Application Of Maximum Entropy In The Data Processing Of Soft Sensor

In recent years, the soft sensor technology plays an important role in the measurementwhere the process parameters are difficult to detect directly. In this paper, it briefly introducessoft sensor technology and the maximum entropy method, and the main innovation pointsinclude the following three aspects:(1) In order to solve the problem of variables selection for soft sensor, a method ofcombining maximum entropy and mutual information is proposed. The mutual informationindirectly reflects the correlation including linear correlation and nonlinear correlationbetween the predicted variable and the secondary variables. Then a threshold value isobtained by t-test approach as a criterion to judge the correlation of variables. When themutual information between the predicted variable and a secondary variable is less than thethreshold value, the secondary variable is not selected. The secondary variables are selectedby combination of the presented method and cross test.(2) Because of complex internal relations and redundant information among secondaryvariables in a soft sensor, it brings negative factors to secondary variables selection andimproving model accuracy and etc. The paper uses a method of fast independent componentanalysis (Fast-ICA) based on negative entropy biggest. This method can effectively eliminatethe phenomenon of information redundancy among secondary variables, and get independentcomponents. According to the internal relation between the predicted variable and eachcomponent by the mutual information, the components which are more associated with thepredicted variable are selected to establish model. The method is applied to a phenol (BF)concentration estimation in an industrial production process.(3) Although a large number of real times operating data can been collected by a DCSfrom a process industry production site, only a small amount of these data can be used assecondary variables sample. Most predicted variables which reflect the quality indicators needto be detected with manual analysis or on-line quality instrument for longer time, which leadsto difficulties of collecting the training sample set for a soft sensor model, and less use of datacollected by the DCS system, as well as affecting the accuracy of machine learning. In thispaper, a maximum entropy method is used to estimate the joint probability distribution of thevariables for a soft sensor and a Bayesian maximum posteriori methods combined clusteringanalysis is applied to estimating the samples which lack of manual analysis values.(4) Possible support vectors of unlabeled samples based on the criteria and markingmethod is used to train SVM regression model, and the estimation accuracy andgeneralization ability of SVM regression model is improved with the increase of supportvectors of unlabeled samples.