An industrial product quality prediction method, device and medium

Abstract

The application provides an industrial product quality prediction method, device and medium, the method comprises the following steps: obtaining multivariate time series data of an industrial process, and dividing the multivariate time series data into a training set and a test set; obtaining enhanced frequency domain features corresponding to the training set through discrete Fourier transform; performing topological preservation and low-dimensional projection processing on the enhanced frequency domain features through a complex frequency domain self-organizing map (SOM) network to obtain SOM projected frequency domain features; designing a time-frequency attention fusion mechanism containing time domain features and SOM projected frequency domain features, and constructing a time-frequency attention fusion network model based on the time-frequency attention fusion mechanism and a time convolution network to capture global-local time patterns; and inputting the test set into the time-frequency attention fusion network model to obtain a predicted value of the industrial product quality. The method, device and medium can solve the problem that existing industrial product quality prediction methods cannot adaptively fuse time domain and frequency domain features, resulting in low accuracy of industrial product quality prediction.

Description

Technical Field

[0001] This invention relates to the field of quality prediction technology, and in particular to a method, apparatus and medium for predicting the quality of industrial products. Background Technology

[0002] In modern industrial production, product quality control is a crucial link in ensuring production efficiency and product competitiveness. Industrial process quality prediction technology, through real-time monitoring and analysis of key parameters in the production process, can provide early warnings of quality anomalies, optimize production processes, and reduce the rate of defective products, which is of great significance for improving the economic benefits of enterprises. Especially in process industries such as petrochemicals, steel, and pharmaceuticals, product quality indicators are often difficult to measure directly online, requiring indirect estimation and prediction using soft sensor modeling techniques.

[0003] Soft measurement technology, as one of the core technologies in the process industry, enables online estimation of key quality indicators by establishing mathematical models between easily measurable auxiliary variables and difficult-to-measurable dominant variables. However, industrial process data typically possesses complex characteristics such as multivariability, strong coupling, nonlinearity, and time-varying nature, containing rich time-frequency domain information. Existing industrial product quality prediction methods cannot adaptively integrate time-domain and frequency-domain features, resulting in low accuracy in industrial product quality prediction. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to address the above-mentioned shortcomings of the prior art by providing an industrial product quality prediction method, device and medium, so as to solve the problem that the existing industrial product quality prediction methods cannot adaptively integrate time domain and frequency domain features, resulting in low accuracy of industrial product quality prediction.

[0005] In a first aspect, the present invention provides a method for predicting the quality of industrial products, comprising:

[0006] Acquire multivariate time-series data of an industrial process, and divide the multivariate time-series data into training set data and test set data;

[0007] Based on the training set data, the enhanced frequency domain features corresponding to the training set data are obtained through discrete Fourier transform;

[0008] The enhanced frequency domain features are topology preserved and low-dimensionally projected by a complex frequency domain self-organizing map (SOM) network to obtain SOM projected frequency domain features.

[0009] Design a time-frequency attention fusion mechanism that incorporates temporal features and the frequency domain features projected by the SOM, and construct a time-frequency attention fusion network model based on the time-frequency attention fusion mechanism and a temporal convolutional network to capture global-local temporal patterns;

[0010] The test set data is input into the time-frequency attention fusion network model to obtain the predicted value of industrial product quality.

[0011] Furthermore, the step of obtaining the enhanced frequency domain features corresponding to the training set data through discrete Fourier transform based on the training set data specifically includes:

[0012] Perform a discrete Fourier transform on the training set data to obtain complex features containing amplitude and phase information;

[0013] The real and imaginary parts of the complex feature are extracted, converted to real numbers, and the amplitude spectrum is normalized to obtain the enhanced frequency domain feature.

[0014] Furthermore, the step of performing a discrete Fourier transform on the training set data to obtain complex features containing amplitude and phase information specifically includes:

[0015] The training set data is subjected to a discrete Fourier transform using the following formula:

[0016] ;

[0017] in, Indicates the first One variable, Indicates the length of the variable. This represents the result of the discrete Fourier transform of the signal. Represents frequency components, This represents the input time series signal. Let be the complex basis functions expressed in Euler's formula.

[0018] Furthermore, the step of performing topology preservation and low-dimensional projection processing on the enhanced frequency domain features through a complex frequency domain self-organizing map (SOM) network to obtain SOM projected frequency domain features specifically includes:

[0019] Based on the enhanced frequency domain features, a complex distance metric function based on Euclidean distance is defined, and the complex frequency domain SOM network is constructed.

[0020] The optimal matching unit (BMU) weight vector of the complex frequency domain SOM network is obtained through a competitive learning mechanism, so as to perform topology preservation and low-dimensional projection processing on the enhanced frequency domain features to obtain the SOM projected frequency domain features.

[0021] Furthermore, the step of obtaining the optimal matching unit (BMU) weight vector of the complex frequency domain SOM network through a competitive learning mechanism to perform topology preservation and low-dimensional projection processing on the enhanced frequency domain features to obtain the SOM projected frequency domain features specifically includes:

[0022] Design competitive learning rules for complex frequency domain SOM networks;

[0023] Based on the competitive learning rules and enhanced frequency domain features, the complex frequency domain SOM network is pre-trained to obtain the BMU weight vector and form the SOM projected frequency domain features.

[0024] Furthermore, the design includes a time-frequency attention fusion mechanism that integrates temporal features with the SOM projected frequency domain features. Based on this time-frequency attention fusion mechanism and a temporal convolutional network to capture global-local temporal patterns, a time-frequency attention fusion network model is constructed, specifically including:

[0025] The SOM projection frequency domain features are used as the input feature matrix, and a linear projection transformation is performed on the input feature matrix to generate a query matrix, a key matrix, and a value matrix.

