Multi-source heterogeneous data fusion method based on graph wavelet depth network

Abstract

The invention discloses a multi-source heterogeneous data fusion method based on a graph wavelet depth network. The method comprises the following steps: preprocessing multi-source data; mapping data elements into vertexes of a graph on the basis of relevance and domain knowledge among the preprocessed data sources, and constructing a heterogeneous graph; on the constructed heterogeneous graph, performing multi-scale feature extraction by using graph wavelet transform to obtain local high-frequency and low-frequency features of each node under multiple scales; constructing a graph wavelet multi-source heterogeneous data fusion model, and training the graph wavelet multi-source heterogeneous data fusion model through local high-frequency and low-frequency features of each node under multiple scales; and fusing the features learned by the nodes through the trained graph wavelet multi-source heterogeneous data fusion model, and outputting a fusion result. According to the method, the heterogeneous graph of the multi-source data is constructed, the multi-scale features are extracted by adopting graph wavelet transform, and joint training and fusion are performed through the deep network, so that the accuracy and robustness of heterogeneous data fusion are improved.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method for fusing multi-source heterogeneous data based on graph wavelet depth networks. Background Technology

[0002] In applications such as the Industrial Internet and smart manufacturing, various sensors, devices, and systems continuously generate massive amounts of heterogeneous data, such as real-time sensor data, machine vision images, process parameters, and production logs. These data sources vary greatly in format and structure, making it difficult to directly integrate and use them using traditional single models. For example, different sensor data may have different sampling rates, dimensions, and statistical characteristics, making it difficult to extract effective joint features through simple splicing or shallow fusion.

[0003] Existing multi-source heterogeneous data fusion methods typically employ feature-level or decision-level fusion approaches. For example, for fusing image and structured data, multiple parallel branch networks can be used to extract features from each data source separately before fusion. While these branch network methods can handle data of different modalities, the features extracted by each branch rely on simple concatenation or linear fusion at the back end, failing to fully utilize the structural relationships between the data. Furthermore, the network architecture of this method has poor adaptability to new data sources, making it difficult to dynamically incorporate new heterogeneous information.

[0004] Furthermore, existing research in image fusion has utilized techniques such as wavelet transform for multi-scale feature extraction. For instance, current techniques employ a support degree transform based on wavelet kernels combined with anti-aliasing contour waves to achieve multi-scale directional extraction of image features. However, these methods primarily target pixel-level image fusion, and their wavelet transform designs are specific to conventional image signals, making them unsuitable for general heterogeneous data fusion scenarios. For general multi-source data fusion, current techniques lack a scheme to model the relationships between data as a graph structure and utilize multi-scale filtering for unified processing.

[0005] Furthermore, Graph Convolutional Neural Networks (GCNs) are essentially based on global Fourier transform for graph signal processing, which limits their focus on local graph features. In heterogeneous data fusion tasks, relying solely on GCN methods can easily lead to overly smoothed features and loss of detailed information, making it difficult to obtain rich local multi-scale representations. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a multi-source heterogeneous data fusion method based on graph wavelet deep networks. This method constructs a graph structure representation of multi-source data, extracts multi-scale features using graph wavelet transform, and performs joint training and fusion through deep networks, thereby improving the accuracy and robustness of heterogeneous data fusion and realizing unified processing and comprehensive utilization of different types of data.

[0007] To achieve the above objectives, the technical solution provided by this invention is as follows:

[0008] A multi-source heterogeneous data fusion method based on graph wavelet depth networks includes:

[0009] S1. Perform multi-source data preprocessing;

[0010] S2. Based on the correlation between the preprocessed data sources and domain knowledge, the data elements are mapped to the vertices of the graph to construct a heterogeneous graph;

[0011] S3. On the constructed heterogeneous graph, multi-scale feature extraction is performed using graph wavelet transform to obtain the local high-frequency and low-frequency features of each node at multiple scales.

[0012] S4. Construct a graph wavelet multi-source heterogeneous data fusion model, and train the graph wavelet multi-source heterogeneous data fusion model using the local high-frequency and low-frequency features of each node at multiple scales obtained in step S3.

[0013] S5. The trained graph wavelet multi-source heterogeneous data fusion model is used to fuse the features learned by each node, and finally the fusion result is output.

[0014] Further, step S1 performs multi-source data preprocessing, including:

[0015] Perform operations on multi-source data, including cleaning, denoising, and normalization.

[0016] For different types of data sources, extract their preliminary features or vector representations respectively.

