Method and system for heterogeneous time series data prediction based on multi-relational graph attention network

Abstract

This invention relates to the field of heterogeneous time-series data prediction technology, and provides a method and system for heterogeneous time-series data prediction based on a multi-relationship graph attention network. The method extracts temporal features from preprocessed multivariate time-series data using a gated recurrent network based on a time-pattern attention mechanism; extracts text embedding features from text data using a pre-trained text embedding model; constructs object relationship graphs and category relationship graphs; and uses a graph attention network to perform attention weighting and aggregation on the features of adjacent nodes in the object relationship graph and category relationship graph, respectively, to obtain object-level relationship features and category-level relationship features. The temporal features, text embedding features, object-level relationship features, and category-level relationship features are fused and input into a multilayer perceptron to obtain the prediction result for the target object. This invention improves the accuracy of heterogeneous time-series data prediction by jointly processing multivariate time-series data, text data, object-level relationship data, and category-level relationship data.

Description

Technical Field

[0001] This invention relates to the field of heterogeneous time series data prediction technology, and in particular to a method and system for heterogeneous time series data prediction based on multi-relation graph attention networks. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] With the development of computer technology, time-series data, which has the characteristics of temporal continuity, diverse sources, and structural differences, is widely used in equipment status monitoring, network operation analysis, traffic status analysis, and other data processing scenarios.

[0004] Existing time series data processing methods typically employ autoregressive models, recurrent neural networks, or long short-term memory networks to model a single numerical sequence. While these methods can extract certain patterns of temporal variation, they are prone to problems such as insufficient extraction of local time series features, unclear distinction of importance between different time steps, and overlapping of long and short-term features when the input data simultaneously contains multiple variables, multiple time scales, and nonlinear dynamic relationships.

[0005] Furthermore, the runtime logs, event records, status descriptions, or other textual data corresponding to the time-series numerical data contain semantic information that reflects the state of the target object. Existing methods typically encode textual data and numerical time-series data separately, lacking an effective cross-modal alignment mechanism. This results in inconsistent feature distributions generated by different data modalities, making it difficult to effectively fuse them within a unified feature space.

[0006] Furthermore, existing data processing methods based on graph neural networks typically only model object relationships at a single level, such as constructing a graph structure based solely on direct connections between objects. In other words, they only model object-level relationships, ignoring category-level relationships of the target object. They do not jointly model dependency features at different levels, and it is difficult to dynamically adjust the influence weights of adjacent nodes on the target node based on the node state, thus limiting the utilization of graph structure information. Summary of the Invention

[0007] To overcome the shortcomings of the prior art, this invention provides a heterogeneous time series data prediction method and system based on multi-relation graph attention network, aiming to solve the problems in the prior art that it is difficult to effectively distinguish the importance of features at different time steps, difficult to uniformly represent numerical time series features and text features, and difficult to collaboratively model relation data at different levels.

[0008] To achieve the above objectives, the technical solution provided by the present invention is as follows: In a first aspect, the present invention provides a method for predicting heterogeneous temporal data based on a multi-relation graph attention network, comprising: Acquire multivariate time-series data and text data of multiple target objects, and preprocess them; A gated recurrent network based on a time-pattern attention mechanism is used to extract features from preprocessed multivariate time-series data to obtain time-series features; and a pre-trained text embedding model is used to encode text data to obtain text embedding features. Construct an object relationship graph based on the relationships between multiple target objects, and construct a category relationship graph based on the relationships between the categories to which multiple target objects belong; The graph attention network is used to perform attention weighting and aggregation on the features of adjacent nodes in the object relationship graph and the category relationship graph respectively to obtain object-level relationship features and category-level relationship features; The temporal features, text embedding features, object-level relationship features, and category-level relationship features are fused to obtain fused features; the multilayer perceptron outputs the prediction result of the target object to be predicted based on the fused features.

[0009] As a further technical solution, the multivariate time series data of the multiple target objects is multivariate time series data of multiple target objects collected within a preset historical time window. The preset historical time window includes multiple consecutive time slices, and the number of consecutive time slices is determined according to the data sampling period. The text data is text data that is temporally correlated with the multivariate time series data.

