Intelligent reflecting surface assisted communication resource allocation method

Abstract

The application relates to an intelligent reflecting surface assisted communication resource allocation method and belongs to the technical field of wireless communication. First, an RIS assisted D2D communication system model is proposed, a feasible solution is found by adopting a DL based algorithm, and the solution is fed into three independent DNN modules, namely a module for intelligent reflecting surface phase, a module for D2D user transmission power and a channel selection module. The method can reduce the training overhead caused by supervised learning, solve the passive beamforming design problem of RIS by training a deep neural network, and solve the large dimension optimization problem with low complexity.

Description

Technical Field

[0001] This invention belongs to the field of wireless communication technology and relates to the design of an intelligent reflector-assisted communication resource allocation algorithm using deep learning in an intelligent reflector-assisted multi-cellular user network. Background Technology

[0002] The rapid growth of mobile devices has brought unprecedented increases in mobile data traffic demand, posing significant challenges to current cellular networks. To address this challenge, Device-to-Device (D2D) technology has been proposed, allowing direct communication between near-end devices. Compared to traditional cellular technologies, D2D technology has two main advantages. First, D2D technology can effectively reduce the burden on base stations through traffic offloading. Second, D2D technology can shorten the signal propagation distance between two devices, thereby significantly improving spectrum efficiency. There are two main methods for D2D to share cellular spectrum: bottom-level D2D and overlay D2D. Typically, allowing D2D links to share uplink spectrum with cellular links can alleviate spectrum shortages. However, unavoidable interference necessitates protecting cellular networks from interference under stringent communication quality requirements.

[0003] Fortunately, in recent years, researchers have developed an innovative and revolutionary technology: the Reconfigurable Intelligent Surface (RIS). This technology can effectively eliminate D2D interference and meet required data rates. An RIS is a plane composed of numerous low-cost passive reflective elements, each capable of independently inducing amplitude and / or phase changes in the incident signal, thereby collaboratively achieving precise three-dimensional (3D) reflected beamforming. For each element, its electromagnetic response—the phase shift of the incident signal—can be adjusted in a software-defined manner. Utilizing its reconfigurability, the incident signal can be superimposed and reflected by adjusting the phase shift, forming a beam steering beneficial to the user and effectively controlling multipath effects. Therefore, by adjusting the amplitude and / or phase shift of the RIS elements, the expected power gain of the receiver and the destructive mitigation of interference can be achieved, thereby improving spectral and energy efficiency. To enhance the performance of D2D communication systems, RIS can be deployed within the system, and the reflection of electromagnetic waves impacting its surface can be controlled by configuring the RIS.

[0004] Currently, most RIS-assisted D2D communication systems utilize relaxation-alternating iterative optimization techniques. The fundamental reason for the signaling problem is that existing solutions optimize for each snapshot of the network state; once the network state changes, the resulting solution is no longer optimal. Deep learning (DL), as a model-free learning technique, holds promise for application in the communication field. Supervised learning is a common method in deep learning, using known real-world data as training data to train models that more closely resemble the actual system environment and possess stronger model generalization capabilities. However, supervised learning requires exhaustively searching for each sample to solve its optimization problem and generate labels during the training data generation phase. With large amounts of codebook data, this training overhead becomes excessively high. Summary of the Invention

[0005] The technical problem to be solved by this invention is:

[0006] To avoid the problems associated with supervised learning, such as the need for a large number of training labels, this invention provides an intelligent reflective surface-assisted communication resource allocation method. This method reduces the training overhead of supervised learning by training a deep neural network (DNN) to solve the passive beamforming design problem of RIS (Resilient Beamforming), thus addressing a large-dimensional optimization problem with lower complexity. Furthermore, due to the dynamic deployment characteristics of UAVs, they can adjust their flight status in real time according to the actual communication scenario and physical environment, exhibiting a certain degree of adaptability to complex environments. This enhances the system's adaptability and anti-interference capabilities; therefore, we use UAVs as aerial base stations.

[0007] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0008] A method for allocating communication resources with an intelligent reflective surface, characterized in that:

[0009] The processing steps at the drone base station transmitter are as follows:

[0010] Step 1: The transmitter includes a drone base station, a RIS with M reflective elements, K channels, and D D2D pairs, where each channel contains a cellular user CUE; in addition, using This indicates whether the i-th D2D pair uses the k-th channel; that is, if the transmitter of the d-th D2D pair transmits data through the k-th channel, then... otherwise Assuming that each D2D pair can only use one channel at a time, i.e. Assume that the drone base station, CUE, and D2D are all equipped with a single antenna. The drone base station is located at (x, y, h), and the l-th CUE is located at (x, y, h). l ,y l,z l ), the RIS position is (0, y ris ,z ris D2D users are randomly distributed within a circle of radius R. The location of the i-th D2D transmitter (DT) is denoted as . The distance between DT and its corresponding receiver DR is L, and the position of the i-th DR is denoted as L.

