Intelligent electromagnetic environment sensing and anti-interference system

Abstract

The invention discloses an intelligent electromagnetic environment perception and anti-interference system, which belongs to the application technology of artificial intelligence in the field of electronic countermeasure, and comprises an electromagnetic signal acquisition module, a multi-scale feature extraction and fusion module, a frequency spectrum situation prediction and decision module, a self-adaptive anti-interference execution module and a performance feedback and optimization module, multi-scale feature extraction is performed through a deep convolutional neural network and a differential transformation network, a spectrum situation is predicted based on a long-short term memory network and an attention mechanism, an anti-interference decision is generated by adopting deep reinforcement learning, and a complete closed-loop system is constructed through performance feedback. According to the method, the signal identification accuracy and the anti-interference performance in a complex electromagnetic environment are remarkably improved, the system communication reliability is improved by more than 50%, and the spectrum utilization efficiency is improved by more than 35%.

Description

Technical Field

[0001] This invention relates to the field of electromagnetic environment sensing and anti-interference technology, specifically to an intelligent electromagnetic environment sensing and anti-interference system based on artificial intelligence, which belongs to the application of artificial intelligence in the field of electronic warfare. Background Technology

[0002] With the rapid development of wireless communication technology and the increasing scarcity of electromagnetic spectrum resources, the electromagnetic environment has become increasingly complex. In both military and civilian fields, wireless communication systems face various complex electromagnetic interference threats, including suppression interference, deceptive interference, and emerging intelligent cognitive interference. These interferences severely affect the reliability and effectiveness of communication systems, posing significant challenges to spectrum management and communication assurance.

[0003] Traditional anti-interference techniques mainly rely on single techniques such as frequency hopping spread spectrum and power control, lacking the ability to adapt to complex electromagnetic environments. These methods are usually based on preset rules or templates for interference detection and countermeasures, making it difficult to cope with rapidly changing interference patterns and unknown interference types. Especially in complex electromagnetic environments with low signal-to-noise ratios, the detection accuracy and anti-interference effectiveness of traditional methods decrease significantly.

[0004] CN119830989B discloses an offline decision-making method, system, device, and storage medium for electronic countermeasures. This patent achieves radar jamming decision-making by modeling the electromagnetic environment as an adversary agent, with the jamming agent interacting and training with the adversary agent. However, this solution has the following shortcomings: First, this solution is mainly for offline decision-making scenarios, relying on historical game sequences to construct the adversary agent, lacking online perception and rapid response capabilities for the real-time electromagnetic environment, making it difficult to achieve real-time decision-making in dynamically changing electromagnetic environments; Second, this solution focuses on the jammer's decision optimization, without fully considering the extraction of multi-scale time-frequency features of electromagnetic signals and the prediction of spectral situation, resulting in insufficient recognition capabilities for complex modulation signals and rapidly changing spectral environments; Third, the agent training in this solution relies on the original dataset and extended dataset, requiring a large amount of data extrapolation and iterative training, resulting in low training efficiency and high data quality requirements; Fourth, this solution lacks a closed-loop optimization mechanism from decision execution to performance feedback, and cannot dynamically adjust the decision strategy according to the actual communication effect, limiting the system's adaptive optimization capability in real environments.

[0005] In recent years, deep learning technology has made groundbreaking progress in signal processing and spectrum sensing. Deep convolutional neural networks have demonstrated powerful features extraction capabilities, automatically learning deep feature representations of signals; long short-term memory networks excel at processing time-series data and can capture the temporal evolution of spectrum situations; reinforcement learning performs well in dynamic decision-making problems, learning optimal strategies through interaction with the environment. However, most existing research applies these technologies in isolation, lacking systematic integration solutions, making it difficult to form a complete closed-loop system from environmental perception and situation prediction to decision execution.

[0006] Furthermore, existing spectrum sensing methods are typically based on traditional techniques such as energy detection and matched filtering, which significantly degrade in performance under low signal-to-noise ratio and complex interference environments. Although some studies have attempted to apply deep learning to spectrum sensing, most methods focus only on feature extraction at a single scale, neglecting the characteristics of electromagnetic signals at different time-frequency scales, resulting in low accuracy in identifying complex modulation signals and diverse interference patterns. Meanwhile, existing anti-interference decision-making methods are often based on static rules or simple optimization algorithms, lacking the ability to predict the dynamic evolution of the electromagnetic environment and failing to achieve forward-looking spectrum resource planning and adaptive parameter adjustment.

