CN-121352633-B - Soft measurement method, device and medium for industrial multi-rate acquisition

CN121352633BCN 121352633 BCN121352633 BCN 121352633BCN-121352633-B

Abstract

The invention provides a soft measurement method, a device and a medium for industrial multi-rate acquisition, which comprise the steps of acquiring historical data in an industrial process, dividing the historical data into a plurality of sampling rate data, extracting data features of different sampling rates respectively, obtaining each scale feature in a regular manner, obtaining each bidirectional cross attention matrix according to each scale feature and a query source, fusing the bidirectional cross attention matrix to obtain bidirectional cross attention features, dynamically calibrating the bidirectional cross attention features based on the historical features to obtain a soft measurement model, and inputting data to be predicted in the industrial process into the soft measurement model to obtain a quality variable prediction result in the industrial process. The method, the device and the medium can solve the problems of low accuracy of industrial process quality variable prediction caused by insufficient feature extraction, insufficient cross-scale association mining and limited utilization of historical target trend information in the conventional soft measurement method of the industrial system when processing multi-sampling rate data.

Inventors

OU CHEN
HUANG XIAO
CHEN JIE
LI SHILING
HOU HAILIANG
PAN ZHUOFU
Quan Qinyi
XU YALAN

Assignees

湖南工商大学

Dates

Publication Date: 20260508
Application Date: 20251218

Claims (10)

1. An industrial multi-rate acquisition oriented soft measurement method, the method comprising: collecting historical data in an industrial process, and dividing the historical data into data in a training set according to sampling frequency to obtain a plurality of sampling rate data, wherein the historical data comprises process variables and quality variables; The method comprises the steps of respectively extracting independent characteristics of sampling rate data with different sampling frequencies through a plurality of independent branches in a parallel multi-scale convolutional neural network module, and carrying out dimension normalization on output after the independent characteristic extraction to obtain scale characteristics of each sampling rate data; Taking the scale features with the highest and lowest sampling frequencies as query sources, obtaining a bidirectional cross attention matrix of each scale feature through a bidirectional cross attention module according to each scale feature and the query sources, and fusing all the bidirectional cross attention matrices according to an average fusion principle to obtain bidirectional cross attention features; Converting a history target value sequence into history features by a history target value guiding attention module, and dynamically calibrating the bidirectional cross attention features based on the history features to obtain a soft measurement model, wherein the history features are consistent with the dimensions of the bidirectional cross attention features; Inputting data to be predicted in the industrial process to the soft measurement model to obtain a quality variable prediction result in the industrial process.
2. The method of claim 1, wherein each of the independent branches comprises, in order, a first 1D convolutional layer, a first batch of normalization layers, a first ReLU activation function, a maximum pooling layer, a second 1D convolutional layer, a second batch of normalization layers, a second ReLU activation function, and an adaptive average pooling layer.
3. The method according to claim 1, wherein the scale features with the highest and lowest sampling frequencies are used as query sources, and a bi-directional cross-attention matrix of each scale feature is obtained through a bi-directional cross-attention module according to each scale feature and the query sources, specifically comprising: sequencing all scale features of the sampling rate data according to the sampling frequency from high to low; calculating a highest frequency query value according to the scale characteristic of the highest sampling frequency; Calculating a lowest frequency query value according to the scale characteristics of the lowest sampling frequency; calculating a key and a value of each scale feature according to each scale feature; And calculating a bidirectional cross attention matrix of each scale feature by adopting a multi-head attention mechanism according to the highest frequency query value, the lowest frequency query value and the key and the value of each scale feature.
4. The method of claim 3, wherein the bi-directional cross attention moment array comprises a high frequency attention matrix and a low frequency attention matrix, and the fusing of all the bi-directional cross attention matrices according to an average fusing principle to obtain bi-directional cross attention features specifically comprises: Fusing the high-frequency attention matrix and the low-frequency attention matrix of each scale feature to obtain a fusion matrix of each scale feature; And splicing each fusion matrix, and performing dimension compression on the spliced matrix to obtain the bidirectional cross attention characteristic.
5. The method according to claim 4, wherein the fusing the high frequency attention matrix and the low frequency attention matrix of each scale feature to obtain a fused matrix of each scale feature specifically comprises: The high frequency attention matrix and the low frequency attention matrix for each of the scale features are fused by the following formula: ; Wherein, the Is a fusion matrix of the r-th scale feature, A high frequency attention matrix for the r-th scale feature, The attention matrix is focused on for the low frequency of the r-th scale feature.
6. The method of claim 1, wherein before the converting the sequence of historical target values into historical features by the historical target value directing attention module, the method further comprises: and constructing a historical target value sequence at the current moment according to samples of the current moment corresponding to the previous preset number of time steps.
7. The method according to claim 6, wherein the directing attention module converts the sequence of historical target values into historical features and dynamically calibrates the bi-directional cross-attention features based on the historical features to obtain a soft measurement model, and specifically comprises: compressing the dimension of the bidirectional cross attention feature through linear embedding, and converting the history target value sequence at the current moment into the history feature consistent with the dimension of the compressed bidirectional cross attention feature through linear embedding; According to the historical characteristics, calculating query values corresponding to the historical characteristics; Calculating keys and values of the compressed bidirectional cross attention feature according to the compressed bidirectional cross attention feature; calculating attention weight guided by the history feature according to the query value, the key and the value; And dynamically calibrating the compressed bidirectional cross attention characteristic according to the attention weight guided by the history characteristic to obtain a characteristic vector guided by the history target value sequence so as to obtain the soft measurement model.
8. An industrial multi-rate acquisition oriented soft measurement device, comprising: The acquisition and division module is used for acquiring historical data in an industrial process and dividing the historical data corresponding to data in a training set according to sampling frequency to obtain a plurality of sampling rate data, wherein the historical data comprises process variables and quality variables; The extraction normalization module is connected with the acquisition dividing module and is used for respectively extracting independent characteristics of the sampling rate data with different sampling frequencies through a plurality of independent branches in the parallel multi-scale convolution neural network module, and carrying out dimension normalization on the output after the independent characteristic extraction to obtain the scale characteristics of each sampling rate data; The acquisition fusion module is connected with the extraction regular module and is used for acquiring a bidirectional cross attention matrix of each scale feature according to each scale feature and the query source by taking the scale feature with the highest and lowest sampling frequency as the query source, and fusing all the bidirectional cross attention matrices according to an average fusion principle to acquire the bidirectional cross attention feature; The conversion calibration module is connected with the acquisition fusion module and used for converting the historical target value sequence into historical characteristics through the historical target value guiding attention module, and dynamically calibrating the bidirectional cross attention characteristics based on the historical characteristics to obtain a soft measurement model, wherein the historical characteristics are consistent with the dimensions of the bidirectional cross attention characteristics; and the input obtaining module is connected with the conversion calibration module and used for inputting data to be predicted in the industrial process to the soft measurement model to obtain a quality variable prediction result in the industrial process.
9. An industrial multi-rate acquisition oriented soft measurement device comprising a memory and a processor, the memory having stored therein a computer program, the processor being arranged to run the computer program to implement the industrial multi-rate acquisition oriented soft measurement method according to any one of claims 1-7.
10. A computer-readable storage medium, characterized in that it has stored thereon a computer program which, when executed by a processor, implements the industrial multi-rate acquisition oriented soft measurement method according to any of claims 1-7.

