A deep learning-based communication and perception fusion method

Abstract

The application discloses a kind of communication and perception fusion methods based on deep learning.The system is composed of a communication-perception dual-function base station, multiple single-antenna mobile terminals and several targets.According to the communication channel state information and the sensing transmission correlation matrix, the base station first designs the sensing transmission waveform and the communication receiving beam based on deep learning.Then, the base station broadcasts the sensing waveform through the downlink to perceive the surrounding environment, and the mobile terminal transmits the communication signal to the base station through the uplink for message transmission.Finally, the base station receives the echo signal from the target reflection and the communication signal transmitted by the mobile terminal, restores the target response matrix in the echo signal to extract the surrounding environment information, and decodes the communication signal through the communication receiving beam.The application provides an effective communication and perception fusion method based on deep learning for wireless networks.

Description

Technical Field

[0001] The present invention relates to the fields of wireless communications and machine learning, and in particular to a communication and perception fusion method based on deep learning. Background Art

[0002] In future wireless networks, communication and perception are essential functions for enabling emerging intelligent services such as autonomous driving, immersive interactions, and smart cities. However, for the past few decades, communication and perception have been implemented separately using different frequency resources and hardware infrastructure. This inevitably wastes wireless resources, leads to low spectrum utilization, and fails to meet the high-performance requirements of future wireless network applications. Because communication and perception can share the same wireless resources and hardware infrastructure, and because of their similarities in signal processing algorithms, the convergence of communication and perception is an inevitable trend to improve wireless network spectrum efficiency and increase overall system throughput.

[0003] On the other hand, artificial intelligence technologies, represented by machine learning, have demonstrated significant advantages in information processing, spurring the application of deep learning in wireless communications. For example, deep learning has been applied to many classic wireless communications problems, such as signal detection, channel estimation, beamforming, and resource management. A key advantage of deep learning is that, with abundant training samples, complex computational tasks are offloaded to the offline training phase. In this case, only simple forward computation is required during the online prediction phase to achieve the desired results. In other words, compared to traditional iterative algorithms, deep learning-based methods achieve fast and efficient computation. Therefore, deep learning offers a new, low-complexity design approach for wireless networks, significantly increasing the potential for the deep integration of communication and perception.

[0004] Therefore, designing a communication and perception fusion method based on deep learning for future wireless networks can make full use of limited wireless resources and expensive hardware resources, effectively improve network spectrum efficiency, reduce computational complexity, and thus meet the high performance requirements of emerging intelligent services. Summary of the Invention

[0005] In order to solve the problems of low spectrum efficiency, separation of communication and perception functions, and high computational complexity in wireless networks, the present invention proposes a communication and perception fusion method based on deep learning.

[0006] The specific technical solutions adopted in the present invention are as follows:

[0007] A communication and perception fusion method based on deep learning, comprising the following steps:

[0008] 1) Deploy a communication-sensing dual-function base station in the center of the cell. The base station has Nt root transmit antenna and N r root receive antenna, K mobile terminals access the base station through a wireless network, and there are several targets, wherein the time slot length of a communication frame is L;

[0009] 2) The base station and the mobile terminal exchange information to obtain communication state information, wherein the communication channel state information of the base station to the kth mobile terminal is h k ,k=1,…,K, and the sensing transmission correlation matrix R T of the target response matrix G is obtained according to the prior information;

[0010] 3) According to the communication channel state information and the sensing transmission correlation matrix, the communication rate and the sensing mutual information are calculated respectively as the performance indicators of communication and sensing;

[0011] 4) The base station designs the sensing transmission waveform S and the communication receiving beam w k of each mobile terminal using deep learning,

[0012] 5) The base station broadcasts the sensing transmission waveform through the downlink to perform surrounding environment sensing, and the mobile terminal transmits the communication signal to the base station through the uplink to perform message transmission;

[0013] 6) The base station receives the echo signal from the target reflection and the communication signal transmitted by the mobile terminal at the same time, on the one hand, the target response matrix in the echo signal is recovered to extract the surrounding environment information, and on the other hand, the communication signal is decoded through the communication receiving beam.