[0007] Therefore, there is an urgent need for a systematic solution that can perceive complex electromagnetic environments in real time, accurately predict spectrum trends, intelligently generate anti-interference decisions, and form closed-loop optimization to meet the reliable communication requirements of modern wireless communication systems in complex electromagnetic environments. Summary of the Invention

[0008] The purpose of this invention is to provide an intelligent electromagnetic environment sensing and anti-interference system to solve the technical problems in the prior art, such as inaccurate electromagnetic environment sensing, insufficient spectrum situation prediction capability, lack of adaptability in anti-interference decision-making, and lack of closed-loop optimization mechanism.

[0009] To achieve the above objectives, the present invention adopts the following technical solution:

[0010] This invention provides an intelligent electromagnetic environment perception and anti-interference system, comprising an electromagnetic signal acquisition module, a multi-scale feature extraction and fusion module, a spectrum situation prediction and decision-making module, an adaptive anti-interference execution module, and a performance feedback and optimization module. The electromagnetic signal acquisition module is responsible for acquiring electromagnetic signals within the target frequency band and generating multi-dimensional raw signal data. The multi-scale feature extraction and fusion module uses a deep convolutional neural network and a differential transform network to perform multi-scale time-frequency feature extraction and feature fusion processing on the raw signal data, generating a fused feature vector containing frequency domain features, time domain features, and modulation features. The spectrum situation prediction and decision-making module, based on a long short-term memory network and an attention mechanism, constructs a spatiotemporal evolution model of spectrum occupancy based on the fused feature vector, predicts the spectrum situation at future times, and generates anti-interference decisions including frequency selection strategies, power adjustment strategies, and waveform design strategies based on a deep reinforcement learning network. The adaptive anti-interference execution module dynamically adjusts communication parameters according to the anti-interference decisions and collects actual communication performance indicators. The performance feedback and optimization module calculates the decision effectiveness evaluation value based on the actual communication performance indicators, and feeds the decision effectiveness evaluation value back to the spectrum situation prediction and decision module to adjust the policy parameters of the deep reinforcement learning network, forming a complete perception-prediction-decision-execution-feedback closed-loop system.

[0011] This invention achieves accurate extraction and efficient fusion of multi-scale time-frequency features of electromagnetic signals by setting up a multi-scale feature extraction and fusion module, combined with deep convolutional neural networks and differential transform networks, significantly improving the recognition accuracy of complex modulation signals and diverse interference modes. By setting up a spectrum situation prediction and decision-making module, and constructing a spatiotemporal evolution model of spectrum occupancy based on long short-term memory networks and attention mechanisms, accurate prediction of future spectrum situations is achieved, providing a reliable basis for forward-looking spectrum resource planning. By setting up a performance feedback and optimization module, a complete closed loop from decision execution to performance evaluation and parameter adjustment is constructed, enabling the system to dynamically optimize decision strategies based on actual communication effects, significantly improving the system's adaptability in real complex electromagnetic environments.

[0012] The beneficial effects of this invention are as follows:

[0013] First, the present invention employs a multi-scale feature extraction method that combines deep convolutional neural networks and differential transform networks, which can simultaneously capture the local spatial features and global dependency features of electromagnetic signals. Compared with traditional single-scale feature extraction methods, it improves the accuracy of complex modulation signal recognition by more than 20% and the anti-interference performance in low signal-to-noise ratio environments by more than 30%.

[0014] Second, this invention constructs a spatiotemporal evolution model of spectrum occupancy through long short-term memory networks and attention mechanisms, which can accurately predict the spectrum situation at future moments with a prediction accuracy of over 92%. This provides the system with a forward-looking spectrum resource planning capability, enabling the system to avoid congested frequency bands in advance and select the optimal communication parameters.

[0015] Third, this invention generates anti-interference decisions based on deep reinforcement learning networks, which can learn the optimal decision strategy in complex and dynamic electromagnetic environments. The decision response time is less than 10ms, which meets the requirements of real-time communication. Compared with traditional rule-driven methods, the anti-interference effect is improved by more than 40%.

[0016] Fourth, this invention constructs a complete closed-loop system through a performance feedback and optimization module, which can dynamically adjust the decision-making strategy according to the actual communication performance, enabling the system to have continuous learning and self-optimization capabilities. During long-term operation, the system performance steadily improves and its adaptability is enhanced.

[0017] Fifth, this invention constructs a complete technical link from electromagnetic signal acquisition, feature extraction, situation prediction, decision generation to execution feedback, realizing integrated intelligent processing of electromagnetic environment perception and anti-interference, improving the overall communication reliability of the system by more than 50% and the spectrum utilization efficiency by more than 35%. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the overall structure of the intelligent electromagnetic environment sensing and anti-interference system of the present invention.

[0019] Figure 2 This is a schematic diagram of the multi-scale feature extraction and fusion module of the present invention.

[0020] Figure 3 This is a schematic diagram of the spectrum situation prediction and decision-making module of the present invention.

[0021] Figure 4 This is a schematic diagram of the adaptive anti-interference execution module of the present invention.

