Research And Application Of Virtual Sample Generation Method For Regression Modeling Of Industrial Process

The precise measurement of industrial process parameters contributes to the realization of intelligent detection,intelligent control and operation optimization,which are crucial to develop the intelligent manufacturing.For important difficult-tomeasure parameters of complex industrial process,the data driven modeling is usually applied to construct their soft-sensor.But it is generally limited by the scarcity of effective modeling samples.“ Small-sample ” exists defects such as insufficient characteristic information,large information interval and unbalanced distribution.These will make measured accuracy of soft-sensor low.Virtual sample generation(VSG)can improve it by expanding modeling samples.But existing VSG methods still have many research difficulties,such as improving the quality,determining the optimal quantity and evaluating the effectiveness of virtual samples.In addition,most methods are studied for classification issues.There is still a lack of in-depth research and application oriented to regression modeling of industrial process.Therefore,it is of great significance to improve the accuracy of soft-sensor to study above task.In response to generating redundancy virtual samples of existing VSG methods,a VSG method based on particle swarm optimization(PSO)was proposed.It can screen virtual samples to improve the soft-sensor performance with them.To further improve the comprehensive quality of virtual samples and determine their optimal quantity,a VSG method based on hybrid optimization with multi-objective PSO(MOPSO)was proposed.Nextly,in order to generate virtual samples conforming to the probability distribution of actual industrial data,a VSG method based on probability density function(PDF)of reduced feature was proposed.Finally,the proposed method was applied to soft-sensor modeling of dioxin(DXN)emission concentration in municipal solid waste incineration(MSWI)process.And a soft measurement system of DXN emission concentration based on VSG was designed and developed.The main work and innovations of the thesis are as follows:1.Most existing VSG methods generate redundant virtual samples.This restricts the effective generalization of soft-sensor.The optimization selection scheme for virtual samples based on PSO is proposed to eliminate redundant virtual samples.Firstly,the sample space is expanded and virtual sample inputs are generated by equidistant interpolation method,and then virtual sample outputs are obtained by mapping model.Then,PSO is used to optimize the selection of virtual samples.During this process,the soft-sensor is constructed with selected virtual samples.Its performance is taken as the evaluation index and optimization objective.So that the virtual sample set with optimal modeling performance is obtained.Finally,the mixing set of it and small samples is used to construct final soft-sensor.The method is verified by using benchmark and real industrial data.2.The improvement of generalization performance of soft-sensor is from the expanding of modeling samples with VSG technology.But too many virtual samples will lead to error accumulation,which results in modeling bias.In addition,the quality difference of virtual samples affects the degree of error accumulation.Thus,a VSG method based on hybrid optimization with MOPSO is proposed.Comprehensive learning PSO(CLPSO)is adopted to simultaneously optimize the generation and selection of virtual samples.The model generalization performance and the quantity of virtual samples are taken as the optimization objective.So that the comprehensive quality of virtual samples can be further improved and their optimal quantity can be determined.In addition,comprehensive evaluation index and distribution similarity index are proposed to evaluate the quality of virtual samples.The method is verified by using benchmark and real industrial data.3.Generally,small samples have the characteristics of sparse and unbalanced spatial distribution.These are the main factors leading to poor generalization performance of soft-sensor.A VSG method based on PDF of reduced feature is proposed to alleviate above defects of small samples.Firstly,principal component analysis(PCA)is used to reduce the features of small samples.So that the main probability distribution features of small samples can be conveniently identified and the problem of “holes” in the high-dimensional spatial of virtual samples can be avoided.Then,the kernel density estimation(KDE)method is used to estimate the PDFs of the reduced features.And then the virtual sample inputs conforming to the PDFs are generated.Nextly,the integrated mapping model is built to balance accuracy and randomness,and the virtual sample outputs are generated.Finally,CLPSO is applied to optimize the hyperparameters of above process.The aim is to obtain virtual samples with high comprehensive quality.In addition,the spatial distribution sparsity index is proposed to evaluate the quality of virtual samples.The method is verified by using benchmark and real industrial data.4.DXN emitted from MSWI process is highly toxic compound.Its detection is difficult,high cost and long time lag.This results in the difficulty of online monitoring and furtherly optimal control.Thus,a soft measurement system of emission concentration of DXN based on the VSG is designed and achieved.It carries out the soft-sensor of the emission concentration of DXN.In addition,it helps the staff to online monitor and make control decisions,and supports the further optimal control.