Method for constructing and deducing electromagnetic situation knowledge graph based on multi-source spectrum data

Abstract

This invention discloses a method for constructing and extrapolating an electromagnetic situation knowledge graph based on multi-source spectrum data, belonging to the field of electromagnetic spectrum construction technology. The method involves feeding situation attribute features into a pre-defined adaptive spatiotemporal feature library, finding matching feature vectors similar to the current situation through time-matching retrieval, and applying the situation attribute features and matching feature vectors together to the adaptive spatiotemporal feature library for correction, generating corrected features. A situation state transition probability matrix is ​​initialized based on the state space, and a probability structure graph is constructed to generate a corrected probability matrix and prediction results. Based on the corrected probability matrix and corrected features, a multi-level knowledge graph is constructed. This construction method effectively improves the accuracy of electromagnetic spectrum situation construction, strengthens the ability to fuse and correlate cross-domain data, and enables prediction and visualization support for the status of spectrum resources and their service impact under complex electromagnetic environments.

Description

Technical Field

[0001] This invention relates to the field of electromagnetic map construction technology, specifically to a method for constructing and extrapolating electromagnetic situation knowledge graphs based on multi-source spectrum data. Background Technology

[0002] With the increasing strategic importance of electromagnetic space in modern information technology and civilian communications, dynamic monitoring, anomaly identification, and risk warning of spectrum resources have become key aspects of ensuring the stable operation of various radio services.

[0003] Existing electromagnetic situation analysis methods lack in-depth exploration of the evolution patterns across time and space, cannot fully incorporate empirical knowledge from similar historical scenarios to correct the current situation representation, and are difficult to dynamically quantify the impact on the state transition process when faced with sudden changes in environmental parameters or adjustments in control strategies, thus limiting the accuracy and timeliness of situation simulation results.

[0004] Therefore, there is an urgent need for a deductive method that can integrate multi-source heterogeneous spectrum data, introduce historical pattern correction and dynamic probability evolution modeling, and construct a multi-level semantic knowledge graph to break through the bottlenecks of existing technologies in cross-domain correlation, environmental adaptation and situation prediction, and achieve efficient cognition and accurate control of complex electromagnetic environments. Summary of the Invention

[0005] The purpose of this invention is to provide a method for constructing and extrapolating an electromagnetic situation knowledge graph based on multi-source spectral data, so as to solve the problems in the background technology.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for constructing and extrapolating an electromagnetic situation knowledge graph based on multi-source spectrum data, wherein the extrapolation method includes the following steps: S1: Collect multi-source heterogeneous spectrum data, map the multi-source heterogeneous spectrum data to the feature space through the feature coding network, generate original discrimination features, input the original discrimination features into the shared fusion layer for cross-domain correlation modeling, obtain situation attribute features, and divide the real-time state of each spatiotemporal frequency energy unit into different state spaces according to the discrete situation state classification mechanism. S2: Input the situation attribute features into the preset adaptive spatiotemporal feature library, find the matching feature vector similar to the current situation through time matching retrieval, apply the situation attribute features and matching feature vector together to the adaptive spatiotemporal feature library for correction, generate correction features, initialize the situation state transition probability matrix based on the state space, and construct a probability structure graph, traverse each connection in the probability structure graph, calculate the disturbance amplitude based on the correlation response strength and node risk factor, iteratively calculate the probability deformation through nonlinear interaction between nodes until the convergence condition is met, and generate the corrected probability matrix and prediction results; S3: Construct a multi-level knowledge graph based on the modified probability matrix and correction features.

[0007] Preferably, in step S2, each connection in the probability structure graph is traversed, the perturbation amplitude is calculated based on the correlation response strength and node risk factor, and the probability deformation is iteratively calculated through the nonlinear interaction between nodes until the convergence condition is met, generating the corrected probability matrix and prediction results, including the following steps: Traverse each connection in the probability structure graph and calculate the perturbation amplitude based on the correlation response strength and node risk factor; The probabilistic deformation is calculated iteratively through the nonlinear interaction between nodes. In each iteration, the current probability value of each node is weighted and aggregated based on the perturbation amplitude of the connected nodes. During the iteration process, the connection weights and node risk factors are continuously updated with the latest calculated probability values. This process is repeated until the probability changes of all connections are less than the preset convergence threshold, generating a corrected probability matrix and the corresponding prediction results.

[0008] Preferably, in step S2: the situation attribute features are fed into a preset adaptive spatiotemporal feature library, and a matching feature vector similar to the current situation is found through time matching retrieval, including the following steps: The adaptive spatiotemporal feature library coarsely aligns the timestamp of the current feature vector with the time labels of historical patterns in the library, and filters out a set of candidate patterns that are close in time. In the candidate pattern set, the spatiotemporal feature vector of each historical pattern is subjected to a dimension-wise similarity measurement with the current feature vector, and the matching degree is comprehensively evaluated by combining the trajectory shape matching algorithm, and several matching feature vectors are output.

[0009] Preferably, in step S2: the situational attribute features and the matching feature vector are applied together to the adaptive spatiotemporal feature library for correction to generate corrected features, including the following steps: Each matching result is assigned a confidence level label. When the confidence level label is higher than the preset confidence level threshold, the feature library correction mechanism is triggered. The fused feature vectors are used to adjust the representation vectors of relevant historical patterns in the feature library, update the position and trajectory parameters of the representation vectors in the spatiotemporal feature space, and generate corrected features.

[0010] Preferably, a confidence level label is assigned to each matching result. When the confidence level label is higher than a preset confidence threshold, a feature library correction mechanism is triggered, including the following steps: The current situation attribute features and the matching feature vector are used together as correction inputs and applied to the online learning module of the feature library; An incremental feature fusion strategy is adopted to concatenate the current feature and the matching feature in a unified dimensional space; The contribution of features from different sources is dynamically adjusted through an attention weight allocation mechanism. If a certain type of environmental interference in the historical scene corresponding to the matching feature is consistent with the current environment, the attention weight of the matching feature is increased, and vice versa.

[0011] Preferably, in step S1: the original discriminative features are input into the shared fusion layer for cross-domain correlation modeling to obtain situational attribute features, including the following steps: For each spatiotemporal frequency energy unit, the feature correlation weights of the spatiotemporal frequency energy unit and other related units are calculated through a self-attention mechanism to identify the local contextual associations that affect the state of the spatiotemporal frequency energy unit. A graph neural network is introduced to construct a global association graph, where nodes are the original discriminative features of all spatiotemporal frequency and energy units, and edge weights are initialized by weights obtained from local association mining. By iteratively updating node features through multi-layer graph convolution, scattered cross-domain associations across the global scope are integrated into the node features; Pooling is performed on the node features output by multi-layer graph convolution, and situational attribute features are generated by combining the semantic constraints of the business requirement text.

[0012] Preferably, in a multi-level knowledge graph, node attributes are embedded with time-frequency-space-energy four-domain features and situational states, and edge weights are integrated with dynamic weights and confidence labels to reflect the cross-domain correlation strength and prediction reliability.

[0013] Preferably, a multi-level knowledge graph is combined with real-time environmental parameters and environmental interference features. The environmental parameters include background noise levels and meteorological propagation influencing factors. These parameters are incorporated into cross-attention calculations, enabling the interaction between situational attribute features and real-time environmental information to generate feature interaction vectors. By analyzing the contribution of each feature in situation prediction and visualization through a weight learning mechanism, multi-source dynamic weights are generated, and the feature interaction vectors are weighted and fused to obtain multi-source coordinated features. Affine transformation operations are performed in conjunction with real-time environmental parameters, and updates are performed using multi-source coordination features.

[0014] Preferably, the multi-level knowledge graph analyzes the contribution of each feature in situation prediction and visualization through a weight learning mechanism, generates multi-source dynamic weights, performs weighted fusion of feature interaction vectors to obtain multi-source coordinated features, performs affine transformation operations in conjunction with real-time environmental parameters, and updates using multi-source coordinated features, including the following steps: In multiple training samples, the situation prediction error and the visualization clarity index are used as supervision signals to construct a weighted learning model. The weighted learning model takes each sub-feature in the feature interaction vector as input and outputs the contribution score of each sub-feature through a trainable scoring network. The contribution scores are normalized to form multi-source dynamic weights, and the feature interaction vectors are weighted and fused according to the multi-source dynamic weights to generate multi-source coordinated features. Based on real-time environmental parameters, an affine transformation matrix and a bias vector are constructed. The multi-source coordination features are regarded as a set of points in a high-dimensional space. The affine transformation is applied to adapt the geometric structure and align the statistical distribution. After the transformation is completed, the final multi-source coordinated features are obtained, written into the attribute fields of the corresponding nodes of the multi-level knowledge graph, and the weights of the edges are updated to complete the dynamic evolution of the graph.

