Research On Problems Of Soft Sensor Based On Information Entropy Theory

Production of process industry is very complex, existing great changes in operating condition, strong nonlinearity, and big disturbance, etc. It put forward higher requirement to detection technology. The appearance of advanced detection instrument has solved a lot of problems in the production process, but it is not enough. For the past few years, establishing a mathematical model for measurement of unpredictable variables indirectly through soft-sensing technology overcome the problem that instruments can’t directly measure unpredictable variables. The paper combines soft-sensing technology and information entropy theory to do system research aimed at problems in the soft sensor procedure.(1) For the problem of variables selection in the implementation process of using soft sensor technology, a method of distributed mutual information is proposed. This method gives joint probability density of random variable in the sample. According to mutual information definition, we can obtain values of mutual information between dominant variable and auxiliary variables. Setting up a threshold is used to judge the relevance between dominant variable and auxiliary variables. Then the auxiliary variables whose mutual information is beyond the threshold value are selected for the next step, achieving selection of relevant variables or discard weak relevant variables. Finally, a soft sensor model is built by support vector machine algorithm with the selected secondary variables. The simulation results show that the method of variable selection for soft sensor based on distributed mutual information is simple, of high reliability and practical, while improving the estimation accuracy of the model.(2) Although a large number of information of auxiliary variables can be collected by DCS systems, but the measurement of dominant variables must be done by professional workers, so there is a time lag in obtaining information of sample, leading to the decrease of model accuracy. Aiming at this issue, the paper proposed a method to estimate missing data based on probabilistic principal component analysis (PPCA). The method is to estimate missing data directly by minimizing the whitened value of the sample. Simulation results show that this method can effectively estimate the missing part of sample set, so that it can increases the number of samples and improves the accuracy of model training.(3) When using a single model to predict, it is generally difficult to express complicated production process, often resulting in low accuracy of prediction and poor performance of generalization. This paper presents a multi-model fusion method based on probability weight and self-adaptive fuzzy gauss kernel clustering. The method determines cluster centers according to dispersion of the samples in a high dimensional space. The weight of every sub-model is given by Bayesian posterior method. The method can not only overcome the limitation of single-model forecast and but also improve traditional multiple-model fusion methods, obtaining higher prediction accuracy.