Maritime electromagnetic spectrum reconstruction method based on spatio-temporal spectrum completion and multi-layer grid fusion of SNN distillation

Abstract

The application discloses a maritime electromagnetic spectrum reconstruction method based on SNN distillation spatiotemporal spectrum completion and multi-layer grid fusion, solves the problems of low spectrum reconstruction accuracy, weak generalization ability and insufficient deployment efficiency, and belongs to the field of communication; including that a UAV flies according to a fixed route to obtain basic cruise data; data exceeding a threshold value is marked as a suspected target, a beacon is generated and broadcast; the optimal standby UAV is selected according to a scoring function to go to a target sea area to obtain tracking data; the tracking data and the basic cruise data are fused to obtain fused data and input into a teacher model; SNN is used to extract spatiotemporal embedding features and the teacher model is trained through a teacher loss function to output a teacher-completed spatiotemporal spectrum; a student-completed spatiotemporal spectrum is output through a knowledge distillation technology; a multi-layer grid fusion is performed through a feature fusion mechanism with local attention to output a reconstructed electromagnetic spectrum graph; and the application realizes completion and reconstruction and accurate restoration of small sample data.

Description

Technical Field

[0001] This invention belongs to the fields of machine learning and marine electromagnetic spectrum sensing technology, and relates to a marine electromagnetic spectrum reconstruction method based on SNN distillation for spatiotemporal spectrum completion and multi-layer grid fusion. Background Technology

[0002] With the rapid development of economic globalization, maritime and river transport-related businesses have experienced explosive growth. The rapid increase in related maritime wireless communication services and electromagnetic equipment has led to an increasingly complex maritime electromagnetic environment, and the scarce maritime electromagnetic spectrum resources face severe challenges such as low utilization and serious interference. Against this backdrop, efficient management and monitoring of the maritime electromagnetic spectrum has become a key means to improve spectrum utilization and ensure the normal operation of maritime radio services.

[0003] Traditional methods of marine electromagnetic spectrum monitoring mainly rely on fixed monitoring stations, mobile monitoring stations, or manned vessels at sea, as well as long-term data acquisition by multi-target satellites (such as publication number: CN118960697A, invention name: Multi-satellite marine observation system). These methods have inherent drawbacks such as high cost, time delay, and low accuracy. Moreover, these methods are difficult to obtain complete and real-time electromagnetic environment information for fixed target sea areas.

[0004] Unmanned Aerial Vehicle (UAV) spectrum monitoring technology has become a research hotspot for maritime electromagnetic spectrum monitoring and electromagnetic signal management due to its advantages such as high flexibility, rapid deployment and relatively low cost.

[0005] like Figure 1 As shown, in the existing field of marine electromagnetic spectrum and signal monitoring based on UAV swarms, the marine spectrum reconstruction method based on UAV swarm cyclic coverage sampling and spatial interpolation is often used to reconstruct the spectrum by collecting data in a distributed manner from multiple UAVs.

[0006] For example, in the invention publication CN119575487A, entitled "An Electromagnetic Retrieval Method for Seabed-Based Data Based on Unmanned Aerial Vehicles," a core architecture of "acquisition, transmission, and centralized processing" is adopted, aiming to achieve highly sensitive detection of weak quantum signals. However, in maritime applications, it relies on long-term fixed patrols, resulting in high randomness and low efficiency in data acquisition. Furthermore, wideband quantum signal receivers are prone to introducing significant quantum clutter interference caused by environmental scattering.

[0007] Existing UAV-based quantum geomagnetic sampling technology, while employing methods such as spectral interpolation to compensate for data sparsity, still suffers a significant decrease in the accuracy of quantum magnetic field reconstruction under extreme conditions where the monitoring area is vast but the number of flight platforms is limited. Particularly when dealing with rapidly changing quantum state signals or complex marine magnetic environments with dramatic spatial fluctuations, the limited number of sampling points makes it difficult to capture the fine characteristics of the quantum field, leading to a smoothing effect in the reconstructed spectrum and the loss of crucial quantum information.

[0008] Furthermore, even when compressed sensing technology is introduced into quantum magnetic field measurement, utilizing the sparsity and low rank of signals in the time, space, and frequency domains to improve acquisition efficiency, its "edge sampling-center reconstruction" technical paradigm still has inherent limitations when facing dynamic targets at sea and complex electromagnetic backgrounds:

[0009] 1) The sparsity assumption fails under strong sea surface clutter: Background electromagnetic clutter scattering from waves, islands, merchant ships, etc., is ubiquitous and powerful over vast sea areas, causing the collected mixed signals to no longer satisfy the sparsity assumption in the transform domain. A "non-sparse" signal will severely degrade the compressed sensing reconstruction model, resulting in reconstruction results filled with a large number of artifacts and noise, with the target signal being submerged.

[0010] 2) Low-rank constraints obliterate dynamic target features: Low-rank matrices and spectral completion techniques tend to recover smooth and stable background environments in signals. However, the electromagnetic signals generated by moving targets at sea exhibit "local singularities" or "subtle dynamic features" in the space-time-frequency three-dimensional tensor. Global low-rank constraint models treat these crucial dynamic trajectories as "noise" or "outliers" and over-smooth them, resulting in blurred target course contours and broken trajectories, failing to meet the requirements for accurate tracking.

[0011] Therefore, existing compressed sensing technology faces three major problems when dealing with complex maritime scenarios:

[0012] 1. Data sparsity: Traditional sampling suffers from data sparsity and background noise, resulting in poor reconstruction results and high computational costs for acquisition and spectral information extraction;

[0013] 2. Environmental dynamics: The complex and ever-changing marine environment makes electromagnetic signal propagation susceptible to signal attenuation and distortion interference caused by changes in the dielectric constant of seawater. However, traditional propagation models cannot accurately characterize these complex effects.

[0014] 3. Poor Spatiotemporal Correlation: Marine spectral characteristics exhibit strong correlations in both time and space. Temporally, influenced by tides, wave periodicity, and meteorological variations, spectral characteristics show clear time-varying patterns. Spatially, the continuity and homogeneity of the marine environment lead to highly similar spectral characteristics in adjacent areas. However, traditional methods such as ARIMA models and simple interpolation algorithms struggle to effectively capture these complex spatiotemporal features, limiting the accuracy of the final spectral reconstruction.

[0015] Current spectrum reconstruction methods offer efficiency advantages at the cost of sacrificing robustness in complex scenarios, sensitivity to dynamic features, and real-time performance. Summary of the Invention

[0016] To address the issues of sparsity in the marine electromagnetic spectrum caused by electromagnetic signal spectrum information acquisition during quantum magnetic field measurements, low spectrum reconstruction accuracy, weak generalization ability, insufficient deployment efficiency, and poor spatiotemporal correlation due to dynamic changes in the sea surface, this invention proposes a marine electromagnetic spectrum reconstruction method based on spatiotemporal spectrum completion and multi-layer grid fusion using SNN distillation. This method uses distributed acquisition by UAVs and combines it with a Spiking Neural Network (SNN) model to learn the spatiotemporal frequency characteristics and dynamic patterns of electromagnetic signals. A knowledge distillation mechanism is introduced to transfer the capabilities of complex models to lightweight models, thereby achieving efficient and accurate completion and reconstruction of small sample data on distributed nodes. Finally, a multi-layer grid strategy is used to achieve accurate restoration from local details to the global scene. This provides a directly transferable solution to problems faced in quantum magnetic field and quantum geomagnetic measurements (such as monitoring blind spots, loss of dynamic targets, and rigid resource allocation), promoting the application of quantum magnetic field measurement technology in real-world environments and improving data acquisition quality, thus ensuring the normal operation of maritime radio services. In terms of swarm intelligence, this method optimizes the collaborative sampling path planning of the UAV swarm through an improved ant colony algorithm, enabling the system to dynamically respond to target signals. This achieves adaptive swarm behavior, transitioning from static wide-area monitoring to dynamic precise tracking, significantly improving the global optimality of the sampling path and system resource efficiency. In terms of deep learning, it innovatively introduces a spiking neural network (SNN) to construct a knowledge distillation framework. This design not only endows the model with the ability to efficiently learn spatiotemporal frequency dynamic features from small sample data but also significantly improves the accuracy of spectrum reconstruction and its robustness in complex sea clutter environments through a two-stage training process using a teacher-student model. Ultimately, these two key technologies work together to overcome the limitations of traditional methods in terms of dynamic feature sensitivity and system real-time performance, providing a new technical solution for achieving high-precision and highly robust marine electromagnetic spectrum reconstruction.

