Knowledge graph-based ultra-low frequency / very low frequency electromagnetic wave cross-domain propagation prediction method

Abstract

This application provides a knowledge graph-based method for predicting the cross-domain propagation of ultra-low frequency / extremely low frequency electromagnetic waves. The method includes: acquiring propagation scenario parameters of the target electromagnetic wave and a pre-constructed knowledge graph, whereby the knowledge graph describes the relationship between target factors influencing the cross-domain propagation of electromagnetic waves and their propagation characteristics; inputting the propagation scenario parameters into a first feedforward neural network to obtain a propagation scenario feature vector, and inputting the knowledge graph into a graph neural network to obtain a knowledge graph feature vector; fusing the two to generate a cross-domain propagation feature vector, and inputting it into a second feedforward neural network to output a potential function used to calculate the prediction result of the first electromagnetic field above the sea surface; determining the conductivity layering structure of seawater based on the knowledge graph, and calculating the prediction result of the second electromagnetic field at a specified depth based on the prediction result of the first electromagnetic field, thereby achieving accurate and reliable propagation prediction of ultra-low frequency / extremely low frequency electromagnetic waves in cross-domain propagation scenarios.

Description

Technical Field

[0001] This application relates to the field of electromagnetic wave propagation prediction technology, and in particular to a knowledge graph-based method for predicting the cross-domain propagation of ultra-low frequency / extremely low frequency electromagnetic waves. Background Technology

[0002] Super Low Frequency (SLF, 30–300 Hz) and Extremely Low Frequency (ELF, 3–30 Hz) electromagnetic waves, due to their strong penetrating power and long propagation distance, have significant application value in fields such as resource exploration, earthquake monitoring, and deep-sea communication. For example, in resource exploration, ELF electromagnetic waves can penetrate deep strata and interact with different electrical media underground, generating detectable electromagnetic signals, thus providing important information for detecting underground mineral resources and oil and gas reservoirs. As another example, in earthquake monitoring and early warning, before an earthquake occurs, the underground medium undergoes a series of physical and chemical changes, leading to alterations in electrical parameters. SLF / ELF electromagnetic waves can capture these weak changes, providing crucial clues for short-term and imminent earthquake prediction. For example, in deep-sea cross-domain and cross-medium communication, ultra-low frequency / extremely low frequency electromagnetic waves have the advantages of long propagation distance and low attenuation. They can penetrate two parallel layers, the ionosphere and the dynamic interface between the atmosphere and the ocean, to achieve stable cross-medium communication within a certain distance and depth range.

[0003] In cross-domain propagation scenarios, ultra-low frequency (ULF) / extremely low frequency (ELF) electromagnetic waves need to traverse multiple non-uniform media such as the ionosphere, atmosphere, ocean, and crust. Their propagation path, attenuation characteristics, and phase changes are affected by the coupling of multiple factors such as changes in the Earth's ionosphere, geological structure, and marine meteorological conditions. The propagation of ULF / ELF electromagnetic waves exhibits high nonlinearity and uncertainty, making it difficult to accurately describe and predict their propagation characteristics.

[0004] In the field of electromagnetic wave propagation prediction in the ultra-low frequency and extremely low frequency bands, physical models and numerical simulations struggle to capture the deep semantic relationships and underlying patterns of electromagnetic waves during cross-domain propagation. This limits the accuracy and generalization ability of the prediction models. While they can describe the propagation process to some extent, they suffer from high computational complexity and poor adaptability to complex environments. Although knowledge graph technology possesses powerful knowledge representation and reasoning capabilities, existing knowledge graph construction methods primarily target general domain knowledge, lacking specific processing for electromagnetic wave propagation expertise. Consequently, they struggle to accurately represent the various entities, attributes, and relationships involved in electromagnetic wave propagation.

[0005] Therefore, this application proposes a knowledge graph-based method for predicting the cross-domain propagation of ultra-low frequency / extremely low frequency electromagnetic waves. By constructing a knowledge graph in the field of electromagnetic wave propagation and combining the reasoning ability of the knowledge graph with the electromagnetic wave propagation prediction model, intelligent cross-domain propagation prediction of ultra-low frequency / extremely low frequency electromagnetic waves based on knowledge graphs can be achieved. Summary of the Invention

[0006] This application provides a knowledge graph-based method for predicting the cross-domain propagation of ultra-low frequency / extremely low frequency electromagnetic waves, which can be used to achieve accurate and reliable propagation prediction of ultra-low frequency / extremely low frequency electromagnetic waves in cross-domain propagation scenarios.

[0007] In a first aspect, embodiments of this application provide a knowledge graph-based method for predicting the cross-domain propagation of ultra-low frequency / extremely low frequency electromagnetic waves, the method comprising: The propagation scenario parameters of the target electromagnetic wave and a pre-constructed knowledge graph are obtained. The target electromagnetic wave is an ultra-low frequency electromagnetic wave or an extremely low frequency electromagnetic wave. The knowledge graph is used to describe the relationship between the target factors affecting the cross-domain propagation of the electromagnetic wave and the propagation characteristics of the electromagnetic wave. The propagation scenario parameters are input into a first feedforward neural network to output a propagation scenario feature vector using the first feedforward neural network. The knowledge graph is input into a graph neural network to output a knowledge graph feature vector using the graph neural network. Based on the propagation scenario feature vector and the knowledge graph feature vector, determine the cross-domain propagation feature vector; The cross-domain propagation feature vector is input into a second feedforward neural network to utilize the output of the second feedforward neural network to characterize the electromagnetic field distribution of the earth-ionospheric waveguide. The potential function is used to calculate the first electromagnetic field prediction result above the sea surface. The conductivity layering structure of seawater along the depth direction is determined based on the knowledge graph. Then, using the first electromagnetic field prediction result and the conductivity layering structure, the second electromagnetic field prediction result at a specified depth below the specified sea surface is calculated. The specified depth is determined based on the propagation scenario parameters.

[0008] Secondly, embodiments of this application provide a knowledge graph-based device for predicting the cross-domain propagation of ultra-low frequency / extremely low frequency electromagnetic waves, the device comprising: The acquisition module is used to acquire the propagation scene parameters of the target electromagnetic wave and a pre-constructed knowledge graph. The target electromagnetic wave is an ultra-low frequency electromagnetic wave or an extremely low frequency electromagnetic wave. The knowledge graph is used to describe the relationship between the target factors affecting the cross-domain propagation of electromagnetic waves and the propagation characteristics of electromagnetic waves. The first prediction module is configured to input the propagation scenario parameters into a first feedforward neural network to output a propagation scenario feature vector using the first feedforward neural network; input the knowledge graph into a graph neural network to output a knowledge graph feature vector using the graph neural network; determine a cross-domain propagation feature vector based on the propagation scenario feature vector and the knowledge graph feature vector; and input the cross-domain propagation feature vector into a second feedforward neural network to output a potential function characterizing the electromagnetic field distribution of the earth-ionospheric waveguide using the second feedforward neural network, wherein the potential function is used to calculate the first electromagnetic field prediction result above the sea surface. The second prediction module is used to determine the conductivity layering structure of seawater along the depth direction based on the knowledge graph, and to calculate the second electromagnetic field prediction result at a specified depth below the specified sea surface using the first electromagnetic field prediction result and the conductivity layering structure, wherein the specified depth is determined according to the propagation scenario parameters.

[0009] Thirdly, embodiments of this application provide an electronic device, including: a memory, a processor, and a communication interface; wherein, the memory stores a computer program, and when the computer program is executed by the processor, the processor can at least implement the knowledge graph-based method for predicting the cross-domain propagation of ultra-low frequency / extremely low frequency electromagnetic waves as described in the first aspect.

[0010] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program that, when executed by a processor of an electronic device, enables the processor to at least implement the knowledge graph-based method for predicting the cross-domain propagation of ultra-low frequency / extremely low frequency electromagnetic waves as described in the first aspect.

