Artificial intelligence based communication protocol and interference rejection system

Abstract

This invention relates to the field of communication technology and discloses an artificial intelligence-based communication protocol and anti-interference system. The protocol includes: real-time acquisition of electromagnetic environment data; multi-scale temporal fusion coding to generate a high-dimensional state vector; outputting communication mode categories and continuous waveform parameters through a hierarchical policy network; jointly training the policy network using a progressive experience replay pool guided by course learning; and dynamically reconstructing physical layer waveforms and adjusting radio frequency parameters. The system includes modules for electromagnetic environment perception, multi-scale state coding, hierarchical policy decision-making, course learning training, and adaptive waveform reconstruction. This invention significantly improves the global optimality of the policy and anti-interference communication performance through hierarchical action modeling, multi-scale feature fusion, and a progressive training mechanism.

Description

Technical Field

[0001] This invention belongs to the field of communication technology, specifically relating to communication protocols and anti-interference systems based on artificial intelligence. Background Technology

[0002] With the deep deployment of wireless communication systems in scenarios such as military confrontation, emergency rescue, and high-density Internet of Things, the electromagnetic environment they face is becoming increasingly complex and highly dynamic. Under conditions of dense coexistence of malicious interference sources and extremely crowded legitimate signal spectrum, traditional anti-jamming communication mechanisms rely on preset rules or static strategy libraries for parameter adjustment, making it difficult to adapt to the rapidly changing interference situation.

[0003] Such methods are typically based on fixed waveform templates, pre-allocated frequency points, or empirical power control. They lack real-time perception and intelligent decision-making capabilities when facing unknown or adaptive interference, which can easily lead to communication link interruptions, information leaks, or resource waste, severely restricting the availability of high-reliability communication in mission-critical scenarios.

[0004] Intelligent anti-jamming methods based on reinforcement learning (RL) have attracted widespread attention in recent years. These methods learn optimal communication strategies through interaction between the agent and the environment, and theoretically can achieve dynamic optimization without the need for a prior interference model. However, in extremely complex electromagnetic environments, the action space is composed of multi-dimensional parameters such as selectable waveforms, frequency hopping points, transmission power, and coding methods, and its dimensionality grows exponentially.

[0005] Classical RL algorithms are limited by problems such as low exploration efficiency and large variance in policy gradient estimation. They often converge slowly and are prone to getting trapped in local optima. They cannot find a globally robust anti-jamming communication strategy within a limited time and are difficult to meet the operational or emergency communication needs under the dual constraints of real-time performance and reliability.

[0006] While introducing deep neural networks enhances the expressive power of strategies, it remains confined to classical computational frameworks and fails to overcome the search bottleneck caused by combinatorial explosion. Furthermore, existing solutions primarily focus on parameter selection, lacking the ability to intelligently generate communication waveforms themselves, and are unable to proactively construct novel signal patterns with low detectability or orthogonality to interference features in the time, frequency, and code domains. Especially when countering intelligent jammers with learning capabilities, fixed or finite sets of waveforms are easily identified, predicted, and targeted for suppression. Summary of the Invention

[0007] This invention provides an artificial intelligence-based communication protocol and anti-interference system, aiming to solve the technical problem that in extremely complex electromagnetic environments, the action space dimension formed by the combination of selectable waveforms, frequencies, and transmission power is extremely high, resulting in slow convergence speed and easy getting trapped in local optima of classical reinforcement learning algorithms, thus making it impossible to generate a globally optimal anti-interference communication strategy.

[0008] To overcome the above-mentioned shortcomings, this invention constructs an intelligent anti-interference communication system architecture that integrates hierarchical action space modeling, multi-scale state-aware encoding, and a policy optimization mechanism based on course learning guidance. By structurally compressing the high-dimensional action space, dynamically extracting key environmental features, and introducing a progressive training strategy, it significantly improves the efficiency of policy search and the ability to guarantee global optimality.

[0009] This invention provides an artificial intelligence-based communication protocol and anti-interference method, which includes: Real-time acquisition of spectrum sensing data, channel status information, and interference source characteristic parameters of the electromagnetic environment in which the current communication link is located; Multi-scale temporal fusion encoding is performed on the spectrum sensing data, channel state information and interference source characteristic parameters to generate a high-dimensional state vector representing the current electromagnetic situation. The high-dimensional state vector is input into a hierarchical policy network, which includes a high-level policy sub-network and a low-level execution sub-network. The high-level policy sub-network outputs a discretized communication mode category, and the low-level execution sub-network generates specific waveform parameters, center frequency offset, and transmit power value in the continuous domain according to the communication mode category. The hierarchical policy network is jointly trained using a progressive experience replay pool built based on a course learning mechanism. The progressive experience replay pool stores and prioritizes the sampling of historical interaction samples according to the task difficulty level. The task difficulty level is determined by interference density, signal-to-noise ratio threshold, and spectrum occupancy rate. Based on the specific waveform parameters, center frequency offset, and transmit power value output by the hierarchical strategy network, the physical layer communication waveform is dynamically reconstructed and the radio frequency front-end parameters are adjusted to complete adaptive anti-interference communication.

