A simulation method and system for intelligent fusion and dynamic prediction of electronic warfare intelligence

Abstract

This application provides a simulation method and system for intelligent fusion and dynamic prediction of electronic warfare intelligence. This simulation method uses multi-domain sensors to collect electromagnetic signals in real time, adopts a dynamic adversarial network (DAN) for spectrum matching and anomaly location, and combines spatiotemporal verification to eliminate interference data. It constructs a semantic association graph of radar, communication, and geographic information, fuses time-domain and spatial-domain features through spatiotemporal attention, and dynamically allocates confidence to achieve multi-source data fusion. It predicts threat evolution trends based on a spatiotemporal transformer, outputs threat level, type, and intent, and generates a dynamic situation map in combination with a hybrid density network. It adopts an online incremental learning framework to adapt to new signal patterns, prioritizes updating the prediction module, and dynamically adjusts push priorities through edge-cloud collaborative optimization of a reinforcement learning-driven multi-path distribution mechanism to achieve low-latency transmission of key threat intelligence.

Description

Technical Field

[0001] The present invention relates to the field of military communication and simulation technology, and in particular to a simulation method and system for intelligent fusion and dynamic prediction of electronic warfare intelligence. Background Art

[0002] Current electronic warfare intelligence processing faces core challenges such as low efficiency in fusion of multi-source heterogeneous data, poor adaptability to dynamic environments, and insufficient real-time threat prediction. Traditional methods typically use static weight allocation strategies in multi-domain sensor data fusion, which makes it difficult to cope with the dynamic changes of signals in complex electromagnetic environments. For example, conventional K-means clustering algorithms are susceptible to noise interference when processing high-dimensional electromagnetic signal features, resulting in reduced sorting accuracy, and lack the ability to quickly adapt to new signal patterns. In addition, existing dynamic prediction models are mostly based on single-modal data, such as radar pulses or communication protocols, and do not fully exploit the synergistic correlation of time-domain and space-domain features, resulting in limited comprehensiveness and accuracy of battlefield situation awareness.

[0003] In terms of technological evolution, artificial intelligence has been gradually applied to electronic warfare in recent years. For example, the US military has achieved rapid identification of unknown radar signals and generation of jamming strategies through cognitive electronic warfare systems such as the F-16 / CADS. However, these algorithms rely on a fixed rule base and are difficult to adapt to rapidly evolving modulation types and signal distributions.

[0004] While multimodal knowledge graph technology has made progress in civilian applications such as semantic search, it lacks the ability to model the dynamic associations between radar signal features and geospatial information in military scenarios. Furthermore, existing incremental learning frameworks often rely on global parameter updates, which can easily lead to catastrophic forgetting when adapting to new signals. Summary of the Invention

[0005] In order to solve the above problems, the present invention proposes a simulation method and system for intelligent fusion and dynamic prediction of electronic warfare intelligence, which is used to improve the real-time, adaptability and battlefield survivability of the electronic warfare intelligence system.

[0006] The present invention specifically provides the following technical solutions:

[0007] A simulation method for intelligent fusion and dynamic prediction of electronic warfare intelligence, characterized by comprising:

[0008] Electromagnetic signal data is collected in real time through multi-domain distributed sensor nodes. Dynamic adversarial network (DAN) is used to perform spectrum fingerprint matching and anomaly location on the electromagnetic signal data. In addition, a spatiotemporal consistency verification mechanism is used to eliminate invalid data and generate a standardized intelligence stream.

[0009] Construct a multimodal knowledge graph, define radar signal features as entities, communication protocol metadata as relationships between entities, and geospatial information as entity attributes, perform semantic association modeling, and generate heterogeneous network representations.

[0010] Based on the anomaly network representation, a spatiotemporal dual-channel attention mechanism is used to fuse the temporal characteristics of radar pulse sequences with the spatial distribution characteristics of communication signal groups to generate a joint semantic vector. A dynamic confidence allocation algorithm is used to achieve the fusion of multi-source heterogeneous data. Based on the joint semantic vector, a spatiotemporal Transformer model is used to dynamically predict threat evolution trends and output multi-task prediction results of threat level, target type, and behavioral intention. A hybrid density network is then combined to generate a dynamic battlefield situation panorama.

[0011] Deploy an online incremental learning framework to detect new signal patterns and fine-tune model parameters in layers, updating the prediction head module to adapt to the dynamic battlefield environment;

[0012] A reinforcement learning-driven multi-path distribution mechanism is implemented through the edge-cloud collaborative architecture, and the intelligence push priority is dynamically optimized based on the channel state matrix and node security assessment to ensure low-latency transmission of critical threat intelligence.

[0013] Optionally, a dynamic adversarial network (DAN) is used to perform spectrum fingerprint matching and anomaly location on the electromagnetic signal data, and a spatiotemporal consistency verification mechanism is combined to eliminate invalid data and generate a standardized intelligence stream, including:

[0014] In the dynamic adversarial network (DAN), an improved spectral clustering algorithm is used and a quantum annealing optimization mechanism is introduced to perform feature sorting and spectrum fingerprint matching on the electromagnetic signal data; the feature sorting is to separate the mixed electromagnetic signal into different groups according to carrier frequency, pulse width, and modulation type; the output result of the feature sorting is used as the input of the spectrum fingerprint matching, and the cluster center vector is dynamically updated to perform similarity comparison with the radiation source characteristics in the fingerprint library to identify abnormal signals and generate standardized intelligence streams; the loss of the spectrum fingerprint matching is Designed to:

[0015]

[0016] in, is the three-dimensional feature vector of the electromagnetic signal sample, is the i-th signal sample, For the The cluster center vector at the iteration, N is the total number of valid signal samples participating in the cluster center update in the current iteration, is the quantum tunneling intensity coefficient; is the adaptive cutoff threshold, To maximize the difference between classes Minimizing intra-class variance The comprehensive measure of is the quantum confinement factor;

[0017] Cluster Center The formula for dynamic update is:

[0018]

[0019] in, is the learning rate, Gaussian kernel bandwidth The specific adjustments are:

[0020]

[0021] in, is the distribution shift sensitivity coefficient, is the KL divergence measure of signal distribution changes, For the At the iteration The joint probability distribution of For the The cluster center vector at the iteration.

