Multi-working-condition industrial process quality index prediction method

Abstract

The embodiment of the invention provides a multi-working-condition industrial process quality index prediction method, and belongs to the technical field of data processing, and the method specifically comprises the steps: determining a key quality index, and screening a process variable related to the key quality index; preprocessing the data, and segmenting the data into a training set by adopting a sliding window technology; using an unsupervised clustering algorithm to perform working condition division according to the training set features; a variational graph attention auto-encoder framework is designed; constructing a comprehensive loss function, and fusing node feature reconstruction errors, graph structure reconstruction errors and potential feature representation loss; training a global shared feature extraction module and a local specific feature extraction module to obtain shared and specific features; fusing the features by using a gating mechanism, and inputting a full-connection layer to generate a predicted value; optimizing a feature extraction module and full connection layer parameters to form an index prediction model; and obtaining a real-time prediction value of the key quality index through the classification model and the prediction index model. Through the scheme of the invention, the accuracy and reliability of prediction are improved.

Description

Technical Field

[0001] The embodiments of the present disclosure relate to the field of data processing technology, and in particular to a method for predicting quality indicators of multi-operating-condition industrial processes. Background Art

[0002] Currently, due to the complex spatial interactions caused by the physical interconnection between sensors and the flow of material and energy, latent spatial dependencies exist between industrial process variables. However, existing deep learning methods typically treat data samples as unified entities and use gridded data as input, which significantly limits the effectiveness of models for industrial quality prediction tasks. Industrial processes are inherently complex nonlinear systems subject to multiple uncertainties. Their data spaces are often affected by noise contamination, high-dimensional coupling, and uncertainty propagation, further hindering the exploration of latent spatial dependencies between process variables. Furthermore, in actual industrial processes, due to multiple sources of uncertainty, such as fluctuations in raw material properties, dynamic adjustments in production demand, and changes in the external environment, production processes often exhibit multiple, significantly different, steady-state operating conditions. Most current data-driven modeling methods assume that data comes from a single production range, and their model construction process strictly relies on the assumption of independent and uniform distribution. Generally speaking, data features under the same operating condition converge to a specific representation space, forming a specific data modal distribution that describes the manufacturing characteristics of that operating condition. However, data features under different operating conditions are inherently different. When a model trained under a single operating condition is transferred to data from other operating conditions, it often faces problems such as feature representation mismatch and distribution shift. Furthermore, the similarities in the underlying physical mechanisms of industrial production dictate the potential for shared information between different modalities. Considering the multimodal nature of data is crucial for building predictive models for industrial process quality indicators.

[0003] In summary, the current data-driven modeling methods are difficult to effectively explore the potential interactions between process variables due to the complexity and uncertainty of industrial process mechanisms, and lack comprehensive consideration of the local specificity and global invariance characteristics of multi-condition industrial process modes, resulting in poor accuracy.

[0004] It can be seen that there is an urgent need for a multi-condition industrial process quality index prediction method with high prediction accuracy and reliability. Summary of the Invention

[0005] In view of this, an embodiment of the present disclosure provides a method for predicting quality indicators of multi-operating industrial processes, which at least partially solves the problem of poor prediction accuracy and reliability in the prior art.

[0006] In a first aspect, an embodiment of the present disclosure provides a method for predicting quality indicators of a multi-operating industrial process, comprising:

[0007] Step 1: Determine the key quality indicators in the industrial process according to the process flow, and select process variables with high correlation with the key quality indicators as auxiliary modeling variables, and use the key quality indicators and auxiliary modeling variables as industrial process data;

[0008] Step 2: preprocess the industrial process data and use the sliding window technology to split the preprocessed data into time windows to obtain the training set;

[0009] Step 3: Use an unsupervised clustering algorithm to divide the working conditions according to the sample characteristics of the training set and generate data pattern labels corresponding to the industrial process data at each historical moment;

[0010] Step 4: Design a variational graph attention autoencoder architecture to dynamically learn the correlation strength between variable nodes, combine variational inference to decouple the latent feature distribution, and extract high-quality spatial topological feature representation;

[0011] Step 5: Construct a comprehensive loss function that integrates the node feature reconstruction error, graph structure reconstruction error, and the loss of latent feature representation to fully capture the nonlinear coupling relationship between industrial variables.

[0012] Step 6: Based on the clustered pattern labels, the variational graph attention autoencoder architecture, and the comprehensive loss function, the global shared feature extraction module and the local specific feature extraction module for each working condition are trained separately to obtain shared features and specific features;

[0013] Step 7: Using the gating mechanism and based on the dynamic weight allocation strategy, the shared features and specific features are fused to generate a multi-modal gated fusion feature, which is then input into the fully connected layer for processing to obtain the predicted value corresponding to the key quality indicator in the industrial process data;

[0014] Step 8: Using the mean square error between the predicted value and the true value of the key quality indicator in the industrial process data as the target loss function, the back propagation algorithm is used to simultaneously optimize the parameters of the global shared feature extraction module, the local specific feature extraction module, and the fully connected layer to form an indicator prediction model;

[0015] Step 9: Collect the values ​​of each modeling auxiliary variable in real time, use the classification model to obtain the data pattern label of the sample through the unsupervised clustering algorithm, and based on the indicator prediction model, extract the shared features between the pattern labels and the specific features of each pattern to obtain the real-time prediction value of the key quality indicator.

