Intelligent reflecting surface control optimization method based on discrete phase shift

Abstract

The application discloses an intelligent reflecting surface regulation optimization method based on discrete phase shift, an intelligent reflecting surface RIS receives a base station target signal, an improved LSTM network is used to extract a first pure signal based on signal timing rules; the target signal is converted to a frequency domain through short-time Fourier transform, a second pure signal is obtained according to channel matrix channel fading rules, a weighted fusion is performed to obtain a denoising signal to offset the influence of interference and channel fading, a complex channel matrix is decomposed to obtain a real characteristic matrix based on Euler formula, a RIS phase shift matrix is generated through multi-head self-attention and cross-attention in the Transform, a quantized phase shift matrix is obtained through a quantized activation function to match hardware, the regulation accuracy of the RIS reflecting unit is improved, and the denoising signal is sent to a target user to realize high-performance wireless communication.

Description

Technical Field

[0001] This invention relates to the field of millimeter-wave wireless communication, and more specifically to a method for intelligent reflector modulation and optimization based on discrete phase shift. Background Technology

[0002] Existing RIS phase shift control schemes all output continuous phase shifts. However, in actual hardware, the phase shift adjustment of the reflector unit is mostly discrete. Therefore, directly applying the continuous phase shift of existing schemes will introduce quantization errors and reduce the accuracy of phase shift control. In addition, during the transmission of signals from the base station to the RIS, signals will interfere with each other, thus affecting the quality of the signal transmitted from the RIS to the target user. Therefore, it is necessary to design an intelligent reflector control optimization method based on discrete phase shifts to improve the signal quality transmitted to the target user via the reflector. Summary of the Invention

[0003] This invention addresses the shortcomings of existing technologies by proposing an intelligent reflective surface control and optimization method based on discrete phase shift.

[0004] The technical solution to achieve the purpose of this invention is as follows:

[0005] The intelligent reflector control optimization method based on discrete phase shift includes the following steps:

[0006] The intelligent reflector RIS receives the target signal sent by the base station. , No. The signal collected at each time-domain sampling point is Then, the channel matrix from the base station to the RIS is obtained. RIS to target user channel matrix and the channel matrix from the base station to the target user Minimum adjustment of the reflection unit on the RIS ;

[0007] Target signal received from RIS using an improved LSTM network The first pure signal is extracted based on the signal timing pattern. The target signal is transformed using short-time Fourier transform. Transform to the frequency domain to obtain the time-frequency matrix And calculate the energy at each frequency point, based on the channel matrix from the base station to the RIS. Extracting channel fading patterns to generate a frequency domain attention weight matrix For the frequency domain attention weight matrix Physical constraints are applied to obtain the optimized frequency domain attention weight matrix. By optimizing the frequency domain attention weight matrix For time-frequency matrix The weighted time-frequency matrix is obtained after weighting The weighted time-frequency matrix is converted to the time domain by inverse short-time Fourier transform to obtain the second pure signal The confidence scores of the first pure signal and the second pure signal are calculated respectively, and the confidence scores are converted into fusion weights, and the first pure signal and the second pure signal are fused by weighting based on the fusion weights to obtain the denoised signal ;

[0008] Based on the Euler formula, the elements in the complex base station-to-RIS channel matrix , the RIS-to-target user channel matrix , and the base station-to-target user channel matrix are decomposed into three real component matrices to obtain real channel matrices , and ; the three real component matrices are spliced in order according to the order of the base station-to-RIS real channel matrix , the RIS-to-target user real channel matrix , and the base station-to-target user real channel matrix to obtain a real feature matrix ; the real feature matrix is input into a Transformer model to generate an RIS phase shift matrix through a multi-head self-attention learning mechanism and a cross-attention mechanism; based on the minimum adjustment amount of the RIS reflection unit , the RIS phase shift matrix is processed by a quantization activation function to obtain a quantized phase shift matrix ;

[0009] After adjusting the reflection units in the RIS according to the quantized phase shift matrix generated by the Transformer model, the denoised signal is sent to the target user.

[0010] Further, the intelligent reflecting surface RIS collects target signals sent by the base station at time domain sampling points , the th element in the target signal represents the time sequence signal data collected by the base station at the th time domain sampling point , which is a vector, specifically denoted as Among them, in the first The signal data collected from each antenna is denoted as . ,in , ;

[0011] Furthermore, there are a total of [number] base stations. Each antenna transmits the target signal to the RIS. The RIS has a total of [number] antennas. There are one reflecting element, and the target user is a single-antenna user who receives the target signal reflected by the RIS; then the target signal is acquired. Channel matrix transmitted from base station to RIS The dimension is The target signal is transmitted to the target user's channel matrix via RIS reflection. The dimension is The channel matrix of the target signal from the base station to the target user The dimension is Simultaneously, the minimum adjustment amount during phase shift adjustment of the reflector unit on the RIS is collected. .

