Symbiotic radio network anti-interference method based on meta learning and RIS

Abstract

The invention discloses a symbiotic radio network anti-interference method based on meta-learning. The method comprises the following steps: acquiring symbiotic radio network model information; respectively deploying reconfigurable intelligent surfaces at secondary user sending ends in the symbiotic radio network model; the reconfigurable intelligent surface identifies and distinguishes the types of external signals through a radio signal identification model; and adjusting the working mode of the reconfigurable intelligent surface according to the type of the external signal. According to the meta-learning-based anti-interference method for the symbiotic radio network, reconfigurable intelligent surfaces are deployed on the surfaces of a plurality of secondary transmitting base stations around a main user to form a'protective cover ', so that secure communication assisted by the reconfigurable intelligent surfaces in the symbiotic radio network is realized in a symbiotic mode and a defense mode; the signal receiving quality of the main user is improved, and the wireless communication safety of the main network is guaranteed.

Description

Technical Field

[0001] The present invention belongs to the technical field of communication networks and relates to an anti-interference method for a symbiotic radio network based on meta-learning and RIS. Background Art

[0002] Meta-learning is a learning method to improve the generalization ability of a model, and its core idea is "learning how to learn". Different from traditional deep learning that focuses on a single task, meta-learning enables the model to quickly adapt to new tasks with only a small amount of data through cross-task training and is widely applied to scenarios such as few-shot learning. Deep reinforcement learning combines deep learning and reinforcement learning and is mainly used to solve complex decision-making problems. Deep learning extracts features of high-dimensional data through multi-layer neural networks, while reinforcement learning is based on the interaction between an agent and the environment and optimizes the decision-making strategy through reward signals. In recent years, deep reinforcement learning has demonstrated powerful capabilities in fields such as games, robot control, and autonomous driving. Especially in dealing with large-scale state spaces and complex strategies, it has significantly improved the learning efficiency and decision-making quality.

[0003] Reconfigurable Intelligent Surface (RIS) is a new type of communication technology. By deploying a large number of adjustable reflection units in the environment, it can intelligently change the propagation characteristics of wireless channels, thereby enhancing signal quality, improving spectrum efficiency, and reducing energy consumption. It is an important part of realizing 6G intelligent wireless communication. Symbiotic radio technology combines cognitive radio and ambient backscatter technology. While improving spectrum efficiency, it eliminates the interference impact of secondary transmit base stations on the primary network and simultaneously enhances the coding security ability of ambient backscatter technology. Deploying a reconfigurable intelligent surface at the secondary transmit base station in a symbiotic radio network can reflect the wireless signals of the primary network, improve the communication quality of the primary network, and at the same time, it can send its own information to the secondary receiver in a symbiotic mode by embedding it into the primary RF signal. However, during the process of the primary network transmitting secret information, there may be an interferer emitting different interference signals to the primary network, thereby damaging the signal reception quality of the primary user and threatening the wireless communication security of the primary network.

[0004] Currently, in the prior art, there is a problem that the signal reception quality of the primary user is damaged when there are interference signals in the primary network of a symbiotic radio network. Summary of the Invention

[0005] The purpose of the present invention is to provide an anti-interference method for a symbiotic radio network based on meta-learning and RIS, which solves the problem in the prior art that the signal reception quality of the primary user is damaged when there are interference signals in the primary network of a symbiotic radio network.

[0006] The technical solution adopted by the present invention is an anti-interference method for a coexisting radio network based on meta-learning, including the following steps: Step 1: Obtain the coexisting radio network model information; Step 2: Deploy reconfigurable intelligent surfaces at the secondary user transmitters in the coexisting radio network model; Step 3: The reconfigurable intelligent surface identifies and distinguishes the types of external signals through a radio signal recognition model; Step 4: Adjust the working mode of the reconfigurable intelligent surface according to the types of external signals.

[0007] The features of the present invention also lie in: The coexisting radio network model information includes primary base station information, primary user information, secondary user transmitter information, secondary user receiver information, and interference signal conditions.

[0008] Step 3 includes: Step 3.1: Deploy and apply the trained radio signal recognition model on the reconfigurable intelligent surface; Step 3.2: The reconfigurable intelligent surface preprocesses the received radio signals through signal framing and transformation using the radio signal recognition model, then extracts the typical features of the received signals, and then classifies the received radio signals into legitimate signals or interference signals.

[0009] The radio signal recognition model uses a model-agnostic meta-learning model to identify and distinguish the types of external signals. The model-agnostic meta-learning model uses a convolutional neural network as the inner model, and the convolutional neural network includes three convolutional layers and two fully connected layers.

