Modeling And Quality Prediction Of Industrial Big Data Based On Different Learning Paradigms

With the advent of the era of intelligent manufacturing and the industrial internet,more opportunities and challenges have been brought to modern industry,which has promoted the continuous development of traditional manufacturing in the direction of digitization,networking and intelligence,and gradually deepened the process of industrial automation.As the core link of intelligent manufacturing and industrial internet platforms,industrial big data analysis has received more and more general attention from academia and industry.How to efficiently mine the high-value information contained in industrial big data and how to use them to solve the problems of actual industrial processes is one of the current hot directions.At the same time,with the advanced development of industrial automation towards knowledge automation,the learning paradigm of industrial big data analysis has also evolved accordingly.Therefore,from the perspective of industrial big data analysis,this thesis studies modeling methods under different data characteristics and process characteristics,which are applied to typical application scenarios such as quality prediction and process monitoring of industrial processes,and actively explores the revolution of learning paradigm under the background of industrial big data.The main research contents of this paper include the following five parts:(1)To deal with the high-dimensional characteristic modeling problem of industrial big data,a parallel modeling framework based on latent variable models is proposed for process monitoring and quality prediction.From the perspective of the autoencoder,compared with the traditional latent variable model,it is extended to the corresponding parallel nonlinear version.In order to further extract deeper nonlinear features from the process data,the basic shallow autoencoder model is extended to a stacked structure,which provides a deep generation struc-ture for nonlinear process monitoring and quality prediction.In the analysis and modeling of industrial big data,the process modeling analysis method combined with parallel computing strategy has higher computational efficiency and performance advantages compared with tradi-tional modeling algorithm.(2)Aiming at the problem of nonlinear characteristics in industrial processes and the situ-ation that data contains noise,a nonlinear variational Bayesian factor regression is proposed for quality prediction.Based on probabilistic modeling,combined with nonlinear mapping tech-nology,the linear probabilistic quality prediction model is extended to a nonlinear form.Since the complexity of parameter update has a great relationship with sample size and variable di-mension,to solve this kind of large-scale and high-dimensional process data modeling task,a nonlinear variational Bayesian factor regression based on parallel framework is proposed.In this way,the introduced parallel strategy effectively transforms the heavy calculation load into multiple subtasks through the two levels of sample parallelism and variable parallelism.In gen-eral,the proposed method not only improves the nonlinear data processing capability,but also further improves the computational efficiency of the model.(3)Aiming at the time-varying and nonlinear characteristics of the process,a parallel just-in-time learning framework is proposed,including parallel searching,parallel modeling,model library management and database management.As the core of the framework,the introduc-tion of parallel computing strategies not only enables just-in-time learning to effectively utilize the full advantages of industrial big data information,but also improves its search capabilities and efficiency under industrial big data.In addition,the adopted model library management strategy can operate on existing similar models by querying similar samples,improving the real-time performance of parallel just-in-time learning.Meanwhile,by selectively adding new data,the developed database management strategy not only alleviates the problem of informa-tion redundancy,but also reduces the search pressure caused by the increase of the database.Furthermore,considering the data noise,taking the traditional variational Bayesian factor re-gression model as an example,it is transformed into a parallel Bayesian just-in-time learning method for streaming industrial big data modeling and analysis.Subsequently,in order to fur-ther improve the performance of the model,the above linear method is extended to a nonlinear form,and a parallel nonlinear Bayesian just-in-time learning method for industrial big data is proposed.(4)Aiming at the problem of adaptive modeling of industrial big data,an adaptive quality prediction method based on streaming variational Bayesian factor regression model is proposed.On the basis of factor regression,this method introduces streaming variational Bayesian,and updates the posterior distribution of model parameters in real time according to the changes of the actual data stream.In order to better adapt to the time-varying nature of the industrial process,the symmetric KL divergence is introduced in the update process to determine the se-lection of the prior distribution,so as to realize the adaptive update of the model.In this way,not only the big data modeling and the timely tracking of the change trend of quality variables are successfully solved,but also the update calculation time is reduced.Subsequently,a parallel computing strategy is introduced,and a more efficient streaming parallel variational Bayesian factor regression is further proposed.In the quality prediction application under the stream-ing industrial big data,the proposed method shows higher training efficiency and prediction accuracy.(5)Aiming at the problem that the traditional learning paradigm cannot effectively ac-cumulate knowledge in process learning under the streaming industrial big data scenario,a lifelong Bayesian learning machines framework is proposed.Combining the Dirichlet process mixture model and lifelong learning ideas,using the nested variational bounds of the infinite non-parametric model,model expansion and model optimization are carried out.The frame-work can not only adaptively create and merge the number of components,but also consider the problem of random noise in the process data.In this continuous learning method,the previ-ous data sets and their knowledge information can be memorized through sufficient statistical learning,and there is no need to revisit the past data sets to complete the retention and accu-mulation of knowledge learning.Taking Dirichlet process Gaussian mixture regression as an example,process modeling is carried out under this framework.Compared with traditional adaptive methods,this method has advantages in modeling efficiency and model performance,and the effectiveness and feasibility of this method are verified through examples.