[0012] Furthermore, regarding the target signal Denoising to obtain a denoised signal Includes the following steps:

[0013] An improved LSTM network is used to retain the first, undisturbed, clean signal by comparing the signal similarity between adjacent time-series signals, based on the signal timing patterns. ;

[0014] Extracting the second pure signal based on the channel fading pattern The target signal is transformed by short-time Fourier transform. Transform to the frequency domain to obtain the time-frequency matrix And calculate the energy at each frequency point, based on the channel matrix from the base station to the RIS. Extracting channel fading patterns to generate a frequency domain attention weight matrix For the frequency domain attention weight matrix Physical constraints are applied to obtain the optimized frequency domain attention weight matrix. By optimizing the frequency domain attention weight matrix For time-frequency matrix After weighting, the weighted time-frequency matrix is ​​obtained. Then, the weighted time-frequency matrix is ​​transformed by inverse short-time Fourier transform. The second pure signal is obtained by converting to the time domain. ;

[0015] First pure signal Second pure signal The weighted fusion obtains a de-noised signal , respectively calculate the confidence score of the first pure signal and the second pure signal and weighted fusion obtains a de-noised signal ;

[0016] Further, in the target signal , the time sequence signal collected at the first time domain sampling point is taken as an example, the input of the improved LSTM network is and , wherein represents the signal hidden state of the first time domain sampling point, and the process of the improved LSTM network obtaining the first pure signal is as follows:

[0017] Firstly, concatenate and and input them to calculate the input proportion of the current signal through the input gate , calculate the forgetting proportion of the historical state through the forgetting gate ; then calculate the time sequence rule weight through the regularity reinforcement gate by comparing the signal similarity between adjacent time points to retain the information of small signal time sequence fluctuation and weaken the information of large signal fluctuation; when the similarity between the current time and the previous time collected signal is high, the regular signal is enhanced; when the similarity between the current time and the previous time collected signal is low, the influence of the interference signal is weakened; calculate the memory cell value according to the input proportion of the current signal , the forgetting proportion of the historical state and the time sequence rule weight ; by introducing the time sequence rule weight when calculating the candidate memory cell value , the memory cell value will preferentially retain the historical information of the regular signal when updating; finally, calculate the activation value of the output gate and calculate the hidden state based on the memory cell value ; the activation value of the output gate is used to control the output proportion of the memory cell, and the signal hidden state of the current first time domain sampling point is calculated;

[0018] According to the calculation method of the improved LSTM network, the signal hidden state of the first As an improvement to LSTM networks from target signals The first pure signal extracted that conforms to the time sequence pattern ;

[0019] Furthermore, the energy distribution of the signal in the frequency domain conforms to the channel fading law; the energy of the clean signal is concentrated in the higher frequency band, while the energy of the interfered signal is concentrated in the lower frequency band. Based on this, the first... Signal at each time-domain sampling point The time-frequency matrix is ​​obtained by transforming the matrix into the frequency domain using a short-time Fourier transform. The energy at each frequency point is calculated to obtain the frequency domain energy matrix. Subsequently, based on the channel matrix from the base station to the RIS... Extracting Channel Fading Patterns Channel fading characteristics This represents the channel matrix from the base station to the RIS. Gain at each frequency; after calculating the channel matrix. After calculating the gain for each frequency, it is also necessary to calculate the sparsity, which reflects the degree of energy concentration. High sparsity indicates that energy is concentrated in a few frequency bands, and the channel matrix sparsity is high. According to the channel matrix The number of non-zero elements in the channel matrix The ratio of the total number of elements in the matrix is ​​calculated; combined with the frequency domain energy matrix. Channel fading characteristics and channel matrix sparsity The frequency domain attention weight matrix is ​​generated through a fully connected layer and a softmax function. Used from time-frequency matrix The system retains signal energy conforming to channel laws located in the channel gain peak frequency band and filters out interference signal energy located in the channel attenuation frequency band. To further increase the weight of signal energy conforming to channel laws located in the channel gain peak frequency band in the frequency domain attention, the frequency domain attention weight matrix is ​​adjusted. Physical constraints are applied to obtain the optimized frequency domain attention weight matrix. ; after obtaining the optimized frequency domain attention weight matrix Then, the time-frequency matrix The weighted time-frequency matrix is ​​obtained by weighting. This filters out energy that does not conform to channel patterns, and then uses inverse short-time Fourier transform to weight the time-frequency matrix. The second pure signal is obtained by converting back to the time domain. ;

[0020] Furthermore, the first pure signal It is obtained by improving the LSTM network based on the signal timing pattern, while the second pure signal is obtained by extracting the signal. This is obtained by extracting the fading pattern characteristics of the channel matrix in the frequency domain. Therefore, each method has its own emphasis and requires adaptive fusion through dynamic weights to obtain a more accurate denoised signal. First, calculate the first pure signal. With the second pure signal The confidence scores, where the first pure signal confidence score The average weight of the temporal regularity of the gate is obtained by calculating the regularity.