[0017] The objective of this invention is specifically achieved through the following technical solutions:

[0018] A method for marine electromagnetic spectrum reconstruction based on SNN distillation for spatiotemporal spectrum completion and multi-layer grid fusion includes:

[0019] S1, combining the flight parameters of UAVs with the number of missions, constructs a route index covering the target sea area, consisting of time, frequency, and space, and assigns the route index to the corresponding UAVs to complete the initialization of the grid system; where UAVs include flying UAVs and standby UAVs;

[0020] S2, the flying UAV cruises along a fixed route and obtains the statistical feature tensor of the broadband marine electromagnetic spectrum signal continuously monitored and collected by the receiver; the missing mask of the statistical feature tensor is determined by a threshold, and after being assigned a value, it is mapped to the grid system, and the statistical feature tensor that passes the verification is retained as the basic cruise data.

[0021] S3 marks basic cruise data exceeding the threshold as suspected targets, generates and broadcasts beacons;

[0022] S4. Based on the scoring function, select the optimal standby UAV to go to the target sea area corresponding to the beacon, perform the refined tracking required by the mission, and obtain the tracking data of the suspected target;

[0023] S5 integrates tracking data with basic cruise data and stores it in the grid system to obtain fused data;

[0024] S6: Input the fused data into the teacher model constructed from the spatiotemporal dynamics term, signal frequency shape term, propagation physics term, and noise term; use the spiking neural network (SNN) to extract spatiotemporal embedding features from the spatiotemporal dynamics term and perform pulse encoding; construct the teacher loss function to train the teacher model and output the teacher-completed spatiotemporal spectrum;

[0025] S7 uses knowledge distillation technology to input the teacher-completed spatiotemporal spectrum and spatiotemporal embedding features into the student model constructed from spatiotemporal dynamics, signal frequency shape, and propagation physics terms; it constructs a student loss function to train the student model and outputs the student-completed spatiotemporal spectrum.

[0026] S8 introduces a feature fusion mechanism based on local attention to perform multi-layer grid fusion on the electromagnetic activity of the student-completed spatiotemporal spectrum and output the reconstructed electromagnetic spectrum map.

[0027] S9 verifies the spectral completion accuracy, physical consistency, algorithm efficiency, and robustness of the reconstructed electromagnetic spectrum, and outputs the verification results.

[0028] The beneficial effects of this invention are:

[0029] 1. A multi-scale grid system is constructed, achieving unified monitoring and precise analysis across the entire domain through the collaboration of coarse-grained and fine-grained grids. The coarse-grained grid, preferably with a spatial resolution of 10-30 km, undertakes wide-area baseline data acquisition and anomaly detection, forming the system's basic sensing layer. The fine-grained grid, preferably with a spatial resolution of 5-1 km, is responsible for refined analysis of signal characteristics and dynamic target tracking in key areas identified by the coarse-grained layer. The two grids work collaboratively through a dynamic resource scheduling mechanism: the coarse-grained layer continuously provides global spectral situational awareness, and when signal anomalies or specific patterns are detected, the fine-grained layer is automatically triggered to initiate high-precision acquisition of the target area. This multi-scale collaborative architecture ensures both the system's monitoring efficiency over vast sea areas and its ability to deeply analyze key areas, achieving an optimal balance between spectral reconstruction accuracy and coverage under resource constraints.

[0030] 2. Constructing a Basic Sensing Grid: Deploying several UAVs along pre-planned fixed routes, their receivers continuously collect broadband electromagnetic signals, forming a routine, wide-coverage monitoring network for the target sea area. This monitoring grid is initialized and managed using a standardized spatiotemporal frequency index, providing a unified fusion framework for all subsequent data (including basic cruise data and tracking data). This basic grid not only forms the skeleton of data acquisition but also constitutes the baseline background for electromagnetic situational awareness acquisition, enabling subsequent dynamic signals and static environments to be compared and fused within a unified coordinate system, fundamentally laying the foundation for data collaboration and sensing.

[0031] 3. Achieving Dynamic and Precise Response: When any UAV in the grid system detects a suspected target signal, it immediately broadcasts a target discovery beacon containing the target's location, characteristics, and intensity via a self-organizing network. Subsequently, based on a scoring function that comprehensively considers multiple dimensions such as distance, received power, and geometric relationships, the optimal node is dynamically selected from the standby UAVs in the region to proceed to the target area and perform refined tracking and acquisition (such as escorting and detours), thereby achieving continuous target locking and high signal-to-noise ratio signal acquisition. This distributed acquisition paradigm based on event triggering and dynamic resource scheduling enables continuous target locking and high signal-to-noise ratio signal acquisition.

[0032] 4. The distributed acquisition paradigm based on event triggering and dynamic resource scheduling disclosed in this invention upgrades the traditional "blind" full-domain patrol to "guided" precise perception. It effectively overcomes the bottlenecks in efficiency and real-time performance of the traditional "fixed patrol-centralized processing" mode, and significantly improves the ability to capture dynamic target signals and the utilization rate of system resources.

[0033] 5. The technical solution disclosed in this invention has a proven and effective collaborative architecture and response logic in the field of electromagnetic signal acquisition, providing a directly transferable solution for addressing similar problems faced in quantum magnetic field and quantum geomagnetic measurements (such as monitoring blind spots, loss of dynamic targets, and rigid resource allocation). For example, applying the above-mentioned distributed collaborative mechanism to a quantum magnetic field measurement network is expected to achieve wide-area early warning and precise tracking of quantum state signals, thereby promoting the application of quantum magnetic field measurement technology in real-world environments and improving data acquisition quality.

[0034] 6. This invention addresses technical challenges such as low spectrum reconstruction accuracy, weak generalization ability, and insufficient deployment efficiency caused by sparse marine electromagnetic spectrum and dynamic sea surface changes through standardized data acquisition and preprocessing at the simulation end (S1-S5), combined with knowledge distillation and multi-layer grid fusion at the algorithm end (S6-S8). Specifically, in the teacher-student distillation framework constructed in S6-S7, a complex teacher model is responsible for learning spatiotemporal dynamic patterns and frequency characteristics with high accuracy from limited effective data, and transferring its knowledge (teacher-completed spatiotemporal spectrum and spatiotemporal embedding features) to a lightweight student model, ensuring high-precision real-time completion capability even when deployed at the edge (e.g., UAV). Subsequently, the feature fusion mechanism with local attention introduced in S8 further intelligently integrates the fine-grained completion results (student-completed spatiotemporal spectrum) output by the student model into the reconstructed electromagnetic spectrum map, achieving accurate restoration from local details to the global scene. This achieves a closed-loop technology chain from raw signal acquisition to high-quality spectrum reconstruction. By deeply integrating physical mechanism environmental parameters, data-driven SNN model architecture, and intelligent fusion methods with local attention, it further provides reliable technical support for maritime electromagnetic situational awareness and spectrum resource management.