[0016] According to a specific implementation of the embodiment of the present disclosure, step 4 specifically includes:

[0017] In step 4.1, the attention mechanism is incorporated into the graph convolutional layer. Based on this, a multi-layer graph attention encoding layer is constructed and used as the encoder part of the graph variational autoencoder model to obtain a high-quality spatial representation of the latent node features. The message passing process of the k-th layer of the graph attention encoding layer includes:

[0018] For node i, the expression of the correlation coefficient between it and its neighbor node j is:

[0019]

[0020] Among them, a (k) represents the attention weight vector of the k-th layer graph attention encoder; W (k) represents the trainable weight matrix of the k-th graph attention encoding layer, which is used to linearly map node i and node j to learn the correlation between them; [·||·] is the concatenation function used to combine the feature representations between nodes; θ(·) represents a shared single-layer feedforward neural network used to calculate the correlation coefficient;

[0021] The correlation coefficient e ij Normalization is performed using the following formula:

[0022]

[0023] Among them, E i Represents the neighbor spatial relationship set of node i;

[0024] Normalized correlation coefficient Update the feature representation of node i by taking the weighted sum of the feature representations of adjacent nodes:

[0025]

[0026] Among them, W (k) represents the weight matrix from the input layer to the hidden layer in the feature map of the k-th layer graph attention encoder, f(·) represents the activation function, is the initial feature representation of the industrial variable node;

[0027] Step 4.2, based on the idea of ​​variational inference, the graph attention encoder layer is used to fit the mean μ and variance σ, where the fitting process is

[0028] μ=GAT μ (X (k) ,E)

[0029] logσ=GAT σ (X (k) ,E)

[0030] Among them, X (k)represents the node feature representation obtained by the k-th graph attention encoding layer, μ and logσ share the same weight;

[0031] After the encoder determines the mean vector matrix and variance matrix of the potential multidimensional Gaussian distribution, the deep potential feature vector of the node is obtained by sampling:

[0032] Z~q(Z|X,E)=N(Z|μ,diag(σ 2 ))

[0033] In step 4.3, the decoder part reconstructs the node features X and spatial relations E through the inner product of the graph attention decoder layer and the latent feature vector:

[0034]

[0035] Among them, GAT (k+1) Indicates that the decoder layer uses the same structure as the corresponding encoder. Represents the inner product of any two node representation vectors to obtain the probability that there is an edge between the two nodes to reconstruct the graph.

[0036] According to a specific implementation of the embodiment of the present disclosure, the expression of the comprehensive loss function is:

[0037]

[0038] Among them, L X Represents the reconstruction loss of node feature X, which is measured by the mean square error between the original feature X and the reconstructed feature The gap, L E Represents the reconstruction loss of the spatial relationship E, by maximizing the conditional log-likelihood function Measures the quality of spatial relationship reconstruction, L KL It represents the divergence loss between the node potential feature distribution q(Z|X,E) and the standard Gaussian distribution p(Z), which is used to constrain the distribution structure of the potential feature space.

[0039] According to a specific implementation of the embodiment of the present disclosure, step 6 specifically includes:

[0040] Step 6.1: Take all training set sample data as input and build a globally shared variational graph attention autoencoder module VGATE shared , used to extract potential global common feature representations in cross-operation data;

[0041] Step 6.2: Based on the data pattern labels of the training sample data, take the training sample data of each pattern as input and construct K local specific variational graph attention autoencoder modules VGATE specific1 ,…,VGATEspecificK , used to extract potential local specific feature representations in each working condition data;

[0042] Step 6.3, in an unsupervised pre-training manner, the globally shared variational graph attention autoencoder module VGATE shared and K local specific variational graph attention autoencoder modules VGATE specific1 ,…,VGATE specificK Through training, a global shared feature extraction module and a local specific feature extraction module for each working condition are obtained, and based on this, feature extraction is performed on the pattern label to obtain shared features and specific features.

[0043] According to a specific implementation of the embodiment of the present disclosure, step 7 specifically includes:

[0044] Step 7.1, design a gating mechanism based on a dynamic weight allocation strategy to dynamically model the importance of shared features. Use the Sigmoid function as a feature-level soft switch to regulate the flow of information and quantify the degree of information retention. The expression of the gating weight corresponding to the shared feature of the ath pattern is:

[0045] g a =f g (W g Z shared_a +b g )

[0046] Among them, Z shared_a represents the shared feature representation of the a-th pattern, W g and b g are the trainable weight matrix and bias term of the gating mechanism, respectively, f g (·) represents the activation function of the gating mechanism;

[0047] Step 7.2, gating weight g based on shared features a , perform weighted fusion of the shared features and specific features under the ath mode to obtain the final multimodal gated fusion feature Z a :

[0048] Z a =g a ·Z shared_a +(1-g a )·Z specific_a

[0049] Among them, Z specific_a Represents the specific feature representation of the a-th pattern;

[0050] Step 7.3, input the multi-modal and multi-modal gated fusion features into the fully connected layer to obtain the predicted value:

[0051]

[0052] Among them, f p represents the activation function of the fully connected layer, W p and b p Represents the trainable weight matrix and bias terms of the fully connected layer.