[0011] Furthermore, assuming the RIS is equipped with M reflective elements, the reflection coefficient matrix Θ is expressed as... The m-th element of this diagonal matrix is ​​represented as Where θ m ∈[0,2π) represents the phase shift of the m-th reflecting element, |β m |≤1 represents the amplitude of the m-th reflecting element; in practice, the phase shift of each element can only have a finite number of discrete values, and the l-th discrete value is expressed as Where L = 2 b , b represents the number of bits used to quantize the phase shift level L; then, it is assumed that RIS performs ideal reflection, so that the signal power of each reflecting element is lossless, i.e., the amplitude reflection coefficient β. m =1; and assume that the transmit power of DT i is expressed as p i And can be quantized as N P A long walk, that is Its maximum value is P D , i.e. P D =P Np ;

[0012] Step 2: The drone base station generates X bit information a[x], x=0,1,…,X-1;

[0013] Step 3: Modulate the information generated in Step 2. The modulated signal is... And satisfy ||s|| 2 =1;

[0014] Step 4: Send the signal s generated in step 3 into the channel;

[0015] The steps for establishing a drone-to-user channel are as follows:

[0016] Step 5: The signal transmitted by the drone base station initially propagates in free space. At this time, there are no tall buildings, trees or other objects in the space to block it, and the drone can directly transmit the signal. At this time, a line-of-sight (LOS) link is formed, and the resulting fading model is mainly a free space loss model. As the signal continues to propagate, there is obstruction between the drone base station and the ground user. At the same time, it is affected by electromagnetic wave reflection and scattering, forming a non-line-of-sight (NLOS) link.

[0017] Step 6: LOS and NLOS in the channel gain occur with a certain probability. Assume the ground node coordinates are (x... i ,y i ,z i The coordinates of the UAV base station are (x, y, h), and the distance between the UAV and the node is... The LOS component probability of the node is:

[0018]

[0019] Where A and B are constants related to the environment. The NLOS component probability of a node is: P NLoS =1-P LoS ;

[0020] Step 7: The UAV to D2D user i-channel gain is:

[0021]

[0022] Where α0 is the path loss coefficient between the user and the drone link, and η is the correlation coefficient with the NLOS link;

[0023] Step 8: The distance between the drone and the lth CUE user is d l Substitute P from step 6 LOS and P NLOS In the process, the Rayleigh distribution coefficient is randomly generated. The channel gain is then:

[0024]

[0025] Step 9: The distance between the drone and the RIS is Ignoring the positional differences of each component in the RIS, d r Substitute P from step 6 LOS and P NLOS In the process, the Rayleigh distribution coefficient is randomly generated. The channel gain is then:

[0026] Step 10: The distance between the drone and the i-th D2D user is d i Substitute P from step 6 LOS and P NLOS In the process, the Rayleigh distribution coefficient is randomly generated. The channel gain is then:

[0027]

[0028] The other link channel construction and processing steps are as follows:

[0029] Step 11: Considering the random Rayleigh distribution and path loss for other link channel gains, the path loss is:

[0030]

[0031] Where d is the distance between nodes, and ρ and υ represent the path loss and path loss coefficient at 1m, respectively.

[0032] Step 12: The distance between RIS and the l-th CUE user is d r,l Substitute the values ​​into the loss function from step 11 and randomly generate the Rayleigh distribution coefficients. The channel gain is then:

[0033]

[0034] Step 13: The distance between RIS and the i-th DR is d ri Substitute the values ​​into the loss function from step 11 and randomly generate the Rayleigh distribution coefficients. The channel gain is then: The distance between RIS and the i-th DT is d ir Substitute the values ​​into the loss function from step 11 and randomly generate the Rayleigh distribution coefficients. The channel gain is then:

[0035]

[0036] Step 14: The distance between the l-th CUE user and the i-th D2D user is d lj Substitute the values ​​into the loss function from step 11 and randomly generate the Rayleigh distribution coefficients. The channel gain is then:

[0037] Step 15: The distance between the i-th DT and the i′-th DR user is d ii′ Substituting into the loss function of step 11, let the loss matrix be denoted as L. D×D The element in the i-th row and i′-th column is denoted as Randomly generate Rayleigh distribution coefficients The channel gain is then: The channel gain between the i-th DT and DR is denoted as h. i,i The channel gain between the i-th DT and the i′-th DR is denoted as h. i,i′ ;

[0038] The processing steps at the receiving end are as follows:

[0039] Step 16: The SINR received by the l-th CUE is:

[0040]

[0041] Where P c p represents the transmit power of the UAV. i For the transmit power of the i-th DT, the additive white Gaussian noise at the k-th CUE user follows:

[0042] Step 17: The SINR received by the i-th DR is:

[0043]

[0044] in and These represent interference from other D2D users and CUE users on the k-th channel, respectively; the additive white Gaussian noise at the i-th DR follows the following order:

[0045] Step 18: Therefore, the achievable rate of the i-th DR in the k-th channel is:

[0046] R i [k] = log2(1 + SINR) i [k])

[0047] Step 19: Calculate the sum rate of all DRs in the k-th channel:

[0048]

[0049] Step 20: To ensure the QoS of the CUE, the CUE's SINR should be greater than the minimum SINR threshold. thr Therefore, resource allocation is formulated as a non-convex optimization problem as follows:

[0050]

[0051]

[0052]

[0053]

[0054]

[0055]

[0056]

[0057]

[0058]

[0059] Where a = [a1, a2, ..., a i ,...a D ], This represents the index vector used for the i-th pair of D2D channels, p = [p1, p2, ... p]. D ] represents the power of D V2V units, I T Indicates the maximum interference limit caused to cellular transmission;

[0060] Step 21: Optimize the above non-convexity problem using a deep learning algorithm based on unsupervised learning. The specific steps are as follows:

[0061] Step 22: Preprocess the input data. The preprocessing function g(·) is:

[0062]

[0063]

[0064] Step 23: Reshape the preprocessed data into a one-dimensional length vector and then feed it into three independent DNN modules: the smart reflector phase module, the D2D user transmit power module, and the channel selection module.

[0065] The specific implementation process for each module is as follows:

[0066] Step 24: The output of the intelligent reflective phase module is divided into M groups, each containing L elements; then, each group's output is input into a separate softmax layer block, and softmax operations are performed on each block, with vector z... i This represents the probability of each reflection phase level being selected, i.e., z i The m-th element corresponds to the m-th reflective element using phase level. The probability of;

[0067] Step 25: The transmit power module calculates the transmit power levels for all D2D operations, and its output is N. p ×D; The output of the transmit power module is divided into D groups, each group containing N... p Elements; using vector q i q represents the probability of each transmit power level being selected. i The l-th element corresponds to D2D TP i using transmit power level P l The probability of;

[0068] Step 26: Feed the reshaped channel gain into the channel selection module with K×D output, then divide its output into D groups and send them to the output of... There are D independent softmax blocks; the k-th output of the i-th softmax block can be interpreted as the probability of D2D TP i choosing channel k; during training, the output of this module It is used directly for channel selection, but during inference, it is set... And set all other k. So that constraints can be satisfied in actual implementation

[0069] Step 27: Obtain the maximum sum rate of D2D users by minimizing the loss function of the entire neural network; the custom neural network loss function is:

[0070]

[0071] Where g1 is the function to eliminate binarization error:

[0072]

[0073] Where τ,λ1,λ2,λ3,λ4,κ are network hyperparameters.

[0074] A computer system is characterized by comprising: one or more processors, and a computer-readable storage medium for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the method described above.

[0075] A computer-readable storage medium is characterized by storing computer-executable instructions, which, when executed, are used to implement the above-described method.

[0076] The beneficial effects of this invention are as follows:

[0077] This invention provides an intelligent reflector-assisted communication resource allocation method. Considering that neural networks can approximate any continuous function, enabling them to simulate the behavior of highly nonlinear and complex systems and serve as universal approximators, this method can be applied to resource allocation in wireless communication systems. It finds feasible solutions by employing a deep learning-based algorithm without using complex mathematical formulas and numerical optimization techniques. The specific algorithm flow is as follows: Figure 2 As shown. The main advantages of this invention are as follows:

[0078] First, due to the flexibility of being unmanned, drones can improve the limitations of poor mobility of traditional base stations by serving as aerial base stations. Therefore, this invention uses drones as aerial base stations. The use of intelligent reflective surfaces can reduce interference caused by LOS links, improve channel quality, and solve the signal attenuation caused by NLOS links.

[0079] Secondly, DL-based optimization schemes can achieve near-optimal performance without explicitly solving complex optimization problems. Then, near-optimal solutions can be found by feeding appropriate input data (e.g., the channel gain of a link in a wireless communication system) into a trained DNN. Compared to traditional methods, DL-based resource allocation schemes are more flexible; the same DNN structure can be trained to achieve different design goals by changing the loss function. Attached Figure Description

[0080] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.

[0081] Figure 1 This is a model diagram of a UAV-D2D system based on a smart reflective surface;

[0082] Figure 2 The schematic diagram of the proposed DL algorithm is shown below.