[0014] As preferred, in step 3), the calculation method of the communication rate and the sensing mutual information is:

[0015] 3.a) The calculation formula of the communication rate is wherein is the signal-to-interference-plus-noise ratio of the base station receiving the signal transmitted by the kth mobile terminal, wherein p k is the power of the signal transmitted by the kth mobile terminal, is the variance of the Gaussian white noise, ||·|| represents the 2-norm, (·) T represents transposition, (·) H represents conjugate transposition;

[0016] 3.b) The calculation formula of the sensing mutual information is wherein is a K-dimensional diagonal matrix whose diagonal elements are p1,…,p K , (·) -1 represents the inverse of a matrix, represents the Kronecker product, represents the dimension of L×N rwhere det(·) represents the determinant and diag(x) represents the diagonal matrix whose diagonal elements are the vector x.

[0017] As a preference, in step 4), the base station uses deep learning to design the perception transmission waveform S and the communication receiving beam w of each mobile terminal. k ,k=1,…,K method is:

[0018] 4.a) Perform singular value decomposition on S and obtain in is the i-th singular value of S, i=1,…,N t , and satisfy U s and V s S is the left singular value matrix and the right singular value matrix, Indicates dimension N t ×(LN t ), (·) 1 / 2 Indicates taking square roots of matrix elements;

[0019] 4.b) R T Perform eigenvalue decomposition and get in σ t,i R T The i-th eigenvalue of t , and meet U T It is R T The eigenvalue matrix of

[0020] 4.c) R H Perform eigenvalue decomposition and get in σ h,j R H The j-th eigenvalue of r , and satisfy U h It is R H The eigenvalue matrix of

[0021] 4.d) Let the right singular value matrix of S be equal to R T The eigenvalue matrix of V s =U T , then the perception mutual information is updated as

[0022] 4.e) Using the Lagrange multiplier method, the optimal communication receiving beam that maximizes the communication rate is obtained as The communication rate is updated to in Indicates h k conjugation of;

[0023] 4.f) According to the perceived signal transmission power limit, let Among them, P s is the maximum transmit power of the sensing signal;

[0024] 4.g) Find the maximum achievable communication rate M c And the maximum achievable perceptual mutual information M s ;

[0025] 4.h) Build and train a deep learning network to find the overall goal of the system, which is the weighted sum of communication rate and perceptual mutual information. The solution with the largest value σ s,i , i=1,…,N t , that is, the perception transmission waveform S and communication receiving beam w are obtained k , where α is the communication weight and 1-α is the perception weight.

[0026] As a preference, in step 4.g), the maximum achievable communication rate is obtained by using the Lagrange multiplier method. Using water injection algorithm to obtain the maximum reachable mutual information in max(x,y) means taking the maximum value of x and y.

[0027] Preferably, the deep learning network constructed in step 4.h) is as follows:

[0028] The network input consists of the communication channel state information h1,…,h K The real and imaginary parts of the matrix are related to the eigenvalues ​​of the perceptual emission correlation matrix The dimension of the composition is 2N r K+N t The real-valued feature vector of the network structure consists of four fully connected layers, and the number of neurons is 16N t , 8N t , 4N t , N t , the activation function of the first three layers is Relu function, the activation function of the fourth layer is Sigmoid function, and a batch normalization layer is added before each fully connected layer; the network output is Lambda layer to meet the transmission power constraint. The function of Lambda function is to normalize the input vector and multiply it by P s , then sort the elements of the vector in descending order; the loss function of the network is Where N is the number of training samples, is the weighted sum of the communication rate and the perceptual mutual information of the qth sample.

[0029] Preferably, during the training phase of the deep learning network, the number of training samples is set to 10,000, the Adam optimizer is used with an initial learning rate of 0.001, the maximum epoch is set to 1,000, and the Early-stopping tool with a patience of 20 is used to improve training efficiency; in addition, to accelerate convergence, the learning rate is decayed by a factor of 0.33 as long as the verification loss does not decrease in 10 consecutive epochs.

[0030] The present invention has the following beneficial effects: The deep learning-based communication and perception fusion method proposed in this invention solves a series of problems caused by low spectrum utilization, high computational complexity, and separation of communication and perception functions in wireless networks. The deep learning-based algorithm for jointly designing the perception transmit waveform and the communication receive beam has the advantages of high communication efficiency, high perception accuracy, and low computational complexity. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a system block diagram of communication and perception fusion;

[0032] Figure 2 This is the proposed deep learning network architecture diagram;

[0033] Figure 3 The performance comparison of the proposed method is shown in the case of different maximum sensing signal transmission powers, where K = 5, L = 10, N t =4,N r =4.