Description

Soft measurement method, device and medium for industrial multi-rate acquisition Technical Field The invention relates to the technical field of quality prediction, in particular to a soft measurement method, a soft measurement device and a soft measurement medium for industrial multi-rate acquisition. Background In the modern industrial process, real-time accurate measurement and monitoring of key product quality is a key for effectively tracking production efficiency and maintaining safe production standards. However, the quality index of the key product is often due to the problems of high cost, complex maintenance and the like of the professional measuring instrument due to high temperature and high pressure in the industrial environment and strong electromagnetic interference, and the real-time direct measurement is difficult to realize. For this reason, soft sensor modeling techniques are the core means of indirectly estimating key quality indicators through easily measured process variables. In order to fully utilize the potential of industrial big data and meet the changing demands of industrial production, the soft measurement method based on deep learning has been widely applied in industrial quality prediction in recent years. However, the sampling frequencies of the various sensors in industrial systems vary widely, resulting in process data exhibiting typical multi-sample rate characteristics. This data characteristic presents significant challenges for industrial quality prediction and soft measurement modeling. The existing quality prediction method mainly relies on single sampling rate data or performs data alignment processing through interpolation, filling and other means, so that the problems of insufficient extraction of multi-sampling rate data characteristics, insufficient cross-scale association mining and the like are caused. In addition, most methods do not fully utilize historical target trend information in the prediction process, resulting in insufficient modeling capabilities for time dependence and long-term trends in industrial processes, thereby affecting prediction accuracy and stability. In summary, the soft measurement method of the existing industrial system has insufficient feature extraction, insufficient cross-scale association mining and limited utilization of historical target trend information when processing multi-sampling rate data, so that the prediction accuracy of industrial process quality variables is not high. Disclosure of Invention The invention aims to solve the technical problems of the prior art, and provides a soft measurement method, a device and a medium for industrial multi-rate acquisition, which are used for solving the problems of insufficient characteristic extraction, insufficient cross-scale associated mining and limited utilization of historical target trend information, resulting in low industrial process quality variable prediction accuracy in the soft measurement method of the existing industrial system when processing multi-sampling rate data. In a first aspect, the present invention provides a soft measurement method for industrial multi-rate acquisition, comprising: collecting historical data in an industrial process, and dividing the historical data into data in a training set according to sampling frequency to obtain a plurality of sampling rate data, wherein the historical data comprises process variables and quality variables; The method comprises the steps of respectively extracting independent characteristics of sampling rate data with different sampling frequencies through a plurality of independent branches in a parallel multi-scale convolutional neural network module, and carrying out dimension normalization on output after the independent characteristic extraction to obtain scale characteristics of each sampling rate data; Taking the scale features with the highest and lowest sampling frequencies as query sources, obtaining a bidirectional cross attention matrix of each scale feature through a bidirectional cross attention module according to each scale feature and the query sources, and fusing all the bidirectional cross attention matrices according to an average fusion principle to obtain bidirectional cross attention features; Converting a history target value sequence into history features by a history target value guiding attention module, and dynamically calibrating the bidirectional cross attention features based on the history features to obtain a soft measurement model, wherein the history features are consistent with the dimensions of the bidirectional cross attention features; Inputting data to be predicted in the industrial process to the soft measurement model to obtain a quality variable prediction result in the industrial process. Further, each independent branch sequentially comprises a first 1D convolution layer, a first normalization layer, a first ReLU activation function, a maximum pooling layer, a second 1