[0035] 7. S9 employs a multi-dimensional evaluation system to objectively measure the overall performance of the proposed method (STFD-SNN) across three core dimensions: reconstruction accuracy, physical consistency, and algorithm efficiency. This system not only addresses numerical errors but also ensures the reconstruction results conform to electromagnetic wave propagation laws through Physical Residual Metric (PRM), assesses the ability to capture key signals through Hotspot Coverage Ratio (HRC), and demonstrates the speedup effect of the lightweight student model compared to the complex teacher model through Reconstruction Efficiency Ratio (DRER). This comprehensive evaluation approach allows for a systematic and objective comparison with baseline methods, comprehensively validating the effectiveness and superiority of the proposed method in terms of accuracy, robustness, and practicality from multiple technical perspectives. Attached Figure Description

[0036] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0037] Figure 1 This is a schematic diagram of the existing distributed sampling spectrum reconstruction process for unmanned aerial vehicles.

[0038] Figure 2 This is the original radio map sampled by the UAV provided in this embodiment of the invention.

[0039] Figure 3 This is a radio map with spectral energy contour lines provided in an embodiment of the present invention.

[0040] Figure 4 This is a radio map processed by SNN spatiotemporal coding, provided in an embodiment of the present invention.

[0041] Figure 5 The radio map is generated by the marine electromagnetic spectrum reconstruction method based on SNN distillation spatiotemporal spectrum completion and multi-layer grid fusion provided in the embodiments of the present invention.

[0042] Figure 6 These are the spectrum distribution map, short-time Fourier transform map, and the result map after Grad-CAM visualization of the signal extraction capability of the UAV at a certain moment provided in the embodiments of the present invention.

[0043] Figure 7 This is a 2-D reconstruction comparison experiment diagram provided in the embodiment of the present invention.

[0044] Figure 8 This is a 3D rendering of the UAV adaptive sampling at different layers provided in an embodiment of the present invention.

[0045] Figure 9 This is a 3D rendering of the image after being restored from different height planes using only the teacher's end, as provided in this embodiment of the invention.

[0046] Figure 10 This is a 3D effect diagram restored by mapping different heights through the student terminal, provided by an embodiment of the present invention.

[0047] Figure 11 This is a 3D image of the STFD-SNN reconstruction sample provided in this embodiment of the invention.

[0048] Figure 12 This is a planar sampling effect diagram provided by an embodiment of the present invention without increasing height mapping.

[0049] Figure 13 This is a 2D rendering of the plane after restoration, provided by the embodiment of the present invention, which is only obtained through the teacher's end.

[0050] Figure 14 This is a 2D effect diagram of the enhanced plane restoration provided by the student end according to an embodiment of the present invention.

[0051] Figure 15This is a 2D image of the STFD-SNN reconstruction sample provided in this embodiment of the invention.

[0052] Figure 16 This is a comparison chart of the recovery accuracy of various models under different resource constraints when different sampling rates are set in the spectrum space, as provided in the embodiments of the present invention.

[0053] Figure 17 This is a comparison chart of the physical residual measure (PRM) results of different algorithms provided in this embodiment of the invention under different environmental conditions.

[0054] Figure 18 This is an example diagram of the maritime target area for UAV cruise sampling area and route planning provided in an embodiment of the present invention.

[0055] Figure 19 This is the optimal sampling route planning diagram for UAV provided in this embodiment of the invention.

[0056] Figure 20 This is an analytical diagram of the UAV optimal sampling 3-D path-following algorithm provided in the embodiment of the present invention.

[0057] Figure 21 This is a comparison chart of the optimal Uav sampling after 200 iterations from a 2-D perspective based on the spectral capability distribution, provided by an embodiment of the present invention.

[0058] Figure 22 This is a 3-D simple sampling path cruise diagram for the start of an iteration provided in an embodiment of the present invention.

[0059] Figure 23 This is an experimental result diagram of the optimal sampling route for 3-D UAV after 200 iterations, provided by an embodiment of the present invention. Detailed Implementation

[0060] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0061] This invention provides a method for marine electromagnetic spectrum reconstruction based on SNN distillation for spatiotemporal spectrum completion and multi-layer grid fusion, comprising:

[0062] S1, combining the flight parameters of UAVs with the number of missions, constructs a route index covering the target sea area, consisting of time, frequency, and space, and assigns the route index to the corresponding UAVs to complete the initialization of the grid system; where UAVs include flying UAVs and standby UAVs;

[0063] S2, the flying UAV cruises along a fixed route and obtains the statistical feature tensor of the broadband marine electromagnetic spectrum signal continuously monitored and collected by the receiver; the missing mask of the statistical feature tensor is determined by a threshold, and after being assigned a value, it is mapped to the grid system. The statistical feature tensor that passes the verification is retained as the basic cruise data, and the collected data is finally put into disk in the original coarse-grained grid.

[0064] S3 marks basic cruise data exceeding the threshold as suspected targets, generates and broadcasts beacons;

[0065] For example, if airborne energy, templates, or GLRT data collection and monitoring exceed the threshold, a UAV beacon (timestamp, UAV number, azimuth / estimated location, frequency band, center frequency, bandwidth, power / fingerprint digest, SNR, confidence level, TTL) is generated and broadcast through a self-organizing network.

[0066] S4. Based on the scoring function, select the optimal standby UAV to go to the target sea area corresponding to the beacon, perform the refined tracking required by the mission, and obtain the tracking data of the suspected target;

[0067] Refined tracking, such as escorting and circling, involves detailed sampling.

[0068] S5 integrates tracking data with basic cruise data and stores it in the grid system to obtain fused data;

[0069] S6: Input the fused data into the teacher model constructed from the spatiotemporal dynamics term, signal frequency shape term, propagation physics term, and noise term; use the spiking neural network (SNN) to extract spatiotemporal embedding features from the spatiotemporal dynamics term and perform pulse encoding; construct the teacher loss function to train the teacher model and output the teacher-completed spatiotemporal spectrum;

[0070] S7 uses knowledge distillation technology to input the teacher-completed spatiotemporal spectrum and spatiotemporal embedding features into the student model constructed from spatiotemporal dynamics, signal frequency shape, and propagation physics terms; it constructs a student loss function to train the student model and outputs the student-completed spatiotemporal spectrum.

[0071] S8 introduces a feature fusion mechanism based on local attention to perform multi-layer grid fusion on the electromagnetic activity of the student-completed spatiotemporal spectrum and output the reconstructed electromagnetic spectrum map.

[0072] In S1, the steps for initializing the grid system are as follows:

[0073] S11, Parameter Input: Set the simulation start time in the time axis parameters. Simulation end time Time sampling step Total number of time sampling points Set the minimum and maximum values ​​of the wideband signal processing frequency band in the frequency axis parameters. and the total number of sub-bands ; Set the latitude and longitude grid coordinates of the observation area in the spatial gridding parameters ( ), For the corresponding grid point number, latitude and longitude grid resolution ( Minimum and maximum altitude of UAV flight trajectory and the total number of planned heights ;

[0074] S12, Generate the time axis: The time axis is determined by the simulation start time, simulation end time, and time sampling step size; the generation method is as follows:

[0075] ;

[0076] ;

[0077] In the formula, For the nth time sampling point, Indicates the sequence number of the signal sampling point; where, The generated timeline ;

[0078] S13, Generate the frequency axis: The frequency axis is determined by the broadband signal processing frequency band, the sub-band center, and the total number of sub-bands; the generation method is as follows:

[0079] ;

[0080] In the formula, Let be the center frequency of the k-th sub-band, and k be the sub-band number. This is the minimum value for the wideband signal processing frequency band. For wideband signal processing, the maximum value of the frequency band is calculated; the frequency band is divided into equal intervals, generating a frequency axis. ;

[0081] S14, Spatial gridding: Based on the latitude and longitude grid coordinates of the observation area ( ) and latitude and longitude grid resolution ( Construct longitude axes separately and latitude axis ; The minimum and maximum altitudes of the UAV's flight path and the total number of planned heights Constructing the height axis via the longitude axis latitude axis and height axis The space is divided into grids; the spatial axes after gridding are mapped and the boundaries of the observation area are saved using GeoSOT-3D regular grid mapping.

