A quality prediction method based on spatio-temporal dynamic graph convolution and mutual attention selection

Abstract

The application relates to a quality prediction method based on spatio-temporal dynamic graph convolution and mutual attention selection. The method comprises the following steps: collecting historical observation values of target variables and process variables in a chemical process, combining the historical observation values into one-dimensional vectors, performing dimension increasing processing on the time dimension of the one-dimensional vectors, and then performing feature embedding and grouping processing to obtain a plurality of grouped features; performing convolution operation on each grouped feature to obtain corresponding time feature representation; obtaining first spatio-temporal representation corresponding to the time feature representation based on an adaptive topological structure and a gating mechanism; aggregating each time feature representation through attention weights to obtain aggregated features; performing average pooling on the aggregated features in the time dimension and the node feature dimension respectively, and performing convolution operation respectively to obtain global features, time local features and node local features, and then obtaining second spatio-temporal representation; and performing residual connection on the second spatio-temporal representation and each first spatio-temporal representation, and then outputting the prediction value of the target variable through a fully connected layer.

Description

Technical Field

[0001] The present application relates to the technical field of chemical process quality prediction, and in particular to a quality prediction method based on spatiotemporal dynamic graph convolution and mutual attention selection. Background Art

[0002] Chemical process quality prediction is to build a quantitative relationship model between measurable process parameters (temperature, pressure, flow, component concentration, etc.) and key quality indicators (product purity, yield, impurity content, etc.) through data-driven methods, and realize online soft measurement and predictive analysis of product quality.

[0003] Existing deep learning methods still face the following challenges in chemical quality prediction scenarios:

[0004] 1. Existing methods mostly use time series model architectures, which make it difficult to effectively model the nonlinear spatiotemporal interaction effects between variables. In particular, for typical chemical equipment such as reactors and distillation towers, the local-global spatiotemporal coupling characteristics formed by material transfer and energy exchange have not been effectively characterized.

[0005] 2. Existing methods usually only rely on weight adjustment between local channels, while ignoring the collaborative interaction between space and layers, which makes it difficult for the model to fully explore the potential correlation between different feature dimensions, thus affecting the final quality prediction performance.

[0006] 3. Existing Transformer-based quality prediction models lack an effective feature filtering mechanism. Similar features extracted at different levels interfere with each other, causing key discriminative information to be buried in redundant data, thereby reducing the model's prediction accuracy and generalization capabilities. Furthermore, overly deep network structures can introduce unnecessary computational overhead, impacting the model's inference efficiency. Summary of the Invention

[0007] Based on this, it is necessary to provide a quality prediction method based on spatiotemporal dynamic graph convolution and mutual attention selection, including:

[0008] S1: Collect historical observations of target variables and process variables in the chemical process, combine the historical observations to obtain a one-dimensional vector, perform dimensionality-upgrading on the time dimension of the one-dimensional vector to obtain a high-dimensional representation of the data; perform feature embedding and grouping on the high-dimensional representation of the data to obtain multiple group features; perform convolution operations on each group feature to obtain the corresponding time feature representation;

[0009] S2: Construct an adaptive topological structure and perform graph convolution on each temporal feature representation based on the adaptive topological structure to obtain the corresponding neighborhood features; fuse the temporal feature representation with the corresponding neighborhood features based on the gating mechanism to obtain the corresponding first spatiotemporal representation;

[0010] S3: Calculate the attention weight based on each temporal feature representation, and aggregate each temporal feature representation based on the attention weight to obtain the aggregated feature; perform average pooling on the aggregated feature in the temporal dimension and the node feature dimension respectively, and weight the aggregated feature based on the average pooling result to obtain the global feature; perform convolution operation on the aggregated feature in the temporal dimension and the node feature dimension respectively, and pass the two convolution results through the attention mechanism respectively to obtain the temporal local feature and the node local feature; add the global feature, the temporal local feature, and the node local feature to obtain the second spatiotemporal representation;

[0011] S4: Perform a residual connection between the second spatiotemporal representation and each first spatiotemporal representation, and then pass it through a fully connected layer to output the predicted value of the target variable.