[0022] Figure 5 This is a schematic diagram of the performance feedback and optimization module of the present invention.

[0023] Figure 6 This is a schematic diagram of the data flow and closed-loop feedback of the system of the present invention.

[0024] Figure 7 This is a schematic diagram of the training process of the system of the present invention. Detailed Implementation

[0025] Please refer to the attached document. Figures 1-7The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the following embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0026] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0027] Reference Figure 1 This invention provides an intelligent electromagnetic environment sensing and anti-interference system, which includes an electromagnetic signal acquisition module 1, a multi-scale feature extraction and fusion module 2, a spectrum situation prediction and decision-making module 3, an adaptive anti-interference execution module 4, and a performance feedback and optimization module 5. The modules are connected via data interfaces to form a complete information processing link and a closed-loop feedback loop.

[0028] The electromagnetic signal acquisition module 1 is used to acquire electromagnetic signals within the target frequency band and generate multi-dimensional raw signal data. In one specific embodiment, the electromagnetic signal acquisition module 1 includes a broadband antenna array, a radio frequency (RF) front-end, and an analog-to-digital converter (ADC). The broadband antenna array covers a frequency range of 2 GHz to 18 GHz and can simultaneously receive electromagnetic signals from multiple frequency bands. The RF front-end performs low-noise amplification, mixing, and intermediate frequency (IF) filtering on the received RF signals, converting the RF signals into IF signals. The ADC uses a high-speed ADC chip with a sampling rate of 100 MSPS and a sampling precision of 14 bits, converting the analog IF signals into digital signals. The raw signal data includes in-phase and quadrature components, stored in IQ data format, with a data sampling window length of 1024 points and sliding window sampling with a 50% overlap rate to ensure the temporal continuity of the signal. In a preferred embodiment, the sampling window length can be dynamically adjusted according to the signal bandwidth and processing requirements, ranging from 512 points to 2048 points.

[0029] The multi-scale feature extraction and fusion module 2 is connected to the electromagnetic signal acquisition module 1, and performs multi-scale time-frequency feature extraction and feature fusion processing on the original signal data to generate a fused feature vector. (Refer to...) Figure 2 The multi-scale feature extraction and fusion module 2 includes a time-frequency transformation unit, a convolutional feature extraction unit, a differential attention unit, and a feature fusion unit.

[0030] The time-frequency transformation unit performs a short-time Fourier transform (SFT) on the original signal data to generate a time-frequency spectrogram. The SFT uses a Hanning window with a length of 256 points, a frame shift of 128 points, and 512 FFT points. Through the SFT, the one-dimensional time-domain signal is converted into a two-dimensional time-frequency representation. The time-frequency spectrogram has a size of 512×64, where 512 represents the frequency dimension and 64 represents the time dimension. The time-frequency spectrogram undergoes logarithmic transformation and normalization to map to a numerical range of 0 to 1, adapting to the input requirements of the subsequent neural network.

[0031] The convolutional feature extraction unit, based on a multi-layer convolutional neural network, performs multi-scale convolution operations on the time-spectrum image to extract spatial features at different scales. The convolutional neural network employs a five-layer convolutional structure. The first layer has a 7×7 kernel size, 64 kernels, a stride of 2, and padding of 3, used to extract coarse-grained global features. The second layer has a 5×5 kernel size, 128 kernels, a stride of 1, and padding of 2, used to extract medium-scale local features. The third to fifth layers all have 3×3 kernels, with 256, 512, and 512 kernels respectively, a stride of 1, and padding of 1, used to extract fine-grained local detail features. Each convolutional layer is followed by a batch normalization layer and a ReLU activation function, with max pooling used for downsampling. The pooling window size is 2×2, and the stride is 2. Through multi-scale convolution operations, the extracted spatial feature dimension is 512×4×8, encompassing the signal's feature representation at different frequencies and time scales.

[0032] The differential attention unit (DIU) is based on a differential transform network (DTN) and suppresses attention noise through a differential self-attention mechanism, extracting global dependency features of the signal. The DTN is one of the core innovations of this invention; its core idea is to suppress high-frequency attention noise by calculating the difference between two attention distributions. Specifically, the DTN flattens the convolutional features into a sequence with a length of 32 and a feature dimension of 512. A first attention distribution and a second attention distribution are calculated for the input features, using different learnable query matrices and key matrices. The formula for calculating the first attention distribution is:

[0033] ,

[0034] in, The first attention distribution matrix, This is the first query matrix. For the first bond matrix, The dimension of the key vector is 64. For the normalized exponential function, the superscript is... This represents the matrix transpose. Second attention distribution. The calculation method and Similar, but using a different learnable parameter matrix. and .

