An anti-interference wireless communication method and apparatus

Abstract

This invention relates to wireless communication technology, and particularly to an anti-interference wireless communication method and apparatus. The method includes acquiring radio frequency data containing interference signals, performing spectral analysis on the data and dividing it into several selectable sub-frequency bands; using an intelligent agent with decision-making capabilities to select a sub-frequency band as the current transmitting end's working channel in each discrete time slot; transmitting data in the selected frequency band at the transmitting end; and receiving data at the receiving end. This invention, through a systematic design combining hardware and algorithms, significantly improves the stability and data transmission accuracy of wireless communication systems under resource-constrained and multi-interference environments, providing a feasible and efficient solution.

Description

Technical Field

[0001] This invention relates to wireless communication technology and signal processing technology, and in particular to an anti-interference wireless communication method and apparatus. Background Technology

[0002] With the widespread application of wireless communication in fields such as the Internet of Things and intelligent transportation, people have placed higher demands on the stability and reliability of communication systems. Currently, improving the anti-interference capability of wireless communication data acquisition mainly relies on the following two technical paths: one is hardware-based anti-interference methods, and the other is adaptive and intelligent methods based on software algorithms.

[0003] At the hardware level, multiple-antenna-machine-machine (MIMO) systems, high-performance filters, and customized RF components are common solutions. For example, patent application CN 119052942A selects channels by measuring noise power in real time and generating unit noise intensity based on bandwidth. It also introduces a cooperative decision-making and fault handling module to enhance the stability and fault tolerance of the wireless communication system. However, this solution relies on a large amount of real-time measurement and information aggregation, placing high demands on hardware resources and channel information interaction. Another example is patent application CN 119094055A, which uses spectrum monitoring and adaptively defines channel mapping and selection mechanisms, enabling wireless devices to maintain relatively smooth communication in variable environments. However, in the process of link quality assessment and adaptive mapping construction, frequent and fine-grained synchronization configuration of communication nodes is still required, increasing implementation complexity. Other hardware measures, such as high-performance filters and multiple-antenna systems, also improve anti-interference levels to some extent, but often lead to increased equipment costs and power consumption, and are difficult to apply in low-power or space-constrained devices (such as drones or portable terminals).

[0004] At the software level, adaptive modulation and coding, interference identification and suppression algorithms, and other methods have been widely explored. Adaptive modulation dynamically switches the coding rate and modulation order according to the channel state to balance data rate and error rate; some researchers have also used advanced signal processing methods to identify and suppress interference signals, improving the utilization efficiency of communication channels. However, these algorithms often have high requirements for computational and storage resources, making it difficult to achieve timely and stable anti-interference effects in low-power, computationally limited scenarios.

[0005] Deep learning-based anti-jamming technologies have emerged in recent years, significantly improving the adaptive capabilities of communication systems under complex interference by automatically extracting channel features and performing interference classification or channel selection. Some solutions have already demonstrated feasibility on software-defined radio (SDR) platforms or high-performance edge computing boards. However, deep learning models generally suffer from complex network structures, large parameter sizes, and significant black-box characteristics, leading to high costs for model training and deployment. Furthermore, the diversity and uncertainty of real-world interference environments, insufficient model generalization ability, and low interpretability of key decision-making processes limit the application of such methods in security-sensitive or real-time-critical scenarios.

[0006] This demonstrates that existing technologies have made numerous attempts at both the hardware and software levels. While hardware technologies have enhanced anti-interference to some extent, they suffer from drawbacks such as high cost, high energy consumption, and space limitations. Software methods, on the other hand, have shortcomings in terms of computational overhead, interpretability, and adaptability to diverse interference environments. Meanwhile, several patent documents have proposed ideas for wireless channel allocation and interference suppression, but they are still not perfect for real-time data acquisition and automatic decision-making in resource-constrained environments. Large-scale deep learning can bring some improvement in anti-interference, but its deployment on low-power devices is difficult. Therefore, how to achieve rapid identification and effective suppression of variable interference while ensuring low power consumption and portability remains a key problem that existing technologies have not fully solved. Summary of the Invention

[0007] To address the aforementioned technical challenges and shortcomings, this invention proposes an anti-interference wireless communication method. This method involves collecting radio frequency data containing interference signals, performing spectrum analysis on the data, and dividing it into several selectable sub-frequency bands. An intelligent agent with decision-making capabilities selects a sub-frequency band as the current transmitting channel in each discrete time slot. Data is transmitted on the selected frequency band at the transmitting end and received at the receiving end.

