Least squares support vector machine soft measurement modeling method based on distribution estimation local optimization

Abstract

The invention discloses a least squares support vector machine soft measurement modeling method based on distribution estimation local optimization, which uses a test set to establish a local target set for least squares support vector machine model parameter optimization, so that the optimization process is more fitted to the test set, the over-fitting phenomenon caused by the traditional use ofthe whole training set as the optimization target is effectively avoided, and strong adaptability to the data with the higher degree of distribution variation is achieved. The distribution estimationalgorithm is used to parameter optimization to effectively improve the convergence accuracy, and a soft measurement model constructed by the method can accurately predict and control the key quality variables.

Description

technical field [0001] The invention belongs to the field of industrial process prediction and control, and relates to a least square support vector machine soft sensor modeling method based on local optimization of distribution estimation. Background technique [0002] Soft sensor modeling technology means that in the actual industrial production process, some process variables and quality variables are difficult to use sensors for direct measurement or the cost of measurement is too high. Therefore, people often use some process variables that are easier to measure and build mathematical models. The methods to estimate those difficult-to-measure process or quality variables can control product quality well and improve production efficiency. [0003] Among the commonly used soft sensor mathematical models, the least squares support vector machine model (hereinafter referred to as LSSVM) is widely used in the soft sensor of various industrial process variables due to its adv...

AI Technical Summary

Problems solved by technology

[0005] However, this method still has certain defects, and it is mainly divided into two parts. On the one hand, if only minimizing the deviation of the training sample set is the optimization goal, the generalization ability of the algorithm will be deteriorated. This situation mainly occurs in the training set. When the data distribution of the test set is different from that of the test set, that is, the model parameters optimized by using the training set will have a better fitting effect on the training set, while the test set may have a large difference; The requirements of the algorithm are also high, that is, the optimization algorithm should have better convergence and optimization effects, but many commonly used intelligent optimization algorithms such as (genetic algorithm) still do not have these characteristics

Images