[0010] The gradient update situation of the inner loop of the model-agnostic meta-learning model is as follows:

[0011] Among them, represents the updated model parameters, represents the current model parameters, represents the learning rate, ( ) represents the loss function for task the gradient of the model parameters .

[0012] The objective function of the outer loop of the model-agnostic meta-learning model is defined as follows: , Among them, represents the model parameters, represents the task distribution for all tasks in Sum denotes using the updated model parameters compute the task on the loss function , denotes the task used during training denotes the distribution set of multiple tasks denotes the task 's loss function denotes on the task after the inner-loop parameter update on the model

[0013] The formula for the outer-loop optimization of the model-agnostic meta-learning model is as follows , where denotes the model parameters denotes the learning rate denotes for the task distribution in all tasks with respect to the model the sum of the loss functions calculated with respect to the parameter 's gradient denotes the task used during training denotes the distribution set of multiple tasks denotes the task 's loss function denotes on the task after the inner-loop update on the model

[0014] The working modes of the reconfigurable intelligent surface include the symbiotic mode and the defense mode, and the working mode of the reconfigurable intelligent surface is adjusted and switched through the reflection matrix coefficient optimization model

[0015] The reflection matrix coefficient optimization model is obtained through the following steps Step A1: Construct a reflection element optimization model by combining the symbiotic radio network model information Step A2: The reconfigurable intelligent surface obtains the reflection matrix coefficient optimization model after training and optimizing the reflection element optimization model through the deep reinforcement learning algorithm The reconfigurable intelligent surface classifies the received external signals into legal signals and interference signals through the radio signal recognition model. When it is recognized as a legal signal, the reconfigurable intelligent surface trains and optimizes the reflection elements of its reflection matrix through the deep reinforcement learning algorithm to obtain the optimal reflection strategy and maximize the signal reception quality of the primary user When it is recognized as an interference signal, the reconfigurable intelligent surface trains and optimizes the reflection elements of its reflection matrix through a deep reinforcement learning algorithm, reflects it to a specific area, obtains the optimal reflection strategy, minimizes the interference impact of the interference signal on the primary user, and obtains a reflection matrix coefficient optimization model by synthesizing the above optimal reflection strategy.

[0016] The reflection element optimization model is a five-tuple multi-objective optimization model based on Markov game; The reflection element optimization model is specifically , where represents the set of agents, including multiple reconfigurable intelligent surfaces; represents the state set, containing the environmental data observed by the reconfigurable intelligent surface; represents the action set, including the selection of the defense mode or the symbiotic mode, and the adjustment of the reflection phase of the reconfigurable intelligent surface; represents the state transition probability, including the possibility that the agent switches to the next state after executing the current action according to the current strategy; represents the reward function, including the reward function of the agent after executing the current action according to the current strategy, reflecting the impact of the current action on the strategy.

[0017] The beneficial effects of the present invention are as follows: The present invention deploys reconfigurable intelligent surfaces on the surfaces of multiple secondary transmitting base stations around the primary user to form a "protective shield", so as to realize secure communication assisted by reconfigurable intelligent surfaces in the symbiotic radio network in the symbiotic mode and the defense mode, improve the signal reception quality at the primary user, and ensure the wireless communication security of the primary network; The present invention combines meta-learning and reinforcement learning at the secondary user transmitter, enabling it to distinguish interference signals and dynamically adjust the reflection matrix of the reconfigurable intelligent surface. When the reconfigurable intelligent surface receives an external signal, it uses a wireless signal recognition model trained by meta-learning to distinguish between legitimate signals and interference signals. If the received signal is recognized as a legitimate signal, the reconfigurable intelligent surface will be in the symbiotic mode, reflecting the legitimate signal and transmitting its own information to the secondary receiver using ambient backscatter technology. On the contrary, if the signal is determined to be an interference signal, the reconfigurable intelligent surface will switch to the defense mode, adjust the reflection matrix, and guide the interference signal to other areas, thereby reducing its impact on the primary user. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 is a schematic structural diagram of the symbiotic radio network assisted by the reconfigurable intelligent surface in the present invention in the symbiotic mode; Figure 2 is a schematic structural diagram of the symbiotic radio network assisted by the reconfigurable intelligent surface in the present invention in the defense mode; Figure 3 It is a comparison schematic diagram of the wireless signal recognition model and the traditional model in terms of the recognition accuracy of interference signals in the present invention; Figure 4 It is a comparison schematic diagram of the wireless signal recognition model and the traditional model in terms of the training convergence speed in the present invention; Figure 5 It is a comparison schematic diagram of the wireless signal recognition model and the traditional model in terms of the signal-to-interference-plus-noise ratio performance in the present invention. Detailed implementation manners

[0019] The present invention will be described in detail below with reference to the accompanying drawings and specific implementation manners.