[0026] The query matrix, key matrix, and value matrix are reshaped into a multi-attention head form, and the query matrix, key matrix, and value matrix are uniformly divided into multiple sub-matrices according to the feature dimension to obtain the query sub-matrix, key matrix, and value matrix corresponding to each attention head;

[0027] Calculate the scaled dot product attention for each attention head based on the query submatrix, key matrix, and value submatrix corresponding to each attention head, as well as the dimension of each attention head;

[0028] The scaled dot product attention of each attention head is concatenated and the final attention feature is generated through the output projection layer.

[0029] The attention features are sequentially processed by temporal convolution and global average pooling, and the pooled attention features are mapped to the target output space through a multi-layer fully connected prediction network to construct the time-frequency attention fusion network model.

[0030] Furthermore, the time-frequency attention fusion network model includes: a SOM feature projection layer, a time-frequency attention fusion layer, a temporal convolutional layer, and a fully connected prediction layer.

[0031] Secondly, the present invention provides an industrial product quality prediction device, comprising:

[0032] The acquisition module is used to acquire multivariate time-series data of industrial processes and divide the multivariate time-series data into training set data and test set data.

[0033] The module is connected to the acquisition module and is used to obtain the enhanced frequency domain features corresponding to the training set data through discrete Fourier transform based on the training set data.

[0034] A processing module, connected to the obtained module, is used to perform topology preservation and low-dimensional projection processing on the enhanced frequency domain features through a complex frequency domain self-organizing map (SOM) network to obtain SOM projected frequency domain features.

[0035] The design and construction module is connected to the processing module and is used to design a time-frequency attention fusion mechanism that includes temporal features and the SOM projection frequency domain features. Based on the time-frequency attention fusion mechanism and the temporal convolutional network to capture global-local temporal patterns, a time-frequency attention fusion network model is constructed.

[0036] The input module, connected to the design and construction module, is used to input the test set data into the time-frequency attention fusion network model to obtain the predicted value of industrial product quality.

[0037] Thirdly, the present invention provides an industrial product quality prediction device, including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to implement the industrial product quality prediction method described in the first aspect above.

[0038] Fourthly, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the industrial product quality prediction method described in the first aspect.

[0039] The present invention provides an industrial product quality prediction method, apparatus, and medium. First, multivariate time-series data of an industrial process is acquired and divided into training and testing sets. Then, based on the training set data, enhanced frequency domain features are obtained through Discrete Fourier Transform. Next, the enhanced frequency domain features are processed with topology preservation and low-dimensional projection using a complex frequency domain self-organizing map (SOM) network to obtain SOM-projected frequency domain features. A time-frequency attention fusion mechanism incorporating both time-domain features and the SOM-projected frequency domain features is designed, and a time-frequency attention fusion network model is constructed based on this mechanism and a temporal convolutional network to capture global-local time patterns. Finally, the testing set data is input into the time-frequency attention fusion network model to obtain predicted values ​​for industrial product quality. This invention converts time-series data into complex frequency domain features using Discrete Fourier Transform (DFT), and utilizes a complex frequency domain self-organizing map network to achieve topology preservation and efficient dimensionality reduction of high-dimensional complex features while retaining amplitude and phase information in the signal. Furthermore, through a designed time-frequency attention fusion network, it adaptively fuses time-domain and frequency-domain features, thereby significantly improving the representational ability of multi-feature combinations and exhibiting stronger perception and capture capabilities for multi-scale characteristics and periodic patterns in industrial signals. In practical applications, it demonstrates higher prediction accuracy, possesses good practical value and promising prospects for widespread application, and solves the problem that existing industrial product quality prediction methods cannot adaptively fuse time-domain and frequency-domain features, resulting in low accuracy in industrial product quality prediction. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in the embodiments of this drawing or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this drawing. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0041] Figure 1 This is a flowchart of an industrial product quality prediction method according to Embodiment 1 of the present invention;

[0042] Figure 2 This is a flowchart of another industrial product quality prediction method according to Embodiment 1 of the present invention;

[0043] Figure 3 This is a diagram of the complex frequency domain self-organizing time-frequency attention fusion network structure according to an embodiment of the present invention;

[0044] Figure 4 This is a comparison chart of predicted and actual values ​​in an embodiment of the present invention;

[0045] Figure 5This is a schematic diagram of the structure of an industrial product quality prediction device according to Embodiment 2 of the present invention;

[0046] Figure 6 This is a schematic diagram of the structure of an industrial product quality prediction device according to Embodiment 3 of the present invention. Detailed Implementation

[0047] To enable those skilled in the art to better understand the technical solution of the present invention, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0048] It is understood that the specific embodiments and accompanying drawings described herein are merely for explaining the invention and are not intended to limit the invention.

[0049] It is understood that, without conflict, the various embodiments and features in the embodiments of the present invention can be combined with each other.

[0050] It is understood that, for ease of description, only the parts related to the present invention are shown in the accompanying drawings, while the parts unrelated to the present invention are not shown in the drawings.

[0051] It is understood that each unit or module involved in the embodiments of the present invention may correspond to only one entity structure, or may be composed of multiple entity structures, or multiple units or modules may be integrated into one entity structure.

[0052] It is understood that, without conflict, the functions and steps marked in the flowcharts and block diagrams of this invention may occur in a different order than that marked in the accompanying drawings.