[0011] Fifthly, embodiments of this application provide a computer program product, including: a computer program or instructions, which, when executed by a processor of an electronic device, enable the processor to at least implement the knowledge graph-based method for predicting the cross-domain propagation of ultra-low frequency / extremely low frequency electromagnetic waves as described in the first aspect.

[0012] The solution provided in this application is applied to the scenario of predicting the propagation of ultra-low frequency / extremely low frequency electromagnetic waves in a marine environment. For the target electromagnetic wave to be predicted (the target electromagnetic wave is an ultra-low frequency electromagnetic wave or an extremely low frequency electromagnetic wave), firstly, the propagation scenario parameters of the target electromagnetic wave and a pre-constructed knowledge graph are obtained. The knowledge graph describes the correlation between target factors affecting the cross-domain propagation of electromagnetic waves and the propagation characteristics of the electromagnetic waves. Then, the propagation scenario parameters are input into a first feedforward neural network to output a propagation scenario feature vector; the knowledge graph is input into a graph neural network to output a knowledge graph feature vector; based on the propagation scenario feature vector and the knowledge graph feature vector, a cross-domain propagation feature vector is determined. Next, the cross-domain propagation feature vector is input into a second feedforward neural network to output a potential function characterizing the electromagnetic field distribution of the earth-ionospheric waveguide. This potential function is used to calculate the first electromagnetic field prediction result above the sea surface. Finally, the conductivity layering structure of seawater along the depth direction is determined based on the knowledge graph, and the second electromagnetic field prediction result at a specified depth below the specified sea surface is calculated using the first electromagnetic field prediction result and the conductivity layering structure. The specified depth is determined based on the propagation scenario parameters.

[0013] In this scheme, on the one hand, a dual-branch fusion architecture of a first feedforward neural network and a graph neural network is used to extract the propagation scene feature vector and knowledge graph feature vector of the propagation scene parameters, respectively, and collaboratively generate cross-domain propagation feature vectors. This drives the second feedforward neural network to output a potential function representing the electromagnetic field distribution of the earth-ionospheric waveguide, so as to solve the electromagnetic field above the sea surface based on the potential function and perform underwater electromagnetic field inference, realizing the joint prediction of the electromagnetic field distribution in the waveguide region above the sea surface and the medium below the sea surface. On the other hand, by introducing a pre-constructed knowledge graph in the field of electromagnetic wave propagation, the physical relationship between the multi-source heterogeneous factors affecting the cross-domain propagation of electromagnetic waves (such as ionospheric state, seawater medium properties, environmental parameters, etc.) and propagation characteristics (such as field strength, attenuation, penetration depth, etc.) is explicitly modeled, and the graph neural network is used to structure and encode it, so that the second feedforward neural network has knowledge-guided reasoning ability. This effectively overcomes the problems of physical inconsistency and poor generalization of pure data-driven methods in small sample or out-of-distribution scenarios, and also ensures that the medium model on which the underwater propagation calculation depends has physical interpretability and environmental adaptability. In summary, this approach not only improves the accuracy and reliability of propagation prediction of ultra-low frequency / extremely low frequency electromagnetic waves in complex cross-domain environments (earth-ionosphere-atmosphere-seawater), but also significantly enhances the model's adaptability and robustness to dynamic marine environments, providing highly reliable technical support for key tasks such as underwater long-range communication, target detection, and electromagnetic situation prediction. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 A flowchart of a knowledge graph-based method for predicting the cross-domain propagation of ultra-low frequency / extremely low frequency electromagnetic waves is provided for embodiments of this application; Figure 2 A schematic diagram illustrating a knowledge graph-based method for predicting the cross-domain propagation of ultra-low frequency / extremely low frequency electromagnetic waves, provided in an embodiment of this application. Figure 3 A schematic diagram illustrating the correlation of a propagation process provided in an embodiment of this application; Figure 4 An environmental impact diagram provided for an embodiment of this application; Figure 5 This is a schematic diagram of the composition of an edge feature matrix provided in an embodiment of this application; Figure 6 A flowchart illustrating a method for training an electromagnetic wave cross-domain propagation model, as provided in this application embodiment; Figure 7 A schematic diagram of the structure of a knowledge graph-based transdomain propagation prediction device for ultra-low frequency / extremely low frequency electromagnetic waves provided in this application embodiment; Figure 8 To and Figure 7 The illustrated embodiment provides a schematic diagram of the electronic device corresponding to the knowledge graph-based ultra-low frequency / extremely low frequency electromagnetic wave cross-domain propagation prediction device. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0017] It should be noted that, in the cases involving user information in the embodiments of this application, the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in the embodiments of this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use, and processing of related data must comply with the relevant laws, regulations, and standards of the relevant countries and regions, and corresponding operation entry points are provided for users to choose to authorize or refuse. In addition, the various models involved in this application (including but not limited to language models or large models) comply with relevant laws and standards.

[0018] Furthermore, the timing of the steps in the following method embodiments is merely an example and not a strict limitation.

[0019] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0020] The knowledge graph-based method for predicting the cross-domain propagation of ultra-low frequency / extremely low frequency electromagnetic waves provided in this application can be executed by an electronic device, such as a PC, laptop, or smartphone, or a server. The server can be a physical server containing an independent host, a virtual server, a cloud server, or a server cluster.

[0021] Figure 1 A flowchart of a knowledge graph-based method for predicting the cross-domain propagation of ultra-low frequency / extremely low frequency electromagnetic waves, as provided in this application embodiment, is shown below. Figure 1 As shown, it may include the following steps: 101. Obtain the propagation scene parameters of the target electromagnetic wave and the pre-constructed knowledge graph. The target electromagnetic wave is an ultra-low frequency electromagnetic wave or an extremely low frequency electromagnetic wave. The knowledge graph is used to describe the relationship between the target factors affecting the cross-domain propagation of electromagnetic waves and the propagation characteristics of electromagnetic waves.

[0022] 102. Input the propagation scenario parameters into the first feedforward neural network to output the propagation scenario feature vector using the first feedforward neural network; input the knowledge graph into the graph neural network to output the knowledge graph feature vector using the graph neural network; determine the cross-domain propagation feature vector based on the propagation scenario feature vector and the knowledge graph feature vector.

[0023] 103. Input the cross-domain propagation feature vector into the second feedforward neural network, so as to use the output of the second feedforward neural network to characterize the electromagnetic field distribution of the earth-ionospheric waveguide. The potential function is used to calculate the prediction result of the first electromagnetic field above the sea surface.

[0024] 104. Determine the conductivity layering structure of seawater along the depth direction based on the knowledge graph, and use the first electromagnetic field prediction result and the conductivity layering structure to calculate the second electromagnetic field prediction result at a specified depth below the specified sea surface. The specified depth is determined based on the propagation scenario parameters.

[0025] In summary, the knowledge graph-based method for predicting the cross-domain propagation of ultra-low frequency / extremely low frequency electromagnetic waves provided in this embodiment addresses the modeling needs of electromagnetic wave cross-domain propagation in complex ocean-atmosphere-ionosphere-earth coupling environments. By integrating domain knowledge of electromagnetic wave cross-domain propagation with data-driven modeling, it accurately characterizes the spatial propagation characteristics of target electromagnetic waves in the earth-ionospheric waveguide. Furthermore, relying on the deep correlation between environmental elements and electromagnetic wave propagation mechanisms contained in the knowledge graph, it achieves joint prediction of electromagnetic field intensity in the waveguide region above the sea surface and in the medium below the sea surface, significantly improving the physical consistency and generalization ability of cross-domain propagation prediction.

[0026] For ease of understanding, combined with Figure 2 To explain, Figure 2 This is a schematic diagram of a knowledge graph-based method for predicting the cross-domain propagation of ultra-low frequency / extremely low frequency electromagnetic waves, provided in an embodiment of this application.

