A deep learning based intelligent reflecting surface assisted wireless computational imaging method

Abstract

The application discloses a kind of wireless computing imaging methods of intelligent reflecting surface assisted based on deep learning, specific steps are as follows: the received signal of base station is extracted using intelligent reflecting surface phase prediction deep neural network, and intelligent reflecting surface phase prediction result is obtained;Target perception area is reconstructed using target perception area detection convolutional neural network, and one-dimensional target scattering coefficient estimation sequence of target perception area is obtained;One-dimensional target scattering coefficient estimation sequence is post-processed, and three-dimensional pixel estimation sequence is obtained, that is, the computing imaging of target perception area is realized.The wireless computing imaging method of intelligent reflecting surface assisted based on deep learning proposed in the application can not only effectively process high-dimensional data under the assistance of intelligent reflecting surface, ensure high-accuracy imaging under the background of big data, but also can automatically extract key features in image or signal, improve the accuracy of imaging.

Description

Technical Field

[0001] The present invention belongs to the field of wireless communications, and in particular relates to a wireless computational imaging method assisted by an intelligent reflective surface based on deep learning. Background Art

[0002] With the continuous advancement of science and technology, wireless computational imaging has been widely used in many fields, such as unmanned driving, drone surveillance, and industrial automation. However, current computational imaging technology still faces several challenges: First, the presence of multiple objects, occlusions, and varying lighting conditions in complex environments makes the accuracy of computational imaging a major problem; second, the need for real-time performance, especially in critical applications such as autonomous driving, where high latency can have serious consequences; third, hardware limitations, as high-precision computational imaging requires a large amount of computing resources and power, poses challenges for deployment on mobile devices and in remote areas; and finally, signal coverage issues. In some complex environments, such as urban core areas or indoors, traditional wireless signals may experience severe attenuation or interference.

[0003] With the rapid development of wireless communication technology, the 6G era is gradually approaching. 6G is not just a communications technology, but also a comprehensive intelligent information system encompassing numerous fields, including cloud computing, big data, and artificial intelligence. Against this backdrop, smart reflector technology has attracted widespread attention and research in the 6G field. As a passive device, smart reflectors can effectively adjust the phase and amplitude of incident waves, creating a superior propagation environment for wireless signals. In 6G communications, faced with limited spectrum resources and the continued growth of connected devices, effectively improving signal quality and expanding coverage have become core technical challenges. Summary of the Invention

[0004] The purpose of the present invention is to solve the problems existing in the prior art and provide a wireless computational imaging method assisted by an intelligent reflective surface based on deep learning.

[0005] The specific technical solutions adopted in the present invention are as follows:

[0006] A wireless computational imaging method assisted by an intelligent reflective surface based on deep learning comprises the following steps:

[0007] S1. The base station transmits a sensing signal into the environment. The sensing signal is reflected by the smart reflecting surface, reaches the target sensing area, and is reflected by the target sensing area to obtain a reflected sensing signal. The reflected sensing signal is returned to the base station along the original path. Channel state information from the base station to the smart reflecting surface, channel state information from the smart reflecting surface to the target sensing area, and a received signal from the base station are obtained.

[0008] S2. Obtain a trained smart reflector phase prediction deep neural network, use the smart reflector phase prediction deep neural network to extract features of the received signal, and obtain a smart reflector phase prediction result;

[0009] S3. Obtain a trained target perception area detection convolutional neural network, reconstruct the target perception area using the target perception area detection convolutional neural network, input the intelligent reflector phase prediction result into the target perception area detection convolutional neural network, and obtain a one-dimensional target scattering coefficient estimation sequence for the target perception area;

[0010] S4. Post-processing the one-dimensional target scattering coefficient estimation sequence, inversely mapping the one-dimensional target scattering coefficient estimation sequence into a three-dimensional pixel estimation sequence, that is, achieving computational imaging of the target perception area.

[0011] Preferably, the method for acquiring the received signal is:

[0012] S11. Obtain the channel matrix from the base station to the smart reflection surface as the first Rice channel Obtain the channel matrix of the smart reflective surface to the target sensing area as the second Rice channel The functional form of the first Ricean channel H is:

[0013]

[0014] The functional form of the second Ricean channel G is:

[0015]

[0016] Where, ∈ is the Rice factor; lp1 is the first channel path loss, lp2 is the second channel path loss; H LoS is the direct part of the first channel, G LoS is the direct part of the second channel; H NLoS is the first channel scattering part that obeys the standard normal distribution, G NLoS is the second scattered part that obeys the standard normal distribution; represents a complex set; M is the number of transmitting and receiving antennas of the base station; N is the number of reflecting units of the smart reflecting surface;