[0010] As a further technical solution, the preprocessing includes data deduplication, missing value imputation, outlier correction, standardization, SMOTE algorithm sample synthesis, and improved K-means clustering.

[0011] As a further technical solution, the objective function in the improved K-means clustering is expressed as: ; in, Indicates the first One cluster, Indicates the first Cluster centers of each cluster. Represents sample points, Indicates the weighting coefficient. Indicates the first The density of each cluster, Indicates the number of cluster centers. This represents the square of the Euclidean distance.

[0012] As a further technical solution, a gated recurrent network based on a time-pattern attention mechanism is used to extract features from the preprocessed multivariate time-series data, resulting in time-series features, represented as: ; ; ; ; in, express Input characteristics at time step express The hidden state at all times express The hidden state at all times Indicates the first The unnormalized time attention score corresponding to each time slice Represents the temporal attention parameter vector transpose, This represents the feature transformation matrix in the temporal attention mechanism. This represents the bias vector in the temporal attention mechanism. Indicates the first Normalized temporal attention weights corresponding to each time slice Indicates the length of the time window. Indicates the first The unnormalized time attention score corresponding to each time slice Indicates the time slice index. This represents the temporal feature obtained by weighted aggregation of the hidden states of each time slice through attention.

[0013] As a further technical solution, a graph attention network is used to perform attention weighting and aggregation on the features of adjacent nodes in the object relationship graph to obtain object-level relationship features, represented as follows: ; ; in, Represents object nodes Updated features This represents the sigmoid activation function. Represents the normalized object node With neighboring nodes Attention weights between them Represents the transformation matrix of object features. Represents the adjacent nodes of an object node Features To represent object nodes The set of adjacent nodes, Represents a collection of object nodes. Represents object-level relational characteristics; The graph attention network is used to perform attention weighting and aggregation on the features of adjacent nodes in the category relationship graph to obtain category-level relationship features, represented as follows: ; ; in, Represents category nodes Updated features This represents the sigmoid activation function. Represents the category nodes after normalization. With neighboring nodes Attention weights Represents the category feature transformation matrix. Represents the neighboring nodes of a category node Features Represents category nodes The set of adjacent nodes, This represents the set of nodes representing the category to which the object belongs. This indicates category-level relationship features.

[0014] Secondly, a heterogeneous temporal data prediction system based on multi-relation graph attention networks includes: Acquisition module: Used to acquire multivariate time-series data and text data of multiple target objects, and preprocess them; Feature extraction module: Used to extract features from preprocessed multivariate time series data using a gated recurrent network based on a time pattern attention mechanism to obtain time series features; and to encode text data using a pre-trained text embedding model to obtain text embedding features; Relationship Graph Construction Module: Used to construct an object relationship graph based on the relationships between multiple target objects, and to construct a category relationship graph based on the relationships between the categories to which multiple target objects belong; Relationship feature extraction module: used to perform attention weighting and aggregation on the features of adjacent nodes in the object relationship graph and the category relationship graph using a graph attention network, respectively, to obtain object-level relationship features and category-level relationship features; The result prediction module is used to fuse temporal features, text embedding features, object-level relationship features, and category-level relationship features to obtain fused features; the multilayer perceptron outputs the prediction result of the target object to be predicted based on the fused features.

[0015] Thirdly, the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps in the heterogeneous time-series data prediction method based on a multi-relation graph attention network as described in the first aspect above.

[0016] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the heterogeneous time-series data prediction method based on a multi-relation graph attention network as described in the first aspect above.

[0017] Fifthly, the present invention provides a computer program product, including software code, wherein the program in the software code performs the steps in the heterogeneous time-series data prediction method based on multi-relation graph attention network described in the first aspect above.

