Soft Sensor Modeling And Iterative Control Based On Os-elm-rpls For Batch Processes

Batch processes play an important role regarded as a typical production mode in industries such as iron and steel, chemical engineering, biopharmaceuticals and semiconductors. Due to the process characteristics of multiple variables, batch-to-batch variation, complexity and multiphase, the mechanism model is difficult to obtain. With development and progress of the data-driven method, soft sensor is becoming a most widely used method in modeling and analyzing for batch processes. Among many soft sensor modeling methods, multivariate statistical analysis techniques, such as PCA and PLS, have become increasingly important on-line tools for online monitoring, fault diagnosis and quality prediction.Batch processes have the characteristics of strong nonlinear and variable correlation. Therefore, using the nonlinear PLS algorithm is regarded to be the appropriate choice for batch processes modeling. Moreover, the neural network is more suitable for the inner relationship of PLS model. Extreme learn machine is a single hidden layer feed forward neural networks, which tends to provide superior generalization performance and extremely fast learning speed by randomly choosing the input weights and analytically determining the output weights. Combining ELM and PLS, the new modeling method has the advantages of both PLS and ELM, which reduces the dimension, handles the nonlinear and shows more rapidly training speed. On the basis of ELM, online sequential ELM exhibits merit of updating the model to cope with batch-to-batch variation and unknown disturbance. Thus, OS-ELM can be used to update the inner model of PLS while weighting the loading matrix for the outer model updating. Combining the OS-ELM and recursive PLS, a nonlinear recursive PLS algorithm, referred as OS-ELM-RPLS algorithm, is proposed which shows superior performance.For batch processes, the final quality of the product at the end of the batch takes an important role in our interests. Due to the model-plant mismatch and batch-to-batch variation, however, the quality at the end of the batch may deviate from the desired quality leading to bad product. Hence, iterative control has been proposed to solve this issue, which exploits the repetitive nature of batch processes and uses process knowledge obtained from previous batch to revise the operating strategy of the current batch so that the output trajectory converges asymptotically to the desired trajectory. In previous literatures, the researchers generally had focus on different models, thus, the model is the key issue for control performance. Based on the proposed OS-ELM-RPLS model which has superior model precision, the batch-to-batch iterative learning control strategy shows better control performance and achieves desire target more quickly.