[0017] S12. Reflection coefficient of the smart reflective surface Among them, e represents a natural constant; is the phase of the reflection coefficient; [ρ1,…,ρ N ] is the amplitude of the reflection coefficient, assuming that the amplitude is 1; where (·) T Represents matrix transpose;

[0018] S13. Preprocessing the target sensing area to obtain the one-dimensional target scattering coefficient sequence;

[0019] The preprocessing process is: first, the target perception area is discretized to obtain a three-dimensional pixel sequence s=[s1,…,s K ] T , and then map the three-dimensional pixel sequence to obtain a one-dimensional target scattering coefficient sequence x=[x1,…,x K ] T , where the i-th three-dimensional pixel s i The scattering coefficient is x i ; If the i-th three-dimensional pixel is a real target, the i-th three-dimensional pixel s i The scattering coefficient x i The value is 0 <x i ≤1; otherwise, the i-th three-dimensional pixel s i The scattering coefficient x i The value is 0, i∈[1,2,…,K]; K is the length of the three-dimensional pixel sequence; x i is the i-th sample of the one-dimensional target scattering coefficient sequence;

[0020] S14. Assuming that the number of real target pixels in the target sensing area is much smaller than the number of ambient target pixels, the one-dimensional target scattering coefficient sequence x=[x1,…,x K ] T The sparse prior probability density has a mean of 0 and a covariance matrix of R x = diag(γ) Gaussian distribution, where diag(·) represents diagonalization, γ represents variance, and α is the inverse of the variance. α obeys the Gamma distribution, and the functional form of α is as follows:

[0021]

[0022] Where Γ(·) is the Gamma function, a and b are parameters that control sparsity; (·) -1 Indicates the inverse operation;

[0023] S15. Obtain the total transmission power P of the base station t , the transmission signal of the base station of length L The one-dimensional target scattering coefficient sequence x is diagonalized to obtain the target sensing area scattering coefficient diagonal matrix X=diag(x); the phase of the reflection coefficient is diagonalized to obtain the smart reflection surface phase diagonal matrix The received signal is constructed by using the total transmit power, the transmit signal, the first Ricean channel, the second Ricean channel, the smart reflector phase diagonal matrix, and the target sensing area scattering coefficient diagonal matrix. The functional form is:

[0024]

[0025] Among them, N is subject to the mean of 0 and variance σ 2 Additive Gaussian white noise; represents the conjugate transpose of the matrix;

[0026] S16. According to the matrix vectorization property, the received signal is vectorized to obtain a vectorized received signal, wherein the function form of the vectorized received signal y is:

[0027]

[0028] Here, vec(·) represents vectorization.

[0029] Preferably, the intelligent reflecting surface phase prediction deep neural network is composed of a first fully connected layer with an activation function, a first Dropout layer, a second fully connected layer with an activation function, a second Dropout layer, and a third fully connected layer without an activation function, which are cascaded in sequence; the number of neurons in the first fully connected layer and the second fully connected layer is 256, and the activation function is Relu; the dropout rate of the first Dropout layer and the second Dropout layer is set to 0.5; the number of neurons in the third fully connected layer is N, and the activation function is Sigmoid.

[0030] Preferably, the target perception area detection convolutional neural network is composed of D sparse Bayesian layers cascaded in sequence, and each of the sparse Bayesian layers is composed of a custom Lambda layer, a first Reshape layer, a first convolutional layer with an activation function, a second convolutional layer with an activation function, and a second Reshape layer cascaded in sequence. The convolution kernel size of the first convolution layer is 5×5×5, and the number of convolution kernels is 8; the convolution kernel size of the second convolution layer is 5×5×5, and the number of convolution kernels is 1; and the activation functions are all Relu.

[0031] As an advantage, the specific process of step S2 is: the total transmission power P of the base station is t, the transmitted signal S, the first Ricean channel H, the second Ricean channel G, and the standard deviation σ of the additive white Gaussian noise are used as hyperparameters of the intelligent reflector phase prediction deep neural network, and the value of each hyperparameter of the intelligent reflector phase prediction deep neural network is fixed; randomly initialize the phase The one-dimensional target scattering coefficient sequence x=[x1,…,x K ] T Input into the intelligent reflector phase prediction deep neural network to obtain the intelligent reflector phase prediction result in represents the intelligent reflector phase prediction deep neural network, represents the set of real numbers.

[0032] As an advantage, the specific process of step S3 is as follows: the phase prediction result of the smart reflector is The smart reflector phase diagonal matrix is ​​updated to obtain the smart reflector phase diagonal update matrix The received signal is updated by the phase diagonal update matrix of the smart reflector to obtain an updated received signal, wherein the updated received signal y opt The functional form is:

[0033]

[0034] Introducing intermediate update variables y opt and A opt As the target perception area detection convolutional neural network hyperparameter, each target perception area detection convolutional neural network hyperparameter has a fixed value; randomly initialize the covariance matrix The one-dimensional target scattering coefficient sequence x=[x1,…,x K ] T Input into the target perception area detection convolutional neural network to obtain the one-dimensional target scattering coefficient estimation sequence in represents the target perception area detection convolutional neural network; ⊙ is the Khatri-Rao product, (·) * represents matrix conjugation.