[0034] Figure 4 The performance comparison of the proposed method is shown in the case of different numbers of mobile terminals and base station antennas, where L = 10, P s =10. DETAILED DESCRIPTION

[0035] In this embodiment, the system block diagram of communication and perception fusion in a wireless network is as follows: Figure 1 As shown in Figure 1, the system consists of a communication-sensing dual-function base station, K single-antenna mobile terminals, and several targets. The base station first uses prior information to build a deep learning network such as Figure 2As shown, a sensing transmit waveform and a communication receive beam are designed. Next, the base station broadcasts the sensing waveform via the downlink to sense the surrounding environment, and the mobile terminal transmits a communication signal to the base station via the uplink to communicate the message. Finally, the base station simultaneously receives the echo signal reflected from the target and the communication signal transmitted by the mobile terminal. It recovers the target response matrix from the echo signal to extract surrounding environment information and decodes the communication signal using the communication receive beam. The specific technical solutions adopted in this embodiment are as follows:

[0036] A communication and perception fusion method based on deep learning, comprising the following steps:

[0037] 1) Deploy a communication-sensing dual-function base station in the center of the cell. The base station has N t Transmitting antennas and N r There are K receiving antennas, K mobile terminals access this base station through the wireless network, and there are several targets at the same time. The time slot length of one communication frame is L.

[0038] 2) The base station and the mobile terminal exchange information to obtain communication status information, where the communication channel status information from the base station to the kth mobile terminal is h k ,k=1,…,K, and at the same time obtain the perception emission correlation matrix R of the target response matrix G according to the prior information T .

[0039] 3) Based on the communication channel state information and the perception-emission correlation matrix, the communication rate and perception mutual information are calculated as the performance indicators of communication and perception respectively.

[0040] In this embodiment, in the above step 3), the calculation method of the communication rate and the perceptual mutual information is:

[0041] 3.a) The formula for calculating the communication rate is in is the signal to interference and noise ratio of the signal transmitted by the kth mobile terminal received by the base station, where p k is the power of the signal transmitted by the kth mobile terminal, is the variance of Gaussian white noise, ||·|| represents the 2-norm, (·) T represents transpose, (·) H represents conjugate transpose;

[0042] 3.b) The calculation formula of perceptual mutual information is: in The diagonal elements are p1,…,p K The K-dimensional diagonal matrix, (·) -1 represents the inverse of the matrix, represents the Kronecker product, Indicates dimension is L×Nr The identity matrix, det(·) represents the determinant, diag(x) represents the diagonal matrix whose diagonal elements are vector x, and vector x is a general variable that can be selected according to actual conditions;

[0043] 4) The base station uses deep learning to design the sensing transmission waveform S and the communication receiving beam w of each mobile terminal k ,k=1,…,K。

[0044] In this embodiment, in the above step 4), the base station uses deep learning to design the perception transmission waveform S and the communication receiving beam w of each mobile terminal. k ,k=1,…,K method is:

[0045] 4.a) Perform singular value decomposition on S and obtain in is the i-th singular value of S, i=1,…,N t , and satisfy U s and V s S is the left singular value matrix and the right singular value matrix, Indicates dimension N t ×(LN t ), (·) 1 / 2 Indicates taking square roots of matrix elements;

[0046] 4.b) R T Perform eigenvalue decomposition and get in σ t,i R T The i-th eigenvalue of t , and meet U T It is R T The eigenvalue matrix of

[0047] 4.c) R H Perform eigenvalue decomposition and get in σ h,j R H The j-th eigenvalue of r , and satisfy U h It is R H The eigenvalue matrix of

[0048] 4.d) Let the right singular value matrix of S be equal to R T The eigenvalue matrix of V s =UT , then the perception mutual information is updated as

[0049] 4.e) Using the Lagrange multiplier method, the optimal communication receiving beam that maximizes the communication rate is obtained as The communication rate is updated to in Indicates h k The conjugation of .

