A high-dimensional maritime data lossy compression method based on graph perception belief propagation

Abstract

The application relates to a high-dimensional maritime data lossy compression method based on graph perception belief propagation, and belongs to the technical field of digital data processing and communication transmission. The method takes a pre-constructed high-dimensional maritime graph signal and a time adjacency matrix as input, maps the original graph signal into low-dimensional latent variables through an encoding network, performs binary quantization on the low-dimensional latent variables, and generates a compressed bit sequence for low-bandwidth transmission; a receiving end remaps the bit sequence into initial log-likelihood ratio soft information, introduces the time adjacency matrix to perform graph perception belief propagation type iterative updating, and fuses node soft information and neighborhood related information; then, the updated soft information is input into a graph attention reconstruction network to restore the high-dimensional maritime graph signal. The application can improve the reliability and graph structure maintaining capability of the compression representation under a low code rate condition, and is suitable for extremely low-bandwidth maritime data transmission.

Description

Technical Field

[0001] This invention belongs to the field of digital data processing and communication transmission technology, specifically involving maritime chart signal processing, lossy data compression, graph-aware belief propagation, and graph neural network reconstruction technology, and in particular, a high-dimensional lossy compression method for maritime data based on graph-aware belief propagation. Background Technology

[0002] With the development of maritime intelligent monitoring, ship behavior analysis, and maritime situational awareness technologies, Automatic Identification System (AIS), shore-based radar, visual perception equipment, and multi-source sensors are constantly generating a large amount of ship status data. This data typically includes ship identity, position, speed, heading, navigation status, and high-dimensional feature representations obtained from multimodal perception, characterized by high dimensionality, strong temporal correlation, and significant structural dependencies between nodes. In actual maritime communication scenarios, shore-based, ship-to-shore, and near-shore monitoring links are often limited by factors such as communication distance, channel resources, shared access, and equipment power consumption, making it difficult to directly transmit complete high-dimensional maritime status features. Therefore, how to achieve effective compression, reliable transmission, and structural preservation and reconstruction of high-dimensional maritime data under low-bandwidth or extremely low-bandwidth conditions has become an urgent problem to be solved in the field of maritime intelligent perception and communication transmission.

[0003] Existing maritime data compression methods mostly employ general data compression, scalar quantization, feature dimensionality reduction, or traditional temporal coding. While these methods can reduce the amount of data to be transmitted to some extent, they typically treat data from different time points as independent samples, making it difficult to fully utilize the temporal continuity and graph structure correlation during ship motion. When high-dimensional maritime data is organized into a graph signal, the temporal adjacency relationships between nodes can reflect the degree of correlation of ship states in temporal evolution. If the compression method ignores this type of graph structure prior, it can easily lead to problems such as structural distortion, decreased temporal smoothness, and loss of key state features in the reconstructed signal under low bitrate conditions.

[0004] Graph signal compression technology offers a new approach to the compressed transmission of high-dimensional structured data. The method presented in "GraphSignal Compression by Joint Quantization and Sampling" proposes a framework for graph signal compression based on joint sampling and quantization. This transforms the graph signal compression problem into a task-driven quantization problem, jointly designing sampling and recovery mechanisms under a fixed quantization mapping, and allocating a finite bit budget through an iterative algorithm. This method is primarily aimed at band-limited graph signals, and its core task is to recover the spectral representation of the graph signal, achieving a compact finite bit representation under certain conditions.

[0005] However, the aforementioned joint sampling and quantization methods still have certain limitations. First, these methods typically rely on the band-limited graph signal assumption, while real maritime chart signals are affected by ship maneuvers, sampling interval variations, observation noise, and complex sea conditions, and may not strictly meet the band-limiting condition. Second, these methods mainly focus on sampling node selection, quantization bit allocation, and spectrum recovery, without establishing soft information representation for the binary compressed bit sequence after low-bandwidth transmission, nor using the time adjacency matrix to iteratively enhance bit reliability at the decoding end. Third, these methods typically do not incorporate the adaptive reconstruction capabilities of graph neural networks, making it difficult to simultaneously achieve low bit rate compression, neighborhood structure preservation, and high-dimensional feature recovery in complex high-dimensional maritime chart signal scenarios.

[0006] Therefore, it is necessary to propose a lossy compression method for high-dimensional maritime map signals. Based on the high-dimensional maritime map signal and temporal adjacency matrix obtained through preceding graph signal modeling, an encoding network maps the high-dimensional graph signal into low-dimensional latent variables, and binarization quantization is used to generate a compressed bit sequence suitable for low-bandwidth transmission. At the receiving end, the bit sequence is remapped into log-likelihood ratio soft information, and a pre-constructed temporal adjacency matrix is ​​introduced to perform graph-aware belief propagation-based iterative updates to fuse the node's own soft information with neighborhood-related information. Finally, a graph attention reconstruction network is used to recover the high-dimensional maritime map signal. This method can enhance the reliability of compressed representation and the ability to preserve graph structure under low bit rate conditions, providing a new technical path for the compressed transmission and reconstruction of high-dimensional situational information in extremely low-bandwidth maritime communications. Summary of the Invention