[0053] It is understood that the flowcharts and block diagrams of this invention illustrate the possible architecture, functions, and operations of systems, apparatuses, devices, and methods according to various embodiments of this invention. Each block in the flowchart or block diagram may represent a unit, module, program segment, or code, containing executable instructions for implementing the specified function. Furthermore, each block or combination of blocks in the block diagram and flowchart can be implemented using a hardware-based system to achieve the specified function, or using a combination of hardware and computer instructions.

[0054] It is understood that the units and modules involved in the embodiments of the present invention can be implemented by software or by hardware. For example, the units and modules can be located in a processor.

[0055] Application Overview

[0056] Industrial process data typically exhibits complex characteristics such as multivariability, strong coupling, nonlinearity, and time-varying nature, containing rich time-frequency domain information. Traditional time-domain analysis methods often struggle to fully extract the potential periodicity and frequency structure within the data. While frequency-domain methods (such as Fourier transform) can effectively extract the frequency components and periodic patterns of signals, they are difficult to directly integrate into the time-series modeling process. In particular, existing soft-sensor modeling methods have significant shortcomings when processing the complex features generated by the Fourier transform. These complex features simultaneously contain amplitude and phase information, exhibiting high dimensionality and complex coupling relationships. Traditional real-domain models struggle to effectively utilize all of this information, limiting the performance of quality prediction models.

[0057] Furthermore, even when time-domain and frequency-domain features are extracted separately, existing fusion strategies mostly involve simple concatenation or weighted averaging, lacking the ability to adaptively learn the interaction between time and frequency. In non-stationary, multimodal, and multi-sampling-rate industrial scenarios, such static fusion methods struggle to capture time-varying spectral structures and cross-frequency-cross-time domain dependencies. In recent years, attention mechanisms have shown advantages in capturing long-term dependencies and adaptive modal fusion, but systematic methods that apply attention mechanisms to adaptive time-frequency fusion of complex frequency domain amplitude and phase features and combine them with dimensionality reduction methods that can preserve topological structure are still lacking.

[0058] To address the aforementioned technical problems, this application provides an industrial product quality prediction method, apparatus, and medium. It converts time-series data into complex frequency domain features using Discrete Fourier Transform (DFT), and utilizes a complex frequency domain self-organizing map network to achieve topology preservation and efficient dimensionality reduction of high-dimensional complex features, while retaining amplitude and phase information in the signal. Furthermore, through a designed time-frequency attention fusion network, it adaptively fuses time-domain and frequency-domain features, thereby significantly improving the representational ability of multi-feature combinations. It also exhibits stronger perception and capture capabilities for multi-scale characteristics and periodic patterns in industrial signals, demonstrating higher prediction accuracy in practical applications. This method possesses good practical value and promising prospects for wider application, at least solving the problem that existing industrial product quality prediction methods cannot adaptively fuse time-domain and frequency-domain features, resulting in low accuracy in industrial product quality prediction.

[0059] After introducing the basic principles of this application, various non-limiting embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0060] Example 1:

[0061] This embodiment provides a method for predicting the quality of industrial products, such as... Figure 1 As shown, the method includes:

[0062] Step S101: Obtain multivariate time series data of the industrial process, and divide the multivariate time series data into training set data and test set data.

[0063] It should be noted that multivariate time series data for industrial processes include time series data of industrial process variables and quality variables.

[0064] Specifically, time series data of industrial processes and quality variables are collected, normalized, and divided into training and testing sets.

[0065] Step S102: Based on the training set data, obtain the enhanced frequency domain features corresponding to the training set data through discrete Fourier transform.

[0066] In one optional embodiment, obtaining the enhanced frequency domain features corresponding to the training set data through discrete Fourier transform based on the training set data specifically includes:

[0067] Perform a discrete Fourier transform on the training set data to obtain complex features containing amplitude and phase information;

[0068] The real and imaginary parts of the complex feature are extracted, converted to real numbers, and the amplitude spectrum is normalized to obtain the enhanced frequency domain feature.

[0069] Specifically, the training set data is subjected to Discrete Fourier Transform (DFT) to obtain complex features containing amplitude and phase information. The real and imaginary parts of the complex features are extracted and converted to real numbers, and the amplitude spectrum is normalized to form an enhanced frequency domain feature representation.

[0070] Specifically, the training set data is subjected to a discrete Fourier transform using the following formula:

[0071] ;

[0072] The number of variables is , Indicates the first One variable, Indicates the length of the variable. This represents the result of the discrete Fourier transform of the signal. Represents frequency components, This represents the input time series signal. Let be the complex basis functions expressed in Euler's formula.

[0073] The Discrete Fourier Transform includes key frequency selection and projection, for the complex spectrum obtained by the Discrete Fourier Transform. The energy of each frequency component is calculated by squared its amplitude:

[0074] ;

[0075] in, Indicates energy value; Indicates taking the real part; This indicates taking the imaginary part.

[0076] All frequency components are sorted in descending order of energy value to obtain a sorted frequency index sequence. The top K frequency components with the highest energy are selected, as these components contain the main information of the signal. Then, complex feature construction and normalization are performed, concatenating the complex frequency features of each variable into a vector.

[0077] ;

[0078] in, Let represent the complex frequency eigenvector of the p-th variable.

[0079] The complex frequency features of all the variables are combined to obtain a global complex frequency feature vector:

[0080] ;

[0081] in, This represents the global complex frequency eigenvector.

[0082] Specifically, the complex frequency feature vector is converted to real numbers, its real and imaginary parts are extracted, and the amplitude spectrum is normalized to obtain an enhanced frequency domain feature representation.