[0010] Furthermore, the real-time acquisition of spectrum sensing data, channel state information, and interference source characteristic parameters of the electromagnetic environment of the current communication link specifically includes: acquiring the signal energy distribution map of the entire frequency band through a broadband spectrum sensing module at a scanning rate of no less than 100,000 scans per second; sending pilot sequences and receiving return signals through a channel detection module to calculate the channel impulse response, multipath delay spread, and coherence bandwidth; and performing modulation type identification, center frequency estimation, bandwidth measurement, and direction of arrival angle calculation on the detected non-cooperative signals through an interference feature extraction module to form an interference source fingerprint database.

[0011] Furthermore, the step of performing multi-scale temporal fusion encoding on the spectrum sensing data, channel state information, and interference source characteristic parameters to generate a high-dimensional state vector representing the current electromagnetic situation specifically includes: dividing the spectrum energy distribution map into multiple sub-bands, performing sliding window statistics on the mean, variance, and rate of change of energy in each sub-band to form a first-level feature sequence; performing multi-scale filtering on the channel impulse response sequence through a 1D convolutional kernel to extract fading characteristics at different time scales to form a second-level feature sequence; aligning the interference source parameters in the interference source fingerprint database by timestamp, and using a graph neural network to model the spatiotemporal correlation between interference sources to generate a third-level graph embedding feature; concatenating the first-level feature sequence, the second-level feature sequence, and the third-level graph embedding feature in the channel dimension, and inputting them into a bidirectional long short-term memory network for context-dependent modeling, finally outputting a high-dimensional state vector with a fixed dimension.

[0012] Furthermore, in the hierarchical policy network, the high-level policy sub-network adopts a classification head structure, and its output is the elements in a predefined set of communication mode categories, including frequency hopping mode, spread spectrum mode, orthogonal frequency division multiplexing mode, chaotic modulation mode, and cognitive radio mode; the low-level execution sub-network configures an independent parameter generator for each communication mode category. Each parameter generator is composed of a fully connected neural network, which receives the high-dimensional state vector as input and outputs the waveform shaping filter coefficients required for the corresponding mode, the offset of the center frequency point relative to the nominal value, and the gain factor of the transmit power relative to the reference power.

[0013] Furthermore, the progressive experience replay pool constructed based on the course learning mechanism specifically includes: setting the initial training stage to only include interactive samples in low interference density scenarios, where the number of interference sources does not exceed 3, the signal-to-noise ratio is not lower than -5 dB, and the spectrum occupancy rate is lower than 40%; as the training rounds increase, medium interference density scenario samples are gradually introduced, with the number of interference sources ranging from 4 to 8, the signal-to-noise ratio ranging from -15 dB to -5 dB, and the spectrum occupancy rate ranging from 40% to 70%; in the later training stage, sampling permissions for high interference density scenario samples are opened, with the number of interference sources greater than 8, the signal-to-noise ratio lower than -15 dB, and the spectrum occupancy rate higher than 70%; all samples are labeled with their difficulty level when stored in the experience replay pool, and each time the policy is updated, priority sampling is performed according to the highest difficulty level allowed in the current training stage, ensuring that the policy network gradually improves its generalization ability under the premise of controllable complexity.

[0014] Furthermore, the dynamic reconstruction of the physical layer communication waveform and adjustment of the RF front-end parameters specifically includes: configuring the pulse shaping filter in the baseband signal processing unit according to the waveform shaping filter coefficients output by the underlying execution sub-network; controlling the frequency synthesizer of the local oscillator to output the corresponding carrier frequency according to the center frequency offset; adjusting the gain control voltage of the power amplifier according to the transmit power gain factor so that the actual output power accurately matches the target value; after completing the above configuration, starting the data frame transmission process, and synchronously recording the success rate, bit error rate and throughput of this communication as reward signals for subsequent strategy evaluation.