[0022] Optionally, a multimodal knowledge graph is constructed, radar signal features are defined as entities, communication protocol metadata as relationships between entities, and geospatial information as entity attributes. Semantic association modeling is performed and a heterogeneous network representation is generated, including:

[0023] In the construction of the multimodal knowledge graph, a multidimensional feature tensor fusion mechanism is added to define the correlation function between the radar signal modulation type and the communication protocol metadata. for:

[0024]

[0025] in, is the total number of modulation types, For the The modulation type encoding matrix of the class modulation signal converts the discrete modulation type Mapped to a three-dimensional vector, For the Standard carrier frequency mean value of the modulated signal, used to dynamically adjust the carrier frequency The weight of For tensor product operation, fusion carrier frequency , pulse width and time difference ; is a parameter matrix used to adjust the pulse width and time difference The concatenated vector [ ; ] for feature mapping;

[0026] The geographic space information is encoded by using spherical harmonic function :

[0027]

[0028] in, is the geographic spatial weight coefficient, are spherical harmonic basis functions, The latitude and longitude coordinates of the signal sources respectively; 、 、 angular quantum number, magnetic quantum number and maximum harmonic order respectively;

[0029] The tensor representation generated by the association function serves as the relationship vector between entities in the multimodal knowledge graph, and the spherical harmonic encoding result serves as the entity attribute, which together constitute the heterogeneous network representation containing topological connections and semantic constraints.

[0030] Optionally, a spatiotemporal dual-channel attention mechanism is used to fuse the time domain features of the radar pulse sequence with the spatial distribution features of the communication signal group to generate a joint semantic vector, including:

[0031] In the spatiotemporal dual-channel attention mechanism, the temporal channel Introducing the carrier frequency-pulse width coupled convolution kernel:

[0032]

[0033] in, is the carrier frequency difference sensitivity coefficient, is the time domain sliding window length, and are the tth moment and the history The carrier frequency value at the time, For history The pulse width value at the moment, is a function for extracting the periodic characteristics of pulse width;

[0034] The spatial channel uses pulse width weighted geodesic distance :

[0035]

[0036] in, is the pulse width difference penalty factor, are the feature vectors of nodes i and j respectively, is the eigenvector angle;

[0037] The dynamic confidence allocation algorithm defines the modulation type entropy weight as follows:

[0038]

[0039] in, For the The information entropy of the signal modulation type, C is the total number of signal categories, is the loop variable;

[0040] The time domain features extracted by the carrier frequency-pulse width coupled convolution kernel and the spatial domain features of the pulse width weighted geodesic distance modeling are dynamically fused through the modulation type entropy weight to generate the joint semantic vector, which is input into the spatiotemporal Transformer model for threat prediction.

[0041] Optionally, in the spatiotemporal Transformer model, a carrier frequency change rate gating mechanism is embedded in the threat prediction module. :

[0042]

[0043] in, is the carrier frequency change rate, is the pulse width stability index, is the linear transformation matrix;

[0044] The conditional probability distribution of the mixture density network Modeled as:

[0045]

[0046] in, is the weight of the kth Gaussian component under the input x, and is the carrier frequency mean and standard deviation of the kth Gaussian component, and Beta distribution describes the statistical characteristics of the pulse width. and are the shape parameters that control the left-biased and right-biased characteristics of the pulse width distribution, and the Cat distribution represents the probability of modulation type classification. is the modulation type probability vector;

[0047] The carrier frequency change rate gating mechanism dynamically adjusts the attention weight distribution of each time step in the spatiotemporal Transformer model based on the generated joint semantic vector, wherein the attention weight of the historical pulse width feature is enhanced when the carrier frequency change rate exceeds a threshold; the conditional probability distribution modeling of the hybrid density network integrates the geographic space coding features and the signal carrier frequency-pulse width joint distribution to generate a dynamic battlefield situation panorama including the threat thermal field and signal propagation path.

[0048] Optionally, in the online incremental learning framework, detecting the new signal pattern includes:

[0049] Adopting modulation type entropy change trigger mechanism Detecting the new signal pattern, the modulation type entropy change trigger mechanism Specifically:

[0050]

[0051] in, (.) is the indicator function, when the modulation type distribution entropy value of the new signal batch and historical distribution entropy The difference exceeds the threshold When , the update mechanism of the incremental learning model is triggered;

[0052] Fine-tune model parameters layer by layer, including:

[0053] Layered fine-tuning applies quantum noise perturbations to the carrier frequency feature layer, specifically:

[0054]

[0055] in, is the weight matrix of the frequency feature layer, is the average offset of the signal carrier frequency, is the quantum noise perturbation intensity coefficient.

[0056] In the edge-cloud collaborative architecture, the multi-path distribution mechanism is used to define the carrier frequency-bit error rate joint reward function

[0057]

[0058] in, 、 、 is the weighting coefficient, The carrier frequency f corresponds to the inverse of the channel bit error rate, is the total transmission time of the signal from the transmitter to the receiver, is the delay sensitivity factor, Provides a security evaluation vector for the node.

[0059] Optionally, based on the quantum annealing optimization, a pulse width modulation type constrained Hamiltonian is added to suppress the interference of carrier frequency drift and pulse width jitter on signal sorting. The Hamiltonian Combining the carrier frequency cluster center constraint and the pulse width distribution prior knowledge, the Hamiltonian The expression is:

[0060]

[0061] in, and are the qubit index and the total number of qubits, respectively. and The constraint coefficients of carrier frequency and pulse width respectively, and Respectively The carrier frequency value of quantum bits and the target carrier frequency value, is the reference pulse width distribution, For the The Pauli-Z operator for 1 qubit.

[0062] Optionally, in the dynamic adversarial network DAN, a carrier frequency-pulse width joint robustness loss is defined in the adversarial defense module. :

[0063]

[0064] in, To find the expectation of the statistical distribution of the input signal x, and are the disturbances for carrier frequency f and pulse width pw, respectively. and are the disturbance amplitudes that limit the carrier frequency and pulse width, To measure the distribution difference between the electromagnetic signal data x and the adversarial sample x+δ in the feature space f(.), is the modulation type classification loss weight, is the cross entropy loss, which quantifies the difference between the modulation type prediction distribution m(.) of the electromagnetic signal data x and the adversarial sample xadv, is a feature extraction function that maps the input signal x to a high-dimensional feature space. is a modulation type classifier that outputs the modulation type probability distribution of the signal x.

[0065] The present invention also provides a simulation system for intelligent fusion and dynamic prediction of electronic warfare intelligence, the simulation system comprising:

[0066] The multi-domain data acquisition and preprocessing module is used to collect electromagnetic signal data in real time through multi-domain distributed sensor nodes. It uses the dynamic adversarial network (DAN) to perform spectrum fingerprint matching and anomaly location on the electromagnetic signal data, and combines the spatiotemporal consistency verification mechanism to eliminate invalid data and generate a standardized intelligence stream.

[0067] A multimodal knowledge graph construction module is used to construct a multimodal knowledge graph, defining radar signal features as entities, communication protocol metadata as relationships between entities, and geospatial information as entity attributes, to perform semantic association modeling and generate heterogeneous network representations;

[0068] The spatiotemporal feature fusion engine uses a spatiotemporal dual-channel attention mechanism to fuse the temporal features of radar pulse sequences with the spatial distribution features of communication signal groups to generate a joint semantic vector. It also uses a dynamic confidence allocation algorithm to achieve the fusion of multi-source heterogeneous data.