[0012] Preferably, the chemical process includes a debutanizer fractionation process or a hydrocracking process in an oil refining process; and the target variable includes the butane content in the debutanizer bottom material or the 10% recovery temperature of the jet fuel in the hydrocracking process.

[0013] Preferably, the process of obtaining the grouping features includes:

[0014] Pass the one-dimensional vector through the fully connected layer and add the mapping result to the positional encoding after sine and cosine encoding to obtain a high-dimensional representation of the data;

[0015] Input the high-dimensional representation of the data into the node feature embedding layer and output the node feature representation;

[0016] The node feature representations are equally grouped based on the node feature dimensions of the node feature representations to obtain multiple grouping features.

[0017] Preferably, building an adaptive topology structure includes:

[0018] Construct a learnable first node feature matrix and a second node feature matrix;

[0019] The values ​​of each element in the first node feature matrix and the second node feature matrix are random values;

[0020] The first node feature matrix and the second node feature matrix are multiplied element by element to obtain a first multiplication result; the first multiplication result is passed through a ReLU activation function to obtain a first activation result; the first activation result is passed through a Softmax function to obtain an adaptive topology structure.

[0021] Preferably, the process of obtaining the first spatiotemporal representation is:

[0022] For any time feature representation,

[0023] Multiply the adaptive topological structure and the temporal feature representation element by element to obtain the corresponding neighborhood features;

[0024] Perform feature splicing on the time feature representation and the corresponding neighborhood feature to obtain a first splicing result;

[0025] Pass the first splicing result through the convolutional network to obtain a first convolution result;

[0026] Pass the first convolution result through the Sigmoid function to obtain the corresponding fusion weight;

[0027] The temporal feature representation is fused with the corresponding neighborhood feature based on the fusion weight to obtain the corresponding first spatiotemporal representation;

[0028] Traverse all temporal feature representations and calculate the first spatiotemporal representation corresponding to each temporal feature representation.

[0029] Preferably, fusing the temporal feature representation with the corresponding neighborhood feature based on the fusion weight includes:

[0030] The fusion weight is multiplied by the time feature representation element by element to obtain the second multiplication result; the neighborhood feature is multiplied by the balance number of the fusion weight about 1 element by element to obtain the third multiplication result;

[0031] The second multiplication result is added to the third multiplication result to obtain the corresponding first space-time representation.

[0032] Preferably, the process of obtaining the aggregated features includes:

[0033] Perform feature splicing on each time feature representation to obtain a second splicing result; perform one-dimensional convolution on the second splicing result to obtain a second convolution result; and pass the second convolution result through the ReLU activation function to obtain channel features;

[0034] The channel features are multiplied element-by-element by the learnable weight vector and then passed through the Softmax function to obtain the attention weight corresponding to each time feature representation;

[0035] Multiply each attention weight by the corresponding temporal feature representation element by element to obtain multiple fourth multiplication results;

[0036] Sum all the fourth multiplication results to get the aggregate feature.

[0037] Preferably, the process of obtaining the global features includes:

[0038] The aggregated features are average-pooled in the time dimension and feature dimension to obtain time global features and node global features respectively;

[0039] Multiply the transposed time global feature and the node global feature to obtain the fifth multiplication result;

[0040] Perform two-dimensional convolution on the fifth multiplication result to obtain the third convolution result; and pass the third convolution result through the Sigmoid function to obtain the weight matrix;

[0041] The weight matrix is ​​multiplied element-wise by the aggregated features to obtain the global features.

[0042] Preferably, the process of obtaining the temporal local features and the node local features includes:

[0043] Convolution operations are performed on the aggregated features in the time dimension and the node feature dimension respectively to obtain the fourth convolution result and the fifth convolution result respectively;

[0044] The fourth convolution result is used as the first query and first value, and the fifth convolution result is used as the first key. The temporal local features are calculated through the attention mechanism.

[0045] The fifth convolution result is used as the second query and second value, and the fourth convolution result is used as the second key. The local features of the node are calculated through the attention mechanism.