[0021] Second pure signal confidence score By calculating the frequency domain attention weight matrix The average value of the attention entropy is obtained; the confidence score of the first pure signal is obtained. Confidence score of the second pure signal Then, the first pure signal was calculated. With the second pure signal The adaptive weights are then fused to obtain the denoised signal. .

[0022] Furthermore, the RIS phase shift matrix is ​​generated using the Transformer model. Includes the following steps:

[0023] Based on Euler's formula, the complex form of the channel matrix from the base station to the RIS is used. RIS to target user channel matrix Channel matrix from base station to target user After the elements in the equation are decomposed into three real components, including the amplitude component, the equation becomes clearer. Cosine phase component Sine phase components The real channel matrix is ​​obtained. , and ;

[0024] According to the real channel matrix from the base station to the RIS Real-number channel matrix from RIS to target user Real-number channel matrix from base station to target user The real component matrices are concatenated in sequence to obtain the real characteristic matrix. ;

[0025] real characteristic matrix The RIS phase shift matrix is generated by a multi-head attention learning mechanism and a cross-attention mechanism in a Transformer model ;

[0026] The minimum adjustment amount of the RIS reflection unit The RIS phase shift matrix is processed by a quantized activation function The quantized phase shift matrix is obtained .

[0027] Further, for the channel matrix of the base station to the RIS , since the dimension of the channel matrix is , let the element in the channel matrix be , which represents the channel coefficient of the th antenna of the base station transmitting to the th reflection unit of the RIS, which is decomposed into three real component matrices by Euler's formula, including the amplitude matrix , the cosine phase matrix , and the sine phase matrix ; each element in the channel matrix is decomposed into three real components to obtain the real channel matrix , since each element in the original channel matrix is decomposed into three real components, the dimension of the real channel matrix becomes ; the elements in the channel matrix of the RIS to the target user and the channel matrix of the base station to the user are decomposed according to the above method to obtain the real channel matrices and , the dimension of the real channel matrix is , and the dimension of the real channel matrix is ;

[0028] Further, the real channel matrix of the base station to the RIS , the real channel matrix of the RIS to the target user , and the real channel matrix of the base station to the target user are spliced according to the signal transmission process to obtain a real feature matrix with a dimension of ; ;

[0029] Further, the real feature matrix is input into a Transformer model to generate the RIS phase shift matrix by an encoder and a decoder.

[0030] Specifically, the encoder includes a 6-layer self-attention network for learning channel global correlation rules in the real feature matrix , extracting cross-channel and cross-dimension correlation features from the features to generate an RIS phase shift matrix for the decoder to provide global context; each layer of the self-attention network includes a multi-head attention module, a feedforward network, and a residual connection.

[0031] The multi-head attention module learns different correlation patterns through multiple attention heads, including the attenuation correlation of channel amplitude between the base station-to-RIS channel matrix , the RIS-to-target user channel matrix , the cumulative offset of phase between the base station-to-RIS channel matrix , the RIS-to-target user channel matrix , and the interference relationship between the base station-to-target user channel matrix and the RIS-to-target user channel matrix ; the feedforward network maps the output of the multi-head attention module, strengthens the feature weights of key correlations, and uses the residual connection to accelerate the convergence trend and ensure stable distribution of features in each layer; after feature encoding through the 6-layer self-attention network, a global feature matrix is obtained, which includes the correlation between amplitude and phase within a single channel, different correlation patterns between multiple channels, and spatial coordination correlation rules between adjacent RIS reflection elements or antennas.

[0032] Specifically, the decoder includes a 6-layer cross-attention network that increases the perception of key channels in the global feature matrix through the cross-attention mechanism, improving the accuracy of the generated RIS phase shift matrix ; an initial phase shift matrix is generated according to the number of reflection elements on the RIS ; the cross-attention mechanism is used to directly associate the phase shift of the reflection elements on the RIS with the global feature matrix ; for the phase shift of the th reflection element, all antennas in the base station-to-RIS channel matrix to the reflection element and the channel from the reflection element to the target user in the RIS-to-target user channel matrix are considered, and after processing through the cross-attention mechanism, the feedforward network and the residual connection are used to strengthen the non-linear mapping ability of the decoder and ensure that the phase shift conforms to the channel phase compensation; after optimizing the values of the initial phase shift matrix through multiple layers of cross-attention network, the RIS phase shift matrix is obtained, taking into account the minimum adjustment amount of the phase shift of the reflection elements on the RIS due to hardware configuration problems To make the RIS phase shift matrix output by the decoder comply with the actual situation, the values in the RIS phase shift matrix are constrained by using a quantized activation function to obtain a quantized phase shift matrix .