[0015] This invention provides an artificial intelligence-based communication protocol and anti-interference system, which includes: The electromagnetic environment sensing module is used to collect spectrum sensing data, channel status information, and interference source characteristic parameters of the electromagnetic environment in which the current communication link is located in real time. The multi-scale state coding module is used to perform multi-scale time-series fusion coding on the spectrum sensing data, channel state information and interference source characteristic parameters to generate a high-dimensional state vector representing the current electromagnetic situation. The hierarchical strategy decision module is used to input the high-dimensional state vector into the hierarchical strategy network and output specific waveform parameters, center frequency offset and transmit power value. The course learning and training module is used to jointly train the hierarchical policy network based on a progressive experience replay pool; The adaptive waveform reconstruction module is used to dynamically reconstruct the physical layer communication waveform and adjust the radio frequency front-end parameters based on the specific waveform parameters, center frequency offset, and transmit power value.

[0016] Furthermore, the electromagnetic environment sensing module includes a broadband spectrum sensing unit, a channel detection unit, and an interference feature extraction unit. The broadband spectrum sensing unit adopts a superheterodyne receiver architecture, possessing an instantaneous bandwidth covering 300 MHz to 6 GHz, and a scanning step accuracy of 100 kHz. The channel detection unit has a built-in pseudo-random noise sequence generator, supports a maximum detection bandwidth of 200 MHz, and has a channel estimation update period of no more than 10 milliseconds. The interference feature extraction unit integrates multiple modulation recognition algorithms, including high-order cumulant analysis, cyclic stationary feature detection, and deep residual classification networks, which can identify no fewer than 15 common communication and radar signal modulation types.

[0017] Furthermore, the multi-scale state coding module includes a sub-band statistics unit, a multi-scale channel feature extraction unit, an interference graph embedding unit, and a context fusion unit. The sub-band statistics unit divides the input spectrogram into 64 sub-bands at equal intervals, and calculates the mean energy, standard deviation, and first-order difference mean for each sub-band within the last 500 milliseconds. The multi-scale channel feature extraction unit is configured with three parallel 1D convolutional branches with kernel lengths of 8, 16, and 32 sampling points, respectively, to capture short-term, medium-term, and long-term channel fading modes. The interference graph embedding unit treats each interference source as a graph node, constructs edge weights based on spatial distance and frequency proximity, generates node embedding vectors through three layers of graph convolutional layers, and obtains a graph-level representation through global average pooling. The context fusion unit uses a bidirectional long short-term memory network with a hidden layer dimension of 512 and an output dimension of 1024.

[0018] Furthermore, the hierarchical policy decision-making module includes a high-level policy sub-network and a low-level execution sub-network. The high-level policy sub-network consists of three fully connected layers, with 1024, 512, and 5 neurons in each layer, corresponding to five communication mode categories. The low-level execution sub-network contains five independent parameter generators, each consisting of four fully connected layers, with output dimensions of 256 for waveform filter coefficients, 1 for frequency offset, and 1 for power gain factor. Each parameter generator is activated and participates in forward propagation only when its corresponding communication mode is selected by the high-level policy.

[0019] Furthermore, the course learning and training module includes a difficulty grading unit, a progressive storage unit, and a priority sampling unit. The difficulty grading unit classifies each empirical sample into three difficulty levels: low, medium, and high, based on the total number of interference sources, the lowest signal-to-noise ratio, and the maximum spectral occupancy rate. The progressive storage unit maintains three independent first-in-first-out queues, storing samples of different difficulty levels, with each queue having a capacity of 500,000 samples. In the nth training cycle, if n is less than 1 / 3 of the total number of cycles, the priority sampling unit samples only from the low-difficulty queue. If n is between 1 / 3 and 2 / 3 of the total number of cycles, it samples from the low- and medium-difficulty queues in a 7:3 ratio. If n is greater than 2 / 3 of the total number of cycles, it samples from the low-, medium-, and high-difficulty queues in a 3:4:3 ratio.

[0020] Furthermore, the adaptive waveform reconstruction module includes a baseband waveform configuration unit, a frequency synthesis control unit, and a power adjustment unit. The baseband waveform configuration unit receives waveform filter coefficients and loads them into a finite impulse response filter core in a field-programmable gate array (FPGA). The filter has 256 taps and a clock frequency of 200 MHz. The frequency synthesis control unit sends control words to the phase-locked loop frequency synthesizer via a serial peripheral interface, achieving a frequency tuning resolution better than 1 Hz. The power adjustment unit outputs an analog voltage from 0 to 3.3 volts via a digital-to-analog converter (DAC) to control the gain of the power amplifier, with an adjustment range covering -30 to +20 dBm and a step accuracy of 0.5 dB.

