Semi-supervised hybrid model-based polypropylene melting index prediction method

Abstract

The invention discloses a semi-supervised hybrid model-based polypropylene melting index prediction method. The method comprises the following steps of: separately processing an auxiliary variable anda melting index, and considering a dependency relationship between the melting index and the auxiliary variable in an explicit manner, so as to establish a randomized mathematic model; and mining information of labeled samples and unlabeled samples, and carrying out automatic model parameter learning and model selection by utilizing an expectation maximization algorithm and a Bayesian informationcriterion. The method is capable of online providing predicted values of melting indexes in real time and assessing credibility of the melting indexes. By applying the method, the correctness of themodel parameter learning and model selection can be improved, and all the parameters do not need to be manually set, so that the melting index prediction precision can be effectively improved, and technical support and guarantee are provided for product quality improvement, cost reduction, process monitor and decision making.

Description

technical field [0001] The invention belongs to the field of process system soft sensor modeling and application, in particular to a polypropylene melt index prediction method based on a semi-supervised mixing model. Background technique [0002] Polypropylene resin has been widely used in many industrial fields of the national economy due to its advantages such as small specific gravity, non-toxic, odorless, easy processing, high impact strength, good twist resistance and good electrical insulation. Melt index is an important index to measure the quality of polypropylene products. It is usually measured by laboratory analysis and the cycle is 2-4 hours. Such a large measurement lag will significantly reduce the dynamics and stability of the closed-loop control system, let alone card edge control, resulting in strong fluctuations in the polypropylene production process and a lot of product waste, which not only increases the production cost of the enterprise, but also aggrav...

AI Technical Summary

Problems solved by technology

First, due to market demand, there are many operating conditions in the polypropylene production process, and the products usually have multiple grades (up to dozens), which makes the polypropylene production process data show strong nonlinearity and non-Gaussian characteristics.

It is difficult for a single global model (such as partial least squares model, neural network model, support vector machine model, etc.) to provide satisfactory prediction accuracy in all grades

Second, current melt index prediction methods are usually supervised learning methods, which rely only on labeled samples with melt index assay values

Unfortunately, most current melt index soft sensor models do not have this capability

Images