A method and system for robust data reconstruction of marine environment based on graph neural network

Abstract

The present invention relates to the field of marine environment technology, and in particular to a method and system for robust data reconstruction of marine environment based on graph neural network. The method comprises acquiring marine environment data; performing data preprocessing on the acquired marine environment data; constructing a graph structure based on the geographical location information in the marine environment data; establishing a marine environment robust data reconstruction model based on graph neural network; and identifying the marine environment data using the marine environment robust data reconstruction model to obtain an identification result. The present invention introduces a structure combining a feature completion network with a denoising network, so that the model can automatically learn deep-level correlation relationships in marine data with high noise and feature loss, and capture key node features through a dynamic attention mechanism, thereby effectively improving the accuracy and robustness of feature representation.

Description

Technical Field

[0001] The present invention relates to the field of marine environment technology, and in particular to a method and system for robust data reconstruction of a marine environment based on a graph neural network. Background Art

[0002] With the development of marine digital twin technology, it has become crucial to simulate and optimize equipment and processes in the marine environment. However, there are significant noise and data missing problems in marine data, mainly due to harsh marine environment, sensor failure, data transmission delay, etc. These data problems have led to challenges in the real-time and accuracy of the twin system, especially in the key links of prediction and analysis, where the demand for data robustness and integrity is particularly prominent.

[0003] Traditional graph neural networks (GNNs) also face certain limitations when dealing with data reconstruction tasks in complex marine environments, especially in dealing with high noise and large amounts of missing data. Existing methods often rely on graph structure preprocessing and simple feature aggregation to reconstruct node representations, but the effect is limited when faced with dynamic environments and noise interference. In addition, with the rapid growth of marine digital twin data, the system needs to efficiently process and analyze large amounts of data, which poses a huge challenge to the processing speed and model stability of traditional GNNs. For example:

[0004] ① During the collection process, marine environmental data is affected by ocean dynamic changes, sensor performance and environmental noise, resulting in a high proportion of noise and missing data. Traditional methods are difficult to effectively complete and purify these data, affecting the authenticity and accuracy of the data.

[0005] ② The traditional graph structure construction method based on geographic location information has limitations in the expression of dynamic multi-dimensional features. It cannot fully characterize the complex correlations between marine equipment or monitoring points, affecting the comprehensiveness and depth of data modeling.

[0006] ③ The current ocean data reconstruction method lacks full utilization of multi-level and heterogeneous features, and cannot effectively distinguish important features from noise during denoising, resulting in insufficient robustness of the data in dealing with environmental complexity and poor quality of the reconstructed data.

[0007] ④ Marine environmental data lacks a robust feature representation mechanism under heterogeneous graph structures. Existing models are difficult to adaptively handle dynamic noise and variable data structures, resulting in their reliability and stability in practical applications being unable to meet high-precision requirements.

[0008] Therefore, a new solution is urgently needed. By using physical locations to build a graph structure, combining the residual network to achieve feature completion, and further using the noise suppression network to improve the robustness of the model in complex environments and the accuracy of data reconstruction, it is necessary to better support real-time analysis and prediction tasks in the marine digital twin system. Summary of the invention

[0009] In order to solve the above-mentioned problems, the present invention provides a method and system for robust data reconstruction of marine environment based on graph neural network. The method first constructs an initial graph structure based on geographic location information and node features, and uses geographic topology information to characterize the basic spatial relationship. In view of the limitations of traditional graph structures, the graph structure is further optimized in combination with node feature content to make it more adaptable to the multi-dimensional expression requirements of data. On the basis of the constructed optimized graph structure, residual networks and graph neural networks are used for feature completion, and residual networks are used to fuse multi-level features, combined with graph neural networks to extract high-order structural dependencies, so as to achieve comprehensive data completion. After completing feature completion, a denoising network is introduced to reduce noise from the features and structures to ensure the accuracy and stability of the data. Finally, the denoised data is subjected to node representation learning through the graph attention network to obtain multi-scale and robust node feature representations, thereby effectively improving the accuracy and consistency of the data.