[0008] Furthermore, the radio frequency data containing interference signals is subjected to spectral analysis and divided into several selectable sub-bands, specifically including the following steps:

[0009] By using the transmission parameters of interference data with different interference intensities and various channel environment combinations provided by the wireless channel simulator, radio frequency data containing interference in different channel environments can be obtained;

[0010] Set a time window and use it to divide the radio frequency data into non-overlapping sub-bands.

[0011] Furthermore, the agent uses the received power vectors at each frequency of the transmitter as state vectors, selects the available sub-frequency bands as action space, and executes the action with the highest reward value in the action space based on the state vectors.

[0012] Furthermore, the agent uses the received power vectors at each frequency of the transmitter as state vectors, selects the available sub-frequency bands as action space, and executes the action with the highest reward value in the action space based on the state vectors.

[0013] Furthermore, the calculation of the reward value includes:

[0014]

[0015] in, This represents the reward value for time slot t; Indicates based on sub-band The throughput of data transmission. This indicates the sub-frequency band where data is transmitted in time slot t; This indicates that the jammer operates at a frequency based on a predefined jamming mode in time slot t; Indicates the cost of sub-band switching; This is a conditional function. The value of this function is 1 when the action in time slot t is different from the action in time slot t-1, and 0 otherwise.

[0016] Furthermore, during training, the agent learns the optimal policy by maximizing a weighted average of cumulative reward and policy entropy, specifically including the following steps:

[0017] During training, the training proceeds in the direction of maximizing the objective function, which is expressed as:

[0018]

[0019] in, The optimal strategy is obtained to maximize the objective function; The strategies that the agent can choose; This indicates a demand for expectation; Represents the state vector of time slot t and action vectors Belongs to the experience pool ; The discount factor represents time slot t; In the state vector Select the action vector to perform. The reward value; As a regulating factor; For strategy In state Entropy of time; For the strategy in the state The probability distribution of all actions.

[0020] Furthermore, during training, the soft Q function directly outputs the Q value of each possible action, that is, the output Q value is a mapping from the state space to the |A|th Cartesian product of the set of real numbers, where |A| represents the number of actions in the action space A.

[0021] Furthermore, the output of the policy network is a mapping of the state space S to the |A|th Cartesian product of [0,1], that is, mapping action selection to the probability of each action being selected in the action space, and the sum of the probability values ​​of all actions is 1.

[0022] This invention also proposes an anti-interference wireless communication device for implementing an anti-interference wireless communication method, comprising an interference signal generator, a wireless channel simulator, a signal receiving module, a signal processing unit, and a data storage module, wherein:

[0023] An interference signal generator is used to produce diverse interference signals based on set power, frequency, and interference duration.

[0024] Wireless channel simulators are used to reproduce different channel characteristics or fading environments, thereby simulating various communication scenarios;

[0025] The signal receiving module is used to capture the target signal and the mixed interference components, and output them to the signal processing unit;

[0026] The signal processing unit is used to filter, amplify, and digitize the received radio frequency signals.

[0027] The data storage module archives the digitized signal data and classifies it according to timestamps, channel numbers, or interference type labels, so that the agent can quickly retrieve and call it during the training and real-time decision-making stages.

[0028] Compared with the prior art, the present invention has the following advantages:

[0029] 1. This invention obtains more comprehensive and high-quality interference samples through a multi-module data acquisition device and a controllable interference environment, laying a solid foundation for reinforcement learning training and inference;

[0030] 2. This invention combines lightweight deep learning with the SAC algorithm, which not only ensures effective interference identification and adaptive scheduling of communication parameters, but also takes into account low power consumption and real-time performance when deployed on edge devices.

[0031] 3. This invention employs a dual-Critic network and entropy regularization strategy to enhance the generalization ability and robustness against variable interference modes, thereby reducing false detections and missed detections under complex channel conditions.

[0032] In summary, this invention, through a systematic design combining hardware and algorithms, significantly improves the stability and data transmission accuracy of wireless communication systems under resource-constrained and multi-interference environments, providing a feasible and efficient solution. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of an anti-interference wireless communication method according to the present invention;

[0034] Figure 2 This is a schematic diagram of an anti-interference wireless communication system according to the present invention.

[0035] Figure 3 A schematic diagram illustrating a single-user, single-jammer scenario.