[0035] Calculate the difference between the first attention distribution and the second attention distribution:

[0036] ,

[0037] in, For differential attention distribution, The balance coefficient is set to 0.5. Differential operations effectively suppress common noise components in the two attention distributions, highlighting the key features of the signal. Input features are weighted based on the differential attention distribution:

[0038] ,

[0039] in, This is a global dependency feature. The value matrix is ​​obtained by linear transformation of the input features. The global dependency feature has a dimension of 32×512 and contains the correlation information of the signal in the global scope. Compared with the traditional self-attention mechanism, the differential attention mechanism can more effectively capture the long-distance dependencies of the signal, while suppressing redundant attention weights and improving the discriminativeness of the feature representation.

[0040] The feature fusion unit adaptively weights and fuses spatial features and global dependency features to generate a fused feature vector. The spatial features are first subjected to global average pooling, pooling the 512×4×8 feature map into a 512-dimensional feature vector. The global dependency features are pooled using global average pooling, transforming the 32×512 feature matrix into a 512-dimensional feature vector. Adaptive weighted fusion employs a gating mechanism:

[0041] ,

[0042] ,

[0043] in, For adaptive weight vectors, For sigmoid activation function, The weight matrix has a dimension of 512×1024. For the bias vector, Indicates feature splicing, This indicates element-wise multiplication. The final fused feature vector has a dimension of 512, integrating local spatial features and global dependency features to fully characterize the multi-scale time-frequency characteristics of the electromagnetic signal. In a preferred embodiment, the fused feature vector also includes the signal's frequency domain statistical features, such as spectral energy distribution, spectral entropy, and spectral correlation. These features are extracted through parallel statistical analysis branches and concatenated into the fused feature vector, further enhancing the expressive power of the features.

[0044] The spectrum situation prediction and decision-making module 3 is connected to the multi-scale feature extraction and fusion module 2. Based on the fused feature vectors, it constructs a spatiotemporal evolution model of spectrum occupancy, predicts the spectrum situation at future times, and generates anti-interference decisions. (Refer to...) Figure 3 The spectrum situation prediction and decision-making module 3 includes a feature sequence construction unit, a temporal coding unit, an attention prediction unit, and a reinforcement learning decision-making unit.

[0045] The feature sequence construction unit organizes the fused feature vectors from historical moments into a feature sequence in chronological order. The system maintains a time window of length 20, storing the fused feature vectors from the most recent 20 moments to form the feature sequence. Each The dimension is 512. As time progresses, the window slides in a first-in, first-out manner, continuously updating the feature sequence. The feature sequence not only contains the current electromagnetic environment state, but also implies the temporal evolution trend of the spectral situation, providing rich historical information for subsequent time series modeling.

[0046] The temporal coding unit, based on a Long Short-Term Memory (LSTM) network, encodes the temporal features of the feature sequence to capture the temporal evolution of the spectral dynamics. The LSTM network employs a bidirectional structure, including forward and backward LSTMs, with a hidden layer dimension of 256. For each time step in the feature sequence... The forward LSTM calculates the forward hidden state based on the forward history information. The backward LSTM calculates the backward hidden state based on the backward history information. The bidirectional hidden states are concatenated to form a complete timing code. The dimension is 512. Long Short-Term Memory (LSTM) networks effectively mitigate the vanishing gradient problem through gating mechanisms, enabling them to capture long-term temporal dependencies. Temporal encoding incorporates the periodic patterns, abrupt changes, and evolutionary trends of spectrum occupancy, providing a reliable temporal representation for spectrum situation prediction.

[0047] The attention prediction unit, based on a multi-head attention mechanism, weights temporal features to predict the spectral occupancy state at future time steps. The multi-head attention mechanism comprises eight attention heads, each with a dimension of 64. For time steps... Timing coding Multi-head attention calculates its relevance to the encodings of all historical time points, generating attention weights:

[0048] ,

[0049] Among them, query matrix Key matrix Sum matrix Each by Obtained through linear transformation, The value is 64. Multi-head attention can capture the diverse correlations of temporal features from different representation subspaces. The output of multi-head attention is passed through a linear transformation and a feedforward neural network to generate a spectral situation prediction vector. The dimension is 128, representing a future moment. The occupancy probability distribution for each frequency band is shown. The spectrum situation prediction vector is mapped to a probability range of 0 to 1 using the sigmoid function; a larger value indicates a higher probability that the corresponding frequency band is occupied. The prediction accuracy is evaluated by comparing it with the actual observed spectrum occupancy status, achieving over 92% on the test dataset.