[0035] Preferably, in the dth sparse Bayesian layer, the updated value of the covariance matrix is ​​output from the (d-1)th sparse Bayesian layer. The update reception signal y opt , the intermediate update variable A opt , the updated value of the covariance matrix Input into the dth sparse Bayesian layer and update the variance γd , update the covariance matrix where d∈[1,…,D].

[0036] Preferably, the custom Lambda layer uses a minimum mean square error filter to perform a preliminary estimate of the one-dimensional target scattering coefficient sequence x according to the sparse prior of the one-dimensional target scattering coefficient sequence x, and obtains an updated posterior mean. and update the first variance

[0037] The updated posterior mean The functional form is:

[0038]

[0039] The updated first variance The functional form is:

[0040]

[0041] Where I is the identity matrix.

[0042] Preferably, the loss function used in training the intelligent reflecting surface phase prediction deep neural network is the noise ratio of the received signal, and the network parameters of the intelligent reflecting surface phase prediction deep neural network are updated. The function form of the noise ratio of the received signal is:

[0043]

[0044] Among them, N S is the total number of training samples; represents the square of the Frobenius-norm of the matrix; Φ j is the jth sample of Φ.

[0045] Preferably, the target perception area detection convolutional neural network adopts the one-dimensional target scattering coefficient sequence x=[x1,…,x K ] T The one-dimensional target scattering coefficient estimation sequence of the target sensing area The mean square error is used as the loss function to update the network parameters of the target perception area detection convolutional neural network. The function form of the mean square error is:

[0046]

[0047] Among them, x j is the jth sample of x; for The jth sample of .

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] The present invention constructs a neural network consisting of two parts: a deep neural network for intelligent reflective surface phase prediction and a convolutional neural network for target perception area detection. In the present invention, the received signal is feature extracted using the deep neural network for intelligent reflective surface phase prediction to obtain an intelligent reflective surface phase prediction result; the target perception area is reconstructed using the convolutional neural network for target perception area detection to obtain a one-dimensional target scattering coefficient estimation sequence for the target perception area, thereby achieving computational imaging of the target perception area. Therefore, the present invention proposes a method for wireless computational imaging assisted by an intelligent reflective surface based on deep learning, which can not only effectively process high-dimensional data with the assistance of an intelligent reflective surface, ensuring high-accuracy imaging in the context of big data, but can also automatically extract key features from images or signals to improve the accuracy of imaging results. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 This is an overall block diagram of a wireless computational imaging method assisted by an intelligent reflective surface based on deep learning;

[0051] Figure 2 This is a schematic diagram of the locations of base stations, target sensing areas, and smart reflective surfaces in a wireless computational imaging method assisted by a smart reflective surface based on deep learning;

[0052] Figure 3 This is a network structure diagram of the intelligent reflector phase prediction deep neural network and target perception area detection convolutional neural network of the wireless computational imaging method assisted by intelligent reflectors based on deep learning;

[0053] Figure 4 This is a detailed diagram of the dth sparse Bayesian layer in the target perception area detection convolutional neural network of a deep learning-based intelligent reflective surface-assisted wireless computational imaging method;

[0054] Figure 5 This is a comparison chart of the mean square error (MSE) of the fixed phase prediction target perception area imaging method and the method proposed in the embodiment of the present invention under different numbers of reflective units on the intelligent reflective surface;

[0055] Figure 6 This is a comparison diagram of the imaging results of the original target perception area in an embodiment of the present invention and the imaging results of the target perception area imaging method using a fixed phase prediction method and the method proposed in an embodiment of the present invention. DETAILED DESCRIPTION

[0056] In order to make the above-mentioned objects, features and advantages of the present invention more clearly understood, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings. In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways than those described herein, and those skilled in the art can make similar improvements without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below. The technical features in the various embodiments of the present invention can be combined accordingly without conflicting with each other.

[0057] In the description of the present invention, it should be understood that when an element is considered to be "connected" to another element, it can be directly connected to the other element or indirectly connected, that is, there are intermediate elements. On the contrary, when an element is said to be "directly" connected to another element, there are no intermediate elements.

[0058] In the description of the present invention, it should be understood that the terms "first" and "second" are used solely for descriptive purposes and are not to be construed as indicating or implying relative importance or implicitly specifying the number of technical features being described. Therefore, features defined as "first" or "second" may explicitly or implicitly include at least one of such features.