[0010] In the first aspect, the present invention provides a method for robust data reconstruction of marine environment based on graph neural network, which adopts the following technical solution:

[0011] A method for robust data reconstruction of marine environment based on graph neural network, comprising:

[0012] Obtain marine environmental data;

[0013] Carry out data preprocessing on the acquired marine environment data;

[0014] Constructing graph structures based on geographic location information in marine environmental data;

[0015] Establish a robust data reconstruction model of marine environment based on graph neural network;

[0016] The ocean environment robust data reconstruction model is used to identify the ocean environment data and obtain the identification results.

[0017] Furthermore, the acquired marine environmental data is preprocessed, including converting categorical variables into binary vectors using one-hot encoding for non-numeric data, normalizing the numerical features after data conversion, reducing variance differences between features, and ensuring that contributions of different features in the model are consistent; the processed data will be saved in CSV file format, and multiple labels will be assigned to each sample.

[0018] Furthermore, the construction of a graph structure based on geographic location information in marine environmental data includes constructing an initial graph G based on the geographic location information in the collected data, treating each monitoring point as a node of the graph, and defining edge connections based on the geographic proximity between nodes and the similarity of functional characteristics. The geographic location coordinates of the nodes are used to initialize the spatial topological structure of the graph, and the weights of the edges are constructed in combination with the equipment status and environmental characteristics.

[0019] Furthermore, the establishment of a robust data reconstruction model for the marine environment based on a graph neural network includes respectively constructing a feature completion layer, a denoising layer and a graph attention layer, wherein the feature completion layer takes the preprocessed marine environment data and the graph structure as input to generate a completed feature vector; the denoising layer receives the completed feature vector and the adjacency matrix as input to remove feature noise, and outputs an optimized feature representation and an updated graph structure; the graph attention layer takes the new features and graph structure output by the denoising layer as input to finally generate a feature representation with high robustness.

[0020] Furthermore, the denoising layer receives the completed feature vector and adjacency matrix as input to remove feature noise, including removing the noise impact on the collected data caused by the complex ocean environment through graph enhancement technology and adversarial training, and strengthening the topological structure of the graph, wherein the kNN graph, the purification graph, the perturbation graph and the negative sample graph are generated respectively through the graph enhancement technology; for the purification graph, the perturbation graph and the negative sample graph, the similarity-preserving convolution operation is performed in turn, the corresponding Laplacian matrix is ​​calculated for each graph, and then L layers of similarity convolution are performed to obtain the purification graph node representation, the perturbation graph node representation and the negative sample graph node representation, and finally the discriminator and the cross entropy loss function are introduced to optimize the network to improve the robustness of the node features and use the three node representations as the subsequent input of the model.

[0021] Furthermore, the graph attention layer uses the new features output by the denoising layer and the graph structure as input to ultimately generate a highly robust feature representation, including using the graph attention layer to further enhance the representation capability of the nodes and the structural representation of the graph based on dynamically aggregated neighborhood features, wherein after concatenating the node features, the attention weight function is used to score them, and the softmax function is introduced to normalize the attention weights of all adjacent nodes to generate a final attention coefficient, a convolution operation is performed based on the obtained attention coefficient, and updated node features are obtained after being processed by the graph attention layer.

[0022] Furthermore, the use of the marine environment robust data reconstruction model to identify marine environment data includes training the marine environment robust data reconstruction model, wherein the marine environment data is divided into data sets, and the training data is brought into the marine environment robust data reconstruction model based on the graph neural network. The feature completion layer error of the model is obtained by measuring the difference between the original data and the completed data, and the loss of the denoising layer is the mutual information loss calculated by the cross entropy loss function. According to the mutual information loss, the gradient descent method is applied to update the network parameters, wherein the mutual information loss It is expressed as:

[0023]

[0024] in, Represents the cosine similarity function; the loss of the graph attention layer is obtained based on the label and cross entropy loss function:

[0025] .

[0026] The second aspect is a robust data reconstruction system for marine environment based on graph neural network, including:

[0027] The data acquisition module is configured to acquire marine environment data;

[0028] The preprocessing module is configured to perform data preprocessing on the acquired marine environment data;

[0029] A graph structure module, configured to construct a graph structure based on geographic location information in the marine environment data;

[0030] A model reconstruction module is configured to establish a robust data reconstruction model of the ocean environment based on a graph neural network;

[0031] The recognition module is configured to recognize the marine environment data using the marine environment robust data reconstruction model to obtain a recognition result.