[0053] The multi-condition industrial process quality indicator prediction scheme in the embodiment of the present disclosure includes: step 1, determining the key quality indicators in the industrial process according to the process flow, and screening process variables with high correlation with the key quality indicators as auxiliary modeling variables, and using the key quality indicators and auxiliary modeling variables as industrial process data; step 2, preprocessing the industrial process data, and using the sliding window technology to divide the preprocessed data according to the time window to obtain a training set; step 3, using an unsupervised clustering algorithm to divide the working conditions according to the sample characteristics of the training set, and generate data pattern labels corresponding to the industrial process data at each historical moment; step 4, designing a variational graph attention autoencoder architecture, dynamically learning the correlation strength between variable nodes, combining variational inference to decouple the potential feature distribution, and extracting high-quality spatial topological feature representation; step 5, constructing a comprehensive loss function, integrating the node feature reconstruction error, the graph structure reconstruction error and the loss of the potential feature representation, and fully capturing the nonlinear coupling relationship between industrial variables; step 6, based on clustering The obtained pattern labels, variational graph attention autoencoder architecture and comprehensive loss function are used to train the global shared feature extraction module and the local specific feature extraction module of each working condition respectively to obtain shared features and specific features; Step 7, using the gating mechanism, based on the dynamic weight allocation strategy, the obtained shared features and specific features are fused to generate multi-mode gated fusion features, and input them into the fully connected layer for processing to obtain the predicted values ​​corresponding to the key quality indicators in the industrial process data; Step 8, using the mean square error between the predicted values ​​and the true values ​​corresponding to the key quality indicators in the industrial process data as the target loss function, the back propagation algorithm is used to synchronously optimize the parameters of the global shared feature extraction module, the local specific feature extraction module and the fully connected layer to form an indicator prediction model; Step 9, the values ​​of each modeling auxiliary variable are collected in real time, and the data pattern labels of the samples are obtained by using the classification model through an unsupervised clustering algorithm, and the real-time predicted values ​​of the key quality indicators are obtained by extracting the shared features between the pattern labels and the specific features of each pattern based on the indicator prediction model.

[0054] The beneficial effects of the embodiments of the present disclosure are as follows: through the scheme of the present disclosure, by incorporating the attention mechanism into the graph convolution layer to adaptively learn the dynamic correlation strength between adjacent nodes, deeply mine the interactive information between variables, and introduce the graph attention network into the variational graph autoencoder, by designing a comprehensive comprehensive loss function for it, so as to maximize the extraction of potential high-quality features in industrial process data, and then construct a cross-modal data-driven joint sharing specific framework, which extracts shared features between modes and specific features of each mode according to the data mode labels, and uses a gating mechanism to fuse the two to comprehensively balance the global laws and working condition characteristics of the data under various working conditions, so that the final constructed indicator prediction model can greatly improve the accuracy and reliability of industrial key quality indicator prediction when predicting industrial key quality indicators. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] In order to more clearly illustrate the technical solutions of the embodiments of the present disclosure, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present disclosure. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0056] Figure 1 A flowchart of a method for predicting quality indicators of a multi-condition industrial process provided by an embodiment of the present disclosure;

[0057] Figure 2 A schematic diagram of a framework of a joint shared specific variational graph attention autoencoder model provided in an embodiment of the present disclosure;

[0058] Figure 3 A simplified flow chart of a potassium salt flotation process provided in an embodiment of the present disclosure;

[0059] Figure 4 A comparison curve of the predicted and actual potassium ion grade of concentrate in the potash flotation process using an LSTNet model provided in an embodiment of the present disclosure;

[0060] Figure 5 A comparison curve of the predicted and actual potassium ion grade of concentrate in a potash flotation process using a LogTrans model provided in an embodiment of the present disclosure;

[0061] Figure 6 A comparison curve of the predicted and actual values ​​of the potassium ion grade of the concentrate in the potash flotation process using an MTGNN model provided in an embodiment of the present disclosure;

[0062] Figure 7 A comparison curve chart of the predicted and actual values ​​of the potassium ion grade of the concentrate in the potash flotation process using the multi-condition industrial process quality index prediction method provided in an embodiment of the present disclosure. DETAILED DESCRIPTION

[0063] The embodiments of the present disclosure are described in detail below with reference to the accompanying drawings.

[0064] The following describes the embodiments of the present disclosure through specific examples, and those skilled in the art can easily understand other advantages and effects of the present disclosure from the contents disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present disclosure, rather than all of the embodiments. The present disclosure can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present disclosure. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in the present disclosure, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present disclosure.

[0065] It should be noted that various aspects of the embodiments within the scope of the appended claims are described below. It should be apparent that the aspects described herein can be embodied in a wide variety of forms, and any specific structure and / or function described herein is merely illustrative. Based on this disclosure, it should be understood by those skilled in the art that an aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects described herein can be used to implement an apparatus and / or practice a method. In addition, other structures and / or functionalities other than one or more of the aspects described herein can be used to implement this apparatus and / or practice this method.

[0066] It should also be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present disclosure. The illustrations only show components related to the present disclosure and are not drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component can be changed at will, and the component layout type may also be more complicated.

[0067] Additionally, in the following description, specific details are provided to provide a thorough understanding of the examples. However, one skilled in the art will appreciate that the aspects described can be practiced without these specific details.

[0068] The embodiments of the present disclosure provide a method for predicting quality indicators of a multi-operating-condition industrial process, which can be applied to the quality indicator prediction process of an industrial production scenario.

[0069] See also Figure 1 , is a flow chart of a method for predicting quality indicators of a multi-condition industrial process provided by an embodiment of the present disclosure. Figure 1As shown, the method mainly includes the following steps:

[0070] Step 1: Determine the key quality indicators in the industrial process according to the process flow, and select process variables with high correlation with the key quality indicators as auxiliary modeling variables, and use the key quality indicators and auxiliary modeling variables as industrial process data;

[0071] In the specific implementation, the example disclosed in this invention takes the industrial scenario of potassium salt flotation as an example. First, the concentrate potassium ion grade is determined as the key quality indicator in the potassium salt flotation process through the process flow mechanism, and multiple process variables with significant correlation with the concentrate potassium ion grade are selected from the potassium salt flotation process data as auxiliary variables for modeling. Among them, the correlation between process variables and key quality indicators can be evaluated based on prior knowledge of the process such as process mechanism and operating experience, and can also be quantitatively calculated using methods such as the Pearson correlation coefficient and the Spearman correlation coefficient. In the implementation case of the present invention, a total of 1,241 industrial process data sample points were obtained, and each sample point included 38 dimensions of features.