[0059] In 6G communications, facing limited spectrum resources and a continuously growing number of connected devices, effectively improving signal quality and expanding coverage has become a core technical challenge. The introduction of smart reflectors addresses this challenge. Deep learning technology has demonstrated outstanding performance in areas such as target detection and image recognition. Its ability to process big data and extract superior features opens new possibilities for computational imaging in complex scenarios. When combined with 6G wireless communication technologies, this technology can achieve higher imaging accuracy and improved real-time performance. Combining smart reflector technology with deep learning offers significant benefits for target detection. Smart reflector phase prediction maximizes the received signal-to-noise ratio, ensuring signal stability and quality in complex environments. By designing a neural network, we can quickly and accurately preprocess and recognize three-dimensional spatial information, effectively improving detection accuracy and real-time performance. Compared to traditional methods, this solution requires no additional sensors or detection equipment, relying solely on existing 6G communication equipment to achieve efficient computational imaging.

[0060] In summary, in the 6G era, the integration of wireless communication technology and deep learning technology has become a major development trend. The wireless computational imaging method assisted by intelligent reflective surfaces based on deep learning is the product of this trend, which brings more flexible and efficient computational imaging solutions to various application scenarios in the future. Therefore, the embodiment of the present invention adopts a wireless computational imaging method assisted by intelligent reflective surfaces based on deep learning, such as Figure 1As shown, it specifically includes the following steps:

[0061] S1. The base station transmits a sensing signal into the environment. The sensing signal is reflected by the smart reflecting surface, reaches the target sensing area, and is reflected by the target sensing area to obtain a reflected sensing signal. The reflected sensing signal is returned to the base station along the original path. Channel state information from the base station to the smart reflecting surface, channel state information from the smart reflecting surface to the target sensing area, and a received signal from the base station are obtained.

[0062] S2. Obtain a trained smart reflector phase prediction deep neural network, use the smart reflector phase prediction deep neural network to extract features of the received signal, and obtain a smart reflector phase prediction result;

[0063] S3. Obtain a trained target perception area detection convolutional neural network, use the target perception area detection convolutional neural network to reconstruct the target perception area, input the phase prediction result of the intelligent reflector into the target perception area detection convolutional neural network, and obtain a one-dimensional target scattering coefficient estimation sequence of the target perception area.

[0064] S4. Post-processing the one-dimensional target scattering coefficient estimation sequence, inversely mapping the one-dimensional target scattering coefficient estimation sequence into a three-dimensional pixel estimation sequence That is to achieve computational imaging of the target perception area.

[0065] The present invention will be further described and illustrated below with reference to the accompanying drawings and specific embodiments.

[0066] In a wireless computing imaging method based on deep learning and assisted by an intelligent reflective surface in this embodiment, the positions of the base station, the target sensing area, and the intelligent reflective surface are shown in FIG. Figure 2 As shown, in this embodiment of the present invention, a pre-deployed system consists of a smart reflective surface with multiple reflective units, a multi-antenna base station with integrated transceivers, and a three-dimensional target sensing area. Specifically, the base station first transmits a sensing signal into the environment. If the direct path—the path from the base station to the target sensing area—is blocked by an obstacle, the sensing signal travels along an auxiliary path from the smart reflective surface, first reflecting off the smart reflective surface to reach the target sensing area. After reflecting from the target sensing area, it returns along the original path to the base station, acquiring channel state information from the base station to the smart reflective surface and from the smart reflective surface to the target sensing area.

[0067] It should be noted that, in order to facilitate those skilled in the art to understand the technical solution, a method for obtaining a received signal is provided in an embodiment of the present invention, specifically:

[0068] S11. Obtain the channel matrix from the base station to the smart reflection surface as the first Rice channel Obtain the channel matrix of the smart reflective surface to the target sensing area as the second Rice channel The functional form of the first Ricean channel H is:

[0069]

[0070] The functional form of the second Ricean channel G is:

[0071]

[0072] Where, ∈ is the Rice factor; lp1 is the first channel path loss, lp2 is the second channel path loss; H LoS is the direct part of the first channel, G LoS is the direct part of the second channel; H NLoS is the first channel scattering part that obeys the standard normal distribution, G NLoS is the second scattered part that obeys the standard normal distribution; represents a complex set; M is the number of transmitting and receiving antennas of the base station; N is the number of reflecting units of the smart reflecting surface;

[0073] S12. Reflection coefficient of the smart reflective surface Among them, e represents a natural constant; is the phase of the reflection coefficient; [ρ1,…,ρ N ] is the amplitude of the reflection coefficient, assuming that the amplitude is 1;

[0074] S13. Preprocessing the target sensing area to obtain the one-dimensional target scattering coefficient sequence;

