Soft-Sensor Modeling And Itsapplication Based On Data Mining

As a result of some key production variables is difficulty to measure, soft_sensor technology has become a hot point in the field of current process measurement as an effective detected method. Most of the actual industrial process mechanism is so difficult to clear that modeling technology has had more rapid progress which bases on the data driven. Due to the incomplete, noise, and fuzzy, relevant, nonlinear features of real data, modeling effect is not ideal.In order to improve generalization performance of the model and the model robust (noise interference resistance ability), this paper applies data mining technology to develop the research of the following several aspects.1、As the characteristics of sample data, there is information interference after clustering. Many algorithms are not good tracking mutation data also, eventually leading to bad estimation accuracy of soft measurement model. An improved Linear Discriminant Analysis (LDA) algorithm is proposed in this paper so as to solve the problem. It redefines the between-class scatter and in-class scatter. The feature vectors are obtained by combining boundary analysis with LDA between the categories. The original sample data are transformed in terms of the feature vectors, and sub-models based on SVM are respectively constructed by transformed data. And then the compositional parameters for sub-models are designed according to the sum of the effectual characteristic values of every feature vector. The simulation results show that the composition model can reduce the information interference among the different data categories and improve the inferential accuracy of the model.2、Methods of soft sensor for multi-model based on traditional clustering algorithms can significantly improve the estimation precision and generalization performance for the model, but its subclasses may be overlap, and the edged classes and abnormal sample points can not be effectively dealt with so that the modeling result is bad. To deal with these programs, this paper presents a method based on feature extraction of the weighted kernel Fisher criterion, which the edge classes and the abnormal sample points are brought closer to other normal sample points by mapping so as to improve the clustering accuracy. At last, establish a combined model with extracted data based on support vector machine. The method is applied to a soft sensor modeling for the quality index in a Bisphenol A production process. The simulation results show that the algorithm is effective improving the data classification result, and improving estimation precision of the soft measurement model.3、Clustering algorithm which is based on the traditional Euclidean distance estimation only reflect the similarity of each category in space. The algorithm neglects the similarity of their geometry characteristic, which leads to insufficient clustering results as well as poor precision of the soft sensor model. In the paper, a clustering algorithm based on minimum energy is promoted. A sample clustering problem is translated into looking for a minimum energy ring in the algorithm. The annealing algorithm is used to search for the minimum energy ring which goes through each sample point in order to clustering. SVM isadopted to establish a regression sub-model for each subclass, and finally, the soft sensor composite model is obtained. This method is applied to a soft sensor modeling for the quality performance in a Bisphenol A production process. The effectiveness of the method is confirmed by simulation results.