[0046] Preferably, a loss function is constructed based on the predicted value of the target variable and the true value of the target variable, and the mean absolute error calculation formula is used to minimize the loss function to train the quality prediction method.

[0047] Beneficial effects: This method collects historical observations of target variables and process variables in the chemical process, combines them into a one-dimensional vector, and performs dimensionality-increasing processing on the time dimension of the one-dimensional vector. It then performs feature embedding and grouping processing to obtain multiple group features. Each group feature is convolved to obtain the corresponding time feature representation. Based on the adaptive topological structure and gating mechanism, the first spatiotemporal representation corresponding to the time feature representation is obtained. Secondly, the time feature representations are aggregated by attention weights to obtain aggregated features. The aggregated features are average-pooled and convolved in the time dimension and node feature dimension respectively to obtain global features, time-local features, and node-local features, thereby obtaining the second spatiotemporal representation. Finally, the second spatiotemporal representation is residually connected with each first spatiotemporal representation and then passed through a fully connected layer to output the predicted value of the target variable. This method effectively suppresses multi-scale redundancy, highlights key features, and significantly improves the prediction performance of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0049] Figure 1 This is a flowchart of the quality prediction method based on spatiotemporal dynamic graph convolution and mutual attention selection in an embodiment of the present application.

[0050] Figure 2 This is a graph showing the predicted butane content at the bottom of the debutanizer during the debutanizer fractionation process in the embodiment of the present application.

[0051] Figure 3 This is a diagram showing the predicted results of the 10% recovery temperature of aviation fuel during the hydrocracking process in the examples of this application. DETAILED DESCRIPTION

[0052] To make the above-mentioned objects, features, and advantages of the present application more clearly understood, the specific embodiments of the present application are described in detail below with reference to the accompanying drawings. The following description sets forth many specific details to facilitate a full understanding of the present application. However, the present application can be implemented in many other ways than those described herein, and those skilled in the art can make similar improvements without violating the scope of the present application. Therefore, the present application is not limited to the specific embodiments disclosed below.

[0053] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of the technical features being referred to. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of such features. Throughout the description of this application, "plurality" means at least two, for example, two, three, etc., unless otherwise specifically defined.

[0054] like Figure 1 As shown, this embodiment provides a quality prediction method based on spatiotemporal dynamic graph convolution and mutual attention selection, including:

[0055] S1: Collect historical observations of target variables and process variables in the chemical process, combine the historical observations to obtain a one-dimensional vector, perform dimensionality-upgrading on the time dimension of the one-dimensional vector to obtain a high-dimensional representation of the data; perform feature embedding and grouping on the high-dimensional representation of the data to obtain multiple group features; perform convolution operations on each group feature to obtain the corresponding time feature representation.

[0056] In this embodiment, the chemical process includes a debutanizer fractionation process or a hydrocracking process in an oil refining process; the target variable includes the butane content in the debutanizer bottom material or the 10% recovery temperature of the jet fuel in the hydrocracking process.

[0057] The process variables in the debutanizer fractionation process are shown in Table 1.

[0058] Table 1 is a description table of process variables in the debutanizer fractionation process;

[0059] ;

[0060] Table 1 shows the seven process variables in the debutanizer fractionation process. Historical observations are taken for all seven process variables. The historical observations of each process variable are arranged in variable order and time order, and combined with the historical observations of the target variable to form a one-dimensional vector to predict the butane content in the debutanizer bottom material.

[0061] The process variables in the hydrocracking process are shown in Table 2.

[0062] Table 2 is a description table of process variables in the hydrocracking process;

[0063] ;

[0064] Table 2 shows the 44 process variables in the hydrocracking process. Historical observations are taken for all 44 process variables. The historical observations of each process variable are arranged in variable order and time order, and combined with the historical observations of the target variable to form a one-dimensional vector to predict the 10% recovery temperature of the jet fuel in the hydrocracking process.