[0020] As Figure 1 and Figure 2 shown, the reconfigurable intelligent surface-assisted coexisting radio network includes a multi-antenna-equipped primary base station, a single-antenna primary user, secondary user transmitters equipped with multi-antenna, each secondary user transmitter is equipped with a reconfigurable intelligent surface, and single-antenna secondary user receivers, and there is a malicious jammer. The secondary user transmitters, reconfigurable intelligent surfaces and secondary user receivers are respectively denoted as , and , where . The primary base station and each secondary user transmitter are respectively equipped with and antennas, and each reconfigurable intelligent surface is equipped with reflection units. While the primary base station transmits confidential information to the primary user, the jammer will emit different interference signals. The secondary user transmitter reflects the confidential signal of the primary base station and the interference signal of the jammer by modulating the matrix coefficients of the reconfigurable intelligent surface, so as to enhance the desired signal received by the primary user and effectively suppress the interference effect, thereby improving the signal-to-interference-plus-noise ratio of the primary user.

[0021] During the process of the primary base station transmitting confidential signals to the primary user, the reconfigurable intelligent surface is deployed on the surface of the secondary user transmitter to enhance the signal-to-interference-plus-noise ratio of the primary user and weaken the interference from the jammer at the same time. According to the different signals received by the reconfigurable intelligent surface, the reconfigurable intelligent surface operates in a coexisting mode and a defense mode. As Figure 1 shown, when the reconfigurable intelligent surface receives a legitimate signal, it executes the coexisting mode, reflects the legitimate signal to the primary user and transmits its own information to the secondary receiver using backscatter technology; as Figure 2 shown, when the reconfigurable intelligent surface receives an interference signal, it executes the defense mode, reflects the interference signal to a specific area, reduces the impact of the jammer on the primary user, and achieves effective anti-interference performance, thereby protecting the wireless communication security of the primary network.

[0022] During the wireless communication of the primary user in the coexisting radio network, the signal received by the primary user can be expressed as: , where represents the signal received by the primary user, represents the conjugate transpose of the channel response from the primary base station to the primary user, represents the beamforming vector of the primary base station, represents the signal transmitted from the primary base station, represents the conjugate transpose of the channel response from the jammer to the primary user, represents the beamforming vector of the jammer, represents the signal transmitted by the jammer, represents the channel response from the primary user to the secondary receiver, represents the modulation coefficient of the secondary transmitter, represents the reflection matrix of the reconfigurable intelligent surface, represents the channel response from the secondary transmitter to the primary user, represents the channel response from the jammer to the secondary receiver, represents additive white Gaussian noise, and M represents the number of secondary users, represents the th secondary user; The reconfigurable intelligent surface reflects the received legitimate signal to the primary user, thereby improving its signal quality. At the same time, the reconfigurable intelligent surface adopts ambient backscatter technology to embed the signal it wants to transmit into the primary RF signal and then forwards the signal to the secondary user receiver.

[0023] The signal received by the secondary user receiver can be expressed as: , where represents the signal received by the th secondary user receiver, represents the conjugate transpose of the channel response from the primary base station to the secondary user, represents the channel response from the primary base station to the reconfigurable intelligent surface of the secondary transmitter, represents the modulation coefficient of the secondary transmitter, represents the reflection matrix of the reconfigurable intelligent surface, represents the channel response from the secondary transmitter to the secondary user, represents the beamforming vector of the primary base station, represents the signal transmitted from the secondary transmitter to the secondary user, represents the channel response from the jammer to the secondary user, Denotes the channel response from the jammer to the reconfigurable intelligent surface at the secondary transmitter Denotes the beamforming vector of the jammer Denotes the signal transmitted by the jammer Denotes ; The channel coefficient from the signal transmitter to the signal receiver is denoted as , where Denotes the primary base station, the jammer, and the reconfigurable intelligent surface Denotes the primary user, the reconfigurable intelligent surface, and the secondary user receiver. The channel model is expressed as: , Where, Denotes a complex channel vector of size Denotes the path loss from channel To The path loss expression is: , where Denotes the distance between the transmitter and the receiver.