[0072] Step 2: preprocess the industrial process data and use the sliding window technology to split the preprocessed data into time windows to obtain the training set;

[0073] In the specific implementation, the historical industrial process data is first preprocessed by filling missing values, removing outliers and filtering data. Then, in order to facilitate data processing and model training, the collected key quality indicator data and its modeling auxiliary variables are normalized. Assume that the time series data set after preprocessing is Where X and Y represent the modeling auxiliary variables and key quality indicators, respectively. N and M represent the total number of samples and the number of auxiliary variables, respectively. The calculation formula for normalizing χ is as follows:

[0074] The modeling auxiliary variable X is normalized using the following formula:

[0075] X=(XX min,i ) / (X max,i -X min,i )

[0076] The key quality indicator Y is normalized by the following formula:

[0077] Y=(YY min ) / (Y max -Y min )

[0078] where X min,i and X max,i Represent the minimum and maximum values ​​of the i-th auxiliary variable, Y min and Ymax Respectively represent the minimum and maximum values ​​of the key quality indicator (i.e., the potassium ion grade of the concentrate). In order to further extract the temporal dynamic features, the sliding window technology is used to cut the industrial process data set into time windows, and the time window data is obtained as the original input data. Among them, R is the lookback window length of the historical sample set required for the model to extract dynamic features, which is set to 42, P is the prediction window length of the key sample indicator, which is set to 1, and the step size of the window sliding is set to 1. The first 80% of the time window data samples are used for training, and the last 20% of the time window data samples are used for subsequent testing of the generalization performance of the trained model.

[0079] Step 3: Use an unsupervised clustering algorithm to divide the working conditions according to the sample characteristics of the training set and generate data pattern labels corresponding to the industrial process data at each historical moment;

[0080] In the specific implementation, considering that the collected industrial process data usually does not contain data pattern labels, the data is first clustered unsupervisedly to assign data pattern labels to the samples based on their characteristics. The K-means algorithm is then used to perform modal segmentation on the preprocessed industrial process data samples to obtain the data pattern labels corresponding to the industrial process data samples at each historical moment.

[0081] The original dataset χ collected O There are K different data patterns M={M1,M2,…,M K}, use K-means algorithm to obtain sample x n (x n ∈χ O ,n∈[1,N]) is expressed as:

[0082]

[0083] Among them, K-means trained (·) represents the trained K-means model. A classification module based on the K-means method is introduced at the bottom of the network to pre-cluster the data and obtain pattern labels.

[0084] Step 4: Design a variational graph attention autoencoder architecture to dynamically learn the correlation strength between variable nodes, combine variational inference to decouple the latent feature distribution, and extract high-quality spatial topological feature representation;

[0085] In the specific implementation, a variational graph attention autoencoder architecture is designed to dynamically learn the correlation strength between variable nodes, combined with variational inference to decouple the potential feature distribution and extract high-quality spatial topological feature representation;

[0086] Based on the above embodiment, step 4 specifically includes:

[0087] In step 4.1, the attention mechanism is incorporated into the graph convolution layer to assist in the construction of the graph attention layer (GAT), and it is used as the encoder part of the graph variational autoencoder model to obtain high-quality spatial representations of latent node features. The message passing process of the k-th layer graph attention convolution encoder is as follows:

[0088] For node i, the expression of the correlation coefficient between it and its neighbor node j is:

[0089]

[0090] where a (k) represents the attention weight vector of the k-th layer graph attention encoder; W (k) represents the trainable weight matrix of the k-th layer graph attention encoder, which is used to linearly map node i and node j to learn the correlation between them; [·||·] is the splicing function used to combine the feature representations between nodes; θ(·) represents a shared single-layer feedforward neural network used to calculate the correlation coefficient.

[0091] In order to facilitate the comparison of the relationship between nodes, e ij Normalization is performed using the following formula:

[0092]

[0093] Among them E i Represents the set of neighbor spatial relationships of node i. Obtain the normalized relative attention coefficient After that, the feature representation of node i is updated by weighted summing of the feature representations of adjacent nodes, and its expression is:

[0094]

[0095] Among them, W (k) represents the weight matrix from the input layer to the hidden layer in the feature map of the k-th layer graph attention encoder, f(·) represents the activation function, is the initial feature representation of the industrial variable node.

[0096] Step 4.2, based on the idea of ​​variational inference, the graph attention encoder layer is used to fit the mean μ and variance σ, where the expression of the process is

[0097] μ=GAT μ (X (k) ,E)

[0098] logσ=GAT σ (X (k) ,E)

[0099] where X (k)represents the node feature representation obtained by the k-th graph attention encoder, μ and logσ share the same weight.

[0100] After the encoder determines the mean vector matrix and variance matrix of the potential multidimensional Gaussian distribution, the deep potential feature vector of the node can be obtained by sampling, as shown below:

[0101] Z~q(Z|X,E)=N(Z|μ,diag(σ 2 ))

[0102] In step 4.3, the decoder part reconstructs the node features X and spatial relations E through the inner product of the graph attention decoder layer and the latent vector, which is expressed as:

[0103]

[0104] Among them, GAT (k+1) Indicates that the decoder layer uses the same structure as the corresponding encoder. Represents the inner product of any two node representation vectors to obtain the probability that there is an edge between the two nodes to reconstruct the graph.