[0082] S15, Mesh system initialization, based on the timeline. Frequency axis Longitude axis latitude axis and height axis A route index covering the target sea area: ;

[0083] In the formula, To unify grid metadata.

[0084] In S2, the statistical feature tensor is:

[0085] ;

[0086] In the formula, The statistical feature tensor represents a gridded raw signal dataset covering the target sea area and observation period; count represents the total number of valid signal sampling points within the current statistical window; Average power; Peak power; For minimum power, The standard deviation of power;

[0087] The method for determining the missing mask of a statistical feature tensor using a threshold is as follows:

[0088] ;

[0089] In the formula, For a binary observation mask, if ,but , is an effective mask for the grid system, indicating the presence of actual signal data; otherwise, , which is the missing mask of the grid system, representing the region in the grid system that needs to be assigned a value;

[0090] The method for verifying the statistical feature tensors mapped to the grid system is as follows:

[0091] ;

[0092] ;

[0093] ;

[0094] In the formula, For time variables, The time for the statistical feature tensor; For frequency variables, The frequency of the statistical feature tensor; For the corresponding position coordinates in the grid system, These are index pairs of all possible spatial locations in the grid system. Let these be the spatial coordinates of the statistical feature tensor. Latitude and longitude coordinates For the data at the j-th index position on the dimension axis, This represents the data at the i-th index position on the longitude axis.

[0095] In S3, the method for marking basic cruise data exceeding a threshold as suspected targets and generating and broadcasting beacons is as follows:

[0096] S31 performs energy detection on basic cruise data. The energy detection method is as follows:

[0097] ;

[0098] S32, for For basic cruise data exceeding a threshold, template matching detection is performed. The template matching detection method is as follows:

[0099] ;Will Basic cruise data exceeding the threshold is marked as a suspected target;

[0100] S33, Generate a beacon for a suspected target. :

[0101] ;

[0102] S34, the method for broadcasting a beacon of a suspected target is as follows:

[0103] ;

[0104] In the formula, This represents the average power of the basic cruise data within the detection window, where N represents the total number of signal sampling points contained in a single detection window. This represents the amplitude of the complex signal at the nth discrete-time sampling point, where n represents the sampling point number. This is a statistical measure for template matching, representing the compatibility and matching between the basic cruise data and the preset template; Where T represents the signal sample size, T denotes transpose, and R is the noise covariance matrix.

[0105] ID is a unique identifier for the beacon. The precise timestamp at which the signal was detected. The unique identifier for the UAV that detected the signal. This refers to the real-time location of the UAV corresponding to the beacon. This represents the estimated location of the signal source obtained by the direction-of-arrival (DOA) estimation algorithm. for The mapped signal frequency domain feature vector, Here, SNR is the signal detection power, and SNR is the signal-to-noise ratio. The detection confidence level is the normalized interval (0,1). Forwarding decisions for beacons suspected of being targets; This represents the ui-th beacon, where ui is the beacon's index, used to distinguish different beacon instances. In a grid system, if multiple UAVs detect multiple suspected targets simultaneously, each detection will generate an independent beacon, and ui is the index used to identify which beacon it is. For flying UAV nodes The set of IDs of processed beacons.

[0106] In S4, the scoring function is:

[0107] ;

[0108] ;

[0109] In the formula, For comprehensive scoring; For normalized distance, To accept the power term, For the geometric cost of observation, To normalize the arrival delay term, This is a penalty term for multi-objective load balancing. ,in, These are the normalized distance weighting coefficients. To accept the weighting coefficients for the power term, The geometric cost weighting coefficient is used to observe the data. The normalized arrival delay term weighting coefficients are used. The weighting coefficients for the multi-objective load balancing penalty term; The best standby UAV indicates the overall score. The lowest scorer was selected as the best standby UAV. This indicates that the UAV is on standby. This refers to a flying UAV.

[0110] In S5, the tracking data and basic cruise data are fused and stored in the grid system. The method for obtaining the fused data is as follows:

[0111] S51, Time Index Mapping of Merged Data:

[0112] ;

[0113] S52, Frequency index mapping of fused data:

[0114] ;

[0115] S53, Spatial index mapping of fused data:

[0116] ;

[0117] ;

[0118] ;

[0119] ;

[0120] S54, Data Fusion and Grid Unification:

[0121] ;

[0122] S55, Mask Generation:

[0123] ;

[0124] S56, Generation of environmental parameter tensors:

[0125] ;

[0126] In the formula, For time index elements of merged data, For each raw data point waiting to be mapped into the grid, For time index mapping functions, To merge the original timestamps of individual data points, This is the nth vector on the time axis; For frequency index elements of fused data, For frequency index mapping functions, The center frequency of the data points for fused data. The first frequency axis One vector; For spatial index elements of merged data, ( () represents the x and y coordinates of the spatial index element. This refers to the height coordinate information within the spatial index element. It is a space mapping function; The first on the longitude axis Data at each index position This is the data at the m-th index position on the height axis. The first in the height axis A specific height layer index marker; To fuse the statistical feature tensors of the data, For a binary observation mask of fused data, if , is the effective mask of the grid system, indicating that there is actual signal data in the grid area, and the judgment condition is . SNR represents the signal-to-noise ratio of the smallest grid cell. The signal-to-noise threshold. The minimum number of valid sampling points is the threshold; otherwise, , which is the missing mask of the grid system, representing the region in the grid system that needs to be assigned a value; The environmental parameter tensor for fused data is used to record environmental feature vectors in the grid cells. For free space path loss, For the two-path reflection gain of the sea surface, For tropospheric refraction and absorption losses, This represents the antenna gain.

[0127] In S6, the teacher model constructed is as follows:

[0128] ;in,

[0129] ;

[0130] The method for extracting spatiotemporal embedding features from spatiotemporal dynamic terms using spiking neural networks (SNNs) is as follows:

[0131] ;

[0132] The method for pulse coding using LIF neurons in SNN is as follows:

[0133] ;

[0134] ;

[0135] The constructed teacher loss function is as follows:

[0136] ;

[0137] The teacher completes the spatiotemporal spectrum as follows:

[0138] ;

[0139] In the formula, The input fused data in time power coordinates ( ) and spatial location ( The observed power value on ) For the spatiotemporal dynamics of the fused data, For the signal frequency shape term of the fused data, For the propagation physical terms of the fused data, For environmental parameter tensors, Noise term; To reduce the free space path loss of the fused data, Spatial location Distance from the radiation source; To fuse the sea surface two-path reflection gain and loss data, The complex reflection coefficient of the sea surface. To integrate tropospheric refraction and absorption losses in the data, For antenna gain of fused data, This represents the azimuth angle information received by the antenna and the transmitting source, indicating the effective gain of the antenna; For the spatiotemporal embedding features of the teacher model, For the teacher model, a spiking neural network, For the spatiotemporal dynamics of the teacher model, For the SNN parameters of the teacher model; For the first The membrane potential of the layer neurons at time t+1. For the first The membrane potential of a layer neuron at time t. For the first The output pulse of the layer neuron at time t. For the first The output pulse of the layer neuron at time t; The membrane potential leakage constant is... For the first The weight matrix of the layer, Set the firing threshold for neurons. Represents the step function; For the teacher's side loss value, For the feature constraints of SNN, For the signal frequency shape term of the teacher model L1 regularization is used to promote frequency domain sparsity and data fidelity terms. , The symbol for calculating L1 regularization terms, For effective masking, For residual calculation, To observe the power value, For the propagation physics term of the teacher model, for nuclear norm number, for Total variation regularity, For the loss weights constrained by the SNN features on the teacher side, For loss weights constrained by frequency sparsity on the teacher side, To improve frequency domain sparsity and reduce the loss weights of data fidelity terms for teachers, The loss weights for the nuclear norm constraint of the teacher model. The weighting coefficients for the loss of the total variational regularization term on the teacher's side; To supplement the spatiotemporal spectrum for teachers.