[0033] Further, after generating the quantized phase shift matrix by the Transformer model, the corresponding reflection units on the RIS are adjusted according to the element values in the quantized phase shift matrix , so that the target signal transmitted by the base station can be accurately transmitted to the target user after denoising processing to complete communication.

[0034] Compared with the prior art, after the intelligent reflecting surface RIS receives the target signal from the base station, the improved LSTM network is used to extract the first pure signal from the target signal based on the signal timing law; after the target signal is converted to the frequency domain by short-time Fourier transform, the second pure signal is obtained according to the channel fading law in the channel matrix; the first pure signal and the second pure signal are weighted and fused to obtain a denoised signal, which eliminates the influence of external interference and channel fading on the signal; based on the Euler formula, the complex channel matrix is decomposed and converted into a real channel matrix, and the RIS phase shift matrix is generated in the Transformer model through the multi-head self-attention learning mechanism and the cross-attention mechanism, and the quantized phase shift matrix is obtained through the quantized activation function, which improves the phase shift control precision by matching the actual hardware conditions; after adjusting the reflection units in the RIS according to the quantized phase shift matrix generated by the Transformer model, the denoised signal is transmitted to the target user. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 is a flowchart of the intelligent reflecting surface regulation and optimization method based on discrete phase shift;

[0036] Figure 2 is a flowchart of the first pure signal acquisition;

[0037] Figure 3 is a flowchart of the second pure signal acquisition;

[0038] Figure 4 is a flowchart of generating a quantized phase shift matrix by a Transformer model. DETAILED DESCRIPTION

[0039] The application will be further described in detail below in combination with the drawings and examples. ​​

[0040] As Figure 1 shown, one specific embodiment of the present application discloses a smart reflecting surface regulation optimization method based on discrete phase shift, including the following steps:

[0041] The smart reflecting surface RIS receives the target signal sent by the base station , and the signal collected by the first time domain sampling point is ; Then the channel matrix from the base station to the RIS , the channel matrix from the RIS to the target user , and the channel matrix from the base station to the target user are obtained, and the minimum adjustment amount of the reflecting unit on the RIS , , is the total number of time domain sampling points;

[0042] An improved LSTM network is used to extract the first pure signal from the target signal received by the RIS based on the signal timing law; The target signal is converted to the frequency domain by short-time Fourier transform to obtain the time-frequency matrix and calculate the energy of each frequency point, extract the channel fading law based on the channel matrix from the base station to the RIS to generate the frequency domain attention weight matrix , and the frequency domain attention weight matrix is physically constrained to obtain the optimized frequency domain attention weight matrix , and the time-frequency matrix is weighted by the optimized frequency domain attention weight matrix to obtain the weighted time-frequency matrix , and the weighted time-frequency matrix is converted to the time domain by inverse short-time Fourier transform to obtain the second pure signal ; The confidence scores of the first pure signal and the second pure signal are calculated respectively, and the confidence scores are converted into fusion weights, and the first pure signal and the second pure signal are weighted and fused based on the fusion weights to obtain the denoising signal ;

[0043] Based on Euler's formula, the elements in the channel matrix from the base station to the RIS , the channel matrix from the RIS to the target user , and the channel matrix from the base station to the target user are decomposed into three real components, including amplitude component , cosine phase component sinusoidal phase component , to obtain a real number channel matrix , and ; according to the order of the real number channel matrix from the base station to the RIS , the real number channel matrix from the RIS to the target user , the real number channel matrix from the base station to the target user , the three real number component matrices are spliced in order to obtain a real number feature matrix ; the real number feature matrix is input into the Transformer model to generate the RIS phase shift matrix through the multi-head self-attention learning mechanism and the cross-attention mechanism; the RIS phase shift matrix is processed through the quantization activation function to obtain a quantized phase shift matrix ; ;

[0044] According to the quantized phase shift matrix generated by the Transformer model , the reflection units in the RIS are adjusted, and the denoised signal is sent to the target user.

[0045] Further, the intelligent reflecting surface RIS collects the target signal sent by the base station with antennas at time domain sampling points , the th element in the target signal represents the time sequence signal data collected by the base station at the th time domain sampling point, is a vector, and is specifically denoted as , wherein the signal data collected at the th antenna is denoted as , wherein , ;

[0046] Further, there are antennas on the base station sending target signals to the RIS, and there are reflection units on the RIS, and the target user is a single-antenna user, receiving the target signal reflected by the RIS; then the target signal is obtained The dimension of the channel matrix from the base station to the RIS , and the dimension of the channel matrix of the target signal reflected by the RIS and transmitted to the target user is The channel matrix of the target signal from the base station to the target user The dimension is Simultaneously, the minimum adjustment amount during phase shift adjustment of the reflector unit on the RIS is collected. .