[0017] Further, step S2 constructs a heterogeneous graph, including:

[0018] Each data sample or feature is treated as a node, and related samples from different data sources or within the same data source are connected by edges. The weight of the edges is determined based on the similarity between data or business rules, thus forming a graph adjacency matrix that reflects the needs of heterogeneous data fusion, converting multi-source data into a graph model, i.e., a heterogeneous graph.

[0019] Further, step S3 includes:

[0020] On the constructed heterogeneous graph, a normalized graph Laplacian operator is constructed to form the spectral domain basis required for subsequent filtering.

[0021] Design a multi-scale bandpass filter in the Laplace spectral domain, with each scale corresponding to a different frequency band, without the need to calculate the complete set of eigenvalues ​​and eigenvectors;

[0022] The designed multi-scale bandpass filter is applied to the graph signal, i.e. the node features, and multi-scale decomposition is achieved through graph wavelet transform to obtain the local high-frequency and low-frequency features of each node at multiple scales.

[0023] Furthermore, in step S4, the constructed graph wavelet multi-source heterogeneous data fusion model includes a multi-layer structure; each layer contains a graph wavelet convolutional layer and a nonlinear activation unit for learning and fusing node features; the graph wavelet convolutional layer adaptively extracts key features from the input graph signal through the parameters obtained by training, thereby achieving deep representation of heterogeneous data.

[0024] Compared with existing technologies, the principles and advantages of this technical solution are as follows:

[0025] 1. By constructing a heterogeneous graph, heterogeneous data is uniformly represented as nodes and edges. Nodes correspond to data elements, and edges represent the relationships between data. In implementation, the system preprocesses different types of data, extracts preliminary features or vector representations, and generates an adjacency matrix of the graph model based on the relationships between data. The design of nodes and edges ensures that complex relationships between data sources are effectively captured, providing a unified structured foundation for subsequent feature extraction and fusion. Through a unified graph representation, the diversity of multi-source data is integrated, enhancing the compatibility and consistency of data processing and supporting the efficient implementation of subsequent multi-scale analysis.

[0026] 2. The system utilizes graph wavelet transform to extract local and global features of nodes from the graph structure, and decomposes the signal through multi-scale filtering to capture multi-level information of the data. In implementation, the system designs filters at different scales to decompose the graph signal, generating high-frequency and low-frequency features to enhance feature representation capabilities. This multi-scale analysis is achieved through spectral domain transformation, efficiently processing graph structure data without complex calculations. This provides rich feature inputs for deep fusion, ensuring that both data details and overall trends are fully utilized, thus improving the comprehensiveness of the fusion results.

[0027] 3. A graph wavelet multi-source heterogeneous data fusion model, trained and completed, is used to perform deep learning on multi-scale features, generating joint representations from various data sources. In implementation, the graph wavelet multi-source heterogeneous data fusion model adopts a multi-layer structure, with each layer containing graph wavelet convolutional layers and nonlinear activation units. Parameter optimization is used to mine deep relationships between data. The convolution process adaptively learns the feature representations of nodes and their neighborhoods, generating a unified feature space. Through the nonlinear transformation of the deep network, the limitations of shallow fusion are overcome, ensuring that the fusion result takes into account the diversity and consistency of data sources, providing high-quality feature support for subsequent decision-making tasks.

[0028] 4. Based on the fused feature representation, the system generates a unified feature vector, supporting tasks such as classification, prediction, and anomaly detection. In implementation, the final layer of the graph wavelet multi-source heterogeneous data fusion model integrates multi-source features into a common representation space through a fully connected layer or other fusion layers, generating high-precision decision results. Through an end-to-end processing flow, multi-scale features are transformed into operable outputs, ensuring analytical efficiency and result reliability. Its unified representation design optimizes the performance of downstream tasks, supports diverse decision-making needs, and enhances the practicality of multi-source data fusion.

[0029] 5. The system supports dynamically expanding new data sources and mitigates the impact of noise and missing data through multi-scale analysis and deep learning. In implementation, the graph structure incorporates new nodes and edges through incremental updates, maintaining the model's adaptability. Multi-scale decomposition of the graph wavelet transform and the network's feature learning capabilities effectively filter noise and compensate for missing data, ensuring the stability of the fusion results. Through flexible architecture design and robust processing mechanisms, the system adapts to changes and complexities in data sources, enhancing its reliability and scalability in dynamic environments. Attached Figure Description

[0030] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the services required in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description 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.

