Soft measurement method for 4-CBA content based on improved AdaBoost (Adaptive Boosting) algorithm

Abstract

The invention discloses a soft measurement method for the 4-CBA content based on an improved AdaBoost (Adaptive Boosting) algorithm, which is characterized in that a double-threshold sample weight updating method is adopted, BP neural network training is selected to act as a weak learner, and a group of acquired weak learners are combined by adopting the AdaBoost algorithm so as to obtain a strong leaner. According to the method, relevant measurable variables of a PTA oxidation process are selected to act as input of a model, the 4-CBA content act as output of the model, and historical acquisition data is selected to act as a training sample. The improved AdaBoost algorithm provided by the invention can reduce influences imposed on the weak learner by samples with a large error, and the precision of prediction for the 4-CBA content is improved.

Description

technical field [0001] The invention relates to a soft-sensing method for 4-CBA content, in particular to a soft-sensing method for 4-CBA content based on an improved AdaBoost algorithm, and belongs to the field of chemical engineering. Background technique [0002] The AdaBoost (Adaptive Boosting) algorithm is an algorithm proposed by Freund and Schapire to promote weak learners to strong learners. This algorithm can not only improve any combination of weak learners, but also improve the prediction accuracy of strong learners. It has been widely used at present. In two types of problems, multi-classification problems and regression problems. The main idea of ​​the AdaBoost algorithm is to assign sample weights to each training sample, and update the sample weights in the process of training the weak learner. The change of the sample weights is determined by the training results of the currently trained weak learner, and the training error is large. The weight of the sample...

AI Technical Summary

Problems solved by technology

[0007] 4-CBA content cannot be measured online by conventional methods, but is obtained through laboratory analysis, and the offline analysis takes a long time, often lagging behind for several hours; at the same time, the cost of experimental analysis and testing is high, and the sampling interval is long and the sampling frequency is small , cannot meet the requirements for real-time monitoring of 4-CBA content. For example, a factory has a sampling period of 8 hours for 4-CBA content, and samples are fixed at 0 o'clock, 8 o'clock and 16 o'clock every day, so there are only three lags at most in one day. Analysis value of 4-CBA content of h

Images