[0032] In a third aspect, the present invention provides a computer-readable storage medium storing a plurality of instructions, wherein the instructions are suitable for being loaded and executed by a processor of a terminal device, for a method for robust data reconstruction of an ocean environment based on a graph neural network.

[0033] In a fourth aspect, the present invention provides a terminal device comprising a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; the computer-readable storage medium is used to store multiple instructions, and the instructions are suitable for being loaded and executed by the processor to implement the method for robust data reconstruction of the marine environment based on a graph neural network.

[0034] In summary, the present invention has the following beneficial technical effects:

[0035] By introducing a structure that combines a feature completion network with a denoising network, the model can automatically learn deep correlation relationships in ocean data with high noise and feature missing, and capture key node features through a dynamic attention mechanism, effectively improving the accuracy and robustness of feature representation.

[0036] By utilizing multi-view comparative learning and multi-label feature enhancement, the model can not only flexibly adapt to complex environments, but also significantly improve its ability to filter noisy data, providing more reliable data reconstruction results, laying a solid foundation for the subsequent analysis and application of marine environmental data. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is a schematic diagram of a method for robust data reconstruction of an ocean environment based on a graph neural network according to Embodiment 1 of the present invention;

[0038] Figure 2 It is a schematic diagram of the model architecture of Example 1 of the present invention. DETAILED DESCRIPTION

[0039] The present invention is further described in detail below in conjunction with the accompanying drawings.

[0040] Example 1

[0041] Reference Figure 1 , a method for robust data reconstruction of an ocean environment based on a graph neural network in this embodiment includes:

[0042] Obtain marine environmental data;

[0043] Carry out data preprocessing on the acquired marine environment data;

[0044] Constructing graph structures based on geographic location information in marine environmental data;

[0045] Establish a robust data reconstruction model of marine environment based on graph neural network;

[0046] The ocean environment robust data reconstruction model is used to identify the ocean environment data and obtain the identification results.

[0047] Specifically:

[0048] Step S1, data collection:

[0049] Collect marine environmental data from a multi-source sensor network, including equipment operating parameters, geographic location coordinates, environmental noise, meteorological parameters, etc. These data usually come from multiple monitoring devices or sites. To ensure the comprehensiveness and multi-dimensional information of the data, the collected data covers spatial location characteristics, real-time environmental data, and equipment characteristic parameters, involving hundreds of characteristic dimensions.

[0050] After data collection, for the marine data environment with high noise and multimodal characteristics, based on the feature correlation and information gain screening methods, the key features related to noise impact and feature missing completion are screened out to build a preliminary input feature set. This process also combines empirical analysis and machine learning feature selection algorithms to ensure that the selected features have strong representativeness and distinguishing ability for subsequent reconstruction and denoising processes.

[0051] Step S2: Data preprocessing:

[0052] In order to unify the data formats from different sources, the collected multimodal data are preprocessed. For non-numeric data, one-hot encoding is used to convert categorical variables into binary vectors, because the model can only process numerical data when processing data.

[0053] The numerical features after data conversion are normalized to reduce the variance differences between features and ensure that the contribution of different features in the model is consistent. Specifically, the minimum-maximum normalization method is used to normalize the data to the range of [0, 1]. The normalization formula is as follows:

[0054]

[0055] in, represents the normalized eigenvalue, and Represents all samples The maximum and minimum values ​​of column features, It is the first Line The original feature value of the column. During normalization, all null values ​​are removed to ensure data integrity.

[0056] The processed data will be saved in CSV file format, and multiple labels will be assigned to each sample, which will be used for feature enhancement in the subsequent reverse optimization process. Each label represents different aspects related to the feature in practical applications. These labels will help further improve data analysis and feature construction in the reconstruction process. The labeled preprocessed data provides accurate and standardized data input for the subsequent feature completion and denoising modules.

[0057] Step S3: Dataset division:

[0058] Load the CSV data file and split the preprocessed data in step S2 into a training set and a test set. The training samples account for 80%, the validation samples account for 10%, and the test samples account for 10%, ensuring that the number of feature samples for each label in the training set and the test set is balanced. In order to enhance the generalization ability of the model, each sample in the training and test sets contains multiple labels, each label corresponds to a different data feature, to ensure that the model can learn from multiple dimensions.