[0031] Figure 1 This is a flowchart illustrating the principle of a multi-source heterogeneous data fusion method based on graph wavelet depth networks according to the present invention. Detailed Implementation

[0032] The present invention will be further described below with reference to specific embodiments:

[0033] like Figure 1 As shown in this embodiment, a multi-source heterogeneous data fusion method based on graph wavelet depth networks includes the following steps:

[0034] S1. Perform multi-source data preprocessing, including:

[0035] Perform operations on multi-source data, including cleaning, denoising, and normalization.

[0036] For different types of data sources (such as time-series sensor data, image data, text descriptions, etc.), preliminary features or vector representations are extracted respectively.

[0037] S2. Based on the correlation between the preprocessed data sources and domain knowledge, the data elements are mapped to the vertices of the graph to construct a heterogeneous graph;

[0038] In this step, the process of constructing the heterogeneous graph includes:

[0039] Each data sample or feature is treated as a node, and related samples from different data sources or within the same data source are connected by edges. The weight of the edges is determined based on the similarity between data (such as Euclidean distance of feature vectors, correlation index, etc.) or business rules, thereby forming a graph adjacency matrix that reflects the needs of heterogeneous data fusion, and converting multi-source data into a graph model, i.e., a heterogeneous graph.

[0040] S3. On the constructed heterogeneous graph, multi-scale feature extraction is performed using graph wavelet transform to obtain the local high-frequency and low-frequency features of each node at multiple scales.

[0041] The specific process for this step is as follows:

[0042] On the constructed heterogeneous graph, a normalized graph Laplacian operator is constructed to form the spectral domain basis required for subsequent filtering.

[0043] Design a multi-scale bandpass filter in the Laplace spectral domain, with each scale corresponding to a different frequency band, without the need to calculate the complete set of eigenvalues ​​and eigenvectors;

[0044] The designed multi-scale bandpass filter is applied to the graph signal, i.e. the node features, and multi-scale decomposition is achieved through graph wavelet transform to obtain the local high-frequency and low-frequency features of each node at multiple scales.

[0045] In this step, the graph wavelet transform is a method that extends traditional wavelet analysis to graph-structured signals. By defining a localized wavelet basis on the graph, it enables multi-scale analysis of the graph signal. This transform projects the graph signal from the vertex domain to the spectral domain, allowing for better focus on local structural features.

[0046] S4. Construct a graph wavelet multi-source heterogeneous data fusion model, and train the graph wavelet multi-source heterogeneous data fusion model using the local high-frequency and low-frequency features of each node at multiple scales obtained in step S3.

[0047] In this step, the constructed graph wavelet multi-source heterogeneous data fusion model comprises a multi-layer structure. Each layer includes a graph wavelet convolutional layer and nonlinear activation units for learning and fusing node features. The graph wavelet convolutional layer adaptively extracts key features from the input graph signal using trained parameters (which can be optimized through backpropagation), thereby achieving deep representation of heterogeneous data. This step enables the graph wavelet multi-source heterogeneous data fusion model to comprehensively utilize the internal features and neighborhood structure information of nodes, achieving effective encoding of multi-dimensional features.

[0048] S5. The trained graph wavelet multi-source heterogeneous data fusion model is used to fuse the features learned by each node, and finally the fusion result is output.

[0049] The method described in this embodiment has the following advantages:

[0050] 1. By constructing a heterogeneous graph, heterogeneous data is uniformly represented as nodes and edges. Nodes correspond to data elements, and edges represent the relationships between data. In implementation, the system preprocesses different types of data, extracts preliminary features or vector representations, and generates an adjacency matrix of the graph model based on the relationships between data. The design of nodes and edges ensures that complex relationships between data sources are effectively captured, providing a unified structured foundation for subsequent feature extraction and fusion. Through a unified graph representation, the diversity of multi-source data is integrated, enhancing the compatibility and consistency of data processing and supporting the efficient implementation of subsequent multi-scale analysis.

[0051] 2. The system utilizes graph wavelet transform to extract local and global features of nodes from the graph structure, and decomposes the signal through multi-scale filtering to capture multi-level information of the data. In implementation, the system designs filters at different scales to decompose the graph signal, generating high-frequency and low-frequency features to enhance feature representation capabilities. This multi-scale analysis is achieved through spectral domain transformation, efficiently processing graph structure data without complex calculations. This provides rich feature inputs for deep fusion, ensuring that both data details and overall trends are fully utilized, thus improving the comprehensiveness of the fusion results.