[0105] In practice, the variational graph attention autoencoder architecture consists of three components: a graph attention encoder, variational inference, and a decoder. The graph attention encoder utilizes a graph attention mechanism to cluster neighboring nodes and adaptively assign different neighbor weights, facilitating interaction-aware feature extraction. The variational inference strategy improves the quality of the latent feature distribution representation, thereby enhancing the overall model performance. The decoder reconstructs node features and spatial relationships through the inner product of the graph attention decoder layer and the latent vector, enhancing the model's ability to represent and reconstruct complex graph-structured data.

[0106] Specifically, in step 4 above, we design a variational graph attention autoencoder architecture, dynamically learn the correlation strength between variable nodes, combine variational inference to decouple the potential feature distribution, and extract high-quality spatial topological feature representations. The specific implementation method is divided into the following steps:

[0107] In the first step, the attention mechanism is incorporated into the graph convolutional layer to assist the construction of the graph attention layer (GAT), and it is used as the encoder part of the variational graph autoencoder model to obtain high-quality spatial representation of latent node features.

[0108] In the embodiments of the present disclosure, it specifically includes:

[0109] First, by the formula Calculate the correlation coefficient between node i and its neighbor node j in the k-th layer graph attention convolution encoder, where a (k) represents the attention weight vector of the k-th layer graph attention encoder; W (k)represents the trainable weight matrix of the k-th layer graph attention encoder, which is used to linearly map node i and node j to learn the correlation between them; [·||·] is the splicing function used to combine the feature representations between nodes; θ(·) represents a shared single-layer feedforward neural network used to calculate the correlation coefficient.

[0110] Secondly, in order to facilitate the comparison of the relationship between nodes, the formula Correlation coefficient Normalized, where E i Represents the set of neighbor spatial relationships of node i.

[0111] Finally, using the normalized relative attention coefficient By formula Update the feature representation of node i by performing weighted summation of the feature representations of adjacent nodes, where W (k) represents the weight matrix from the input layer to the hidden layer in the feature map of the k-th layer graph attention encoder, f(·) represents the activation function, is the initial feature representation of the industrial variable node.

[0112] In the second step, based on the idea of ​​variational inference, the graph attention encoder layer is used to fit the mean μ and variance σ. After the encoder determines the mean vector matrix and variance matrix of the potential multi-dimensional Gaussian distribution, the deep potential feature vector of the node is obtained by sampling.

[0113] In the embodiment of the present disclosure, the formula μ=GAT can be used to calculate μ (X (k) ,E) and logσ=GAT σ (X (k) ,E) respectively fit μ and σ, and use the formula Z~q(Z|X,E)=N(Z|μ,diag(σ 2 )) Get the deep potential features of the node, where X (k) represents the node feature representation obtained by the k-th graph attention encoder, μ and logσ share the same weight.

[0114] In the third step, the decoder part reconstructs the node features X and spatial relations E through the inner product of the graph attention decoder layer and the latent vector.

[0115] In this embodiment, the formula and To reconstruct node features X and spatial relations E respectively, where GAT (k+1) Indicates that the decoder layer uses the same structure as the corresponding encoder. Represents the inner product of any two node representation vectors to obtain the probability that there is an edge between the two nodes to reconstruct the graph.

[0116] Step 5: Construct a comprehensive loss function that integrates the node feature reconstruction error, graph structure reconstruction error, and the loss of latent feature representation to fully capture the nonlinear coupling relationship between industrial variables.

[0117] In the specific implementation, in order to obtain a well-trained variational graph attention autoencoder pre-training model, a comprehensive loss function is designed, including the node feature reconstruction error, graph structure reconstruction error and the loss of potential feature representation. The specific expression is:

[0118]

[0119] Among them, L X Represents the reconstruction loss of node feature X, and measures the difference between the original node feature X and the reconstructed feature by the mean square error (MSE) The gap; L E Represents the reconstruction loss of the spatial relationship E, by maximizing the conditional log-likelihood function Measures the quality of spatial relationship reconstruction; L KL It represents the divergence loss between the node potential feature distribution q(Z|X,E) and the standard Gaussian distribution p(Z), which is used to constrain the distribution structure of the potential feature space.

[0120] The above loss function uses the mean square error between the reconstructed node features and the original input node features as the node feature reconstruction loss, and the log-likelihood difference between the reconstructed graph structure and the original graph structure as the graph structure reconstruction loss. At the same time, KL divergence is introduced as a regularization term to measure the difference between the posterior distribution learned by the encoder and the standard Gaussian prior distribution. The three are combined to construct an overall comprehensive loss function to obtain high-quality representation of the reconstructed node features and graph topology structure, and skillfully capture the deep topological correlation between industrial variables.

[0121] Specifically, in step 5 above, a comprehensive loss function is constructed to integrate the node feature reconstruction error, graph structure reconstruction error, and the loss of potential feature representation, fully capturing the nonlinear coupling relationship between industrial variables. The specific implementation method is as follows:

[0122] In order to obtain a well-trained variational graph attention autoencoder pre-training model, a comprehensive loss function is designed, including node feature reconstruction error, graph structure reconstruction error and loss of latent feature representation.

