Joint beam forming optimization method for RIS-assisted low-altitude communication system

Abstract

The invention discloses a joint beam forming optimization method for an RIS auxiliary low-altitude communication system. The problem of performance bottleneck caused by interaction of active beam forming and passive beam forming is solved by adopting a genetic algorithm based on a dynamic adaptive theory. Aiming at the limitation of a standard optimization method, a hybrid strategy of genetic algorithm optimization and a full-connection neural network is explored, and intelligent collaborative setting of system parameters is realized. Simulation results show that the hybrid algorithm not only simplifies the calculation process to obtain the optimal network parameters, but also effectively optimizes the system structure by dynamically adjusting the RIS reflection configuration. On the basis, the signal transmission quality is improved, interference is reduced, and it is guaranteed that the RIS-assisted low-altitude unmanned aerial vehicle intelligent communication system stably and efficiently operates in a complex wireless transmission environment.

Description

Technical Field

[0001] The invention belongs to the technical field of wireless communications and relates to a joint beamforming optimization method for a RIS-assisted low-altitude communication system. Background Art

[0002] With the continuous development of communication technology, especially the strong development of the fifth generation (5G) technology, a series of applications have become more and more comprehensive, such as smart manufacturing, smart cities, telemedicine, etc. In this context, intelligent metasurface (RIS), as an emerging wireless communication technology, shows significant potential for future 6G communication systems. Specifically, RIS consists of many adjustable passive reflection units. By adaptively controlling the phase and amplitude of each reflection unit, the beamforming and direction control of the incident signal can be achieved, thereby enhancing the signal reception quality and effectively suppressing multipath interference. In addition, through precise beamforming, RIS is able to reduce interference between users and achieve efficient reuse of spatial resources. Moreover, the beam can be dynamically adjusted to adapt to changes in user positions, ensuring a continuous interference suppression effect.

[0003] Unlike established intelligent massive multiple-input multiple-output (MIMO) communication systems, RIS introduces a passive reflective surface that can be integrated into modern mobile communication systems without any transmitting or receiving functional modules. The key advantage of RIS integration lies in its unparalleled ability to dynamically manipulate the propagation trajectory of the incident signal. Through fine-tuning, efficient spatial resource reuse is achieved, and phase offsets and amplitude changes are imparted by the numerous reflective elements embedded within the RIS surface. This process provides accuracy and adaptability beyond traditional signal control methods. Based on this, signal coverage and communication efficiency can be significantly improved without increasing energy consumption. Therefore, the potential performance of RIS cannot be achieved only through close cooperation between the base station (BS), user terminals and RIS. In addition, the joint optimization of beamforming parameters to improve the performance of outage probability and symbol error ratio (SER) is studied. However, due to its highly non-convex nature, the joint beamforming problem is complex and difficult to obtain a direct closed-form solution. Traditional methods are often hindered by falling into local optimality, which hinders the realization of high performance. Therefore, scientists in academia and industry are studying suboptimal methods. Among these schemes, optimization methods based on quadratic programming are explored due to their simplicity. Through convexification and linearization, complex non-convex problems are transformed into tractable quadratic programming problems, thereby reducing computational complexity. However, its local optimal solution is inherently limited, and it is difficult to obtain global optimization in scenarios with high performance requirements. Summary of the invention

[0004] The purpose of the present invention is to provide a joint beamforming optimization method for a RIS-assisted low-altitude communication system to solve the problems raised in the above background technology.

[0005] To achieve the above object, the present invention provides the following technical solution: a joint beamforming optimization method for a RIS-assisted low-altitude communication system, comprising the following steps: Step S1: construct a downlink intelligent MIMO wireless communication system model including a base station, an intelligent metasurface, and a receiving end, and clarify the signal transmission process and the joint beamforming optimization goal to maximize the signal-to-noise ratio and minimize the bit error rate; Step S2: Use FCNN to build an autoencoder, update it through the Adam optimizer, calculate the gradient, back propagate, and minimize the cross entropy loss function through multiple layers of transformation and processing. At the same time, the user-side signal flows through two layers of fully connected layers and activation function processing; Step S3: Optimizing network parameters using genetic algorithm to reduce bit error rate; Step S4: Normalize the optimization result of the bit error rate to meet the amplitude limit of the inactive beamforming matrix and ensure the effectiveness of the algorithm.