[0021] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention constructs a hierarchical policy network, which decomposes the high-dimensional continuous-discrete hybrid action space into two levels: a high-level discrete mode selection and a low-level continuous parameter generation. This effectively reduces the dimensionality curse of policy search and avoids the slow convergence problem caused by the low exploration efficiency of classic reinforcement learning algorithms in a huge action space.

[0022] 2. The multi-scale state coding mechanism integrates the spatiotemporal dynamic characteristics of spectrum, channel and interference, and models the correlation between interference sources through graph neural networks, which significantly improves the ability to represent complex electromagnetic situations and provides a more comprehensive and accurate environmental context for policy decisions.

[0023] 3. The progressive training mechanism guided by the course learning, through difficulty grading and priority sampling, enables the policy network to focus on solvable subproblems in the early stage of training and gradually transition to extremely complex scenarios, fundamentally avoiding training collapse or local optimum traps caused by poor initial policy performance.

[0024] 4. The adaptive waveform reconstruction module implements end-to-end closed-loop control from strategy output to physical layer parameters, ensuring the executability and real-time performance of the anti-interference strategy. Combining the above technical means, this invention can stably and efficiently generate globally near-optimal anti-interference communication strategies in extreme environments of dense malicious interference and extremely congested spectrum, significantly improving the connectivity, throughput, and anti-interception capability of communication links, and resolving the fundamental contradiction of existing technologies in balancing global strategy optimality and real-time response capability. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the overall technical solution architecture of the communication protocol and anti-interference system based on artificial intelligence proposed in this invention; Figure 2 This is a schematic diagram of the core principle framework of the hierarchical policy network in this invention; Figure 3 This is a logical flowchart of the multi-scale state perception and encoding in this invention; Figure 4 This is a logical flowchart of the progressive experience playback training mechanism for course learning guidance in this invention. Figure 5 This is a schematic diagram of the multi-source information acquisition and feature extraction framework of the electromagnetic environment sensing module in this invention; Figure 6 This is a schematic diagram of the multi-level interaction relationship and data flow between adaptive waveform reconstruction and closed-loop control of radio frequency parameters in this invention. Detailed Implementation

[0026] Please refer to the attached document. Figures 1 to 6 This invention provides an artificial intelligence-based communication protocol and anti-interference method. Its core lies in achieving efficient, globally near-optimal anti-interference communication control in extremely complex electromagnetic environments by structurally modeling a high-dimensional action space, fusing environmental state information across multiple scales, and introducing a curriculum learning mechanism to guide policy network training. The method includes the following steps: S1 collects spectrum sensing data, channel status information, and interference source characteristic parameters of the electromagnetic environment in which the current communication link is located in real time. S2, perform multi-scale time-series fusion encoding on the spectrum sensing data, channel state information and interference source characteristic parameters to generate a high-dimensional state vector representing the current electromagnetic situation; S3, the high-dimensional state vector is input to the hierarchical policy network, which includes a high-level policy sub-network and a low-level execution sub-network. The high-level policy sub-network outputs a discretized communication mode category, and the low-level execution sub-network generates specific waveform parameters, center frequency offset and transmit power value in the continuous domain according to the communication mode category. S4, the hierarchical policy network is jointly trained using a progressive experience replay pool constructed based on the course learning mechanism. The progressive experience replay pool stores and prioritizes the historical interaction samples according to the task difficulty level. The task difficulty level is determined by interference density, signal-to-noise ratio threshold and spectrum occupancy rate. S5. Based on the specific waveform parameters, center frequency offset, and transmit power value output by the hierarchical strategy network, dynamically reconstruct the physical layer communication waveform and adjust the radio frequency front-end parameters to complete adaptive anti-interference communication.

[0027] In step S1, spectrum sensing data, channel state information, and interference source characteristic parameters of the electromagnetic environment of the current communication link are acquired in real time. This step is completed collaboratively by three parallel functional units: a broadband spectrum sensing unit, a channel detection unit, and an interference feature extraction unit. The broadband spectrum sensing unit adopts a superheterodyne receiver architecture, possessing an instantaneous bandwidth covering 300 MHz to 6 GHz, a scanning step accuracy of 100 kHz, and acquires the signal energy distribution spectrum across the entire frequency band at a scanning rate of no less than 100,000 scans per second. This spectrum is output in the form of a three-dimensional time-frequency-energy tensor, with a time dimension sampling interval of 10 microseconds and a frequency dimension divided into 57,000 frequency points, each corresponding to an energy value in dBm. The channel detection unit has a built-in pseudo-random noise sequence generator, supports a maximum detection bandwidth of 200 MHz, and periodically transmits a pilot sequence of 1024 points. After the receiver synchronously captures the returned signal, it calculates the channel impulse response using a minimum mean square error algorithm, and then derives the multipath delay spread and coherence bandwidth.

