Research On Batch Process Quality Prediction Method Based On Multi-source Domain Transfer Learning Strategy

With the increasingly fierce competition in the international product market,the quality standards and requirements of batch process products are becoming higher and higher.In recent years,data-driven modeling method has been widely used in batch process product quality prediction,and sufficient data is the key to its accurate modeling.However,in the actual industrial production process,there is often a problem that it is difficult to accurately model due to insufficient data.Although transfer learning can effectively use source domain process data to assist target domain process modeling,improper transfer learning methods can easily lead to "negative transfer".This paper mainly focuses on "solving the problem of product quality prediction in batch process with insufficient data by using transfer learning".The main research contents are as follows:(1)Aiming at the problem that batch process is difficult to model accurately due to lack of data and strong nonlinear and multi-scale characteristics,this paper combined the advantages of transfer learning method and multi-scale kernel learning method,and proposed a batch process product quality prediction method based on JYMKPLS(Joint-Y Multi-scale Kernel Partial Least Squares)single-source domain transfer learning strategy.This method improves the modeling efficiency and quality prediction accuracy of target domain process by using the old process data of similar source domain through transfer learning.On this basis,aiming at the strong nonlinear and multi-scale characteristics of batch process data,the multi-scale kernel function is introduced to better fit the trend of data change,so as to improve the prediction accuracy of the model.In addition,the strategy of online model updating and data elimination is proposed to continuously improve the matching degree of the transfer model to the new batch process,so as to eliminate the adverse effects of the differences between similar processes on the transfer learning,and continuously improve the prediction accuracy.Finally,the effectiveness of the proposed method was verified by the simulation experiment of quality prediction in penicillin production process.(2)Aiming at the problems of low utilization of batch process data and "negative transfer" easily caused by transfer learning,this paper proposed a multi-source domain transfer learning strategy based on similarity measure for batch process product quality prediction.This method judges and counts the similarity and the amount of data between the source domain process and the target domain process,and takes it as the basis of transfer.Then,by setting judgment conditions,three alternative transfer strategies are provided,which can effectively reduce the possibility of "negative transfer" and improve the utilization rate of data resources and the efficiency of transfer learning.At the same time,a scheme is designed to update the transfer model and strategy in time by using online data,so as to better ensure the timeliness and reliability of the transfer learning strategy.Finally,the Mixed-effect Gaussian Process modeling method is taken as an example to verify the effectiveness of the proposed method through simulation experiment on quality prediction of cobalt oxalate synthesis process.(3)Aiming at the problem that the labeled data of batch processes are relatively small,and the unlabeled data are relatively large and unused,this paper proposed a product quality prediction method for batch process based on semi-supervised multi-source domain transfer learning strategy.Based on the multi-source domain transfer learning strategy,the semi-supervised learning is introduced in this method,and the pseudo labels are labeled for the unlabeled data of the source domain process and the target domain process respectively by the multi-source domain transfer learning modeling method,which can effectively improve the initial accuracy of the pseudo labels.Then,through the method of confidence determination and iterative learning,the pseudo labels with high confidence are continuously added to the labeled data set for retraining,which can further improve the prediction accuracy of the model.Finally,the effectiveness of the proposed method is verified by the simulation of the quality prediction of cobalt oxalate synthesis process.