[0083] ;

[0084] in, This represents the normalized complex frequency eigenvector (i.e., the enhanced frequency domain feature).

[0085] Step S103: Perform topology preservation and low-dimensional projection processing on the enhanced frequency domain features through a complex frequency domain SOM (Self-Organizing Map) network to obtain SOM projected frequency domain features.

[0086] In an optional embodiment, the step of performing topology preservation and low-dimensional projection processing on the enhanced frequency domain features through a complex frequency domain self-organizing map (SOM) network to obtain SOM projected frequency domain features specifically includes:

[0087] Based on the enhanced frequency domain features, a complex distance metric function based on Euclidean distance is defined, and the complex frequency domain SOM network is constructed.

[0088] The weight vector of the BMU (Best Matching Unit) of the complex frequency domain SOM network is obtained through a competitive learning mechanism, so as to perform topology preservation and low-dimensional projection processing on the enhanced frequency domain features to obtain the SOM projected frequency domain features.

[0089] Specifically, based on the enhanced frequency domain features, a complex distance metric function based on Euclidean distance is defined, a self-organizing map (SOM) network is constructed, and the weight vector of the best matching unit (BMU) is obtained through a competitive learning mechanism to achieve topology preservation and low-dimensional projection of the frequency domain features, thereby obtaining a compact feature representation (i.e., SOM projected frequency domain features) that can be used for time-frequency fusion modeling.

[0090] In an optional embodiment, the step of obtaining the best matching unit (BMU) weight vector of the complex frequency domain SOM network through a competitive learning mechanism to perform topology preservation and low-dimensional projection processing on the enhanced frequency domain features to obtain the SOM projected frequency domain features specifically includes:

[0091] Design competitive learning rules for complex frequency domain SOM networks;

[0092] Based on the competitive learning rules and enhanced frequency domain features, the complex frequency domain SOM network is pre-trained to obtain the BMU weight vector and form the SOM projected frequency domain features.

[0093] Specifically, the training process of the complex frequency domain self-organizing map network includes the following steps:

[0094] Step 1: Define a complex distance metric function based on Euclidean distance, wherein the complex distance metric function is:

[0095] ;

[0096] in, This indicates the row index of the node in the SOM grid; This represents the column index of a node in the SOM network; Represents the m-th complex component; This represents the weight vector associated with the m-th complex component; Indicates the weights controlling the phase difference; It represents the distance metric in the complex space; P is the total number of variables, K is the number of frequency components selected for each variable, and PK represents the total number of complex components; Indicates position The current weight vector of the neuron.

[0097] Step 2: Design the competitive learning rules for the complex frequency domain self-organizing map network. The competitive learning rules are as follows:

[0098] ;

[0099] in, Used to identify the location of each neuron node in the SOM grid; This indicates the location of the winning neuron in the SOM grid.

[0100] The weight update rule is as follows:

[0101] ;

[0102] ;

[0103] in, Indicates the learning rate; Let be the domain function, representing the activation characteristics of the self-organizing map network; Indicates the radius of the domain (used to control the radius of the domain). Indicates position The update amount of the neuron weight vector.

[0104] Step 3: Pre-train the complex domain self-organizing map network based on complex features, obtain the weight vector of the winning neuron (i.e. the weight vector of the best matching unit), and form the SOM topological feature representation (i.e., the SOM projection frequency domain feature).

[0105] Step S104: Design a time-frequency attention fusion mechanism that includes temporal features and the frequency domain features of the SOM projection, and construct a time-frequency attention fusion network model based on the time-frequency attention fusion mechanism and the capture of global-local temporal patterns by a temporal convolutional network.

[0106] Specifically, a time-frequency attention fusion mechanism incorporating temporal features and SOM projection frequency domain features is designed. By combining a temporal convolutional network to capture global-local temporal patterns, an end-to-end neural network architecture is constructed, comprising an SOM feature projection layer, a time-frequency attention fusion layer, a temporal convolutional layer, and a fully connected prediction layer, resulting in a time-frequency attention fusion network model based on complex frequency domain self-organization.

[0107] It is worth mentioning that the time-frequency attention fusion network, which is constructed by using a time-frequency attention fusion mechanism that includes time-domain features and SOM projection frequency domain features, and a time-frequency attention fusion network that captures global-local time patterns through a temporal convolutional network, can fully explore and model the joint time-frequency features contained in the data, thereby improving the accuracy and reliability of industrial process quality prediction.

[0108] In an optional embodiment, the design includes a time-frequency attention fusion mechanism that integrates temporal features and the SOM projected frequency domain features. Based on this time-frequency attention fusion mechanism and a temporal convolutional network capturing global-local temporal patterns, a time-frequency attention fusion network model is constructed, specifically including:

[0109] The SOM projection frequency domain features are used as the input feature matrix, and a linear projection transformation is performed on the input feature matrix to generate a query matrix, a key matrix, and a value matrix.

[0110] The query matrix, key matrix, and value matrix are reshaped into a multi-attention head form, and the query matrix, key matrix, and value matrix are uniformly divided into multiple sub-matrices according to the feature dimension to obtain the query sub-matrix, key matrix, and value matrix corresponding to each attention head;

[0111] Calculate the scaled dot product attention for each attention head based on the query submatrix, key matrix, and value submatrix corresponding to each attention head, as well as the dimension of each attention head;

[0112] The scaled dot product attention of each attention head is concatenated and the final attention feature is generated through the output projection layer.

[0113] The attention features are sequentially processed by temporal convolution and global average pooling, and the pooled attention features are mapped to the target output space through a multi-layer fully connected prediction network to construct the time-frequency attention fusion network model.