[0123] In the embodiment, the formula

[0124] Define the comprehensive loss function of the variational graph attention autoencoder model, where L X Represents the reconstruction loss of node feature X, and measures the difference between the original node feature X and the reconstructed feature by the mean square error (MSE) The gap; L E Represents the reconstruction loss of the spatial relationship E, by maximizing the conditional log-likelihood function Measures the quality of spatial relationship reconstruction; L KL It represents the divergence loss between the node potential feature distribution q(Z|X,E) and the standard Gaussian distribution p(Z), which is used to constrain the distribution structure of the potential feature space.

[0125] Step 6: Based on the clustered pattern labels, the variational graph attention autoencoder architecture, and the comprehensive loss function, the global shared feature extraction module and the local specific feature extraction module for each working condition are trained separately to obtain shared features and specific features;

[0126] In specific implementation, the variational graph attention autoencoder is considered a feature extraction model. Considering that a single data-driven model, based on a single-mode distribution assumption, weakens its ability to capture data specificity under specific working conditions, and that the underlying mechanisms of the production process are constrained by similar physical mechanisms and production environments, hidden connections exist in data from multiple modes of the actual production process. A cross-modal data-driven joint shared specific framework is proposed. Through a cross-modal data joint learning mechanism, it captures global common information across all modalities, enabling collaborative mining and knowledge transfer of potential common information across multiple modalities. Furthermore, through a modal decoupling independent modeling framework, it accurately reveals the differentiated feature representations of each mode, enhancing the accuracy of characterizing the local distribution characteristics of heterogeneous working condition data. All training sample data and training sample data for each mode are used to train the global shared feature extraction module and K local specific feature extraction modules, respectively. The parameters of the pre-trained encoder are used as initialization parameters for the hidden layer features of the prediction model, thereby helping the network find a good initialization point.

[0127] Through the cross-modal data joint learning mechanism, the global common information of all modalities is captured, and the collaborative mining and knowledge transfer of potential common information among multiple modalities are realized.

[0128] Specifically, in step 6 above, based on the pattern labels obtained by clustering, the global shared feature extraction module and the local specific feature extraction module for each working condition are trained separately to realize the parallel mining of cross-modal common features and modality-specific features. The specific implementation method is divided into the following steps:

[0129] The first step is to take all training sample data as input and train a globally shared variational graph attention autoencoder module VGATE shared , used to extract potential global common feature representations in cross-operating condition data.

[0130] The second step is to train K local specific variational graph attention autoencoder modules VGATE based on the data pattern labels of the training sample data and take the training sample data of each pattern as input. specific1 ,…,VGATE specificK , which is used to extract the potential local specific feature representations in each working condition data.

[0131] The third step is to use unsupervised pre-training to train the globally shared variational graph attention autoencoder module VGATE shared and K local specific variational graph attention autoencoder modules VGATE specific1 ,…,VGATE specificK The training is performed and the parameters of the encoder part are used as the initialization parameters of the hidden layer features of the subsequent prediction model, so as to extract high-quality representations of the potential shared features and specific features of the training sample data in the industrial process.

[0132] Step 7: Using the gating mechanism and based on the dynamic weight allocation strategy, the shared features and specific features are fused to generate a multi-modal gated fusion feature, which is then input into the fully connected layer for processing to obtain the predicted value corresponding to the key quality indicator in the industrial process data;

[0133] In the specific implementation, in order to achieve the adaptive fusion of multi-modal features, a gating mechanism is designed and based on the dynamic weight allocation strategy, the extracted shared features Z are transformed into shared and specific feature Z specific By integrating and comprehensively balancing the working characteristics and global laws of the data under various working conditions, more comprehensive and effective data pattern information can be obtained to improve the predictive performance of the model.

[0134] It is understandable that after obtaining the multi-modal and multi-modal gated fusion features, in order to predict the key quality indicators corresponding to the original data, an additional readout layer is added at the end of the network. The expression of the prediction process is:

[0135]

[0136] Among them, f p represents the activation function of the readout layer, W p and b p Represents the trainable weight matrix and bias term of the readout layer.

[0137] Specifically, in step 7 above, the shared features and specific features are fused using a gating mechanism based on a dynamic weight allocation strategy to generate a multi-modal gated fusion feature, which is then input into a fully connected layer for processing to obtain the predicted value of the key quality indicator corresponding to the original data. The specific implementation method is divided into the following steps:

[0138] In the first step, a gating mechanism is designed based on a dynamic weight allocation strategy to dynamically model the importance of shared features. The Sigmoid function is used as a feature-level soft switch to regulate the flow of information and quantify the degree of information retention.

[0139] In the embodiment of the present disclosure, specifically through formula g a =f g (W g Z shared_a +b g ) to obtain the shared feature gating value of the a-th mode, where Z shared_a represents the shared feature representation of the a-th pattern, W g and b g are the trainable weight matrix and bias term of the gating module, respectively, f g (·) represents the activation function of the gating module, which is set to the Sigmoid activation function to obtain the weight coefficient for preserving shared features.

[0140] In the second step, the obtained shared feature gating weight g is used a , perform weighted fusion of the shared features and specific features in the ath mode.

[0141] In the embodiment of the present disclosure, specifically through the formula Z a =g a ·Z shared_a +(1-g a )·Z specific_a Get the final multimodal gated fusion feature Z a , where Z specific_a Represents the specific feature representation of the a-th pattern.

[0142] The third step is to use the multi-modal multi-modal gated fusion feature as Z a Input, an additional readout layer is added at the end of the network to predict the key quality indicators corresponding to the original data.