[0027] like Figure 2 As shown, in the scenario of predicting the cross-domain propagation of ultra-low frequency / extremely low frequency electromagnetic waves, the propagation scenario parameters of the target electromagnetic wave and the pre-constructed knowledge graph are first obtained.

[0028] Among them, the propagation scenario parameters are a set of predefined input information used to describe the physical environment in which the target electromagnetic wave is located in a specific cross-domain propagation task. As an external condition of the entire prediction process, they directly determine the spatial path, boundary constraints, and environmental context of the target electromagnetic wave from transmission to reception.

[0029] Optionally, the propagation scenario parameters can be obtained through actual measurement / observation using ocean buoys, ionospheric radar, satellite remote sensing, or underwater sensor networks.

[0030] Optionally, the specific composition of the propagation scenario parameters can be customized according to actual application requirements. For example, when the target electromagnetic wave is an ultra-low frequency electromagnetic wave (e.g., 30–300 Hz), the propagation scenario parameters may include: the location of the transmitting source (e.g., 30.5°N, 120.2°E, 80 m below sea level), the location of the receiving point (e.g., 32.1°N, 121.7°E, 200 m below sea level), the propagation path length (e.g., 300 km), the operating frequency (e.g., 75 Hz), the transmission time (e.g., 08:00 UTC on June 15, 2025), the antenna height or burial depth, and the initial polarization direction in the reference coordinate system. When the target electromagnetic wave is an extremely low frequency electromagnetic wave (e.g., 3–30 Hz), propagation scenario parameters may include, for example, the global-scale geographic coordinates of the transmitter-receiver pair (e.g., from an Antarctic transmitter to a moored receiver near the equator), the medium across which the signal propagates (e.g., including the D layer of the ionosphere, the troposphere, the sea surface, and deep-sea strata), the center frequency (e.g., 22 Hz), the signal duration, the local time of propagation, and environmental parameters such as the season and solar activity index.

[0031] Among them, the knowledge graph is used to describe the relationship between the target factors affecting the cross-domain propagation of electromagnetic waves and the propagation characteristics of electromagnetic waves.

[0032] In an optional embodiment, the knowledge graph can be constructed as follows: Define a set of entity types and a set of relation types related to the cross-domain propagation task of ultra-low frequency / extremely low frequency electromagnetic waves, wherein the set of entity types contains multiple entity types related to electromagnetic wave propagation, and the set of relation types is used to describe the interaction relationships between multiple entity types; based on the set of entity types and the set of relation types, extract triple instances that conform to the entity types and relation types from the pre-acquired multi-source data related to cross-domain propagation of electromagnetic waves, and organize the triple instances into a graph structure to form a knowledge graph.

[0033] In practice, the entity types in the entity type set and the relationship types in the relationship type set can be customized based on usage requirements.

[0034] For example, the main entity types in the knowledge graph can be customized to include at least one of the following: electromagnetic waves (subdivided by frequency band into extremely low frequency electromagnetic waves and ultra-low frequency electromagnetic waves), propagation medium (earth, ionosphere, air, sea surface, and different underwater regions), geological structure (submarine faults, ore bodies, oil and gas reservoirs, etc.), marine environmental elements (water temperature, salinity, ocean currents, etc.), meteorological factors (wind speed, air pressure, etc.), propagation path (the propagation trajectory of electromagnetic waves at different geographical locations and times), etc.

[0035] For example, the relationships between entities in a knowledge graph can be customized to include at least one of the following: a "propagation" relationship between electromagnetic waves and the propagation medium, indicating that electromagnetic waves propagate in a specific medium; an "interaction" relationship between electromagnetic waves and geological structures, reflecting the mutual influence between electromagnetic waves and seabed geological bodies; an "influence" relationship between marine environmental factors and the propagation medium, explaining how the marine environment changes the properties of the medium; a "correlation" relationship between meteorological factors and the air medium, reflecting the influence of meteorological conditions on air electrical parameters; and a "correlation" relationship between the propagation path and time and geographical location, recording the spatial and temporal information of the propagation path, etc.

[0036] Table 1 provides examples of entities and attributes in a knowledge graph of electromagnetic wave cross-domain propagation provided in this application embodiment. As shown in Table 1, the entity type set can be defined to include at least one of the following types of entities: electromagnetic wave propagation entities, environmental entities, propagation characteristic entities, etc. Further, each type of entity can contain at least one entity. For example, electromagnetic wave propagation entities include: electromagnetic wave attributes, time, location, etc.; environmental entities include: terrestrial environment, sea surface environment, underwater environment, ionospheric environment, etc.; propagation characteristic entities include: propagation path, attenuation characteristics, etc. In addition, each entity can be expressed through corresponding attributes. Specific attributes corresponding to different entities can be found in Table 1, and will not be elaborated further here.

[0037] Table 1. Examples of entities and attributes in the knowledge graph of electromagnetic wave transdomain propagation.

[0038] In this embodiment, the relationship types included in the relationship type set can be used to describe physical interactions, environmental influences, propagation paths, etc., during the electromagnetic wave propagation process. They may include at least one of the following types: propagation process correlation, environmental influence correlation, physical process correlation, experimental and measurement correlation, etc.

[0039] Among them, the correlation in the propagation process is used to describe the causal or dependent relationships formed between different entities during the cross-domain propagation of ultra-low frequency / extremely low frequency electromagnetic waves due to signal transmission paths, medium effects, or field quantity evolution.

[0040] Figure 3 This application provides a schematic diagram illustrating the correlation of a propagation process in an embodiment of the present application, such as... Figure 3As shown, the longitude and latitude of the transmitting point represent its location, and their changes directly affect the propagation path. Changes in the propagation path usually indicate changes in the propagation environment, which in turn determine the attenuation characteristics of the electromagnetic wave. The attenuation characteristics of the electromagnetic wave affect the amplitude, phase, and waveform of the receiving point field. Based on this, the entities involved in this example and their relationships during propagation can be transformed into triples (entity 1, relationship, entity 2), which can be represented as (transmitting point longitude, influence, propagation path), (transmitting point latitude, influence, propagation path), (propagation path, influence, attenuation characteristics), (attenuation characteristics, influence, amplitude), (attenuation characteristics, influence, phase), (attenuation characteristics, influence, waveform), etc.

[0041] Environmental impact relationships describe the direct or indirect effects of the physical properties of environmental media such as the ionosphere and ocean on the propagation characteristics of ultra-low frequency (ULF) / extremely low frequency (ELF) electromagnetic waves during transdomain propagation. Environmental impact relationships encompass at least one of the following types of relationships: the influence of time on environmental properties, the influence of geographical location on environmental properties, the influence of environmental properties on propagation characteristics, and the interaction between environmental properties.

[0042] Figure 4 An environmental impact diagram provided for an embodiment of this application, such as... Figure 4 As shown, the red directed line segments represent attributes belonging to the ionospheric environment, the purple directed line segments represent the influence of time on attributes pointing to the ionospheric environment, and the gray directed line segments represent the influence relationship between attributes belonging to the ionospheric environment and the attributes being pointed to. Specifically, when an entity is in the ionospheric environment, excluding the inherent characteristics of electromagnetic waves, ionospheric parameters will be related to time, latitude, longitude, and altitude. Free electrons, ions, and neutral particles present in the ionosphere, such as NO+, O2+, O+, and oxygen atoms, will also affect the propagation characteristics of radio electromagnetic waves. For example, the concentration of various particles in the ionosphere and the average kinetic energy of free electrons affect the conductivity, dielectric constant (or dielectric tensor), and collision frequency in the ionospheric environment. Furthermore, changes in collision frequency will directly affect conductivity and dielectric constant. The relationships between these exemplified entities can also be converted into triples, such as (ionosphere, including, nitrogen density) (time, influence, ionospheric dielectric constant), etc.