[0024] In the received signal expressions at the primary user and the secondary user receiver, the beamforming vectors of the primary base station and the jammer belong to the complex spaces And . Denotes the primary signal from the primary base station to the primary user Denotes the interference signal from the jammer Denotes the secondary signal from the secondary signal transmitter to the secondary user receiver. The symbol Denotes the additive white Gaussian noise, whose two-sided power spectral density is . Assume Complies with a Gaussian distribution with a mean of zero and a variance of , where Denotes the channel bandwidth. Assume that all channels experience independent Rayleigh fading.

[0025] The reconfigurable intelligent surface deployed on the surface of the secondary user receiver adjusts its reflection coefficient through the reflection matrix , expressed as: ,

[0026] Where, Denotes the reflection matrix of the reconfigurable intelligent surface Denotes the amplitude coefficient of the th reflection element Denotes the phase coefficient of the th reflection element is the number of reflecting elements of the reconfigurable intelligent surface.

[0027] Under the assumption of imperfect channel state information, this invention studies the performance of anti-interference secure communication in the primary network of a co-radio network. This invention evaluates the anti-interference communication performance of the system based on the signal-to-interference ratio of the primary user. To improve the communication quality of the primary network and mitigate the impact of interference, this invention proposes a method combining meta-learning and reinforcement learning to maximize the signal-to-interference-plus-noise ratio of the primary user, thereby achieving robust anti-interference communication.

[0028] Given the transmission powers of the primary base station and the interferer and , the signal-to-interference-plus-noise ratio of the primary user can be expressed as:

[0029] where represents the beamforming matrix of the primary base station, represents the beamforming matrix of the interferer, represents the conjugate transpose of the channel response from the primary base station to the primary user, represents the channel response from the primary user to the secondary receiver, represents the modulation coefficient of the secondary transmitter, represents the reflection matrix of the reconfigurable intelligent surface, represents the channel response from the secondary transmitter to the primary user, represents the channel response from the jammer to the secondary receiver, represents the noise power at the primary user; In the co-radio network, the objective of this invention is to enhance the signal quality of the primary user by optimizing the beamforming vector of the primary base station and the reflection coefficients of the reconfigurable intelligent surface. Therefore, the problem of maximizing the signal-to-interference-plus-noise ratio can be formulated as follows:

[0030] where represents the signal-to-interference-plus-noise ratio of the primary user, represents the beamforming matrix of the primary base station, represents the reflection matrix of the reconfigurable intelligent surface, represents the maximum transmission power of the primary base station, represents the maximum transmission power of the interferer, represents the rank of represents the rank of

[0031] In the present invention, the secondary signal transmission end provides computing resources for model training and reinforcement learning optimization, and the reconfigurable intelligent surface executes the trained model and the optimization results. The wireless signal recognition model trained by meta-learning is used to identify the type of signals received by the reconfigurable intelligent surface, and the beamforming vector of the primary base station and the reflection matrix coefficient of the reconfigurable intelligent surface are optimized through a deep reinforcement learning algorithm, so as to realize the reconfigurable intelligent surface-assisted secure communication for the coexisting radio network.

[0032] The first stage of creating the wireless signal recognition model is the task that the recognition model must execute. Each task is defined as a classification task among different categories. For each category, the model can be trained with a fixed number of samples (denoted as ). In addition, the present invention divides the categories into two groups, one group for creating tasks in the meta-training stage and the other group for tasks in the meta-testing stage. Since the categories in the meta-testing tasks are unknown to the model, the present invention can examine the flexibility of the model. The model-agnostic meta-learning method (MAML) is used as the basis of the wireless signal recognition model, and the model uses a convolutional neural network model with three convolutional layers and two fully connected layers as the inner model.

[0033] To solve the optimization problems of the beamforming vector of the primary base station and the reflection matrix coefficient of the reconfigurable intelligent surface, the present invention constructs it as a multi-objective optimization problem based on Markov game. First, a five-tuple Markov game formulates the optimization problem of the reflection elements, denoted as . Among them, represents the set of agents, represents the set of states, represents the set of actions, represents the state transition probability, represents the reward function. Since multiple reconfigurable intelligent surfaces participate as intelligent agents, the present invention regards the communication network with interference sources as the environment. The specific details are as follows: 1) State space: contains the environmental data observed by the reconfigurable intelligent surface, such as past channel state data, the characteristics of the signals received by the reconfigurable intelligent surface, and the current reflection matrix of the reconfigurable intelligent surface.

[0034] 2) Action space: For each reconfigurable intelligent surface, the action space includes selecting a defense mode or a symbiotic mode and adjusting the reflection phase of the reconfigurable intelligent surface to enhance the legitimate signals or suppress the interference signals.

