CN-122022598-A - Soft measurement modeling method, device and medium in industrial process

CN122022598ACN 122022598 ACN122022598 ACN 122022598ACN-122022598-A

Abstract

The invention provides a soft measurement modeling method, a device and a medium in an industrial process, wherein the method comprises the steps of collecting historical data in the industrial process, adopting a forward backtracking strategy to obtain a multi-sampling-rate process variable sequence and a quality variable sequence, carrying out parallel feature extraction on each sampling-rate process variable sequence to obtain a multi-sampling-rate feature sequence, obtaining local features and global features from the multi-sampling-rate feature sequence, fusing the local features and the global features to obtain feature sequences after feature fusion, taking the feature sequences after feature fusion of the quality variable sequence and the feature sequences as query vector and key value pairs respectively, and obtaining a quality prediction result based on the query vector and the key value pairs. The method, the device and the medium can solve the problems of low quality prediction precision in industrial soft measurement caused by complex multi-sampling rate data fusion and insufficient key information focusing in the existing industrial soft measurement modeling under multi-sampling rate acquisition.

Inventors

HOU HAILIANG
LI CONG
OU CHEN
CHEN JIE
LI SHILING
PAN ZHUOFU
ZHU SHENGYU

Assignees

湖南工商大学

Dates

Publication Date: 20260512
Application Date: 20260413

Claims (10)

1. A method of modeling soft measurements in an industrial process, the method comprising: Acquiring historical data in an industrial process, and adopting a forward backtracking strategy to obtain sample data according to the historical data, wherein the sample data comprises a multi-sampling-rate process variable sequence and a quality variable sequence; Parallel feature extraction is carried out on each sampling rate process variable sequence in the multi-sampling rate process variable sequence by adopting a grouping multi-head attention mechanism, so as to obtain a multi-sampling rate feature sequence; Obtaining local features and global features of the relation between all sampling rate feature sequences from the multi-sampling rate feature sequences through a pre-built local-global dual-branch feature fusion module, and fusing the local features and the global features of the relation between all sampling rate feature sequences to obtain feature sequences after feature fusion; And respectively taking the quality variable sequence and the feature sequence after feature fusion as a query vector and a key value pair, and obtaining a quality prediction result through a cross attention mechanism guided by the quality variable sequence based on the query vector and the key value pair.
2. The method according to claim 1, wherein the obtaining sample data using a forward backtracking strategy based on the history data specifically comprises: backtracking a preset number of historical quality variable values forward from the effective quality variable values in the historical data; and obtaining the sample data from the historical data according to the moments corresponding to the effective quality variable value and the historical quality variable values of the preset quantity.
3. The method according to claim 1, wherein the parallel feature extraction is performed on each of the multiple sample rate process variable sequences using a packet multi-head attention mechanism to obtain a multiple sample rate feature sequence, and specifically comprises: Creating an attention header group for each sample rate process variable sequence, wherein the attention header group comprises a plurality of attention headers; And inputting each sampling rate process variable sequence into a corresponding attention head group, and carrying out parallel feature extraction on each sampling rate process variable sequence through each attention head group to obtain the multi-sampling rate feature sequence.
4. A method according to claim 3, wherein the parallel feature extraction is performed on each sample rate process variable sequence by each attention header group to obtain the multi-sample rate feature sequence, and the method specifically comprises: generating a query matrix, a key matrix and a value matrix corresponding to the attention head groups according to the sampling rate process variable sequences corresponding to the attention head groups for each attention head group; Calculating the output of each attention head in the attention head group according to the query matrix, the key matrix and the value matrix corresponding to the attention head group; Splicing the outputs of all the attention heads in the attention head group, and obtaining the output of the attention head group through a linear layer; And taking the output of all the attention head groups as the multi-sampling rate characteristic sequence.
5. The method according to claim 1, wherein the obtaining, by the pre-constructed local-global dual-branch feature fusion module, local features and global features of relationships between all sample rate feature sequences from the multi-sample rate feature sequences, and fusing the local features and global features of relationships between all sample rate feature sequences, to obtain feature fused feature sequences specifically includes: Splicing all the sampling rate characteristic sequences in the multi-sampling rate characteristic sequences to obtain spliced characteristic sequences; copying the spliced characteristic sequences to obtain a first copied characteristic sequence and a second copied characteristic sequence; processing the first replication feature sequence through local branches to obtain local features of relations among all sampling rate feature sequences; processing the second replication feature sequence through a global branch to obtain global features of the relation among all the sampling rate feature sequences; Performing feature compression on the spliced feature sequence to obtain the spliced feature after feature compression, inputting the spliced feature after feature compression into a preset dynamic weight generator, and obtaining dynamic weights corresponding to the spliced feature after feature compression through the dynamic weight generator; And based on the dynamic weight, fusing the local features and the global features to obtain the feature sequence after feature fusion.
6. The method according to claim 1, wherein the cross-attention mechanism guided by the quality variable sequence based on the query vector and key value pairs yields quality prediction results, specifically comprising: obtaining a characteristic sequence guided by a quality variable sequence according to the query vector and the key value pair; optimizing the feature sequence guided by the quality variable sequence to obtain the feature sequence guided by the optimized quality variable sequence; And carrying out quality prediction through the last time step feature in the feature sequence guided by the optimized quality variable sequence to obtain the quality prediction result.
7. The method of claim 6, wherein the key-value pairs comprise a key matrix and a value matrix; The step of obtaining the feature sequence guided by the quality variable sequence according to the query vector and the key value pair specifically comprises the following steps: According to the query vector and the key matrix, calculating the attention score between each quality time step in the quality variable sequence and each characteristic time step in the characteristic sequence after characteristic fusion; normalizing the attention score between each quality time step in the quality variable sequence and each characteristic time step in the characteristic sequence after characteristic fusion to obtain the attention score between each quality time step and each characteristic time step after normalization; And calculating the feature sequence guided by the quality variable sequence according to the attention score between each quality time step and each feature time step after normalization processing and the value matrix.
8. A soft measurement modeling apparatus in an industrial process, comprising: The acquisition module is used for acquiring historical data in an industrial process and acquiring sample data by adopting a forward backtracking strategy according to the historical data, wherein the sample data comprises a multi-sampling-rate process variable sequence and a quality variable sequence; The characteristic extraction module is connected with the acquisition module and is used for carrying out parallel characteristic extraction on each sampling rate process variable sequence in the multi-sampling rate process variable sequence by adopting a grouping multi-head attention mechanism to obtain a multi-sampling rate characteristic sequence; The acquisition fusion module is connected with the feature extraction module and is used for acquiring local features and global features of the relation between all sampling rate feature sequences from the multi-sampling rate feature sequences through a pre-constructed local-global double-branch feature fusion module, and fusing the local features and the global features of the relation between all sampling rate feature sequences to obtain feature sequences after feature fusion; And the obtaining module is connected with the acquisition fusion module and is used for respectively taking the quality variable sequence and the feature sequence after feature fusion as a query vector and a key value pair, and obtaining a quality prediction result through a cross attention mechanism guided by the quality variable sequence based on the query vector and the key value pair.
9. A soft measurement modeling apparatus in an industrial process, comprising a memory having a computer program stored therein and a processor configured to run the computer program to implement the soft measurement modeling method in an industrial process as claimed in any of claims 1-7.
10. A computer readable storage medium, characterized in that the computer readable storage medium has stored thereon a computer program which, when executed by a processor, implements a soft measurement modeling method in an industrial process according to any of the claims 1-7.