[0043] In practical applications, there are a series of physical conclusions and experiments regarding electromagnetic wave propagation. Based on these conclusions and experiments, the relationships can be categorized into physical process correlations and experimental-measurement correlations. For example, higher frequency / longer path / daytime compared to nighttime / land compared to seawater → greater attenuation rate → smaller electric field strength, etc. These physical process correlations and experimental-measurement correlations provide concise rules and reasoning paths for predicting the cross-domain propagation of electromagnetic waves, contributing to the dynamic prediction and analysis of electromagnetic wave propagation.

[0044] After defining the entity type set and relation type set, the system further extracts triple instances that conform to the entity type and relation type from the pre-acquired multi-source data related to the cross-domain propagation of electromagnetic waves, and organizes the triple instances into a graph structure to form a knowledge graph.

[0045] Regarding multi-source data, the data sources include, but are not limited to: research literature, experimental data, technical white papers, and industry standards on the propagation of extremely low frequency (ULF) / ultra-low frequency (ULF) electromagnetic waves; propagation data generated through simulation using professional electromagnetic wave propagation simulation software; classic cases and records of ULF / ULF electromagnetic wave propagation; and observational data collected by ULF / ULF receiving equipment (such as underwater electromagnetic sensor arrays, surface buoy receiving stations, etc.), marine geological exploration vessels, marine observation satellites, buoys, and underwater moorings deployed in deep-sea areas. In terms of data type, this includes, but is not limited to: structured data, semi-structured data, and unstructured data.

[0046] As an optional method for extracting triple instances that conform to entity and relation types from multi-source data, for structured data such as geological exploration reports and marine environmental monitoring databases, database querying and rule matching methods can be used to extract entity and relation information and generate corresponding triple instances. For semi-structured data such as marine scientific research literature and weather forecast texts, natural language processing techniques, such as named entity recognition and relation extraction algorithms, can be used to identify entities and their relationships from the text. For example, information such as "abnormal electromagnetic wave propagation near a certain submarine fault, possibly related to changes in the fault's electrical properties" can be extracted from the literature to construct corresponding triple instances of entities and relations. For unstructured data such as marine remote sensing images and electromagnetic wave propagation waveforms, image recognition and signal processing techniques can be used to extract key information. For example, by analyzing marine remote sensing images, the sea surface temperature distribution in different sea areas can be identified, converted into corresponding entities and relations, and corresponding triple instances can be generated. These triple instances are organized into a graph structure, namely the knowledge graph constructed in this application.

[0047] As an alternative method to extract triple instances matching entity and relation types from multi-source data, the natural language understanding capabilities of a Large Language Model (LLM) can be leveraged to directly extract entities and relations from multi-source data and generate corresponding triple instances without distinguishing data types. Specifically, the entity type set and relation type set are encoded into structured prompt word templates; target prompt words are generated based on the prompt word templates and multi-source data; these target prompt words are then input into the LLM, enabling the LLM to extract entities and relations matching the entity type and relation type sets from the multi-source data under the guidance of the target prompt words, and output corresponding triple instances for use in constructing a knowledge graph.

[0048] In practical applications, due to differences in expression habits, terminology, and data precision among different data sources, entities and relations extracted from multi-source data may have issues such as homonyms, synonyms, conflicting relationships, or redundancy. To improve the accuracy, consistency, and completeness of the constructed knowledge graph, the triples extracted from multi-source data are called candidate triples. These candidate triples are then further processed through knowledge fusion to obtain the final triple instances used to construct the knowledge graph.

[0049] Optionally, the knowledge fusion process includes at least one of the following: performing semantic similarity calculation on entities in candidate triples (e.g., similarity measurement based on context embedding or domain dictionary), normalizing multiple entities with semantic similarity greater than a preset similarity threshold to the same entity identifier, thereby achieving entity alignment; identifying multiple triples in candidate triples where the head and tail entities are the same but the relation type or relation value conflicts, and filtering out triples whose data source credibility is lower than a preset credibility threshold by combining the credibility assessment of each triple's corresponding data source (e.g., source authority, collection time, measurement accuracy, etc.), thereby completing relation fusion and conflict resolution, and thus achieving relation alignment.

[0050] After knowledge fusion processing, a set of triple instances with consistent structure, accurate semantics, and low redundancy can be obtained. Subsequently, a graph database (such as Neo4j) is selected to store the constructed knowledge graph, and the fused entity, attribute, and relation information is persistently stored in the form of a graph structure, supporting efficient subgraph query, path reasoning, and semantic retrieval, providing reliable knowledge support for subsequent prediction of electromagnetic wave cross-domain propagation.

[0051] like Figure 2After obtaining the propagation scene parameters of the target electromagnetic wave and the pre-constructed knowledge graph, the propagation scene parameters are further input into a first feedforward neural network (FNN) to output a propagation scene feature vector. h The knowledge graph is input into a graph neural network (GNN) to output a feature vector of the knowledge graph. .

[0052] Optionally, the graph neural network can be a message-passing neural network (MPNN).

[0053] In an optional embodiment, the knowledge graph is input into a graph neural network to output a knowledge graph feature vector using the graph neural network, including: First, the knowledge graph is represented as a directed graph G=(V, E) consisting of nodes and directed edges, where node V represents the set of entities in the knowledge graph, and edge E represents the set of relationships between entities. Next, for each node, based on the feature vectors of each node and its corresponding neighbor nodes at the current time step, as well as the edge features of the directed edges connecting each node and its corresponding neighbor nodes, the message to be passed from the neighbor nodes in the next time step is calculated through the message function.

[0054] To facilitate understanding, let's take an example. Suppose that the characteristic of the "collision frequency" of entity nodes in the directed graph G corresponding to the knowledge graph is: At the current time step t, the feature vector of the collision frequency node v is: The adjacent nodes of the collision frequency are the entity "electron density" and "NO". + The eigenvectors of the adjacent nodes at time step t are: "ion density" and "ionospheric dielectric constant". Based on this assumption, at time step t+1, the collision frequency node v receives messages from its neighboring nodes. , can be represented as: Where N(v) represents all neighboring nodes of the collision frequency node v. The collision frequency is node v and node w Edge features. Vector This means putting all nodes w The feature vectors are concatenated in one dimension, that is: Among them, message function The expression is: in, It is a learnable edge feature matrix that describes all nodes. w It indicates and points out the situation. Figure 5 This is a schematic diagram of the composition of an edge feature matrix provided in an embodiment of this application, such as... Figure 5 As shown, based on the above assumptions, the node w Including: collision frequency v, electron density, NO + Ion density and ionospheric dielectric constant. Its edge characteristics can be described by two features: outward pointing and inward pointing. Specifically, NO + Ion density and electron density, respectively, are derived from the outward-pointing collision frequency v, corresponding to B. 14 and B 24 The collision frequency v is pointed out from the inner index to the dielectric constant of the ionosphere, corresponding to B. 43 .

[0055] Then, using the vertex update function, the feature vector of each node at the current time step is updated in the next time step by combining the feature vector of each node at the current time step with the messages passed from neighboring nodes.

[0056] Continuing with the example above, the feature vector (or node state) corresponding to the collision frequency node v at time step t+1 can be represented as: in, U t The vertex update function.

[0057] Optionally, the vertex update function can be implemented using a gated recurrent unit (GRU). Correspondingly, .

[0058] Finally, after completing message passing at the preset time step, the knowledge graph feature vector is determined based on the final feature vector of each node in the directed graph. .

[0059] In the specific implementation process, assuming the preset time step is T, for any entity node v in the directed graph G corresponding to the knowledge graph, the read function is used. R Calculate the feature vector corresponding to entity node v at time step T of the last node update. .

[0060] The reading function R can be expressed as: Among them, operations This indicates element-wise multiplication. i , j These represent two fully connected networks. i There is an outer sigmoid function σ Neural networks i The input is the initial state of the node. and final state The input to j is only the final state of the node.

[0061] Using the above method, we can obtain the feature vector of each entity node in the directed graph G corresponding to the knowledge graph at time step T. These feature vectors constitute the feature vector of the knowledge graph. .