[0047] Furthermore, traditional LSTM networks control information retention and forgetting through input gates, forget gates, and output gates to capture long-term dependencies in sequence data. Building upon this, considering the strict temporal regularity of communication signals, an improved LSTM network adds a regularity enhancement gate to strengthen the LSTM network's memory of signal temporal regularities and weaken interference in the signal. For example, taking the target signal... In the first Time-series signals acquired at time-domain sampling points For example, the input to the improved LSTM network is and ,in Indicates the first The hidden state of the signal at each time-domain sampling point;

[0048] like Figure 2 As shown, the improved LSTM network obtains the first clean signal. The process is as follows:

[0049] First, the input ratio of the current signal is calculated through the input gate. The forgetting rate of historical states is calculated using the forgetting gate. The calculation method is as follows:

[0050] ,

[0051] ,

[0052] in, , It is a weight matrix. , It is a bias vector. It is an activation function; subsequently, a pattern reinforcement gate is used to compare the signal similarity between adjacent time points to calculate the temporal pattern weights. To retain information with small timing fluctuations and weaken information with large fluctuations, the calculation method is as follows:

[0053] ,

[0054] in, It is a weight matrix. It is a bias vector. is the error between the current time and the previous time, which can effectively reflect the similarity between the current time and the previous time, is the similarity factor, when the similarity between the current time and the previous time is high, the value of tends to 1, enhancing the retention of regular signals; when the similarity between the current time and the previous time is low, the value of tends to 0, weakening the influence of interference signals; according to the input proportion of the current signal , the forgetting proportion of the historical state and the timing regularity weight , the memory cell value of the current time is calculated , the calculation method is as follows:

[0055] ,

[0056] ,

[0057] Among them, is the memory cell value of the previous time, is the weight matrix, is the bias vector, is a nonlinear transformation function, represents the Hadamard product of element-by-element multiplication, is the candidate memory cell value; by introducing the timing regularity weight when calculating the candidate memory cell value , the memory cell value will preferentially retain the historical information of regular signals when updating; finally, the activation value of the output gate is calculated and the hidden state is calculated based on the memory cell value , the calculation method is as follows:

[0058] ,

[0059] ,

[0060] Among them, is the weight matrix, is the bias vector; the activation value of the output gate is used to control the output proportion of the memory cell, and the signal hidden state of the current time domain sampling point is calculated ;

[0061] According to the calculation method of the improved LSTM network, the signal hidden state of the first time domain sampling point is calculated , as an improved LSTM network extracts a first pure signal from the target signal that conforms to the timing law .

[0062] Further, as shown in Figure 3 , the energy distribution of the signal in the frequency domain conforms to the channel fading law, and the energy of the pure signal is concentrated in the higher frequency band, while the energy of the interference signal is concentrated in the lower frequency band; on this basis, the signal of the first time domain sampling point is converted to the frequency domain by short-time Fourier transform to obtain a time-frequency matrix and the energy of each frequency point is calculated to obtain a frequency energy matrix , the calculation method is as follows:

[0063] ,

[0064] ,

[0065] where, denotes the short-time Fourier transform; then based on the channel matrix from the base station to the RIS extract the channel fading law feature , the calculation method is as follows:

[0066] ,

[0067] where, denotes the Fourier transform; the channel fading law feature indicates the channel matrix from the base station to the RIS gain for each frequency; after calculating the gain of the channel matrix for each frequency, the sparsity reflecting the energy concentration degree also needs to be calculated, when the sparsity is high, it indicates that the energy is concentrated in a few frequency bands, the calculation method of the channel matrix sparsity is as follows:

[0068] ,

[0069] where, denotes the number of non-zero elements in the channel matrix , denotes the element in the channel matrix , denotes the total number of elements in the channel matrix ; combined with the frequency energy matrix , the channel fading law feature and the channel matrix sparsity generate a frequency domain attention weight matrix , which is used to extract the target signal from the time-frequency matrix​ The signal energy conforming to the channel law in the channel gain peak frequency band is reserved, and the interference signal energy in the channel attenuation frequency band is filtered, a fully connected layer and a softmax function are used to generate a frequency domain attention weight matrix The calculation method is as follows:

[0070] ,

[0071] Among them, is a maximum value normalization operation of the frequency energy matrix , which is used to eliminate the scale effect, The fully connected layer is used to further improve the weight of the signal energy conforming to the channel law in the channel gain peak frequency band in the frequency domain attention, and the frequency domain attention weight matrix is subjected to physical constraint to obtain an optimized frequency domain attention weight matrix The calculation method is as follows:

[0072] ;

[0073] After obtaining the optimized frequency domain attention weight matrix , the time-frequency matrix is weighted to obtain a weighted time-frequency matrix , so as to filter the energy not conforming to the channel law, and then the weighted time-frequency matrix is converted back to the time domain through inverse short-time Fourier transform to obtain a second pure signal The calculation method is as follows:

[0074] ,

[0075] ,

[0076] Among them, indicates inverse short-time Fourier transform;

