CN-121435155-B - Industrial product quality prediction method, device and medium

CN121435155BCN 121435155 BCN121435155 BCN 121435155BCN-121435155-B

Abstract

The invention provides an industrial product quality prediction method, device and medium, which comprise the steps of obtaining multivariable time sequence data of an industrial process, dividing the multivariable time sequence data into a training set and a testing set, obtaining enhanced frequency domain features corresponding to the training set through discrete Fourier transform, performing topology maintenance and low-dimensional projection processing on the enhanced frequency domain features through a complex frequency domain self-organizing map SOM network to obtain SOM projection frequency domain features, designing a time-frequency attention fusion mechanism comprising the time domain features and the SOM projection frequency domain features, capturing a global-local time mode based on the time-frequency attention fusion mechanism and a time-frequency convolution network, constructing a time-frequency attention fusion network model, and inputting the testing set into the time-frequency attention fusion network model to obtain a predicted value of industrial product quality. The method, the device and the medium can solve the problem that the accuracy of industrial product quality prediction is not high because the existing industrial product quality prediction method cannot adaptively fuse time domain and frequency domain characteristics.

Inventors

HOU HAILIANG
ZHANG YING
OU CHEN
CHEN JIE
LI SHILING
PAN ZHUOFU
LI JUNJIE
LI YANYI

Assignees

湖南工商大学

Dates

Publication Date: 20260508
Application Date: 20251229

Claims (9)