[0140] In S7, the student model constructed is as follows:

[0141] ;

[0142] The constructed student loss function is as follows:

[0143] ;

[0144] The student completes the spatiotemporal spectrum as follows:

[0145] ;

[0146] In the formula, To complete the spatiotemporal spectrum for students, For the spatiotemporal dynamics of the student model, For the signal frequency shape term of the student model, For the propagation physics term in the student model; For student-side loss values, For mission losses, For output distillation, For characteristic distillation, For physical consistency, For smoothing regularization, The weighting coefficients for the task loss items. To output the weighting coefficients of the distillation term, These are the weighting coefficients for the characteristic distillation term. The weighting coefficient for the physical consistency term. These are the weighting coefficients for the smoothing regularization term; For the spatiotemporal embedding features of the student model, It is a student-model spiking neural network. These are the SNN parameters for the student model.

[0147] In S8, the feature fusion mechanism of local attention is as follows:

[0148] ;

[0149] The method for calculating the electromagnetic activity degree of students completing the spatiotemporal spectrum is as follows:

[0150] ;

[0151] In the formula, For the reconstructed electromagnetic spectrum, To complete the fine-grained spatial units of the spatiotemporal spectrum for students, Let p be the set of all p. For local attention weights, To complete the spatiotemporal spectrum for students after multi-layer grid fusion; To complete the spatiotemporal spectrum for students, the spatial coordinates are provided at a fine-grained grid level; where fine-grained refers to the grid size set according to the dynamic task scenario.

[0152] Electromagnetic activity, To help students complete the longitudinal and transverse coordinates of fine-grained spatial units in the spatiotemporal spectrum, This represents the total number of time sampling points. The total number of sub-bands. To complete the spatiotemporal spectrum for students in the corresponding Power value at location, This represents the average background power level.

[0153] It also includes S9, a step that verifies the spectral completion accuracy, physical consistency, algorithm efficiency, and robustness of the reconstructed electromagnetic spectrum, and outputs the verification results; among which,

[0154] The method for verifying the spectral completion accuracy of the reconstructed electromagnetic spectrum is as follows:

[0155] ;

[0156] ;

[0157] ;

[0158] The method for verifying the physical consistency of the reconstructed electromagnetic spectrum is as follows:

[0159] ;

[0160] ;

[0161] The method for verifying the algorithm efficiency and robustness of the reconstructed electromagnetic spectrum is as follows:

[0162] ;

[0163] In the formula, The root mean square error, For effective masking, For a single data point in a grid system, for The flattened representation, i.e., the reconstructed electromagnetic spectrum, is the first... The value at each position, for In the grid space, the first A flattened representation of the point location, i.e., the actual observed data at the [number]th [location]. The value of each position; The actual observed spectral power; For normalized mean square error, Peak signal-to-noise ratio, For signal sampling quantity, Mean square error; For physical residual measurement, For coverage of hotspot areas, To count the number of grid cells that satisfy the conditions within parentheses, For the fused tensor The channel output student completes the spatiotemporal spectrum. For hotspot power threshold, The sampling rate error curve SREC is the ratio of model reconstruction efficiency to the model reconstruction error. This represents the total number of data tests. For the number of data tests, Let m_time be the time of the teacher model's inference. Let m_time be the time taken for the student model to make its inference.

[0164] Verification experiment:

[0165] To verify the effectiveness of the technical solution of the present invention, a verification experiment is provided for illustration:

[0166] Comparative experimental results of reconstruction based on two-dimensional spectrograms Figure 2 The original radio map, constructed from samples collected by a drone swarm, visually presents the distribution and spatial coverage of the raw data collected from the maritime electromagnetic signals. These data points form the basis for all subsequent processing and analysis. Figure 3 Based on this, equivalent spectral energy lines were added, which clearly revealed the spatial gradient of signal energy and the location of the core radiation source, verifying that the basic cruise data can effectively support the preliminary identification of signal hotspot areas. Figure 4 The spectrum after SNN spatiotemporal coding is shown. Its significant feature is that the background noise is effectively suppressed, while the spatiotemporal continuity structure of the signal is enhanced, which proves the effectiveness of the SNN model in extracting and purifying spatiotemporal features. Figure 5 This visually demonstrates the complete two-dimensional radio map generated by the STFD-SNN reconstruction method, and... Figure 2 Compared with the original data, the reconstructed map retains key signal features while filling in data from sparsely sampled areas, achieving high-completeness spectrum reconstruction and verifying the core capability of this method to recover the global spectrum situation from sparse sampling. Figure 6 It provides multi-dimensional analysis of instantaneous signals. The spectral energy distribution plot shows the frequency domain characteristics of the signal, the short-time Fourier transform plot reveals the dynamic characteristics of the signal in the time-frequency joint domain, and the Grad-CAM visualization clearly marks the key time-frequency regions that the SNN model focuses on when making decisions, enhancing the interpretability of the model and proving that the features it focuses on are consistent with the physical signal characteristics. Figure 7 The first column shows the true distribution of the desired spectrum; the second column presents the distribution at the sampling rate. The data presented are observational data under the given conditions, with white pixels representing missing data areas. The remaining columns show the spectrum reconstruction results of the baseline comparison algorithm and the proposed technical solution (hereinafter referred to as STFD-SNN). Based on the marine spectrum reconstruction scenario, three types of regions need to be focused on: (i) strong reflection / echo areas near the shore and at the intersection of shipping channels; (ii) the influence zone of moving radiation sources around the ship's trajectory; and (iii) the significant multipath region near the sea-air interface (lower altitude). In these regions, STFD-SNN exhibits fewer holes and signal scattering diffraction waves, with coherent edge structures, and maintains topological consistency in hotspot shape when crossing missing measurement zones. In contrast, baseline methods (such as LRTC-TNN and RTCD) are more prone to over-smoothing or texture drift under sparse conditions. Comparing the reconstructed images, it can be seen that BRTF, due to its global low-rank factor and the contraction bias generated by Bayesian modeling of sparse noise under MNAR conditions, is more prone to amplitude underestimation and boundary blunting around hotspots. The reconstructed REM image shows flattening rather than true completion around the signal hotspot region, resulting in a ring-shaped enhancement at the hotspot edge, further leading to insufficient local completion. HaL-RTC relies on the quasi-periodicity and stationary trend after Hankelization, resulting in a non-periodic discontinuity in the 2D reconstruction result, disrupting boundaries and length truncation. RTCD's robust loss depends on correct rank and rank allocation, causing local rank to change with location in maritime scenes, and local overfitting with fixed rank. Its reconstructed 2D REM image shows edge breaks, re-adhesion, and a patchwork feel, altering the hotspot topology. LRTC-TNN, due to its global tubal low-rank preference for isotropic smoothing, causes hotspot haloing in the 2D spectrogram to appear as continuous bands at the edges, resulting in blurred boundary regions. The STFD-SNN proposed in this invention combines the SNN temporal features of the teacher with the frequency spatiotemporal attention of FSTA to extract saliency and suppress it in the two-dimensional plane, and learns the two-dimensional spectral features better through student distillation alignment. This makes the two-dimensional REM graph superior to the baseline experiment in terms of hot spot amplitude preservation, boundary geometry and connectivity.

[0167] Figures 8 to 11 This presentation showcases a visualization of the stepwise completion process of STFD-SNN at a sampling rate of r=20% from a 3D perspective: 2D completion from the teacher's end, completion from the student's end, and completion results after adding physics guidance. Finally, the results are fused to generate a complete 3D REM map reconstruction of STFD-SNN. The addition of... and Subsequently, the near-sea surface two-path enhancement was correctly preserved, the rapid attenuation in the mid-to-high atmosphere was physically consistent with the vertical continuity, and the layer mean and layer variance of the completion result decreased synchronously.