[0035] 3) State transition probability: At time , when the action is selected, the probability of transitioning from state to the subsequent state . All and satisfy the following constraints: , where represents the probability of transitioning from state to after performing action , and represents the sum of the probabilities of transitioning to all possible states after performing action from state being 1. 4) Reward function: To maximize the signal-to-interference-plus-noise ratio of the primary user, the reflection matrix coefficients of the reconfigurable intelligent surface are optimized in the coexisting radio network. Therefore, the immediate reward function of an agent can be written in the following form: , where represents the immediate reward function of the agent, and represents the signal-to-interference-plus-noise ratio of the primary user.

[0036] The present invention introduces a multi-agent reinforcement learning algorithm based on the multi-agent deep deterministic gradient algorithm to solve the optimization problem of P1. The multi-agent deep deterministic gradient algorithm is an improvement of the actor-critic and deep deterministic policy gradient algorithms. It adopts an operation mode of centralized training and distributed execution. Each agent includes an actor network, a critic network, an actor target network, and a critic target network. In the multi-agent deep deterministic gradient algorithm, the actor network operates only using local information, the critic network is improved using global information, and each agent considers the influence of other agents when making decisions.

[0037] The present invention determines as the policy of agent in the joint trajectory design and power allocation algorithm. By modifying the parameters and of the evaluation network, the optimal policy can be achieved. During this process, the parameters and of the evaluation network are modified continuously in time. Specifically, the operation experience obtained through the interaction between the agent and the environment is retained in the experience replay pool . During the training process, by sampling from the experience replay buffer Extract a small batch of samples to update the parameters of the evaluation network. By minimizing the loss function, the critic network modifies the parameters of the evaluation network . The formula of the loss function can be expressed as follows: , where represents the state–action value function of the target network, represents the agent 's critic network loss function, represents the parameters of the critic network; represents the agent 's critic network's Q value at state , its own action and other agents' actions , with parameters ; represents the target Q value; represents the expected value; represents the th agent's immediate reward obtained at time.

[0038] To modify the parameters of the actor network , it is necessary to maximize the policy objective function, and the expression of the policy objective function can be expressed as follows: , where represents the policy objective function, is the actor evaluation network function representing the deterministic policy , represents the Q value function output by the critic network (Critic) of the agent ; In the present invention, instead of directly transferring the parameters and to the target network, they are gradually adjusted as the parameters of the evaluation network and are modified, as follows: , where , represents the target network parameters of the actor network of the agent , represents the soft update mixing coefficient of the actor network, represents the target network parameters of the critic network of the agent, represents the soft update mixing coefficient of the critic network.

[0039] Example 1 This example presents an anti-jamming method for a coexisting radio network based on meta-learning, including the following steps: Step 1: Obtain the information of the coexisting radio network model; Step 2: Deploy reconfigurable intelligent surfaces at the secondary user transmitters in the coexisting radio network model; Step 3: The reconfigurable intelligent surface identifies and differentiates the types of external signals through a radio signal recognition model; Step 4: Adjust the working mode of the reconfigurable intelligent surface according to the types of external signals.

[0040] Example 2 This example presents an anti-jamming method for a coexisting radio network based on meta-learning, including the following steps: Step 1: Obtain the information of the coexisting radio network model; The information of the coexisting radio network model includes primary base station information, primary user information, secondary user transmitter information, secondary user receiver information, and interference signal conditions; Step 2: Deploy reconfigurable intelligent surfaces at the secondary user transmitters in the coexisting radio network model; Step 3: The reconfigurable intelligent surface identifies and differentiates the types of external signals through a radio signal recognition model; Step 4: Adjust the working mode of the reconfigurable intelligent surface according to the types of external signals.

[0041] Example 3 This example presents an anti-jamming method for a coexisting radio network based on meta-learning, including the following steps: Step 1: Obtain the information of the coexisting radio network model; Step 2: Deploy reconfigurable intelligent surfaces at the secondary user transmitters in the coexisting radio network model; Step 3: The reconfigurable intelligent surface identifies and differentiates the types of external signals through a radio signal recognition model; Step 3.1: Deploy and apply the trained radio signal recognition model on the reconfigurable intelligent surface; The radio signal recognition model uses a model-agnostic meta-learning model to identify and differentiate the types of external signals. The model-agnostic meta-learning model uses a convolutional neural network as the inner model, and the convolutional neural network includes three convolutional layers and two fully connected layers; Step 3.2: The reconfigurable intelligent surface preprocesses the received radio signals through signal framing and transformation through the radio signal recognition model, then extracts the typical features of the received signals, and then classifies the received radio signals as legal signals or interference signals; Step 4. Adjust the working mode of the reconfigurable intelligent surface according to the type of external signal.