[0006] Preferably, the specific steps of step S1 are: Step 101: The signal transmitted by the base station passes through the channel Arrives at RIS, and then RIS uses the reflection phase matrix Reflected back, the reflected signal matrix on RIS is expressed as:

[0007] in, represents the transmitted signal vector; Step 102: Output obtained through direct and indirect paths at the client It is expressed as:

[0008] in, represents the direct link channel matrix, represents the channel matrix connecting the base station and RIS, is the channel between RIS and the receiver, is the reflection phase matrix of RIS, is the base station’s transmit beamforming vector, is the transmitted signal vector; the channel response in the downlink follows an independent and identically distributed distribution with zero mean, represents the additive white Gaussian noise vector at the receiver; Step 103: Calculate the system signal-to-noise ratio, and the received signal power is , the noise power is Therefore, SNR is expressed as: ; Step 104: Calculate the maximum signal-to-noise ratio, expressed as:

[0009] in, Indicates The phase adjustment of the reflective components (in radians), is the base station’s transmit beamforming vector, is the maximum power, is the square root of the sum of the squares of the vector elements, yes The square of the Frobenius norm of .

[0010] Preferably, the specific steps of step S2 are: Step 201: Use FCNN to implement the transmitter, channel and receiver as autoencoders and define the trainable weight values ​​as and , used to connect adjacent layers, adjust the network model parameters of the sender and receiver to simultaneously reproduce the sender's input at the receiver, and use the Adam optimizer to update the transmitter weights , RIS reflection phase matrix and receiver weights , the weights are calculated based on the derivative of the cross entropy loss function The gradient of , specifically expressed as:

[0011]

[0012]

[0013] in represents the learning rate, is the CE loss function set by the algorithm, is the gradient vector of the function at each point in space; Step 202: FCNN is constructed based on the connection weights, the bias between neurons and the nonlinear characteristics of the activation function. Through the linear combination of weights and biases and the nonlinear transformation provided by the activation function, complex feature extraction and pattern recognition are performed on the data. The signal received by the user end passes through two fully connected layers, which is specifically expressed as follows:

[0014]

[0015] in, Is a user The first weight vector output, Is a user The second weight vector output, It is the signal that users obtain through direct and indirect connections. Is a user The first bias of the output, Is a user The second bias of the output, v is the sigmoid activation function, Indicates user Output after the second layer.

[0016] Preferably, the specific steps of step S3 are: Step 301: Represent the FCNN as a weight matrix and the bias vector in the form of Represents the number of layers in the network. For each weight and bias, first convert them to the range The floating point numbers in the data are then converted into binary strings of length 4. The encoding of individuals and populations is as follows:

[0017]

[0018] in, Indicates training individuals, parameters represents the population size, Indicates the population size; Step 302: Use a neural network to train each individual , calculated by forward propagation output , therefore, the loss function is given by:

[0019] in, represents the actual value that matches the th sample, is the predicted value associated with the th sample; Step 303: Evaluate the fitness and save the optimal chromosome, and use the roulette algorithm to select the individual with the highest fitness: ; Step 304: Perform crossover and mutation operations, re-evaluate fitness and update the optimal chromosome. If the fitness and chromosome do not meet the termination conditions, continue to perform selection, crossover and mutation operations, re-evaluate fitness and update the optimal chromosome. Otherwise, output the result; the new population is: ; Step 305: When the maximum allowed number of iterations is reached, the training process ends, and finally the neural network parameters corresponding to the individual with the best fitness are obtained: .

[0020] Preferably, the specific method of the normalization process in step S4 is as follows:

[0021] in, Indicates The iteration coefficient, parameter express , , ..., The maximum value in .