[0050] The reinforcement learning decision unit takes the spectrum occupancy state as the environmental state input and generates anti-interference actions based on a deep Q-network. The deep Q-network employs a dual-network structure, including an online network and a target network. The network structure is a three-layer fully connected network with 256, 256, and 128 neurons in each layer, and the activation function is ReLU. The anti-interference action space includes three dimensions: frequency selection, power adjustment, and waveform design. Frequency selection has 10 candidate frequency bands, power adjustment has 5 power levels, and waveform design has 4 modulation schemes, totaling 200 discrete actions. The reinforcement learning decision unit uses a partially observable Markov decision process model, taking into account the partial observability of the spectrum situation in the actual environment.

[0051] Specifically, the current spectrum occupancy status, historical observation information, and channel quality information are used to construct the belief state. The belief state is represented in vector form and consists of three parts: the first part is the predicted spectrum occupancy probability distribution. The first part is a 128-dimensional belief state vector; the second part is the action sequence encoding of historical moments, where the 10 most recent actions are encoded into a 128-dimensional vector using a single-layer LSTM; the third part is the channel quality metrics, including normalized values ​​of average signal-to-noise ratio, bit error rate, and latency, with a dimension of 16. These three parts are concatenated to form a 272-dimensional belief state vector. .

[0052] Based on belief state Given the current policy network parameters, calculate the Q-value of each candidate anti-interference action:

[0053] ,

[0054] in, For the first One candidate action, For the parameters of a deep Q-network, Indicates the state Next action The expected cumulative return. The action with the largest Q-value is selected as the robust decision output for the current moment:

[0055] ,

[0056] in, The optimal anti-interference action is defined as follows: The optimal action is analyzed into specific frequency selection, power adjustment, and waveform design strategies. The frequency selection strategy determines the target frequency for system operation, prioritizing frequency bands with less interference and better channel quality. The power adjustment strategy dynamically adjusts the transmit power based on interference intensity and channel fading, ensuring communication quality while avoiding energy waste and detection risks caused by excessive power. The waveform design strategy selects appropriate modulation methods and coding schemes to achieve a balance between spectral efficiency and anti-interference performance. In a preferred embodiment, the deep Q-network employs a priority experience replay technique, assigning higher sampling probabilities to important samples to accelerate network convergence; a dual deep Q-network algorithm is used to decouple action selection and value evaluation, reducing the problem of Q-value overestimation.

[0057] The adaptive anti-interference execution module 4 is connected to the spectrum situation prediction and decision-making module 3. It dynamically adjusts communication parameters based on anti-interference decisions and collects actual communication performance indicators. (Refer to...) Figure 4 The adaptive anti-interference execution module 4 includes a parameter parsing unit, a multi-domain collaborative adjustment unit, and a performance monitoring unit.

[0058] The parameter analysis unit is used to analyze the frequency selection strategy, power adjustment strategy, and waveform design strategy in anti-interference decision-making. Optimal action. The mapping table is used to convert the parameters into specific configurations: the frequency selection strategy corresponds to one of 10 candidate frequency bands, with the center frequency range of 2.4GHz to 5.8GHz and a bandwidth of 20MHz; the power adjustment strategy corresponds to 5 transmit power levels, namely 10dBm, 15dBm, 20dBm, 25dBm and 30dBm; the waveform design strategy corresponds to 4 modulation methods, namely BPSK, QPSK, 16QAM and 64QAM, and the coding scheme adopts convolutional code with a coding rate of 0.5 or 0.75.

[0059] The multi-domain coordinated adjustment unit adjusts the operating parameters of the communication system in a coordinated manner based on the parsed strategy parameters. Frequency adjustment achieves rapid frequency hopping by controlling the local oscillator frequency of the RF front-end, with a frequency switching time of less than 5ms; power adjustment dynamically controls the transmit power through a programmable gain amplifier, with a power adjustment accuracy of 0.5dB and an adjustment time of less than 1ms; waveform adjustment is achieved by reconfiguring the modulation and demodulation parameters of the baseband processor, including the symbol mapping table and channel encoder configuration, with an adjustment time of less than 2ms. Multi-domain coordinated adjustment ensures the synchronous updating of frequency, power, and waveform parameters, avoiding performance degradation caused by parameter mismatch. In a preferred embodiment, the multi-domain coordinated adjustment unit also supports dynamic bandwidth adjustment and adaptive channel coding rate, further enhancing the system's flexibility.

[0060] The performance monitoring unit collects real-time data on bit error rate (BER), signal-to-noise ratio (SNR), throughput, and latency during the communication process, generating actual communication performance metrics. The BER is calculated by comparing transmitted and received data, with a statistical window length of 10,000 symbols. The SNR is calculated as the ratio of received signal power to noise power, using an exponentially weighted moving average filter with a smoothing coefficient of 0.9. Throughput is measured in Mbps (memory per second) and the round-trip time from transmission to confirmed reception is measured in milliseconds. These performance metrics are updated in real-time at a sampling rate of 10Hz, providing real-time data support for the performance feedback and optimization module.