[0018] Compared with the prior art, the beneficial effects of the present invention are: This invention enhances the collaborative representation capabilities among features at different time steps, modalities, and hierarchical levels by jointly processing multivariate time-series data, text data, object-level relationship data, and category-level relationship data; extracts dynamic features of key time slices through a gated recurrent network TPA-GRU based on a time-pattern attention mechanism; performs weighted aggregation of object-level and category-level relationships through a graph attention network; and improves the feature representation completeness and state prediction capabilities of heterogeneous time-series data by fusing multi-source features. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating the heterogeneous time-series data prediction method based on a multi-relation graph attention network provided in an embodiment of the present invention. Detailed Implementation

[0021] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0022] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments of the present invention. The terms “comprising” and “having”, and any variations thereof, are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0023] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0024] All data acquisition in this embodiment is carried out in accordance with laws and regulations and with user consent, and the data is used legally.

[0025] Example 1 In the field of heterogeneous time-series data prediction, existing methods typically suffer from problems such as a single information source, difficulty in effectively integrating textual features and time-series numerical features, and insufficient modeling of object-level and category-level relationships. This results in prediction models struggling to comprehensively extract effective features from different time steps, data modalities, and hierarchical relationships. In this embodiment, by constructing object relationship graphs and category relationship graphs, and utilizing deep learning models based on Time Pattern Attention Mechanism Gated Recurrent Network (TPA-GRU) and Graph Attention Network (GAT), along with pre-trained text embedding technology, deep fusion of multi-source heterogeneous data and prediction of target object states are achieved.

[0026] like Figure 1 As shown, this embodiment provides a method for predicting heterogeneous time-series data based on a multi-relationship graph attention network, including: Step 1: Obtain multivariate time-series data and text data of multiple target objects, and preprocess them; Step 2: A gated recurrent network based on a time-pattern attention mechanism is used to extract features from the preprocessed multivariate time-series data to obtain time-series features; and a pre-trained text embedding model is used to encode the text data to obtain text embedding features. Step 3: Construct an object relationship diagram based on the relationships between multiple target objects, and construct a category relationship diagram based on the relationships between the categories to which the multiple target objects belong; Step 4: Use a graph attention network to perform attention weighting and aggregation on the features of adjacent nodes in the object relationship graph and the category relationship graph respectively to obtain object-level relationship features and category-level relationship features; Step 5: Fuse the temporal features, text embedding features, object-level relationship features, and category-level relationship features to obtain fused features; the multilayer perceptron outputs the prediction result of the target object to be predicted based on the fused features.

[0027] The following content provides a more detailed introduction to the heterogeneous time-series data prediction method based on multi-relation graph attention network provided in this embodiment.

[0028] In step 1, multivariate time-series data and text data of multiple target objects are obtained and preprocessed.

[0029] In one specific implementation, the multivariate time-series data of the multiple target objects is multivariate time-series data of multiple target objects collected within a preset historical time window. The preset historical time window includes multiple consecutive time slices, and the number of consecutive time slices is determined according to the data sampling period. The text data is text data that is temporally correlated with the multivariate time-series data.

[0030] Furthermore, the multivariate time-series data and text data of multiple target objects can be multivariate time-series data and text data of multiple target devices. Specifically, the multivariate time-series data includes temperature, vibration, current, voltage, speed and power; the text data includes operation logs, alarm records and maintenance records.

[0031] In one specific implementation, to ensure data quality, the multivariate time-series data of the collected target device is preprocessed. The preprocessing includes data deduplication, missing value imputation, outlier correction, standardization, SMOTE algorithm sample synthesis, and improved K-means clustering.

[0032] For minority class samples, such as abnormal state samples or rare state samples, the SMOTE algorithm is used for sample synthesis, as follows: ; in, This represents a new minority class sample generated using the SMOTE algorithm. Represents the original minority class samples, This represents the nearest neighbor of the original minority class sample in the feature space. This represents the weight parameter, which controls the position of the new minority class sample on the line connecting the original minority class sample and its nearest neighbor. This represents the sample density of the original minority class samples in a local region, used to normalize the interpolation step size and avoid generating samples that deviate excessively from the true distribution in sparse regions.

[0033] The sample synthesis step of the SMOTE algorithm described above can increase the number of minority class samples and mitigate the impact of class imbalance on model training.