[0022] Compared with the prior art, the present invention has the following advantages: (1) An innovative optimization strategy is proposed. A hybrid optimization strategy (GAO-FCNN) based on genetic algorithm (GA) and fully connected neural network (FCNN) is proposed. It combines the advantages of both and effectively solves the bit error rate optimization problem of joint beamforming in RIS-assisted low-altitude UAV intelligent communication system, avoiding the limitation of traditional methods converging to local optimality. (2) The performance is significantly improved. Numerical results show that compared with the traditional RIS-assisted low-altitude UAV intelligent communication system, the algorithm achieves a signal-to-noise ratio gain of about 2.5 dB, significantly improves the signal transmission quality of the system, reduces the bit error rate, and enhances the reliability and stability of the system. (3) Adapting to complex environments, the algorithm can dynamically adjust the RIS reflection configuration and optimize the system structure, ensuring the efficient and stable operation of the RIS-assisted low-altitude UAV intelligent communication system in complex wireless transmission environments, effectively reducing interference and improving the overall performance of the system; (4) Efficient parameter acquisition simplifies the calculation process of obtaining the optimal network parameters. Through the feature extraction and pattern recognition capabilities of FCNN and the global search capabilities of GA, the optimal parameter configuration can be quickly found, thereby improving the efficiency of system optimization. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 : Schematic diagram of the intelligent downlink MIMO communication system assisted by RIS; Figure 2 :BER and SNR relationship curves of different systems; Figure 3 :Relationship curve between BER and user moving distance of different systems when SNR=0; Figure 4: Loss value change curve during GAO-FCNN algorithm training. DETAILED DESCRIPTION

[0024] The present invention is further explained below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, various equivalent forms of modifications to the present invention by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0025] In the scheme of the present invention, the main purpose is to establish a beamforming method for a RIS-assisted low-altitude UAV intelligent communication system optimized by a GAO (genetic optimization algorithm)-FCNN algorithm.

[0026] It includes the following 4 steps: The first step is to build a downlink of the intelligent MIMO wireless communication system for the RIS-assisted intelligent MIMO wireless communication system model. The base station (BS) has antennas, the intelligent metasurface (RIS) has reflection unit, the receiving end has The optimization goal of the joint beamforming problem is to optimize the system performance by collaboratively adjusting the transmission parameters (including the BS transmit beamforming vector and the RIS reflection phase matrix), maximizing the signal-to-noise ratio (SNR) and minimizing the bit error rate (BER) as indicators for improving performance, and optimizing the SNR of the received signal to achieve the optimization goal. The specific steps are: Step 101: The signal transmitted by the base station passes through the channel Arrives at RIS, and then RIS uses the reflection phase matrix Reflected back; the reflection signal matrix on RIS (i.e. the signal reflected by RIS) can be expressed as:

[0027] in, Represents the transmitted signal vector.

[0028] Step 102: Output obtained through direct and indirect paths at the client It is expressed as:

[0029] in, represents the direct link channel matrix, represents the channel matrix connecting the base station and RIS, is the channel between RIS and the receiver, is the reflection phase matrix of RIS, is the base station’s transmit beamforming vector. The channel response in the downlink follows an independent and identically distributed distribution with zero mean, represents the additive white Gaussian noise (AWGN) vector at the receiver; Step 103: Calculate the system signal-to-noise ratio (SNR). SNR is defined as the ratio of the received signal strength to the noise level. In the case of multiple antennas, the received signal power is , the noise power is (The noise power at each receiving antenna is ), Refers to the noise power at the receiving end. Therefore, SNR is expressed as: ; Step 104: Calculate the maximum signal-to-noise ratio SNR, which can be expressed as:

[0030] in, Indicates The phase adjustment of the reflective components (in radians), is the base station’s transmit beamforming vector, is the maximum power, is the square root of the sum of the squares of the vector elements, yes The square of the Frobenius norm of .

[0031] The second step is to process the signal by constructing and training FCNN (fully connected neural network). The transceiver is constructed as an autoencoder using FCNN. After determining the parameters, it is updated using the Adam optimizer. The gradient backpropagation is calculated based on the cross entropy loss function. Weight regularization is used to prevent overfitting. The loss function is minimized during training to obtain the bit sequence. The input bit sequence is processed by three layers of fully connected layers. The user-side signal stream is also processed by two layers of fully connected layers and activation functions to complete the construction and training of FCNN. The specific steps are as follows: Step 201: Implement the transmitter, channel, and receiver as autoencoders using FCNN. Define the trainable weight values ​​as and , used to connect adjacent layers. Fine-tune the network model parameters of the sender and receiver to reproduce the sender's input at the receiver; use the Adam optimizer to update the transmitter weights , RIS reflection phase matrix and receiver weights ; Calculate weights based on the derivative of the cross entropy (CE) loss function The gradient of Specifically expressed as:

[0032]

[0033] , in represents the learning rate, is the CE loss function set by the algorithm, is the gradient vector of the function at each point in space; Step 202: FCNN is constructed based on the connection weights, the bias between neurons and the nonlinear characteristics of the activation function. Through the linear combination of weights and biases and the nonlinear transformation provided by the activation function, complex feature extraction and pattern recognition are performed on the data. The signal received by the user end passes through two fully connected layers, which is specifically expressed as: , in, is the communication connection channel of the first group, Is a user The first weight vector output, Is a user The second weight vector output, Is a user The first bias of the output, Is a user The second bias of the output, v is the sigmoid activation function, Indicates user Output after the second layer.

[0034] The third step is to use genetic algorithms to optimize network parameters to reduce the bit error rate. In the preparation stage, the parameters are converted and the population is initialized. Then the individuals are trained and the fitness is evaluated. Then, the selection, crossover, and mutation operations are performed in sequence to update the population and update the optimal value. At the same time, the optimization variables are normalized during iteration to meet the amplitude limit until the termination condition is met. Finally, the optimization parameters are obtained to achieve bit error rate optimization. The specific steps are as follows: Step 301: Represent the FCNN as a weight matrix and the bias vector in the form of Represents the number of layers in the network. For each weight and bias, first convert them to the range The floating point number in is then converted into a binary string of length 4. The encoding form of individuals and populations is:

[0035]

[0036] in, Indicates training individuals, parameters represents the population size, Indicates the population size.

[0037] Step 302: Use a neural network to train each individual , calculated by forward propagation, the output . Therefore, the loss function is given by:

[0038] in, represents the actual value that matches the th sample, is the predicted value associated with the th sample; Step 303: Evaluate the fitness and save the optimal chromosome, and use the roulette algorithm to select the individual with the highest fitness: ; in, is the fitness value of the i-th individual; Step 304: Perform crossover and mutation operations, re-evaluate fitness and update the optimal chromosome. If the fitness and chromosome do not meet the termination conditions, continue to perform selection, crossover, and mutation operations, re-evaluate fitness and update the optimal chromosome. Otherwise, output the GAO result; the new population is: in, Refers to the kth worthy new individual selected according to the individual's fitness and cumulative probability in the roulette selection method; Step 305: When the maximum allowed number of iterations is reached, the training process ends, and finally the neural network parameters corresponding to the individual with the best fitness are obtained: .

[0039] The fourth step is to normalize the optimization results in iterations to meet the amplitude limit of the inactive beamforming matrix and ensure the effectiveness of the algorithm.

[0040] Optimization Normalization is performed at each iteration:

[0041] in, Indicates The iteration coefficient, parameter express , , ..., The maximum value in .

[0042] Based on the above derivation, the system contains 80 base station antennas, and the base station adopts a uniform linear array. There are 256 RIS units and 20 user antennas. The training cycle is set to 200 rounds. The base station is located at the coordinates (0 m, -40 m), and the user terminal is located between (0 m, 0 m) and (100 m, 0 m). The RIS is located at the coordinates (60 m, 10 m). Figure 2 The figure shows the variation of BER performance with signal-to-noise ratio (SNR) in the intelligent MIMO communication system without RIS, the RIS-assisted low-altitude UAV intelligent communication system, the RIS-assisted low-altitude UAV intelligent communication system using FCNN, and the RIS-assisted low-altitude UAV intelligent communication system using the GAO-FCNN algorithm. The green line corresponds to the MIMO communication system without RIS as the baseline. The blue line shows the improved BER performance of the MIMO intelligent system assisted by RIS. The red line shows the RIS-assisted intelligent MIMO system using FCNN. The pink line depicts the BER performance of the RIS-assisted low-altitude UAV intelligent communication system using GAO-FCNN. At that time, the RIS-assisted low-altitude UAV intelligent communication system used the GAO-FCNN algorithm. The performance curves show that as the SNR increases, all BERs of the three intelligent communication systems decrease. However, the BER of FCNN remains basically unchanged at different SNR values. Although the RIS-assisted MIMO communication system is supported by FCNN, it can still maintain a stable BER in a high noise value environment. It is worth noting that a high BER means that the data transmission quality of the system may not be efficient enough, thus affecting the user experience and system reliability. The beamforming performance of the RIS-assisted low-altitude UAV intelligent communication system using GAO-FCNN is slightly improved when the SNR varies between 3 and 4 and slowly decreases to 5. When the BER performance value is 0.001, a performance gain of 2.5dB is achieved when 256 RIS elements are established in the system. The results show that the GAO-FCNN algorithm optimizes the signal processing in the BS, RIS, and users for RIS-assisted intelligent communication. It surpasses traditional methods by discovering the optimal network parameters. In addition, it mitigates the bit error rate spikes caused by relying on neural networks.