[0083] Figure 3 For RIS phase optimization modules;

[0084] Figure 4 For power distribution modules;

[0085] Figure 5 For the channel selection module. Detailed Implementation

[0086] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0087] This invention proposes a smart reflector-assisted communication optimization method, including the following:

[0088] The processing steps at the drone base station transmitter are as follows:

[0089] Step 1: The transmitter includes a drone base station, a RIS with M reflective elements, K channels, and D D2D pairs, where each channel contains a cellular user CUE. Additionally, using... This indicates whether the i-th D2D pair uses the k-th channel; that is, if the transmitter of the d-th D2D pair transmits data through the k-th channel, then... otherwise Assuming that each D2D pair can only use one channel at a time, i.e. Assume that the drone base station, CUE, and D2D are all equipped with a single antenna. The drone base station is located at (x, y, h), and the l-th CUE is located at (x, y, h). l ,y l ,z l ), the RIS position is (0, y ris ,z ris D2D users are randomly distributed within a circle of radius R. The location of the i-th D2D transmitter (DT) is denoted as . The distance between DT and its corresponding receiver DR is L, and the position of the i-th DR is denoted as L.

[0090] Furthermore, assuming the RIS is equipped with M reflective elements, the reflection coefficient matrix Θ is expressed as... The m-th element of this diagonal matrix is ​​represented as Where θ m ∈[0,2π) represents the phase shift of the m-th reflecting element, |β m |≤1 represents the amplitude of the m-th reflecting element. In practice, the phase shift of each element can only have a finite number of discrete values, and the l-th discrete value is represented as... Where L = 2 b , where b represents the number of bits used to quantize the phase shift level L. Then, we assume that RIS performs ideal reflection, so that the signal power of each reflecting element is lossless, i.e., the amplitude reflection coefficient β. m =1. Also, assume that the transmit power of DT i is expressed as p. i And can be quantized as N P The distance between the two walks is p i ∈Π={P1,P1,…,P NP}, its maximum value is P D ,Right now

[0091] Step 2: The drone base station generates X bit information a[x], x=0,1,…,X-1;

[0092] Step 3: Modulate the information generated in Step 2. The modulated signal is... And satisfy ||s|| 2 =1;

[0093] Step 4: Send the signal s generated in step 3 into the channel;

[0094] The steps for establishing a drone-to-user channel are as follows:

[0095] Step 5: The signal transmitted by the drone base station initially propagates in free space. At this time, there are no tall buildings, trees, or other objects in the space to block it, and the drone can directly transmit the signal, forming a line-of-sight (LOS) link. The resulting fading model is mainly a free space loss model. As the signal continues to propagate, there is obstruction between the drone base station and the ground user, and it is also affected by electromagnetic wave reflection and scattering, forming a non-line-of-sight (NLOS) link.

[0096] Step 6: LOS and NLOS in the channel gain occur with a certain probability. Assume the ground node coordinates are (x... i ,y i ,z i The coordinates of the UAV base station are (x, y, h), and the distance between the UAV and the node is... The LOS component probability of the node is:

[0097]

[0098] Where A and B are constants related to the environment. The NLOS component probability of a node is: P NLoS =1-P LoS .

[0099] Step 7: The UAV to D2D user i-channel gain is:

[0100]

[0101] Where α0 is the path loss coefficient between the user and the drone link, and η is the correlation coefficient with the NLOS link.

[0102] Step 8: The distance between the drone and the lth CUE user is d l Substitute P from step 6 LOS and P NLOS In the process, the Rayleigh distribution coefficient is randomly generated. The channel gain is then:

[0103]

[0104] Step 9: The distance between the drone and the RIS is Ignoring the positional differences of each component in the RIS, d r Substitute P from step 6 LOS and P NLOS In the process, the Rayleigh distribution coefficient is randomly generated. The channel gain is then:

[0105] Step 10: The distance between the drone and the i-th D2D user is d i Substitute P from step 6 LOS and P NLOS In the process, the Rayleigh distribution coefficient is randomly generated. The channel gain is then:

[0106]

[0107] The other link channel construction and processing steps are as follows:

[0108] Step 11: Considering the random Rayleigh distribution and path loss for other link channel gains, the path loss is:

[0109]

[0110] Where d is the distance between nodes, and ρ and υ represent the path loss and path loss coefficient at 1m, respectively.

[0111] Step 12: The distance between RIS and the l-th CUE user is d r,l Substitute the values ​​into the loss function from step 11 and randomly generate the Rayleigh distribution coefficients. The channel gain is then:

[0112]

[0113] Step 13: The distance between RIS and the i-th DR is d ri Substitute the values ​​into the loss function from step 11 and randomly generate the Rayleigh distribution coefficients. The channel gain is then: The distance between RIS and the i-th DT is d ir Substitute the values ​​into the loss function from step 11 and randomly generate the Rayleigh distribution coefficients. The channel gain is then:

[0114]

[0115] Step 14: The distance between the l-th CUE user and the i-th D2D user is d lj Substitute the values ​​into the loss function from step 11 and randomly generate the Rayleigh distribution coefficients. The channel gain is then:

[0116] Step 15: The distance between the i-th DT and the i′-th DR user is d ii′ Substituting into the loss function of step 11, let the loss matrix be denoted as L. D×D The element in the i-th row and i′-th column is denoted as Randomly generate Rayleigh distribution coefficients The channel gain is then: The channel gain between the i-th DT and DR is denoted as h. i,i The channel gain between the i-th DT and the i′-th DR is denoted as h. i,i′ .