[0015] Preferably, the sources of multi-source heterogeneous spectrum data include fixed monitoring stations, mobile detection platforms, space remote sensing receivers, and service requirement description texts; The state space includes a stable and controllable region, an anomaly warning region, and a region with severe interference / failure risk.

[0016] The technical effects and advantages provided by the present invention in the above technical solution are as follows: 1. This application introduces an adaptive spatiotemporal feature library for time matching retrieval and confidence level discrimination. This allows for feature correction when the current situation is highly similar to historical patterns, significantly enhancing the model's adaptability and representation accuracy to new and marginal scenarios, and avoiding underfitting of rare or abrupt situations. Based on the perturbation propagation and iterative correction mechanism of the probability matrix and probability structure diagram initialized in the state space, it can quantify external factors such as environmental changes and policy strength adjustments as nonlinear effects on state transition probabilities. This overcomes the limitations of traditional static probability models in reflecting the coupling effect of dynamic regulation and complex environments, thereby outputting a corrected probability matrix and prediction results that are closer to the actual evolution law, providing a reliable quantitative basis for early identification and graded response of situational risks.

[0017] 2. This application constructs a multi-level knowledge graph, organically organizing four-domain features, situational states, cross-domain association strength, and confidence into an interpretable and navigable semantic network, achieving a panoramic mapping from micro-signal features to macro-business impacts. By combining cross-attention calculation of real-time environmental parameters and multi-source dynamic weight learning, it enables bidirectional deep interaction between situational attribute features and environmental information, ensuring that the graph can dynamically reflect the immediate impact of environmental factors such as background noise, meteorological propagation, and sudden interference. Furthermore, by correcting the geometric and statistical properties of features through environment-driven affine transformations, it ensures the consistency and comparability of graph nodes and edge attributes in time-varying environments, thereby enabling graph updates to retain the stability of historical knowledge while possessing the ability to respond quickly to real-time disturbances. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0019] Figure 1 This is a flowchart of the derivation method of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Example: This example provides a method for constructing and extrapolating an electromagnetic situation knowledge graph based on multi-source spectrum data. Please refer to [link / reference]. Figure 1 As shown, the deduction method includes the following steps: S1: In response to the start signal of the electromagnetic spectrum situation building task (such as the detection of cross-domain anomalies), it collects heterogeneous spectrum data from multiple sources, including fixed monitoring stations, mobile detection platforms, space remote sensing receivers, and service requirement description texts. Through a feature encoding network, the data from different sources are mapped to a unified time-frequency-spatial-energy domain feature space to generate raw discriminative features. These raw discriminative features are then input into a shared fusion layer for cross-domain correlation modeling to obtain situation attribute features. Based on a discrete situation state classification mechanism, the real-time state of each spatiotemporal frequency-energy unit is divided into a stable and controllable zone, an anomaly warning zone, and a severe interference / failure risk zone, providing different state spaces for subsequent probabilistic evolution and spectrum construction.

[0022] S2: The situation attribute features are fed into a preset adaptive spatiotemporal feature library, which contains historical electromagnetic situation patterns and their spatiotemporal evolution trajectories. The matching feature vectors similar to the current situation are found through time matching retrieval and assigned a confidence level label. If the confidence level label is higher than the preset confidence level threshold, the situation attribute features and the matching feature vectors are applied together to the feature library for correction to generate correction features, thereby enhancing the accuracy of the representation of the current scene.

[0023] The state transition probability matrix is ​​initialized based on the state space to describe the evolution probability of each state under the existing control strategy. To assess the impact of environmental changes and strategy strength adjustments, a probability structure graph is constructed (nodes represent state units, and lines represent state transition relationships). Each line in the probability structure graph is traversed, and the disturbance amplitude is calculated based on the associated response strength and node risk factors. The probability deformation variables are iteratively calculated through nonlinear interactions between nodes until the convergence condition is met, generating a corrected probability matrix and prediction results.

[0024] S3: Based on the corrected probability matrix and correction features, a multi-level knowledge graph is constructed, encompassing abnormal situation elements, monitoring equipment, spatial location, and operational impact. Node attributes are embedded with features from the four domains of time, frequency, space, and energy, along with situation status. Edge weights are fused with dynamic weights and confidence markers to reflect cross-domain correlation strength and prediction reliability. Combining real-time environmental parameters (such as background noise levels and meteorological propagation impact factors) with environmental interference features, cross-attention calculations are performed to enable deep interaction between situation attribute features and real-time environmental information, generating feature interaction vectors. A weight learning mechanism is used to analyze the contribution of each feature to situation prediction and visualization, generating multi-source dynamic weights. These weights are then used to weight and fuse the feature interaction vectors, resulting in multi-source coordinated features. Finally, affine transformation operations are performed based on real-time environmental parameters to achieve dynamic correction of geometric and statistical properties, generating the final multi-source coordinated features for graph updates.

[0025] In this application, the detailed steps of each process are as follows: The S1 process revolves around the systematic acquisition of multi-source heterogeneous spectrum data, unified feature mapping, cross-domain correlation modeling, and discrete situational state classification, forming a complete processing link from raw observation to structured situational attributes.

[0026] Specifically, upon receiving the start signal for the electromagnetic spectrum situation building task, the trigger scenario is that the monitoring subsystem detects cross-domain abnormal signals, such as unlicensed occupancy of multiple frequency bands simultaneously, sudden changes in cross-regional signal parameters, or a sharp drop in service carrying performance—abnormal events with significant cross-domain coupling characteristics. The system immediately activates the multi-source data acquisition module to synchronously aggregate data from four core data sources: Fixed monitoring stations deployed in key areas or with wide coverage acquire signals with high temporal resolution and stable spatial reference; mobile detection platforms with mobility, such as vehicle-mounted, ship-mounted, or UAV-borne detection units, can capture spectrum distributions that change with spatial location in dynamic scenarios; space remote sensing receivers, typically mounted on low-Earth orbit or geostationary orbit satellites, can acquire panoramic spectrum information over a large spatial scale, compensating for the limited field of view of ground and near-Earth platforms; and service requirement description texts, which are data derived from explicit requirements of users or command and decision-making levels for specific frequency bands, service types, or security levels, including semantic constraints such as service priority, anti-interference thresholds, and available bandwidth ranges.

[0027] The above four types of data exhibit significant heterogeneity in modality (time-series sampled signals, spatial images, text semantics), structure (structured numerical matrices, unstructured character sequences), and spatiotemporal reference (fixed coordinates, moving trajectories, satellite nadir points, and no explicit spatial anchor points), and need to be uniformly represented through subsequent feature encoding.

[0028] Specifically, for the unified mapping of multi-source heterogeneous data, the system uses a feature encoding network for processing: The first layer is modal feature extraction. For time-series spectral signals collected by fixed monitoring stations and mobile detection platforms, the signals are first segmented into local time-domain segments using a sliding window. For each segment, time-domain envelope analysis (extracting the trend of signal strength change over time, such as rising slope and fluctuation period), frequency-domain spectral peak detection (identifying main lobe frequency, bandwidth, and adjacent channel leakage ratio), and energy centroid calculation after short-time Fourier transform (determining the distribution center of energy in the frequency domain) are performed in parallel. For spectral data from space remote sensing receivers, spatial grid division is first performed (e.g., dividing into several pixel units according to latitude and longitude or nadir projection). Within each grid, time-frequency analysis similar to the above is performed, and spatial gradient calculation is additionally superimposed (the difference in spectral parameters between adjacent grids reflects the abrupt change in spatial distribution). For business requirement description text, key terms (such as military frequency band, emergency communication, and anti-interference level A) are first extracted through word segmentation and entity recognition in natural language processing. Then, based on a pre-trained domain semantic embedding model, the terms are mapped into high-dimensional semantic vectors, and the overall semantics of the text are weighted and aggregated through an attention mechanism to highlight constraints strongly related to the current monitoring task.

[0029] The second layer is cross-modal dimensional alignment. For the differences in the dimensionality of feature vectors output by different modalities (e.g., time-frequency features are 128-dimensional, spatial gradient features are 64-dimensional, and text semantic vectors are 256-dimensional), a learnable linear projection layer is used to uniformly project each modal feature to the same intermediate dimension (e.g., 512-dimensional). At the same time, the projection process is optimized by contrastive learning loss function, so that semantically similar features of different modalities (e.g., high-priority business in text and high-energy time-frequency features of corresponding frequency bands) are closer in the projection space, thus enhancing cross-modal semantic consistency.