[0143] In this embodiment, specifically through the formula Get the predicted value of the final key quality indicator where f p represents the activation function of the readout layer, W p and b p Represents the trainable weight matrix and bias term of the readout layer.

[0144] Step 8: Using the mean square error between the predicted value and the true value of the key quality indicator in the industrial process data as the target loss function, the back propagation algorithm is used to simultaneously optimize the parameters of the global shared feature extraction module, the local specific feature extraction module, and the fully connected layer to form an indicator prediction model;

[0145] In specific implementation, based on the predicted values ​​of key quality indicators To obtain the predicted value The mean square error between the true value y is the loss function, and its expression is:

[0146]

[0147] Among them, θ represents the learnable parameters in the model, and N represents the number of samples. Select the appropriate optimizer, learning rate, and number of iterations, and then use the obtained loss function to supervise the reverse update of the weight parameters of the key indicator prediction model composed of the global shared feature extraction module, the local specific feature extraction module, and the fully connected prediction layer until the model converges. The trained key indicator prediction model is obtained as follows Figure 2 shown.

[0148] Step 9: Collect the values ​​of each modeling auxiliary variable in real time, use the classification model to obtain the data pattern label of the sample through the unsupervised clustering algorithm, and based on the indicator prediction model, extract the shared features between the pattern labels and the specific features of each pattern to obtain the real-time prediction value of the key quality indicator.

[0149] In specific implementation, after obtaining the trained key quality indicator prediction model, the values ​​of all modeling auxiliary variables collected in real time can be input into the trained indicator prediction model to perform real-time prediction of the key quality indicators.

[0150] The multi-operating condition industrial process quality indicator prediction method provided in this embodiment incorporates the attention mechanism into the graph convolution layer to adaptively learn the dynamic correlation strength between adjacent nodes, deeply mine the interactive information between variables, and introduces the graph attention network into the variational graph autoencoder. By designing a comprehensive integrated loss function for it, it can maximize the extraction of potential high-quality features in industrial process data. Subsequently, a cross-modal data-driven joint sharing specific framework is constructed to extract shared features between modes and specific features of each mode according to the data mode labels, and a gating mechanism is used to fuse the two to comprehensively balance the global laws and operating condition characteristics of the data under each operating condition, so that the final constructed indicator prediction model can greatly improve the accuracy and reliability of industrial key quality indicator prediction when used for industrial key quality indicator prediction.

[0151] The method will be further described below with reference to a specific embodiment. Figure 3 As shown, the above-mentioned multi-condition industrial process quality index prediction method is further explained below using the potassium salt flotation process as an example.

[0152] Specifically, you can use Figure 2A multi-condition industrial process quality indicator prediction method based on a joint shared specific variational graph attention autoencoder is shown.

[0153] The standard mean absolute error (MAE), mean absolute percentage error (MAPE), root mean square error (RMSE) and coefficient of determination (R 2 ) is used to quantitatively evaluate the prediction performance of the obtained key indicator prediction model. The calculation formula of each indicator is as follows:

[0154]

[0155]

[0156] in Represents the average value of all test set data samples, y i and They represent the true value and predicted value of the key quality indicator of the i-th sample in the test set data.

[0157] In the embodiment of the present invention, the LSTNet model, LogTrans model and MTGNN model are used to compare with the multi-condition industrial process quality index prediction method of the embodiment of the present disclosure to verify the effectiveness of the proposed method. The prediction results are as follows: Figure 4 、 5 , 6, and 7, the standard mean absolute error (MAE), mean absolute percentage error (MAPE), root mean square error (RMSE), and coefficient of determination (R 2 ) as shown in Table 1.

[0158] Table 1

[0159]

[0160] In the embodiments of the present invention, it can be seen that the multi-condition industrial process quality index prediction method based on the joint shared specific variational graph attention autoencoder provided by the present invention has achieved the best prediction effect in various evaluation indicators compared with traditional deep learning prediction methods such as the LSTNet model, LogTrans model and MTGNN model, which also verifies the effectiveness of the method disclosed in the present invention.

[0161] It should be understood that various parts of the present disclosure can be implemented in hardware, software, firmware, or a combination thereof.

[0162] The above description is merely a specific embodiment of the present disclosure, but the scope of protection of the present disclosure is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this disclosure should be included in the scope of protection of the present disclosure. Therefore, the scope of protection of the present disclosure should be based on the scope of protection of the claims.

Claims

1. A method for predicting quality indicators of multi-condition industrial processes, characterized in that: include: Step 1: Determine the key quality indicators in the industrial process according to the process flow, and select process variables with high correlation with the key quality indicators as auxiliary modeling variables, and use the key quality indicators and auxiliary modeling variables as industrial process data; Step 2: preprocess the industrial process data and use the sliding window technology to split the preprocessed data into time windows to obtain the training set; Step 3: Use an unsupervised clustering algorithm to divide the working conditions according to the sample characteristics of the training set and generate data pattern labels corresponding to the industrial process data at each historical moment; Step 4: Design a variational graph attention autoencoder architecture to dynamically learn the correlation strength between variable nodes, combine variational inference to decouple the latent feature distribution, and extract high-quality spatial topological feature representation; Step 5: Construct a comprehensive loss function that integrates the node feature reconstruction error, graph structure reconstruction error, and the loss of latent feature representation to fully capture the nonlinear coupling relationship between industrial variables. Step 6: Based on the clustered pattern labels, the variational graph attention autoencoder architecture, and the comprehensive loss function, the global shared feature extraction module and the local specific feature extraction module for each working condition are trained separately to obtain shared features and specific features; Step 7: Using the gating mechanism and based on the dynamic weight allocation strategy, the shared features and specific features are fused to generate a multi-modal gated fusion feature, which is then input into the fully connected layer for processing to obtain the predicted value corresponding to the key quality indicator in the industrial process data; Step 8: Using the mean square error between the predicted value and the true value of the key quality indicator in the industrial process data as the target loss function, the back propagation algorithm is used to simultaneously optimize the parameters of the global shared feature extraction module, the local specific feature extraction module, and the fully connected layer to form an indicator prediction model; Step 9: Collect the values ​​of each modeling auxiliary variable in real time, use the classification model to obtain the data pattern label of the sample through the unsupervised clustering algorithm, and based on the indicator prediction model, extract the shared features between the pattern labels and the specific features of each pattern to obtain the real-time prediction value of the key quality indicator.