[0069] A threat prediction and situation generation module is used to dynamically predict threat evolution trends based on the joint semantic vector using a spatiotemporal Transformer model, output multi-task prediction results of threat level, target type, and behavioral intention; and generate a dynamic battlefield situation panorama in combination with a hybrid density network;

[0070] An online incremental learning agent module, which deploys an online incremental learning framework to detect new signal patterns and fine-tune model parameters hierarchically, updating the prediction head module to adapt to the dynamic battlefield environment;

[0071] The edge-cloud collaborative distribution control module is used to implement a reinforcement learning-driven multi-path distribution mechanism through the edge-cloud collaborative architecture, dynamically optimize the intelligence push priority based on the channel state matrix and node security assessment, and ensure low-latency transmission of critical threat intelligence.

[0072] The present invention has the following beneficial technical effects: The present invention provides a simulation method and system for intelligent fusion and dynamic prediction of electronic warfare intelligence. By introducing quantum annealing optimization mechanism through dynamic adversarial network DAN to improve spectrum clustering algorithm, high robustness signal sorting is achieved in carrier frequency, pulse width and modulation type feature dimensions, and quantum constraint factor and adaptive truncation threshold are combined to suppress feature drift and enhance the anti-interference capability of spectrum fingerprint matching; a heterogeneous network representation is constructed that integrates radar signals, communication protocols and geographic space information, and a multi-dimensional feature tensor fusion mechanism and spherical harmonic function coding technology are used to achieve cross-domain entity dynamic association modeling, support accurate signal source positioning and threat intention reasoning; based on the spatiotemporal dual-channel attention mechanism, the carrier frequency-pulse width coupling convolution kernel is used to extract the time domain frequency hopping law, and the pulse width weighted geodesic distance is combined to analyze the spatial distribution characteristics, and the use of the placement The dynamic allocation algorithm of credibility realizes the differentiated weight fusion of multi-source heterogeneous data and improves the accuracy of threat identification. It adopts the modulation type entropy change threshold to trigger the model update mechanism, prioritizes fine-tuning the prediction head module and applies quantum noise perturbation, quickly adapts to new signal patterns, and reduces the risk of forgetting historical knowledge. The multi-path distribution mechanism driven by reinforcement learning combines the channel state matrix and node security assessment to dynamically optimize the transmission path, and improves the channel utilization and anti-interference capability through the carrier frequency-bit error rate joint reward function. The pulse width-modulation type constrained Hamiltonian and the carrier frequency-pulse width joint robustness loss function are introduced to construct a feature space energy minimization mechanism at the quantum bit level, suppress the interference of adversarial samples and improve the efficiency of abnormal response. The present invention comprehensively solves the problems of low efficiency of multi-source data fusion, poor dynamic adaptability and lag in threat prediction of electronic warfare systems in complex electromagnetic environments, significantly improving battlefield survivability and intelligent decision-making support efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0074] Figure 1 The present invention provides a flowchart of a simulation method for intelligent fusion and dynamic prediction of electronic warfare intelligence.

[0075] Figure 2 A framework diagram of the dynamic adversarial network (DAN) model provided in an embodiment of the present invention.

[0076] Figure 3 A block diagram of a simulation system for intelligent fusion and dynamic prediction of electronic warfare intelligence provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0077] The embodiments of the present application are described in detail below with reference to the accompanying drawings.

[0078] The following describes the embodiments of the present application through marked specific examples, and those skilled in the art can easily understand other advantages and effects of the present application from the contents disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. The present application can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present application. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without making creative work are within the scope of protection of this application.

[0079] It should be noted that various aspects of the embodiments within the scope of the appended claims are described below. It should be apparent that the aspects described herein can be embodied in a wide variety of forms, and any structure and / or function described herein is illustrative only. Based on this application, one skilled in the art will appreciate that an aspect described herein can be implemented independently of any other aspect, and that two or more of these aspects can be combined in various ways. For example, any number and aspect described herein can be used to implement an apparatus and / or practice a method.

[0080] It should also be noted that the diagrams provided in the following embodiments are only used to schematically illustrate the basic concept of the present application.

[0081] Additionally, in the following description, specific details are provided to facilitate a thorough understanding of the examples, however, one skilled in the art will understand that the examples can be practiced without these marked details.

[0082] The purpose of the present invention is to provide a simulation method and system for intelligent fusion and dynamic prediction of electronic warfare intelligence, aiming to improve battlefield survivability and intelligent decision-making support.

[0083] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0084] Reference Figure 1 , shows a simulation method for intelligent fusion and dynamic prediction of electronic warfare intelligence according to an embodiment of the present application, the simulation method comprising:

[0085] Electromagnetic signal data is collected in real time through multi-domain distributed sensor nodes, and the dynamic adversarial network (DAN) is used to perform spectrum fingerprint matching and anomaly positioning on the electromagnetic signal data. In addition, the spatiotemporal consistency verification mechanism is combined to eliminate invalid data and generate a standardized intelligence stream.

[0086] Specifically, sensor nodes distributed across land, sea, and air, such as airborne radars and shipborne communication detectors, are deployed to collect raw electromagnetic signal data in real time. In a dynamic adversarial network (DAN), an improved spectral clustering algorithm is used, along with a quantum annealing optimization mechanism, to perform feature sorting on the electromagnetic signal data. Feature dimensions include carrier frequency, pulse width, and modulation type. Mixed electromagnetic signals are separated into different groups based on carrier frequency, pulse width, and modulation type. The output of feature sorting serves as the input for spectral fingerprint matching. By dynamically updating the cluster center vector and performing a similarity comparison with the radiation source features in the fingerprint library, abnormal signals are identified and a standardized intelligence stream is generated. The loss of spectral fingerprint matching is designed to be:

[0087]

[0088] in, is the three-dimensional feature vector of the electromagnetic signal sample, For the The cluster center vector at the iteration, N is the total number of valid signal samples participating in the cluster center update in the current iteration, is the quantum tunneling intensity coefficient, which is 0.8; is the adaptive cutoff threshold, To maximize the difference between classes Minimizing intra-class variance The comprehensive measure of is the quantum confinement factor, Controlling the influence of quantum tunneling effect on feature sorting, smaller It can suppress excessive jumps of quantum bits and balance global exploration and local convergence in complex electromagnetic environments.

[0089] Among them, two core functions are realized simultaneously through the improved spectral clustering algorithm and quantum annealing optimization mechanism. The first is feature sorting, which is to separate the mixed electromagnetic signals into different groups according to carrier frequency, pulse width, and modulation type; the second is spectrum fingerprint matching: through the loss function design, the sorting results are dynamically matched with the preset fingerprint library, and the matching accuracy is optimized.

[0090] Cluster center movement The formula for state update is:

[0091]

[0092] in, is the learning rate, which is 0.002; Gaussian kernel bandwidth The specific adjustments are:

[0093]

[0094] in, is the distribution shift sensitivity coefficient, is the KL divergence measure of signal distribution changes, For the At the iteration The joint probability distribution of For the The cluster center vector at the iteration.

