Research On Just-in-time Learning Based Modeling Methods For Quality Prediction Of Industrial Production Process

With the rapid development of industrial technology,the scale and complexity of industrial production system have increased greatly.Many factors in production process could affect product quality,and the process mechanism is not clear.In addition,with the market demand for personalized and diversified products,multi-variety and small batch production mode gradually becomes popular.Many kinds of products need to be produced in small quantity,and the delivery time is short,which makes the operation condition or process state switch frequently in production process.All of these ask higher requirements for quality control of production process.On the other hand,with the rapid integration of information technology and manufacturing industry,various sensors and distributed control systems are widely used in production process,which makes a large number of industrial process data can be easily obtained.Under this background,datadriven quality analysis and modeling technologies have developed rapidly.Among them,quality prediction technology can realize fast detection of key quality indicators that are difficult to measure by establishing a mathematical model between easily measured process variables and key quality indicators,so as to provide important information for real-time optimization and management decision-making of production process.And it has developed into a key technology to solve the quality control problems of complex industrial production process.Traditional quality prediction models are mostly obtained by offline training based on historical data,which can not adapt to the time-varying characteristics of industrial processes,resulting in degradation or even failure of model performance over time.And retraining the model will increase time and economic cost.The just-in-time learning(JITL)based quality prediction modeling method has strong adaptive ability to process timevarying characteristics.And because of its simple modeling principle,easy implementation,the JITL method has attracted extensive attention of many scholars.However,due to the large scale and complex mechanism of modern industrial production system,the production process often presents nonlinear,time-varying,dynamic and other complex characteristics.The quality modeling data may also have complex characteristics such as small labelled sample size,unbalanced data distribution,highdimensional collinearity of input variables.Therefore,the JITL-based quality prediction modeling still faces many challenges.This dissertation mainly aims at the quality control of two kinds of typical production processes,i.e.process with small sample size and multi-operating conditions,and nonlinear large-scale continuous process.On the basis of comprehensively considering the complex characteristics of production process and modeling data,some improved modeling methods for quality prediction based on JITL are proposed from the aspects of modeling parameter optimization,similarity measurement improvement and adaptive update management of modeling data.The main research contents are as follows:(1)Just-in-time learning modeling based on PSO optimization for quality prediction of processes with small sample size and multi-operating conditionsUnder the production mode of multi-variety and small batch,the same production line of industrial process often needs to produce a variety of products with limited quantity,and the demand for different products may vary greatly,which leads to multi-operating condition characteristics of process,and also results in quality modeling data of small sample size and uneven distribution.For the quality prediction of the production process with small sample size and multi-operating conditions,the locally weighted regression(LWR)of JITL modeling methods is adopted to reduce the modeling requirements for sample size.Meanwhile,to deal with the characteristics of uneven data distribution,a locally weighted bandwidth parameter optimization scheme is designed by combining particle swarm optimization(PSO)algorithm and K-nearest neighbor method.Therefore,by adaptively selecting the optimal bandwidth parameter for each query,the LWR model can achieve good and stable prediction performance under uneven data distribution.In the calculation process,the off-line and on-line calculation contents are cleverly designed to ensure that the model can quickly respond to the prediction query.The effectiveness of the proposed method is verified by a numerical case,and industrial cases for detection of chemical composition of molten iron and tensile strength of grey castings in casting production process.(2)Improved JITL modeling method for quality prediction of complex nonlinear large-scale continuous processLarge-scale continuous production process usually has many complex characteristics,such as nonlinear,high-dimensional and multicollinear,large range of operating conditions,time-varying and so on.Aiming at the quality prediction modeling problem for such complex nonlinear processes,a double weighted JITL modeling method,with the similarity measure improved by combining mutual information(MI)and partial least squares(PLS),is proposed.The proposed method can make full use of input and output information of samples for similarity measure,overcome the influence of nonlinearity and collinearity,and effectively ensure the accuracy in nearest neighbor sample selection.In local modeling,considering the different importances of variables or samples to the model,a double weighted partial least squares modeling method is proposed,in which the variables are weighted by MI and the samples are weighted by similarity measure,so as to effectively enhance the ability of the model to describe complex nonlinear characteristics of process.A detailed two-stage calculation framework is designed to realize this modeling method.Finally,through a numerical simulation case and debutanizer process case,it is verified that the proposed modeling method has good prediction performance and can guarantee fast prediction response speed.(3)Adaptive updating management of quality modeling data based on density under operating condition changesWhen the process operating condition changes to a new state beyond the coverage of historical data,the JITL-based quality prediction model will fail because it cannot select the accurate nearest neighbor samples for the new query.To solve this problem,an adaptive updating management method of quality modeling data based on density is proposed.Firstly,the data density is defined based on sample similarity measure calculated in JITL algorithm,and a method to determine data density threshold parameters is proposed based on the cross validation prediction errors.Then,to effectively control the modeling sample size,a data management mechanism for selectively updating new samples and eliminating old samples based on density is designed,which ensure the information richness of the modeling data and avoid data redundancy.Finally,the proposed data updating management method based on density is applied to the JITL-based modeling for quality control.Through a numerical case and a sulfur recovery industrial case,it's verified that the proposed method can guarantee the JITL-based model to adapt to new working conditions quickly,and effectively enhance the adaptive tracking ability of the quality prediction model to operating condition changes,so as to maintain the long-term effectiveness of the model in practical use.Aiming at the complex characteristics of industrial production process and quality modeling data,this dissertation mainly explores improved JITL-based modeling methods for process quality prediction by comprehensively considering several key performance indicators such as prediction accuracy,adaptability and calculation efficiency.Some improved modeling methods are proposed,including JITL-based quality prediction model with bandwidth parameter optimized by PSO,the double weighted JITL modeling method with improved similarity measure based on mutual information and partial least squares,and adaptive updating management method of quality modeling data based on density under operating condition changes.In order to verify the effectiveness of the proposed modeling methods,several numerical simulation cases and industrial production process quality control cases are used.The experimental results show that the proposed modeling methods can better deal with the complex process characteristics,and has strong adaptability.The related research achievements can provide some solutions for quality control of complex time-varying industrial processes.