[0062] In this scheme, graph neural networks are used to extract feature vectors from knowledge graphs. During the process, each round of message passing dynamically aggregates information from neighboring nodes based on the directed graph structure of the knowledge graph, and iteratively optimizes the embedding representation of each entity node through a vertex update function. Since the factors involved in the cross-domain propagation of electromagnetic waves (such as ionospheric state, seawater conductivity, and emission parameters) are structurally organized in the knowledge graph as entities and relationships, the graph neural network can capture semantic dependencies in the local neighborhood and even the global topology layer by layer through multiple rounds of message passing. This ensures that the final representation of each node not only includes its own attributes but also incorporates contextual information from other physically related entities. This graph-based deep information propagation mechanism enables the graph neural network to naturally encode the complex causal chains and environmental coupling relationships in the electromagnetic wave propagation process, thereby mapping the entire knowledge graph completely and compactly into a continuous vector space, generating a physically interpretable knowledge graph feature vector K.

[0063] After obtaining the feature vector of the propagation scene h and knowledge graph feature vectors Then, further, based on the feature vector of the propagation scenario... h and knowledge graph feature vectors Determine the cross-domain propagation feature vector This is used as input to a second feedforward neural network, which outputs a potential function characterizing the electromagnetic field distribution of the ground-ionospheric waveguide.

[0064] In this embodiment of the application, in order to realize the propagation scene feature vector h and knowledge graph feature vectors The effective integration of these technologies enhances the sensitivity and interpretability of cross-domain propagation prediction models to key parameters, and employs an attention mechanism to dynamically calculate the feature vector of the propagation scenario. h and knowledge graph feature vectors The feature correlation between the features is determined, and the feature vectors of the propagation scene are weighted and fused based on the feature correlation to obtain the fused feature vector C; the fused feature vector C and the propagation scene feature vector are then combined. h The vectors are concatenated to obtain the cross-domain propagation feature vector. .

[0065] In the specific implementation process, firstly, the feature vector of the propagation scene is... h The propagation scenario parameters are divided into multiple scenario sub-vectors based on parameter categories. For example, the propagation scenario feature vectors can be divided according to parameter categories such as transmitter height, receiver height, and frequency. h It is divided into scene sub-vectors corresponding to the height of the transmitting point, scene sub-vectors corresponding to the height of the receiving point, and scene sub-vectors corresponding to the frequency, etc.

[0066] Parameters for propagation scenarios Any category parameter in x n (e.g., launch point height data), and its corresponding scene sub-vector is called the target scene sub-vector h. n ,in, It is understood that the target scene sub-vector mentioned in this embodiment is the scene sub-vector corresponding to any category of parameters in the propagation scene parameters, that is, the propagation scene feature vector. h Any scene sub-vector among the multiple scene sub-vectors obtained by partitioning.

[0067] Next, determine the sub-vector h of the target scene within the feature vector of the knowledge graph. n The knowledge subvector corresponding to the corresponding parameter category k ,in, For example, when the target scene subvector h n When the corresponding parameter category is launch point height data, the knowledge sub-vector k For knowledge graph feature vectors The subvector corresponding to the launch point height data.

[0068] Next, the target scene sub-vector h is calculated using a pre-defined weighted model. n With knowledge subvectors k The correlation between target features is calculated, and the attention weights corresponding to the target scene sub-vectors are determined based on this correlation. Here, the target scene sub-vector h... n With knowledge subvectors k The correlation between the target features is also the attention distribution between the two.

[0069] Optionally, the target scene subvector h n With knowledge subvectors k The correlation between target features (i.e., attention distribution) can be calculated using the following formula: in, N represents the total number of parameter categories included in the propagation scenario parameters, and S represents the attention scoring function.

[0070] Alternatively, the attention mechanism score can be obtained using the following weighted model:

[0071] in, v , W , U These are the learnable network parameters of the weighted model.

[0072] The propagation scenario parameters can be calculated using the methods described above. The feature correlation between the scene sub-vectors corresponding to each type of parameter and the corresponding knowledge sub-vectors is determined. Then, based on the multiple attention weights corresponding to the multiple scene sub-vectors, the multiple scene sub-vectors are weighted and summed to obtain the fused feature vector C, which is the propagated scene feature vector. h Attention values ​​for all scene subvectors.

[0073] The calculation process of the fused feature vector C can be expressed as follows: Finally, the feature vector C and the propagation scene feature vector will be fused. h The vectors are concatenated to obtain the cross-domain propagation feature vector. .

[0074] In this embodiment, to achieve efficient fusion of the propagation scenario feature vector and the knowledge graph feature vector, and to improve the cross-domain propagation prediction model's ability to follow physical laws and its sensitivity to key parameters, an attention mechanism is employed to dynamically adjust the intensity of knowledge graph information introduction in the intermediate layers of the model. This mechanism can adaptively select, weight, and integrate relevant semantic information from the knowledge graph during the training and usage phases, thereby avoiding model convergence difficulties caused by introducing too much irrelevant knowledge in the early stages, while highlighting parameters that have a decisive impact on the current propagation scenario.

[0075] Specifically, during the training phase, the attention mechanism automatically assigns attention weights by learning the semantic association strength between various parameter categories in the propagation scenario parameters (such as the latitude, longitude, altitude, and frequency of the transmission point) and corresponding entities in the knowledge graph (such as ionospheric state, surface conductivity, and seawater salinity). For example, in the extremely low frequency (ELF) band (3–30 Hz), electromagnetic wave wavelengths can reach thousands of kilometers, while the typical transmission source height (such as tens of meters below the sea surface or a ground antenna) is much smaller than the wavelength. Theoretically, the transmission point can be approximated as being located on the Earth's surface. In this case, the influence of the transmission point height on propagation characteristics is significantly weaker than that of the transmission point latitude and longitude—the latter determining the geometries of the Earth-ionospheric waveguide through which the signal traverses and the environmental properties of the medium it passes through. The attention mechanism can capture this physical prior and automatically assign higher weights to latitude and longitude-related features during training, while reducing the contribution of altitude-related features.

[0076] This mechanism also functions during the usage (inference) phase: given new propagation scenario parameters, the model dynamically focuses on the most relevant substructures in the knowledge graph based on the learned attention distribution, enabling on-demand access to domain knowledge. For example, when the predicted path crosses the high-latitude auroral zone, the model will increase its focus on ionospheric disturbances; while in the equatorial calm zone, it will place greater emphasis on seawater conductivity profiles and crustal conductivity information.

[0077] Through the above methods, the attention mechanism not only enables fine-grained control of knowledge graph information in the intermediate layer, but also enhances the physical consistency, interpretability, and generalization ability of the end-to-end cross-domain radio wave propagation prediction model. Especially in scenarios with sparse data or sudden environmental changes, the model can still maintain reasonable predictions based on knowledge guidance, significantly improving the reliability and robustness of predicting the propagation status of ultra-low frequency / extremely low frequency electromagnetic waves in complex ocean-atmosphere-ionospheric coupled environments.

[0078] like Figure 2 As shown, after obtaining the cross-domain propagation feature vector Subsequently, the cross-domain propagation feature vector is input into a second feedforward neural network to output a potential function (U, V) characterizing the electromagnetic field distribution of the earth-ionospheric waveguide. Here, the potential functions U and V correspond to the scalar potential and vector magnetic potential, respectively, and their spatial distribution is determined by the cross-domain propagation feature vector. The determination is used to describe the propagation characteristics of target electromagnetic waves in the region above the atmosphere-ocean boundary.

[0079] Based on the physical relationship between the potential function (U, V) and the electromagnetic field, the electric field above the sea surface can be calculated using standard electromagnetic field theory formulas. E P ,magnetic field H P That is, the prediction result of the first electromagnetic field.