2. The method according to claim 1, characterized in that The step 4 specifically includes: In step 4.1, the attention mechanism is incorporated into the graph convolutional layer. Based on this, a multi-layer graph attention encoding layer is constructed and used as the encoder part of the graph variational autoencoder model to obtain a high-quality spatial representation of the latent node features. The message passing process of the k-th layer of the graph attention encoding layer includes: For node i, the expression of the correlation coefficient between it and its neighbor node j is: Among them, a (k) represents the attention weight vector of the k-th layer graph attention encoder; W (k) represents the trainable weight matrix of the k-th graph attention encoding layer, which is used to linearly map node i and node j to learn the correlation between them; [·||·] is the concatenation function used to combine the feature representations between nodes; θ(·) represents a shared single-layer feedforward neural network used to calculate the correlation coefficient; The correlation coefficient e ij Normalization is performed using the following formula: Among them, E i Represents the set of neighbor spatial relationships of node i; Normalized correlation coefficient Update the feature representation of node i by taking the weighted sum of the feature representations of adjacent nodes: Among them, W (k) represents the weight matrix from the input layer to the hidden layer in the feature map of the k-th layer graph attention encoder, f(·) represents the activation function, is the initial feature representation of the industrial variable node; Step 4.2, based on the idea of ​​variational inference, the graph attention encoder layer is used to fit the mean μ and variance σ, where the fitting process is μ=GAT μ (X (k) ,AND) logσ=GAT σ (X (k) ,A) Among them, X (k) represents the node feature representation obtained by the k-th graph attention encoding layer, μ and logσ share the same weight; After the encoder determines the mean vector matrix and variance matrix of the potential multidimensional Gaussian distribution, the deep potential feature vector of the node is obtained by sampling: Z~q(Z|X,E)=N(Z|μ,diag(σ 2 )) In step 4.3, the decoder part reconstructs the node features X and spatial relations E through the inner product of the graph attention decoder layer and the latent feature vector: Among them, GAT (k+1) Indicates that the decoder layer uses the same structure as the corresponding encoder. Represents the inner product of any two node representation vectors to obtain the probability that there is an edge between the two nodes to reconstruct the graph.

3. The method according to claim 2, characterized in that The expression of the comprehensive loss function is: Among them, L X Represents the reconstruction loss of node feature X, which is measured by the mean square error between the original feature X and the reconstructed feature The gap, L E Represents the reconstruction loss of the spatial relationship E, by maximizing the conditional log-likelihood function Measures the quality of spatial relationship reconstruction, L KL It represents the divergence loss between the node potential feature distribution q(Z|X,E) and the standard Gaussian distribution p(Z), which is used to constrain the distribution structure of the potential feature space.

4. The method according to claim 3, characterized in that The step 6 specifically includes: Step 6.1: Take all training set sample data as input and build a globally shared variational graph attention autoencoder module VGATE shared , used to extract potential global common feature representations in cross-operation data; Step 6.2: Based on the data pattern labels of the training sample data, take the training sample data of each pattern as input and construct K local specific variational graph attention autoencoder modules VGATE specific1 ,…,VGATE specificK , used to extract potential local specific feature representations in each working condition data; Step 6.3, in an unsupervised pre-training manner, the globally shared variational graph attention autoencoder module VGATE shared and K local specific variational graph attention autoencoder modules VGATE specific1 ,…,VGATE specificK Through training, a global shared feature extraction module and a local specific feature extraction module for each working condition are obtained, and based on this, feature extraction is performed on the pattern label to obtain shared features and specific features.

5. The method according to claim 4, characterized in that The step 7 specifically includes: Step 7.1, design a gating mechanism based on a dynamic weight allocation strategy to dynamically model the importance of shared features. Use the Sigmoid function as a feature-level soft switch to regulate the flow of information and quantify the degree of information retention. The expression of the gating weight corresponding to the shared feature of the ath pattern is: g a =f g (W g Z shared_a +b g ) Among them, Z shared_a represents the shared feature representation of the a-th pattern, W g and b g are the trainable weight matrix and bias term of the gating mechanism, respectively, and f g (·) represents the activation function of the gating mechanism; Step 7.2, gating weight g based on shared features a , perform weighted fusion of the shared features and specific features under the ath mode to obtain the final multimodal gated fusion feature Z a : Z a =g a ·Z shared_a +(1-g a )·Z specific_a Among them, Z specific_a Represents the specific feature representation of the a-th pattern; Step 7.3, input the multi-modal and multi-modal gated fusion features into the fully connected layer to obtain the predicted value: Among them, f p represents the activation function of the fully connected layer, W p and b p Represents the trainable weight matrix and bias terms of the fully connected layer.

Images