[0007] The purpose of this invention is to address the technical problems of large data volume of high-dimensional maritime map signals in low-bandwidth maritime communication scenarios, the inability of traditional compression methods to fully utilize the correlation of time graph structure, and the tendency for structural distortion and loss of key information to occur when reconstructing signals under low bit rate conditions. This invention provides a lossy compression method for high-dimensional maritime data based on graph-aware belief propagation. The goal is to achieve low bit rate compressed transmission and structurally preserved reconstruction of high-dimensional maritime map signals by using coding networks, binarization and quantization, graph-aware belief propagation-based soft information iterative updates, and graph attention reconstruction networks, based on the high-dimensional maritime map signal and its temporal adjacency matrix obtained through preceding graph signal modeling.

[0008] To achieve the above objectives, the technical solution of the present invention is: a lossy compression method for high-dimensional maritime data based on graph-aware belief propagation, comprising:

[0009] Acquire pre-constructed high-dimensional maritime map signals and its temporal adjacency matrix The high-dimensional maritime chart signal The time adjacency matrix is ​​used to characterize the high-dimensional state features of maritime targets at multiple discrete time points. Used to characterize the temporal correlation between different time points;

[0010] The high-dimensional maritime chart signal The input is encoded by a network that maps the original high-dimensional graph signal into low-dimensional latent variables. ;

[0011] For the low-dimensional latent variables Perform binarization and quantization to generate a compressed bit sequence. ;

[0012] The compressed bit sequence Transmission is carried out via low-bandwidth maritime communication links;

[0013] The compressed bit sequence at the receiving end Remapping to the initial log-likelihood ratio soft information ;

[0014] Introducing the time adjacency matrix As a graph structure prior, the initial log-likelihood ratio soft information Perform graph-aware belief propagation-based iterative updates to obtain a soft information representation that incorporates relevant neighborhood information. ;

[0015] Representing the iteratively updated soft information The input graph attention reconstruction network recovers high-dimensional maritime map signals through graph structure-aware information aggregation and feature mapping. .

[0016] Furthermore, the method specifically includes the following steps:

[0017] Step S1: Obtain the pre-constructed high-dimensional maritime chart signal and its corresponding temporal adjacency matrix The high-dimensional maritime chart signal The time adjacency matrix is ​​used to characterize the high-dimensional state features of maritime targets at multiple discrete time points. Pre-generated by the preceding graph signal modeling process, it is used to characterize the temporal correlation between different time points. This invention does not include a temporal adjacency matrix. Instead of constructing the graph step, it is used as a priori input of a known graph structure.

[0018] Step S2: Transfer the high-dimensional maritime chart signal The input is encoded by a network, which maps the original high-dimensional graph signal into low-dimensional latent variables. The encoding network employs a multilayer perceptron structure, using nonlinear mapping to transform the input dimension... High-dimensional node features are compressed to a dimension of 1. The low-dimensional representation is used to reduce the size of the data transmitted subsequently.

[0019] Step S3: For low-dimensional latent variables Perform binarization and quantization to generate a compressed bit sequence. The binarization quantization maps continuous latent variables to binary representations through symbolic decision, enabling the compression results to adapt to low-bandwidth or extremely low-bandwidth maritime communication links. During the training phase, a pass-through estimation method is used to approximate the gradient propagation of the binarization quantization process to achieve end-to-end optimization.

[0020] Step S4: Compress the bit sequence It is used for low-bandwidth transmission of data to be transmitted. Since the transmitted object is transformed from the original high-dimensional graph signal into a low-dimensional binary bit sequence, the burden of maritime data transmission can be significantly reduced.

[0021] Step S5: At the receiving end, compress the bit sequence Remapping to the initial log-likelihood ratio soft information By using a learnable soft-information scaling factor to represent binary bits with soft information, the subsequent decoding process can not only utilize hard decision results, but also form an iteratively updated confidence information representation.

[0022] Step S6: Input the pre-entered time adjacency matrix As a priori information for graph structure, the initial log-likelihood is compared to soft information. The graph-aware belief propagation-based iterative update is performed. In each iteration, the current soft information undergoes a nonlinear transformation, utilizing the time adjacency matrix. The messages from neighboring nodes are weighted and aggregated, and the initial soft information and the neighbor propagation messages are updated with damping to obtain the soft information representation after fusing time-related neighborhood information. .

[0023] Step S7: Represent the iteratively updated soft information. The input graph attention reconstruction network recovers high-dimensional maritime map signals through graph structure-aware information aggregation and feature mapping. The graph attention reconstruction network is based on the temporal adjacency matrix. By limiting the scope of information aggregation between nodes and introducing graph structure priors in the attention weight calculation, nodes with strong temporal correlation can obtain higher information contributions during the high-dimensional feature reconstruction process.