Description

Soft measurement modeling method, device and medium in industrial process Technical Field The present invention relates to the field of soft measurement technologies, and in particular, to a soft measurement modeling method, apparatus, and medium in an industrial process. Background In modern process industry, real-time monitoring and prediction of product quality is critical to ensuring production efficiency and product quality. The quality variable is difficult to acquire due to the problems of expensive special measuring hardware, offline detection lag and the like, and the soft measuring technology can build a model through the associated auxiliary variable easy to acquire, so that the quality variable can be estimated online at low cost. However, typical multi-sample rate data acquisition problems exist in industrial processes in that different sensors and sensing devices have different sampling frequencies, forming complex multi-sample rate time series data structures. Soft measurement modeling at multiple sample rate acquisition has mainly the following challenges: 1. The multi-sampling rate data fusion is complex, namely, the time scale and the characteristic properties are different among different sampling rates, so that the information fusion is necessary to be carried out on the data with different sampling rates, but the existing method does not realize the separate extraction of the characteristics with different sampling rates and lacks an effective multi-sampling rate data fusion mechanism. 2. The key information focusing is insufficient, the difference of contribution degrees of different time steps to quality prediction is obvious under a multi-sampling-rate acquisition environment, but a key information screening mechanism guided by quality variables is not established in the traditional method, and the core information utilization efficiency is low. In summary, the existing industrial soft measurement modeling under multi-sampling rate acquisition has complex multi-sampling rate data fusion and insufficient key information focusing, so that the quality prediction precision in the industrial soft measurement is not high. Disclosure of Invention The invention aims to solve the technical problems of the prior art, and provides a soft measurement modeling method, a device and a medium in an industrial process, which are used for solving the problems of complex multi-sampling rate data fusion and insufficient key information focusing in the existing industrial soft measurement modeling under multi-sampling rate acquisition, so that the quality prediction precision in the industrial soft measurement is not high. In a first aspect, the present invention provides a method of modeling soft measurements in an industrial process, comprising: Acquiring historical data in an industrial process, and adopting a forward backtracking strategy to obtain sample data according to the historical data, wherein the sample data comprises a multi-sampling-rate process variable sequence and a quality variable sequence; Parallel feature extraction is carried out on each sampling rate process variable sequence in the multi-sampling rate process variable sequence by adopting a grouping multi-head attention mechanism, so as to obtain a multi-sampling rate feature sequence; Obtaining local features and global features of the relation between all sampling rate feature sequences from the multi-sampling rate feature sequences through a pre-built local-global dual-branch feature fusion module, and fusing the local features and the global features of the relation between all sampling rate feature sequences to obtain feature sequences after feature fusion; And respectively taking the quality variable sequence and the feature sequence after feature fusion as a query vector and a key value pair, and obtaining a quality prediction result through a cross attention mechanism guided by the quality variable sequence based on the query vector and the key value pair. Further, the obtaining sample data by adopting a forward backtracking strategy according to the historical data specifically includes: backtracking a preset number of historical quality variable values forward from the effective quality variable values in the historical data; and obtaining the sample data from the historical data according to the moments corresponding to the effective quality variable value and the historical quality variable values of the preset quantity. Further, the method for extracting parallel characteristics of each sampling rate process variable sequence in the multi-sampling rate process variable sequence by adopting a packet multi-head attention mechanism to obtain the multi-sampling rate characteristic sequence specifically comprises the following steps: Creating an attention header group for each sample rate process variable sequence, wherein the attention header group comprises a plurality of attention headers; And inputting each sampling r