[0065] In this embodiment, the historical observation values ​​are sampled using a sliding window, that is, the historical observation values ​​of the target variable and the process variable are taken in the time dimension with a preset sliding window size. l (The preset sliding window size can be set according to the actual application) The observed values ​​of the target variable and process variable in the step are used to predict the future value of the target variable.

[0066] The one-dimensional vector is embedded through the time dimension embedding layer to extract key information in the time dimension, providing a basis for subsequent feature processing.

[0067] Furthermore, the process of obtaining grouping features includes:

[0068] Pass the one-dimensional vector through the fully connected layer and add the mapping result to the positional encoding after sine and cosine encoding to obtain a high-dimensional representation of the data;

[0069] Input the high-dimensional representation of the data into the node feature embedding layer and output the node feature representation;

[0070] Node feature representations are grouped in equal quantities based on their node feature dimensions to obtain multiple grouping features. The node feature dimension of the node feature representation is denoted by Z. Node feature representations are grouped in equal quantities based on dimension Z. Each resulting grouping feature contains g node feature representations. For example, if Z is 4 and g is 2, the first and second node feature representations form a group corresponding to the first grouping feature, and the third and fourth node feature representations form a group corresponding to the second grouping feature.

[0071] Convolution processing is performed on each time feature representation in the time dimension to capture the dynamic features at different time scales, effectively modeling the dynamic dependency relationship in the time dimension, and providing a basis for subsequent spatiotemporal representation.

[0072] S2: Construct an adaptive topological structure, perform graph convolution processing on each temporal feature representation based on the adaptive topological structure to obtain the corresponding neighborhood features; fuse the temporal feature representation with the corresponding neighborhood features based on the gating mechanism to obtain the corresponding first spatiotemporal representation.

[0073] Specifically, building an adaptive topology structure includes:

[0074] Construct a learnable first node feature matrix and a second node feature matrix;

[0075] The values ​​of each element in the first node feature matrix and the second node feature matrix are random values;

[0076] The first node feature matrix and the second node feature matrix are multiplied element by element to obtain a first multiplication result; the first multiplication result is passed through a ReLU activation function to obtain a first activation result; the first activation result is passed through a Softmax function to obtain an adaptive topology structure.

[0077] Furthermore, the process of obtaining the first space-time representation is:

[0078] For any time feature representation,

[0079] Multiply the adaptive topological structure and the temporal feature representation element by element to obtain the corresponding neighborhood features;

[0080] Perform feature splicing on the time feature representation and the corresponding neighborhood feature to obtain a first splicing result;

[0081] Pass the first splicing result through the convolutional network to obtain a first convolution result;

[0082] Pass the first convolution result through the Sigmoid function to obtain the corresponding fusion weight;

[0083] The temporal feature representation is fused with the corresponding neighborhood feature based on the fusion weight to obtain the corresponding first spatiotemporal representation, which has multi-scale characteristics;

[0084] Traverse all temporal feature representations and calculate the first spatiotemporal representation corresponding to each temporal feature representation.

[0085] Furthermore, the temporal feature representation is fused with the corresponding neighborhood feature based on the fusion weight, including:

[0086] The fusion weight is multiplied by the time feature representation element by element to obtain the second multiplication result; the neighborhood feature is multiplied by the balance number of the fusion weight about 1 element by element to obtain the third multiplication result;

[0087] The second multiplication result is added to the third multiplication result to obtain the corresponding first space-time representation.

[0088] By dynamically aggregating feature information between different nodes, complex spatial dependencies can be modeled. A gating mechanism is then used to dynamically adjust how information is fused between a node's own features and those of its neighbors. This design not only improves the model's ability to model dynamic spatial relationships, but also effectively mitigates oversmoothing issues through the gating mechanism, enhancing feature discriminability.

[0089] S3: Calculate the attention weight based on each temporal feature representation, and aggregate each temporal feature representation based on the attention weight to obtain the aggregated feature; perform average pooling on the aggregated feature in the temporal dimension and the node feature dimension respectively, and weight the aggregated feature based on the average pooling result to obtain the global feature; perform convolution operation on the aggregated feature in the temporal dimension and the node feature dimension respectively, and pass the two convolution results through the attention mechanism respectively to obtain the temporal local feature and the node local feature; add the global feature, the temporal local feature, and the node local feature to obtain the second spatiotemporal representation.

