Research And Application Of Data-driven Modeling Methods For Difficult-to-measure Parameters In The Process Of Tire Production

In the process of tire industry,there are some parameters which are difficult to measure in real time.However,these parameters directly affect the production efficiency and tire quality.For example,the rubber mixing process is an intermittent process,and the running time is short,it is difficult to realize the on-line measurement of the relevant parameters based on universal sensors.Hence,this dissertation focus on just-in-time soft sensor modeling approach in the field of “rubber mixing process”.The main research work can be summarized as follows:(1)In the process of industrial intermittent rubber mixing process,multiple operating phases can be determined according to the detailed technical formulas.The process has obvious similarity and regularity within each phase.Based on the analysis of the implicit feature information in different operation periods,a just-in-time learning soft sensor modeling approach is proposed.First,history samples from process are trained based on the Gaussian Mixture Model.Multiple Gaussian components represented these phases can be obtained.Mahalanobis distance for each Gauss component of the test samples and the historical samples is calculated.The integrated distance between test sample and historical sample can be formed by ensemble these Mahalanobis distances.Local samples are selected according to the integrated distance,and then local models are built.Data from industrial rubber mixing process is used to model and predict the parameters.The results of the proposed method and the comparison with other soft sensing method have verified the effectiveness of the proposed algorithm.(2)The rubber mixing process has obvious non-linear and time-varying characteristics.The performance of the model is affected by the large amount of redundant information contained in the high dimensional process data.In order to solve the problem,a soft sensor modeling approach based on local learning and ensemble learning has been proposed.First,multiple groups of samples are built by repeated random resampling in the history sample sets.Related variables are selected using partial mutual information theory based on each group.Using these selected variables,historical samples are predicted through just-in-time Gaussian process regression method.Several variable groups with best prediction performance can be determined as final variable groups.Then,multiple local outputs can be obtained by modeling based on each variable group using testing samples.Finally,the prediction value is get by ensemble the local model output using Bayesian method.Experimental results using field data show that the proposed method has obvious advantages.(3)Proper variable selection and variable grouping of industrial data can improve the prediction ability of soft sensor model for complex industrial data.In the process of determining the variable group,if we can try to ensure that each variable group is independent,and reduce the information redundancy between variables,we can further improve the reliability and accuracy of the model.Although Gauss process regression can be used to deal with nonlinear problems,the amount of computation load is relatively heavy.To solve this problem,a soft sensor modeling approach based on independent principal component analysis is proposed.The independent principal component analysis is used to separate multiple variable groups in the probability sense.The information redundancy among variable groups can be reduced.The kernel learning method is used to deal with the nonlinear characteristics of the process and reduce the computational complexity of the prediction model.The validity of the method is verified by the prediction of the quality property of semi-finished products in the process of rubber mixing.(4)In the process of tread rubber manufacture,the length of tread rubber is difficult to be measured in real time.The traditional contact measuring approach always cannot guarantee the measuring accuracy because of the slippage between the tread rubber and the conveyor belt.It is also difficult to meet the measurement requirements of the production process by the common non-contact measuring method.In the image recognition process,there are several technical difficulties such as: The tread of the tread rubber is inclined,irregular,difficult to identify and fit,the color of the conveyor belt is similar to the tread rubber after the long term use,the edge of tread is difficult to recognition,etc.In order to solve these problems,this dissertation studies the optical recognition and fitting algorithm.An optical measurement scheme of CCD camera driven by servo motor is designed.The effectiveness of the algorithm and the scheme is verified by the testing results in the industrial field.In the research process of soft sensor modeling method and optical measuring approach,a data acquisition and information management system software has been developed and applied to manufacture process according to the requirements of industrial manufacture.The system can quite well meet the requirements of the production process and the operating habits of the using company.