Online Soft Sensor Research Based On Outlier Detection

As we all know, during the petrol-chemistry, oil refining, papermaking and other chemical processes, there are many variables that cannot be directly measured with instruments. As a new technology, soft sensor technology was put forward to solve the above mentioned problem since its being introduced. By establishing the mathematical relationship between the secondary variables that can be easily measured and the primary variables that cannot be directly measured, it successfully achieves the solution to the problem. As far as we know, it is necessary to eliminate some useless outliers, because the outliers resulted from the instrument’s failure and sensors affected by the chemical environment will have an impact on the prediction accuracy. However, the problem is that the replacement of raw materials or the change of process may lead to the data migration, and we often consider the data mistakenly as outliers which may be eliminated. In order to solve this problem, the concept of classification of outliers is proposed in this paper, with the dry point of aviation kerosene oil and the polypropylene melt index as objects in the simulation study. The main content of this paper are as follows:Firstly, to the problem of outliers existing in the operating process, this paper introduces the method of outlier detection based on support vector data description(SVDD). The method will get the center and radius of the super ball by training the data when offline, and the two parameters obtained will be the basis of deciding whether the data is outliers when online.Secondly, to determine whether the outliers we get is authentic, the concept of outliers classification is put forward. Based on the Bayesian classification principle, varieties of outliers include the impulse outliers, the short-step type outliers and the step outliers, among which the impulse outliers and the short-step outliers are required to be eliminated while the step outliers reveals the transformation of chemical engineering.Thirdly, aiming at the outliers required to be adjusted, the method of time series correction is adopted in this paper. Because all the secondary variables are multivariable, so it is necessary to analyze and decide which variable or variables are in need of correction by adopting the method of contributive rate analysis before we do the correction. So by adjusting the data in need of correction, we can avoid the wrong correction which may result in some bigger errors.Fourthly, when it has been decided that the chemical process is a process change, we are required to renew the model. So in this paper, the model renewal method based on the discount moving window recursive PLS is employed. From the simulation results, it can be realized if the prediction and tracking effect of the model can reach our requirements.Finally, the final part of this paper summarizes the research working on soft measurement technology of outliers and puts forward some questions in need of further research.