[0061] The performance feedback and optimization module 5 is connected to the adaptive anti-interference execution module 4 and the spectrum situation prediction and decision module 3. It calculates the decision effectiveness evaluation value based on actual communication performance indicators and feeds this evaluation value back to the spectrum situation prediction and decision module 3 for adjusting the policy parameters of the deep reinforcement learning network. (Refer to...) Figure 5 The performance feedback and optimization module 5 includes a performance evaluation unit, an error calculation unit, and a parameter update unit.

[0062] The performance evaluation unit calculates a weighted comprehensive score based on actual communication performance indicators to determine the decision performance evaluation value. The specific calculation steps are as follows: For the bit error rate... After taking the reciprocal and normalizing, the first evaluation component is obtained:

[0063] ,

[0064] in, Take the maximum tolerance value for the bit error rate. , To find the optimal value for the bit error rate The signal-to-noise ratio (SNR) is normalized to obtain the second evaluation component:

[0065] ,

[0066] in, The minimum signal-to-noise ratio is set to 0 dB. The maximum signal-to-noise ratio is set to 30dB. For throughput... After normalization, the third evaluation component is obtained:

[0067] ,

[0068] in, The minimum throughput is set to 1 Mbps. The maximum throughput is set to 100 Mbps. Regarding latency... After taking the reciprocal and normalizing, the fourth evaluation component is obtained:

[0069] ,

[0070] in, The maximum tolerance for delay is set to 100ms. The optimal delay value is 1ms. The decision effectiveness evaluation value is obtained by weighted summation of the four evaluation components.

[0071] ,

[0072] in, As a value for evaluating decision-making effectiveness, , , , These are weighting coefficients, with values ​​of 0.3, 0.2, 0.3, and 0.2 respectively. These weighting coefficients are set according to the priority requirements of different application scenarios. Decision effectiveness evaluation value. The value ranges from 0 to 1. The larger the value, the higher the efficiency of the current decision and the better the communication quality.

[0073] The error calculation unit calculates the deviation between the decision performance evaluation value and the preset target value. (Preset target value) The value is set according to the quality requirements of the communication task, and is usually between 0.8 and 0.95. The deviation is calculated as follows:

[0074] ,

[0075] in, This represents the deviation value; a positive value indicates that the current performance has not met the target and the decision-making strategy needs to be optimized.

[0076] The parameter update unit generates gradient information based on the bias and feeds this gradient information back to the deep reinforcement learning network of the spectrum situation prediction and decision module 3 to update the policy network parameters and value network parameters. The deep Q-network uses a temporal difference learning algorithm for parameter updates. For time steps... State-Action Pairs The target Q value is calculated as follows:

[0077] ,

[0078] in, The discount factor is set to 0.99. The parameters of the target network are copied from the online network every 100 steps. The loss function is defined as:

[0079] ,

[0080] in, The minimum batch size is set to 32. The gradient is calculated using the backpropagation algorithm.

[0081] ,

[0082] Parameter updates use the Adam optimizer with a learning rate of [missing information]. The estimated decay rate of the first moment is 0.9, and the estimated decay rate of the second moment is 0.999. The parameter update formula is:

[0083] ,

[0084] in, For learning rate, and These are the first-order moment estimates and second-order moment estimates of the gradient, respectively. Take the numerical stability constant as Through continuous parameter updates, deep Q-networks continuously optimize decision-making strategies, enabling the system to gradually learn the optimal anti-interference strategy in complex electromagnetic environments.

[0085] Reference Figure 6 The data flow and closed-loop feedback process of the system of this invention is as follows: the raw signal data acquired by the electromagnetic signal acquisition module 1 flows to the multi-scale feature extraction and fusion module 2; the extracted fused feature vector flows to the spectrum situation prediction and decision module 3; the generated anti-interference decision flows to the adaptive anti-interference execution module 4; the acquired actual communication performance indicators flow to the performance feedback and optimization module 5; and the calculated decision effectiveness evaluation value and gradient information are fed back to the spectrum situation prediction and decision module 3, forming a complete perception-prediction-decision-execution-feedback closed loop. This closed loop enables the system to have autonomous learning and continuous optimization capabilities, and can continuously improve decision strategies based on actual communication effects to adapt to the dynamically changing electromagnetic environment.

