Research And Application On The Method Of Soft Sensor Based On Multi-model Fusion

With a minimum cost to produce the most high quality product is the purpose of all industrial production. The difficulty of achieving this goal is that it needs real-time analysis and strictly control the quality of products or the important process variables closely related with the quality of products. Along with the rapid development of computer technology, the informationization of process industry continues to improve, soft measurement techniques emerged. It is able to real-time estimate the key production parameters which is affecting the quality of products. Then, it can provide reference for monitoring the production process, so as to improve the production efficiency, ensure the quality of products.Just as many application cases, the use of a single model remains widespread. But it also exposed two problems. The single model is hard to fully describe complex systems'global properties. Its prediction accuracy and robustness are unsatisfactory. According to different work attributes and demand of actual industrial system, this paper provides three kinds of multi-modeling soft sensor method based on multiple model thought.First of all, a soft sensor modeling method based on multi-modeling dynamic Gauss-Markov estimation fusion is proposed. It separately uses the static model RBF network, the dynamic model OELM and OLS-SVM modeling, and then fuses the estimated values by dynamic Gauss-Markov estimation. The accuracy of this method is higher than any sub-model. It is able to track the dynamic characteristics of time-varying systems.Secondly, this paper provides a new soft-sensor modeling method based on mixed kernel function and Sparse LS-SVM. The mixed kernel is constituted by Polynomial kernel and RBF kernel. It forms multi-model structure in the model internal. It can be both the global fitting ability and local fitting ability of least squares support vector machine. An algorithm called vector base learning has been used to improve the sparseness of the LS-SVM.Finally, this paper provides a new soft-sensing method based on ensemble purning to predict the polymerization conversion of styrene butadiene rubber. We first construct multiple LS-SVMs as weak learners in a bagging ensemble, and then use AdaBoost.RT to pure bagging ensemble. The final output of the pruned ensemble is constructed as weighted combination of the outputs from each selected weak learner. This method has higher precision of prediction and overcomes the problem that original ensemble algorithm requires much memory for storage and predicts slow. And to some extent, it improves the sparseness and robustness issue of LS-SVM.These three soft sensor modeling methods have been simulated. The first two kinds of modeling methods are applicable to the system whose operation is more complicated. The first method can adapt to the emergence of new set-point in the sample inadequate circumstances. Under the condition of the enough samples, the second method has many advantages. The structure of Model is simple. Its online computation is less. It is easy to engineering realization. The third method is more suitable for stable industrial system which requires higher prediction accuracy.