[0117] The processing steps at the receiving end are as follows:

[0118] Step 16: The SINR received by the l-th CUE is:

[0119]

[0120] Where P c This indicates the UAV's transmit power. p i For the transmit power of the i-th DT, the additive white Gaussian noise at the k-th CUE user follows:

[0121] Step 17: The SINR received by the i-th DR is:

[0122]

[0123] in and Inter C [k]=|(h r,i [k]) H Θh r [k]+h i [k]| 2 P c These represent interferences from other D2D users and CUE users on the k-th channel, respectively. The additive white Gaussian noise at the i-th DR follows the following formula:

[0124] Step 18: Therefore, the achievable rate of the i-th DR in the k-th channel is:

[0125] R i [k] = log2(1 + SINR) i [k])

[0126] Step 19: Finally, we can calculate the sum rate of all DRs in the k-th channel as follows:

[0127]

[0128] Step 20: To ensure the QoS of the CUE, the CUE's SINR should be greater than the minimum SINR threshold. thr Therefore, resource allocation is formulated as a nonconvex optimization problem as follows:

[0129]

[0130]

[0131]

[0132]

[0133]

[0134]

[0135]

[0136]

[0137]

[0138] Where a = [a1, a2, ..., a i ,...a D ], This represents the index vector used for the i-th pair of D2D channels, p = [p1, p2, ... p]. D ] represents the power of D V2V units, I T This indicates the maximum interference limit for cellular transmission.

[0139] Step 21: Due to the non-convexity of the objective function, the optimization problem in the above equation is a non-convex mixed-integer programming problem. The performance of the suboptimal solution obtained by traditional optimization methods cannot be guaranteed, and it usually has high complexity. Based on these problems, we propose a deep learning algorithm based on unsupervised learning.

[0140] The proposed DL algorithm is as follows Figure 2 As shown, the specific steps are as follows:

[0141] Step 22: Preprocess the input data. The preprocessing function g(•) is:

[0142]

[0143]

[0144] Step 23: Reshape the preprocessed data into a one-dimensional length vector, and then feed it into three independent DNN modules, namely the intelligent reflector phase module (e.g., Figure 3 As shown), the module for D2D user transmit power (such as...) Figure 4 As shown), and the channel selection module (as shown) Figure 5 (As shown).

[0145] The specific implementation process for each module is as follows:

[0146] Step 24: As Figure 3 As shown, the output of the intelligent reflector phase module is divided into M groups, each containing L elements. Then, each group's output is input into a separate softmax layer block, where softmax operations are performed on the vector z. i This represents the probability of each reflection phase level being selected, i.e., z i The m-th element corresponds to the m-th reflective element using phase level. The probability of.

[0147] Step 25: As Figure 4 As shown, the transmit power module calculates the transmit power levels for all D2D operations, and its output is N. p ×D. The output of the transmit power module is divided into D groups, each group containing N... p Elements. Using vector q i q represents the probability of each transmit power level being selected. i The l-th element corresponds to D2D TP i using transmit power level P l The probability of.

[0148] Step 26: As Figure 5 As shown, the reshaped channel gain is fed into a channel selection module with K×D output, and then its output is divided into D groups, which are then sent to the output of... There are D independent softmax blocks. The k-th output of the i-th softmax block can be interpreted as the probability of choosing channel k in a D2D TP i. During training, the output of this module... It is used directly for channel selection, but during inference, it is set... And set all other k. So that constraints can be satisfied in actual implementation

[0149] Step 27: Obtain the maximum sum rate of D2D users by minimizing the loss function of the entire neural network. The custom neural network loss function is:

[0150]

[0151] Where g1 is the function to eliminate binarization error:

[0152]

[0153] Where τ,λ1,λ2,λ3,λ4,κ are network hyperparameters.

[0154] Example 1:

[0155] like Figure 1 The diagram shows the applicable scenarios for this invention.