[0086] In one specific embodiment, the system further includes an interference type identification module connected to the multi-scale feature extraction and fusion module 2. The interference type identification module uses a multi-classification neural network to identify the type of interference signal based on the fused feature vector. Interference types include suppression interference, deception interference, and intelligent cognitive interference. Suppression interference overwhelms the target signal by transmitting high-power broadband or narrowband noise; deception interference misleads the receiver by transmitting false target signals; and intelligent cognitive interference learns the target system's communication pattern and adaptively adjusts its interference strategy. The multi-classification neural network uses a three-layer fully connected structure, with the output layer using a softmax activation function to output the probability distribution of each interference type. The identification results are input to the spectrum situation prediction and decision module 3 as auxiliary information for generating targeted anti-interference decisions. For different interference types, the system employs differentiated countermeasures: for suppression interference, frequency hopping and power control are prioritized; for deception interference, signal authentication and anomaly detection are enhanced; and for intelligent cognitive interference, a randomization strategy is used to increase the difficulty of prediction for the adversary.

[0087] Reference Figure 7 The training process of this invention includes two stages: feature extraction network training and reinforcement learning network training. In the feature extraction network training stage, signal samples and annotation information under different electromagnetic environments are collected to construct a training dataset. The training dataset contains 100,000 signal samples, covering different modulation methods, signal-to-noise ratios, and interference types. The annotation information includes signal type, modulation method, and spectrum occupancy status. Supervised learning is used to train the convolutional neural network parameters and differential attention network parameters of the multi-scale feature extraction and fusion module 2. The loss function is cross-entropy loss, the optimizer is SGD, the initial learning rate is 0.01, and it decays to 0.1 every 30 epochs, with a total training epoch count of 150. In the reinforcement learning network training stage, the feature extraction network parameters are fixed, and reinforcement learning is used to train the policy network and value network of the spectrum situation prediction and decision-making module 3. An electromagnetic environment simulator is constructed to simulate interference scenarios of different intensities and types. The reinforcement learning uses a deep Q-learning algorithm, and the exploration rate is... - A greedy strategy is used, with an initial exploration rate of 1.0, which decays linearly to 0.01 with each training step, for a total of 100,000 steps. The experience replay buffer has a capacity of 100,000 steps, and 32 samples are randomly sampled from the buffer each step for network updates. The total training steps are 1 million, with policy performance evaluated every 1000 steps. During the joint training phase, the parameters of the feature extraction network, policy network, and value network are trained iteratively and alternately. In each iteration, the policy network parameters are first fixed, and the feature extraction network is fine-tuned using new samples collected through reinforcement learning. Then, the feature extraction network parameters are fixed again, and the policy network continues to be trained. This alternating iteration is repeated for 10 rounds until the overall system performance converges. The final trained system, in the test environment, achieves a 92% accuracy rate in spectral situation prediction, an average reward of 0.85 for anti-interference decisions, and a 50% improvement in overall communication reliability compared to the benchmark method.

[0088] In practical applications, the system is deployed on wireless communication terminals or base stations, operating in real time for electromagnetic environment perception and anti-interference decision-making. The system's processing latency is less than 50ms, meeting real-time communication requirements. The system supports an online learning mode, continuously collecting new sample data and periodically updating network parameters during operation, enabling the system to adapt to new electromagnetic environment characteristics and interference patterns. In applications such as military communication, emergency communication, and unmanned system communication, this invention significantly improves communication reliability and anti-interference capabilities, ensuring the effective transmission of critical information.

[0089] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An intelligent electromagnetic environment perception and anti-jamming system, characterized in that , comprising: an electromagnetic signal acquisition module, configured to acquire electromagnetic signals in a target frequency band and generate multi-dimensional original signal data; a multi-scale feature extraction and fusion module, connected with the electromagnetic signal acquisition module, configured to perform multi-scale time-frequency feature extraction and feature fusion processing on the original signal data based on a deep convolutional neural network and a differential transformation network, to generate a fusion feature vector, the fusion feature vector containing frequency domain features, time domain features and modulation features of the signal; a spectrum situation prediction and decision module, connected with the multi-scale feature extraction and fusion module, configured to construct a spectrum occupation space-time evolution model according to the fusion feature vector based on a long short-term memory network and an attention mechanism, to predict a spectrum situation at a future time, and to generate an anti-interference decision based on a deep reinforcement learning network, the anti-interference decision including a frequency selection strategy, a power adjustment strategy and a waveform design strategy; an adaptive anti-interference execution module, connected with the spectrum situation prediction and decision module, configured to perform dynamic adjustment of communication parameters according to the anti-interference decision, and to acquire actual communication performance indicators; a performance feedback and optimization module, connected with the adaptive anti-interference execution module and the spectrum situation prediction and decision module, configured to calculate a decision performance evaluation value according to the actual communication performance indicators, and to feed back the decision performance evaluation value to the spectrum situation prediction and decision module, to adjust strategy parameters of the deep reinforcement learning network.