[0095] In the test experiment, the quantum confinement factor was introduced in the carrier frequency dimension = 0.1, by adaptive truncation threshold =0.05 dynamically eliminates noise interference, improving signal sorting accuracy by 12% compared to traditional methods. In the spatiotemporal consistency verification mechanism, the signal propagation speed error is set to ≤5% by verifying the physical constraint relationship between the pulse arrival time difference and the geographic coordinates, thus eliminating invalid data.

[0096] The output of this feature sorting serves as input for spectral fingerprint matching. By dynamically updating the cluster center vectors and comparing them with the radiation source features in the fingerprint library, anomalous signals are identified and a standardized intelligence stream is generated. Specifically, the sorted signal groups are matched against a preset radiation source fingerprint library for similarity. The matching confidence level is calculated based on the dynamically updated cluster center vectors. When the confidence level falls below a preset threshold, the signal is identified as an anomaly and an alarm is triggered.

[0097] Based on the quantum annealing optimization mechanism, the pulse width-modulation type constrained Hamiltonian is introduced to suppress the interference of carrier frequency drift and pulse width jitter on signal sorting. The Hamiltonian Combining the carrier frequency cluster center constraint and the pulse width distribution prior knowledge, its expression is:

[0098]

[0099] in, and are the qubit index and the total number of qubits, respectively. and The constraint coefficients of carrier frequency and pulse width are preferably 0.1 and 0.08 respectively; and Respectively The carrier frequency value of quantum bits and the target carrier frequency value, is the reference pulse width distribution, For the The Pauli-Z operator for 1 qubit.

[0100] In the actual experiment, the reference distribution Assuming Gaussian distribution, specifically μ = 1.2μs, σ = 0.3μs, KL divergence penalty makes the pulse width estimation mean square error MSE from 0.15μs 2 Down to 0.07μs 2 The signal clustering purity of the same modulation type is increased from 82% to 95%, the false correlation rate is less than 3%, the number of quantum annealing iterations is reduced to 1 / 3 of the traditional method, and the energy consumption of a single task is reduced by 58%.

[0101] Figure 2The framework of the dynamic adversarial network (DAN) is presented. Designed around the requirements of spectrum fingerprint matching and anomaly localization in electromagnetic signal processing, this model combines multimodal feature fusion with adversarial training to construct an end-to-end dynamic adaptive network. The model takes the raw electromagnetic signal as input and first converts the time-domain signal into a time-frequency spectrogram using a short-time Fourier transform (STFT) preprocessing module, achieving a joint time-frequency representation of the signal features. The parameters of a 256-window length and a 75% overlap rate balance temporal accuracy and frequency resolution. In a shared encoder, the preprocessed spectrogram is fed into a multi-scale convolution block combined with multi-head attention. The multi-scale convolution block uses kernel sizes of 3×3, 5×5, and 7×7 to extract local frequency band features, including the instantaneous frequency features of radar pulses. The Transformer encoding layer uses a multi-head self-attention mechanism to capture global cross-band correlations, including focusing on carrier-switching regions in frequency-hopping communication signals to suppress noise interference. In the multi-task collaborative branch, encoded features are split into two parallel task branches. The first branch is the spectral fingerprint matching branch. This branch compares known signal templates, such as radar emitter fingerprints, against a dynamically updated memory library. A dynamic prototype alignment loss is used to calculate the cosine similarity between the input signal and the template, with a threshold of ≥0.9 considered a match. The signal type / ID and match probability are output, enabling high-precision recognition in small sample sizes. The second branch is the anomaly localization branch. This branch uses a U-Net decoder to upsample the encoded features to restore spatial resolution in the time-frequency domain. A pixel-level anomaly heatmap is generated, including areas of pulse loss or frequency deviation anomalies. Combined with a binary mask, the anomaly frequency band and timestamp coordinates are output, with a localization error of ≤0.5ms / 2MHz. In adversarial training optimization, a gradient reversal layer (GRL) reverses the sign of the feature extractor's gradient, forcing the encoder to generate robust features that are independent of the interference environment. The multi-domain discriminator is designed with independent classification heads to address different electromagnetic interference scenarios, including strong noise and multipath effects. The adversarial loss weights are dynamically adjusted, starting with λ = 0.3 and increasing to 1.2 during training, achieving cross-domain adaptation. The final output layer integrates the results of the two branches, including fingerprint matching results: identifying signal types in the form of classification probabilities, including radar models LFM / FSK; anomaly location masks: binary thermal images mark abnormal areas, supporting real-time threat warnings in electronic warfare, and measured end-to-end latency ≤50ms.

[0102] In the dynamic adversarial network (DAN), the robustness of carrier frequency and pulse width features is jointly optimized to improve the defense capability against adversarial samples.

[0103] Define the carrier frequency-pulse width joint robustness loss in the adversarial defense module :

[0104]

[0105] in, To find the expectation of the statistical distribution of the input signal x, ensure that the robustness loss covers all possible signal scenarios; and are the disturbances for carrier frequency f and pulse width pw, respectively. and is the maximum allowable boundary for anti-disturbance, which is the disturbance amplitude that limits the carrier frequency and pulse width, and the value is =0.1GHz, =5μs; To measure the distribution difference between the electromagnetic signal data x and the adversarial sample x+δ in the feature space f(.), is the modulation type classification loss weight, which is 0.3 after tuning; is the cross entropy loss, which quantifies the electromagnetic signal data x and the adversarial sample x adv The difference in the modulation type prediction distribution m(.), is a feature extraction function that maps the input signal x to a high-dimensional feature space. is a modulation type classifier that outputs the modulation type probability distribution of the signal x.

[0106] A multimodal knowledge graph is constructed, radar signal features are defined as entities, communication protocol metadata as relationships between entities, and geographic spatial information as entity attributes. Semantic association modeling is performed and a heterogeneous network representation is generated.

[0107] Specifically, in the construction of the multimodal knowledge graph, a multidimensional feature tensor fusion mechanism is added to define the correlation function between the radar signal modulation type and the protocol metadata. for:

[0108]

[0109] in, is the total number of modulation types, The modulation type encoding matrix is ​​the discrete modulation type Mapped to a three-dimensional vector, For the Standard carrier frequency mean value of the modulated signal, used to dynamically adjust the carrier frequency The weight of For tensor product operation, fusion carrier frequency , pulse width and time difference ; is a parameter matrix used to adjust the pulse width and time difference The concatenated vector [ ; ] for feature mapping;

[0110] The geographic space information is encoded by using spherical harmonic function :

[0111]

[0112] in, is the geographic spatial weight coefficient, are spherical harmonic basis functions, The latitude and longitude coordinates of the signal sources respectively; 、 、 angular quantum number, magnetic quantum number and maximum harmonic order respectively.