[0080] Specifically, the electric and magnetic fields can be derived from the gradient and curl of the potential function, respectively, for example: Where A is a vector potential, denoted by V, the above process realizes the mapping from the high-dimensional features output by the neural network to physically interpretable field quantities, ensuring the physical consistency of the prediction results.

[0081] Furthermore, the conductivity stratification structure of seawater along the depth direction is determined based on the knowledge graph, and the electric field at a specified depth h below the specified sea surface is calculated using the first electromagnetic field prediction results and the conductivity stratification structure. E Ph ,magnetic field H PhThis refers to the second electromagnetic field prediction result. The specified depth is determined based on the propagation scenario parameters. For example, the specified depth can be the depth corresponding to the target electromagnetic wave receiving position in the propagation scenario parameters. For instance, if the receiving point is located 100 meters below the sea surface, then the specified depth h = 100m is set.

[0082] Specifically, since the knowledge graph encompasses the physical relationship between marine environmental parameters (such as salinity, temperature, and pressure) and conductivity, as well as the distribution patterns of typical conductivity profiles in different sea areas, semantic reasoning can be performed on this knowledge using methods such as graph neural networks to deduce the variation law of seawater conductivity with depth under current propagation conditions. Based on this variation law, a conductivity stratification method that can accurately reflect the inhomogeneity of the seawater medium can be selected to divide the depth-conductivity variation relationship, thereby obtaining the conductivity stratification structure of seawater along the depth direction.

[0083] Furthermore, the predicted results of the first electromagnetic field above the sea surface (i.e., the electric field) E P ,magnetic field H P The model uses a pre-constructed air-conductivity-layered seawater cross-medium propagation calculation model, based on layered medium propagation theories (such as the transmission line matrix method, finite element method, or numerical solution method for wave equations), to calculate the attenuation, refraction, and reflection characteristics of electromagnetic waves in each conductivity layer. It is worth noting that during this process, the dielectric constant, conductivity, and other parameters of each layer are assigned values ​​based on the conductivity layered structure derived from the aforementioned knowledge graph.

[0084] Finally, the electric field at a specified depth h is obtained through iterative solutions or direct analytical calculations. E Ph ,magnetic field H Ph That is, the prediction result of the second electromagnetic field.

[0085] In this scheme, on the one hand, a dual-branch fusion architecture of a first feedforward neural network and a graph neural network is used to extract the propagation scene feature vector and knowledge graph feature vector of the propagation scene parameters, respectively, and collaboratively generate cross-domain propagation feature vectors. This drives the second feedforward neural network to output a potential function representing the electromagnetic field distribution of the earth-ionospheric waveguide, so as to solve the electromagnetic field above the sea surface based on the potential function and perform underwater electromagnetic field inference, realizing the joint prediction of the electromagnetic field distribution in the waveguide region above the sea surface and the medium below the sea surface. On the other hand, by introducing a pre-constructed knowledge graph in the field of electromagnetic wave propagation, the physical relationship between the multi-source heterogeneous factors affecting the cross-domain propagation of electromagnetic waves (such as ionospheric state, seawater medium properties, environmental parameters, etc.) and propagation characteristics (such as field strength, attenuation, penetration depth, etc.) is explicitly modeled, and the graph neural network is used to structure and encode it, so that the second feedforward neural network has knowledge-guided reasoning ability. This effectively overcomes the problems of physical inconsistency and poor generalization of pure data-driven methods in small sample or out-of-distribution scenarios, and also ensures that the medium model on which the underwater propagation calculation depends has physical interpretability and environmental adaptability. In summary, this approach not only improves the accuracy and reliability of propagation prediction of ultra-low frequency / extremely low frequency electromagnetic waves in complex cross-domain environments (earth-ionosphere-atmosphere-seawater), but also significantly enhances the model's adaptability and robustness to dynamic marine environments, providing highly reliable technical support for key tasks such as underwater long-range communication, target detection, and electromagnetic situation prediction.

[0086] Optionally, the knowledge graph-based method for predicting the cross-domain propagation of ultra-low frequency / extremely low frequency electromagnetic waves provided in this application can be executed by an electromagnetic wave cross-domain propagation prediction model. The electromagnetic wave cross-domain propagation prediction model may include: a first feedforward neural network, a graph neural network, and a second feedforward neural network.

[0087] Next, combined Figure 6 The training process of the electromagnetic wave cross-domain propagation prediction model in the embodiments of this application is described.

[0088] Figure 6 A flowchart of a method for training an electromagnetic wave cross-domain propagation model provided in this application embodiment is shown below. Figure 6 As shown, it may include the following steps: 601. Obtain training samples, which include propagation scenario parameter samples, knowledge graph, and measured electromagnetic field data above the sea surface corresponding to the sample propagation scenario parameters.

[0089] 602. Input the propagation scenario parameter samples and knowledge graph into the electromagnetic wave cross-domain propagation prediction model, so as to use the electromagnetic wave cross-domain propagation prediction model to output the sample potential function that characterizes the electromagnetic field distribution of the earth-ionospheric waveguide. The sample potential function is used to calculate the predicted electromagnetic field above the sea surface.

[0090] It is understood that the working principle of the model during use is the same or similar to that of the model during training. Therefore, in this embodiment, the process of obtaining the sample bit function through the first feedforward neural network, the graph neural network and the second feedforward neural network will not be described again. The specific process can be referred to the aforementioned embodiment.

[0091] The data-driven loss term is calculated based on the difference between the predicted electromagnetic field and the measured electromagnetic field data above the sea surface.

[0092] Calculate the Helmholtz equation residuals for the sample potential function in the Earth-ionospheric waveguide region; calculate the boundary condition residuals based on the predicted electromagnetic field and the impedance boundary conditions at the air-seawater interface; and use the Helmholtz equation residuals and the boundary condition residuals as physical constraint loss terms.

[0093] A loss function is constructed based on the data-driven loss term and the physical constraint loss term.

[0094] 606. With the goal of minimizing the loss function, perform end-to-end training on the first feedforward neural network, the graph neural network, and the second feedforward neural network until convergence.

[0095] In an optional embodiment, the loss function constructed based on the data-driven loss term and the physical constraint loss term can be expressed as:

[0096] in, This represents the data-driven loss term. Represents the physical constraint loss term. These are the weighting coefficients corresponding to the data-driven loss term. These are the weighting coefficients corresponding to the physical constraint loss term. and Used to balance the proportion of loss in the loss function between data-driven loss terms and physical constraint loss terms.

[0097] The electromagnetic wave transdomain propagation prediction model trained in this scheme incorporates the Helmholtz equation residuals and impedance boundary conditions at the air-seawater interface as physical constraints. This allows the model to not only learn the statistical regularities between the propagation scenario parameter samples and the correlation between target factors influencing electromagnetic wave propagation and propagation characteristics in the knowledge graph during training, but also internalize the physical essence that electromagnetic waves must follow to propagate in conductive marine media. This fusion mechanism enables the model to effectively suppress the non-physical interpretations easily generated by pure data-driven methods (i.e., machine learning methods) when facing dynamic, non-uniform, and spatiotemporally continuously evolving marine environments, while overcoming the systematic biases caused by the oversimplification of marine environments in traditional analytical or numerical simulations. Physical constraints ensure that the output field satisfies the fundamental laws of electromagnetic fields at any spatiotemporal point, while knowledge guidance allows the network to focus on environmental variables and their evolution paths that play a dominant role in electromagnetic wave propagation from the input information. Thus, even in complex transdomain propagation scenarios, it can accurately characterize the changing trends of key characteristics such as attenuation, phase shift, and propagation depth of ultra-low frequency / extremely low frequency electromagnetic waves. Thus, while maintaining the strong fitting ability of deep learning, the model achieves physical interpretability and extrapolation robustness, enabling accurate and reliable prediction of electromagnetic wave propagation behavior in cross-domain scenarios.

[0098] The following describes in detail one or more embodiments of the knowledge graph-based transdomain propagation prediction device for ultra-low frequency / extremely low frequency electromagnetic waves according to this application. Those skilled in the art will understand that these devices can all be configured using commercially available hardware components through the steps taught in this solution.