[0034] In one specific implementation, the data preprocessing further includes selecting representative samples using improved K-means clustering to reduce data redundancy and improve training efficiency; the objective function in the improved K-means clustering is expressed as: ; in, Indicates the first One cluster, Indicates the first Cluster centers of each cluster. Represents sample points, Represents the balance coefficient. Indicates the first The density of each cluster, Indicates the number of cluster centers. This represents the square of the Euclidean distance.

[0035] In this embodiment, the balance coefficient The value of is determined through grid search and cross-validation. The specific process is as follows: Set a set of candidate values ​​for β on the validation set, typically in the range of [0, 0.001, 0.01, 0.05, 0.1, 0.5, 1.0]. Perform a complete clustering, sample selection, model training, validation and evaluation process for each β value; The F1 score of the model on the validation set is used as the main evaluation metric, supplemented by training convergence speed and inference latency as auxiliary considerations; the β value with the best overall performance is selected as the final hyperparameter.

[0036] The improved K-means clustering described above minimizes the objective function. Select the best cluster center to make the sample points as close to the center as possible, while selecting more samples in sparse areas.

[0037] In step 2, a gated recurrent network based on a time-pattern attention mechanism is used to extract features from the preprocessed multivariate time-series data to obtain time-series features. The text data is then encoded using a pre-trained text embedding model to obtain text embedding features. .

[0038] Furthermore, a gated recurrent network based on a time-pattern attention mechanism is used to extract features from the preprocessed multivariate time-series data to obtain time-series features. , represented as: ; ; ; ; in, express Input characteristics at time step express The hidden state at all times express The hidden state at all times Indicates the first The unnormalized time attention score corresponding to each time slice Represents the temporal attention parameter vector transpose, This represents the feature transformation matrix in the temporal attention mechanism. This represents the bias vector in the temporal attention mechanism. Indicates the first Normalized temporal attention weights corresponding to each time slice Indicates the length of the time window. Indicates the first The unnormalized time attention score corresponding to each time slice Indicates the time slice index. This represents the temporal feature obtained by weighted aggregation of the hidden states of each time slice through attention.

[0039] Furthermore, the pre-trained text embedding model can be a text encoding model based on the Transformer structure.

[0040] In step 4, an object relationship graph is constructed based on the relationships between multiple target objects, and a category relationship graph is constructed based on the relationships between the categories to which multiple target objects belong.

[0041] Furthermore, the object relationship diagram is represented as follows: ; in, Represents a collection of object nodes. Represents the set of edges between object nodes.

[0042] Furthermore, the category relationship diagram is represented as follows: ; in, This represents the set of nodes representing the category to which the object belongs. This represents the set of edges between nodes belonging to the same category as the object.

[0043] In this embodiment, when constructing the object relationship graph, the nodes can be devices, and the edge weights are determined based on the temporal correlation between different devices. In one implementation, when the temporal correlation between two devices is greater than a preset threshold, a connection edge is established between the corresponding device nodes.

[0044] In this embodiment, when constructing the category relationship diagram, its nodes can be equipment categories. Equipment categories can be determined based on equipment type, function type, or production unit to which they belong. For example, motor equipment, pump equipment, compressor equipment, and conveying equipment can be divided into different equipment categories.

[0045] Furthermore, the edge weights in the object relationship graph are determined based on the temporal correlation between target objects; the edge weights in the category relationship graph are determined based on the category association between target objects.

[0046] In step 5, a graph attention network is used to perform attention weighting and aggregation on the features of adjacent nodes in the object relationship graph and the category relationship graph to obtain object-level relationship features and category-level relationship features.

[0047] For nodes and its adjacent nodes First, calculate the attention weights: ; in, Represents a node With neighboring nodes Attention weights between them Indicates the attention parameter. Represents the characteristic transformation matrix. Represents a node Features Indicates adjacent nodes Features This indicates a splicing operation.

[0048] Attention weights Normalize: ; in, Represents the nodes after normalization. With neighboring nodes Attention weights between them Represents a node With neighboring nodes Attention weights between them Represents a node With neighboring nodes Attention weights between them Represents a node The set of adjacent nodes.