[0168] Figures 12 to 15 The diagram shows the reconstruction results of different parts of the model's performance. The reconstruction results were verified mainly through a systematic ablation experiment using controlled variables. Figure 12 The demonstration shows sampling solely through drones and Figure 13 The reconstruction results (for the teacher model only) serve as a baseline, demonstrating the limitations of planar reconstruction or the computational redundancy of complex models. Figure 14 The results of the (student-side enhancement) show that the lightweight student model, guided by the teacher model, achieves significantly higher reconstruction accuracy and efficiency than the baseline. Ultimately, Figure 15 The provided complete STFD-SNN model reconstruction results achieve the best balance between detail recovery and global consistency, fully demonstrating the comprehensive advantages of the knowledge distillation and multi-layer grid fusion mechanism in this method.

[0169] To comprehensively evaluate the impact of sampling rate on spectrum recovery accuracy, we simulated different resource constraints by adjusting the sampling rate settings for the spectrum space. Figure 16 The image shows a comparison of the recovery accuracy evaluation results between STFD-SNN and four benchmark algorithms. For example... Figure 16 The comparison results in (a) show that all algorithms exhibit the expected trend of monotonically decreasing reconstruction error (RMSE) as the sampling rate increases. However, the STFD-SNN algorithm demonstrates a significant and consistent advantage across the entire test range. Under the most challenging low sampling rate condition (r=0.1), it reduces the RMSE by approximately 2.3% compared to the best benchmark algorithm LRTC-TNN. At r=0.2, the RMSE of STFD-SNN (0.529) is reduced by 31.4% and 24.8% compared to BRTF (0.771) and HaL-RTC (0.703), respectively. This superior performance under sparse data conditions can be attributed to the spatiotemporal dynamic modeling capability of the SNN encoder in STFD-SNN, combined with the focusing mechanism of the FSTA module on key spectral features, enabling the algorithm to effectively extract and utilize the spatiotemporal correlations in the environment with limited observation data.

[0170] A thorough analysis of the algorithm's convergence characteristics reveals another significant advantage of STFD-SNN. When the sampling rate r ≥ 0.5, STFD-SNN exhibits fast and stable convergence performance, with the RMSE dropping below 0.297. In contrast, traditional algorithms such as BRTF still have an RMSE higher than 0.422 under the same conditions. This phenomenon verifies the crucial role of the physical guidance mechanism in improving the algorithm's convergence speed and final accuracy. Particularly noteworthy is that, near full sampling (r = 1.0), the RMSE of STFD-SNN further decreases to 0.103, indicating that the proposed algorithm is not only suitable for sparse data conditions but also fully realizes its performance potential when data is relatively abundant.

[0171] Figure 16 (b) is an experimental result figure that uses hotspot area coverage as the core indicator for evaluating the practicality of spectrum reconstruction. It directly reflects the algorithm's ability to reconstruct areas with significant signal strength. Figure 16 (b) shows the coverage performance of each algorithm at different sampling rates. Although all algorithms show a similar improvement trend to RMSE, STFD-SNN has a more significant advantage in coverage. At the lowest sampling rate (r=0.1), STFD-SNN achieves a coverage of 49.15%, a 3.5% improvement over the best benchmark algorithm LRTC-TNN (47.47%). This further demonstrates that STFD-SNN has a more comprehensive spectral environment awareness capability with limited sampling resources. At near-full sampling (r=1.0), STFD-SNN achieves the best performance of 94.72%, mainly benefiting from the SNN spatiotemporal feature encoding of the STFD-SNN model and the FSTA module's ability to focus on key spectral features. The coverage curve of STFD-SNN exhibits a smooth and monotonically increasing characteristic. This difference in stability reflects the robustness of the hierarchical architecture of STFD-SNN, where the teacher model is responsible for capturing basic spatiotemporal features, while the student model ensures the rationality of the reconstruction results through physical constraints.

[0172] Figure 17This paper presents the Physical Residual Metrics (PRM) of different algorithms under various electromagnetic propagation conditions. The analysis shows that different algorithms exhibit significant differences in performance when dealing with variations in electromagnetic propagation parameters. The STFD-SNN proposed in this invention, by explicitly embedding the physical propagation model into the completion network, accurately characterizes multipath propagation and signal attenuation in the marine environment, thus exhibiting the lowest PRM values ​​under all three electromagnetic propagation conditions. In contrast, traditional tensor completion methods, such as BRTF and LRTC-TNN, primarily rely on data-driven low-rank assumptions, making it difficult to accurately model electromagnetic propagation processes in complex marine channels, resulting in higher PRM values. Although RTCD and HaL-RTC methods introduce certain regularization constraints, they still lack explicit modeling of the physical propagation mechanism, thus limiting their performance in handling complex phenomena such as sea surface reflection.

[0173] The experimental results and comparisons above verify the advantages of the physical consistency framework in the STFD-SNN reconstruction network and its key role in improving reconstruction performance. Experimental results show that physical constraints help recover realistic electromagnetic propagation characteristics in sparse data scenarios, and parameters such as the path loss exponent, sea surface reflection coefficient, and atmospheric attenuation coefficient have a significant impact on reconstruction performance. Physical consistency not only improves the vertical continuity of 3D reconstruction but also effectively captures complex electromagnetic propagation phenomena in the marine environment.

[0174] Figure 18 It shows the maritime target monitoring area and initial route, providing a realistic scenario context for the simulation experiment. Figure 19 The demonstration shows the optimal sampling route for UAVs, reflecting the comprehensive decision-making result of the scoring function after balancing distance, signal reception power, and geometric relationships. The route can effectively cover signal hotspots. Figure 20 The algorithm provides an analytical diagram of the optimal path from a 3D perspective, clearly showing the three-dimensional sampling trajectory of the UAV in space, and verifying the algorithm's ability to perform efficient and intelligent path planning in complex three-dimensional space.

[0175] Figure 21 The heatmap comparison clearly demonstrates the performance differences of different algorithms in path optimization, verifying the superiority and practicality of the adaptive ACO algorithm in complex spectrum environments. Figure 21 (a) shows a heatmap of the spectrum signal intensity distribution in the simulated sea area under dynamic changes, with the signal intensity ranging from -110dB to -50dB. Figure 21 (b) shows the sampling path planning results of uniformly distributed sampling paths with regular gridded cruise paths. Although this strategy can guarantee basic spatial coverage, it is obviously lacking in the ability to respond to spatial variability of signal strength. Figure 21(c) The path optimization effect of the traditional ACO algorithm is presented. Compared with traditional grid sampling, traditional ACO shows certain adaptive characteristics. The path trajectory begins to avoid some weak signal areas and moves closer to medium-intensity signal areas. Although the path still fails to fully focus on the strongest hotspot areas, it exhibits certain randomness and instability in the path planning process. Figure 21 (d) shows the path optimization results of the proposed adaptive ACO algorithm. The algorithm's path is significantly more concentrated in the hotspot areas with the highest signal strength, and has advantages such as avoiding ineffective wandering in low-value areas, a more compact and efficient path structure, and more comprehensive coverage of hotspot areas.

[0176] Figure 22 The evolution of the UAV flight trajectory is shown during the initial measurement phase and optimization iteration. Figure 23 By setting the ant colony size to 24 ants and the pheromone evaporation coefficient ρ=0.46, adaptive sampling of the spectral characteristics of the three-dimensional sea area was completed after 200 iterations. The visualization results clearly show that the UAV initially adopts a regular gridded cruising path, but as the ACO iterations deepen, the flight trajectory gradually converges to the spectral hotspot region, exhibiting significant path adaptation characteristics, especially in sea areas with strong signal intensity.