[0156] The processing steps at the drone base station transmitter are as follows:

[0157] Step 1: The transmitting end includes a UAV base station, a RIS with 4 reflectors, 2 cellular user devices (CUEs), and 3 D2D pairs. The UAV base station, CUEs, and D2D pairs are each equipped with a single antenna. The UAV base station is located at (0,0,30)m, the RIS is located at (0,10,2)m, and the cellular users and D2D pairs are randomly distributed within a circle with a radius of 10m. The location of the first D2D transmitter (DT) is denoted as (20,60,1)m, and the distance between the DT and its corresponding receiver (DR) is 5m. The location of the kth cellular user is (x... k ,y k 1). The maximum transmit power of the base station is 30W, and the maximum transmit power of the DT is P. D =10W, noise variance In addition, the power quantization level N p =5, phase quantization bits b=3, then L=2 3 =8. The path loss index α0 between the user and the drone link is 3, and the correlation coefficient η with the NLOS link is 20dB. The path loss coefficient ρ at 1m in other links is -30dB, and the path loss index υ under this link is 2.5.

[0158] Step 2: The drone base station generates 128 bits of information;

[0159] Step 3: Modulate the information generated in Step 2. The modulated signal is... And satisfy ||s|| 2 =1;

[0160] Step 4: Send the signal s generated in step 3 into the channel;

[0161] The signals transmitted by the drone base station will form LOS and NLOS links. The drone-to-user channel construction process in the k-th channel is as follows:

[0162] Step 5: LOS and NLOS in the channel gain occur with a certain probability. Assume the ground node coordinates are (x... i ,y i ,z iThe coordinates of the UAV base station are (x, y, h), and the distance between the UAV and the node is... The LOS component probability of the node is:

[0163]

[0164] in The NLOS component probability of a node is: P NLoS =1-P LoS .

[0165] Step 6: The path loss coefficient α0 between the user and the drone link is 3, and the correlation coefficient with the NLOS link is η = 20dB. Therefore, the channel gain from the drone to user i is:

[0166] h i =d -3 P LOS +100d -3 P NLOS

[0167] Step 7: The distance between the drone and the CUE user in the k-th channel is... but P NLoS =1-P LoS Randomly generate Rayleigh distribution coefficients The channel gain is then: h c [k]=((d k ] -3 ×P LoS +100×[d k ) -3 ×P NLoS )×fad UC .

[0168] Step 8: The distance between the drone and the RIS is Ignoring the differences in RIS location, then P NLoS =0.01, randomly generated Rayleigh distribution coefficient The channel gain is then: h r [k]=((29.73) -3 ×0.99+100×(29.73) -3 ×0.01)×fad UR =3.8e-05×fad UR .

[0169] Step 9: The distance between the drone and the i-th D2D user is but P NLoS =1-PLoS Randomly generate Rayleigh distribution coefficients The channel gain is then:

[0170] The other link channel construction and processing steps are as follows:

[0171] Step 10: Considering the path loss coefficient ρ = -30dB at 1m, the path loss exponent υ = 2.5 for this link. For other links, considering the random Rayleigh distribution of channel gain and path loss, the path loss is:

[0172] loss = 0.01 × (d) -2.5

[0173] Step 11: The distance between RIS and the k-th CUE user is but Randomly generate Rayleigh distribution coefficients The channel gain is then: h r,c [k]=loss rk ×fad RC .

[0174] Step 12: The distance between RIS and the i-th DR is but Randomly generate Rayleigh distribution coefficients The channel gain is then: The distance between RIS and the i-th DT is but Randomly generate Rayleigh distribution coefficients The channel gain is then:

[0175]

[0176] Step 13: The distance between the k-th CUE and the i-th D2D user is but Randomly generate Rayleigh distribution coefficients The channel gain is then:

[0177]

[0178] Step 14: The distance between the i-th DT and the i′-th DR user is but Let the loss matrix be L D×D The element in the i-th row and i′-th column is denoted as Randomly generate Rayleigh distribution coefficients The channel gain is then: H = L × fad tr The channel gain between the i-th DT and DR is denoted as h. i,i[k], the channel gain between the i-th DT and the i′-th DR is denoted as h. i,i [k].

[0179] The processing steps at the receiving end are as follows:

[0180] Step 15: The SINR received by the k-th CUE is:

[0181]

[0182] Step 16: The SINR received by the i-th DR in the k-th channel is:

[0183]

[0184] Step 17: The achievable rate of the i-th DR is:

[0185] R i [k] = log2(1 + SINR) i [k])

[0186] Step 18: The sum rate of all DRs in the k-th channel is:

[0187]

[0188] Step 19: To improve the achievable rate of DUE while limiting interference to CUE, an optimization algorithm is needed to perform optimal phase design for RIS and power allocation for D2D users. A minimum SINR threshold is set for D2D users. thr =12dB, minimum achievable data rate threshold R for cellular users d_thr =20dB. Therefore, resource allocation is formulated as a non-convex optimization problem as follows:

[0189]

[0190]

[0191]

[0192]

[0193]

[0194]

[0195]

[0196]

[0197]

[0198] Step 20: This invention proposes a deep learning algorithm based on unsupervised learning to solve this optimization problem. The specific optimization process is as follows: Figure 2 As shown. The entire training process consists of 1000 episodes. The algorithm inputs the current channel information. To prevent the channel gain from being too large or too small and affecting the training results, the data should be normalized using the function g(·), and the preprocessed channel gain should be denoted as... Right now

[0199]

[0200] Step 21: Reshape the preprocessed data into a one-dimensional length vector, and then feed it into three independent DNN modules, namely the intelligent reflector phase module (e.g., Figure 3 As shown), the module for D2D user transmit power (such as...) Figure 4 As shown), and the channel selection module (as shown) Figure 5 (As shown).