[0024] Furthermore, in step S1, the high-dimensional maritime chart signal Represented as:

[0025]

[0026] in, The number of time points. High-dimensional maritime state feature dimensions corresponding to each time point; time adjacency matrix Represented as:

[0027]

[0028] in, Indicates the first The time node and the first The time association weights between time nodes; the time adjacency matrix This is one of the input data for this method, which does not include the time adjacency matrix. The construction steps.

[0029] Furthermore, in step S2, the encoding network High-dimensional maritime chart signals The mapping is to a low-dimensional latent variable z, which is calculated as follows:

[0030]

[0031] in, These are the learnable parameters of the encoding network; the encoding network is a multilayer perceptron structure, including a first linear mapping layer, A non-linear activation function and a second linear compression layer are used to reduce the input dimension. Mapped to compressed dimensions ,and .

[0032] Furthermore, the intermediate hidden dimensions of the encoding network From input dimension and compressed dimensions The optimal calculation method is jointly determined as follows:

[0033]

[0034] in, The original high-dimensional graph signal dimension, The dimension of the low-dimensional latent variable.

[0035] Furthermore, in step S3, the binarization quantization process uses a sign-based decision method, and its calculation method is as follows:

[0036]

[0037] Where, when an element in z is greater than or equal to 0, the corresponding The value of each element in z is +1; when an element in z is less than 0, the corresponding value is... The value of the element is -1.

[0038] Furthermore, during the model training phase, the gradient propagation of the binary quantization process is approximated using a direct-pass estimation quantization method. Its forward propagation process is expressed as follows:

[0039]

[0040] in, The random perturbation term follows a uniform distribution and is used to smooth the binarization and quantization process. During the backpropagation stage, the end-to-end joint optimization of the encoding network, the graph-aware belief propagation module, and the graph attention reconstruction network is achieved through approximate gradients.

[0041] Furthermore, in step S5, the compressed bit sequence is... Remapping to the initial log-likelihood ratio soft information The calculation method is as follows:

[0042]

[0043] in, This is a learnable soft information scaling factor. Preferably, The initial value is 2.0.

[0044] Furthermore, in step S6, the graph-aware confidence propagation iterative update includes neighborhood message generation, graph structure aggregation, and damped update, and its calculation method is as follows:

[0045]

[0046]

[0047]

[0048]

[0049] in, For the first Soft information representation in the next iteration The message is propagated in the neighborhood after nonlinear transformation. For the time adjacency matrix The aggregated neighborhood messages, This is the message amplification factor. This is the scaling factor for soft information. To plot the propagation intensity coefficient, is the damping coefficient.

[0050] Furthermore, the number of iterations for the graph-aware belief propagation iterative update... The value range is 5 to 20, with a preferred value of 10; damping coefficient The value range is 0.3 to 0.7, with a preferred value of 0.5; message amplification factor. The initial value is 2.0; the soft information scaling factor. The initial value is 2.0; the graph propagation intensity coefficient. The preferred value is 1.0.

[0051] Furthermore, in step S7, the graph attention reconstruction network includes at least one graph attention layer, which is configured according to the temporal adjacency matrix. The scope of information aggregation between nodes is limited, and the features of neighboring nodes are adaptively fused through attention weights.

[0052] Furthermore, the calculation method for the graph attention layer includes:

[0053]

[0054] in, Let K be the updated feature representation of the i-th node, and K be the number of attention heads. Let be the linear mapping matrix corresponding to the k-th attention head. Let j be the attention weight of node i under the k-th attention head. Let A be the neighborhood set of node i determined by the temporal adjacency matrix A, and || denote the multi-head feature concatenation operation.

[0055] Furthermore, the attention weights A time adjacency matrix is ​​introduced during the calculation process. The prior edge weights are calculated as follows:

[0056]

[0057]

[0058] in, For nodes With nodes In the Attention score under each attention level For learnable attention vectors, To prevent overflow of extremely small positive numbers in logarithmic operations.

[0059] Furthermore, the hidden dimension of the graph attention reconstruction network is preferably 256, the number of attention heads is preferably 4, the number of graph attention layers is preferably 2, and the dropout rate is preferably 0.1.

[0060] Furthermore, the method employs a combination of reconstruction error and bitrate constraint to form the training objective, and its loss function is expressed as:

[0061]

[0062] in, For reconstruction error, For bitrate constraints, This is the weight for bitrate constraints.

[0063] Furthermore, the reconstruction error and bit rate constraint terms are respectively expressed as:

[0064]

[0065]

[0066] in, The original high-dimensional maritime chart signal, For the reconstructed high-dimensional maritime chart signal, These are low-dimensional latent variables.

[0067] Furthermore, the training objective also includes a graph structure preservation constraint, whose extended loss function is expressed as:

[0068]

[0069] in, To preserve the constraint weights of the graph structure, This is the Laplace structural loss.

[0070] Furthermore, the Laplace structural loss is expressed as:

[0071]

[0072] in, The graph Laplacian matrix is ​​constructed from the time adjacency matrix A and is used to constrain the smoothness difference between the reconstructed graph signal and the original graph signal on the time adjacency graph.