[0075] The preprocessing process is: first, the target perception area is discretized to obtain a three-dimensional pixel sequence s=[s1,…,s K ] T , and then map the three-dimensional pixel sequence to obtain a one-dimensional target scattering coefficient sequence x=[x1,…,x K ] T , where the i-th three-dimensional pixel s i The scattering coefficient is x i ; If the i-th three-dimensional pixel is a real target, the i-th three-dimensional pixel s i The scattering coefficient x i The value is 0 <x i ≤1; otherwise, the i-th three-dimensional pixel s i The scattering coefficient x iThe value is 0, i∈[1,2,…,K]; K is the length of the three-dimensional pixel sequence; x i is the i-th sample of the one-dimensional target scattering coefficient sequence;

[0076] S14. Assuming that the number of real target pixels in the target sensing area is much smaller than the number of ambient target pixels, the one-dimensional target scattering coefficient sequence x=[x1,…,x K ] T The sparse prior probability density has a mean of 0 and a covariance matrix of R x = diag(γ) Gaussian distribution, where diag(·) represents diagonalization, γ represents variance, and α is the inverse of the variance. α obeys the Gamma distribution, and the functional form of α is as follows:

[0077]

[0078] Where Γ(·) is the Gamma function, a and b are parameters that control sparsity; (·) -1 Indicates the inverse operation;

[0079] S15. Obtain the total transmission power P of the base station t , the transmission signal of the base station of length L The one-dimensional target scattering coefficient sequence x is diagonalized to obtain the target sensing area scattering coefficient diagonal matrix X=diag(x); the phase of the reflection coefficient is diagonalized to obtain the smart reflection surface phase diagonal matrix The received signal is constructed by using the total transmit power, the transmit signal, the first Ricean channel, the second Ricean channel, the smart reflector phase diagonal matrix, and the target sensing area scattering coefficient diagonal matrix. The functional form is:

[0080]

[0081] Among them, N is subject to the mean of 0 and variance σ 2 Additive Gaussian white noise; represents the conjugate transpose of the matrix;

[0082] S16. Introducing intermediate variables According to the matrix vectorization property, vec(ABC)=(C T ⊙A)diag(B), vectorize the received signal to obtain a vectorized received signal, where the function form of the vectorized received signal y is:

[0083]

[0084] in, ⊙ is Khatri-Rao product, (·) * denotes matrix conjugation, and vec(·) denotes vectorization.

[0085] It should be noted that the intelligent reflecting surface phase prediction deep neural network of the present invention is composed of a first fully connected layer with an activation function, a first Dropout layer, a second fully connected layer with an activation function, a second Dropout layer, and a third fully connected layer without an activation function, which are cascaded in sequence; the number of neurons in the first fully connected layer and the second fully connected layer is 256, and the activation function is Relu; the dropout rate of the first Dropout layer and the second Dropout layer is set to 0.5; the number of neurons in the third fully connected layer is N, and the activation function is Sigmoid.

[0086] In the embodiment of the present invention, the intelligent reflector phase prediction deep neural network is used to extract the features of the received signal, and the specific process of obtaining the intelligent reflector phase prediction result is as follows: the total transmit power P of the base station is t , the transmitted signal S, the first Ricean channel H, the second Ricean channel G, and the standard deviation σ of the additive white Gaussian noise are used as hyperparameters of the intelligent reflector phase prediction deep neural network, and the value of each hyperparameter of the intelligent reflector phase prediction deep neural network is fixed; randomly initialize the phase The one-dimensional target scattering coefficient sequence x=[x1,…,x K ] T Input into the intelligent reflector phase prediction deep neural network to obtain the intelligent reflector phase prediction result in represents the intelligent reflector phase prediction deep neural network, represents the set of real numbers.

[0087] It should be noted that, since the input and output of the above-mentioned intelligent reflector phase prediction deep neural network are both real numbers, in the embodiment of the present invention, the intelligent reflector phase prediction result is converted to As the output of the deep neural network for intelligent reflector phase prediction.

[0088] The target perception area detection convolutional neural network of the present invention is composed of D sparse Bayesian layers cascaded in sequence, and each of the sparse Bayesian layers is composed of a custom Lambda layer, a first Reshape layer, a first convolutional layer with an activation function, a second convolutional layer with an activation function, and a second Reshape layer cascaded in sequence. The convolution kernel size of the first convolutional layer is 5×5×5, and the number of convolution kernels is 8; the convolution kernel size of the second convolutional layer is 5×5×5, and the number of convolution kernels is 1; and the activation functions are all Relu.