[0090] Specifically, the process of obtaining aggregate features includes:

[0091] Perform feature splicing on each time feature representation to obtain a second splicing result; perform one-dimensional convolution on the second splicing result to obtain a second convolution result; and pass the second convolution result through the ReLU activation function to obtain channel features;

[0092] The channel features are multiplied element-by-element by the learnable weight vector and then passed through the Softmax function to obtain the attention weight corresponding to each time feature representation;

[0093] Multiply each attention weight by the corresponding temporal feature representation element by element to obtain multiple fourth multiplication results;

[0094] Sum all the fourth multiplication results to get the aggregate feature.

[0095] Furthermore, the process of obtaining global features includes:

[0096] The aggregated features are average-pooled in the time dimension and feature dimension to obtain time global features and node global features respectively;

[0097] Multiply the transposed time global feature and the node global feature to obtain the fifth multiplication result;

[0098] Perform two-dimensional convolution on the fifth multiplication result to obtain the third convolution result; and pass the third convolution result through the Sigmoid function to obtain the weight matrix;

[0099] The weight matrix is ​​multiplied element-wise by the aggregated features to obtain the global features.

[0100] Furthermore, the process of obtaining temporal local features and node local features includes:

[0101] Convolution operations are performed on the aggregated features in the time dimension and the node feature dimension respectively to obtain the fourth convolution result and the fifth convolution result respectively;

[0102] The fourth convolution result is used as the first query and first value, and the fifth convolution result is used as the first key. The temporal local features are calculated through the attention mechanism.

[0103] The fifth convolution result is used as the second query and second value, and the fourth convolution result is used as the second key. The local features of the node are calculated through the attention mechanism.

[0104] Step S3 first adjusts the weights of the channel dimension through the channel attention mechanism, highlighting important channel variable features and suppressing redundant information. Since global features are obtained through operations such as average pooling, the mutual influence between local features is ignored. Therefore, a query-key-to-interchange dual-tower mutual attention mechanism is designed to perform dynamic feature adjustment and fusion in the time dimension and node feature dimension respectively, realizing multi-level feature optimization selection and obtaining time-local features and node-local features. Finally, the global features, time-local features, and node-local features are added together to obtain a second spatiotemporal representation after the interaction of the time dimension and node feature dimension under global and local conditions.

[0105] S4: Perform a residual connection between the second spatiotemporal representation and each first spatiotemporal representation, and then pass it through a fully connected layer to output the predicted value of the target variable.

[0106] In this embodiment, based on the predicted value of the target variable and the true value of the target variable, a loss function is constructed using the mean absolute error calculation formula, and the loss function is minimized to train the quality prediction method to achieve accurate prediction of the value of the target variable.

[0107] After training, this embodiment learns the mapping relationship from the historical observation values ​​of the process variables and the target variables to the predicted values ​​of the target variables. The mapping relationship can be expressed as:

[0108] ;

[0109] in, Indicates the t +1 The predicted value of the target variable at time; Indicates a mapping relationship; Represents the time from 0 to t Historical observations of the target variable at each moment; Represents the time from 0 to t The historical observation values ​​of the process variables at time m.

[0110] In order to evaluate and verify the performance of the method provided in this embodiment, this embodiment provides input and output data of a debutanizer fractionation process to test the quality prediction method of this embodiment.

[0111] The debutanizer fractionation process in the oil refining process, the debutanizer is a fractionating tower used to recover butane and the above components of natural gas in the shallow cold light hydrocarbon cycle process. The tower takes advantage of the difference in boiling points of different hydrocarbons by heating the mixture and providing precise temperature control in the tower. The debutanizer mainly consists of six parts: a heat exchanger, an overhead condenser, a bottom reboiler, an overhead reflux pump, a separator feed pump and a reflux tank. The purpose of this process is to remove the propane and butane contained in the naphtha stream, hoping to minimize the butane content in the debutanizer bottom material. This example selects the butane content in the debutanizer bottom material as the target variable, and its experimental prediction results are as follows. Figure 2 As shown in the figure, the predicted value of butane content by this method maintains good consistency with the true value of butane content, showing good predictive ability.