[0114] Specifically, constructing a time-frequency attention fusion network based on complex frequency domain self-organization includes the following steps:

[0115] Step 1: Use the obtained SOM topological feature representation as the input feature matrix. Perform a linear projection transformation on it to generate the query matrix. Key matrix Sum matrix :

[0116] ;

[0117] in, It is obtained by linear projection transformation and dimension reshaping of the weight vector of BMU; Batch size; For time step; For feature dimensions; , , is a learnable linear projection parameter matrix.

[0118] Step 2: , , The matrix is ​​reshaped into a multi-head form, with each attention head having a dimension of [missing value]. Query matrix The attention head is uniformly divided into h sub-matrices according to its feature dimensions, and the query sub-matrix corresponds to the i-th attention head. Obtained by truncating the corresponding dimension range, where i ranges from 0 to h-1. Key matrix. Sum matrix The same segmentation method is used. Scaled dot product attention is calculated for each attention head:

[0119] ;

[0120] Wherein, scaling factor Used to prevent gradient vanishing; Indicates the first Calculate the scaled dot product attention for each attention head; It is an activation function.

[0121] The outputs of all attention heads are concatenated, and the final attention features are generated through an output projection layer. The output of each attention head is defined as follows:

[0122] ;

[0123] in, Indicates the first Pay attention to the output of the head.

[0124] Then, the images are stitched together and projected:

[0125] ;

[0126] in, To output the projection matrix; Indicates splicing; This represents the final output of multi-head attention.

[0127] Step 3: Perform temporal convolution and global average pooling on the output of the multi-head attention module in sequence. Temporal convolution is used to capture local temporal dependencies in the sequence, and global average pooling is used to aggregate feature information in the temporal dimension into a fixed-length feature vector.

[0128] Step 4: Map the pooled feature vectors to the target output space through a multi-layer fully connected prediction network, realizing a non-linear transformation from high-dimensional feature representation to predicted values, and constructing a time-frequency attention fusion network based on complex frequency domain self-organization:

[0129] ;

[0130] ;

[0131] ;

[0132] in, , , These represent the weight matrices for each layer; , , For the corresponding bias term; This represents the feature vector after pooling; This indicates the final predicted output; Representation layer normalization; It is an activation function; This indicates the output of the first hidden state. Indicates random deactivation regularization; This indicates the output of the second hidden state.

[0133] Step S105: Input the test set data into the time-frequency attention fusion network model to obtain the predicted value of industrial product quality.

[0134] Specifically, the data in the test set are sequentially input into the time-frequency attention fusion network model to obtain the prediction results.

[0135] It should be noted that the topology preservation loss can be used to evaluate the performance of the SOM model:

[0136] ;

[0137] in, This represents the topology preservation loss; This represents a neuron in the SOM model (i.e., it is used to identify the location of each neuron node in the SOM grid). This indicates the location of the winning neuron within the SOM grid; Represents neurons Associated weight vector (i.e., representing position) (the current weight vector of the neuron). Indicates the winning neuron The associated weight vector; Indicates the radius of the domain.

[0138] The mean squared error can be used to evaluate the predictive performance of the model:

[0139] ;

[0140] in, Indicates mean square error; Indicates the actual value; This represents an estimated value.

[0141] The performance of the entire model can be evaluated using a composite loss function:

[0142] ;

[0143] in, Represents the composite loss function; This represents the balance coefficient, controlling the weights of unsupervised and supervised losses, and its range is... .

[0144] It is worth mentioning that the industrial product quality prediction method based on a self-organizing time-frequency attention fusion network in the complex frequency domain provided by this invention addresses the complex time-frequency characteristics caused by the multivariate, strongly coupled, nonlinear, and time-varying nature of industrial process data. This invention extends the traditional SOM to the complex domain, thereby achieving unified topology preservation of amplitude and phase information during dimensionality reduction. At the same time, it further constructs a time-frequency attention fusion network to fully explore and model the joint time-frequency features contained in the data, thereby improving the accuracy and reliability of industrial process quality prediction and meeting the requirements of modern industrial production for high-precision quality control.

[0145] In a specific embodiment, such as Figure 2 As shown, this industrial product quality prediction method may specifically include:

[0146] S1: Normalize the collected time series data of industrial processes and quality variables, and divide them into training set and test set data;

[0147] S2: Perform Discrete Fourier Transform (DFT) on the training set data to obtain complex features containing amplitude and phase information. Extract the real and imaginary parts of the complex features respectively, convert them to real numbers, and normalize the amplitude spectrum to form an enhanced frequency domain feature representation.

[0148] S3: Based on the enhanced frequency domain features, a complex distance metric function based on Euclidean distance is defined to construct a self-organizing map (SOM) network. The weight vector of the best matching unit (BMU) is obtained through a competitive learning mechanism to achieve topology preservation and low-dimensional projection of the frequency domain features, thereby obtaining a compact feature representation (i.e., SOM projected frequency domain features) that can be used for time-frequency fusion modeling.

[0149] S4: Design a time-frequency attention fusion mechanism that includes temporal features and SOM projection frequency domain features. Combine a temporal convolutional network to capture global-local temporal patterns. Construct an end-to-end neural network architecture that includes an SOM feature projection layer, a time-frequency attention fusion layer, a temporal convolutional layer, and a fully connected prediction layer to obtain a complex frequency domain self-organizing time-frequency attention fusion network model.