[0073] Furthermore, the preferred number of training iterations for the method is 12,000, and the preferred learning rate is [value missing]. It is used to achieve a balance between compression ratio and reconstruction quality under low bitrate conditions.

[0074] Furthermore, the method is applicable to high-dimensional graph signal compression transmission, maritime situation information compression reconstruction, and compact representation of ship temporal state characteristics in low-bandwidth or extremely low-bandwidth maritime communication scenarios.

[0075] Compared to existing technologies, this invention offers the following advantages: The method takes a pre-constructed high-dimensional maritime map signal and a temporal adjacency matrix as input. First, it maps the original map signal into low-dimensional latent variables through an encoding network and binarizes and quantizes them to generate a compressed bit sequence for low-bandwidth transmission. The receiving end remaps the bit sequence into initial log-likelihood ratio soft information, introduces the temporal adjacency matrix to perform graph-aware belief propagation-based iterative updates, and fuses node soft information with neighborhood-related information. Subsequently, the updated soft information is input into a graph attention reconstruction network to recover the high-dimensional maritime map signal. This invention improves the reliability of compressed representation and graph structure preservation under low bitrate conditions, making it suitable for extremely low-bandwidth maritime data transmission. Attached Figure Description

[0076] Figure 1 This is a schematic diagram of the overall process of a lossy compression method for high-dimensional maritime data based on graph-aware belief propagation provided in an embodiment of this application;

[0077] Figure 2 This is a schematic diagram of the high-dimensional maritime data compression and reconstruction process provided in the embodiments of this application;

[0078] Figure 3 This is a performance analysis chart of lossy reconstruction provided in the embodiments of this application, including the reconstruction mean square error and cosine similarity.

[0079] The curve of Laplace error as a function of compression dimension;

[0080] Figure 4 This is a comparison chart of the reconstruction performance of the method of the present invention and the GFT low-pass graph signal compression method under different compression volume conditions provided in the embodiments of this application;

[0081] Figure 5 This is a comparison chart of the reconstruction performance of the method of the present invention and the JQS graph signal compression method under different compression volume conditions provided in the embodiments of this application;

[0082] Figure 6 This is a comparison chart of the reconstruction performance of the method of the present invention and the standard BP method provided in the embodiments of this application. Detailed Implementation

[0083] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are all within the protection scope of the present invention.

[0084] Numerous specific details are set forth in the following description to provide a thorough understanding of the invention; however, the invention may also be practiced in other ways different from those described herein. Those skilled in the art can make similar extensions or equivalent substitutions without departing from the spirit of the invention, and therefore the invention is not limited to the specific embodiments disclosed below.

[0085] like Figure 1 and Figure 2 As shown, this invention discloses a lossy compression method for high-dimensional maritime data based on graph-aware belief propagation. This method uses a pre-constructed high-dimensional maritime map signal... and its temporal adjacency matrix Using the input as input, low-bit-rate transmission and structure-preserving reconstruction of high-dimensional maritime map signals are achieved through encoding network compression mapping, binarization quantization, low-bandwidth transmission, graph-aware belief propagation-based soft information iterative update, and graph attention reconstruction.

[0086] S1. Obtain the pre-constructed high-dimensional maritime chart signal. and its temporal adjacency matrix .

[0087] In this embodiment, the high-dimensional maritime chart signal The structured vector-valued graph signal matrix output from the preceding graph signal modeling process is represented as:

[0088]

[0089] in, The time adjacency matrix represents the number of time points, where D represents the high-dimensional maritime state feature dimension corresponding to each time point. Represented as:

[0090]

[0091] in, Indicates the first The time node and the first The temporal correlation weights between time nodes are also included. It should be noted that the temporal adjacency matrix A has been pre-generated during the preceding graph signal modeling process; this invention does not include the temporal adjacency matrix. Instead of constructing the graph itself, it is used as a priori input of the known graph structure for subsequent graph-aware belief propagation and graph attention reconstruction.

[0092] S2. Transform high-dimensional maritime chart signals through an encoding network. Mapping to low-dimensional latent variables .

[0093] In this embodiment, the high-dimensional maritime chart signal is... Input encoding network To obtain low-dimensional latent variables :

[0094]

[0095] in, These are the learnable parameters of the encoding network. The encoding network preferably employs a multilayer perceptron structure, including a first linear mapping layer, A nonlinear activation function and a second linear compression layer. Their function is to map a high-dimensional graph signal with input dimension D to a compressed graph with dimension D. The low-dimensional latent variables, and Preferably, the hidden dimension h in the middle of the encoding network is set according to the following formula:

[0096]

[0097] in, The original high-dimensional graph signal dimension, This is a low-dimensional latent variable dimension. Through this coding network, high-dimensional maritime chart signals are mapped into a more compact low-dimensional continuous representation, providing a foundation for subsequent binarization, quantization, and low-bandwidth transmission.

[0098] S3, for low-dimensional latent variables Perform binarization and quantization to generate a compressed bit sequence. .