[0049] node Features through adjacent nodes Weighted aggregation is represented as: ; in, For nodes Updated features This represents the sigmoid activation function. Indicates adjacent nodes Features Represents the characteristic transformation matrix. For nodes The set of adjacent nodes, For nodes after normalization With neighboring nodes Attention weights between them.

[0050] By analyzing the object relationship diagram Graph attention processing is performed to obtain object-level relation features. ; through the category relationship diagram Graph attention processing is performed to obtain category-level relationship features. GAT can dynamically adjust node weights based on the degree of feature correlation between different nodes, enabling multi-level modeling of object-level and category-level dependencies.

[0051] In one specific implementation, a graph attention network is used to analyze the object relationship graph. The features of adjacent nodes are weighted and aggregated by attention to obtain object-level relationship features. , is represented as: ; ; in, Represents object nodes Updated features This represents the sigmoid activation function. Represents the normalized object node With neighboring nodes Attention weights between them Represents the transformation matrix of object features. Represents the adjacent nodes of an object node Features To represent object nodes The set of adjacent nodes, Represents a collection of object nodes. Represents object-level relationship characteristics.

[0052] In one specific implementation, a graph attention network is used to process the category relationship graph. The features of adjacent nodes are weighted and aggregated by attention to obtain category-level relationship features. , is represented as: ; ; in, Represents category nodes Updated features This represents the sigmoid activation function. Represents the category nodes after normalization. With neighboring nodes Attention weights Represents the category feature transformation matrix. Represents the neighboring nodes of a category node Features Represents category nodes The set of adjacent nodes, This represents the set of nodes representing the category to which the object belongs. This indicates category-level relationship features.

[0053] In this embodiment, the device node features and device category node features are updated by a graph attention network to obtain device-level relationship features and device category-level relationship features, which represent the operational associations between different devices and the hierarchical relationships between different device categories.

[0054] In step 6, the temporal features, text embedding features, object-level relationship features, and category-level relationship features are fused to obtain fused features; the multilayer perceptron outputs the prediction result of the target object to be predicted based on the fused features.

[0055] Furthermore, the fusion feature is represented as: ; in, Representing temporal characteristics, Represents text embedding features, Represents object-level relational characteristics. This indicates category-level relationship features.

[0056] Furthermore, the multilayer perceptron outputs the prediction result of the target object based on the fused features, which is represented as: ; in, Indicates fusion features, This represents a multilayer perceptron. This represents the model's predicted output.

[0057] In this embodiment, the model prediction output is the equipment operating status, which can include normal status, slightly abnormal status, and severely abnormal status. The equipment operating status can also be classified according to the type of abnormality into temperature abnormality, vibration abnormality, current abnormality, and comprehensive abnormality.

[0058] For example, when the vibration and temperature data of the target device continue to rise over multiple consecutive sampling periods, and the corresponding text descriptions of bearing noise or temperature rise appear in the equipment maintenance records, and other devices connected to the target device also exhibit abnormal vibration, the prediction model integrates the device's own time series features, text embedding features, device relationship features, and device category relationship features to output the probability that the target device will be in an abnormal state in the next sampling period.

[0059] In this embodiment, a hybrid loss method is used for model optimization during model training. Including classification loss and alignment loss ; The classification loss is calculated by comparing the model's predicted output with the true label. The alignment loss is obtained by calculating the cosine similarity between temporal features and text embedding features, and is expressed as: ; in, Indicates alignment loss. Representing temporal characteristics, This represents text embedding features.

[0060] This invention optimizes the model's predictive performance through cross-validation and hyperparameter tuning, including learning rate, batch size, training epochs, time window length, and the number of graph attention network layers.

[0061] In addition, the heterogeneous time-series data prediction method based on multi-relation graph attention network of the present invention can be widely applied to time-series data prediction scenarios such as network operation analysis, traffic status analysis, and financial risk control analysis.