[0177] Through the above sampling-based experiments and comparative analyses, the technical solution proposed in this invention demonstrates significant technical advantages in adaptive sampling and automatic pathfinding algorithms for maritime spectrum monitoring tasks. At the path planning level, the traditional two-dimensional path planning problem is extended to a three-dimensional spectrum space, and the curse of dimensionality is further addressed through a hierarchical optimization network. This invention not only considers traditional factors such as path length but also incorporates real-time spectrum characteristics as heuristic information, enabling UAVs to intelligently respond to environmental changes and providing an efficient and reliable solution for maritime spectrum monitoring and sampling tasks.

[0178] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for spatiotemporal spectrum completion and multi-layer grid fusion for marine electromagnetic spectrum reconstruction based on SNN distillation, characterized in that, include: S1, combining the flight parameters of UAVs with the number of missions, constructs a route index covering the target sea area, consisting of time, frequency, and space, and assigns the route index to the corresponding UAVs to complete the initialization of the grid system; where UAVs include flying UAVs and standby UAVs; S2, the flying UAV cruises along a fixed route and obtains the statistical feature tensor of the broadband marine electromagnetic spectrum signal continuously monitored and collected by the receiver; the missing mask of the statistical feature tensor is determined by a threshold, and after being assigned a value, it is mapped to the grid system, and the statistical feature tensor that passes the verification is retained as the basic cruise data. S3 marks basic cruise data exceeding the threshold as suspected targets, generates and broadcasts beacons; S4. Based on the scoring function, select the optimal standby UAV to go to the target sea area corresponding to the beacon, perform the refined tracking required by the mission, and obtain the tracking data of the suspected target; S5 integrates tracking data with basic cruise data and stores it in the grid system to obtain fused data; S6: Input the fused data into the teacher model constructed from the spatiotemporal dynamics term, signal frequency shape term, propagation physics term, and noise term; use the spiking neural network (SNN) to extract spatiotemporal embedding features from the spatiotemporal dynamics term and perform pulse encoding; construct the teacher loss function to train the teacher model and output the teacher-completed spatiotemporal spectrum; S7 uses knowledge distillation technology to input the teacher-completed spatiotemporal spectrum and spatiotemporal embedding features into the student model constructed from spatiotemporal dynamics, signal frequency shape, and propagation physics terms; it constructs a student loss function to train the student model and outputs the student-completed spatiotemporal spectrum. S8 introduces a feature fusion mechanism based on local attention to perform multi-layer grid fusion on the electromagnetic activity of the student-completed spatiotemporal spectrum and output the reconstructed electromagnetic spectrum map.

2. The marine electromagnetic spectrum reconstruction method based on SNN distillation for spatiotemporal spectrum completion and multi-layer grid fusion as described in claim 1, characterized in that, In S1, the steps for initializing the grid system are as follows: S11, Parameter Input: Set the simulation start time in the time axis parameters. Simulation end time Time sampling step Total number of time sampling points Set the minimum and maximum values ​​of the wideband signal processing frequency band in the frequency axis parameters. and the total number of sub-bands ; Set the latitude and longitude grid coordinates of the observation area in the spatial gridding parameters ( ), For the corresponding grid point number, latitude and longitude grid resolution ( Minimum and maximum altitude of UAV flight trajectory and the total number of planned heights ; S12, Generate the time axis: The time axis is determined by the simulation start time, simulation end time, and time sampling step size; the generation method is as follows: ； ； In the formula, For the nth time sampling point, Indicates the sequence number of the signal sampling point; where, The generated timeline ; S13, Generate the frequency axis: The frequency axis is determined by the broadband signal processing frequency band, the sub-band center, and the total number of sub-bands; the generation method is as follows: ； In the formula, Let k be the center frequency of the k-th sub-band, and k be the sub-band number. This is the minimum value for the wideband signal processing frequency band. For wideband signal processing, the maximum value of the frequency band is calculated; the frequency band is divided into equal intervals, generating a frequency axis. ; S14, Spatial gridding: Based on the latitude and longitude grid coordinates of the observation area ( ) and latitude and longitude grid resolution ( Construct longitude axes separately and latitude axis ; The minimum and maximum altitudes of the UAV's flight path and the total number of planned heights Construct height axis via the longitude axis latitude axis and height axis Divide the space into grids; S15, Mesh system initialization, based on the timeline. Frequency axis Longitude axis latitude axis and height axis A route index covering the target sea area: ; In the formula, To unify grid metadata.

3. The marine electromagnetic spectrum reconstruction method based on SNN distillation for spatiotemporal spectrum completion and multi-layer grid fusion as described in claim 2, characterized in that, In S2, the statistical feature tensor is: ； In the formula, The statistical feature tensor represents a gridded raw signal dataset covering the target sea area and observation period; count represents the total number of valid signal sampling points within the current statistical window. Average power; Peak power; For minimum power, The standard deviation of power; The method for determining the missing mask of a statistical feature tensor using a threshold is as follows: ； In the formula, For a binary observation mask, if ,but , is an effective mask for the grid system, indicating the presence of actual signal data; otherwise, , which is the missing mask of the grid system, representing the region in the grid system that needs to be assigned a value; The method for verifying the statistical feature tensors mapped to the grid system is as follows: ； ； ； In the formula, For time variables, The time for the statistical feature tensor; For frequency variables, The frequency of the statistical feature tensor; For the corresponding position coordinates in the grid system, These are index pairs of all possible spatial locations in the grid system. Let these be the spatial coordinates of the statistical feature tensor. Latitude and longitude coordinates For the data at the j-th index position on the dimension axis, This represents the data at the i-th index position on the longitude axis.

4. The marine electromagnetic spectrum reconstruction method based on SNN distillation for spatiotemporal spectrum completion and multi-layer grid fusion as described in claim 3, characterized in that, In S3, the method for marking basic cruise data exceeding a threshold as suspected targets and generating and broadcasting beacons is as follows: S31 performs energy detection on basic cruise data. The energy detection method is as follows: ； S32, for For basic cruise data exceeding a threshold, template matching detection is performed. The template matching detection method is as follows: ;Will Basic cruise data exceeding the threshold is marked as a suspected target; S33, Generate a beacon for a suspected target. : ； S34, the method for broadcasting a beacon of a suspected target is as follows: ； In the formula, This represents the average power of the basic cruise data within the detection window, where N represents the total number of signal sampling points contained in a single detection window. This represents the amplitude of the complex signal at the nth discrete-time sampling point, where n represents the sampling point number. This is a statistical measure for template matching, representing the compatibility and matching between the basic cruise data and the preset template; The value is the signal sample size, and T represents the transpose. R is the noise covariance matrix; ID is a unique identifier for the beacon. The precise timestamp at which the signal was detected. The unique identifier for the UAV that detected the signal. This refers to the real-time location of the UAV corresponding to the beacon. This represents the estimated location of the signal source obtained by the direction-of-arrival (DOA) estimation algorithm. for The mapped signal frequency domain feature vector, Here, SNR is the signal detection power, and SNR is the signal-to-noise ratio. The detection confidence level is the normalized interval (0,1). Forwarding decisions for beacons suspected of being targets; This represents the ui-th beacon, where ui is the beacon's index, used to distinguish different beacon instances. In a grid system, if multiple UAVs detect multiple suspected targets simultaneously, each detection will generate an independent beacon, and ui is the index used to identify which beacon it is. For flying UAV nodes The set of IDs of processed beacons.

5. The marine electromagnetic spectrum reconstruction method based on SNN distillation for spatiotemporal spectrum completion and multi-layer grid fusion as described in claim 4, characterized in that, In S4, the scoring function is: ； ； In the formula, For comprehensive scoring; For normalized distance, To accept the power term, For the geometric cost of observation, To normalize the arrival delay term, This is a penalty term for multi-objective load balancing. ,in, These are the normalized distance weighting coefficients. To accept the weighting coefficients for the power term, The geometric cost weighting coefficient is used to observe the data. The normalized arrival delay term weighting coefficients are used. The weighting coefficients for the multi-objective load balancing penalty term; The best standby UAV indicates the overall score. The lowest scorer was selected as the best standby UAV. This indicates that the UAV is on standby. Indicates a flying UAV.