[0099] In this embodiment, symbol decision quantization is performed on the low-dimensional latent variable z to obtain a compressed bit sequence. :

[0100]

[0101] Among them, when When an element is greater than or equal to 0, the corresponding The value of the element is +1; when When the element in the middle is less than 0, the corresponding The element in the middle takes a value of -1. Through this binarization and quantization process, continuous low-dimensional latent variables are converted into binary compressed bit sequences suitable for low-bandwidth transmission.

[0102] During the model training phase, in order to enable the binarization and quantization process to participate in end-to-end optimization, a pass-through estimation method is preferred to approximate the gradient propagation of the binarization and quantization process. Its forward propagation process is expressed as:

[0103]

[0104] in, The random perturbation term is used to smooth the binarization and quantization process; during the backpropagation stage, the joint optimization of the encoding network, the graph-aware belief propagation module, and the graph attention reconstruction network is achieved through approximate gradient.

[0105] S4. Compress the bit sequence Used for low-bandwidth transmission.

[0106] After step S3, the original high-dimensional maritime chart signal It has been converted into a low-dimensional binary compressed bit sequence .because The dimension of the compressed data is significantly smaller than that of the original high-dimensional graph signal, and each element only needs to be represented in binary form. Therefore, the amount of data to be transmitted can be significantly reduced, making it suitable for low-bandwidth or extremely low-bandwidth maritime communication links. The transmitting end only needs to send a compressed bit sequence. The receiving end according to Subsequent soft information recovery and graph structure perception reconstruction will be carried out.

[0107] S5. Compress the bit sequence at the receiving end. Remapping to the initial log-likelihood ratio soft information .

[0108] At the receiving end, the received compressed bit sequence Mapped to initial log-likelihood ratio soft information The calculation method is as follows:

[0109]

[0110] in, This is a learnable soft information scaling factor. Preferably, The initial value is 2.0. Through this step, the binary bit sequence in hard-decision form is converted into an iteratively updatable soft-information representation, so that the subsequent decoding process not only depends on the hard bit result, but also can use belief propagation updates to enhance the reliability of the bit representation.

[0111] S6. Introduce a time adjacency matrix The execution graph-aware belief propagation iterative update is performed.

[0112] In this embodiment, the time adjacency matrix is... As a graph structure prior, it is used to guide the initial log-likelihood ratio soft information. Propagation and updates on temporal adjacency graphs. Graph-aware confidence propagation iterative updates include neighborhood message generation, graph structure aggregation, and damped updates. The iteration process can be represented as:

[0113]

[0114]

[0115]

[0116]

[0117] in, For the first Soft information representation in the next iteration The message is propagated in the neighborhood after nonlinear transformation. For the time adjacency matrix The aggregated neighborhood messages, This is the message amplification factor. This is the scaling factor for soft information. To plot the propagation intensity coefficient, is the damping coefficient.

[0118] Preferably, the number of iterations for graph-aware belief propagation iterative updates is... The value is 10; damping coefficient The value is 0.5; message amplification factor. The initial value is 2.0; the soft information scaling factor. The initial value is 2.0; the graph propagation intensity coefficient The value is 1.0. Through the above iterative updates, the node's own soft information and temporal neighborhood information are fused, enabling the compressed representation under low bitrate conditions to achieve stronger structure preservation capabilities.

[0119] S7. Reconstruct the network by inputting the updated soft information into the graph attention, and restore the high-dimensional maritime graph signal. .

[0120] After completing the graph-aware belief propagation-based iterative update, the updated soft information representation will be... Input a graph attention reconstruction network. The graph attention reconstruction network is based on the temporal adjacency matrix. The scope of information aggregation between nodes is limited, and features of neighboring nodes are adaptively fused through an attention mechanism. The feature update process of the graph attention layer can be represented as follows:

[0121]

[0122] in, For the first The updated feature representation of each node. To focus on the number of heads, For the first The linear mapping matrix corresponding to each attention head For the first Each attention node For nodes Attention weights Let A be the neighborhood set of node i determined by the temporal adjacency matrix A, and || denote the multi-head feature concatenation operation.

[0123] In this embodiment, the edge weight prior of the temporal adjacency matrix A is introduced during the attention weight calculation process, and its calculation method is as follows:

[0124]

[0125]

[0126] in, For nodes With nodes In the Attention score under each attention level For learnable attention vectors, To prevent overflow of extremely small positive numbers in logarithmic operations, a method is introduced... As a priori edge weight, nodes with strong time correlation can obtain higher information contribution during the reconstruction process.

[0127] Preferably, the hidden dimension of the graph attention reconstruction network is 256, the number of attention heads is 4, the number of graph attention layers is 2, and the dropout rate is 0.1. Finally, the output of the graph attention network is mapped back to the original high-dimensional graph signal space through a linear mapping layer to obtain the reconstructed high-dimensional maritime map signal. .

[0128] S8. Model training objectives and optimization methods.