[0150] Specifically, such as Figure 3 As shown, Figure 3 The diagram shows the structure of a complex frequency domain self-organizing time-frequency attention fusion network. The main process is as follows: the input is first normalized, then sequentially processed through discrete Fourier transform, complex frequency domain self-organizing mapping, and time-frequency attention fusion network to produce the output.

[0151] S5: Input the data from the test set into the prediction model in sequence to obtain the prediction results.

[0152] In another specific embodiment, this invention uses a dataset of bottom product quality from a butanizer process for experimental purposes. The butanizer is a crucial separation device in the petroleum refining process, separating mixed alkanes into light and heavy components through distillation under heating and pressure conditions. In the butanizer process, multivariate time-series data is collected in real-time from the tower using industrial sensors. This data covers key indicators in the butanizer process, including reactor temperature (…). ), system pressure ( ), feed flow rate ( ), reflux ratio ( ), tower top temperature ( ), tower bottom temperature ( ), steam flow rate ( There are a total of 7 indicators, including (e.g., ), which comprehensively reflect the operating status of the butanizer process and provide rich information for subsequent data analysis and model building. The time series data of these 7 indicators correspond to process variables, while the time series data of the bottom product quality (y) corresponds to the quality variable. Multivariate time series data were collected using continuous sampling to form a complete operating dataset. After eliminating dimensional differences through MinMax normalization, the data was divided into training and testing sets using a sliding window method (window size 24). The quality variable value at the current moment was predicted using the process variable data from the past 24 time points. The comparison between the predicted and actual values ​​is shown in the figure below. Figure 4 As shown.

[0153] The industrial product quality prediction method provided in this invention first acquires multivariate time-series data of an industrial process and divides the multivariate time-series data into training set data and test set data. Then, based on the training set data, enhanced frequency domain features corresponding to the training set data are obtained through Discrete Fourier Transform. Next, the enhanced frequency domain features are processed by topology preservation and low-dimensional projection through a complex frequency domain self-organizing map (SOM) network to obtain SOM projected frequency domain features. Then, a time-frequency attention fusion mechanism that includes time domain features and the SOM projected frequency domain features is designed, and a time-frequency attention fusion network model is constructed based on the time-frequency attention fusion mechanism and a temporal convolutional network to capture global-local time patterns. Finally, the test set data is input into the time-frequency attention fusion network model to obtain the predicted value of industrial product quality. This invention converts time-series data into complex frequency domain features using Discrete Fourier Transform (DFT), and utilizes a complex frequency domain self-organizing map network to achieve topology preservation and efficient dimensionality reduction of high-dimensional complex features while retaining amplitude and phase information in the signal. Furthermore, through a designed time-frequency attention fusion network, it adaptively fuses time-domain and frequency-domain features, thereby significantly improving the representational ability of multi-feature combinations and exhibiting stronger perception and capture capabilities for multi-scale characteristics and periodic patterns in industrial signals. In practical applications, it demonstrates higher prediction accuracy, possesses good practical value and promising prospects for widespread application, and solves the problem that existing industrial product quality prediction methods cannot adaptively fuse time-domain and frequency-domain features, resulting in low accuracy in industrial product quality prediction.

[0154] Example 2:

[0155] like Figure 5 As shown, this embodiment provides an industrial product quality prediction device for executing the above-described industrial product quality prediction method, including:

[0156] The acquisition module 11 is used to acquire multivariate time series data of industrial processes and divide the multivariate time series data into training set data and test set data;

[0157] The module 12 is connected to the acquisition module 11 and is used to obtain the enhanced frequency domain features corresponding to the training set data through discrete Fourier transform based on the training set data.

[0158] The processing module 13 is connected to the obtained module 12 and is used to perform topology preservation and low-dimensional projection processing on the enhanced frequency domain features through the complex frequency domain self-organizing map (SOM) network to obtain SOM projected frequency domain features.

[0159] Design module 14, connected to the processing module 13, is used to design a time-frequency attention fusion mechanism that includes temporal features and the SOM projection frequency domain features, and to construct a time-frequency attention fusion network model based on the time-frequency attention fusion mechanism and the temporal convolutional network to capture global-local temporal patterns.

[0160] The input module 15, connected to the design and construction module 14, is used to input the test set data into the time-frequency attention fusion network model to obtain the predicted value of industrial product quality.

[0161] Furthermore, the obtaining module 12 specifically includes:

[0162] The unit is used to perform a discrete Fourier transform on the training set data to obtain complex features containing amplitude and phase information;

[0163] The processing normalization unit is used to extract the real and imaginary parts of the complex feature, perform realization processing, and perform amplitude spectrum normalization to obtain the enhanced frequency domain feature.

[0164] Furthermore, the obtaining unit specifically includes:

[0165] The training set data is subjected to a discrete Fourier transform using the following formula:

[0166] ;

[0167] in, Indicates the first One variable, Indicates the length of the variable. This represents the result of the discrete Fourier transform of the signal. Represents frequency components, This represents the input time series signal. Let be the complex basis functions expressed in Euler's formula.

[0168] Furthermore, the processing module 13 specifically includes:

[0169] Define a building unit for defining a complex distance metric function based on Euclidean distance based on the enhanced frequency domain features, and construct the complex frequency domain SOM network;

[0170] The obtained unit is used to obtain the best matching unit (BMU) weight vector of the complex frequency domain SOM network through a competitive learning mechanism, so as to perform topology preservation and low-dimensional projection processing on the enhanced frequency domain features to obtain the SOM projected frequency domain features.