[0113] The tensor representation generated by the association function serves as the relationship vector between entities in the knowledge graph, and the spherical harmonic encoding results serve as entity attributes, together forming a heterogeneous network representation that includes topological connections and semantic constraints. Specifically, the association function is used to define dynamic coordination rules between radar signals and communication protocols, generating interpretable semantic relationships; while spherical harmonic encoding is used to achieve high-precision modeling of geospatial attributes and embedding signal propagation characteristics.

[0114] For example, the radar signal entity is: {ID: RS001, Type: Linear Frequency Modulation, Carrier Frequency: 9.5GHz, Pulse Width: 1.2μs}; the communication protocol entity is: {ID: CP002, Type: Link-16, Encryption Level: AES-256, Rate: 1Mbps}; and the geographic entity is: {ID: G003, Longitude: 116.4°, Latitude: 39.9°, Signal Strength: -80dBm}. The knowledge inference engine mines cross-domain entity relationships. For example, when a radar signal group in a certain area with a carrier frequency of 9.5GHz±0.2GHz appears simultaneously with abnormal communication protocol metadata, such as an unknown encryption algorithm, a potential electronic warfare attack alert is triggered. In the experiment, The three-dimensional vector mapping represents the signal's time-frequency-modulation characteristics. The BPSK mapping is [0.2, 0.7, -0.1], and the QPSK mapping is [0.5, -0.3, 0.8]. The parameter matrix W is initialized using Xavier with a dimension of 2×d, where d is the concatenated vector dimension. In testing, feature mapping achieved optimal results when d = 128. When L = 6, the average positioning error was 0.48°. As the order increases, the positioning error decreases more slowly, while the computational complexity increases exponentially. When L = 6, the error is 41% higher than when L = 3, which achieved an error of 0.82°.

[0115] A spatiotemporal dual-channel attention mechanism is adopted to fuse the time domain features of the radar pulse sequence with the spatial distribution features of the communication signal group to generate a joint semantic vector; the fusion of multi-source heterogeneous data is achieved through a dynamic confidence allocation algorithm.

[0116] In the spatiotemporal dual-channel attention mechanism, the temporal channel Introducing the carrier frequency-pulse width coupled convolution kernel:

[0117]

[0118] in, is the carrier frequency difference sensitivity coefficient, which is 0.15; is the time domain sliding window length, and are the tth moment and the history The carrier frequency value at the time, For history The pulse width value at the moment, is a function for extracting the periodic characteristics of pulse width;

[0119] The spatial channel uses pulse width weighted geodesic distance :

[0120]

[0121] in, is the pulse width difference penalty factor, which is set to 0.6; are the feature vectors of nodes i and j respectively, is the eigenvector angle;

[0122] The dynamic confidence allocation algorithm defines the modulation type entropy weight as follows:

[0123]

[0124] in, For the The information entropy of the signal modulation type, C is the total number of signal categories, Is the loop variable.

[0125] The time domain features extracted by the carrier frequency-pulse width coupled convolution kernel and the spatial domain features of the pulse width weighted geodesic distance model are dynamically fused through the modulation type entropy weight to generate a joint semantic vector, which is input into the spatiotemporal Transformer model for threat prediction.

[0126] For example, in a dual-channel spatiotemporal attention mechanism, a carrier-pulse-width coupled convolution kernel w=5 is used in the time domain channel to capture carrier frequency hopping sequences, with a frequency-agile radar hopping interval of ≤1ms. In the spatial channel, the signal population density is quantified using a pulse-width-weighted geodesic distance α=0.3, with a frequency of ≥50 pulses per square kilometer. A dynamic confidence allocation algorithm assigns differentiated weights to heterogeneous data based on the entropy weight of the modulation type, such as ω=0.8 for conventional communication signals and ω=1.2 for frequency-hopping signals. Simulation tests show that this mechanism improves the discrimination between communication jammers and real threats by 2.7 times.

[0127] Based on the joint semantic vector, the spatiotemporal Transformer model is used to dynamically predict the threat evolution trend and output multi-task prediction results of threat level, target type and behavioral intention; combined with the hybrid density network, a dynamic battlefield situation panorama is generated.

[0128] Specifically, in the spatiotemporal Transformer model, a carrier frequency change rate gating mechanism is embedded in the threat prediction module. :

[0129]

[0130] in, is the carrier frequency change rate, It is the pulse width stability index;

[0131] The conditional probability distribution of the mixture density network Modeled as:

[0132]

[0133] in, is the weight of the kth Gaussian component under the input x, and is the carrier frequency mean and standard deviation of the kth Gaussian component, and Beta distribution describes the statistical characteristics of the pulse width. and are the shape parameters that control the left-biased and right-biased characteristics of the pulse width distribution, and the Cat distribution represents the probability of modulation type classification. is the modulation type probability vector.

[0134] The carrier frequency change rate gating mechanism dynamically adjusts the attention weight distribution of each time step in the spatiotemporal Transformer model based on the generated joint semantic vector, wherein the attention weight of the historical pulse width feature is enhanced when the carrier frequency change rate exceeds a threshold; the conditional probability distribution modeling of the hybrid density network integrates the geographic space coding features and the signal carrier frequency-pulse width joint distribution to generate a dynamic battlefield situation panorama that includes the threat thermal field and signal propagation path.

[0135] For example, In order to control the contribution weight of carrier frequency change rate to threat prediction, the Sigmoid function output range is [0,1]. =0.8 indicates a high level of concern about carrier frequency mutations; is a linear transformation matrix that maps the input features to the gated space with a dimension of 128×64, 64 input features, 128 hidden layers, and Xavier initialization; is the carrier frequency change per unit time, the frequency agile radar signal =0.5 MHz / ms; is the statistical stability of the pulse width sequence, calculated as the inverse of the standard deviation of the pulse width within a sliding window of length r, for a stable radar pulse σ = 0.1 μs → Stability = 8.0; is the mixture weight of the k-th Gaussian component, generated by Softmax. When K=3, π=[0.6, 0.3, 0.1], indicating that the main threat comes from category 1; is the carrier frequency center value of the kth Gaussian component, μ1=3.0 GHz for S-band radar and μ2=9.5 GHz for X-band radar; is the discrete degree of carrier frequency distribution, σ=50MHz for conventional radar and σ=200MHz for frequency hopping radar; To control the left / right deviation of the pulse width distribution, when κ>λ, the pulse width is left-biased. Specifically, the short pulse width is concentrated. For normal pulses, κ=2.0, λ=5.0 (right-biased); for burst pulses, κ=5.0, λ=2.0 (left-biased). is the probability distribution of the modulation type, such as the confidence level of [BPSK, QPSK, LFM]. p = [0.7, 0.2, 0.1] indicates a BPSK probability of 70%. The mixture density network (MDN) models the conditional probability distribution using a Gaussian mixture model (GMM).

