Transfer Learning Modeling Method And Its Application For Multi-grade Processes

The characteristics of multi-grade are universally appeared in chemical production processes due to diversified market demand.The influence of multi-grade processes on model performance degradation can not be ignored,especially,in most working conditions,insufficient labeled data makes it difficult to establish a robust prediction model,which seriously affects smooth operation of the production process and product quality.Transfer learning modeling method is rarely applied in the field of chemical multi-grade processes.Therefore,it is of great research value and significance to use transfer learning to solve the modeling problem of multi-grade processes.In this paper,the existing problem of multi-grade processes and the limitations of corresponding methods are reviewed.As a novel modeling method,transfer learning method is proposed to solve the modeling problem of multi-grade processes.First,For the target domain condition with a small amount of labeled data,a parameter transfer learning method is proposed to achieve more accurate prediction result.Furthermore,in view of the large discrepancy of data distribution among diffenent operation conditions,an adversarial transfer learning method is proposed to establish a great prediction model.Finally,in order to effectively combine the advantages of both transfer learning and just-in-time learning,a just-in-time transfer modeling method is proposed.The main work and contributions are as follows:(1)For the target domain operation condition which has too little labeled data to establish a good prediction model,a parameter transfer modeling method based on domain adaptation extreme learning machine is proposed.The model parameters are effectively selected by combining the fast leave-one-out cross-validation,so as to solve the modeling problem of multi-grade processes.The experimental results show that the proposed method can effectively improve the prediction performance.(2)Given the problem that parameter transfer modeling method is not suitable for multi-grade processes with large data distribution discrepancy.In order to effectively reduce the distribution discrepancy among different operation conditions,an adversarial transfer modeling method based on cycle feature transformation adversarial network is proposed.The experimental results demonstrate that the proposed method cannot only improve the performance of traditional modeling method,but also further improve the performance of parameter transfer modeling method.(3)Considering that just-in-time learning has better prediction performance than traditional method by selecting high similar sample,but it can be still affected by large distribution discrepancy.At the same time,the transfer learning method do not have adaptive adjustment ability.To overcome the above problems,a just-in-time transfer modeling method by combining transfer learning and just-in-time learning is proposed.The experimental results demonstrate that the proposed method can further improve prediction performance without available labeled data in target domain condition.