[0030] The third layer is a unified spatial projection, which concatenates the modal features after dimensional alignment into a joint feature vector. This vector is then input into a multilayer perceptron (MLP) for nonlinear transformation, ultimately outputting the original discriminative features in a unified four-dimensional feature space encompassing the time domain, frequency domain, spatial domain, and energy domain. The time domain dimension characterizes the temporal correlation of features (such as duration and rate of change), the frequency domain dimension characterizes the frequency location and distribution characteristics, the spatial domain dimension characterizes the distribution patterns on geographical or spatial coordinates, and the energy domain dimension characterizes the signal power intensity and its dynamic range. These four dimensions together constitute a fine-grained description of the spectral state.

[0031] The generated original discriminative features are input into the shared fusion layer for cross-domain association modeling. The processing logic is as follows: For each spatiotemporal frequency-energy unit (i.e., the smallest analytical unit with a specific time slice, spatial location, frequency range, and energy interval), the feature correlation weights of this unit and other related units (such as adjacent time slices, nearby spatial locations, adjacent frequency bands, and units with similar energy levels) are calculated through a self-attention mechanism. Local contextual associations that have a significant impact on the state of this unit are identified (for example, a sudden increase in energy in a certain region during a certain period may be causally related to a frequency jump in an adjacent region at the previous moment). A graph neural network (GNN) is introduced to construct a global association graph. The nodes are the original discriminative features of all spatiotemporal frequency energy units. The edge weights are initialized by weights obtained from local association mining. The node features are iteratively updated through multi-layer graph convolution, so that scattered cross-domain associations (such as the risk of service interruption caused by the coordination of abnormal signals in different regions and frequency bands) can be integrated into the node features. Pooling operations are performed on the node features output by GNN (such as global pooling based on attention weighting to highlight the contribution of high-risk areas), and semantic constraints of business requirement text are combined (such as using the anti-interference threshold extracted from the text as a filtering condition to strengthen the weight of security-related attributes). Finally, situational attribute features covering core elements such as anomaly correlation, risk transmission path, and business impact are generated, realizing the attribute sublimation of spectrum status from single-point observation to global situation.

[0032] A fixed monitoring station collected a data segment of length 1 at a certain time slice. The time-series spectrum signal of seconds is passed through a sliding window (window length 0.2). Seconds, step size 0.1 The signal was segmented into 10 local time-domain segments (0.5 seconds). One segment was selected for time-domain envelope analysis, and the measured signal strength ranged from 0.5... W increased to 1.5 W, time taken 0.05 If the time is seconds, then the slope of the rising edge is calculated as (1.5-0.5) / 0.05. = 20 W / s indicates a rapid increase in signal strength in this segment; the main lobe center frequency was detected to be 2.45 MHz. GHz, bandwidth 20 MHz, adjacent channel leakage ratio -35 dB; the energy centroid after short-time Fourier transform is calculated as ∑(E ·f ) / ∑E Where E1=0.8 and f1=2.44 GHz, E2=1.2, f2=2.46 GHz, the energy centroid ≈ (0.8 × 2.44 + 1.2 × 2.46) / (0.8 + 1.2) = 2.452 GHz indicates that energy is concentrated in the vicinity of that frequency point.

[0033] The space remote sensing receiver divides the corresponding area into four grids with latitude and longitude of 0.1° × 0.1°. Time-frequency characteristics are also obtained within a certain grid, and the frequency domain main lobe difference Δf = 15 is calculated between the receiver and the eastward adjacent grid. MHz, as a spatial gradient, reflects abrupt changes in spectral parameters. The business requirement text, after word segmentation and entity recognition, extracts the military frequency band and anti-interference level A. This is then mapped to 256-dimensional vectors v1=[0.2,0.5,…] and v2=[0.6,0.1,…] using a domain semantic embedding model. After weighting via an attention mechanism (with weights α1=0.7 and α2=0.3), the text semantic vector V is obtained. =α1·v1+α2·v2=[0.34,0.38,…].

[0034] In the cross-modal dimension alignment stage, the 128-dimensional time-frequency features, 64-dimensional spatial gradients, and 256-dimensional textual semantics are mapped to 512 dimensions through a learnable linear projection layer. For example, time-frequency features x∈ 128 via W1∈ 512×128Projecting x' = W1·x, similarly we obtain the spatial gradient y' and text z'. By comparing the learning loss function L = max(0,m + D(x',z') - D(x',y')) (where m = 0.5 and D is the Euclidean distance), we optimize the high-priority business text and the corresponding high-energy time-frequency features to be closer in the projection space.

[0035] During unified spatial projection, the three elements are concatenated into a 1536-dimensional joint feature vector, which is then input into an MLP (two fully connected layers + ReLU) to obtain the original discriminative features f=MLP([x';y';z']). Its four dimensional components are: temporal correlation T=0.78 (indicating a relatively long duration and moderate rate of change), frequency domain position F=2.452, and so on. GHz (energy centroid), spatial coordinates S=(lat=30.25°, lon=120.15°) (grid center), energy intensity E=1.42 W (reflects power level).

[0036] Input this f into the shared fusion layer, and for a certain spatiotemporal frequency energy unit u, calculate the correlation weight softmax((Q_u·K_v) using self-attention. T In the expression ) / √d), Q_u and K_v are the query and key matrices of u and other unit features, respectively, d is the feature dimension, and the weight of u and its neighboring unit v at the previous time step is 0.85, which is identified as a significant local association. A global GNN is constructed, with nodes representing all unit features and edge weights initialized to local attention values. After two layers of graph convolution, the feature h_u of node u is updated by fusing information from the distant unit w (weight 0.62), reflecting cross-regional abnormal collaboration. Attention-weighted global pooling yields H=Σβ_i·h_i (β_i is the normalized weight of risk factors for each node; if u belongs to a high-risk area, β_u=0.4), combined with the anti-interference threshold of 0.5 extracted from the text. After filtering with W, the final generated situational attribute features g=[abnormal correlation=0.82, risk transmission path length=3, business impact=0.76] are significantly improved in terms of global semantic and threat assessment capabilities compared to single-point observation.

[0037] Specifically, after extracting the situational attribute features, the system classifies the real-time state of each spatiotemporal frequency energy unit according to a discrete situational state hierarchical mechanism. This mechanism achieves objective classification based on a preset multi-dimensional criterion set and hierarchical logic, as follows: For the situational attribute characteristics of each unit, key quantitative indicators are extracted, including the strength of abnormal correlation (measured by the average weight of the edges in the correlation graph, reflecting the degree of coupling between the unit and other abnormal units), the probability of risk transmission (based on the proportion of similar feature combinations in historical data that cause serious interference, combined with Bayesian inference to update the current probability), and the deviation of business assurance (the degree of deviation between the actual spectrum parameters and the threshold in the business requirement text, such as the proportion of actual bandwidth exceeding the allowable limit).

[0038] A three-tiered threshold system is established. When the abnormal correlation strength is below the weak correlation threshold, the risk transmission probability is below the low risk threshold, and the service assurance deviation is within an acceptable range, the unit is determined to be in a stable and controllable zone, indicating that the current spectrum status is stable and poses no significant threat to service operation. When the abnormal correlation strength reaches the medium correlation threshold, or the risk transmission probability enters the medium risk range, or the service assurance deviation exceeds the mild alarm threshold but does not reach the severe level, it is determined to be an abnormal warning zone, indicating that monitoring needs to be strengthened and intervention measures need to be prepared. When the abnormal correlation strength is above the strong correlation threshold, the risk transmission probability exceeds the high risk threshold, and the service assurance deviation reaches the severe violation threshold (such as the critical service frequency band being completely occupied by unauthorized signals), it is determined to be a severe interference / failure risk zone, meaning that the spectrum resources can no longer support the service normally, and an emergency response needs to be initiated immediately.

[0039] Each unit is labeled with its state category and the key indicator values ​​for classification are recorded to form a structured state space. This state space not only contains discrete state labels, but also implies the evolutionary tendency between states (such as the possibility that an abnormal warning area may migrate to a serious interference / failure risk area if not dealt with in time).