[0099] Figure 7 A schematic diagram of a knowledge graph-based transdomain propagation prediction device for ultra-low frequency / extremely low frequency electromagnetic waves is provided as an embodiment of this application. Figure 7 As shown, the device includes: an acquisition module 11, a first prediction module 12, and a second prediction module 13.

[0100] The acquisition module 11 is used to acquire the propagation scene parameters of the target electromagnetic wave and a pre-constructed knowledge graph. The target electromagnetic wave is an ultra-low frequency electromagnetic wave or an extremely low frequency electromagnetic wave. The knowledge graph is used to describe the relationship between the target factors affecting the cross-domain propagation of electromagnetic waves and the propagation characteristics of electromagnetic waves.

[0101] The first prediction module 12 is used to input the propagation scene parameters into a first feedforward neural network to output a propagation scene feature vector using the first feedforward neural network; input the knowledge graph into a graph neural network to output a knowledge graph feature vector using the graph neural network; determine a cross-domain propagation feature vector based on the propagation scene feature vector and the knowledge graph feature vector; and input the cross-domain propagation feature vector into a second feedforward neural network to output a potential function characterizing the electromagnetic field distribution of the earth-ionospheric waveguide using the second feedforward neural network, wherein the potential function is used to calculate the first electromagnetic field prediction result above the sea surface.

[0102] The second prediction module 13 is used to determine the conductivity layering structure of seawater along the depth direction based on the knowledge graph, and to calculate the second electromagnetic field prediction result at a specified depth below the specified sea surface using the first electromagnetic field prediction result and the conductivity layering structure, wherein the specified depth is determined according to the propagation scenario parameters.

[0103] Optionally, the first prediction module 12 is specifically used to: determine the feature correlation between the propagation scene feature vector and the knowledge graph feature vector through an attention mechanism, and perform weighted fusion processing on the propagation scene feature vector according to the feature correlation to obtain a fused feature vector; and concatenate the fused feature vector and the propagation scene feature vector to obtain a cross-domain propagation feature vector.

[0104] Optionally, the first prediction module 12 is further configured to: divide the propagation scene feature vector into multiple scene sub-vectors according to the parameter categories of the propagation scene parameters; determine the knowledge sub-vector in the knowledge graph feature vector that corresponds to the parameter category of the target scene sub-vector, wherein the target scene sub-vector is any one of the multiple scene sub-vectors; calculate the target feature correlation between the target scene sub-vector and the knowledge sub-vector using a preset weighted model, and determine the attention weight corresponding to the target scene sub-vector based on the target feature correlation; and perform weighted summation on the multiple scene sub-vectors based on the multiple attention weights corresponding to the multiple scene sub-vectors to obtain a fused feature vector.

[0105] Optionally, in the process of acquiring the pre-constructed knowledge graph, the acquisition module 11 is specifically used to: define a set of entity types and a set of relation types related to the cross-domain propagation task of ultra-low frequency / extremely low frequency electromagnetic waves, wherein the set of entity types contains multiple entity types related to electromagnetic wave propagation, and the set of relation types is used to describe the interaction relationships between the multiple entity types; based on the set of entity types and the set of relation types, extract triple instances that conform to the entity types and relation types from the pre-acquired multi-source data related to cross-domain propagation of electromagnetic waves, and organize the triple instances into a graph structure to form the knowledge graph.

[0106] Optionally, the acquisition module 11, in the process of acquiring the pre-built knowledge graph, is further specifically used for: encoding the entity type set and the relation type set into a structured prompt word template; generating target prompt words according to the prompt word template and the multi-source data, inputting the target prompt words into a large language model, so that the large language model, under the guidance of the target prompt words, extracts entities and relations matching the entity type set and the relation type set from the multi-source data, and outputs candidate triples; performing knowledge fusion processing on the candidate triples to obtain triple instances; wherein, the knowledge fusion processing includes at least one of the following: performing semantic similarity calculation on the entities in the candidate triples, normalizing entities with a similarity threshold greater than a set similarity threshold to the same entity identifier; identifying multiple triples in the candidate triples that have the same entity but conflicting relations, and filtering out target triples in the multiple triples whose corresponding data source credibility is lower than a set credibility threshold.

[0107] Optionally, the first prediction module 12 is further configured to: represent the knowledge graph as a directed graph consisting of nodes and directed edges, wherein the nodes correspond to entities in the knowledge graph and the directed edges correspond to relationships between entities; for each node, calculate the message transmitted from the neighboring nodes in the next time step using a message function based on the feature vectors of each node and its corresponding neighboring nodes at the current time step and the edge features of the directed edges connecting each node and its corresponding neighboring nodes; update the feature vector of each node in the next time step using a vertex update function, combining the feature vector of each node at the current time step and the message transmitted from the neighboring nodes; and after completing the message transmission for a preset time step, determine the knowledge graph feature vector based on the final feature vector of each node in the directed graph.

[0108] Optionally, the device further includes a model training module for training an electromagnetic wave cross-domain propagation prediction model, the electromagnetic wave cross-domain propagation prediction model including: a first feedforward neural network, the graph neural network, and a second feedforward neural network. Specifically, the training process of the electromagnetic wave cross-domain propagation prediction model includes: acquiring training samples, the training samples including propagation scenario parameter samples, the knowledge graph, and measured electromagnetic field data above the sea surface corresponding to the sample propagation scenario parameters; inputting the propagation scenario parameter samples and the knowledge graph into the electromagnetic wave cross-domain propagation prediction model, so as to use the electromagnetic wave cross-domain propagation prediction model to output a sample potential function characterizing the electromagnetic field distribution of the earth-ionospheric waveguide, the sample potential function being used to calculate the predicted electromagnetic field above the sea surface; and calculating the predicted electromagnetic field and the measured electromagnetic field above the sea surface. The differences between the data are used to calculate the data-driven loss term; the Helmholtz equation residuals of the sample potential functions in the Earth-ionospheric waveguide region are calculated; the boundary condition residuals are calculated based on the predicted electromagnetic field and the impedance boundary conditions at the air-seawater interface; the Helmholtz equation residuals and the boundary condition residuals are used as physical constraint loss terms; a loss function is constructed based on the data-driven loss term and the physical constraint loss term; and the first feedforward neural network, the graph neural network, and the second feedforward neural network are trained end-to-end with the goal of minimizing the loss function until convergence.

[0109] Figure 7 The device shown can perform the steps described in the foregoing embodiments. For detailed execution process and technical effects, please refer to the description in the foregoing embodiments, which will not be repeated here.

[0110] In one possible design, the above Figure 7 The structure of the knowledge graph-based transdomain propagation prediction device for ultra-low frequency / extremely low frequency electromagnetic waves shown can be implemented as an electronic device, such as... Figure 8 As shown, the electronic device may include: a memory 21, a processor 22, and a communication interface 23. The memory 21 stores a computer program, which, when executed by the processor 22, enables the processor 22 to at least implement the knowledge graph-based method for predicting the cross-domain propagation of ultra-low frequency / extremely low frequency electromagnetic waves as provided in the foregoing embodiments.

[0111] The aforementioned memory 21 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random-Access Memory (SRAM), Electrically Erasable Programmable Read Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0112] Accordingly, embodiments of this application also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, enables the processor to implement the steps in the above-described method embodiments. The computer-readable storage medium includes volatile or non-volatile components, or a combination thereof, and can be removable or non-removable. Examples of computer-readable storage media include, but are not limited to, phase-change random access memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random-access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), flash memory or other memory technologies, CD-ROM, Digital Video Disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transmission medium. Accordingly, this application also provides a computer program product, which includes a computer program or instructions. When the computer program or instructions are executed by a processor, the processor is able to implement the steps in the above method embodiments. It should be understood that each step or combination of steps in the above method flow can be implemented by the computer program or instructions. In addition, these computer programs or instructions can be applied to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable knowledge graph-based ultra-low frequency / extremely low frequency electromagnetic wave cross-domain propagation prediction device, so that the processor of the general-purpose computer, special-purpose computer, embedded processor, or other programmable knowledge graph-based ultra-low frequency / extremely low frequency electromagnetic wave cross-domain propagation prediction device can be implemented as a means to implement the corresponding functions in the above method embodiments.