[0028] The channel estimation update period is no more than 10 milliseconds to ensure the timeliness of state information. The interference feature extraction unit performs modulation type identification, center frequency estimation, bandwidth measurement, and direction-of-arrival (AOA) calculation on detected non-cooperative signals. This unit integrates three modulation identification algorithms: high-order cumulant analysis, cyclostationary feature detection, and deep residual classification network. It can identify no fewer than 15 common communication and radar signal modulation types, including BPSK, QPSK, 16QAM, 64QAM, OFDM, LFM, and FSK. Each identified interference source is assigned a unique identifier, and its modulation type, center frequency, bandwidth, power spectral density, AOA, and first appearance timestamp are recorded, forming a structured interference source fingerprint database. All acquired data has a unified timestamp, and the time synchronization accuracy is better than 1 microsecond, guaranteed by a high-stability crystal oscillator and the IEEE 1588 precision time protocol.

[0029] In step S2, the spectrum sensing data, channel state information, and interference source characteristic parameters are subjected to multi-scale temporal fusion encoding to generate a high-dimensional state vector representing the current electromagnetic situation. This process is executed sequentially by four sub-modules: a sub-band statistics unit, a multi-scale channel feature extraction unit, an interference map embedding unit, and a context fusion unit. The sub-band statistics unit divides the input spectrum energy distribution map into 64 sub-bands at equal intervals, each with a bandwidth of 92.1875 MHz. For each sub-band, three statistics are calculated within the most recent 500 millisecond window: energy mean, standard deviation, and first-order difference mean.

[0030] The mean energy reflects the average activity level of the sub-band, the standard deviation characterizes the intensity of energy fluctuations, and the first-order difference mean indicates the energy change trend. This generates 64 sub-bands with their respective three-dimensional features, forming the first-level feature sequence with a dimension of 64×3. The multi-scale channel feature extraction unit is configured with three parallel 1D convolutional branches, used to capture short-time, medium-time, and long-time channel fading modes, respectively. The kernel lengths of the three branches are 8, 16, and 32 sampling points, corresponding to time scales of approximately 80 nanoseconds, 160 nanoseconds, and 320 nanoseconds. The input is a channel impulse response sequence with a length of 1024 points. Each branch outputs a feature map, which, after global max pooling, yields a fixed-length vector. These three vectors are concatenated to form the second-level feature sequence with a dimension of 384. The interference map embedding unit treats each interference source in the interference source fingerprint database as a node in a graph neural network.

[0031] The node feature vector consists of a one-hot encoding of the modulation type (15-dimensional), a normalized center frequency (1-dimensional), a normalized bandwidth (1-dimensional), a power spectral density (1-dimensional), and an angle of arrival (1-dimensional), totaling 19 dimensions. Edge weights are determined by two factors: spatial distance and frequency proximity. If the difference in the angle of arrival between two interfering sources is less than 10 degrees and the difference in their center frequencies is less than 5 MHz, an edge connection is established, with the weight being the weighted sum of the reciprocals of their frequency and angle differences. This is then processed through three graph convolutional layers, with output dimensions of 64, 128, and 256 for each layer, respectively.

[0032] Finally, through global average pooling, all node embedding vectors are aggregated into a single graph-level representation, namely the third-level graph embedding feature, with a dimension of 256. The context fusion unit receives the above three levels of features: the first-level feature sequence (64×3), the second-level feature sequence (384-dimensional vector), and the third-level graph embedding feature (256-dimensional vector). First, the first-level feature sequence is flattened into a 192-dimensional vector, and then concatenated with the second and third-level features along the channel dimension to form a fusion input vector with a total dimension of 832. This vector is input into a bidirectional long short-term memory network, whose forward and backward hidden layers each have a dimension of 512. The output, after concatenation, is 1024-dimensional. This output is the final high-dimensional state vector, used for subsequent policy decisions.