[0201] The specific implementation process for each module is as follows:

[0202] Step 24: As Figure 3 As shown, the output of the intelligent reflector phase module is divided into four groups, each containing eight elements. Then, each group's output is input into a separate softmax layer block, where a softmax operation is performed. Vector z i This represents the probability of each reflection phase level being selected, i.e., z i The m-th element corresponds to the m-th reflective element using phase level. The probability of.

[0203] Step 25: As Figure 4 As shown, the transmit power module calculates the transmit power levels for all D2D operations, with an output count of 5×3. The outputs of the transmit power module are divided into three groups, each containing 5 elements. Then, each group of outputs is input into a separate softmax layer block, where a softmax operation is performed. Vector q i q represents the probability of each transmit power level being selected. i The l-th element corresponds to D2D TP i using transmit power level P l The probability of.

[0204] Step 26: As Figure 5 As shown, the reshaped channel gain is fed into a channel selection module with 2×3 outputs, and then its output is divided into 3 groups, which are respectively sent to the output of... It consists of three independent softmax layers. During training, the output of this module... It is used directly for channel selection, but during inference, it is set... And set all other k. So that constraints can be satisfied in actual implementation

[0205] Step 27: Obtain the maximum sum rate of D2D users by minimizing the loss function of the entire neural network. Setting δ = 0.01, the custom neural network loss function is:

[0206]

[0207] Where g1 is the function to eliminate binarization error:

[0208]

[0209] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the scope of the technology disclosed in the present invention, and such modifications or substitutions should all be covered within the scope of protection of the present invention.