2. The intelligent electromagnetic environment perception and anti-jamming system according to claim 1, characterized in that The multi-scale feature extraction and fusion module comprises: a time-frequency transformation unit, configured to perform short-time Fourier transform on the original signal data to generate a time-frequency spectrum; a convolution feature extraction unit, configured to perform multi-scale convolution operation on the time-frequency spectrum based on a multi-layer convolutional neural network, to extract spatial features of different scales; a differential attention unit, configured to suppress attention noise by a differential self-attention mechanism based on a differential transformation network, to extract global dependency feature of the signal; a feature fusion unit, configured to adaptively weight and fuse the spatial features and the global dependency feature to generate the fusion feature vector.

3. The intelligent electromagnetic environment perception and anti-jamming system according to claim 2, characterized in that The differential attention unit realizes the differential self-attention mechanism in the following manner: first attention distribution and second attention distribution are calculated for input features; a differential value of the first attention distribution and the second attention distribution is calculated; input features are weighted based on the differential value, to suppress high-frequency attention noise and highlight key feature components of the signal.

4. The intelligent electromagnetic environment perception and anti-jamming system according to claim 1, characterized in that The spectrum situation prediction and decision module comprises: a feature sequence construction unit, configured to organize fusion feature vectors at historical time into a feature sequence in chronological order; a time sequence encoding unit, configured to perform time sequence feature encoding on the feature sequence based on a long short-term memory network, to capture time evolution law of the spectrum situation; an attention prediction unit, configured to perform weighting processing on the time sequence feature based on a multi-head attention mechanism, to predict a spectrum occupation state at a future time; a reinforcement learning decision unit, configured to input the spectrum occupation state as an environment state, to generate an anti-interference action based on a deep Q network, the anti-interference action corresponding to strategy parameters of the anti-interference decision.

5. The intelligent electromagnetic environment perception and anti-jamming system according to claim 4, characterized in that The reinforcement learning decision unit models a partially observable Markov decision process: The current spectrum occupation state, historical observation information and channel quality information are constructed into a belief state; Based on the belief state and current policy network parameters, the Q values of each candidate anti-jamming action are calculated; The anti-jamming action with the maximum Q value is selected as the anti-jamming decision output at the current time.

6. The intelligent electromagnetic environment perception and anti-jamming system according to claim 1, characterized in that The adaptive anti-jamming execution module includes: A parameter analysis unit for analyzing the frequency selection strategy, power adjustment strategy and waveform design strategy in the anti-jamming decision; A multi-domain cooperative adjustment unit for selecting a target operating frequency according to the frequency selection strategy, determining a transmission power level according to the power adjustment strategy, and determining a modulation mode and coding scheme according to the waveform design strategy; A performance monitoring unit for collecting bit error rate, signal-to-noise ratio, throughput and latency data in real time during the communication process to generate the actual communication performance indicators.

7. The intelligent electromagnetic environment perception and anti-jamming system of claim 1, wherein The performance feedback and optimization module includes: An efficiency evaluation unit for calculating a weighted comprehensive score based on the actual communication performance indicators to determine the decision efficiency evaluation value; An error calculation unit for calculating the deviation between the decision efficiency evaluation value and a preset target value; A parameter updating unit for generating gradient information based on the deviation, feeding the gradient information back to the deep reinforcement learning network of the spectrum situation prediction and decision module, and updating the policy network parameters and value network parameters.

8. The intelligent electromagnetic environment perception and anti-jamming system according to claim 7, characterized in that The efficiency evaluation unit calculates the decision efficiency evaluation value in the following manner: The bit error rate is taken as the reciprocal and normalized to obtain a first evaluation component; The signal-to-noise ratio is normalized to obtain a second evaluation component; The throughput is normalized to obtain a third evaluation component; The latency is taken as the reciprocal and normalized to obtain a fourth evaluation component; The first evaluation component, the second evaluation component, the third evaluation component and the fourth evaluation component are weighted and summed to obtain the decision efficiency evaluation value.

9. The intelligent electromagnetic environment perception and anti-jamming system of claim 1, wherein The multi-scale feature extraction and fusion module and the spectrum situation prediction and decision module are jointly trained in the training stage in the following manner: Signal samples and labeled information under different electromagnetic environments are collected to construct a training data set; The convolutional neural network parameters of the multi-scale feature extraction and fusion module are trained using supervised learning; The policy network and value network of the spectrum situation prediction and decision module are trained using reinforcement learning while the convolutional neural network parameters are fixed; The convolutional neural network parameters, the policy network parameters and the value network parameters are alternately iteratively trained until the overall system performance converges.

10. The intelligent electromagnetic environment perception and anti-jamming system of claim 1, wherein The system further includes an interference type identification module connected to the multi-scale feature extraction and fusion module, which identifies the type of interference signal based on the fusion feature vector, the type of interference signal including suppressive jamming, deceptive jamming and intelligent cognitive jamming, and the identification result is input to the spectrum situation prediction and decision module to assist in generating targeted anti-jamming decisions.

Images