[0112] In order to evaluate and verify the generalization ability of the method provided in this embodiment, this embodiment provides the input and output data of the hydrocracking process to test the quality prediction method of this embodiment. This embodiment selects the 10% recovery temperature of the jet fuel in the hydrocracking process as the target variable, and the experimental prediction results are as follows: Figure 3 As shown, the method maintains good consistency between the predicted value of the 10% recovery temperature of aviation fuel and the actual value of the recovery temperature, showing good prediction ability and proving the generalization ability of the method provided in this embodiment.

[0113] The quality prediction method based on spatiotemporal dynamic graph convolution and mutual attention selection provided in this embodiment has the following beneficial effects:

[0114] To address the shortcomings of traditional prediction models in capturing spatiotemporal relationships, this paper captures spatiotemporal features by grouping node features, performing temporal convolution, and aggregating graph structure information. To address the challenges of feature redundancy caused by insufficient feature interaction mechanisms and multi-scale operations, a dual-tower mutual attention mechanism with query-key-value exchange was designed. A multi-level feature selection network based on this dual-tower mutual attention mechanism was constructed. This network performs feature selection fusion across three dimensions: variable, time, and node features, effectively suppressing multi-scale redundancy, highlighting key features, and significantly improving the model's predictive performance.

[0115] The technical features of the above-mentioned embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above-mentioned embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0116] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent application. It should be noted that a person of ordinary skill in the art may make various modifications and improvements without departing from the spirit of the present application, and these modifications and improvements fall within the scope of protection of the present application. Therefore, the scope of protection of the present patent application shall be determined by the appended claims.

Claims

1. A quality prediction method based on spatiotemporal dynamic graph convolution and mutual attention selection, characterized in that: include: S1: Collect historical observations of target variables and process variables in the chemical process, combine the historical observations to obtain a one-dimensional vector, perform dimensionality-upgrading on the time dimension of the one-dimensional vector to obtain a high-dimensional representation of the data; perform feature embedding and grouping on the high-dimensional representation of the data to obtain multiple group features; perform convolution operations on each group feature to obtain the corresponding time feature representation; S2: Build an adaptive topology structure and perform graph convolution on each temporal feature representation based on the adaptive topology structure to obtain the corresponding neighborhood features; Based on the gating mechanism, the temporal feature representation is fused with the corresponding neighborhood features to obtain the corresponding first spatiotemporal representation; S3: Calculate the attention weight based on each temporal feature representation, and aggregate each temporal feature representation based on the attention weight to obtain the aggregated feature; perform average pooling on the aggregated feature in the temporal dimension and the node feature dimension respectively, and weight the aggregated feature based on the average pooling result to obtain the global feature; perform convolution operation on the aggregated feature in the temporal dimension and the node feature dimension respectively, and pass the two convolution results through the attention mechanism respectively to obtain the temporal local feature and the node local feature; add the global feature, the temporal local feature, and the node local feature to obtain the second spatiotemporal representation; The process of obtaining temporal local features and node local features includes: Convolution operations are performed on the aggregated features in the time dimension and the node feature dimension respectively to obtain the fourth convolution result and the fifth convolution result respectively; The fourth convolution result is used as the first query and first value, and the fifth convolution result is used as the first key. The temporal local features are calculated through the attention mechanism. The fifth convolution result is used as the second query and second value, and the fourth convolution result is used as the second key. The local features of the node are calculated through the attention mechanism. S4: Perform a residual connection between the second spatiotemporal representation and each first spatiotemporal representation, and then pass it through a fully connected layer to output the predicted value of the target variable.

