Some Applications Of Machine Learning Algorithms For The Soft-Sensor Modeling

Soft sensor technology is one of the most important research directions in the area of the process monitoring and process control. Because the actual industrial process showed the characteristics of process variables with strong correlation, non-linearity and time-varying operation condition, the methods of soft sensor based on multiple linear regression model can not meet the need of the soft sensor modeling for complex industrial process. The development of machine learning provides a new direction for the actual industrial process modeling. In order to improve the generalization performance of the model, this paper proposed a number of new methods of modeling for soft sensor based on machine learning. The main research is described as follows.1. Using integrated learning, this paper proposes an integration algorithm for fuzzy support vector regression based on improved Adaboost.RT to improve the generalization performance of the single support vector machine (SVM). The sample weight function in the Adaboost.RT algorithm is taken as the fuzzy membership function of fuzzy support vector machine (FSVM). And a series of weak FSVM regression machines are generated through the iteration of Adaboost.RT. And then an upper-layer SVM is used to optimize and integrate the output results of the lower-layer FSVMs. The proposed algorithm is used for the soft-sensor model to estimate the content of phenol at the outlet of the dehydration tower. Simulation results show that the integration algorithm can improve the effect of regression estimation and the generalization performance of the soft-sensor model.2. Local learning is the new trend in the field of machine learning. To deal with complex industrial process variables with strong correlation, non-linearity and time-varying characteristics of operation condition, a new soft sensor modeling method is proposed based on local Kernel Partial Least Squares (KPLS) feature extraction and on-line Least Squares Support Vector Machines (LSSVM). Firstly some similar samples are found out with the current test sample from the whole sample space, and features of the subspace are extracted, and then a local soft sensor model based on LSSVM is built to estimate the current output. Experimental results show that this method can effectively realize feature extraction., and have a better generalization ability than off-line LSSVM based on global feature extraction with KPLS as well as global LSSVM without feature extraction.3. Probability kernel learning machine is a new direction in the area of machine learning. A multi-model modeling method based on affinity propagation clustering and Gaussian processes is presented because a single model usually suffers from bad accuracy.. This method can cluster samples according to different work modes by affinity propagation clustering with a new similarity. It trains sub-cluster by Gaussian processes, and obtains the final result by the"Switch"way. The proposed algorithm is used for the soft-sensor model to estimate the content of acetone at the outlet of the reaction vessel. The experimental results indicate that the proposed algorithm has a superior accuracy and certain application value.