[0129] In this embodiment, the model uses reconstruction error and bitrate constraint together to form the training objective, and its loss function is expressed as:

[0130]

[0131] in, For reconstruction error, Here, μ represents the bitrate constraint term, and μ is the bitrate constraint weight. Preferably, the reconstruction error and the bitrate constraint term are expressed as follows:

[0132]

[0133]

[0134] in, The original high-dimensional maritime chart signal, For the reconstructed high-dimensional maritime chart signal, This represents the low-dimensional latent variables encoded by the network output. Using the loss function described above, the model can achieve a balance between compression ratio and reconstruction quality.

[0135] In an optional embodiment, to further enhance the graph structure preservation capability of the reconstructed signal, a graph structure preservation constraint can be introduced, the extended loss function of which is expressed as:

[0136]

[0137] in, To preserve the constraint weights of the graph structure, This is the Laplace structure loss. The Laplace structure loss can be expressed as:

[0138]

[0139] in, For the pre-input temporal adjacency matrix The obtained graph Laplacian matrix is ​​used to constrain the smoothness difference between the reconstructed graph signal and the original graph signal on the temporal adjacency graph.

[0140] In this embodiment, the preferred number of model training epochs is 12,000, and the preferred learning rate is... Bitrate constraint weights The preferred value is 0.01. After training, a compressed bit sequence is generated during the testing phase using a deterministic symbol decision method, and the bit sequence is processed according to... The process of "initial LLR soft information - graph-aware confidence propagation iterative update - graph attention reconstruction" restores the high-dimensional maritime map signal.

[0141] Verification of the effectiveness of the embodiments of the present invention:

[0142] To verify the effectiveness of the method described in this invention, this embodiment experimentally analyzes the compression and reconstruction performance of high-dimensional maritime chart signals. The experimental evaluation metrics include the mean square error of reconstruction. Cosine similarity and Laplace error .in, Cosine similarity is used to measure the directional consistency between the original and reconstructed image signals, while cosine similarity is used to measure the numerical reconstruction error. Used to measure the ability of a reconstructed signal to preserve its structure on a time adjacency graph.

[0143] 1. Performance analysis of lossy reconstruction.

[0144] like Figure 3 As shown, this embodiment analyzes the lossy reconstruction performance under different compression dimensions. Experimental results show that under different compression dimensions, the method of this invention can maintain relatively stable MSE and cosine similarity, while the Laplace error is generally within a controllable range. This result indicates that the present invention, through encoding networks, binarization quantization, graph-aware belief propagation-based soft information iterative update, and graph attention reconstruction networks, can recover high-dimensional maritime map signals under low-dimensional compressed representation, and to a certain extent maintain the temporal structure characteristics of the original map signal. Specifically, the Laplace error fluctuates under some compression dimensions, indicating that changes in compression dimensions affect the structural smoothness of the reconstructed signal on the temporal adjacency graph; however, under most compression dimension settings, the method of this invention can still maintain relatively stable reconstruction mean square error and high cosine similarity, indicating that the method has good reconstruction stability under low-dimensional compressed representation.

[0145] 2. Comparative analysis with GFT low-pass graph signal compression method.

[0146] like Figure 4 As shown, this embodiment compares the method of the present invention with the GFT low-pass graph signal compression method under different compression volume conditions. Experimental results show that the traditional GFT low-pass method can gradually reduce the reconstruction error under higher compression volume conditions, but under conditions of smaller compression volume and stronger low-bandwidth constraints, its... The large Laplace error indicates its sensitivity to bit rate conditions. In contrast, the method of this invention exhibits more stable reconstruction error and structure preservation under low compression volume conditions, making it suitable for stable compression and reconstruction in low-bandwidth maritime communication scenarios.

[0147] 3. Comparative analysis with JQS graph signal compression method.

[0148] like Figure 5As shown, this embodiment compares the method of the present invention with the JQS graph signal compression method. The JQS method achieves graph signal compression through joint sampling and quantization, and can achieve low reconstruction error under some higher bit rate conditions. However, under extremely low bandwidth conditions, the JQS method is quite sensitive to changes in compression volume, and the reconstruction error and Laplace error fluctuate significantly. The method of the present invention maps the compressed bit sequence into soft information and combines it with graph-aware confidence propagation iterative updates using the time adjacency matrix, which can maintain more stable reconstruction performance under lower compression volume conditions.

[0149] 4. Comparative analysis with the standard BP method.

[0150] like Figure 6 As shown, this embodiment compares the reconstruction performance of the method of the present invention with that of the standard BP method. Experimental results show that, under the same scenario and similar compression conditions, the method of the present invention achieves better reconstruction performance. The cosine similarity is closer to 1 and the Laplace error is also lower than that of the standard BP method. This result shows that it is difficult to fully utilize the temporal adjacency structure in high-dimensional maritime graph signals by simply using the standard BP method. However, the method of this invention can enhance the neighborhood-related information in low-bit-rate compressed representations through graph-aware confidence propagation updates guided by the temporal adjacency matrix, thereby improving reconstruction accuracy and graph structure preservation.