[0059] Step S4: constructing a graph structure based on geographic location information:

[0060] Construct an initial map based on the geographic location information in the collected data ,Each monitoring point or device is regarded as a node of the graph, and the edge connection is defined based on the geographical proximity relationship between nodes and the similarity of functional characteristics. The geographical location coordinates of the nodes are used to initialize the spatial topology of the graph, and the weight of the edge is constructed in combination with the device status and environmental characteristics.

[0061] Step S5: Establish a robust data reconstruction model of the ocean environment based on graph neural network:

[0062] This model consists of three parts: feature completion layer, denoising layer and graph attention layer. The feature completion layer takes the preprocessed data features and the constructed graph adjacency matrix as input to generate the completed feature vector. The denoising layer receives the completed feature vector and adjacency matrix as input to remove feature noise, and outputs the optimized feature representation and updated graph structure. The graph attention layer takes the new features and graph structure output by the denoising layer as input, and finally generates a feature representation with high robustness.

[0063] The feature completion layer combines the preprocessed data features with the constructed graph adjacency matrix using the residual connection mechanism, and generates features with higher integrity and expression ability through the graph neural network and residual mechanism. The specific operations are as follows:

[0064] The preprocessed data features and the adjacency matrix Input feature completion layer. In order to enhance the expressive power of the graph, add the self-cyclization term I to the adjacency matrix A and perform normalization to obtain the normalized adjacency matrix :

[0065]

[0066] in, yes The degree matrix of represents the degree distribution of nodes. The layer graph convolutional network realizes feature completion. Layer, the update formula of node features is:

[0067]

[0068] in, Indicates The node characteristics of the layer, represents the hidden layer dimension, It is The learnable parameter matrix of the layer, represents a non-linear activation function.

[0069] After each layer of graph convolution calculation, the output features are processed using normalization operations to avoid excessive feature value amplitude or information imbalance. Then, the layer-by-layer residual connection is introduced to convert the output of the current layer Superimposed with the initial input X:

[0070]

[0071] in, represents the normalization operation, It is The residual connection weight coefficient of the layer is used to control the influence of the initial features. The final feature completion representation is the sum of all layer outputs:

[0072]

[0073] The denoising layer uses graph enhancement technology and adversarial training to remove the noise impact on the collected data caused by the complex ocean environment and strengthen the topological structure of the graph. The specific operations are as follows:

[0074] Four graph views are generated through graph enhancement technology:

[0075] (1) kNN graph: A k-nearest neighbor graph is constructed based on cosine similarity to represent the topological relationship of similar nodes in the feature space:

[0076]

[0077] in, and Is a node and nodes The characteristic vector of Represents the cosine similarity value. According to the descending order of similarity, the first k nodes are selected as neighbors to construct the kNN graph , whose adjacency matrix Representation structure. The kNN graph is used to help aggregate the features of high-similarity neighbors and enhance the similarity retention ability of node representation.

[0078] (2) Graph purification: By deleting the edges between pairs of nodes with low similarity, the perturbation of the graph is cleared, and the similarity between nodes is expressed using Jaccard similarity:

[0079]

[0080] in, is the Jaccard similarity function, Is a node and nodes By comparing the set threshold , the pruning similarity is lower than Connect the edges of , whose adjacency matrix is ​​expressed as The purified graph is used to provide more robust structural information and high-quality neighbors for neighbor aggregation.

[0081] (3) Perturbation graph: By Add slight perturbations to randomly restore some deleted edges, helping the model learn features that are insensitive to perturbations. Randomly sample a mask matrix from a Bernoulli distribution ,in Controls the proportion of restored edges. Based on the mask matrix , restore some of the deleted edges in the purified graph to obtain the perturbation graph :

[0082]

[0083] in, represents the Hadamard product, is the adjacency matrix with the perturbed graph.

[0084] (4) Negative sample image: randomly shuffle the purified image The feature order of the middle nodes generates a negative sample graph , its topological structure and purification diagram Topology The kNN view and the negative sample view are compared and the model is optimized by contrastive learning.