1. A method for predicting quality of an industrial product, the method comprising: Acquiring multi-variable time sequence data of an industrial process, and dividing the multi-variable time sequence data into training set data and test set data; obtaining the enhanced frequency domain characteristics corresponding to the training set data through discrete Fourier transform according to the training set data; Performing topology maintenance and low-dimensional projection processing on the enhanced frequency domain features through a complex frequency domain self-organizing map (SOM) network to obtain SOM projection frequency domain features; Designing a time-frequency attention fusion mechanism comprising time domain features and SOM projection frequency domain features, capturing a global-local time mode based on the time-frequency attention fusion mechanism and a time-frequency convolution network, and constructing a time-frequency attention fusion network model; Inputting the test set data into the time-frequency attention fusion network model to obtain a predicted value of industrial product quality; the design comprises a time-frequency attention fusion mechanism of time domain features and SOM projection frequency domain features, and captures a global-local time mode based on the time-frequency attention fusion mechanism and a time convolution network to construct a time-frequency attention fusion network model, and the method specifically comprises the following steps: Taking the SOM projection frequency domain features as an input feature matrix, and performing linear projection transformation on the input feature matrix to generate a query matrix, a key matrix and a value matrix; remolding the query matrix, the key matrix and the value matrix into a multi-attention head form, and respectively and uniformly dividing the query matrix, the key matrix and the value matrix into a plurality of submatrices according to characteristic dimensions to obtain a query submatrix, a key submatrix and a value submatrix corresponding to each attention head; Calculating the scaling dot product attention of each attention head according to the query submatrix, the key submatrix and the value submatrix corresponding to each attention head and the dimension of each attention head; Splicing the zoom dot product attention of each attention head, and generating a final attention characteristic through an output projection layer; Sequentially carrying out time sequence convolution processing and global average pooling processing on the attention features, and mapping the pooled attention features to a target output space through a multi-layer fully-connected prediction network so as to construct the time-frequency attention fusion network model; The time sequence convolution processing is used for capturing local time sequence dependency relations in the sequence, and the global average pooling processing is used for aggregating characteristic information in a time sequence dimension into characteristic vectors with fixed lengths.
2. The method according to claim 1, wherein the obtaining, according to the training set data, the enhanced frequency domain feature corresponding to the training set data through discrete fourier transform specifically includes: Performing discrete Fourier transform on the training set data to obtain complex characteristics containing amplitude and phase information; And extracting the real part and the imaginary part of the complex feature to carry out real-number processing, and carrying out amplitude spectrum normalization to obtain the enhanced frequency domain feature.
3. The method according to claim 2, wherein said performing a discrete fourier transform on said training set data results in complex features comprising amplitude and phase information, comprising in particular: Performing discrete fourier transform on the training set data by the following formula: ; Wherein, the Representing the first The number of variables that can be used, Representing the length of the variable(s), Representing the discrete fourier transform result of the signal, The frequency component is represented by a frequency component, A time-series signal representing the input is provided, A complex basis function expressed as an euler equation.
4. The method according to claim 1, wherein the topology preserving and low-dimensional projection processing is performed on the enhanced frequency domain features through a complex frequency domain self-organizing map SOM network to obtain SOM projection frequency domain features, specifically including: Defining a complex distance measurement function based on Euclidean distance based on the enhanced frequency domain characteristics, and constructing a complex frequency domain SOM network; And obtaining an optimal matching unit BMU weight vector of the complex frequency domain SOM network through a competition learning mechanism so as to realize topology maintenance and low-dimensional projection processing on the enhanced frequency domain features and obtain the SOM projection frequency domain features.
5. The method of claim 4, wherein the obtaining, by a contention learning mechanism, a best matching unit BMU weight vector of the complex frequency domain SOM network to implement topology preserving and low-dimensional projection processing on the enhanced frequency domain feature, to obtain the SOM projection frequency domain feature specifically includes: Designing competition learning rules of a complex frequency domain SOM network; and based on the competition learning rule and the enhanced frequency domain feature, pre-training the complex frequency domain SOM network to obtain the BMU weight vector, and forming the SOM projection frequency domain feature.
6. The method of claim 1, wherein the time-frequency attention fusion network model comprises a SOM feature projection layer, a time-frequency attention fusion layer, a time-sequence convolution layer and a full-connection prediction layer.
7. An industrial product quality prediction apparatus, comprising: the acquisition module is used for acquiring multi-variable time sequence data of the industrial process and dividing the multi-variable time sequence data into training set data and test set data; the obtaining module is connected with the obtaining module and is used for obtaining the enhanced frequency domain characteristics corresponding to the training set data through discrete Fourier transform according to the training set data; The processing obtaining module is connected with the obtaining module and is used for carrying out topology maintenance and low-dimensional projection processing on the enhanced frequency domain features through a complex frequency domain self-organizing map (SOM) network to obtain SOM projection frequency domain features; The design construction module is connected with the processing obtaining module and is used for designing a time-frequency attention fusion mechanism containing time domain features and SOM projection frequency domain features, capturing a global-local time mode based on the time-frequency attention fusion mechanism and a time-frequency convolution network and constructing a time-frequency attention fusion network model; The input obtaining module is connected with the design construction module and used for inputting the test set data into the time-frequency attention fusion network model to obtain a predicted value of industrial product quality; Further, the design building module specifically includes: The transformation unit is used for taking the SOM projection frequency domain features as an input feature matrix, and performing linear projection transformation on the input feature matrix to generate a query matrix, a key matrix and a value matrix; The remodelling and dividing unit is used for remodelling the query matrix, the key matrix and the value matrix into a multi-attention head form, and respectively and uniformly dividing the query matrix, the key matrix and the value matrix into a plurality of submatrices according to characteristic dimensions to obtain a query submatrix, a key submatrix and a value submatrix corresponding to each attention head; the calculation unit is used for calculating the scaling dot product attention of each attention head according to the query submatrix, the key submatrix and the value submatrix corresponding to each attention head and the dimension of each attention head; The splicing generation unit is used for splicing the zoom dot product attention of each attention head and generating a final attention characteristic through an output projection layer; The processing mapping unit is used for sequentially carrying out time sequence convolution processing and global average pooling processing on the attention features, and mapping the pooled attention features to a target output space through a multi-layer fully-connected prediction network so as to realize the construction of the time-frequency attention fusion network model; The time sequence convolution processing is used for capturing local time sequence dependency relations in the sequence, and the global average pooling processing is used for aggregating characteristic information in a time sequence dimension into characteristic vectors with fixed lengths.
8. An industrial product quality prediction 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 product quality prediction method of any one of claims 1-6.
9. 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 the industrial product quality prediction method according to any of claims 1-6.