[0151] 5. Comparison of compression performance with methods that process AIS data and image data separately.

[0152] To further verify the applicability of this invention in extremely low-bandwidth maritime communication scenarios, this embodiment compares the method of this invention with traditional methods that process AIS data and image data separately. Traditional separate processing methods typically compress and transmit image data and AIS data separately, making it difficult to utilize the complementary information and temporal graph structure correlation between the two. The method of this invention, on the other hand, first uniformly expresses multi-source maritime information as a high-dimensional maritime graph signal, and then performs low-dimensional bit compression and graph structure perception reconstruction on the graph signal.

[0153] Table 1. Comparison of compression performance between the method of the present invention and methods for separate processing of AIS data and image data.

[0154] method Parameter settings Compressed volume / KB Cosine similarity Laplace error Image data and AIS data are processed separately. JPEG quality = 3, AIS quantization bits = 1 516.94 0.022741 0.977143 61.5201 Image data and AIS data are processed separately. JPEG quality = 5, AIS quantization bits = 3 589.84 0.021082 0.978786 44.7954 Method of the present invention Compressed dimensions 20.19 0.000286 0.999713 3.7351

[0155] Table 1 shows a conservative comparison of the configuration with the largest compressed file size in the method of this invention. The transmission volume can be further reduced under a low-compression dimensionality configuration. As shown in Table 1, in the method of processing image data and AIS data separately, even with the parameter combination that minimizes the compressed file size, the compressed file size is still 516.94KB, with a reconstruction mean square error of 0.022741 and a Laplacian error of 61.5201. In contrast, even with the configuration with the largest compressed file size in the experiment, the method of this invention only has a compressed file size of 20.19KB, significantly lower than the 516.94KB of the method of processing image data and AIS data separately. Simultaneously, the reconstruction mean square error is reduced to 0.000286, the cosine similarity is increased to 0.999713, and the Laplacian error is reduced to 3.7351. Experimental results demonstrate that the method of this invention can maintain high reconstruction accuracy and graph structure preservation capability under extremely low transmission volume conditions, making it more suitable for low-bandwidth maritime communication scenarios.

[0156] Table 2. Comparison of average performance of the method of the present invention and existing compression methods under low compression volume conditions in multiple scenarios.

[0157] method Comparison Settings Average compressed volume / KB average Mean cosine similarity Mean Laplace error AIS data and image data are processed separately. The average value of the minimum compressed file size for each of the six scenarios 427.86 0.023990 0.975809 42.3242 GFT low-pass graph signal compression method The average of the smallest compressed file size for each of the six scenes. 0.58 0.002645 0.998843 154.2150 JQS graph signal compression method The average of the smallest compressed file size for each of the six scenes. 0.50 0.002644 0.998667 140.9664 Method of the present invention Average result of compression dimension d=8 in 6 scenarios 0.50 0.000716 0.999285 14.8383

[0158] Table 2 shows that, in the average results of the six maritime compression scenarios, even with the parameter combination that minimizes the compressed file size in each scenario, the method of processing AIS data and image data separately still has an average compressed file size of 427.86KB; while the method of this invention has an average compressed file size of only 0.50KB under the condition of compression dimension d=8. Compared with the GFT low-pass graph signal compression method and the JQS graph signal compression method, the method of this invention achieves a lower average MSE, a higher average cosine similarity, and a lower average Laplacian error under similar low compression volume conditions. The above results indicate that the method of this invention can better balance compression efficiency, reconstruction accuracy, and graph structure preservation under low compression volume conditions.

[0159] Table 3. Average performance comparison of the graph-aware belief propagation method of the present invention and the standard BP method in six scenarios.

[0160] method Compressed dimensions Average compressed volume / KB average Mean cosine similarity Mean Laplace error Standard BP+GAT 6.2480 0.004311 0.995976 24.1906 Image Sensing BP+GAT 6.2480 0.001372 0.998622 18.4923

[0161] As shown in Table 3, under the same compression dimensionality and compression volume, the graph-aware BP + GAT method of this invention achieves lower average MSE, higher average cosine similarity, and lower average Laplace error compared to the standard BP + GAT method. This result indicates that introducing the temporal adjacency matrix A for graph-aware soft information propagation helps enhance the reliability of low-bit-rate compressed representation and the ability to preserve graph structure.

[0162] As can be seen from Tables 2 and 3, this invention does not simply combine encoding compression, binarization quantization, confidence propagation, and graph attention networks in parallel. Instead, in low-bandwidth transmission scenarios, it remaps the compressed bit sequence to initial log-likelihood ratio soft information, uses the temporal adjacency matrix to perform graph-aware confidence propagation-based iterative updates on the soft information, and then recovers the high-dimensional maritime map signal through a graph attention reconstruction network. This processing link enhances the reliability of the binary compressed bits, fuses temporal neighborhood-related information, and simultaneously demonstrates lower reconstruction mean square error, higher cosine similarity, lower Laplace error, and smaller compression volume in the average results across multiple scenarios, indicating that the various technical features have a synergistic effect on the compression and reconstruction of low-bandwidth maritime map signals.