[0042] Embodiment 4 This embodiment proposes an anti-interference method for a symbiotic radio network based on meta-learning, including the following steps: Step 1. Obtain the symbiotic radio network model information; Step 2. Deploy reconfigurable intelligent surfaces at the secondary user transmitters in the symbiotic radio network model; Step 3. The reconfigurable intelligent surface identifies and distinguishes the types of external signals through a radio signal recognition model; Step 3.1. Deploy and apply the trained radio signal recognition model on the reconfigurable intelligent surface; The radio signal recognition model uses a model-agnostic meta-learning model to identify and distinguish the types of external signals. The model-agnostic meta-learning model uses a convolutional neural network as the inner model, and the convolutional neural network includes three convolutional layers and two fully connected layers; The gradient update of the inner loop of the model-agnostic meta-learning model is as follows:

[0043] Among them, represents the updated model parameters, represents the current model parameters, represents the learning rate, ( ) represents the loss function for the task with respect to the model parameters gradient; Step 3.2. The reconfigurable intelligent surface preprocesses the received radio signals through signal framing and transformation by the radio signal recognition model, then extracts the typical features of the received signals, and then classifies the received radio signals into legal signals or interference signals; Step 4. Adjust the working mode of the reconfigurable intelligent surface according to the type of external signal.

[0044] Embodiment 5 This embodiment proposes an anti-interference method for a symbiotic radio network based on meta-learning, including the following steps: Step 1. Obtain the symbiotic radio network model information; Step 2. Deploy reconfigurable intelligent surfaces at the secondary user transmitters in the symbiotic radio network model; Step 3. The reconfigurable intelligent surface identifies and distinguishes the types of external signals through a radio signal recognition model; Step 3.1. Deploy and apply the trained radio signal recognition model on the reconfigurable intelligent surface; The radio signal recognition model uses a model-agnostic meta-learning model to identify and distinguish different types of external signals. The model-agnostic meta-learning model uses a convolutional neural network as the inner model, and the convolutional neural network includes three convolutional layers and two fully connected layers; The objective function of the outer loop of the model-agnostic meta-learning model is defined as follows: , where, represents the model parameters, represents the sum over all tasks in the task distribution , represents calculating the loss function on task using the updated model parameters , , represents the task used during training, represents the set of distributions of multiple tasks, represents the task 's loss function, represents the model after the inner loop parameter update on task ; The formula for optimizing the outer loop of the model-agnostic meta-learning model is as follows: , where, represents the model parameters, represents the learning rate, represents the gradient of the sum of the loss functions calculated for all tasks in the task distribution with respect to the model with respect to the parameter , represents the task used during training, represents the set of distributions of multiple tasks, represents the task 's loss function, represents the model after the inner loop update on task ; Step 3.2: The reconfigurable intelligent surface preprocesses the received radio signal by signal framing and transformation through the radio signal recognition model, then extracts the typical features of the received signal, and then classifies the received radio signal as a legitimate signal or an interference signal; Step Four: Adjust the working mode of the reconfigurable intelligent surface according to the type of external signal.

[0045] Example 6 This embodiment proposes an anti-interference method for coexisting radio networks based on meta-learning, including the following steps: Step 1: Obtain the coexisting radio network model information; Step 2: Deploy reconfigurable intelligent surfaces at the secondary user transmitters in the coexisting radio network model; Step 3: The reconfigurable intelligent surface identifies and distinguishes the types of external signals through a radio signal recognition model; Step 4: Adjust the working mode of the reconfigurable intelligent surface according to the types of external signals.