[0077] Further, the first pure signal is extracted according to the signal time sequence law through an improved LSTM network, and the second pure signal is obtained by extracting the channel matrix fading law feature in the frequency domain, so that the two have different focuses, and need to be adaptively fused through dynamic weights to obtain a more accurate denoising signal ; first, the confidence scores of the first pure signal and the second pure signal are calculated, wherein the confidence score of the first pure signal is obtained by calculating the time sequence law weight average of the law strengthening gate, and the calculation method is as follows:

[0078]

[0079] in, Indicates the number of time-domain sampling points. Indicates the number of antennas in the base station. Indicates the first At the nth time domain sampling point The weighting of the timing pattern of signals transmitted by each antenna;

[0080] Second pure signal confidence score By calculating the frequency domain attention weight matrix The average value of the attention entropy is obtained, and the calculation method is as follows:

[0081] ,

[0082] in, Represent a non-zero constant to avoid the occurrence of... The situation renders the logarithm meaningless; the above equation represents the frequency domain attention weight matrix. The average attention entropy of each element; the confidence score of obtaining the first pure signal. Confidence score of the second pure signal Then, the first pure signal was calculated. With the second pure signal The adaptive weights are then fused to obtain the denoised signal. The calculation method is as follows:

[0083] ,

[0084] The denoised signal obtained by the above fusion method It can suppress two different types of interference signals at the same time, making the signals received and reflected by the RIS to the target user more accurate, thus improving communication accuracy.

[0085] Furthermore, such as Figure 4 As shown, the RIS phase shift matrix is ​​generated using the Transformer model. Includes the following steps:

[0086] Based on Euler's formula, the complex form of the channel matrix from the base station to the RIS is used. RIS to target user channel matrix Channel matrix from base station to target user After the elements in the equation are decomposed into three real components, including the amplitude component, the equation becomes clearer. Cosine phase component Sine phase components The real channel matrix is ​​obtained. , and ;

[0087] According to the real channel matrix from the base station to the RIS Real-number channel matrix from RIS to target user Real-number channel matrix from base station to target user The real component matrices are concatenated in sequence to obtain the real characteristic matrix. ;

[0088] real characteristic matrix The input is fed into the Transformer model, where a RIS phase shift matrix is ​​generated through multi-head attention learning and cross-attention mechanisms. ;

[0089] Minimum adjustment based on RIS reflection unit Processing the RIS phase shift matrix using a quantized activation function Obtain the quantization phase shift matrix .

[0090] Furthermore, the channel matrix from the base station to the RIS RIS to target user channel matrix Channel matrix from base station to target user The elements in the matrix are complex numbers, representing the channel coefficients from the transmitter to the receiver, reflecting the signal transmission characteristics along a specific path. Since the Transformer model cannot be directly used to process complex data, the complex elements in the channel matrix must first be decomposed into real component matrices. For example, consider the channel matrix from the base station to the RIS. For example, due to the channel matrix The dimension is Let the channel matrix be denoted. The elements in are Indicates the number on the base station The first antenna was transmitted to the RIS. The channel coefficients of each reflecting unit are decomposed into three real component matrices using Euler's formula, including the amplitude matrix. Cosine phase matrix Sine phase matrix The decomposition method is as follows:

[0091] ,

[0092] ,

[0093] ,

[0094] ,

[0095] wherein, denotes the amplitude component, denotes the cosine phase component, for representing the phase of periodicity, denotes the sine phase component, for representing the phase direction, which is combined with the cosine phase component to determine the unique phase , denotes the real part of the channel coefficient, denotes the imaginary part of the channel coefficient; after decomposing each element in the channel matrix into three real components, the real channel matrix is obtained, since each element in the original channel matrix is decomposed into three real components, the dimension of the real channel matrix becomes ; the elements in the channel matrix from RIS to target user, the channel matrix from base station to target user are decomposed according to the above method to obtain the real channel matrix and respectively, the dimension of the real channel matrix is , and the dimension of the real channel matrix is ;

[0096] Furthermore, the real channel matrix from base station to RIS, the real channel matrix from RIS to target user, and the real channel matrix from base station to target user are spliced according to the signal transmission process to obtain the real feature matrix ; since the target signal is transmitted to the target user after being reflected by the reflecting unit of the RIS after being emitted from the base station antenna, the calculation method of the real feature matrix spliced according to the signal transmission process is as follows:

[0097] ,

[0098] wherein, denotes the splicing of the real channel matrix and in the matrix row to obtain a matrix with the dimension of , denotes the splicing of each row of the real channel matrix and to obtain the real feature matrix with the dimension of ​ ;

[0099] Further, the real feature matrix is input into the Transformer model to generate the RIS phase shift matrix through the encoder and the decoder.