[0171] Furthermore, the acquisition unit specifically includes:

[0172] Design unit, used to design competitive learning rules for complex frequency domain SOM networks;

[0173] The training unit is used to pre-train the complex frequency domain SOM network based on the competitive learning rules and enhanced frequency domain features to obtain the BMU weight vector and form the SOM projected frequency domain features.

[0174] Furthermore, the design construction module 14 specifically includes:

[0175] As a transformation unit, it is used to take the SOM projected frequency domain features as the input feature matrix and perform linear projection transformation on the input feature matrix to generate a query matrix, a key matrix and a value matrix;

[0176] The reshaping segmentation unit is used to reshape the query matrix, key matrix, and value matrix into a multi-attention head form, and to uniformly divide the query matrix, key matrix, and value matrix into multiple sub-matrices according to the feature dimensions, so as to obtain the query sub-matrix, key matrix, and value matrix corresponding to each attention head;

[0177] The computing unit is used to calculate the scaled dot product attention of each attention head based on the query submatrix, key submatrix, and value submatrix corresponding to each attention head and the dimension of each attention head;

[0178] The stitching generation unit is used to stitch together the scaled dot product attention of each attention head and generate the final attention feature through the output projection layer;

[0179] The processing mapping unit is used to sequentially perform temporal convolution and global average pooling on the attention features, and then map the pooled attention features to the target output space through a multi-layer fully connected prediction network to realize the construction of the time-frequency attention fusion network model.

[0180] Furthermore, the time-frequency attention fusion network model includes: a SOM feature projection layer, a time-frequency attention fusion layer, a temporal convolutional layer, and a fully connected prediction layer.

[0181] Example 3:

[0182] refer to Figure 6 This embodiment provides an industrial product quality prediction device, including a memory 21 and a processor 22. The memory 21 stores a computer program, and the processor 22 is configured to run the computer program to execute the industrial product quality prediction method in Embodiment 1.

[0183] The memory 21 is connected to the processor 22. The memory 21 can be a flash memory, a read-only memory or other memory, and the processor 22 can be a central processing unit or a microcontroller.

[0184] Example 4:

[0185] This embodiment provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the industrial product quality prediction method in Embodiment 1 above.

[0186] The computer-readable storage medium includes volatile or non-volatile, removable or non-removable media implemented in any method or technology for storing information, such as computer-readable instructions, data structures, computer program modules or other data. Computer-readable storage media include, but are not limited to, RAM (Random Access Memory), ROM (Read-Only Memory), EEPROM (Electrically Erasable Programmable Read-Only Memory), flash memory or other memory technologies, CD-ROM (Compact Disc Read-Only Memory), DVD or other optical disc storage, cartridges, magnetic tapes, disk storage or other magnetic storage devices, or any other medium that can be used to store desired information and is accessible to a computer.

[0187] In summary, the industrial product quality prediction method, apparatus, and medium provided in this embodiment of the invention first acquire multivariate time-series data of an industrial process and divide the multivariate time-series data into training set data and test set data. Then, based on the training set data, enhanced frequency domain features corresponding to the training set data are obtained through Discrete Fourier Transform. Next, the enhanced frequency domain features are processed by topology preservation and low-dimensional projection through a complex frequency domain self-organizing map (SOM) network to obtain SOM projected frequency domain features. Then, a time-frequency attention fusion mechanism that includes time domain features and the SOM projected frequency domain features is designed, and a time-frequency attention fusion network model is constructed based on the time-frequency attention fusion mechanism and a temporal convolutional network to capture global-local time patterns. Finally, the test set data is input into the time-frequency attention fusion network model to obtain the predicted value of industrial product quality. This invention converts time-series data into complex frequency domain features using Discrete Fourier Transform (DFT), and utilizes a complex frequency domain self-organizing map network to achieve topology preservation and efficient dimensionality reduction of high-dimensional complex features while retaining amplitude and phase information in the signal. Furthermore, through a designed time-frequency attention fusion network, it adaptively fuses time-domain and frequency-domain features, thereby significantly improving the representational ability of multi-feature combinations and exhibiting stronger perception and capture capabilities for multi-scale characteristics and periodic patterns in industrial signals. In practical applications, it demonstrates higher prediction accuracy, possesses good practical value and promising prospects for widespread application, and solves the problem that existing industrial product quality prediction methods cannot adaptively fuse time-domain and frequency-domain features, resulting in low accuracy in industrial product quality prediction.

[0188] It is understood that the above embodiments are merely exemplary implementations used to illustrate the principles of the present invention, and the present invention is not limited thereto. For those skilled in the art, various modifications and improvements can be made without departing from the spirit and essence of the present invention, and these modifications and improvements are also considered to be within the scope of protection of the present invention.