[0136] Deploy an online incremental learning framework to detect new signal patterns and hierarchically fine-tune model parameters, prioritizing the update of the prediction head module to adapt to the dynamic battlefield environment.

[0137] The prediction head module is a built-in component of the online incremental learning framework. Its update logic, parameter adjustment, and function execution all rely on the framework's detection and scheduling mechanisms. This design ensures the model's rapid adaptability in dynamic battlefield environments while balancing stability and flexibility through a layered update strategy.

[0138] Specifically, in the online incremental learning framework, the detection of the new signal adopts a modulation type entropy change trigger mechanism Specifically:

[0139]

[0140] When the modulation type distribution entropy value of the new signal batch and historical distribution entropy The difference exceeds the threshold When , the update mechanism of the incremental learning model is triggered;

[0141] Layered fine-tuning applies quantum noise perturbations to the carrier frequency feature layer, specifically:

[0142]

[0143] in, is the weight matrix of the frequency feature layer, is the average offset of the signal carrier frequency, is the quantum noise perturbation intensity coefficient.

[0144] For example, the modulation type entropy value The value range is 0~log2(C), where C is the total number of modulation types. For example, when C=5, the maximum entropy is ≈2.32; the modulation type entropy change trigger threshold =0.2, it can effectively capture significant changes in the modulation type distribution, such as the low probability of intercept signal of the newly emerged LPI radar, while avoiding false triggering caused by noise. =1.0, corresponding to BPSK / QPSK mixed mode; new batch entropy value , newly added LFM signal, difference Δ=0.4> =0.2→trigger update. Quantum noise perturbation intensity coefficient =0.5, the noise disturbance amplitude is about 5% of the weight. In the X-band radar signal, the carrier frequency is 9.5GHz, and the frequency deviation interference of ±0.5MHz can be tolerated. The initial weight of the carrier frequency feature layer is =[0.8, -0.3, 1.2]; after adding noise, it is updated to =[0.8+0.05·N(0,1), -0.3+0.05·N(0,1), 1.2+0.05·N(0,1)]; learning rate =0.005, the model can adapt to new signals, such as FHSS frequency hopping signals, within 10 iterations, the convergence speed is increased by 40%, and the average carrier frequency offset is =2.5MHz→ =0.92; weight update amount: ΔW=0.005×N(0,0.05²)×0.92≈±0.0023; =3.0MHz, Sigmoid(3.0)=0.95, and the noise perturbation weight is increased to 95%; the enemy radar carrier frequency jumps from 9.5GHz to 9.503GHz→ =3.0MHz → triggers high sensitivity update.

[0145] A reinforcement learning-driven multi-path distribution mechanism is implemented through the edge-cloud collaborative architecture, and the intelligence push priority is dynamically optimized based on the channel state matrix and node security assessment to ensure low-latency transmission of critical threat intelligence.

[0146] Specifically, in the edge-cloud collaborative architecture, the multi-path distribution mechanism is used to define the carrier frequency-bit error rate joint reward function

[0147]

[0148] in, 、 、 is the weighting coefficient, The carrier frequency f corresponds to the inverse of the channel bit error rate, is the total transmission time of the signal from the transmitter to the receiver, is the delay sensitivity factor, Provides a security evaluation vector for the node.

[0149] For example, in a conventional tactical communication scenario: =0.3 represents the bit error rate weight, and the tolerance BER≤e -4 ,avoid frequent channel switching affecting timeliness; =0.5 represents the delay weight, emphasizing low-latency transmission and requiring end-to-end delay ≤50ms; =0.2 represents the security weight, and the tolerance for node security assessment differences is high. ≤0.3; =50ms, when the delay exceeds 50ms, the reward value decays rapidly; =e 4 , =40ms, =[0.8, 0.9]. In the critical threat intelligence transmission scenario, =0.2 represents the bit error rate weight, requiring the channel bit error rate to be extremely low BER≤e -6 , select high-quality links; =0.3 represents the delay weight, which allows slightly higher delay and requires end-to-end delay ≤80ms, but requires stable transmission; =0.5 represents the safety weight, strictly limiting the safety difference, ≤0.1; =100ms, the delay penalty decays slowly, avoiding excessive sacrifice of stability; =e 6 , =70ms, =[1.0, 1.0]. In the emergency scenario of strong interference environment, =0.4 represents the bit error rate weight, giving priority to anti-interference links (such as spread spectrum communication, BER≤e -3 ; =0.4 represents the delay weight, requiring the delay to be ≤100ms to support fast response; =0.2 represents the security weight, temporarily relaxing security requirements, such as enabling emergency communication protocols; =30ms, =e 3 , =90ms, =[0.6, 0.7], indicating a temporarily enabled emergency node with a degraded security level but available.

[0150] In one embodiment of the present application, Figure 3 As shown, a simulation system for intelligent fusion and dynamic prediction of electronic warfare intelligence is also provided, and the simulation system includes:

[0151] The multi-domain data acquisition and preprocessing module is used to collect electromagnetic signal data in real time through multi-domain distributed sensor nodes. It uses the dynamic adversarial network (DAN) to perform spectrum fingerprint matching and anomaly location on the electromagnetic signal data, and combines the spatiotemporal consistency verification mechanism to eliminate invalid data and generate a standardized intelligence stream.

[0152] A multimodal knowledge graph construction module is used to construct a multimodal knowledge graph, perform semantic association modeling on radar signal features, communication protocol metadata, and geospatial information, and generate a heterogeneous network representation containing entities, relationships, and attributes;

[0153] The spatiotemporal feature fusion engine uses a spatiotemporal dual-channel attention mechanism to fuse the temporal features of radar pulse sequences with the spatial distribution features of communication signal groups to generate a joint semantic vector. It also uses a dynamic confidence allocation algorithm to achieve the fusion of multi-source heterogeneous data.

[0154] A threat prediction and situation generation module is used to dynamically predict threat evolution trends based on the joint semantic vector using a spatiotemporal Transformer model, output multi-task prediction results of threat level, target type, and behavioral intention; and generate a dynamic battlefield situation panorama in combination with a hybrid density network;

[0155] An online incremental learning agent module, which deploys an online incremental learning framework, detects new signal patterns, and fine-tunes model parameters hierarchically, prioritizing updates to the prediction head module to adapt to dynamic battlefield environments.

[0156] The edge-cloud collaborative distribution control module is used to implement a reinforcement learning-driven multi-path distribution mechanism through the edge-cloud collaborative architecture, dynamically optimize the intelligence push priority based on the channel state matrix and node security assessment, and ensure low-latency transmission of critical threat intelligence.

[0157] Based on the same inventive concept, according to another aspect of the present invention, an embodiment of the present invention further provides a computer device, including:

[0158] at least one processor; and

[0159] The memory stores a computer program that can be run on the processor, and the processor executes the steps of any one of the above simulation methods when executing the program.

