Enhanced Multi-Scale Convolutional Neural Network Soft-Sensing Method

Abstract

The invention discloses an enhanced multi-scale convolutional neural network soft-sensing method. The method performs data enhancement by reconstructing coarse-grained features, and uses the convolutional neural network to construct a regression model. The process variable data of adjacent time periods are spliced ​​to obtain a data matrix, which is used as a new input of the sample to replace the original fine-grained features, so as to solve the problem of unbalanced sampling of process variables and inaccurate correspondence between input and output of samples; Expansion, data enhancement of the original data set by quantitatively generating samples to solve the problem of local missing process data; use one-dimensional convolution kernel and two-dimensional convolution kernel to extract features of coarse-grained features at the same time, considering the dynamics of variables The time difference between sex and each other solves the delay problem between process variables; the data is reduced to one dimension through the alternating operation of convolution and pooling, and a prediction model is established through a multi-layer perceptron to realize the target quality variable. Real-time soft measurement.

Description

technical field [0001] The invention belongs to the field of industrial process soft measurement, in particular to an enhanced multi-scale convolutional neural network soft measurement method. Background technique [0002] In the era of industrial big data, data-driven soft sensors have become an important means of guiding industrial production and control. However, in actual industry, different data acquisition methods are different. Most of the process variables in production can be directly measured by the distributed control system (DCS), while some quality variables need to be analyzed by chemical means or professional instruments to obtain value. Meanwhile, the process variable data and the quality variable data are an input value and an output value of one sample, respectively. Due to its high measurement cost, the quantity of quality variable data of industrial products is far less than that of process variable data. Therefore, the unbalanced sampling of different...

AI Technical Summary

Problems solved by technology

[0004] Based on the above three problems, fine-grained features (process variable data at a certain moment) cannot provide effective process information for soft sensor modeling, so good soft sensor results cannot be obtained.

At the same time, when the absolute quantity of quality variable data is considered to be small, and local missing due to some reasons, it will further adversely affect the model

Images