[0040] After extracting the situational attribute features of a certain unit, three key quantitative indicators are obtained. First, the anomaly correlation strength A is calculated by averaging the weights of all edges connected to the unit in its correlation graph. Let the edge weights between this unit and the five surrounding anomaly units be 0.12, 0.18, 0.25, 0.31, and 0.40, respectively. Then, A = (0.12 + 0.18 + 0.25 + 0.31 + 0.40) / 5 = 0.252, reflecting a moderate to weak coupling degree between this unit and the anomaly units. Next, the risk transmission probability R is calculated based on historical data. According to statistics, the proportion of similar feature combinations that caused serious interference in the past 100 observations was 20%, i.e., the prior probability P0=0.20. Combining Bayesian inference with the current environmental factors (such as the likelihood correction coefficient L=1.5 for the increase in the number of sudden interference signals detected in the neighboring area), the posterior probability is approximately R=(P0×L) / (P0×L+(1-P0))=(0.20×1.5) / (0.20×1.5+0.80)=0.23, indicating that the risk of transmission to serious interference under the current conditions is still low to moderate. Calculate the service assurance deviation D, assuming the service requirement text specifies that the maximum allowable bandwidth occupied by this unit in the frequency band is 10. MHz, actual measured bandwidth usage is 13 If the value is MHz, then D = (measured value - threshold) / threshold = (13 - 10) / 10 = 0.30, which means it exceeds the allowable upper limit by 30%. An example of a preset three-level criterion threshold is as follows: The thresholds are: weak correlation A1 = 0.30, medium correlation A2 = 0.50, low risk threshold R1 = 0.25, high risk threshold R2 = 0.60, acceptable deviation range for business assurance D < 0.10, minor alarm threshold 0.10 ≤ D < 0.40, and serious violation threshold D ≥ 0.40. Based on the calculated values ​​A = 0.252 (lower than A1 = 0.30), R = 0.23 (lower than R1 = 0.25), and D = 0.30 (falling into the minor alarm range but not reaching the serious violation threshold), D triggers the minor alarm condition. According to the hierarchical logic, this unit is determined to be in an abnormal warning zone, indicating the need for strengthened real-time monitoring and prepared intervention measures. The system labels the state of this unit as an abnormal warning zone and records (A=0.252, R=0.23, D=0.30) as the basis for classification, storing it in the structured state space. At the same time, based on the historical evolution statistical model, it infers that if this state is not dealt with in subsequent time slices, the probability of it migrating to the severe interference / failure risk zone is about 0.45, providing state transition tendency information for subsequent probability evolution and map construction.

[0041] The execution of the S2 process focuses on further introducing historical knowledge guidance and dynamic probabilistic evolution modeling based on the discrete state space constructed in S1. This aims to achieve a continuous characterization of the electromagnetic spectrum situation over time and predictable analysis under the influence of policy and environmental changes. It consists of two main stages: First, it uses an adaptive spatiotemporal feature library to complete similarity matching and feature correction of the current situation, improving the accuracy and adaptability of the situation representation; second, it initializes the probability transition framework based on the state space and quantifies and iteratively corrects environmental and policy disturbances by constructing a probability structure diagram, ultimately outputting a corrected probability matrix and situation evolution trend that can be used for prediction.

[0042] Specifically, in the first stage, the system sends the situational attribute feature vector output by S1 into a pre-set adaptive spatiotemporal feature library. This feature library has systematically collected and encoded a large number of historical electromagnetic situational patterns during the construction phase. It not only saves multi-dimensional feature snapshots under different time slices, but also records the evolution trajectory of these patterns in the spatiotemporal dimension. For example, it records the complete process chain of a certain area gradually evolving from a stable and controllable area to an abnormal warning area and then to a serious interference / failure risk area during a specific period, as well as the correlation records of accompanying environmental factors (such as enhanced solar activity, sudden dense occurrence of military exercise signals) and control strategies (such as temporary frequency reallocation, power limitation instructions).

[0043] The feature library employs a spatiotemporal joint index structure, enabling it to automatically perform time-matching retrieval after receiving current situation attribute features. The timestamp of the current feature vector is coarsely aligned with the timestamps of historical patterns in the library to filter out a set of candidate patterns that are temporally close (such as those from the same season, the same day-night cycle, or the same task background window). Then, in the candidate set, a dimension-wise similarity measure is performed between the spatiotemporal feature vector of each historical pattern and the current feature vector. The measurement methods include cosine similarity to determine directional consistency, Euclidean distance to capture amplitude differences, and a trajectory shape matching algorithm (comparing the morphological similarity of two evolutionary trajectories in the state sequence, such as the degree of consistency between the state transition order and the interval time) to comprehensively evaluate the matching degree. Finally, several matching feature vectors are output and a confidence label is assigned to each matching result. This confidence is calculated by weighted fusion of the normalized result of the similarity measure value and the trajectory consistency, ensuring that both instantaneous feature similarity and evolutionary path consistency are considered.

[0044] The current situation attribute feature vector output by S1 is F_c, with its timestamp being the spring afternoon task background window. The feature library has stored a large number of historical patterns during construction, and each pattern contains a time label, spatiotemporal feature vector, and state evolution trajectory. During retrieval, coarse-grained time alignment is performed, and a set of four candidate patterns with close temporal proximity is obtained, corresponding to spring afternoon, spring afternoon, same time of the same season in the same year, and autumn morning, with different backgrounds.

[0045] For the spatiotemporal feature vectors F_hi and F_c of candidate pattern i, calculate the cosine similarity Sim_cos(i) and the Euclidean distance class metric Sim_euc(i) respectively (here, the normalized similarity form Sim_euc = 1 - normalized Euclidean distance is used), and design the calculation results: Pattern 1: Sim_cos(1)=0.92, Sim_euc(1)=0.88; Pattern 2: Sim_cos(2)=0.85, Sim_euc(2)=0.80; Mode 3: Sim_cos(3)=0.96, Sim_euc(3)=0.93; Pattern 4: Sim_cos(4)=0.60, Sim_euc(4)=0.55; Next, the trajectory shape matching algorithm is executed to compare the evolution of the historical patterns with the current situation. For example, the evolution trajectory of pattern 3 is stable and controllable zone (10:00) → abnormal warning zone (11:30) → severe interference / failure risk zone (13:00). The trajectory of the current situation before 14:30 is stable and controllable zone (12:00) → abnormal warning zone (13:45). The two are highly consistent in terms of the state transition order (stable → abnormal warning) and the time interval ratio. The trajectory consistency Match_trj(3) = 0.94. The trajectory consistency of pattern 1 is 0.75, that of pattern 2 is 0.68, and that of pattern 4 is 0.40.

[0046] The confidence score Conf(i) is calculated by weighted fusion of normalized similarity measures and trajectory matching scores. Let the cosine similarity weight w1 = 0.4, the Euclidean similarity weight w2 = 0.3, and the trajectory matching score weight w3 = 0.3. Each similarity measure is first normalized to the [0,1] interval (this is already a normalized value). The calculation formula is: Conf(i) = w1·Sim_cos(i) + w2·Sim_euc(i) + w3·Match_trj(i), Substituting the data for pattern 3: Conf(3) = 0.4 × 0.96 + 0.3 × 0.93 + 0.3 × 0.94 = 0.384 + 0.279 + 0.282 = 0.945, Pattern 1: Conf(1) = 0.4 × 0.92 + 0.3 × 0.88 + 0.3 × 0.75 = 0.368 + 0.264 + 0.225 = 0.857, Pattern 2: Conf(2 =0.4×0.85+0.3×0.80+0.3×0.68=0.340+0.240+0.204=0.784, Mode 4: Conf(4)=0.4×0.60+0.3×0.55+0.3×0.40=0.240+0.165+0.120=0.525, and finally output matching feature vectors as Mode 3, Mode 1, and Mode 2, and assign confidence labels of 0.945, 0.857, and 0.784 respectively. Mode 4 is removed because its confidence is lower than the commonly used preset confidence threshold of 0.80.

[0047] Specifically, when the confidence level is higher than the preset confidence threshold, it means that the current situation and a certain historical pattern are highly similar in structure and evolution trend. At this time, the system triggers the feature library correction mechanism: the current situation attribute features and the matching feature vector are used as correction inputs and applied to the online learning module of the feature library.

[0048] An incremental feature fusion strategy is adopted, which first concatenates the current feature and the matching feature in a unified dimensional space, and then dynamically adjusts the contribution of features from different sources through an attention weight allocation mechanism. For example, if a certain type of environmental disturbance (such as ionospheric disturbance) in the historical scene corresponding to the matching feature is highly consistent with the current environment, the attention weight of that part of the feature is increased, and vice versa. The fused feature vectors are used to fine-tune the representation vectors of relevant historical patterns in the feature library, and update their position and trajectory parameters in the spatiotemporal feature space to avoid representation lag caused by insufficient learning of new scene characteristics, thereby generating corrective features.