[0160] Based on the same inventive concept, according to another aspect of the present invention, an embodiment of the present invention further provides a computer-readable storage medium, which stores computer program instructions. When the computer program instructions are executed by a processor, the steps of any of the above simulation methods are performed.

[0161] Finally, it should be noted that those skilled in the art will understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The program can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. The storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM). The above-mentioned computer program embodiments can achieve the same or similar effects as the corresponding embodiments of any of the above-mentioned methods.

[0162] Furthermore, the apparatuses and devices disclosed in the embodiments of the present invention may typically be various electronic terminal devices, such as mobile phones, personal digital assistants (PDAs), tablet computers (PADs), smart televisions, etc., or large terminal devices, such as servers. Therefore, the scope of protection disclosed in the embodiments of the present invention should not be limited to a specific type of apparatus or device. The client disclosed in the embodiments of the present invention may be implemented in any of the above-mentioned electronic terminal devices in the form of electronic hardware, computer software, or a combination of both.

[0163] In addition, the method disclosed in the embodiment of the present invention can also be implemented as a computer program executed by a CPU, and the computer program can be stored in a computer-readable storage medium. When the computer program is executed by the CPU, the above functions defined in the method disclosed in the embodiment of the present invention are performed.

[0164] In addition, the above method steps and system units can also be implemented using a controller and a computer-readable storage medium for storing a computer program that enables the controller to implement the above steps or unit functions.

[0165] The above are exemplary embodiments disclosed in the present invention, but it should be noted that various changes and modifications may be made without departing from the scope of the embodiments disclosed in the claims. The functions, steps and / or actions of the method claims according to the disclosed embodiments described herein do not need to be performed in any marked order. In addition, although the elements disclosed in the embodiments of the present invention may be described or required in individual form, they may also be understood as multiple unless expressly limited to the singular.

[0166] In this specification, the same or similar parts between the various embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the embodiments described later, the description is relatively simple, and the relevant parts can be referred to the partial description of the previous embodiments.

[0167] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A simulation method for intelligent fusion and dynamic prediction of electronic warfare intelligence, characterized in that: The simulation method comprises: Electromagnetic signal data is collected in real time through multi-domain distributed sensor nodes. Dynamic adversarial network (DAN) is used to perform spectrum fingerprint matching and anomaly location on the electromagnetic signal data. In addition, a spatiotemporal consistency verification mechanism is used to eliminate invalid data and generate a standardized intelligence stream. Construct a multimodal knowledge graph, define radar signal features as entities, communication protocol metadata as relationships between entities, and geospatial information as entity attributes, perform semantic association modeling, and generate heterogeneous network representations. Based on the heterogeneous network representation, a spatiotemporal dual-channel attention mechanism is adopted to fuse the time domain features of the radar pulse sequence with the spatial distribution features of the communication signal group to generate a joint semantic vector; and a dynamic confidence allocation algorithm is used to achieve the fusion of multi-source heterogeneous data. Based on the joint semantic vector, the spatiotemporal Transformer model is used to dynamically predict the threat evolution trend and output multi-task prediction results of threat level, target type and behavioral intention; combined with the hybrid density network to generate a dynamic battlefield situation panorama; Deploy an online incremental learning framework to detect new signal patterns and fine-tune model parameters in layers, updating the prediction head module to adapt to the dynamic battlefield environment; Implementing a reinforcement learning-driven multi-path distribution mechanism through an edge-cloud collaborative architecture, dynamically optimizing intelligence push priorities based on the channel state matrix and node security assessment, ensuring low-latency transmission of critical threat intelligence; The dynamic adversarial network (DAN) is used to perform spectrum fingerprint matching and anomaly location on the electromagnetic signal data. In combination with the spatiotemporal consistency verification mechanism, invalid data is eliminated to generate a standardized intelligence stream, including: In the dynamic adversarial network (DAN), an improved spectral clustering algorithm is used and a quantum annealing optimization mechanism is introduced to perform feature sorting and spectrum fingerprint matching on the electromagnetic signal data; the feature sorting is to separate the mixed electromagnetic signal into different groups according to carrier frequency, pulse width, and modulation type; the output result of the feature sorting is used as the input of the spectrum fingerprint matching, and the cluster center vector is dynamically updated to perform similarity comparison with the radiation source characteristics in the fingerprint library to identify abnormal signals and generate standardized intelligence streams; the loss of the spectrum fingerprint matching is Designed to: in, is the three-dimensional feature vector of the electromagnetic signal sample, is the i-th signal sample, For the The cluster center vector at the iteration, N is the total number of valid signal samples participating in the cluster center update in the current iteration, is the quantum tunneling intensity coefficient; is the adaptive cutoff threshold, To maximize the difference between classes Minimizing intra-class variance The comprehensive measure of is the quantum confinement factor; Cluster Center The formula for dynamic update is: in, is the learning rate, Gaussian kernel bandwidth The specific adjustments are: in, is the distribution shift sensitivity coefficient, is the KL divergence measure of signal distribution changes, For the At the iteration The joint probability distribution of For the The cluster center vector at the iteration, the DAN is an end-to-end dynamic adaptation network.

2. The simulation method for intelligent fusion and dynamic prediction of electronic warfare intelligence according to claim 1, characterized in that: Construct a multimodal knowledge graph, define radar signal features as entities, communication protocol metadata as relationships between entities, and geospatial information as entity attributes, perform semantic association modeling, and generate heterogeneous network representations, including: In the construction of the multimodal knowledge graph, a multidimensional feature tensor fusion mechanism is added to define the correlation function between the radar signal modulation type and the communication protocol metadata. for: in, is the total number of modulation types, For the The modulation type encoding matrix of the class modulation signal converts the discrete modulation type Mapped to a three-dimensional vector, For the Standard carrier frequency mean value of the modulated signal, used to dynamically adjust the carrier frequency The weight of For tensor product operation, fusion carrier frequency , pulse width and time difference ; is a parameter matrix used to adjust the pulse width and time difference The concatenated vector [ ; ] for feature mapping; The geographic space information is encoded by using spherical harmonic function : in, is the geographic spatial weight coefficient, are spherical harmonic basis functions, The latitude and longitude coordinates of the signal sources respectively; 、 、 angular quantum number, magnetic quantum number and maximum harmonic order respectively; The tensor representation generated by the association function serves as the relationship vector between entities in the multimodal knowledge graph, and the spherical harmonic encoding result serves as the entity attribute, which together constitute the heterogeneous network representation containing topological connections and semantic constraints.