[0043] like Figure 3As shown, the BER performance changes when the user moves at different distances in these four RIS deployment scenarios when the SNR is 0 are shown. The green line represents the traditional RIS-free intelligent MIMO system. The blue line shows the BER performance of the RIS-assisted intelligent MIMO system after enhancement. The red line highlights the RIS-assisted MIMO intelligent system enhanced by FCNN. The pink line depicts the BER performance of the RIS-assisted low-altitude UAV intelligent communication system using the GAO-FCNN algorithm. In the RIS-free system, the bit error rate remains basically unchanged as the distance increases in system utilization. It is worth noting that in the system using FCNN, the BER also remains basically unchanged as the distance increases, showing a certain stability. However, the error rate will increase during the same data transmission process, affecting the user experience and system reliability. Then the results show that the BER characteristics of the RIS-assisted low-altitude UAV intelligent communication system using GAO-FCNN gradually decrease with increasing distance, reaching the lowest value at 60m, and then gradually increase. And it is shown that the BER characteristics of the RIS-assisted low-altitude UAV intelligent communication system under the GAO-FCNN algorithm are improved compared with the other three traditional communication systems. It can be observed that the scheme proposed here can achieve a lower BER compared with the existing alternating schemes. Therefore, it is concluded that the RIS-assisted low-altitude UAV intelligent communication system under the GAO-FCNN algorithm is superior to the other three systems.

[0044] like Figure 4 As shown in the figure, the GAO-FCNN algorithm is used to train the RIS-assisted low-altitude UAV intelligent communication system with a training SNR of 6dB. In the blue part, the system maintains a fixed noise power level for the standard deviation and the mean. The green part has an adjustable noise power setting with two parameters. The red part represents the dynamic adjustment of the standard deviation and the mean. In addition, it is shown that the GAO-FCNN algorithm is trained with a training SNR of 6dB in the RIS-assisted low-altitude UAV intelligent communication system. As the training cycle progresses, the loss value of the optimized learning first decreases rapidly, then gradually decreases, and finally reaches almost the optimal performance in the 64th round. To prevent the loss value from increasing, the optimization can be stopped once the training reaches 64 rounds. Compared with the two learning methods of fixed noise power and adjustable noise power, although the optimized learning has a higher loss value at the beginning, it converges to the best performance in only 9 cycles, and achieves a smaller standard deviation and faster training speed, and then remains stable.

[0045] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations of the present invention. A person skilled in the art may change, modify, replace and vary the above embodiments within the scope of the present invention.

Claims

1. A joint beamforming optimization method for RIS-assisted low-altitude communication system, characterized in that: The following steps are involved: Step S1: construct a downlink intelligent MIMO wireless communication system model including a base station, an intelligent metasurface, and a receiving end, and clarify the signal transmission process and the joint beamforming optimization goal to maximize the signal-to-noise ratio and minimize the bit error rate; Step S2: Use FCNN to build an autoencoder, update it through the Adam optimizer, calculate the gradient, back propagate, and minimize the cross entropy loss function through multiple layers of transformation and processing. At the same time, the user-side signal flows through two layers of fully connected layers and activation function processing; Step S3: Optimizing network parameters using genetic algorithm to reduce bit error rate; Step S4: Normalize the optimization result of the bit error rate to meet the amplitude limit of the inactive beamforming matrix and ensure the effectiveness of the algorithm.