[0049] The generation of corrective features not only improves the accuracy of the current scene representation, but also enables the feature library to adapt and improve itself as the environment evolves, providing basic data that is closer to the real situation for the input of subsequent probabilistic models.

[0050] The current situational attribute feature vector is F_c (dimension 512). A matching feature vector F_m (dimension 512) is retrieved from the adaptive spatiotemporal feature library. The corresponding historical scene has experienced ionospheric disturbances, with a confidence score of Conf=0.945, which is higher than the preset confidence threshold of 0.80. The system triggers the incremental feature fusion strategy of the online learning module. First, F_c and F_m are concatenated in a unified dimensional space to obtain the joint feature vector F_concat=[F_c;F_m] (dimension 1024).

[0051] An attention weighting mechanism is used to dynamically adjust the contributions of the two parts: Let the environment matching score be e = 0.88 (derived from the current environment parameters and the environment label matching algorithm of the historical scene, e ∈ [0,1], the higher the e, the more consistent the environmental interference type), then the current feature attention weight α_c and the matching feature attention weight α_m are calculated using softmax normalization: α_c = exp(e·γ) / (exp(e·γ) + exp((1-e)·γ)), α_m = 1 - α_c, where γ is the sharpening coefficient (γ = 5 is taken to enhance the difference). Substituting e = 0.88, we get α_c ≈ exp(4.4) / (exp(4.4) + exp(0.6)) ≈ 81.45 / (81.45 + 1.82) ≈ 0.978, α_m ≈ 0.022. This indicates that due to the highly consistent ionospheric disturbance environment, although the matching features come from history, their contribution weight is still appropriately retained to enhance the common features.

[0052] The fused feature vector F_fuse is calculated as follows: F_fuse=α_c·F_c+α_m·F_m (After dimension alignment, the weighted sum can be directly calculated; if concatenated, attention gating is used to select the splicing subspaces, which is simplified to weighted fusion here). Let F_c=[0.52,0.33,…], F_m=[0.48,0.36,…], then F_fuse[1]=0.978×0.52+0.022×0.48≈0.519+0.011≈0.530. Similarly, the remaining dimensions can be obtained to form the fusion features before correction.

[0053] Then, F_fuse is used to fine-tune the representation vectors of the historical pattern and its neighboring patterns in the feature library. Let the original representation of the historical pattern be H_old. The fine-tuning adopts gradient descent to update in one step: H_new=H_old+η·(F_fuse-H_old), learning rate η=0.1. If H_old=[0.50,0.30,…], then H_new[1]=0.50+0.1×(0.530-0.50)=0.503. After updating dimension by dimension, a new representation vector is formed, so that the position of the mode in the spatiotemporal feature space is closer to the current scene. The trajectory parameters (such as the state transition time point) are also finely adjusted according to the residual ratio to avoid lag due to the new scene characteristics not being learned.

[0054] The final generated corrected feature is the updated H_new (or the result of its re-fusion with F_fuse). It not only more accurately depicts the current situation, but also allows the feature library to directly call upon more realistic representations when encountering similar environmental disturbances in the future, thereby improving online adaptability and the input quality of subsequent probabilistic models.

[0055] Specifically, in the second stage, the system constructs an initial version of the state transition probability matrix based on the discrete situational state space initialized by S1. Each element of this matrix describes the probability that a certain state unit (such as the state label of a certain spatiotemporal frequency energy unit) will transition to another state in the next time step under a specific existing control strategy framework. The initialization is based on historical statistical data and expert rules. For example, the probability of a stable and controllable zone maintaining its own state under conventional strategies is relatively high, while the probability of a direct jump to a severely disturbed / failure risk zone is relatively low.

[0056] To assess the impact of environmental changes (such as an increase in sudden signal sources or a rise in natural noise background) and policy strength adjustments (such as increasing monitoring frequency or implementing stricter power control) on the probability of state transitions, the system further constructs a probabilistic structure diagram: The nodes in the graph represent individual state units, and the lines represent possible transition relationships between states. Each line carries a basic transition probability value and associated contextual attributes (such as the frequency band range involved, spatial adjacency, and energy change amplitude). Traverse each connection in the probability structure diagram, and calculate the disturbance amplitude based on the correlation response strength (reflecting the intensity of the effect of environmental and strategic factors on this transfer path, which can be obtained through correlation analysis of real-time monitoring data and strategy execution records, such as a significant increase in the triggering frequency of related connections after the implementation of a certain strategy) and node risk factor (a comprehensive assessment of the risk transmission probability in the historical abnormal frequency and current situation attribute characteristics of the node). The meaning of the disturbance amplitude is the amount of offset driving force of this environmental and strategic change on the basic transfer probability. The probability deformation variables are calculated iteratively through nonlinear interactions between nodes. Specifically, in each iteration, the current probability value of each node and the disturbance amplitude of connected nodes are weighted and converged. The interaction function between nodes is designed as a saturable nonlinear form to avoid probability overflow from a reasonable range due to extreme disturbances. During the iteration process, the connection weights and node risk factors are continuously updated with the latest calculated probability values. This process is repeated until the probability change of all connections is less than the preset convergence threshold. At this point, a corrected probability matrix and corresponding prediction results are generated. The prediction results include the probability distribution map of each state unit in the future for several time steps, which can be directly used to identify high-risk evolution paths and formulate forward-looking control plans.

[0057] S1 has divided a certain region into three spatiotemporal frequency energy units, with initial states of a stable and controllable zone (S), an anomaly warning zone (W), and a severe interference / failure risk zone (R), respectively. The state transition probability matrix P0 is initialized based on historical statistical data and expert rules, where S→S=0.80, S→W=0.15, S→R=0.05, W→S=0.10, W→W=0.70, W→R=0.20, and R→S=0.02, R→W=0.18, R→R=0.80. To assess the impact of environmental changes and strategy intensity adjustments, the system constructs a probability structure diagram: Nodes N1, N2, and N3 correspond to the three units mentioned above, with connections L12 representing S→W, L23 representing W→R, etc. Each connection carries the basic transition probability p_base and contextual attributes (e.g., L23 involves frequency band 3.5). GHz, spatial adjacency, energy amplification 8 When traversing the connections, first calculate the disturbance amplitude Δp, with the formula Δp(L)=λ·I(L)·Ω(N_src,N_dst), where λ is the disturbance gain coefficient (taken as 0.1), I(L) is the correlation response strength (real-time monitoring shows that the frequency of interference triggering in this band has increased by 50% compared to the historical average recently, so I(L23)=0.5), and Ω(N_src,N_dst) is the node risk factor (N2 has a historical abnormal frequency ratio of 0.3, the current risk transmission probability is 0.23, and after normalization Ω(N2,N3)=(0.3+0.23) / 2≈0.265). Substituting these values, we get Δp(L23)=0.1×0.5×0.265=0.01325.

[0058] Then, the iterative calculation of probability shape variables is performed: Let the transition probability of N2→N3 in round t be p_t(L23). The interaction function adopts a saturable nonlinear form f(x)=x / (1+|x|) to prevent overflow. In each round, it is updated as p_{t+1}(L)=p_t(L)+f(Σ_jw_j·Δp_j), where w_j is the contribution weight of the connected node j to the connection (for L23, it is mainly affected by the probability of state N2 and the risk factor of N3, let w_N2=0.6, w_N3=0.4). In the first round, Σ_jw_j·Δp_j=0.6×0.01325+0.4×0.010 (L23 is affected by another connected edge Δp=0.010)=0.01195, f(0.01195)≈0.01189, so p1(L23)=0.20+0.01189=0.21189. After updating the connection weights and node risk factors, proceed to the next round. Stop when all connections |p_{t+1}-p_t| < ε (convergence threshold ε = 0.0005). After three rounds of iteration, p(L23) converges to 0.216.

[0059] The final corrected probability matrix P_c is generated, where W→R increases from 0.20 to 0.216, indicating that the probability of this path transition increases under the influence of enhanced environmental interference and strategy adjustment. The prediction results provide the probability distribution of the state of each unit at the next three time steps. For example, the probability of N2 being in R at time t+1 increases from 0.20 to 0.216, which can identify this as a high-risk evolution path in advance and guide the control plan to prioritize strengthening the monitoring and power control of this frequency band and region, thereby improving the foresight of the response and the effectiveness of decision-making.