[0052] 3. A graph wavelet multi-source heterogeneous data fusion model, trained and completed, is used to perform deep learning on multi-scale features, generating joint representations from various data sources. In implementation, the graph wavelet multi-source heterogeneous data fusion model adopts a multi-layer structure, with each layer containing graph wavelet convolutional layers and nonlinear activation units. Parameter optimization is used to mine deep relationships between data. The convolution process adaptively learns the feature representations of nodes and their neighborhoods, generating a unified feature space. Through the nonlinear transformation of the deep network, the limitations of shallow fusion are overcome, ensuring that the fusion result takes into account the diversity and consistency of data sources, providing high-quality feature support for subsequent decision-making tasks.

[0053] 4. Based on the fused feature representation, the system generates a unified feature vector, supporting tasks such as classification, prediction, and anomaly detection. In implementation, the final layer of the graph wavelet multi-source heterogeneous data fusion model integrates multi-source features into a common representation space through a fully connected layer or other fusion layers, generating high-precision decision results. Through an end-to-end processing flow, multi-scale features are transformed into operable outputs, ensuring analytical efficiency and result reliability. Its unified representation design optimizes the performance of downstream tasks, supports diverse decision-making needs, and enhances the practicality of multi-source data fusion.

[0054] 5. The system supports dynamically expanding new data sources and mitigates the impact of noise and missing data through multi-scale analysis and deep learning. In implementation, the graph structure incorporates new nodes and edges through incremental updates, maintaining the model's adaptability. Multi-scale decomposition of the graph wavelet transform and the network's feature learning capabilities effectively filter noise and compensate for missing data, ensuring the stability of the fusion results. Through flexible architecture design and robust processing mechanisms, the system adapts to changes and complexities in data sources, enhancing its reliability and scalability in dynamic environments.

[0055] The above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Therefore, any changes made in accordance with the shape and principle of the present invention should be covered within the protection scope of the present invention.

Claims

1. A method for fusing multi-source heterogeneous data based on graph wavelet depth networks, characterized in that, include: S1. Perform multi-source data preprocessing; S2. Based on the correlation between the preprocessed data sources and domain knowledge, the data elements are mapped to the vertices of the graph to construct a heterogeneous graph; S3. On the constructed heterogeneous graph, multi-scale feature extraction is performed using graph wavelet transform to obtain the local high-frequency and low-frequency features of each node at multiple scales. S4. Construct a graph wavelet multi-source heterogeneous data fusion model, and train the graph wavelet multi-source heterogeneous data fusion model using the local high-frequency and low-frequency features of each node at multiple scales obtained in step S3. S5. The trained graph wavelet multi-source heterogeneous data fusion model is used to fuse the features learned by each node, and finally the fusion result is output.

2. The multi-source heterogeneous data fusion method based on graph wavelet depth networks according to claim 1, characterized in that, Step S1 involves multi-source data preprocessing, including: Perform operations on multi-source data, including cleaning, denoising, and normalization. For different types of data sources, extract their preliminary features or vector representations respectively.

3. The multi-source heterogeneous data fusion method based on graph wavelet depth networks according to claim 1, characterized in that, Step S2 constructs a heterogeneous graph, including: Each data sample or feature is treated as a node, and related samples from different data sources or within the same data source are connected by edges. The weight of the edges is determined based on the similarity between data or business rules, thus forming a graph adjacency matrix that reflects the needs of heterogeneous data fusion, converting multi-source data into a graph model, i.e., a heterogeneous graph.

4. The multi-source heterogeneous data fusion method based on graph wavelet depth networks according to claim 1, characterized in that, Step S3 includes: On the constructed heterogeneous graph, a normalized graph Laplacian operator is constructed to form the spectral domain basis required for subsequent filtering. Design a multi-scale bandpass filter in the Laplace spectral domain, with each scale corresponding to a different frequency band, without the need to calculate the complete set of eigenvalues ​​and eigenvectors; The designed multi-scale bandpass filter is applied to the graph signal, i.e. the node features, and multi-scale decomposition is achieved through graph wavelet transform to obtain the local high-frequency and low-frequency features of each node at multiple scales.

5. The multi-source heterogeneous data fusion method based on graph wavelet depth networks according to claim 1, characterized in that, In step S4, the constructed graph wavelet multi-source heterogeneous data fusion model includes a multi-layer structure; each layer contains a graph wavelet convolutional layer and a nonlinear activation unit, which are used to learn and fuse node features. Graph wavelet convolutional layers adaptively extract key features from input graph signals using trained parameters, thereby achieving deep representation of heterogeneous data.

Images