[0046] The working modes of the reconfigurable intelligent surface include a coexisting mode and a defense mode, and the working mode of the reconfigurable intelligent surface is adjusted and switched through a reflection matrix coefficient optimization model; The reflection matrix coefficient optimization model is obtained through the following steps: Step A1: Construct a reflection element optimization model in combination with the coexisting radio network model information; The reflection element optimization model is a five-tuple multi-objective optimization model based on Markov game; the reflection element optimization model is specifically , where represents the set of agents, including multiple reconfigurable intelligent surfaces; represents the state set, containing the environmental data observed by the reconfigurable intelligent surface; represents the action set, including the selection of the defense mode or the coexisting mode, and the adjustment of the reflection phase of the reconfigurable intelligent surface; represents the state transition probability, including the possibility that the agent switches to the next state after executing the current action according to the current strategy; represents the reward function, including the reward function of the agent after executing the current action according to the current strategy, reflecting the impact of the current action on the strategy; Step A2: The reconfigurable intelligent surface trains and optimizes the reflection element optimization model through a deep reinforcement learning algorithm to obtain a reflection matrix coefficient optimization model; The reconfigurable intelligent surface classifies the received external signals into legal signals and interference signals through a radio signal recognition model. When it is identified as a legal signal, the reconfigurable intelligent surface trains and optimizes the reflection elements of its reflection matrix through a deep reinforcement learning algorithm to obtain an optimal reflection strategy to maximize the signal reception quality of the primary user; When it is identified as an interference signal, the reconfigurable intelligent surface trains and optimizes the reflection elements of its reflection matrix through a deep reinforcement learning algorithm, reflects it to a specific area, obtains an optimal reflection strategy, and minimizes the interference impact of the interference signal on the primary user. The optimal reflection strategy above is integrated to obtain a reflection matrix coefficient optimization model.

[0047] Such asFigure 3 As shown, under different signal-to-interference-plus-noise ratio (SINR) conditions, the performance comparison of the meta-learning model proposed in this application with traditional deep learning models and transfer learning models in terms of the accuracy of interference signal recognition is presented. First, it can be observed that as the SINR increases, the recognition accuracy of all three models improves. This is because stronger legitimate signals enhance the feature-based recognition accuracy. In addition, the meta-learning model demonstrates stronger adaptability by being trained on tasks containing different interference signals. Therefore, it can quickly adapt to new tasks and efficiently recognize interference signals in different environments, achieving higher accuracy even with limited training data.

[0048] As Figure 4 shown, the comparison of the training convergence speed of the meta-learning model proposed in this invention with traditional deep learning models and transfer learning models in a new interference environment is presented. The meta-learning model can achieve convergence with only 5 to 15 rounds of training, demonstrating its excellent adaptability under new interference conditions. In contrast, the transfer learning model requires 100 to 500 rounds of training to converge, while the traditional deep learning model requires thousands of rounds of training. This highlights the significant advantage of the meta-learning method in quickly adapting to dynamic interference environments.

[0049] As Figure 5 shown, as the transmit power of the primary base station increases, the SINR of the meta-learning model proposed in this invention, traditional deep learning models, and transfer learning models all show an upward trend. This is because higher transmit power improves the signal strength received by the primary user and also amplifies the signal reflected by the reconfigurable intelligent surface, thus improving the overall signal quality. In addition, the meta-learning model proposed in this invention achieves a significantly higher SINR compared to traditional deep learning models and transfer learning models. This improvement benefits from the combination of meta-learning and reinforcement learning, where meta-learning effectively identifies legitimate signals and interference signals, and reinforcement learning dynamically optimizes the reflection coefficient of the reconfigurable intelligent surface, ultimately achieving better SINR performance.

[0050] This invention proposes a reconfigurable intelligent surface-assisted anti-interference strategy in a coexisting radio network. In the primary network, interference sources transmit various interference signals, leading to a decline in the communication quality of primary users. To improve the signal quality of primary users, this invention deploys reconfigurable intelligent surfaces on the surfaces of multiple secondary signal transmitters, using the reconfigurability of the reconfigurable intelligent surfaces to enhance the signals received by primary users and, at the same time, reducing the interference effect by reflecting interference signals. Specifically, this invention combines meta-learning and reinforcement learning on secondary user transmitters to enable them to distinguish interference signals and dynamically adjust the reflection matrix of the reconfigurable intelligent surface. When the reconfigurable intelligent surface receives an external signal, it uses a wireless signal recognition model trained through meta-learning to distinguish between legitimate signals and interference signals. If the received signal is recognized as a legitimate signal, the reconfigurable intelligent surface will be in a coexisting mode, reflecting the legitimate signal while transmitting its own information to the secondary receiver using ambient backscatter technology. On the contrary, if the signal is determined to be an interference signal, the reconfigurable intelligent surface will switch to a defensive mode, adjust the reflection matrix, and direct the interference signal to other areas, thereby reducing its impact on primary users. The simulation results verify the accuracy of the wireless signal recognition model trained through meta-learning and demonstrate the effectiveness of the proposed anti-interference strategy in dealing with new types of interference attacks.

Claims

1. Anti-interference method for symbiotic radio network based on meta-learning, characterized in that It includes the following steps: Step 1: Obtain the information of the coexisting radio network model; Step 2: Deploy reconfigurable intelligent surfaces at the secondary user transmitters in the coexisting radio network model; Step 3: The reconfigurable intelligent surface identifies and differentiates the types of external signals through a radio signal recognition model; Step 4: Adjust the working mode of the reconfigurable intelligent surface according to the types of external signals.