3. The simulation method for intelligent fusion and dynamic prediction of electronic warfare intelligence according to claim 2, characterized in that: A spatiotemporal dual-channel attention mechanism is used to fuse the temporal features of the radar pulse sequence with the spatial distribution features of the communication signal group to generate a joint semantic vector, including: In the spatiotemporal dual-channel attention mechanism, the temporal channel Introducing the carrier frequency-pulse width coupled convolution kernel: in, is the carrier frequency difference sensitivity coefficient, is the time domain sliding window length, and are the tth moment and the history The carrier frequency value at the time, For history The pulse width value at the moment, is a function for extracting the periodic characteristics of pulse width; The spatial channel uses pulse width weighted geodesic distance : in, is the pulse width difference penalty factor, are the feature vectors of nodes i and j respectively, is the eigenvector angle; The dynamic confidence allocation algorithm defines the modulation type entropy weight Specifically: in, For the The information entropy of the signal modulation type, C is the total number of signal categories, is the loop variable; The time domain features extracted by the carrier frequency-pulse width coupled convolution kernel and the spatial domain features of the pulse width weighted geodesic distance modeling are dynamically fused through the modulation type entropy weight to generate the joint semantic vector, which is input into the spatiotemporal Transformer model for threat prediction.

4. The simulation method for intelligent fusion and dynamic prediction of electronic warfare intelligence according to claim 3, characterized in that: In the spatiotemporal Transformer model, a carrier frequency change rate gating mechanism is embedded in the threat prediction module. : in, is the carrier frequency change rate, is the pulse width stability index, is the linear transformation matrix; The conditional probability distribution of the mixture density network Modeled as: in, is the weight of the kth Gaussian component under the input x, and is the carrier frequency mean and standard deviation of the kth Gaussian component, and Beta distribution describes the statistical characteristics of the pulse width. and are the shape parameters that control the left-biased and right-biased characteristics of the pulse width distribution, and the Cat distribution represents the probability of modulation type classification. is the modulation type probability vector; The carrier frequency change rate gating mechanism dynamically adjusts the attention weight distribution of each time step in the spatiotemporal Transformer model based on the generated joint semantic vector, wherein the attention weight of the historical pulse width feature is enhanced when the carrier frequency change rate exceeds a threshold; the conditional probability distribution modeling of the hybrid density network integrates the geographic space coding features and the signal carrier frequency-pulse width joint distribution to generate a dynamic battlefield situation panorama including the threat thermal field and signal propagation path.

5. The simulation method for intelligent fusion and dynamic prediction of electronic warfare intelligence according to claim 4, characterized in that: In the online incremental learning framework, detecting the new signal pattern includes: Adopting modulation type entropy change trigger mechanism Detecting the new signal pattern, the modulation type entropy change trigger mechanism Specifically: in, (.) is the indicator function, when the modulation type distribution entropy value of the new signal batch and historical distribution entropy The difference exceeds the threshold When , the update mechanism of the incremental learning model is triggered; Fine-tune model parameters layer by layer, including: Layered fine-tuning applies quantum noise perturbations to the carrier frequency feature layer, specifically: in, is the weight matrix of the frequency feature layer, is the average offset of the signal carrier frequency, is the quantum noise perturbation intensity coefficient.

6. The simulation method for intelligent fusion and dynamic prediction of electronic warfare intelligence according to claim 5, characterized in that: Implementing a reinforcement learning-driven multi-path distribution mechanism through an edge-cloud collaborative architecture, including: In the edge-cloud collaborative architecture, the multi-path distribution mechanism is used to define the carrier frequency-bit error rate joint reward function Specifically: in, 、 、 is the weighting coefficient, The carrier frequency f corresponds to the inverse of the channel bit error rate, is the total transmission time of the signal from the transmitter to the receiver, is the delay sensitivity factor, is the node security assessment vector, is the reference safety vector.

7. The simulation method for intelligent fusion and dynamic prediction of electronic warfare intelligence according to claim 1, characterized in that: On the basis of the quantum annealing optimization, a pulse width-modulation type constrained Hamiltonian is added to suppress the interference of carrier frequency drift and pulse width jitter on signal sorting. The Hamiltonian Combining the carrier frequency cluster center constraint and the pulse width distribution prior knowledge, the Hamiltonian The expression is: in, and are the qubit index and the total number of qubits, respectively. and The constraint coefficients of carrier frequency and pulse width respectively, and Respectively The carrier frequency value of quantum bits and the target carrier frequency value, is the reference pulse width distribution, For the The Pauli-Z operator for 1 qubit.

8. The simulation method for intelligent fusion and dynamic prediction of electronic warfare intelligence according to claim 1, characterized in that: In the dynamic adversarial network DAN, the carrier frequency-pulse width joint robustness loss is defined in the adversarial defense module. : in, To find the expectation of the statistical distribution of the input signal x, and are the disturbances for carrier frequency f and pulse width pw, respectively. and are the disturbance amplitudes that limit the carrier frequency and pulse width, To measure the distribution difference between the electromagnetic signal data x and the adversarial sample x+δ in the feature space f(.), is the modulation type classification loss weight, is the cross entropy loss, which quantifies the electromagnetic signal data x and the adversarial sample x adv The difference in the modulation type prediction distribution m(.), is a feature extraction function that maps the input signal x to a high-dimensional feature space. is a modulation type classifier that outputs the modulation type probability distribution of the signal x.

9. A simulation system for intelligent fusion and dynamic prediction of electronic warfare intelligence, used to implement a simulation method for intelligent fusion and dynamic prediction of electronic warfare intelligence according to any one of claims 1 to 8, characterized in that: The simulation system comprises: The multi-domain data acquisition and preprocessing module is used to collect electromagnetic signal data in real time through multi-domain distributed sensor nodes, use the dynamic adversarial network (DAN) to perform spectrum fingerprint matching and anomaly location on the electromagnetic signal data, and combine the spatiotemporal consistency verification mechanism to eliminate invalid data and generate a standardized intelligence stream; A multimodal knowledge graph construction module is used to construct a multimodal knowledge graph, defining radar signal features as entities, communication protocol metadata as relationships between entities, and geospatial information as entity attributes, to perform semantic association modeling and generate heterogeneous network representations; The spatiotemporal feature fusion engine uses a spatiotemporal dual-channel attention mechanism to fuse the temporal features of radar pulse sequences with the spatial distribution features of communication signal groups to generate a joint semantic vector. It also uses a dynamic confidence allocation algorithm to achieve the fusion of multi-source heterogeneous data. A threat prediction and situation generation module is used to dynamically predict threat evolution trends based on the joint semantic vector using a spatiotemporal Transformer model, output multi-task prediction results of threat level, target type, and behavioral intention; and generate a dynamic battlefield situation panorama in combination with a hybrid density network; An online incremental learning agent module, which deploys an online incremental learning framework to detect new signal patterns and fine-tune model parameters hierarchically, updating the prediction head module to adapt to the dynamic battlefield environment; The edge-cloud collaborative distribution control module is used to implement a reinforcement learning-driven multi-path distribution mechanism through the edge-cloud collaborative architecture, dynamically optimize the intelligence push priority based on the channel state matrix and node security assessment, and ensure low-latency transmission of critical threat intelligence.

Images