6. The marine electromagnetic spectrum reconstruction method based on SNN distillation for spatiotemporal spectrum completion and multi-layer grid fusion as described in claim 5, characterized in that, In S5, the tracking data and basic cruise data are fused and stored in the grid system. The method for obtaining the fused data is as follows: S51, Time Index Mapping of Merged Data: ； S52, Frequency index mapping of fused data: ； S53, Spatial index mapping of fused data: ； ； ； ； S54, Data Fusion and Grid Unification: ； S55, Mask Generation: ； S56, Generation of environmental parameter tensors: ； In the formula, For time index elements of merged data, For each raw data point waiting to be mapped into the grid, For time index mapping functions, To merge the original timestamps of individual data points, This is the nth vector on the time axis; For frequency index elements of fused data, For frequency index mapping functions, The center frequency of the data points for fused data. The first frequency axis One vector; For spatial index elements of merged data, ( () represents the x and y coordinates of the spatial index element. This refers to the height coordinate information within the spatial index element. It is a space mapping function; The first on the longitude axis Data at each index position This is the data at the m-th index position on the height axis. The first in the height axis A specific height layer index marker; To fuse the statistical feature tensors of the data, For a binary observation mask of fused data, if , is the effective mask of the grid system, indicating that there is actual signal data in the grid area, and the judgment condition is . SNR represents the signal-to-noise ratio of the smallest grid cell. The signal-to-noise threshold. The minimum number of valid sampling points is the threshold; otherwise, , which is the missing mask of the grid system, representing the region in the grid system that needs to be assigned a value; The environmental parameter tensor for fused data is used to record environmental feature vectors in the grid cells. For free space path loss, For the two-path reflection gain of the sea surface, For tropospheric refraction and absorption losses, This represents the antenna gain.

7. The marine electromagnetic spectrum reconstruction method based on SNN distillation for spatiotemporal spectrum completion and multi-layer grid fusion as described in claim 6, characterized in that, In S6, the teacher model constructed is as follows: ;in, ； The method for extracting spatiotemporal embedding features from spatiotemporal dynamic terms using spiking neural networks (SNNs) is as follows: ； The method for pulse coding using LIF neurons in SNN is as follows: ； ； The constructed teacher loss function is as follows: ； The teacher completes the spatiotemporal spectrum as follows: ； In the formula, The input fused data in time power coordinates ( ) and spatial location ( The observed power value on ) For spatiotemporal dynamic items of fused data, For the signal frequency shape term of the fused data, For the propagation physical terms of the fused data, For environmental parameter tensors, Noise term; To reduce the free space path loss of the fused data, Spatial location Distance from the radiation source; To fuse the sea surface two-path reflection gain and loss data, The complex reflection coefficient of the sea surface. To integrate tropospheric refraction and absorption losses in the data, For antenna gain of fused data, This represents the azimuth angle information received by the antenna and the transmitting source, indicating the effective gain of the antenna; For the spatiotemporal embedding features of the teacher model, For the teacher model, a spiking neural network, For the spatiotemporal dynamics of the teacher model, For the SNN parameters of the teacher model; For the first The membrane potential of the layer neurons at time t+1. For the first The membrane potential of a layer neuron at time t. For the first The output pulse of the layer neuron at time t. For the first The output pulse of the layer neuron at time t; The membrane potential leakage constant is... For the first The weight matrix of the layer, Set the firing threshold for neurons. Represents the step function; For the teacher's side loss value, For the feature constraints of SNN, For the signal frequency shape term of the teacher model L1 regularization is used to promote frequency domain sparsity and data fidelity terms. , The symbol for calculating L1 regularization terms, For effective masking, For residual calculation, To observe the power value, For the propagation physics term of the teacher model, for nuclear norm number, for Total variation regularity, For the loss weights constrained by the SNN features on the teacher side, For loss weights constrained by frequency sparsity on the teacher side, To improve frequency domain sparsity and reduce the loss weights of data fidelity terms for teachers, The loss weights for the nuclear norm constraint of the teacher model. The weighting coefficients for the loss of the total variational regularization term on the teacher's side; To supplement the spatiotemporal spectrum for teachers.

8. The marine electromagnetic spectrum reconstruction method based on SNN distillation for spatiotemporal spectrum completion and multi-layer grid fusion as described in claim 7, characterized in that, In S7, the student model constructed is as follows: ； The constructed student loss function is as follows: ； The student completes the spatiotemporal spectrum as follows: ； In the formula, To complete the spatiotemporal spectrum for students, For the spatiotemporal dynamics of the student model, For the signal frequency shape term of the student model, For the propagation physics term in the student model; For student-side loss values, For mission losses, For output distillation, For characteristic distillation, For physical consistency, For smoothing regularization, The weighting coefficients for the task loss items. To output the weighting coefficients of the distillation term, These are the weighting coefficients for the characteristic distillation term. The weighting coefficient for the physical consistency term. These are the weighting coefficients for the smoothing regularization term; For the spatiotemporal embedding features of the student model, It is a student-model spiking neural network. These are the SNN parameters for the student model.

9. The marine electromagnetic spectrum reconstruction method based on SNN distillation for spatiotemporal spectrum completion and multi-layer grid fusion as described in claim 8, characterized in that, In S8, the feature fusion mechanism of local attention is as follows: ； The method for calculating the electromagnetic activity degree of students completing the spatiotemporal spectrum is as follows: ； In the formula, For the reconstructed electromagnetic spectrum, To complete the fine-grained spatial units of the spatiotemporal spectrum for students, Let p be the set of all p. For local attention weights, To complete the spatiotemporal spectrum for students after multi-layer grid fusion; To complete the spatiotemporal spectrum for students, the spatial coordinates are provided at a fine-grained grid level; where fine-grained refers to the grid size set according to the dynamic task scenario. Electromagnetic activity, To help students complete the longitudinal and transverse coordinates of fine-grained spatial units in the spatiotemporal spectrum, This represents the total number of time sampling points. The total number of sub-bands. To complete the spatiotemporal spectrum for students in the corresponding Power value at location, This represents the average background power level.

10. The marine electromagnetic spectrum reconstruction method based on SNN distillation for spatiotemporal spectrum completion and multi-layer grid fusion as described in claim 9, characterized in that, It also includes S9, a step that verifies the spectral completion accuracy, physical consistency, algorithm efficiency, and robustness of the reconstructed electromagnetic spectrum, and outputs the verification results; among which, The method for verifying the spectral completion accuracy of the reconstructed electromagnetic spectrum is as follows: ； ； ； The method for verifying the physical consistency of the reconstructed electromagnetic spectrum is as follows: ； ； The method for verifying the algorithm efficiency and robustness of the reconstructed electromagnetic spectrum is as follows: ； In the formula, The root mean square error, For effective masking, For a single data point in a grid system, for The flattened representation, i.e., the reconstructed electromagnetic spectrum, is the first... The value at each position, for In the grid space, the first A flattened representation of the point location, i.e., the actual observed data at the [number]th [location]. The value of each position; The actual observed spectral power; For normalized mean square error, Peak signal-to-noise ratio, For signal sampling quantity, Mean square error; For physical residual measurement, For coverage of hotspot areas, To count the number of grid cells that satisfy the conditions within parentheses, For the fused tensor The channel output student completes the spatiotemporal spectrum. For hotspot power threshold, The sampling rate error curve SREC is the ratio of model reconstruction efficiency to the model reconstruction error. This represents the total number of data tests. For the number of data tests, Let m_time be the time of the teacher model's inference. Let m_time be the time taken for the student model to make its inference.

Images