[0089] In the embodiment of the present invention, the target perception area is reconstructed using the target perception area detection convolutional neural network to obtain a one-dimensional target scattering coefficient estimation sequence of the target perception area. The specific process is as follows: the phase prediction result of the smart reflector The smart reflector phase diagonal matrix is ​​updated to obtain the smart reflector phase diagonal update matrix The received signal is updated by the phase diagonal update matrix of the smart reflector to obtain an updated received signal, wherein the updated received signal y opt The functional form is:

[0090]

[0091] Update intermediate variables y opt and A opt As the target perception area detection convolutional neural network hyperparameter, the target perception area detection convolutional neural network hyperparameter value is fixed; randomly initialize the covariance matrix The one-dimensional target scattering coefficient sequence x=[x1,…,x K ] T Input into the target perception area detection convolutional neural network to obtain the one-dimensional target scattering coefficient estimation sequence in represents the target perception area detection convolutional neural network; ⊙ is the Khatri-Rao product, (·) * represents matrix conjugation.

[0092] It should be noted that, in the dth sparse Bayesian layer in the embodiment of the present invention, the updated value of the covariance matrix is ​​output by the (d-1)th sparse Bayesian layer. The update reception signal y opt , the intermediate update variable A opt , the updated value of the covariance matrix Input into the dth sparse Bayesian layer and update the variance γ d , update the covariance matrix where d∈[1,…,D].

[0093] The custom Lambda layer in the embodiment of the present invention uses the minimum mean square error filter to perform a preliminary estimate of the one-dimensional target scattering coefficient sequence x based on the sparse prior of the one-dimensional target scattering coefficient sequence x, and obtains an updated posterior mean. and update the first variance

[0094] The updated posterior mean The functional form is:

[0095]

[0096] The updated first variance The functional form is:

[0097]

[0098] Where I is the identity matrix.

[0099] It should be noted that, in this embodiment, each intermediate process in the above-mentioned sparse Bayesian layer and the dimensions of each parameter are further described so that those skilled in the art can better understand the principle of using the sparse Bayesian layer. For each sparse Bayesian layer, the updated received signal, the intermediate updated variable, and the updated value of the covariance matrix output by the previous sparse Bayesian layer are first taken as input. The following takes the dth sparse Bayesian layer as an example to describe its process in detail. First, the updated received signal y opt , intermediate update variable A opt , updated value of the covariance matrix Input into the custom Lambda layer of the dth sparse Bayesian layer to obtain the posterior mean μ d and variance θ d , use them as the first feature F1 and the second feature F2 respectively, input the first feature F1 and the second feature F2 into the first Reshape layer respectively, and perform shape transformation on the first feature F1 and the second feature F2 in the first Reshape layer, changing them from (K, 1) to (P, P, P, 1); next, input the first feature F1 and the second feature F2 after shape transformation into two connected 3D convolutional layers (first convolutional layer and second convolutional layer), set the padding attribute to SAME to ensure that the network input and output sizes are equal, and perform feature extraction by the 3D convolutional layer. Then, input the features extracted by the convolutional layer into the second Reshape layer, transform the extracted features into (P, P, P, 1) into (K, 1), and finally output the updated variance γ d and covariance matrix

[0100] The loss function used by the intelligent reflecting surface phase prediction deep neural network in the embodiment of the application during training is a received signal signal-to-noise ratio of the base station, and the network parameters of the intelligent reflecting surface phase prediction deep neural network are updated, and the function form of the received signal signal-to-noise ratio is:

[0101]

[0102] Wherein, N S is the total number of training samples; represents the square of the Frobenius-norm of a matrix; Φ j is the jth sample of Φ.

[0103] The target perception zone detection convolutional neural network in the embodiment of the application uses the one-dimensional target scattering coefficient sequence x = [x1,..., x K ] T and the mean square error of the one-dimensional target scattering coefficient estimation sequence of the target perception zone is the loss function, and the network parameters of the target perception zone detection convolutional neural network are updated, and the function form of the mean square error is:

[0104]

[0105] Wherein, x j is the jth sample of x; is the jth sample of .