[0033] In step S3, the high-dimensional state vector is input to the hierarchical policy network, which outputs specific waveform parameters, center frequency offset, and transmit power value. The hierarchical policy network consists of a high-level policy sub-network and a low-level execution sub-network. The high-level policy sub-network adopts a classification head structure, consisting of three fully connected layers with an input dimension of 1024, 1024 neurons in the first hidden layer, 512 neurons in the second hidden layer, and 5 neurons in the output layer, corresponding to five predefined communication mode categories: frequency hopping mode, spread spectrum mode, orthogonal frequency division multiplexing mode, chaotic modulation mode, and cognitive radio mode.

[0034] The output layer uses the Softmax activation function to generate probability distributions for five categories, selecting the category with the highest probability as the current communication mode. The underlying execution subnetwork contains five independent parameter generators, each dedicated to a specific communication mode. Each parameter generator consists of four fully connected layers, with inputs being high-dimensional state vectors (1024 dimensions). The number of neurons in each hidden layer is 512, 256, and 180, respectively, and the output layer dimension is set according to the mode requirements.

[0035] For frequency hopping mode, the output consists of waveform shaping filter coefficients (256 dimensions), center frequency offset (1 dimension, in Hertz, ranging from -50 MHz to +50 MHz), and transmit power gain factor (1 dimension, in dB, ranging from -10 to +10). Other modes follow the same logic, only the waveform parameter dimensions may differ. Each parameter generator is activated and participates in forward propagation only when its corresponding communication mode is selected by the higher-level policy; the remaining generators remain dormant and do not consume computational resources. This design ensures a structured decomposition of the action space, avoiding the curse of dimensionality that can occur with direct optimization in a continuous-discrete hybrid space.

[0036] In step S4, the hierarchical policy network is jointly trained using a progressive experience replay pool constructed based on a course learning mechanism. This mechanism is implemented collaboratively by a difficulty grading unit, a progressive storage unit, and a priority sampling unit. The difficulty grading unit evaluates each experience sample (including state, action, reward, and next state) based on three metrics: the total number of interference sources N, and the lowest signal-to-noise ratio. Maximum spectrum utilization .

[0037] If N≤3, ≥-5dB If the percentage is ≤40%, it is classified as a low difficulty level; If 4≤N≤8, -15dB≤ <-5dB, 40% If the score is ≤70%, it is considered a medium difficulty level. If N>8, <-15dB A score of >70% indicates a high level of difficulty.

[0038] The incremental storage unit maintains three independent first-in-first-out queues, each with a capacity of 500,000 entries, storing low, medium, and high difficulty samples respectively. The priority sampling unit is based on the current training epoch n and the total training epochs. The ratio determines the sampling strategy. If If so, sampling will only be performed from the low-difficulty queue; if Then, samples are mixed from the low-difficulty and medium-difficulty queues in a 7:3 ratio; if Then, samples are mixed from low, medium, and high difficulty queues in a 3:4:3 ratio. Each policy update uses experience samples with a batch size of 256. The reward signal is calculated based on the communication success rate, bit error rate, and throughput, using the following formula: in, This is a function that indicates successful communication (1 for success, 0 for failure). For bit error rate, For actual throughput, For the theoretical maximum throughput, =0.5、 =0.3、 =0.2 is the weight coefficient. The policy network is updated using the proximal policy optimization algorithm, with an initial learning rate of 0.0003 that decays exponentially with each training epoch.

[0039] In step S5, based on the specific waveform parameters, center frequency offset, and transmit power value output by the hierarchical strategy network, the physical layer communication waveform is dynamically reconstructed and the RF front-end parameters are adjusted. This process is executed by the baseband waveform configuration unit, the frequency synthesis control unit, and the power adjustment unit. After receiving the waveform filter coefficients, the baseband waveform configuration unit loads them into the finite impulse response filter core in the field-programmable gate array.

[0040] The filter has 256 taps, operates at a clock frequency of 200 MHz, and supports real-time coefficient updates. The frequency synthesis control unit sends a 32-bit control word to the phase-locked loop frequency synthesizer via a serial peripheral interface. The control word is calculated based on the center frequency offset, achieving a frequency tuning resolution better than 1 Hz and a switching time of less than 100 microseconds.

[0041] The power regulation unit outputs an analog voltage of 0 to 3.3 volts via a 12-bit digital-to-analog converter. This voltage serves as the gain control signal for the power amplifier, with an adjustment range of -30 to +20 dBm and a step accuracy of 0.5 dB. After completing the above configuration, the system initiates the data frame transmission process and records the performance indicators of this communication as a reward signal for the next round of strategy evaluation, forming a closed-loop control.