[0085] For the purified graph, perturbed graph, and negative sample graph, similarity-preserving convolution operations are performed in sequence. First, the corresponding Laplacian matrix is ​​calculated for each graph, taking the purified graph as an example:

[0086]

[0087] in, and Represent the adjacency matrices of the kNN graph and the purified graph, respectively. represents the corresponding degree matrix, It is the balance coefficient used to adaptively adjust the graph information fusion. Based on the obtained Laplace matrix Perform similarity convolution operation:

[0088]

[0089] in, Represents the feature matrix of the previous layer node, initial value is the input feature, Indicates The weight parameters of the layer.

[0090] The three enhanced images are Layer similarity convolution to obtain purified graph node representation , perturbation graph node representation And negative sample graph node representation . Construct positive and negative sample sets, optimize the network by introducing discriminator and cross entropy loss function to improve the robustness of node features and As the input of the model.

[0091] The node similarity obtained from the above process and , set a new threshold , in the original image Add Greater than , delete the connecting edge Less than The connected edges of , whose adjacency matrix is ​​expressed as . The robust nodes are represented by and graph structure as input to the next layer.

[0092] Furthermore, the graph attention layer further enhances the representation ability of nodes and the structural representation of the graph by dynamically aggregating neighborhood features. The specific operations are as follows:

[0093] Will No. Row and The row node features are spliced ​​and combined with the attention weight function To rate:

[0094]

[0095] in, Representation Node and nodes The initial attention weight between , the softmax function is introduced to normalize the attention weights of all adjacent nodes to generate the final attention coefficient :

[0096]

[0097] in, Representation Node All neighbor nodes of . Perform convolution operation based on the obtained attention coefficient:

[0098]

[0099] Updated node features after processing by the graph attention layer , which is used for subsequent data analysis.

[0100] Step S6: training model:

[0101] (1) Preprocess the raw data and divide the data set;

[0102] (2) The training data is introduced into the robust data reconstruction model of the marine environment based on the graph neural network established in step S5. The error of the feature completion layer of the model is obtained by measuring the difference between the original data and the completed data. The loss of the denoising layer is the mutual information loss calculated by the cross entropy loss function. :

[0103]

[0104] in, Represents the cosine similarity function. The loss of the graph attention layer is obtained based on the label and cross entropy loss function:

[0105] ;

[0106] (3) Based on the loss, apply the gradient descent method to update the network parameters;

[0107] (4) Determine whether an iteration is completed. If not, return to step (2); if so, go to step (5);

[0108] (5) Determine whether the early stopping condition is met or the maximum number of iterations is reached. If not, return to step (2); if so, complete the network training.

[0109] Step S7, verify the training results: use the test set to verify the trained ocean environment robust data reconstruction model, and judge the recognition accuracy of the model according to the recognition results; if the recognition accuracy reaches more than 80%, the trained ocean environment robust data reconstruction model will be used for subsequent data analysis and other operations, otherwise return to step S3 for further training.

[0110] Example 2

[0111] This embodiment provides a marine environment robust data reconstruction system based on a graph neural network, including:

[0112] The data acquisition module is configured to acquire marine environment data;

[0113] The preprocessing module is configured to perform data preprocessing on the acquired marine environment data;

[0114] A graph structure module, configured to construct a graph structure based on geographic location information in the marine environment data;

[0115] A model reconstruction module is configured to establish a robust data reconstruction model of the ocean environment based on a graph neural network;

[0116] The recognition module is configured to recognize the marine environment data using the marine environment robust data reconstruction model to obtain a recognition result.

[0117] A computer-readable storage medium storing a plurality of instructions, wherein the instructions are suitable for being loaded and executed by a processor of a terminal device, a method for robust data reconstruction of an ocean environment based on a graph neural network.

[0118] A terminal device includes a processor and a computer-readable storage medium, the processor is used to implement various instructions; the computer-readable storage medium is used to store multiple instructions, and the instructions are suitable for being loaded and executed by the processor to describe a method for robust data reconstruction of an ocean environment based on a graph neural network.