[0062] Example 2 This embodiment provides a heterogeneous time-series data prediction system based on a multi-relation graph attention network, including: Acquisition module: Used to acquire multivariate time-series data and text data of multiple target objects, and preprocess them; Feature extraction module: Used to extract features from preprocessed multivariate time series data using a gated recurrent network based on a time pattern attention mechanism to obtain time series features; and to encode text data using a pre-trained text embedding model to obtain text embedding features; Relationship Graph Construction Module: Used to construct an object relationship graph based on the relationships between multiple target objects, and to construct a category relationship graph based on the relationships between the categories to which multiple target objects belong; Relationship feature extraction module: used to perform attention weighting and aggregation on the features of adjacent nodes in the object relationship graph and the category relationship graph using a graph attention network, respectively, to obtain object-level relationship features and category-level relationship features; The result prediction module is used to fuse temporal features, text embedding features, object-level relationship features, and category-level relationship features to obtain fused features; the multilayer perceptron outputs the prediction result of the target object to be predicted based on the fused features.

[0063] It should be noted that the acquisition module, feature extraction module, temporal feature extraction module, relationship graph construction module, relationship feature extraction module, and result prediction module mentioned above correspond to steps 1 to 5 in Embodiment 1. The examples and application scenarios implemented by the above modules and their corresponding steps are the same, but they are not limited to the content disclosed in Embodiment 1. It should be noted that the above modules, as part of the system, can be executed in a computer system such as a set of computer-executable instructions.

[0064] The descriptions of each embodiment in the above embodiments have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0065] The proposed system can be implemented in other ways. For example, the system embodiments described above are merely illustrative, and the division of modules described above is only a logical functional division. In actual implementation, there may be other division methods. For example, multiple modules may be combined or integrated into another system, or some features may be ignored or not executed.

[0066] Example 3 This embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the heterogeneous time-series data prediction method based on a multi-relation graph attention network as described in Embodiment 1 above.

[0067] It should be understood that in this embodiment, the processor can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc.

[0068] Memory may include read-only memory and random access memory, and provides instructions and data to the processor. A portion of memory may also include non-volatile random access memory. For example, memory may also store information about the device type.

[0069] In the implementation process, each step of the above method can be completed by the integrated logic circuits in the processor hardware or by software instructions.

[0070] The method in Embodiment 1 can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor. The software modules can reside in readily available storage media in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory; the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, a detailed description is not provided here.

[0071] Those skilled in the art will recognize that the units and algorithm steps described in conjunction with the embodiments herein can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0072] Example 4 This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps in the heterogeneous time-series data prediction method based on a multi-relation graph attention network as described in Embodiment 1 above.

[0073] Example 5 This embodiment provides a computer program product, including software code, wherein the program in the software code executes the steps in the heterogeneous time-series data prediction method based on multi-relation graph attention network described in Embodiment 1 above.

[0074] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for predicting heterogeneous temporal data based on multi-relation graph attention networks, characterized in that, include: Acquire multivariate time-series data and text data of multiple target objects, and preprocess them; A gated recurrent network based on a time-pattern attention mechanism is used to extract features from preprocessed multivariate time-series data to obtain time-series features; The text data is encoded using a pre-trained text embedding model to obtain text embedding features; Construct an object relationship graph based on the relationships between multiple target objects, and construct a category relationship graph based on the relationships between the categories to which multiple target objects belong; A graph attention network is used to perform attention weighting and aggregation on the features of adjacent nodes in the object relationship graph and the category relationship graph, respectively, to obtain object-level relationship features and category-level relationship features; By fusing temporal features, text embedding features, object-level relationship features, and category-level relationship features, a fused feature is obtained. The multilayer perceptron outputs the prediction result of the target object based on the fused features.

2. The heterogeneous time-series data prediction method based on multi-relation graph attention network as described in claim 1, characterized in that, The multivariate time-series data of the multiple target objects are multivariate time-series data of the multiple target objects collected within a preset historical time window. The preset historical time window includes multiple consecutive time slices, and the number of consecutive time slices is determined according to the data sampling period. The text data is text data that is temporally correlated with the multivariate time-series data. The multivariate time-series data includes temperature, vibration, current, voltage, rotational speed, and power; the text data includes operation logs, alarm records, and maintenance records.