[0060] The core objective of the S3 process is to establish a multi-level knowledge graph that can simultaneously express the intrinsic correlation of abnormal situation elements, the layout of monitoring methods, spatial distribution, and business impact chains, based on the corrected probability matrix and correction features obtained in S2. Through deep integration with real-time environmental information and dynamic weight adjustment, it generates multi-source coordinated features with spatiotemporal consistency and predictive reliability to continuously update the graph and support subsequent situation visualization and decision support.

[0061] Specifically, nodes and edges are organized into a four-layer structure: abnormal situation elements, monitoring equipment, spatial location, and business impact, forming a knowledge system rich in both semantics and topology. Node attribute embedding follows the four-domain characteristics defined in S1: time domain, frequency domain, spatial domain, and energy domain. It also includes situational state labels generated by S1 and S2 (such as stable and controllable zone, abnormal warning zone, and severe interference / failure risk zone) and the transition probability distribution of the corresponding state unit in the corrected probability matrix. The processing logic is as follows: Each type of node has a defined set of attributes. For example, the abnormal situation element node includes the detection time window, dominant frequency range, spatial coverage, peak energy and current status of the element; the monitoring equipment node includes the equipment type, deployment coordinates, sensitivity parameters and historical detection success rate; the spatial location node is indexed by geographic grid or 3D coordinates and records the background electromagnetic environment baseline of the location; the business impact node is bound to the business type, priority, required frequency band and power tolerance, historical impact frequency, etc.

[0062] Edge weights combine two parts of information: First, the cross-domain correlation strength is jointly determined by the cross-domain correlation modeling output of the S1 shared fusion layer and the perturbation amplitude of the connection in the S2 probability structure diagram, reflecting the degree of coupling between different nodes in dimensions such as signal propagation, interference transmission, and monitoring coverage. Second, the confidence level label, derived from the confidence level assigned to historical similar patterns in S2 time-matching retrieval, is used to characterize the predictive reliability of the association.

[0063] The edge weights are normalized and fused, that is, the association strength and confidence are first linearly mapped to a unified score range, and then weighted and summed according to the learnable proportional coefficient to ensure that the weights can reflect the true strength of physical coupling while taking into account the credibility of historical experience.

[0064] An anomalous situation element E (node ​​type: anomalous situation element) was detected in a coastal area, with a dominant frequency range of 2.4–2.45. GHz, spatial coverage radius 5 km, peak energy 1.8 W, currently labeled as an abnormal warning zone, has a transition probability of 0.216 from this cell to a severe interference / failure risk zone in the corrected probability matrix. This feature was captured by a type D monitoring device M (node ​​type: monitoring device), with deployment coordinates (lat=30.25°, lon=120.15°) and a sensitivity parameter of -90. dBm, historical detection success rate 0.93. The spatial location of the anomaly, node L (node ​​type: spatial location), is grid index G(3025, 12015), and the mean baseline energy of the background electromagnetic environment is 0.4. W. The affected service node B (node ​​type: service impact) is for maritime emergency communications, with high priority, and requires the frequency band 2.4–2.48. GHz, power tolerance limit 2.0 W, historically affected 7 times.

[0065] Edges such as E–M, M–L, and L–B are established in the graph, and the edge weights integrate two parts of information: First, the cross-domain correlation strength C is calculated by combining the cross-domain correlation modeling weights output by the S1 shared fusion layer (e.g., E–M signal propagation coupling degree 0.82) and the perturbation amplitude of the corresponding connection line in the S2 probability structure diagram (e.g., M–L perturbation due to environmental noise rise 0.035). Here, the multiplication and normalization method is used to obtain C(E–M)=0.82, C(M–L)=0.78, and C(L–B)=0.65. Second, the confidence label Conf is derived from the historical similarity pattern confidence of S2 time-matching retrieval (E–M matching historical interference event confidence is 0.91, M–L is 0.87, and L–B is 0.83). The edge weights are normalized and fused: first, C and Conf are linearly mapped to the [0,1] interval (in this case, it's already in the 0–1 range), then weighted and summed using the learnable proportional coefficients θ1=0.6 (association strength weight) and θ2=0.4 (confidence weight), resulting in the formula W=θ1·C+θ2·Conf. Substituting: E–M: W(E–M)=0.6×0.82+0.4×0.91=0.492+0.364=0.856; M–L: W(M–L)=0.6×0.78+0.4×0.87=0.468+0.348=0.816; L–B: W(L–B)=0.6×0.65+0.4×0.83=0.390+0.332=0.722.

[0066] This results in a set of edge weights that reflect both the physical coupling strength of signal propagation and interference transmission, and the reliability of historical experience. These weights are then assigned to the corresponding edges, forming a knowledge graph that is rich in both semantics (node ​​attributes include four-domain features, situational status, and business constraints) and topology (cross-layer associations and weights). This graph not only depicts the detection source and spatial location of abnormal situational elements, but also reveals the path and strength of their transmission to key businesses through monitoring equipment and spatial location, providing an interpretable and structured basis for subsequent situational visualization and risk assessment.

[0067] Specifically, real-time environmental parameters and interference characteristics are introduced into the cross-attention calculation stage to achieve deep interaction between situational attributes and the environment. Real-time environmental parameters include background noise levels (which can be measured in real time by fixed monitoring stations and mobile platforms and quantiles can be calculated), meteorological propagation influencing factors (such as atmospheric humidity and precipitation rate correction coefficients for shortwave propagation loss), ionospheric disturbance index, etc.; environmental interference characteristics cover the number of sudden non-cooperative signal sources, known interference pattern matching scores, and power density distribution of interference sources in space.

[0068] Cross-attention design is a dual-path query-key-value matching structure: One approach uses situational attribute features as query vectors, and real-time environmental parameters and interference features form a set of key-value pairs. By calculating the similarity score between the query and each key (which can be achieved by using an inner product followed by a normalized soft maximization operation), the attention distribution of environmental information on each dimension of the current situation can be obtained. The other approach uses environmental features as query vectors and situational attribute features as key-value pairs to capture the feedback influence of situational state on environmental parameters.

[0069] The outputs of the two attention channels are concatenated along the feature dimension and then nonlinearly integrated via a feedforward network to generate a feature interaction vector. This process ensures that environmental information is not only passively injected into the situational representation, but the situation itself also actively responds to environmental changes, forming a two-way coupled interactive representation.

[0070] The situational attribute feature vector of a certain spatiotemporal frequency energy unit is F_s (dimension 512). The real-time environmental parameter vector E_p includes background noise level 0.72 (normalized value, 0 represents extremely low noise, 1 represents extremely high noise), meteorological propagation influence factor 0.58 (reflecting the correction of shortwave loss by humidity and precipitation), and ionospheric disturbance index 0.41, synthesized as E_p=[0.72,0.58,0.41] (extendable to high-dimensional embedding); the environmental interference feature vector E_d includes the number of sudden non-cooperative signal sources 8 (normalized 0.8), known interference pattern matching score 0.76, and mean power density distribution of spatial interference sources 0.65, synthesized as E_d=[0.8,0.76,0.65]. Concatenating E_p and E_d and mapping them to the same dimension 512 as F_s, we obtain the environmental key-value matrix K_env=V_env (dimension 512).

[0071] First Route (Situation → Environment): Using F_s as the query vector Q_s, the inner product similarity S1(i) = Q_s·K_env(i) with K_env is calculated. Soft maximization normalization is performed on all i to obtain the attention weight α1(i) = exp(S1(i)) / Σ_jexp(S1(j)). Calculation shows that α1 has the highest value in the noise dimension (0.45), followed by the interference pattern dimension (0.35), and the lowest value in the propagation factor dimension (0.20). The output feature O1 = Σ_iα1(i)·V_env(i) represents the differentiated impact of environmental information on each dimension of the current situation.

[0072] Second approach (Environment → Situation): The query vector Q_e (dimension 512) is obtained by linearly mapping the concatenated vector E_all=[E_p;E_d] of environmental features. The key-value matrix K_s=V_s is constructed using F_s, and the inner product S2(i)=Q_e·K_s(i) is calculated. Soft maximization yields α2(i)=exp(S2(i)) / Σ_jexp(S2(j)). α2 has the largest weight in the frequency domain feature dimension of the situation (0.50), followed by the energy domain (0.30), and the smallest weight in the time domain (0.20). The output feature O2=Σ_iα2(i)·V_s(i) reflects the feedback response of the situation state to environmental parameters.

[0073] Concatenate the two outputs along the feature dimension: [O1;O2] (dimension 1024) is nonlinearly integrated by a feedforward network (two fully connected layers + ReLU) to obtain the feature interaction vector F_int (dimension 512). For example, after the network merges the value of O1 in the noise response channel (0.63) and the value of O2 in the frequency domain feedback channel (0.58), the dimension of F_int may increase to 0.71, indicating that when the background noise is high and the situation frequency domain features are significant, the interaction strengthens the contribution of this dimension to subsequent predictions.