[0106] The above deep learning-based intelligent reflecting surface assisted wireless computing imaging method will be applied to a specific embodiment below to demonstrate the technical effects it can achieve. Figure 3 The network structure diagram of the intelligent reflecting surface phase prediction deep neural network and the target perception zone detection convolutional neural network of a deep learning-based intelligent reflecting surface assisted wireless computing imaging method; Figure 4 The specific details diagram of the dth sparse Bayesian layer in the target perception zone detection convolutional neural network of a deep learning-based intelligent reflecting surface assisted wireless computing imaging method; Figure 5The present invention compares the mean square error (MSE) of the fixed phase prediction target perception area imaging method and the method proposed in the embodiment of the present invention under different numbers of intelligent reflection surface reflection units. The results show that as the number N of intelligent reflection surface units increases, the detection mean square error (MSE) of the proposed method is always lower than the MSE of the fixed intelligent reflection surface, and the MSE shows a gradually decreasing trend. Among them, the fixed phase prediction target perception area imaging method uses a random phase as the phase prediction result of the intelligent reflection surface, and then uses the target perception area detection convolutional neural network to reconstruct the target perception area, and then perform computational imaging of the target perception area. Figure 6 : is a comparison diagram of the imaging results of the method for predicting the target perception area using a fixed phase in an embodiment of the present invention and the method proposed in an embodiment of the present invention, Figure 6 The three imaging figures are the imaging results of the original target perception area, the imaging results of the fixed phase prediction target perception area imaging method, and the imaging results of the method proposed in the embodiment of the present invention. Figure 6 As can be seen from the figure, the imaging effect of the method proposed in this embodiment of the present invention is the best, approaching the original image. However, the imaging effect of the target perception area predicted by the fixed phase is poor. Therefore, this invention combines deep learning to propose a deep learning-based intelligent reflective surface-assisted wireless computational imaging method, providing a better technical solution for wireless imaging in the field of wireless communications.

[0107] The embodiment described above is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Persons skilled in the art may make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, any technical solution obtained by equivalent substitution or equivalent transformation falls within the scope of protection of the present invention.