[0163] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be included within the scope of protection of the present invention.

Claims

1. A lossy compression method for high-dimensional maritime data based on graph-aware belief propagation, characterized in that, include: Acquire pre-constructed high-dimensional maritime map signals and its temporal adjacency matrix The high-dimensional maritime chart signal The time adjacency matrix is ​​used to characterize the high-dimensional state features of maritime targets at multiple discrete time points. Used to characterize the temporal correlation between different time points; The high-dimensional maritime chart signal The input is encoded by a network that maps the original high-dimensional graph signal into low-dimensional latent variables. ; For the low-dimensional latent variables Perform binarization and quantization to generate a compressed bit sequence. ; The compressed bit sequence Transmission is carried out via low-bandwidth maritime communication links; The compressed bit sequence at the receiving end Remapping to the initial log-likelihood ratio soft information ; Introducing the time adjacency matrix As a graph structure prior, the initial log-likelihood ratio soft information Perform graph-aware belief propagation-based iterative updates to obtain a soft information representation that incorporates relevant neighborhood information. ; Representing the iteratively updated soft information Input graph attention reconstruction network to recover high-dimensional maritime graph signals .

2. The lossy compression method for high-dimensional maritime data based on graph-aware belief propagation according to claim 1, characterized in that, The high-dimensional maritime chart signal Represented as: in, The number of time points. The high-dimensional maritime state feature dimension corresponding to each time point; The time adjacency matrix Represented as: in, Indicates the first The time node and the first Time correlation weights between time nodes.

3. The lossy compression method for high-dimensional maritime data based on graph-aware belief propagation according to claim 1, characterized in that, Coding Network High-dimensional maritime chart signals The mapping is to a low-dimensional latent variable z, which is calculated as follows: in, These are the learnable parameters of the encoding network; the encoding network is a multilayer perceptron structure, including a first linear mapping layer, A non-linear activation function and a second linear compression layer are used to reduce the input dimension. Mapped to compressed dimensions ,and .

4. The lossy compression method for high-dimensional maritime data based on graph-aware belief propagation according to claim 3, characterized in that, The intermediate hidden dimensions of the coding network From input dimension and compressed dimensions The optimal calculation method is jointly determined as follows: in, The original high-dimensional graph signal dimension, The dimension of the low-dimensional latent variable.

5. The lossy compression method for high-dimensional maritime data based on graph-aware belief propagation according to claim 1, characterized in that, Binarization quantization uses a sign-decision method, and its calculation method is as follows: Where, when an element in z is greater than or equal to 0, the corresponding The value of each element in z is +1; when an element in z is less than 0, the corresponding value is... The value of the element is -1.

6. The lossy compression method for high-dimensional maritime data based on graph-aware belief propagation according to claim 5, characterized in that, During the model training phase, the gradient propagation of the binary quantization process is approximated using a direct-pass estimation quantization method. Its forward propagation process is represented as follows: in, The random perturbation term follows a uniform distribution and is used to smooth the binarization and quantization process. During the backpropagation stage, the end-to-end joint optimization of the encoding network, the graph-aware belief propagation module, and the graph attention reconstruction network is achieved through approximate gradients.

7. The lossy compression method for high-dimensional maritime data based on graph-aware belief propagation according to claim 1, characterized in that, Compress the bit sequence Remapping to the initial log-likelihood ratio soft information The calculation method is as follows: in, It is a learnable soft information scaling factor.

8. The lossy compression method for high-dimensional maritime data based on graph-aware belief propagation according to claim 1, characterized in that, Graph-aware confidence propagation-based iterative updates include neighborhood message generation, graph structure aggregation, and damped updates. The calculation method is as follows: in, For the first Soft information representation in the next iteration The message is propagated in the neighborhood after nonlinear transformation. For the time adjacency matrix The aggregated neighborhood messages This is the message amplification factor. This is the scaling factor for soft information. To plot the propagation intensity coefficient, is the damping coefficient.

9. The lossy compression method for high-dimensional maritime data based on graph-aware belief propagation according to claim 1, characterized in that, The graph attention reconstruction network includes at least one graph attention layer, where the updated feature representation of the i-th node in the graph attention layer is... Calculated by the following formula: in, Let K be the updated feature representation of the i-th node, and K be the number of attention heads. Let be the linear mapping matrix corresponding to the k-th attention head. Let j be the attention weight of node i under the k-th attention head. Let A be the neighborhood set of node i determined by the temporal adjacency matrix A, and || denote the multi-head feature concatenation operation.

10. A lossy compression method for high-dimensional maritime data based on graph-aware belief propagation according to claim 9, characterized in that, Attention weight A time adjacency matrix is ​​introduced during the calculation process. The prior edge weights are calculated as follows: in, For nodes With nodes In the Attention score under each attention level For learnable attention vectors, To prevent overflow of extremely small positive numbers in logarithmic operations.

Images