[0042] The AI-based communication protocol and anti-interference system includes an electromagnetic environment sensing module, a multi-scale state coding module, a hierarchical strategy decision-making module, a course learning and training module, and an adaptive waveform reconstruction module. The electromagnetic environment sensing module integrates a broadband spectrum sensing unit, a channel detection unit, and an interference feature extraction unit to achieve synchronous acquisition of multi-source heterogeneous data.

[0043] The multi-scale state encoding module generates high-dimensional state vectors through sub-band statistics, multi-scale convolution, graph embedding, and bidirectional temporal modeling. The hierarchical policy decision module deploys high-level policy sub-networks and low-level execution sub-networks to achieve hierarchical decoupling of the action space. The course learning and training module ensures the stability and convergence of policy training through difficulty grading and progressive sampling.

[0044] The adaptive waveform reconstruction module accurately maps digital policy commands to analog RF parameters, ensuring the physical executability of the policy. All modules are interconnected via a high-speed data bus, with a status update cycle of no more than 20 milliseconds, meeting real-time anti-interference communication requirements.

Claims

1. A communication protocol based on artificial intelligence, characterized in that, include: Real-time acquisition of spectrum sensing data, channel status information, and interference source characteristic parameters of the electromagnetic environment in which the current communication link is located; Multi-scale temporal fusion encoding is performed on the spectrum sensing data, channel state information and interference source characteristic parameters to generate a high-dimensional state vector representing the current electromagnetic situation. The high-dimensional state vector is input into a hierarchical policy network, which includes a high-level policy sub-network and a low-level execution sub-network. The high-level policy sub-network outputs a discretized communication mode category, and the low-level execution sub-network generates specific waveform parameters, center frequency offset, and transmit power value in the continuous domain according to the communication mode category. The hierarchical policy network is jointly trained using a progressive experience replay pool built based on a course learning mechanism. The progressive experience replay pool stores and prioritizes the sampling of historical interaction samples according to the task difficulty level. The task difficulty level is determined by interference density, signal-to-noise ratio threshold, and spectrum occupancy rate. Based on the specific waveform parameters, center frequency offset, and transmit power value output by the hierarchical strategy network, the physical layer communication waveform is dynamically reconstructed and the radio frequency front-end parameters are adjusted to complete adaptive anti-interference communication.

2. The artificial intelligence-based communication protocol according to claim 1, characterized in that, Real-time acquisition of spectrum sensing data, channel state information, and interference source characteristic parameters of the current communication link's electromagnetic environment, including: The signal energy distribution map of the entire frequency band is obtained by a broadband spectrum sensing module at a scanning rate of no less than 100,000 times per second; By sending pilot sequences and receiving feedback signals through the channel sounding module, the channel impulse response, multipath delay spread, and coherence bandwidth are calculated. The interference feature extraction module identifies the modulation type, estimates the center frequency, measures the bandwidth, and calculates the direction of arrival angle of the detected non-cooperative signals to form an interference source fingerprint database.

3. The artificial intelligence-based communication protocol according to claim 2, characterized in that, The spectrum sensing data, channel state information, and interference source characteristic parameters are subjected to multi-scale time-series fusion encoding to generate a high-dimensional state vector representing the current electromagnetic situation, including: The spectral energy distribution map is divided into multiple sub-bands. The mean, variance and rate of change of energy in each sub-band are statistically analyzed using a sliding window to form a first-level feature sequence. The channel impulse response sequence is filtered at multiple scales using a 1D convolution kernel to extract fading characteristics at different time scales, forming a second-level feature sequence. After aligning the parameters of each interference source in the interference source fingerprint database according to the timestamp, a graph neural network is used to model the spatiotemporal correlation between interference sources and generate third-level graph embedding features. The first-level feature sequence, the second-level feature sequence, and the third-level graph embedding feature are concatenated along the channel dimension and input into a bidirectional long short-term memory network for context dependency modeling, ultimately outputting a high-dimensional state vector with fixed dimensions.

4. The artificial intelligence-based communication protocol according to claim 3, characterized in that, The spectral energy distribution map is divided into multiple sub-bands. The mean, variance, and rate of change of energy in each sub-band are statistically analyzed using a sliding window to form a first-level feature sequence, including: dividing the input spectrum map into 64 sub-bands at equal intervals; For each sub-band, the mean energy, standard deviation, and first-order difference mean within the most recent 500-millisecond window are calculated to form the three-dimensional features corresponding to each of the 64 sub-bands, resulting in a first-level feature sequence with a dimension of 192.