Claims

1. A wireless computational imaging method assisted by an intelligent reflective surface based on deep learning, characterized in that: The steps include: S1. The base station transmits a sensing signal into the environment. The sensing signal is reflected by the smart reflecting surface, reaches the target sensing area, and is reflected by the target sensing area to obtain a reflected sensing signal. The reflected sensing signal is returned to the base station along the original path. Channel state information from the base station to the smart reflecting surface, channel state information from the smart reflecting surface to the target sensing area, and a received signal from the base station are obtained. S2. Obtain a trained smart reflector phase prediction deep neural network, use the smart reflector phase prediction deep neural network to extract features of the received signal, and obtain a smart reflector phase prediction result; S3. Obtain a trained target perception area detection convolutional neural network, reconstruct the target perception area using the target perception area detection convolutional neural network, input the intelligent reflector phase prediction result into the target perception area detection convolutional neural network, and obtain a one-dimensional target scattering coefficient estimation sequence for the target perception area; S4. Post-processing the one-dimensional target scattering coefficient estimation sequence, inversely mapping the one-dimensional target scattering coefficient estimation sequence into a three-dimensional pixel estimation sequence, thereby achieving computational imaging of the target perception area; The method for obtaining the received signal is: S11. Obtain the channel matrix from the base station to the smart reflection surface as the first Rice channel Obtain the channel matrix of the smart reflective surface to the target sensing area as the second Rice channel The functional form of the first Ricean channel H is: The functional form of the second Ricean channel G is: Where, ∈ is the Rice factor; lp1 is the first channel path loss, lp2 is the second channel path loss; H LoS is the direct part of the first channel, G LoS is the direct part of the second channel; H NLoS is the first channel scattering part that obeys the standard normal distribution, G NLoS is the second scattered part that obeys the standard normal distribution; represents a complex set; M is the number of transmitting and receiving antennas of the base station; N is the number of reflecting units of the smart reflecting surface; S12. Reflection coefficient of the smart reflective surface Among them, e represents a natural constant; is the phase of the reflection coefficient; [ρ1,…,ρ N ] is the amplitude of the reflection coefficient, assuming that the amplitude is 1; where (·) T Represents matrix transpose; S13. Preprocessing the target sensing area to obtain the one-dimensional target scattering coefficient sequence; The preprocessing process is: first, the target perception area is discretized to obtain a three-dimensional pixel sequence s=[s1,…,s K ] T , and then map the three-dimensional pixel sequence to obtain a one-dimensional target scattering coefficient sequence x=[x1,…,x K ] T , where the i-th three-dimensional pixel s i The scattering coefficient is x i ; If the i-th three-dimensional pixel is a real target, the i-th three-dimensional pixel s i The scattering coefficient x i The value is 0 <x i ≤1; otherwise, the i-th three-dimensional pixel s i The scattering coefficient x i The value is 0, i∈[1,2,…,K]; K is the length of the three-dimensional pixel sequence; x i is the i-th sample of the one-dimensional target scattering coefficient sequence; S14. Assuming that the number of real target pixels in the target sensing area is much smaller than the number of ambient target pixels, the one-dimensional target scattering coefficient sequence x=[x1,…,x K ] T The sparse prior probability density has a mean of 0 and a covariance matrix of R x = diag(γ) Gaussian distribution, where diag(·) represents diagonalization, γ represents variance, and α is the inverse of the variance. α obeys the Gamma distribution, and the functional form of α is as follows: Where Γ(·) is the Gamma function, a and b are parameters that control sparsity; (·) -1 Indicates the inverse operation; S15. Obtain the total transmission power P of the base station t , the transmission signal of the base station of length L The one-dimensional target scattering coefficient sequence x is diagonalized to obtain the target sensing area scattering coefficient diagonal matrix X=diag(x); the phase of the reflection coefficient is diagonalized to obtain the smart reflection surface phase diagonal matrix The received signal is constructed by using the total transmit power, the transmit signal, the first Ricean channel, the second Ricean channel, the smart reflector phase diagonal matrix, and the target sensing area scattering coefficient diagonal matrix. The functional form is: Among them, N is subject to the mean of 0 and variance σ 2 Additive Gaussian white noise; represents the conjugate transpose of the matrix; S16. According to the matrix vectorization property, the received signal is vectorized to obtain a vectorized received signal, wherein the function form of the vectorized received signal y is: Among them, vec(·) represents vectorization; The intelligent reflector phase prediction deep neural network is composed of a first fully connected layer with an activation function, a first dropout layer, a second fully connected layer with an activation function, a second dropout layer, and a third fully connected layer without an activation function, which are cascaded in sequence; the number of neurons in the first fully connected layer and the second fully connected layer is 256, and the activation function is Relu; the dropout rate of the first dropout layer and the second dropout layer is set to 0.5; the number of neurons in the third fully connected layer is N, and the activation function is Sigmoid; The target perception area detection convolutional neural network is composed of D sparse Bayesian layers cascaded in sequence, and each of the sparse Bayesian layers is composed of a custom Lambda layer, a first Reshape layer, a first convolutional layer with an activation function, a second convolutional layer with an activation function, and a second Reshape layer cascaded in sequence. The convolution kernel size of the first convolutional layer is 5×5×5, and the number of convolution kernels is 8; the convolution kernel size of the second convolutional layer is 5×5×5, and the number of convolution kernels is 1; the activation function is ReLU; The specific process of step S2 is: the total transmission power P of the base station t , the transmitted signal S, the first Ricean channel H, the second Ricean channel G, and the standard deviation σ of the additive white Gaussian noise are used as hyperparameters of the intelligent reflector phase prediction deep neural network, and the value of each hyperparameter of the intelligent reflector phase prediction deep neural network is fixed; randomly initialize the phase The one-dimensional target scattering coefficient sequence x=[x1,…,x K ] T Input into the intelligent reflector phase prediction deep neural network to obtain the intelligent reflector phase prediction result in represents the intelligent reflector phase prediction deep neural network, represents the set of real numbers; The specific process of the step S3 is: The smart reflector phase diagonal matrix is ​​updated to obtain the smart reflector phase diagonal update matrix The received signal is updated by the phase diagonal update matrix of the smart reflector to obtain an updated received signal, wherein the updated received signal y opt The functional form is: Introducing intermediate update variables y opt and A opt As the target perception area detection convolutional neural network hyperparameter, each target perception area detection convolutional neural network hyperparameter has a fixed value; randomly initialize the covariance matrix The one-dimensional target scattering coefficient sequence x=[x1,…,x K ] T Input into the target perception area detection convolutional neural network to obtain the one-dimensional target scattering coefficient estimation sequence in represents the target perception area detection convolutional neural network; ⊙ is the Khatri-Rao product, (·) * represents matrix conjugation; In the dth sparse Bayesian layer, the updated value of the covariance matrix is ​​output from the (d-1)th sparse Bayesian layer. The update reception signal y opt , the intermediate update variable A opt , the updated value of the covariance matrix Input into the dth sparse Bayesian layer and update the variance γ d , update the covariance matrix Where, d∈[1,…,D]; The custom Lambda layer uses the minimum mean square error filter to perform a preliminary estimate of the one-dimensional target scattering coefficient sequence x based on the sparse prior of the one-dimensional target scattering coefficient sequence x, and obtains an updated posterior mean. and update the first variance The updated posterior mean The functional form is: The updated first variance The functional form is: Where I is the identity matrix; The loss function used in the training of the intelligent reflector phase prediction deep neural network is the noise ratio of the received signal. The network parameters of the intelligent reflector phase prediction deep neural network are updated. The function form of the noise ratio of the received signal is: Among them, N S is the total number of training samples; represents the square of the Frobenius-norm of the matrix; Φ j is the j-th sample of Φ; The target perception area detection convolutional neural network adopts the one-dimensional target scattering coefficient sequence x=[x1,…,x K ] T The one-dimensional target scattering coefficient estimation sequence of the target sensing area The mean square error is used as the loss function to update the network parameters of the target perception area detection convolutional neural network. The function form of the mean square error is: Among them, x j is the jth sample of x; for The jth sample of .

Images