[0074] Specifically, the process proceeds to the multi-source dynamic weight learning and weighted fusion stage, aiming to analyze the contribution of each feature to the situation prediction and visualization tasks, and optimize the feature fusion strategy accordingly. The processing logic is as follows: In multiple training samples, the situation prediction error (such as the difference between the probability distribution of future states and the true distribution) and the visualization clarity index (such as the cluster compactness of graph nodes and the identification rate of important edges) are used as supervision signals to construct a weighted learning model.

[0075] The model takes each sub-feature (time domain component, frequency domain component, spatial component, energy component, environmental noise component, interference feature component, etc.) in the feature interaction vector as input, and outputs the contribution score of each sub-feature through a trainable scoring network. The scoring network can use a gating mechanism to adjust the importance of sub-features using independent gating vectors for different tasks (prediction vs. visualization), thereby achieving task-oriented dynamic weight allocation.

[0076] After obtaining the contribution score, it is normalized to form a multi-source dynamic weight, and then the feature interaction vector is weighted and fused according to the weight to generate multi-source coordinated features. The key to this step is that the weight is not statically preset, but adaptively adjusted with changes in environment and situation, ensuring that in a complex and ever-changing electromagnetic environment, the map update can focus on the most informative feature dimensions.

[0077] The feature interaction vector F_int is obtained through the cross-attention process in the previous stage. It has a dimension of 512 and is semantically divided into 6 sub-feature components: The dataset includes a time-domain component f_t, a frequency-domain component f_f, a spatial component f_s, an energy component f_e, an environmental noise component f_n, and a disturbance feature component f_d. Each sample set contains the current F_int and two supervisory signals, along with the situation prediction error Err_pred (measured by KL divergence, etc., to reflect the difference between the future state probability distribution and the true distribution; a smaller value is better) and a visualization clarity index Clarity_viz (e.g., a weighted average of node cluster compactness and important edge identification rate; a larger value is better). During training or online learning, a weighted learning model is constructed. The input consists of the sub-feature vectors of F_int (each with a dimension of 512 / 6≈85). These are passed through a trainable scoring network (fully connected layer + ReLU) to obtain the original contribution scores s_t, s_f, s_s, s_e, s_n, s_d. To introduce task orientation, a gating mechanism is used. Let the prediction task gating vector be g_pred=[0.9,0.7,0.5,0.8,0.6,0.4], and the visualization task gating vector be g_viz=[0.5,0.6,0.9,0.7,0.8,0.7], corresponding to the adjustment coefficients of each sub-feature. The task-oriented contribution is calculated as: s_task(i)=s_i×g_task(i), where task∈{pred,viz}. The scoring network outputs s_t=0.80, s_f=0.65, s_s=0.70, s_e=0.75, s_n=0.55, s_d=0.60. Therefore, for the prediction task: s_pred(t)=0.80×0.9=0.72; s_pred(f)=0.65×0.7=0.455; s_pred(s)=0.70×0.5=0.35; s_pred(e)=0.75×0.8=0.60; s_pred(n)=0.55×0.6=0.33; s_pred(d)=0.60×0.4=0.24. Normalization yields the multi-source dynamic weights for the prediction task. w_pred(i)=s_pred(i) / Σ_js_pred(j), denominator Σ=0.72+0.455+0.35+0.60+0.33+0.24=2.695, w_pred(t)≈0 .267, w_pred(f)≈0.169, w_pred(s)≈0.130, w_pred(e)≈0.223, w_pred(n)≈0.122, w_pred(d)≈0.089. The sub-features of F_int are weighted and fused according to this weight to generate a multi-source coordinated feature F_coord_pred = Σ_iw_pred(i)·f_i, which is guided by the prediction task. For example, in the time domain, if the mean of f_t is 0.62, the contribution is 0.267×0.62≈0.166; in the frequency domain, if the mean of f_f is 0.58, the contribution is 0.169×0.58≈0.098, and so on. The new coordinated feature vector is obtained by accumulating the features of each dimension. Similarly, for the visualization task, g_viz is used to calculate s_viz(i) and the normalized weight w_viz(i) to obtain F_coord_viz.

[0078] The weights in this process are dynamically adjusted by the supervision signal and task gating, rather than being fixed presets. For example, when heavy rain causes a significant increase in environmental noise, the product of g_pred(n) and s_n increases, and w_pred(n) increases accordingly, making the fused features pay more attention to the noise component. This ensures that the map update always focuses on the feature dimension with the most information in complex electromagnetic environments, thereby improving prediction accuracy and the semantic recognition of visualization.

[0079] Specifically, affine transformations are performed in conjunction with real-time environmental parameters to dynamically correct the geometric and statistical properties of multi-source coordinated features, generating the final features used for map updates. Here, the affine transformation is not a simple linear scaling and translation, but a composite transformation constructed based on environmental parameters to handle feature distribution drift and scale changes caused by the environment. Based on real-time environmental parameters, affine transformation matrices and bias vectors are constructed. For example, an increase in background noise level may correspond to an overall compression transformation of characteristic amplitude (reducing the effective signal weight under noise masking). Meteorological propagation influencing factors can be transformed into local stretching or rotation parameters of spatial coordinates (reflecting spatial feature reconstruction caused by propagation path distortion). Interference characteristic statistics can be mapped to nonlinear translation of energy domain components.

[0080] The multi-source coordinated features are treated as a set of points in a high-dimensional space. By applying this affine transformation, the geometric structure can be adapted (such as the relative position correction of nodes in the spatial grid) and the statistical distribution can be aligned (such as making the feature mean and variance regress to a reasonable range that matches the environment).

[0081] After the transformation, the features are subjected to validity checks (such as value range constraints and normalization) to obtain the final multi-source coordinated features. These features have the ability to represent the real-time state of the environment and can be directly written into the attribute fields of the corresponding nodes in the multi-level knowledge graph, updating the edge weights, thereby completing the dynamic evolution of the graph.

[0082] The multi-source coordinated feature vector for the prediction task is F_coord (dimension 512), and its instantiation values ​​are arranged according to semantic sub-features as follows: mean of time domain component 0.62, mean of frequency domain component 0.58, mean of spatial component 0.55, mean of energy component 0.70, mean of environmental noise component 0.48, and mean of interference feature component 0.51. Real-time environmental parameters are: background noise level N = 0.80 (normalized, higher values ​​indicate stronger noise), meteorological propagation influence factor M = 0.65 (reflecting shortwave propagation path distortion caused by humidity and precipitation), and interference feature statistics I = 0.75 (e.g., high mean power density of interference sources). A composite affine transformation is constructed based on the environmental parameters: An increase in background noise N corresponds to overall compression of feature amplitudes. A scaling matrix S is constructed as a diagonal matrix, with diagonal elements s_i = 1 - k·N, where k is the compression coefficient (taken as 0.3). Therefore, s_i ≈ 1 - 0.3 × 0.80 = 0.76, meaning all feature dimensions are multiplied by 0.76 to reduce the weights masked by noise. The meteorological propagation influence factor M is transformed into local stretching and rotation parameters of the spatial components. Construct a spatial transformation matrix T_s. In the two-dimensional spatial projection, apply a stretching coefficient (1+M)=1.65 to the spatial components in the principal direction and a rotation angle θ=arcsin(M-0.5)=arcsin(0.15)≈8.62°. Use the two-dimensional rotation matrix R(θ) and the stretching matrix together to apply to the spatial component coordinates to reconstruct the spatial features caused by the propagation path distortion. The interference characteristic statistics I are mapped to the nonlinear translation of the energy domain components, b_e=τ·tanh(I), where τ=0.2, tanh(0.75)≈0.635, resulting in b_e≈0.127, that is, add a constant bias to the energy components to reflect the energy baseline drift caused by the interference. Combine the above transformations into an affine transformation matrix A and a bias vector b: Except for the spatial components, A is a diagonal matrix with scalar multipliers of 0.76. The spatial components are replaced with corresponding blocks of T_s, and b is 0.127 at the energy component position and 0 elsewhere. Treating F_coord as a high-dimensional point set, we execute F_affine = A·F_coord + b: Time-domain component: 0.62 × 0.76 ≈ 0.471; Frequency domain component: 0.58 × 0.76 ≈ 0.441; After being subjected to T_s, the coordinates of the spatial component change from (0.55, 0.55) to the new coordinates after stretching and rotation (≈0.79, ≈0.61); Energy component: 0.70 × 0.76 + 0.127 ≈ 0.659; Environmental noise component: 0.48 × 0.76 ≈ 0.365; Interference characteristic components: 0.51 × 0.76 ≈ 0.388; Ensure all components fall within the preset value range [0,1], and perform normalization on the entire vector to match the mean and variance with the environmental state. For example, the energy component may be adjusted to 0.68 after normalization to reflect a reasonable energy distribution under the current high interference background. Finally, the final multi-source coordinated feature F_final is obtained. Its geometric structure (such as the relative position of spatial nodes) has been adapted to the propagation path distortion, and its statistical distribution (mean / variance) has been aligned with the real-time environment. It can be directly written into the four-domain feature attribute fields of the corresponding nodes in the multi-level knowledge graph, and the edge weights are updated based on the new feature similarity.