2. The joint beamforming optimization method for RIS-assisted low-altitude communication system according to claim 1, characterized in that: The specific steps of step S1 are: Step 101: The signal transmitted by the base station passes through the channel Arrives at RIS, and then RIS uses the reflection phase matrix Reflected back, the reflected signal matrix on RIS is expressed as: ； in, represents the transmitted signal vector; Step 102: Output obtained through direct and indirect paths at the client It is expressed as: ； in, represents the direct link channel matrix, represents the channel matrix connecting the base station and RIS, is the channel between RIS and the receiver, is the reflection phase matrix of RIS, is the base station’s transmit beamforming vector, is the transmitted signal vector; the channel response in the downlink follows an independent and identically distributed distribution with zero mean, represents the additive white Gaussian noise vector at the receiver; Step 103: Calculate the system signal-to-noise ratio, and the received signal power is , the noise power is Therefore, SNR is expressed as: ； in, is the noise power at each receiving antenna, is the noise power at the receiving end; Step 104: Calculate the maximum signal-to-noise ratio, expressed as: ； ； in, Indicates The phase adjustment of the reflective components (in radians), is the base station’s transmit beamforming vector, is the maximum power, is the square root of the sum of the squares of the vector elements, yes The square of the Frobenius norm of .

3. The joint beamforming optimization method for RIS-assisted low-altitude communication system according to claim 1, characterized in that: The specific steps of step S2 are: Step 201: Use FCNN to implement the transmitter, channel and receiver as autoencoders and define the trainable weight values ​​as and , used to connect adjacent layers, adjust the network model parameters of the sender and receiver to simultaneously reproduce the sender's input at the receiver, and use the Adam optimizer to update the transmitter weights , RIS reflection phase matrix and receiver weights , the weights are calculated based on the derivative of the cross entropy loss function The gradient of , specifically expressed as: ； ； ； in represents the learning rate, is the CE loss function set by the algorithm, is the gradient vector of the function at each point in space; Step 202: FCNN is constructed based on the connection weights, the bias between neurons and the nonlinear characteristics of the activation function. Through the linear combination of weights and biases and the nonlinear transformation provided by the activation function, complex feature extraction and pattern recognition are performed on the data. The signal received by the user end passes through two fully connected layers, which is specifically expressed as follows: ； ； in, is the communication connection channel of the first group, Is a user The first weight vector output, Is a user The second weight vector output, It is the signal that users obtain through direct and indirect connections. Is a user The first bias of the output, Is a user The second bias of the output, v is the sigmoid activation function, Indicates user Output after the second layer.

4. The joint beamforming optimization method for RIS-assisted low-altitude communication system according to claim 1, characterized in that: The specific steps of step S3 are: Step 301: Represent the FCNN as a weight matrix and the bias vector in the form of Represents the number of layers in the network. For each weight and bias, first convert them to the range The floating point numbers in the data are then converted into binary strings of length 4. The encoding of individuals and populations is as follows: ； ； in, Indicates training individuals, parameters represents the population size, Indicates the population size; Step 302: Use a neural network to train each individual , calculated by forward propagation output , therefore, the loss function is given by: ； in, represents the actual value that matches the th sample, is the predicted value associated with the th sample; Step 303: Evaluate the fitness and save the optimal chromosome, and use the roulette algorithm to select the individual with the highest fitness: ； in, is the probability of the ith individual, is the fitness value of the i-th individual; Step 304: Perform crossover and mutation operations, re-evaluate fitness and update the optimal chromosome. If the fitness and chromosome do not meet the termination conditions, continue to perform selection, crossover and mutation operations, re-evaluate fitness and update the optimal chromosome. Otherwise, output the result; the new population is: ； Step 305: When the maximum allowed number of iterations is reached, the training process ends, and finally the neural network parameters corresponding to the individual with the best fitness are obtained: ， in, is the unit with the maximum score, is the probability of the kth unit.

5. The joint beamforming optimization method for RIS-assisted low-altitude communication system according to claim 1, characterized in that: The specific method of the normalization process in step S4 is as follows: ； in, Indicates The iteration coefficient, parameter express , , ..., The maximum value in .

Images