[0036] Figure 4 This is a schematic diagram illustrating the structure and main functions of the modules in the invention model. Detailed Implementation

[0037] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0038] This invention proposes an anti-interference wireless communication method, such as... Figure 1 This includes: collecting radio frequency data containing interference signals, performing spectrum analysis on it and dividing it into several selectable sub-frequency bands; using an intelligent agent with decision-making capabilities to select a sub-frequency band as the current working channel of the transmitter in each discrete time slot, transmitting data in the frequency band selected by the transmitter, and receiving data at the receiver.

[0039] This embodiment focuses on four key aspects: the composition and function of the data acquisition device, the data acquisition process and data feature extraction, the model structure, and the reinforcement learning algorithm (SAC). The aim is to construct an efficient, reliable, and scalable anti-interference communication system. By providing diverse and accurate channel and interference information at the data level and performing adaptive decision-making at the reinforcement learning level, the anti-interference performance and transmission accuracy of the communication system can be significantly improved in diverse wireless scenarios.

[0040] The composition and functions of data acquisition equipment are as follows: Figure 2 As shown, the overall hardware mainly consists of an interference signal generator, a wireless channel simulator, a signal receiving module, a signal processing unit, and a data storage module, among which:

[0041] Interference signal generator: Used to generate controllable interference signals, providing diverse interference sources for subsequent data acquisition by adjusting output power, frequency, and interference duration;

[0042] Wireless channel simulator: Simulates various real wireless channel transmission environments and simulates the channel state of interference signals in real transmission environments;

[0043] Signal receiving module: It has high sensitivity and wide bandwidth receiving capability, and is used to capture target signals after propagation through the interference environment in real time, so as to ensure that complete received data can still be obtained under multi-frequency band and diverse interference.

[0044] Signal processing unit: performs filtering, amplification and digital conversion on the received raw radio frequency signal, including operations such as removing high-frequency noise and low-frequency noise, and enhancing the effective signal amplitude, laying a high-quality input foundation for subsequent data analysis and feature extraction;

[0045] Data storage module: It classifies and stores digital signals according to various tags such as timestamp, channel number or interference type, and supports fast retrieval and recall, which facilitates the training and performance evaluation of subsequent algorithm models;

[0046] The modules are connected via a high-speed digital interface to achieve low-loss, high-fidelity data transmission. By controlling the parameters of the background signal transmitter and the signal receiver, multi-source signals can be collected under interference environments with different power and bandwidth, providing rich training and test samples for the anti-interference algorithm of this invention.

[0047] The data acquisition process and data feature extraction specifically include the following steps:

[0048] First, the interference signal generator outputs an interference signal with preset power, frequency and duration parameters, and then superimposes it into the target communication channel through a wireless channel simulator to create diverse interference scenarios with varying intensities and types.

[0049] Next, the signal receiving module monitors and captures the target signal mixed with interference in real time, and transmits the obtained radio frequency data to the signal processing unit for amplification and digital conversion;

[0050] The processed digital signals are sent to the data storage module for on-demand classification and labeling, including key information such as interference intensity, interference mode, and experimental conditions. High-quality data samples are continuously collected based on the flexibly scalable interference power and channel status.

[0051] After data collection, further feature extraction and statistical analysis can be performed to select features such as power, gain, or channel state that are of significant reference value for interference detection or system optimization. This provides a unified and controllable input for subsequent deep learning or reinforcement learning models. It can effectively reproduce complex interference phenomena and construct diverse training and testing datasets in a short time, thereby enabling timely testing and iterative improvement of the system's anti-interference performance during the algorithm development phase.

[0052] After analyzing and initially cleaning the binary IQ signals, considering the practical needs of interference identification and noise suppression, attributes that do not significantly contribute to current interference identification or have excessively high noise values ​​(such as freq and rcvpwrdBm) are first removed. The remaining power gain-related indicators are then uniformly converted to dB form to maintain comparability and stability during multi-channel interference analysis. To characterize the overall interference intensity, the "total gain" is defined as the sum of the power of the entire IQ signal followed by the logarithm. This allows for a more intuitive and clearly contrasting energy distribution even in complex noise backgrounds or variable channel conditions. Specifically, this includes:

[0053] Let the length be The time-domain sampling sequence is as follows:

[0054]

[0055] Its time domain amplitude It can be recorded as:

[0056]

[0057] From the maximum amplitude in the time domain Converted to dB values, it can be written as:

[0058]

[0059] The "total gain" is defined as the logarithm of the sum of the power (amplitude squared) of the entire IQ signal:

[0060]

[0061] This value can be considered as the total power over the entire sampling segment, and is used to measure the intensity of interference or the energy level of communication signals.