[0100] Specifically, the encoder includes 6 layers of self-attention networks for learning the channel global correlation in the real feature matrix and extracting cross-channel and cross-dimension correlation features from the features to provide global context for the decoder to generate the RIS phase shift matrix ; each layer of the self-attention network includes a multi-head attention module, a feedforward network, and a residual connection.

[0101] The multi-head attention module learns different correlation patterns through multiple attention heads, including the attenuation correlation of channel amplitudes between the base station-to-RIS channel matrix and the RIS-to-target user channel matrix , the cumulative offset of phases between the base station-to-RIS channel matrix and the RIS-to-target user channel matrix , and the interference relationship between the base station-to-target user channel matrix and the RIS-to-target user channel matrix ; the feedforward network maps the output of the multi-head attention module, strengthens the feature weights of key correlations, and then uses the residual connection to accelerate the convergence trend and ensure stable distribution of features in each layer; after feature encoding through the 6 layers of self-attention networks, the global feature matrix is obtained, which includes the correlation between amplitudes and phases within a single channel, different correlation patterns between multiple channels, and spatial coordination correlation rules between adjacent RIS reflection units or antennas.

[0102] Specifically, the decoder includes 6 layers of cross-attention networks that increase the perception of key channels in the global feature matrix through the cross-attention mechanism to improve the accuracy of the generated RIS phase shift matrix ; an initial phase shift matrix is generated according to the number of reflection units on the RIS ; the cross-attention mechanism is used to directly associate the phase shifts of the reflection units on the RIS with the global feature matrix ; for the phase shift of the th reflection unit, all channels from the antennas to the reflection unit in the base station-to-RIS channel matrix and the channel from the RIS to the target user in the RIS-to-target user channel matrix need to be considered.The channel from the reflection unit to the target user is processed through the cross-attention mechanism, and then the feedforward network and the residual connection strengthen the decoding ability of the nonlinear mapping of the decoder and ensure that the phase shift conforms to the channel phase compensation; the value of the initialized phase shift matrix is optimized through the multi-layer cross-attention network to obtain the RIS phase shift matrix Considering the minimum adjustment amount of the reflection unit on the RIS due to the hardware configuration problem when adjusting the phase shift In order to make the value of the RIS phase shift matrix output by the decoder conform to the actual situation, based on the minimum adjustment amount The value in the RIS phase shift matrix is constrained by using a quantization activation function to obtain a quantized phase shift matrix The calculation method is as follows:

[0103] ,

[0104] Among them, represents rounding off, to ensure that the phase shift of each reflection unit in the final output quantized phase shift matrix is an integer multiple of the minimum adjustment amount of the reflection unit under actual hardware conditions , realizing the quantized adjustment of the phase shift of each reflection unit.

[0105] Further, after generating the quantized phase shift matrix through the Transformer model, the corresponding reflection unit on the RIS is adjusted according to the element value in the quantized phase shift matrix , so that the target signal sent by the base station can accurately send the de-noising signal obtained after de-noising processing to the target user after reflection on the RIS to complete the communication.

[0106] The application discloses an intelligent reflecting surface regulation optimization method based on discrete phase shift. After an intelligent reflecting surface (RIS) receives a target signal from a base station, an improved LSTM network is used to extract a first pure signal from the target signal based on signal timing rules. After the target signal is converted to the frequency domain through short-time Fourier transform, a second pure signal is obtained according to the channel fading law in the channel matrix. The first pure signal and the second pure signal are weighted and fused to obtain a denoised signal, which eliminates the influence of external interference and channel fading on the signal. Based on the Euler formula, the complex channel matrix is decomposed and converted into a real channel matrix. In the Transformer model, the RIS phase shift matrix is generated through a multi-head self-attention learning mechanism and a cross-attention mechanism, and the quantized phase shift matrix is obtained through a quantization activation function, which improves the phase shift regulation accuracy by matching the actual hardware conditions. After adjusting the reflecting units in the RIS according to the quantized phase shift matrix generated by the Transformer model, the denoised signal is sent to the target user.

[0107] The above is only the preferred embodiment of the present application, and the protection scope of the present application is not limited to the above-mentioned embodiments. Any technical solution falling within the concept of the present application shall be deemed to fall within the protection scope of the present application. It should be noted that, for ordinary skilled persons in the art, some improvements and refinements without departing from the principles of the present application shall also be deemed to fall within the protection scope of the present application.