[0119] The above are all preferred embodiments of the present invention, and are not intended to limit the protection scope of the present invention. Therefore, any equivalent changes made based on the structure, shape, and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for robust data reconstruction of marine environment based on graph neural network, characterized in that: include: Obtain marine environmental data; Carry out data preprocessing on the acquired marine environment data; Constructing graph structures based on geographic location information in marine environmental data; Establish a robust data reconstruction model of marine environment based on graph neural network; The marine environment data is identified using the marine environment robust data reconstruction model to obtain the identification results; The acquired marine environment data is preprocessed, including converting categorical variables into binary vectors using one-hot encoding for non-numeric data, normalizing the converted numeric features, reducing variance differences between features, and ensuring that contributions of different features in the model are consistent; the processed data will be saved in CSV file format, and multiple labels will be assigned to each sample; The construction of the graph structure based on the geographic location information in the marine environment data includes constructing an initial graph G based on the geographic location information in the collected data, treating each monitoring point as a node of the graph, and defining the connection of edges according to the geographic proximity relationship between the nodes and the similarity of functional characteristics. The geographic location coordinates of the nodes are used to initialize the spatial topological structure of the graph, and the weights of the edges are constructed in combination with the equipment status and environmental characteristics; The establishment of a robust data reconstruction model for the marine environment based on a graph neural network includes constructing a feature completion layer, a denoising layer and a graph attention layer respectively, wherein the feature completion layer uses the preprocessed marine environment data and the graph structure as input to generate a completed feature vector; the denoising layer uses the completed feature vector and the adjacency matrix as input to remove feature noise, and outputs an optimized feature representation and an updated graph structure; the graph attention layer uses the new features and the graph structure output by the denoising layer as input to finally generate a feature representation with high robustness; The denoising layer is used to receive the completed feature vector and adjacency matrix as input to remove feature noise, including removing the noise impact on the collected data caused by the complex ocean environment through graph enhancement technology and adversarial training, and strengthening the topological structure of the graph, wherein the kNN graph, the purification graph, the perturbation graph and the negative sample graph are respectively generated through the graph enhancement technology; for the purification graph, the perturbation graph and the negative sample graph, similarity-preserving convolution operations are performed in turn, the corresponding Laplacian matrix is ​​calculated for each graph, and then L layers of similarity convolution are performed to obtain the purification graph node representation, the perturbation graph node representation and the negative sample graph node representation, and finally the discriminator and the cross entropy loss function are introduced to optimize the network to improve the robustness of the node features and use the three node representations as the subsequent input of the model; The graph attention layer is used to take the new features output by the denoising layer and the graph structure as input, and finally generates a feature representation with high robustness, including using the graph attention layer to further enhance the representation ability of the node and the structural representation of the graph according to the dynamic aggregation neighborhood features, wherein after the node features are spliced, the attention weight function is used to score, and the softmax function is introduced to normalize the attention weights of all adjacent nodes to generate the final attention coefficient, and a convolution operation is performed based on the obtained attention coefficient, and the updated node features are obtained after being processed by the graph attention layer; The method of using the marine environment robust data reconstruction model to identify marine environment data includes training the marine environment robust data reconstruction model, wherein the marine environment data is divided into data sets, and the training data is introduced into the marine environment robust data reconstruction model based on the graph neural network. The feature completion layer error of the model is obtained by measuring the difference between the original data and the completed data. The loss of the denoising layer is the mutual information loss calculated by the cross entropy loss function. According to the mutual information loss, the gradient descent method is applied to update the network parameters, wherein the mutual information loss It is expressed as: in, Represents the cosine similarity function; the loss of the graph attention layer is obtained based on the label and cross entropy loss function: 。 2. A marine environment robust data reconstruction system based on graph neural network, executing the method according to claim 1, characterized in that: include: The data acquisition module is configured to acquire marine environment data; The preprocessing module is configured to perform data preprocessing on the acquired marine environment data; A graph structure module, configured to construct a graph structure based on geographic location information in the marine environment data; A model reconstruction module is configured to establish a robust data reconstruction model of the ocean environment based on a graph neural network; The recognition module is configured to recognize the marine environment data using the marine environment robust data reconstruction model to obtain a recognition result.

3. A computer-readable storage medium storing a plurality of instructions, characterized in that: The instructions are suitable for being loaded by a processor of a terminal device and executing the method according to claim 1 .

4. A terminal device, comprising a processor and a computer-readable storage medium, wherein the processor is used to implement each instruction; and the computer-readable storage medium is used to store multiple instructions, characterized in that: The instructions are suitable for being loaded by a processor and executing the method as claimed in claim 1 .

Images