Claims

1. A method for predicting the quality of industrial products, characterized in that, The method includes: Acquire multivariate time-series data of an industrial process, and divide the multivariate time-series data into training set data and test set data; Based on the training set data, the enhanced frequency domain features corresponding to the training set data are obtained through discrete Fourier transform; The enhanced frequency domain features are topology preserved and low-dimensionally projected by a complex frequency domain self-organizing map (SOM) network to obtain SOM projected frequency domain features. Design a time-frequency attention fusion mechanism that incorporates temporal features and the frequency domain features projected by the SOM, and construct a time-frequency attention fusion network model based on the time-frequency attention fusion mechanism and a temporal convolutional network to capture global-local temporal patterns; The test set data is input into the time-frequency attention fusion network model to obtain the predicted value of industrial product quality; The design includes a time-frequency attention fusion mechanism that integrates temporal features and the frequency domain features projected by the SOM. Based on this time-frequency attention fusion mechanism and a temporal convolutional network to capture global-local temporal patterns, a time-frequency attention fusion network model is constructed, specifically including: The SOM projection frequency domain features are used as the input feature matrix, and a linear projection transformation is performed on the input feature matrix to generate a query matrix, a key matrix, and a value matrix. The query matrix, key matrix, and value matrix are reshaped into a multi-attention head form, and the query matrix, key matrix, and value matrix are uniformly divided into multiple sub-matrices according to the feature dimension to obtain the query sub-matrix, key matrix, and value matrix corresponding to each attention head; Calculate the scaled dot product attention for each attention head based on the query submatrix, key matrix, and value submatrix corresponding to each attention head, as well as the dimension of each attention head; The scaled dot product attention of each attention head is concatenated and the final attention feature is generated through the output projection layer. The attention features are sequentially processed by temporal convolution and global average pooling, and the pooled attention features are mapped to the target output space through a multi-layer fully connected prediction network to realize the construction of the time-frequency attention fusion network model. The temporal convolution process is used to capture local temporal dependencies in the sequence, and the global average pooling process is used to aggregate feature information in the temporal dimension into a feature vector of fixed length.

2. The method according to claim 1, characterized in that, The step of obtaining the enhanced frequency domain features corresponding to the training set data through discrete Fourier transform specifically includes: Perform a discrete Fourier transform on the training set data to obtain complex features containing amplitude and phase information; The real and imaginary parts of the complex feature are extracted, converted to real numbers, and the amplitude spectrum is normalized to obtain the enhanced frequency domain feature.

3. The method according to claim 2, characterized in that, The step of performing a discrete Fourier transform on the training set data to obtain complex features containing amplitude and phase information specifically includes: The training set data is subjected to a discrete Fourier transform using the following formula: ； in, Indicates the first One variable, Indicates the length of the variable. This represents the result of the discrete Fourier transform of the signal. Represents frequency components, This represents the input time series signal. Let be the complex basis functions expressed in Euler's formula.

4. The method according to claim 1, characterized in that, The process of performing topology preservation and low-dimensional projection processing on the enhanced frequency domain features through a complex frequency domain self-organizing map (SOM) network to obtain SOM-projected frequency domain features specifically includes: Based on the enhanced frequency domain features, a complex distance metric function based on Euclidean distance is defined, and a complex frequency domain SOM network is constructed. The optimal matching unit (BMU) weight vector of the complex frequency domain SOM network is obtained through a competitive learning mechanism, so as to perform topology preservation and low-dimensional projection processing on the enhanced frequency domain features to obtain the SOM projected frequency domain features.

5. The method according to claim 4, characterized in that, The step of obtaining the optimal matching unit (BMU) weight vector of the complex frequency domain SOM network through a competitive learning mechanism, in order to perform topology preservation and low-dimensional projection processing on the enhanced frequency domain features to obtain the SOM projected frequency domain features, specifically includes: Design competitive learning rules for complex frequency domain SOM networks; Based on the competitive learning rules and enhanced frequency domain features, the complex frequency domain SOM network is pre-trained to obtain the BMU weight vector and form the SOM projected frequency domain features.

6. The method according to claim 1, characterized in that, The time-frequency attention fusion network model includes: a SOM feature projection layer, a time-frequency attention fusion layer, a temporal convolutional layer, and a fully connected prediction layer.

7. An industrial product quality prediction device, characterized in that, include: The acquisition module is used to acquire multivariate time-series data of industrial processes and divide the multivariate time-series data into training set data and test set data. The module is connected to the acquisition module and is used to obtain the enhanced frequency domain features corresponding to the training set data through discrete Fourier transform based on the training set data. A processing module, connected to the obtained module, is used to perform topology preservation and low-dimensional projection processing on the enhanced frequency domain features through a complex frequency domain self-organizing map (SOM) network to obtain SOM projected frequency domain features. The design and construction module is connected to the processing module and is used to design a time-frequency attention fusion mechanism that includes temporal features and the SOM projection frequency domain features. Based on the time-frequency attention fusion mechanism and the temporal convolutional network to capture global-local temporal patterns, a time-frequency attention fusion network model is constructed. The input module, connected to the design and construction module, is used to input the test set data into the time-frequency attention fusion network model to obtain the predicted value of industrial product quality. Furthermore, the design building module specifically includes: As a transformation unit, it is used to take the SOM projected frequency domain features as the input feature matrix and perform linear projection transformation on the input feature matrix to generate a query matrix, a key matrix and a value matrix; The reshaping segmentation unit is used to reshape the query matrix, key matrix, and value matrix into a multi-attention head form, and to uniformly divide the query matrix, key matrix, and value matrix into multiple sub-matrices according to the feature dimensions, so as to obtain the query sub-matrix, key matrix, and value matrix corresponding to each attention head; The computing unit is used to calculate the scaled dot product attention of each attention head based on the query submatrix, key submatrix, and value submatrix corresponding to each attention head and the dimension of each attention head; The stitching generation unit is used to stitch together the scaled dot product attention of each attention head and generate the final attention feature through the output projection layer; The processing mapping unit is used to sequentially perform temporal convolution and global average pooling on the attention features, and then map the pooled attention features to the target output space through a multi-layer fully connected prediction network to realize the construction of the time-frequency attention fusion network model. The temporal convolution process is used to capture local temporal dependencies in the sequence, and the global average pooling process is used to aggregate feature information in the temporal dimension into a feature vector of fixed length.

8. An industrial product quality prediction device, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to implement the industrial product quality prediction method as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the industrial product quality prediction method as described in any one of claims 1-6.

Images