Claims

1. A method for smart reflector control optimization based on discrete phase shift, characterized in that, The method comprises the steps of: After receiving the target signal sent by the base station, the intelligent reflecting surface RIS obtains the channel matrix from the base station to the RIS, the channel matrix from the RIS to the target user, and the minimum adjustment amount of the reflecting unit on the RIS; The improved LSTM network is adopted to extract a first pure signal from the target signal based on the signal timing law; the target signal is converted to the frequency domain through short-time Fourier transform to obtain a time-frequency matrix, and a second pure signal is obtained based on the channel fading law; the fusion weight of the first pure signal and the second pure signal is calculated, and the first pure signal and the second pure signal are weighted and fused based on the fusion weight to obtain a denoising signal; Based on the Euler formula, the elements in the collected channel matrix are decomposed into real components to obtain a real channel matrix, which is spliced to obtain a real feature matrix; the real feature matrix is input into a Transformer model to generate a quantized phase shift matrix through a multi-head self-attention learning mechanism, a cross-attention mechanism and a quantized activation function; After adjusting the reflecting unit in the RIS according to the quantized phase shift matrix generated by the Transformer model, the denoising signal is sent to the target user; The method comprises the steps of: The improved LSTM network calculates the input proportion of the current signal through the input gate and calculates the forgetting proportion of the historical state through the forgetting gate; then, the signal similarity between adjacent time points is compared through the law reinforcement gate to calculate the timing law weight; the memory cell value is calculated according to the input proportion of the current signal, the forgetting proportion of the historical state and the timing law weight; the activation value of the output gate is calculated, and the first pure signal is calculated based on the memory cell value; The method comprises the steps of: The target signal is converted to the frequency domain through short-time Fourier transform to obtain a time-frequency matrix, and the energy of each frequency point is calculated; the channel fading law is extracted based on the channel matrix from the base station to the RIS to generate a frequency domain attention weight matrix; the optimized frequency domain attention weight matrix is obtained by performing physical constraint on the frequency domain attention weight matrix; the weighted time-frequency matrix is obtained by weighting the time-frequency matrix through the optimized frequency domain attention weight matrix; and the second pure signal is obtained by converting the weighted time-frequency matrix to the time domain through inverse short-time Fourier transform; The method comprises the steps of: The confidence score of the first pure signal is obtained by calculating the average value of the timing law weight of the law reinforcement gate in the improved LSTM network; the confidence score of the second pure signal is obtained by calculating the average value of the attention entropy of the frequency domain attention weight matrix; and the denoising signal is obtained by weighted fusion based on the confidence scores of the first pure signal and the second pure signal; The method comprises the steps of: ​ Based on the Euler formula, elements in a complex number form of a base station to RIS channel matrix, a RIS to target user channel matrix, and a base station to target user channel matrix are decomposed into real number components to obtain a base station to RIS real number channel matrix, a RIS to target user real number channel matrix, and a base station to target user real number channel matrix; a matrix obtained by splicing each row of the base station to RIS real number channel matrix with a matrix obtained by splicing the RIS to target user real number channel matrix and the base station to target user real number channel matrix is spliced to obtain a real number feature matrix.

2. The discrete phase shift based smart reflective surface steering optimization method of claim 1, wherein, The real number feature matrix is input into the Transformer model to generate a quantized phase shift matrix through a multi-head self-attention learning mechanism, a cross-attention mechanism, and a quantized activation function, including: The real number feature matrix is input into the Transformer model, and the Transformer model includes an encoder and a decoder; The encoder performs feature encoding on the real number feature matrix through a self-attention network to obtain a global feature matrix; The decoder processes the global feature matrix through a cross-attention network to obtain an RIS phase shift matrix and restricts values in the RIS phase shift matrix based on a minimum adjustment amount using a quantized activation function to obtain a quantized phase shift matrix.

3. The method of claim 2, wherein the optimization is based on a discrete phase shift. 3 The global feature matrix is obtained by repeatedly performing feature encoding through a 6-layer self-attention network, wherein each layer of the self-attention network learns the attenuation correlation between channel matrices, the cumulative offset of the phase between channel matrices, and the interference relationship between channel matrices through a multi-head attention module, maps the output of the multi-head attention module through a feedforward network, and strengthens the feature weight of the key association through a residual connection. The RIS phase shift matrix is obtained by processing the global feature matrix through the cross-attention network, and values in the RIS phase shift matrix are restricted based on a minimum adjustment amount using a quantized activation function to obtain a quantized phase shift matrix, including:

4. The discrete phase shift based smart reflective surface steering optimization method of claim 2, wherein, The decoder directly associates the phase shift of the reflection unit on the RIS with the global feature matrix using a cross-attention mechanism; The phase shift is learned through a feedforward network and a residual connection to conform to the channel phase compensation; The value of the initialized phase shift matrix is optimized through a multi-layer cross-attention network to obtain a phase shift matrix; Values in the phase shift matrix are restricted based on a minimum adjustment amount using a quantized activation function to obtain a quantized phase shift matrix. The de-noised signal is sent to the target user, including:

5. The discrete phase shift based smart reflective surface optimization method of claim 1, wherein, The corresponding reflection unit on the RIS is adjusted according to the element value in the quantized phase shift matrix; The de-noised signal obtained after the target signal sent by the base station is subjected to de-noising processing is reflected by the adjusted reflection unit and then sent to the target user to complete communication. ​

Images