[0083] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0084] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for constructing and extrapolating an electromagnetic situation knowledge graph based on multi-source spectrum data, characterized by: The deduction method includes the following steps: S1: Collect multi-source heterogeneous spectrum data, map the multi-source heterogeneous spectrum data to the feature space through the feature coding network, generate original discrimination features, input the original discrimination features into the shared fusion layer for cross-domain correlation modeling, obtain situation attribute features, and divide the real-time state of each spatiotemporal frequency energy unit into different state spaces according to the discrete situation state classification mechanism. S2: Input the situation attribute features into the preset adaptive spatiotemporal feature library, find the matching feature vector similar to the current situation through time matching retrieval, apply the situation attribute features and matching feature vector together to the adaptive spatiotemporal feature library for correction, generate correction features, initialize the situation state transition probability matrix based on the state space, and construct a probability structure graph, traverse each connection in the probability structure graph, calculate the disturbance amplitude based on the correlation response strength and node risk factor, iteratively calculate the probability deformation through nonlinear interaction between nodes until the convergence condition is met, and generate the corrected probability matrix and prediction results; S3: Construct a multi-level knowledge graph based on the modified probability matrix and correction features.

2. The method for constructing and inferring an electromagnetic situation knowledge graph based on multi-source spectrum data according to claim 1, characterized in that: In step S2, each connection in the probability structure graph is traversed, and the perturbation amplitude is calculated based on the correlation response strength and node risk factor. The probability deformation is iteratively calculated through the nonlinear interaction between nodes until the convergence condition is met, generating the corrected probability matrix and prediction results. This includes the following steps: Traverse each connection in the probability structure graph and calculate the perturbation amplitude based on the correlation response strength and node risk factor; The probabilistic deformation is calculated iteratively through the nonlinear interaction between nodes. In each iteration, the current probability value of each node is weighted and aggregated based on the perturbation amplitude of the connected nodes. During the iteration process, the connection weights and node risk factors are continuously updated with the latest calculated probability values. This process is repeated until the probability changes of all connections are less than the preset convergence threshold, generating a corrected probability matrix and the corresponding prediction results.

3. The method for constructing and inferring an electromagnetic situation knowledge graph based on multi-source spectrum data according to claim 2, characterized in that: In step S2: the situation attribute features are fed into a preset adaptive spatiotemporal feature library, and matching feature vectors similar to the current situation are found through time matching retrieval, including the following steps: The adaptive spatiotemporal feature library coarsely aligns the timestamp of the current feature vector with the time labels of historical patterns in the library, and filters out a set of candidate patterns that are close in time. In the candidate pattern set, the spatiotemporal feature vector of each historical pattern is subjected to a dimension-wise similarity measurement with the current feature vector, and the matching degree is comprehensively evaluated by combining the trajectory shape matching algorithm, and several matching feature vectors are output.

4. The method for constructing and inferring an electromagnetic situation knowledge graph based on multi-source spectrum data according to claim 3, characterized in that: In step S2: the situational attribute features and the matching feature vector are applied together to the adaptive spatiotemporal feature library for correction, generating corrected features, including the following steps: Each matching result is assigned a confidence level label. When the confidence level label is higher than the preset confidence level threshold, the feature library correction mechanism is triggered. The fused feature vectors are used to adjust the representation vectors of relevant historical patterns in the feature library, update the position and trajectory parameters of the representation vectors in the spatiotemporal feature space, and generate corrected features.

5. The method for constructing and inferring an electromagnetic situation knowledge graph based on multi-source spectrum data according to claim 4, characterized in that: Each matching result is assigned a confidence level label. When the confidence level label is higher than a preset confidence threshold, a feature library correction mechanism is triggered, including the following steps: The current situation attribute features and the matching feature vector are used together as correction inputs and applied to the online learning module of the feature library; An incremental feature fusion strategy is adopted to concatenate the current feature and the matching feature in a unified dimensional space; The contribution of features from different sources is dynamically adjusted through an attention weight allocation mechanism. If a certain type of environmental interference in the historical scene corresponding to the matching feature is consistent with the current environment, the attention weight of the matching feature is increased, and vice versa.

6. The method for constructing and inferring an electromagnetic situation knowledge graph based on multi-source spectrum data according to claim 2, characterized in that: In step S1: the original discriminative features are input into the shared fusion layer for cross-domain correlation modeling to obtain situational attribute features, including the following steps: For each spatiotemporal frequency energy unit, the feature correlation weights of the spatiotemporal frequency energy unit and other related units are calculated through a self-attention mechanism to identify the local contextual associations that affect the state of the spatiotemporal frequency energy unit. A graph neural network is introduced to construct a global association graph, where the nodes are the original discriminative features of all spatiotemporal frequency and energy units, and the edge weights are initialized by weights obtained from local association mining. By iteratively updating node features through multi-layer graph convolution, scattered cross-domain associations across the global scope are integrated into the node features; Pooling is performed on the node features output by the multi-layer graph convolution, and situational attribute features are generated by combining the semantic constraints of the business requirement text.

7. The method for constructing and inferring an electromagnetic situation knowledge graph based on multi-source spectrum data according to claim 1, characterized in that: In the multi-level knowledge graph, node attributes are embedded with time-frequency-space-energy four-domain features and situational states, and edge weights are fused with dynamic weights and confidence labels to reflect the cross-domain correlation strength and prediction reliability.

8. The method for constructing and inferring an electromagnetic situation knowledge graph based on multi-source spectrum data according to claim 7, characterized in that: The multi-level knowledge graph combines real-time environmental parameters and environmental interference features. The environmental parameters include background noise levels and meteorological propagation influencing factors. These parameters are incorporated into cross-attention calculations, enabling the interaction between situational attribute features and real-time environmental information to generate feature interaction vectors. By analyzing the contribution of each feature in situation prediction and visualization through a weight learning mechanism, multi-source dynamic weights are generated, and the feature interaction vectors are weighted and fused to obtain multi-source coordinated features. Affine transformation operations are performed in conjunction with real-time environmental parameters, and updates are performed using multi-source coordination features.

9. The method for constructing and inferring an electromagnetic situation knowledge graph based on multi-source spectrum data according to claim 8, characterized in that: The multi-level knowledge graph analyzes the contribution of each feature in situation prediction and visualization through a weight learning mechanism, generates multi-source dynamic weights, performs weighted fusion of feature interaction vectors to obtain multi-source coordinated features, performs affine transformation operations in conjunction with real-time environmental parameters, and updates the graph using the multi-source coordinated features, including the following steps: In multiple training samples, the situation prediction error and the visualization clarity index are used as supervision signals to construct a weighted learning model. The weighted learning model takes each sub-feature in the feature interaction vector as input and outputs the contribution score of each sub-feature through a trainable scoring network. The contribution scores are normalized to form multi-source dynamic weights, and the feature interaction vectors are weighted and fused according to the multi-source dynamic weights to generate multi-source coordinated features. Based on real-time environmental parameters, an affine transformation matrix and a bias vector are constructed. The multi-source coordination features are regarded as a set of points in a high-dimensional space. The affine transformation is applied to adapt the geometric structure and align the statistical distribution. After the transformation is completed, the final multi-source coordinated features are obtained, written into the attribute fields of the corresponding nodes of the multi-level knowledge graph, and the weights of the edges are updated to complete the dynamic evolution of the graph.

10. The method for constructing and inferring an electromagnetic situation knowledge graph based on multi-source spectrum data according to claim 1, characterized in that: The sources of the multi-source heterogeneous spectrum data include fixed monitoring stations, mobile detection platforms, space remote sensing receivers, and service requirement description texts; The state space includes a stable and controllable region, an anomaly warning region, and a severe interference / failure risk region.

Images