5. The artificial intelligence-based communication protocol according to claim 4, characterized in that, The channel impulse response sequence is filtered at multiple scales using a 1D convolution kernel to extract fading characteristics at different time scales, forming a second-level feature sequence, including: Configure three parallel 1D convolutional branches with kernel lengths of 8, 16, and 32 sampling points, respectively; The channel impulse response sequence is input into three convolutional branches, and after global max pooling, three feature vectors are obtained. The three feature vectors are concatenated to form a second-level feature sequence with a dimension of 384.

6. The artificial intelligence-based communication protocol according to claim 5, characterized in that, After aligning the parameters of each interference source in the interference source fingerprint database by timestamp, a graph neural network is used to model the spatiotemporal correlation between interference sources, generating third-level graph embedding features, including: Each interference source is considered as a graph node, and the node feature vector consists of the modulation type one-hot coding, the center frequency normalization value, the bandwidth normalization value, the power spectral density, and the direction of arrival angle. Edge weights are constructed based on spatial distance and frequency proximity. If the difference in the direction of arrival angle between two interference sources is less than 10 degrees and the difference in the center frequency is less than 5 MHz, then an edge connection is established. Then, through three layers of graph convolutional layers, after global average pooling, a third-level graph embedding feature with a dimension of 256 is output.

7. The artificial intelligence-based communication protocol according to claim 6, characterized in that, In the hierarchical policy network, the high-level policy sub-network adopts a classification header structure, and its output is an element of a predefined set of communication mode categories. The set of communication mode categories includes frequency hopping mode, spread spectrum mode, orthogonal frequency division multiplexing mode, chaotic modulation mode, and cognitive radio mode. The underlying execution sub-network is configured with an independent parameter generator for each communication mode category. Each parameter generator is composed of a fully connected neural network, which receives the high-dimensional state vector as input and outputs the waveform shaping filter coefficients required for the corresponding mode, the offset of the center frequency point relative to the nominal value, and the gain factor of the transmit power relative to the reference power.

8. The artificial intelligence-based communication protocol according to claim 7, characterized in that, The hierarchical policy network is jointly trained using a progressive experience replay pool constructed based on a curriculum learning mechanism, including: The initial training phase is set to include only interactive samples in low-interference-density scenarios, with no more than 3 interference sources, a signal-to-noise ratio of no less than -5 dB, and a spectrum occupancy rate of less than 40%. As the training rounds increase, scene samples with medium interference density are gradually introduced, with 4 to 8 interference sources, a signal-to-noise ratio range of -15 dB to -5 dB, and a spectrum occupancy rate of 40% to 70%. In the later stages of training, sampling permissions for high-interference-density scene samples are granted, with more than 8 interference sources, a signal-to-noise ratio below -15 dB, and a spectrum occupancy rate above 70%. All samples are labeled with their difficulty level when stored in the experience replay pool, and are sampled preferentially according to the highest difficulty level allowed in the current training phase during each policy update.

9. The artificial intelligence-based communication protocol according to claim 8, characterized in that, Based on the specific waveform parameters, center frequency offset, and transmit power value output by the hierarchical strategy network, the physical layer communication waveform is dynamically reconstructed and the RF front-end parameters are adjusted, including: Configure the pulse shaping filter in the baseband signal processing unit based on the waveform shaping filter coefficients output by the underlying execution sub-network; Based on the center frequency offset, the frequency synthesizer of the local oscillator is controlled to output the corresponding carrier frequency; based on the transmit power gain factor, the gain control voltage of the power amplifier is adjusted so that the actual output power accurately matches the target value. After completing the above configuration, start the data frame sending process and simultaneously record the success rate, bit error rate and throughput of this communication as reward signals for subsequent policy evaluation.

10. A communication anti-interference system based on artificial intelligence, characterized in that, include: The electromagnetic environment sensing module is used to collect spectrum sensing data, channel status information, and interference source characteristic parameters of the electromagnetic environment in which the current communication link is located in real time. The multi-scale state coding module is used to perform multi-scale time-series fusion coding on the spectrum sensing data, channel state information and interference source characteristic parameters to generate a high-dimensional state vector representing the current electromagnetic situation. The hierarchical strategy decision module is used to input the high-dimensional state vector into the hierarchical strategy network and output specific waveform parameters, center frequency offset and transmit power value. The course learning and training module is used to jointly train the hierarchical policy network based on a progressive experience replay pool; The adaptive waveform reconstruction module is used to dynamically reconstruct the physical layer communication waveform and adjust the radio frequency front-end parameters based on the specific waveform parameters, center frequency offset, and transmit power value.

Images