Research And Application On The Optimization Theory Of Gaussian Process For The Manufacturing Process

This thesis further expand the application of research in the field of the optimization of the manufacturing process and sugar processing based on the full study of the method of GP modeling for the optimization of injection molding process and the cooperative control field of parameters.According to the volatility characteristics of the manufacturing process and the problem that it is not effective for real-time monitoring and forecasting for the quality of multiple target and the real-time control for production parameters,this thesis explores the use of the machine learning of GP as the axis and puts forward a set of overall solution for the trends of manufacturing process and the optimization of parameters based on the field data modeling and data processing from the manufacturing process.According to the acquisition process of the data of the sugar production process,this thesis introduces the method of the chaos fractal theory for analyzing the time series and studies the problem of the trend forecasting and quality control in the process of production and further expands GP for the practical applications in the field of the complex manufacturing combined with in-depth study of GP model.For the problem of the field data containing noise and interference in the process of production,the chaos fractal theory is introduced to data for time series analysis.Given the problem that the traditional detrended fluctuation analysis(DFA)can only separate the trend of the discrete polynomial,this thesis uses an adaptive fractal(AFA)model to reconstruct the phase space of time series data and portray around the fluctuation of the overall trend and the same time maintain the characteristic of the scale in order to achieve the separation of the input data containing noise and interference and provide effective sample space for the establishment of the model.At the same time,this thesis use the AFA model for each process parameters and the data set of quality indexes to analyze the change of Hurst index in order to characterize the long-range correlation in the time series and achieve the trend forecast of the production process.For the industrialized GP modeling that require huge amount of data information and high operational efficiency,this thesis proposes improved model for GP.Firstly,the thesis uses enhanced translation spread Latin Hypercube experimental design methods(ETPLHD)to improve the quality and efficiency of the traditional Latin hypercube sampling methods in order to achieve the construction of uniformity and effectiveness of the GP agent model sample set.Then,the matrix approximation is in the course of the training of the agent model and the caching method is used to speed up the training of algorithm.Finally,using the improving EPI method in maximizing the probability to meet the GP regression model obtaining the global optimal solution in order to meet the actual needs of the improved GP regression model applied to manufacturing process.Contraposing the existing problem that classical time series model can not perform efficiently when analyzing prediction problems of non-Gaussian process data series.This thesis improved the covariance matrix form of the Gaussian process model which would be applied to analyze the time series model later,making this improved model perform well for data analysis of stationary time series,non-stationary time series and non-Gaussian process series.And Mackey-Glass(MG)series which is one of the non-Gaussian process was employed to validate the prediction properties of the improved model for both single-step and multi-steps test.The manufacturing process optimization,whose modeling process needs to sufficiently utilize multiple information and the cross correlativity of information in manufacturing process,in essence is one of the multi-objective optimization problem.A multi-objective collaborative optimization model based on the Gaussian process was proposed in this thesis.Firstly,by combining the non-finite correlation matrix with the covariance function of Gaussian process model,the responding covariance function contains the effective definitions of both spatial correlativity of input variables and cross correlativity of output variables.Then construct the non-finite correlation matrix by means of the decomposition of hypersphere parameterization.It avoids directly solving the matrix positive constraints in an optimization problem and reduces the computation complexity at the same time.At last,the consequence of numerical research indicates that the proposed multi-objective optimization model based on Gaussian process can not only acquire a less root mean square error but also fully reveal the cross correlativity of each quality objective.The solution process of classical multi-objective optimization methods can not give consideration to both the convergence of Pareto optimum and the diversity of non-dominated solution sets meanwhile.To address this problem,an enhanced multi-objective PSO optimization algorithm was proposed in this thesis.For improving the convergence,this proposed algorithm employed an efficient region searching strategy proposed in this thesis which is combined with Gaussian mutation operator,and utilized mathematical programming method to quickly converge to the Pareto optimum solution.On the other hand,merging the common advantages of Maximin algorithm and crowding distance,an enhanced strategy based on Maximin fitness function,in which both the distance information from the former to latter particle and particle itself biased information were considered,was proposed to realize elitism selection and the diversity of non-dominated solution sets.This thesis chose the machine learning theory of Gaussian process as its main research line to study the problem of the analytical method of time series and the theory of multi-objective collaborative optimization problem.By employing the time series model based on chaos and fractal theory,the preprocess problem of noise reduction of a set of field data from realistic manufacturing has been successfully addressed.Additionally,both the ion and prediction of the global trends of key parameters and quality objectives were realized.The cross correlativity between parametric variable and quality objective in manufacturing process was studied intensively,and at the meanwhile a more reasonable modeling method for nonlinear multi-objective collaborative optimization was realized in this thesis.Further more,the utilization of region searching strategy and enhanced Maximin fitness function strategy have successfully addressed both the convergence problem and the diversity problem when searching non-dominated solution.The proposed method was applied to predict and control the trend of the clarity as well as the color value in sugar clarification process so as to acquire a group of optimal process parameters.The consequences of the practical production verification test prove that the proposed method can not only effectively and process production data but also implement a better prediction of the trend.What's more,the proposed method can robustly realize the quality control and the process optimization with a high computation efficiency.In conclusion,the research in this thesis shows a highly theoretical value and scientific significance in these aspect such as the intelligent manufacturing,the prediction and the optimal control of quality fluctuation,etc.