[0062] The reference power is measured by the equipment and can be approximated as the average time-domain power under certain "noise-free" or "noise-background" conditions. , represented as:

[0063]

[0064] Received Signal Strength Indication (RSSI) is expressed as:

[0065]

[0066] Relative power (relpwr) is expressed as:

[0067]

[0068] Average power (avgpwr) is obtained by averaging the squares of the time-domain amplitudes and then expressed in dB. Specifically, it includes:

[0069]

[0070] The above metrics, obtained by performing amplitude calculations, summation, and logarithmic transformations on the IQ signals in the time domain, yield the interference intensity and signal energy characteristics. The dB representation of power gain-related metrics is preserved and standardized, thus maintaining good comparability and stability in multi-channel interference analysis, signal evaluation, and subsequent algorithms.

[0071] To reduce the impact of temporal fluctuations, the minimum, maximum, average, standard deviation, and several quantiles (25%, 50%, 75%) of metrics such as max_magndB, total_gain_dB, base_pwr, rssi, relpwr, and avgpwr_dB can be calculated within a certain sampling window. These statistical values ​​are then concatenated column-wise to form a unified feature vector, and scene annotations (interference type, frequency, and intensity) are retained within the same data structure, providing directly usable input for subsequent model training or classification tasks.

[0072] After feature extraction and scene annotation are completed, the processed data (spectral_data) is imported into the pre-training process to improve the model's generalization ability and training efficiency. First, the data is sliced ​​according to a certain length and stride. In multi-channel scenarios, the temporal segments of each channel are aggregated into batches, forming a three-dimensional tensor suitable for input to deep learning or reinforcement learning models. Several temporal windows within each batch can be considered as training samples, enabling temporal stitching and comprehensive utilization of multiple scan results.

[0073] This system is designed for single-user and single-jammer scenarios (see appendix). Figure 3This system aims to perform frequency band selection and anti-interference decisions based on existing interference signal data. Unlike the traditional transmitter-receiver-jammer synchronous operation, this system treats a segment of radio frequency data containing interference signals as an "interference scenario." Specifically, the system first performs spectrum analysis and segmentation on the collected interference data to form several selectable sub-frequency bands; then, an agent with decision-making capabilities selects one sub-frequency band in each discrete time slot as the working channel for the current transmitter user.

[0074] Figure 3 This is a single-user, single-jammer scenario. The diagram illustrates a wireless communication system architecture including a transmitter, a receiver, and a jammer. The transmitter (left) transmits data to the receiver (right) via radio frequency signals, while the jammer, located above, sends interference signals to both the transmitter and receiver, attempting to disrupt or block the normal communication link between them. This invention collects and analyzes multi-channel interference signals, uses dual-end selection of the interference signal to replace the data transmission process from transmitter to receiver, and combines this with an adaptive decision-making algorithm to identify and avoid interference signals. This ensures high communication reliability and transmission efficiency even in dynamic and malicious interference environments.

[0075] In this invention, if the energy or signal-to-interference ratio (SIR) of the sub-frequency band selected by the agent in the corresponding data segment is lower than a predetermined threshold, it is considered an "interference conflict"—that is, the frequency band is actually within the coverage area of ​​the jammer and is determined to be subject to interference. If the measured value corresponding to the sub-frequency band is higher than the threshold, it indicates that the transmitter user has successfully avoided the interference or the interference intensity is weak and will not cause communication interruption. By repeatedly selecting frequency bands and performing "conflict checks" (i.e., measuring whether the sub-frequency band is interfered with) in different time slots, the system can collect a large amount of decision-feedback data to train or update the agent, thereby gradually improving anti-interference performance. This process is logically equivalent to traditional user-jammer confrontation, but offline collected interference data replaces the real-time interference process, making the study of multiple interference scenarios and their statistical characteristics more flexible and efficient.

[0076] In its implementation, the jammer can employ various strategies within the data frequency band, such as persistent, frequency-sweeping, or random transmission of jamming power. This system divides the entire jamming data segment into a grid based on frequency and time, then uses these grids as potential "channel selection" candidates. The agent selects a sub-band from several candidate channels within the current discrete time slot to attempt communication. If the power intensity or SINR value corresponding to that frequency band in the sampled data is insufficient to reach the communication threshold, it indicates that the sub-band is occupied by the jammer or the jamming intensity is too high. Conversely, if the test result is better than the threshold, it means that the frequency band can be safely used and generate effective "throughput" benefits. By continuously iterating, recording, and updating the "conflict" or "safety" results, the system can generate training samples that can be used for reinforcement learning, evolutionary algorithms, or other decision-making mechanisms, ultimately achieving more accurate and robust anti-jamming strategy selection.