3. The heterogeneous time-series data prediction method based on multi-relation graph attention network as described in claim 1, characterized in that, The preprocessing includes data deduplication, missing value imputation, outlier correction, standardization, SMOTE algorithm sample synthesis, and improved K-means clustering.

4. The heterogeneous time-series data prediction method based on multi-relation graph attention network as described in claim 3, characterized in that, The objective function in the improved K-means clustering is expressed as: ； in, Indicates the first One cluster, Indicates the first Cluster centers of each cluster. Represents sample points, Indicates the weighting coefficient. Indicates the first The density of each cluster, Indicates the number of cluster centers. This represents the square of the Euclidean distance.

5. The heterogeneous time-series data prediction method based on multi-relation graph attention network as described in claim 1, characterized in that, A gated recurrent network based on a time-pattern attention mechanism is used to extract features from the preprocessed multivariate time-series data, resulting in time-series features, represented as: ； ； ； ； in, express Input characteristics at time step express The hidden state at all times express The hidden state at all times Indicates the first The unnormalized time attention score corresponding to each time slice Represents the temporal attention parameter vector transpose, This represents the feature transformation matrix in the temporal attention mechanism. This represents the bias vector in the temporal attention mechanism. Indicates the first Normalized temporal attention weights corresponding to each time slice Indicates the length of the time window. Indicates the first The unnormalized time attention score corresponding to each time slice Indicates the time slice index. This represents the temporal feature obtained by weighted aggregation of the hidden states of each time slice through attention.

6. The heterogeneous time-series data prediction method based on multi-relation graph attention network as described in claim 1, characterized in that, The graph attention network is used to perform attention weighting and aggregation on the features of adjacent nodes in the object relationship graph to obtain object-level relationship features, represented as follows: ； ； in, Represents object nodes Updated features This represents the sigmoid activation function. Represents the normalized object node With neighboring nodes Attention weights between them Represents the transformation matrix of object features. Represents the adjacent nodes of an object node Features To represent object nodes The set of adjacent nodes, Represents a collection of object nodes. Represents object-level relational characteristics; The graph attention network is used to perform attention weighting and aggregation on the features of adjacent nodes in the category relationship graph to obtain category-level relationship features, represented as follows: ； ； in, Represents category nodes Updated features This represents the sigmoid activation function. Represents the category nodes after normalization. With neighboring nodes Attention weights Represents the category feature transformation matrix. Represents the neighboring nodes of a category node Features Represents category nodes The set of adjacent nodes, This represents the set of nodes representing the category to which the object belongs. This indicates category-level relationship features.

7. A heterogeneous time-series data prediction system based on multi-relation graph attention networks, characterized in that, include: Acquisition module: Used to acquire multivariate time-series data and text data of multiple target objects, and preprocess them; Feature extraction module: Used to extract features from preprocessed multivariate time series data using a gated recurrent network based on a time pattern attention mechanism, to obtain time series features; The text data is encoded using a pre-trained text embedding model to obtain text embedding features; Relationship Graph Construction Module: Used to construct an object relationship graph based on the relationships between multiple target objects, and to construct a category relationship graph based on the relationships between the categories to which multiple target objects belong; Relationship feature extraction module: used to perform attention weighting and aggregation on the features of adjacent nodes in the object relationship graph and the category relationship graph using a graph attention network, respectively, to obtain object-level relationship features and category-level relationship features; The result prediction module is used to fuse time-series features, text embedding features, object-level relationship features, and category-level relationship features to obtain fused features. The multilayer perceptron outputs the prediction result of the target object based on the fused features.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the heterogeneous time-series data prediction method based on multi-relation graph attention network as described in any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the heterogeneous time-series data prediction method based on multi-relation graph attention network as described in any one of claims 1-6.

10. A computer program product, comprising software code, characterized in that, The program in the software code executes the steps in the heterogeneous time-series data prediction method based on multi-relation graph attention network according to any one of claims 1-6.

Images