Description

Industrial product quality prediction method, device and medium Technical Field The present invention relates to the field of quality prediction technologies, and in particular, to a method, an apparatus, and a medium for predicting quality of an industrial product. Background In modern industrial production processes, product quality control is a key link to ensure production efficiency and product competitiveness. The industrial process quality prediction technology can early warn quality abnormality in advance, optimize the production process, reduce the rate of unqualified products and have important significance for improving the economic benefit of enterprises by carrying out real-time monitoring and analysis on key parameters in the production process. In particular, in the process industries of petrifaction, steel, pharmacy and the like, the product quality index is often difficult to directly measure on line, and indirect estimation and prediction are needed to be carried out by depending on a soft measurement modeling technology. The soft measurement technology is used as one of core technologies in the process industry, and the on-line estimation of the key quality index is realized by establishing a mathematical model between the easily-measured auxiliary variable and the difficult-to-measure dominant variable. However, industrial process data generally has complex characteristics of multiple variables, strong coupling, nonlinearity, time-varying and the like, contains rich time-frequency domain information, and the existing industrial product quality prediction method cannot adaptively fuse time domain and frequency domain characteristics, so that the industrial product quality prediction accuracy is not high. Disclosure of Invention The technical problem to be solved by the invention is to provide an industrial product quality prediction method, device and medium for solving the problem that the accuracy of industrial product quality prediction is not high because the existing industrial product quality prediction method cannot adaptively fuse time domain and frequency domain characteristics. In a first aspect, the present invention provides a method for predicting quality of an industrial product, comprising: Acquiring multi-variable time sequence data of an industrial process, and dividing the multi-variable time sequence data into training set data and test set data; obtaining the enhanced frequency domain characteristics corresponding to the training set data through discrete Fourier transform according to the training set data; Performing topology maintenance and low-dimensional projection processing on the enhanced frequency domain features through a complex frequency domain self-organizing map (SOM) network to obtain SOM projection frequency domain features; Designing a time-frequency attention fusion mechanism comprising time domain features and SOM projection frequency domain features, capturing a global-local time mode based on the time-frequency attention fusion mechanism and a time-frequency convolution network, and constructing a time-frequency attention fusion network model; And inputting the test set data into the time-frequency attention fusion network model to obtain a predicted value of the quality of the industrial product. Further, the obtaining, according to the training set data, the enhanced frequency domain feature corresponding to the training set data through discrete fourier transform specifically includes: Performing discrete Fourier transform on the training set data to obtain complex characteristics containing amplitude and phase information; And extracting the real part and the imaginary part of the complex feature to carry out real-number processing, and carrying out amplitude spectrum normalization to obtain the enhanced frequency domain feature. Further, the discrete fourier transform is performed on the training set data to obtain complex features including amplitude and phase information, which specifically includes: Performing discrete fourier transform on the training set data by the following formula: ; Wherein, the Representing the firstThe number of variables that can be used,Representing the length of the variable(s),Representing the discrete fourier transform result of the signal,The frequency component is represented by a frequency component,A time-series signal representing the input is provided,A complex basis function expressed as an euler equation. Further, the topology maintenance and low-dimensional projection processing are performed on the enhanced frequency domain features through a complex frequency domain self-organizing map SOM network to obtain SOM projection frequency domain features, which specifically includes: Defining a complex distance measurement function based on Euclidean distance based on the enhanced frequency domain characteristics, and constructing the complex frequency domain SOM network; And obtaining an optimal matching