[0113] The device embodiments described above are merely illustrative, and the units described as separate components may or may not be physically separate. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0114] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of a necessary general-purpose hardware platform, or by a combination of hardware and software. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a computer product. This application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0115] Finally, it should be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0116] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A knowledge graph-based method for predicting the cross-domain propagation of ultra-low frequency / extremely low frequency electromagnetic waves, characterized in that, include: The propagation scenario parameters of the target electromagnetic wave and a pre-constructed knowledge graph are obtained. The target electromagnetic wave is an ultra-low frequency electromagnetic wave or an extremely low frequency electromagnetic wave. The knowledge graph is used to describe the relationship between the target factors affecting the cross-domain propagation of the electromagnetic wave and the propagation characteristics of the electromagnetic wave. The propagation scenario parameters are input into a first feedforward neural network to output a propagation scenario feature vector using the first feedforward neural network. The knowledge graph is input into a graph neural network to output a knowledge graph feature vector using the graph neural network. Based on the propagation scenario feature vector and the knowledge graph feature vector, determine the cross-domain propagation feature vector; The cross-domain propagation feature vector is input into a second feedforward neural network to utilize the output of the second feedforward neural network to characterize the electromagnetic field distribution of the earth-ionospheric waveguide. The potential function is used to calculate the first electromagnetic field prediction result above the sea surface. The conductivity layering structure of seawater along the depth direction is determined based on the knowledge graph. Then, using the first electromagnetic field prediction result and the conductivity layering structure, the second electromagnetic field prediction result at a specified depth below the specified sea surface is calculated. The specified depth is determined based on the propagation scenario parameters.

2. The method according to claim 1, characterized in that, The step of determining the cross-domain propagation feature vector based on the propagation scenario feature vector and the knowledge graph feature vector includes: The feature correlation between the propagation scenario feature vector and the knowledge graph feature vector is determined by an attention mechanism, and the propagation scenario feature vector is weighted and fused according to the feature correlation to obtain a fused feature vector. The fused feature vector and the propagation scenario feature vector are concatenated to obtain the cross-domain propagation feature vector.

3. The method according to claim 2, characterized in that, The step of determining the feature relevance between the propagation scenario feature vector and the knowledge graph feature vector through an attention mechanism, and then performing a weighted fusion process on the propagation scenario feature vector based on the feature relevance to obtain a fused feature vector, includes: The propagation scene feature vector is divided into multiple scene sub-vectors according to the parameter categories of the propagation scene parameters; Determine the knowledge sub-vector corresponding to the parameter category of the target scene sub-vector in the knowledge graph feature vector, wherein the target scene sub-vector is any one of the plurality of scene sub-vectors; The target feature relevance between the target scene sub-vector and the knowledge sub-vector is calculated using a preset weighted model, and the attention weight corresponding to the target scene sub-vector is determined based on the target feature relevance. Based on the multiple attention weights corresponding to the multiple scene sub-vectors, the multiple scene sub-vectors are weighted and summed to obtain a fused feature vector.

4. The method according to claim 1, characterized in that, The knowledge graph is constructed in the following manner: Define a set of entity types and a set of relationship types related to the cross-domain propagation task of ultra-low frequency / extremely low frequency electromagnetic waves. The set of entity types contains multiple entity types related to electromagnetic wave propagation, and the set of relationship types is used to describe the interaction relationships between the multiple entity types. Based on the set of entity types and the set of relation types, triple instances that conform to the entity type and relation type are extracted from the pre-acquired multi-source data related to the cross-domain propagation of electromagnetic waves, and the triple instances are organized into a graph structure to form the knowledge graph.

5. The method according to claim 4, characterized in that, The step of extracting triple instances that conform to the entity type and relation type from pre-acquired multi-source data based on the entity type set and relation type set includes: The entity type set and the relation type set are encoded into structured prompt word templates; Target prompt words are generated based on the prompt word template and the multi-source data. The target prompt words are then input into a large language model, so that the large language model, guided by the target prompt words, extracts entities and relations from the multi-source data that match the entity type set and the relation type set, and outputs candidate triples. The candidate triples are subjected to knowledge fusion processing to obtain triple instances; The knowledge fusion processing includes at least one of the following: performing semantic similarity calculation on entities in the candidate triples, normalizing entities with similarity thresholds greater than a set similarity threshold to the same entity identifier; identifying multiple triples in the candidate triples that have the same entity but conflicting relationships, and filtering out target triples in the multiple triples whose corresponding data source credibility is lower than a set credibility threshold.

6. The method according to claim 1, characterized in that, The step of inputting the knowledge graph into a graph neural network to output a knowledge graph feature vector using the graph neural network includes: The knowledge graph is represented as a directed graph consisting of nodes and directed edges, where the nodes correspond to entities in the knowledge graph and the directed edges correspond to relationships between entities. For each node, based on the feature vectors of each node and its corresponding neighbor nodes at the current time step, as well as the edge features of the directed edges connecting each node and its corresponding neighbor nodes, the message to be passed from the neighbor nodes in the next time step is calculated using the message function. Using the vertex update function, and combining the feature vector of each node at the current time step with the messages passed from neighboring nodes, update the feature vector of each node at the next time step. After the message transmission at the preset time step is completed, the knowledge graph feature vector is determined based on the final feature vector of each node in the directed graph.

7. The method according to any one of claims 1-6, characterized in that, The method is executed through an electromagnetic wave cross-domain propagation prediction model, which includes: a first feedforward neural network, a graph neural network, and a second feedforward neural network. The training of the electromagnetic wave cross-domain propagation prediction model includes: Acquire training samples, which include propagation scenario parameter samples, the knowledge graph, and measured electromagnetic field data above the sea surface corresponding to the propagation scenario parameters of the samples; The propagation scenario parameter samples and the knowledge graph are input into the electromagnetic wave cross-domain propagation prediction model, so as to use the electromagnetic wave cross-domain propagation prediction model to output a sample potential function that characterizes the electromagnetic field distribution of the earth-ionospheric waveguide. The sample potential function is used to calculate the predicted electromagnetic field above the sea surface. The data-driven loss term is calculated based on the difference between the predicted electromagnetic field and the measured electromagnetic field data above the sea surface. Calculate the Helmholtz equation residuals of the sample potential function in the Earth-ionospheric waveguide region; Calculate the boundary condition residuals based on the predicted electromagnetic field and the impedance boundary conditions at the air-seawater interface; The residuals of the Helmholtz equation and the residuals of the boundary conditions are used as physical constraint loss terms. Based on the data-driven loss term and the physical constraint loss term, a loss function is constructed; With the goal of minimizing the loss function, the first feedforward neural network, the graph neural network, and the second feedforward neural network are trained end-to-end until convergence.

8. An electronic device, characterized in that, include: The system includes a memory, a processor, and a communication interface; wherein the memory stores a computer program, which, when executed by the processor, causes the processor to perform the knowledge graph-based method for predicting the cross-domain propagation of ultra-low frequency / extremely low frequency electromagnetic waves as described in any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor of an electronic device, causes the processor to perform the knowledge graph-based method for predicting the cross-domain propagation of ultra-low frequency / extremely low frequency electromagnetic waves as described in any one of claims 1 to 7.

10. A computer program product, characterized in that, include: A computer program or instruction, when executed by a processor of an electronic device, causes the processor to perform the knowledge graph-based method for predicting the transdomain propagation of ultra-low frequency / extremely low frequency electromagnetic waves as described in any one of claims 1 to 7.

Images