[0077] This approach inherits the core idea of ​​user-jammer confrontation, namely, the transmitter makes a "frequency band selection" action in the same time slot, while the jammer may or may not interfere in that frequency band. On the other hand, by using pre-collected interference data to segment and check for conflicts, it can effectively avoid real-time confrontation platforms with high hardware costs. At the same time, it can flexibly build a variety of interference scenarios in offline or simulated environments, which is convenient for evaluating and iteratively optimizing the performance of the intelligent agent.

[0078] As attached Figure 4 As shown, the system first acquires environmental conditions through a data acquisition module, including channel characteristics, interference intensity, and other key indicators affecting communication quality. This status information is simultaneously input into both the Actor network and the Critic network, where:

[0079] After receiving the state vector from the environment, the Actor network generates a probability distribution for each possible action (e.g., selecting different transmit power, switching channels, or changing the modulation scheme). During the inference phase, the system can sample based on this probability distribution or directly select the action with the highest probability as the final decision, thereby outputting the action command to the communication device to perform real-time anti-interference operations.

[0080] The Critic network evaluates the value (Q-value) of a given state-action pair, measuring the long-term reward of the decision. To reduce the bias of the value estimation and improve stability, the Critic network adopts a dual network structure (Q_1, Q_2). At each update, two sets of Q-values ​​are calculated for the input state-action pair, and the smaller one is used in the calculation of the target Q-value.

[0081] The parameters of both the Actor and Critic networks are updated by minimizing the loss function. The Critic network adjusts its Q-value estimate based on the actual reward and the state at the next time step; the Actor network adjusts its policy based on the Critic's Q-value feedback and entropy regularization term, balancing the policy's exploratory nature and exploitation.

[0082] This invention employs two key loss functions of SAC during the optimization process: Critic loss and Actor loss. An entropy regularization term is introduced. To improve the robustness of the strategy in a variable and disturbed environment, specific measures include:

[0083] The goal of a Critic network is to perform actions on a given state-action pair. At that time, accurately estimate its Q value to guide the Actor in adjusting its strategy. This applies to immediate rewards. and the next state The present invention defines the target Q value (Q_target) as:

[0084]

[0085] in, This represents a conservative estimate under the dual-Critic structure. Entropy regularization encourages strategy exploration. The discount factor is denoted by mask, which represents the mask indicating whether the state is terminated. Indicates an immediate reward; This indicates the first Critic network's response to the next state. and actions Value estimation; This indicates the second Critic network's response to the next state. and actions Value estimation; Representing state Take action below The probability distribution;

[0086] The Critic loss function is the mean squared error between the target Q value and the predicted Qpred value:

[0087]

[0088] The Actor network, based on the policy gradient concept, optimizes its parameters by minimizing the following objective:

[0089]

[0090] in, Controlling the strength of entropy regularization encourages the randomness and exploration of strategies; The value assessment from the Critic is used to guide the Actor toward convergence towards high-value actions.

[0091] By employing the aforementioned two-part loss and iteratively updating the Actor and Critic network parameters, this invention achieves rapid convergence of adaptive decision-making and optimal communication strategies in complex interference environments. The dual-Critic structure reduces value bias caused by overestimation in a single network, while entropy regularization enhances the exploratory nature of the strategy in uncertain scenarios, thereby significantly improving the robustness and optimality of the system in multi-interference environments.

[0092] This embodiment also proposes a method for implementing the learning objective of an intelligent agent. The SAC (soft actor-critic) algorithm is a reinforcement learning algorithm that aims to learn the optimal policy by maximizing the weighted sum of cumulative reward and policy entropy. This method performs well not only in continuous action spaces but also in discrete action spaces. The following are the learning objectives of an intelligent agent based on the SAC algorithm:

[0093] The goal of the SAC algorithm is to find a strategy that maximizes the following objective function:

[0094]

[0095] in, The optimal strategy is obtained to maximize the objective function; The strategies that the agent can choose; This indicates a demand for expectation; Represents the state vector of time slot t and action vectors Belongs to the experience pool ; The discount factor represents time slot t; In the state vector Select the action vector to perform. The reward value; This is a temperature parameter used to adjust the balance between reward and entropy. as a strategy In state The entropy of time is calculated using the following formula: ; For the strategy in the state The probability distribution of all actions.