Claims

1. A method for allocating communication resources assisted by an intelligent reflective surface, characterized in that: The processing steps at the drone base station transmitter are as follows: Step 1: The transmitter includes a drone base station, a RIS with M reflective elements, K channels, and D D2D pairs, where each channel contains a cellular user CUE; in addition, using Indicates the first i Whether to use the first method for D2D The first channel, that is, if the first channel... i The transmitter for D2D is transmitted via the first Data is transmitted through each channel. ,otherwise Assume that each D2D pair can only use one channel at a time, i.e. ; Assuming the drone base station, CUE, and D2D are all equipped with a single antenna, and the drone base station is located at... , No. l The CUE positions are The RIS position is D2D users are randomly distributed within a circle of radius R. i The location of each D2D transmitter (DT) is denoted as The distance between DT and the corresponding receiver DR is L , No. i Each DR position is denoted as ; Furthermore, assuming the RIS is equipped with One reflecting element, reflection coefficient matrix Represented as The first of this diagonal matrix Each element is represented as ,in For the first Phase shift of each reflecting element, Indicates the first The amplitude of each reflecting element; in practice, the phase shift of each element can only have a finite number of discrete values, the l-th discrete value is expressed as ,in , Indicates the level used for quantizing phase shift. The number of bits; then, assuming that RIS performs ideal reflection, so that the signal power of each reflecting element is lossless, i.e., the amplitude reflection coefficient. Meanwhile, assume that the transmit power of DTi is expressed as... And can be quantified as A walk of a distance, that is Its maximum value is ,Right now ; Step 2: Generation of drone base station X bit information ; Step 3: Modulate the information generated in Step 2. The modulated signal is... and satisfy ; Step 4: Transfer the signal generated in step 3 s Send into the channel; The steps for establishing a drone-to-user channel are as follows: Step 5: The signal transmitted by the drone base station initially propagates in free space. At this time, there are no tall buildings, trees or other objects in the space to block it, and the drone can directly transmit the signal. At this time, a line-of-sight (LOS) link is formed, and the resulting fading model is mainly a free space loss model. As the signal continues to propagate, there is obstruction between the drone base station and the ground user. At the same time, it is affected by electromagnetic wave reflection and scattering, forming a non-line-of-sight (NLOS) link. Step 6: LOS and NLOS in the channel gain occur with a certain probability. Assume the ground node coordinates are... The coordinates of the UAV base station are The distance between the drone and the node is Then the LOS component probability of the node is: Where A and B are constants related to the environment. The NLOS component probabilities of a node are: ; Step 7: Drone to D2D User i The channel gain is: in It is the path loss coefficient between the user and the drone link. It is the correlation coefficient with the NLOS link; Step 8: Drones and the first l The distance between CUE users is ,Will Substitute into step 6 and In the process, the Rayleigh distribution coefficient is randomly generated. If the channel gain is: ; Step 9: The distance between the drone and the RIS is Ignoring the positional differences of each component in the RIS, Substitute into step 6 and In the process, the Rayleigh distribution coefficient is randomly generated. Then the channel gain is: ; Step 10: Drones and the first i The distance between D2D users is ,Will Substitute into step 6 and In the process, the Rayleigh distribution coefficient is randomly generated. Then the channel gain is: ; The other link channel construction and processing steps are as follows: Step 11: Considering the random Rayleigh distribution and path loss for other link channel gains, the path loss is: in d It is the distance between nodes. These represent the path loss and path loss coefficient at 1m, respectively. Step 12: RIS and the first l The distance between CUE users is ,Will Substitute the values ​​into the loss function from step 11 and randomly generate the Rayleigh distribution coefficients. Then the channel gain is: ; Step 13: RIS and the first i The distance between each DR is ,Will Substitute the values ​​into the loss function from step 11 and randomly generate the Rayleigh distribution coefficients. Then the channel gain is: ;RIS and the i The distance between each DT is ,Will Substitute the values ​​into the loss function from step 11 and randomly generate the Rayleigh distribution coefficients. Then the channel gain is: ; Step 14: l The CUE user and the first i The distance between D2D users is ,Will Substitute the values ​​into the loss function from step 11 and randomly generate the Rayleigh distribution coefficients. Then the channel gain is: ; Step 15: i The first DT and the first The distance between DR users is ,Will Substituting into the loss function of step 11, the loss matrix is ​​denoted as... , where the i-th row is the first Column elements are denoted as Randomly generate Rayleigh distribution coefficients Then the channel gain is: , of which i The channel gain between DT and DR is denoted as , No. i The first DT and the first The channel gain between the DRs is denoted as ; The processing steps at the receiving end are as follows: Step 16: The SINR received by the l-th CUE is: in Indicates the UAV's transmission power. For the first i The transmit power of the DT, the first The additive white Gaussian noise at each CUE user follows the following rules: ; Step 17: i The SINR received by each DR is: in and They represent from the first k Interference from other D2D users and CUE users on the first channel; i The additive white Gaussian noise at each DR follows: ; Step 18: Therefore, the first k The first channel i The achievable rate for each DR is: Step 19: Calculate the first... k The sum rate of all DRs in each channel is: Step 20: To ensure the QoS of the CUE, the SINR of the CUE should be greater than the minimum SINR threshold. Therefore, resource allocation is formulated as a non-convex optimization problem as follows: in , Indicates the first i Use index vectors for D2D channels. This represents the power of D D2D units. Indicates the maximum interference limit caused to cellular transmission; Step 21: Optimize the above non-convexity problem using a deep learning algorithm based on unsupervised learning. The specific steps are as follows: Step 22: Preprocess the input data; the preprocessing execution function is used. for: Step 23: Reshape the preprocessed data into a one-dimensional length vector and then feed it into three independent DNN modules: the smart reflector phase module, the D2D user transmit power module, and the channel selection module. The specific implementation process for each module is as follows: Step 24: The output of the intelligent reflective phase module is divided into M groups, each group containing... Each set of outputs is then fed into a separate softmax layer block, where the softmax operation is performed on each element. This represents the probability of each reflection phase level being selected, i.e. The m-th element corresponds to the first element. m Each reflective element uses phase level The probability of; Step 25: The transmit power module calculates the transmit power levels for all D2D operations, and its output is: The output of the transmit power module is divided into D groups, each group containing... Elements; using vectors This represents the probability of each transmit power level being selected, i.e. The l Each element corresponds to a D2D TP. i Use transmit power level The probability of; Step 26: Feed the reshaped channel gain into the system with... The output channel selection module then divides its output into D groups, which are then sent to the output... D independent softmax blocks; the first i The first softmax layer block k The output can be interpreted as D2D TP i Select Channel k The probability; the output of this module during training. It is used directly for channel selection, but during inference, it is set... , and set all others k of So as to satisfy constraints in actual implementation ; Step 27: Obtain the maximum sum rate of D2D users by minimizing the loss function of the entire neural network; the custom neural network loss function is: in, The function to eliminate binarization error: in, These are network hyperparameters.

2. A computer system, characterized in that... include: One or more processors, a computer-readable storage medium for storing one or more programs, wherein, when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the method of claim 1.

3. A computer-readable storage medium, characterized in that... The device stores computer-executable instructions, which, when executed, are used to implement the method of claim 1.

Images