reconstructed based on the time delay parameter.Considering the weight of the new sample relative to the training samples,a weighted Gaussian model is established as well as a joint probability density function of input and output variables.Finally,the output variables are estimated in real time based on the conditional distribution function.The effectiveness of the proposed method is validated through a numerical example and the industrial debutanizer column process.(2)Taking into account the problem of local modeling for each working point and the large amount of modeling calculation during local modeling,an improved online soft-sensor algorithm based on weighted Gaussian model is proposed.First,the data for modelling is reconstructed by time delay information extracted from the database to solve the problem of the industrial data timing mismatch.Then,the cumulative similarity factor is introduced into the selection rules of the model training set to improve the real-time performance of the model.When the new query sample arrives,an adaptive similarity threshold was used to determine criteria of updating local model of the current operating point,so as to reduce the model update frequency.The improved modeling method is applied to the prediction of butane concentration in the debutanizer column process.Simulation analysis shows that the improved method has higher prediction accuracy.(3)Traditional soft sensor modeling methods usually overlook time characteristic of data samples,which may result in poor performance in real-time and accurate quality prediction.Therefore,this paper proposes a weighed Gaussian model regression based adaptive soft sensor for handling time series data.Based on the Euclidean distance,the concept of time series is introduced into selection criterion of historical dataset to determine modeling neighborhoods of the current query sample.The locally weighted standardization(LWS)method is employed to transform the original collected data into an approximate Gaussian distribution to satisfy the model Gaussian distribution hypothesis.Accordingly,an adaptive similarity threshold is established to update parameters recursively for model update frequency reduction.In the experiment of sulfur recovery process,the time-ordered modeling method shows higher prediction accuracy and real-time performance.