[0096] The process of finding the optimal strategy includes the following steps:

[0097] When setting a function that can be randomly initialized when the state space is discrete, Begin by repeatedly applying the modified Bellman backup operator. Thus, the soft Q function is obtained, expressed as:

[0098]

[0099] in, Bellman backup operator; Indicates the regulating factor; This indicates the next state. Calculation of expected value, where The distribution of Decide; The soft state value function representing the next time step;

[0100] The strategy is updated to maximize rewards, using a soft Q function to guide policy changes. Specifically, the updated strategy is as follows:

[0101]

[0102] in, Indicates a new strategy; Represents the KL divergence operation; It's an old strategy. The soft function; It is a partitioning function.

[0103] By alternating between policy evaluation and policy improvement, SAC learns an optimal policy that balances exploration and exploitation in a discrete action space. This approach demonstrates performance comparable to state-of-the-art model-free reinforcement learning algorithms across a variety of discrete action environments.

[0104] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method of interference-free wireless communication, characterized by, The radio frequency data containing interference signals is collected, spectrum analysis is performed thereon, and the radio frequency data is divided into a plurality of selectable sub-frequency bands; a decision-making intelligent agent selects a sub-frequency band as a working channel of a current transmitting end in each discrete time slot, data is transmitted in the selected frequency band at the transmitting end, and data is received at the receiving end; the intelligent agent takes the received power vector at each frequency of the transmitting end as a state vector, takes the selectable sub-frequency bands as an action space, and selects an action with the maximum reward value in the action space according to the state vector, and the calculation of the reward value includes: wherein, represents a reward value of time slot t; represents the throughput of data transmission according to the sub-band represents the throughput of data transmission according to the sub-band represents the sub-band of time slot t for data transmission; represents the frequency of time slot t interferer according to the predefined interference pattern; represents the sub-band switching cost; is a conditional function, which is 1 when the action of time slot t is different from the action of time slot t-1, otherwise 0.

2. The anti-jamming wireless communication method of claim 1, wherein, The radio frequency data containing interference signals is collected, spectrum analysis is performed thereon, and the radio frequency data is divided into a plurality of selectable sub-frequency bands, specifically including the following steps: The radio frequency data containing different interference intensities under different channel environments is obtained through the transmission parameters of the interference data with different interference intensities and various channel environment combinations provided by the wireless channel simulator. A time window is set, and the radio frequency data is divided into non-overlapping sub-frequency bands by using the time window.

3. The anti-jamming wireless communication method of claim 1, wherein, During training, the intelligent agent learns the optimal strategy by maximizing the weighted sum of the cumulative reward and the policy entropy in the training process, specifically including the following steps: During the training process, the training is performed in the direction of maximizing the target function, and the target function is represented as: wherein, is the optimal policy resulting from maximizing the objective function; is a policy that the agent can select; denotes taking the expectation; denotes the state vector at time slot t and action vector belongs to the experience pool ; denotes the discount factor at time slot t; denotes the reward value for selecting the action vector in state vector ; is a tuning factor; is the entropy of the policy at state ; is the probability distribution of the policy over all actions in state .

4. The anti-jamming wireless communication method of claim 3, wherein, During the training process, the soft Q function directly outputs the Q value of each possible action, that is, the output Q value is a mapping from the state space to the |A| times Cartesian product of the real set, and |A| represents the number of actions in the action space A.

5. The anti-jamming wireless communication method of claim 3, wherein, During the training process, the output of the policy network is a mapping from the state space S to the |A| times Cartesian product of the [0, 1] set, that is, the action selection is mapped to the probability of selecting each action in the action space, and the sum of the probability values of all actions is 1.

6. An interference resistant wireless communication device, comprising: A method for implementing the anti-interference wireless communication method of claim 1, comprising an interference signal generator, a wireless channel simulator, a signal receiving module, a signal processing unit and a data storage module, wherein: The interference signal generator is used to generate diversified interference signals according to the set power, frequency and interference time length; The wireless channel simulator is used to reproduce different channel characteristics or fading environments, and thus simulate various communication scenarios; The signal receiving module is used to capture the target signal and the mixed interference component, and output them to the signal processing unit; The signal processing unit is used to filter, amplify and digitize the received radio frequency signals; The data storage module archives the digitized signal data, classifies and manages them according to the time stamp, channel number or interference type label, so that the intelligent agent can quickly retrieve and call them during the training and real-time decision-making stages.

Images