Researches On Batch Process Quality Prediction Based On Local Partial Least Squares

As one of the sub items of the second industry,batch process is a relatively representative production form that belongs to mass production in modern society.Due to the good flexibility,high value-added products and its long history of development,batch processes have been applied in fields such as pharmacy,food production,chemical and material industry and so on.However,because of its different characteristics comparing with continuous processes,it brings complex production mechanism,variety of working conditions,as well as different production situations.With the application and development of modern sensors and those corresponding measuring techniques,large amounts of data have been accumulated along with cycles of production activities,and inevitably,such data contain information that has not been excavated.By utilizing these data,data driven based multivariate statistical algorithms can implement process monitoring,fault detection and diagnosis,as well as quality prediction and soft sensing on corresponding processes,which has an unneglectable supplement to process performance evaluation,production safety,problem analysis,process optimization and real-time control for industrial processes.Based on the requirements of the on-line prediction for batch production process,batch process characteristics,including production rules and forms of data is analyzed using multivariate statistical algorithms in this paper,a series of quality prediction methods forced on batch processes are proposed,which provides a strength for process optimization and safety production in batch processes.The main research contents are as follows:(1)Aiming at the multi-stage local model selection problem of batch processes,a batch process quality prediction method based on accelerated Just in Time Learning-Multiway Partial Least Squares(JITL-MPLS)is proposed.Taking PLS as the main step,the method unfolds and expands three-way data of batch process into two-way data variable-wisely,dealing with on-line computational load issue caused by excessive data by pre-clustering process data after introducing fuzzy C-means and just in time learning.By considering the potential large periodic characteristics of some batch processes,the real-time quality of online quality prediction process is improved under the premise of maintaining all the historical data as candidate data at the stage of online model training.The application of penicillin simulation platform and Escherichia coli in industrial process show,compared with similar algorithms,the proposed algorithm can improve the system operation speed in the similar situation,while maintaining the algorithm in low sensitivity to those new parameters to improve the robustness of the algorithm itself,maintaining the prediction accuracy in the same extent.(2)Aiming at the problem that the PLS algorithm,especially its local algorithm,lacks the model reselection mechanism under fault situation,a MPLS quality prediction method based on local sample reselection under process fault situation is proposed.If proceeding under minor faults,the similarity measure mechanism of traditional local algorithm can cause selection error when selecting online training sample during fault stage due to the drift of the variables,which can lead to the reduction of model accuracy.By considering this problem,the proposed method uses statistical limits derived from the contribution chart of process monitoring,reselects training data from the historical data under the guidance of dynamic time warping index based on the stable variables selected during offline stage,and proceeds with quality prediction during fault alarm.The test of the penicillin simulation platform and the fermentation process of E.coli shows that compared with the traditional methods,the proposed algorithm can maintain the quality prediction accuracy of the process to an extent under non-serious faults.(3)By analyzing the selection trend of local JITL method to online training database and considering the different data amount at different stages of the batch process,a local adaptive JITL-MPLS method for batch data is proposed.The strategy combines the unique batch characteristics of the batch process and with the collaboration of the original similarity measure function to dynamically restrict the just in time learning algorithm in a reasonable range,in order to control the amount of training sample when selecting online training database,and adjust each built model.Experiments on penicillin show that the proposed algorithm can improve the prediction accuracy of quality data compared with traditional methods.Then,through the analysis of the experimental results,a new model predictive effect evaluation index is proposed.Compared with the prediction accuracy,the new evaluation index focuses on the stability of the algorithm prediction effect.(4)The nonlinear problems inherent in batch process are analyzed,focused on the problem that traditional kernel technique can easily cause dimension disaster as well as the application potential of the kernel trick on the quality data,the reduced dual kernel multiway partial least squares algorithm for the quality prediction of nonlinear batch processes is proposed.First,extract the feature vectors that carry relatively more information through the feature vector selection.Then construct the dual kernel matrices by projecting both measured data and quality data into two kernel spaces,a soft sensing model is established using data in high-dimensional kernel space to improve the compatibility of the algorithm in the systems especially that with low computational capacity.Finally,the high-dimensional prediction data in the kernel space are restored to the original space by the online reverse projection algorithm,solving the problem that the data in the high-dimensional space cannot be read directly.After tests of both numerical examples and the batch fermentation process of the actual Escherichia coli,results show that the proposed algorithm can obviously improve the accuracy of the online quality prediction,the system requirements and the robustness comparing to similar algorithms.(5)Discuss and analysis the situation that the amount of online measured data is lager than that of quality data,the problem of relative insufficient quality data is put forward,and have its reasons are discussed,describe the characteristics of measured data which without corresponding quality data and its potential of the improvements on the prediction accuracy.The proximity MPLS quality prediction method focused on lack of quality data is proposed.The penicillin data experiments show that compared with the traditional methods,the proposed method can not only improve the prediction accuracy of the local method,but also improve the prediction accuracy of the global method.After that,the routes of kernel tricks and local algorithms are discussed on the aspect of dealing with nonlinear problems,and illustrate that the two methods have different directions in solving the same problem.Further more,clarify the conflict between traditional local algorithm and PLS's concept of quality prediction when selecting online training sample from history database.Based on this,the kernel trick is applied to the just in time learning strategy,and experiments are carried out.The experimental results verify the hypothesis and the improvement of the prediction accuracy.