2. The quality prediction method based on spatiotemporal dynamic graph convolution and mutual attention selection according to claim 1, characterized in that The chemical process includes a debutanizer fractionation process or a hydrocracking process in an oil refining process; the target variable includes a butane content in a debutanizer bottom material or a 10% recovery temperature of jet fuel in a hydrocracking process.

3. The quality prediction method based on spatiotemporal dynamic graph convolution and mutual attention selection according to claim 1, characterized in that The process of obtaining grouping features includes: Pass the one-dimensional vector through the fully connected layer and add the mapping result to the positional encoding after sine and cosine encoding to obtain a high-dimensional representation of the data; Input the high-dimensional representation of the data into the node feature embedding layer and output the node feature representation; The node feature representations are equally grouped based on the node feature dimensions of the node feature representations to obtain multiple grouping features.

4. The quality prediction method based on spatiotemporal dynamic graph convolution and mutual attention selection according to claim 1, characterized in that Building an adaptive topology includes: Construct a learnable first node feature matrix and a second node feature matrix; The values ​​of each element in the first node feature matrix and the second node feature matrix are random values; The first node feature matrix and the second node feature matrix are multiplied element by element to obtain a first multiplication result; the first multiplication result is passed through a ReLU activation function to obtain a first activation result; the first activation result is passed through a Softmax function to obtain an adaptive topology structure.

5. The quality prediction method based on spatiotemporal dynamic graph convolution and mutual attention selection according to claim 1, characterized in that The process of obtaining the first space-time representation is: For any time feature representation, Multiply the adaptive topological structure and the temporal feature representation element by element to obtain the corresponding neighborhood features; Perform feature splicing on the time feature representation and the corresponding neighborhood feature to obtain a first splicing result; Pass the first splicing result through the convolutional network to obtain a first convolution result; Pass the first convolution result through the Sigmoid function to obtain the corresponding fusion weight; The temporal feature representation is fused with the corresponding neighborhood feature based on the fusion weight to obtain the corresponding first spatiotemporal representation; Traverse all temporal feature representations and calculate the first spatiotemporal representation corresponding to each temporal feature representation.

6. The quality prediction method based on spatiotemporal dynamic graph convolution and mutual attention selection according to claim 5, characterized in that The fusion of temporal feature representation and corresponding neighborhood features based on fusion weights includes: The fusion weight is multiplied by the time feature representation element by element to obtain the second multiplication result; the neighborhood feature is multiplied by the balance number of the fusion weight about 1 element by element to obtain the third multiplication result; The second multiplication result is added to the third multiplication result to obtain the corresponding first space-time representation.

7. The quality prediction method based on spatiotemporal dynamic graph convolution and mutual attention selection according to claim 1, characterized in that The process of obtaining aggregate features includes: Perform feature splicing on each time feature representation to obtain a second splicing result; perform one-dimensional convolution on the second splicing result to obtain a second convolution result; and pass the second convolution result through the ReLU activation function to obtain channel features; The channel features are multiplied element-by-element by the learnable weight vector and then passed through the Softmax function to obtain the attention weight corresponding to each time feature representation; Multiply each attention weight by the corresponding temporal feature representation element by element to obtain multiple fourth multiplication results; Sum all the fourth multiplication results to get the aggregate feature.

8. The quality prediction method based on spatiotemporal dynamic graph convolution and mutual attention selection according to claim 1, characterized in that The process of obtaining global features includes: The aggregated features are average-pooled in the time dimension and feature dimension to obtain time global features and node global features respectively; Multiply the transposed time global feature and the node global feature to obtain the fifth multiplication result; Perform two-dimensional convolution on the fifth multiplication result to obtain the third convolution result; and pass the third convolution result through the Sigmoid function to obtain the weight matrix; The weight matrix is ​​multiplied element-wise by the aggregated features to obtain the global features.

9. The quality prediction method based on spatiotemporal dynamic graph convolution and mutual attention selection according to claim 1, characterized in that Based on the predicted value of the target variable and the true value of the target variable, the loss function is constructed using the mean absolute error calculation formula, and the loss function is minimized to train the quality prediction method.

Images