2. The anti-interference method for a symbiotic radio network based on meta-learning according to claim 1, characterized in that The information of the coexisting radio network model includes primary base station information, primary user information, secondary user transmitter information, secondary user receiver information, and interference signal conditions.

3. The anti-interference method for symbiotic radio networks based on meta-learning according to claim 1, wherein, The said Step 3 includes: Step 3.1: Deploy and apply the trained radio signal recognition model on the reconfigurable intelligent surface; Step 3.2: The reconfigurable intelligent surface preprocesses the received radio signals through signal framing and transformation using the radio signal recognition model, then extracts the typical features of the received signals, and then classifies the received radio signals into legitimate signals or interference signals.

4. The anti-interference method for a symbiotic radio network based on meta-learning according to claim 3, wherein The radio signal recognition model uses a model-agnostic meta-learning model to identify and differentiate the types of external signals. The model-agnostic meta-learning model uses a convolutional neural network as the inner model, and the convolutional neural network includes three convolutional layers and two fully connected layers.

5. The anti-interference method for symbiotic radio networks based on meta-learning according to claim 4, wherein The gradient update of the inner loop of the model-agnostic meta-learning model is as follows: Among them, represents the updated model parameters, represents the current model parameters, represents the learning rate, ( ) represents the loss function for the task the gradient of the model parameters .

6. The anti-interference method for a symbiotic radio network based on meta-learning according to claim 4, wherein The objective function of the outer loop of the model-agnostic meta-learning model is defined as follows: ， Among them, represents the model parameters, represents the summation of all tasks in the task distribution represents using the updated model parameters to calculate the loss function on the task , represents the task used during training, represents the set of task distributions of multiple tasks, represents the task 's loss function, represents the model after the inner-loop parameter update on the task.

7. The anti-interference method for a symbiotic radio network based on meta-learning according to claim 4, characterized in that The formula for optimizing the outer loop of the model-agnostic meta-learning model is as follows: ， Among them, represents the model parameters, represents the learning rate, represents the task distribution for all tasks in the model the sum of the loss functions calculated with respect to the parameters of the gradient, represents the tasks used during training, represents a set of distributions of multiple tasks, represents the task of the loss function, represents the model after being updated through the inner loop on the task ​ 8. The anti-interference method for symbiotic radio networks based on meta-learning according to claim 1, characterized in that The working modes of the reconfigurable intelligent surface include a coexisting mode and a defense mode, and the working mode of the reconfigurable intelligent surface is adjusted and switched through a reflection matrix coefficient optimization model.

9. The anti-interference method for a symbiotic radio network based on meta-learning according to claim 8, wherein The reflection matrix coefficient optimization model is obtained through the following steps: Step A1: Construct a reflection element optimization model in combination with the information of the coexisting radio network model; Step A2: The reconfigurable intelligent surface trains and optimizes the reflection element optimization model through a deep reinforcement learning algorithm to obtain the reflection matrix coefficient optimization model; The reconfigurable intelligent surface classifies the received external signals into legitimate signals and interference signals through the radio signal recognition model. When it is identified as a legitimate signal, the reconfigurable intelligent surface trains and optimizes the reflection elements of its reflection matrix through a deep reinforcement learning algorithm to obtain the optimal reflection strategy to maximize the signal reception quality of the primary user; When it is identified as an interference signal, the reconfigurable intelligent surface trains and optimizes the reflection elements of its reflection matrix through a deep reinforcement learning algorithm to reflect it to a specific area to obtain the optimal reflection strategy to minimize the interference impact of the interference signal on the primary user. The optimal reflection strategies above are integrated to obtain the reflection matrix coefficient optimization model.

10. The anti-interference method for symbiotic radio networks based on meta-learning according to claim 9, characterized in that, The reflection element optimization model is a five-tuple multi-objective optimization model based on Markov game; The specific reflection element optimization model is , where represents the set of agents, including multiple reconfigurable intelligent surfaces; represents the set of states, containing the environmental data observed by the reconfigurable intelligent surfaces; represents the set of actions, including the selection of the defense mode or the symbiotic mode, and the adjustment of the reflection phase of the reconfigurable intelligent surface; represents the state transition probability, including the possibility that the agent switches to the next state after executing the current action according to the current policy; represents the reward function, including the reward function of the agent after executing the current action according to the current policy, reflecting the impact of the current action on the policy.

Images