[0050] 4.f) According to the perceived signal transmission power limit, let Among them, P s is the maximum transmit power of the sensing signal;

[0051] 4.g) Use the Lagrange multiplier method to find the maximum achievable communication rate Using water injection algorithm to obtain the maximum reachable mutual information in max(x,y) means taking the maximum value of x and y;

[0052] 4.h) Build and train a deep learning network to find the overall goal of the system, which is the weighted sum of communication rate and perceptual mutual information. The solution with the largest value σ s,i , i=1,…,N t , that is, the perception transmission waveform S and communication receiving beam w are obtained k , where α is the communication weight and 1-α is the perception weight. In this embodiment, the deep learning network and training method are constructed as follows: the network input is composed of the communication channel state information h1,...,h K The real and imaginary parts of the matrix are related to the eigenvalues ​​of the perceptual emission correlation matrix The dimension of the composition is 2N r K+N t The real-valued feature vector of the network structure consists of four fully connected layers, and the number of neurons is 16N t , 8N t , 4N t , N t , the activation function of the first three layers is Relu function, and the activation function of the fourth layer is Sigmoid function. In order to ensure convergence, a batch normalization layer is added before each fully connected layer; a Lambda layer is set at the network output to meet the transmission power constraint. The function of the Lambda function is to normalize the input vector and multiply it by P s , then sort the elements of the vector in descending order; the loss function of the network is Where N is the number of training samples, is the weighted sum of the communication rate and perceptual mutual information of the qth sample. During network training, the number of training samples was set to 10,000, the Adam optimizer was used with an initial learning rate of 0.001, the maximum epoch number was set to 1,000, and early-stopping with a patience of 20 was used to improve training efficiency. Furthermore, to accelerate convergence, the learning rate was decayed by a factor of 0.33 if the validation loss did not decrease for 10 consecutive epochs.

[0053] 5) The base station broadcasts a sensing waveform via the downlink to sense the surrounding environment, and the mobile terminal transmits a communication signal to the base station via the uplink to transmit messages.

[0054] 6) The base station simultaneously receives the echo signal reflected from the target and the communication signal transmitted by the mobile terminal. On the one hand, it recovers the target response matrix in the echo signal to extract the surrounding environment information, and on the other hand, it decodes the communication signal through the communication receiving beam.

[0055] Computer simulations show that Figure 3 As shown, in the communication and perception fusion method based on deep learning proposed by the present invention, the greater the maximum perception transmission power, the higher the perception rate and the lower the communication rate. Figure 4 This shows that, with the proposed method, the greater the number of mobile terminals, the higher the communication rate and the lower the perception rate. Furthermore, as the number of antennas increases, perception and communication performance can be significantly improved. Therefore, this invention provides an effective deep learning-based communication and perception fusion method for wireless networks.

Claims

1. A communication and perception fusion method based on deep learning, characterized in that: The steps include: 1) Deploy a communication-sensing dual-function base station in the center of the cell. The base station has N t Transmitting antennas and N r There are K receiving antennas, K mobile terminals access this base station through the wireless network, and there are several targets at the same time. The time slot length of one communication frame is L; 2) The base station and the mobile terminal exchange information to obtain communication status information, where the communication channel status information from the base station to the kth mobile terminal is h k ,k=1,…,K, and at the same time obtain the perception emission correlation matrix R of the target response matrix G according to the prior information T ; 3) Based on the communication channel state information and the sensing-emission correlation matrix, the communication rate and sensing mutual information are calculated as the performance indicators of communication and sensing respectively; 4) The base station uses deep learning to design the sensing transmission waveform S and the communication receiving beam w of each mobile terminal k ,k=1,…,K; 5) The base station uses the downlink broadcast sensing waveform to sense the surrounding environment, and the mobile terminal transmits communication signals to the base station through the uplink for message transmission; 6) The base station simultaneously receives the echo signal reflected from the target and the communication signal transmitted by the mobile terminal. On the one hand, it recovers the target response matrix in the echo signal to extract the surrounding environment information, and on the other hand, it decodes the communication signal through the communication receiving beam; In step 4), the base station uses deep learning to design the perception transmission waveform S and the communication receiving beam w of each mobile terminal k ,k=1,…,K The method is: 4.a) Perform singular value decomposition on S and obtain in is the i-th singular value of S, i=1,…,N t , and satisfy U s and V s S is the left singular value matrix and the right singular value matrix, Indicates dimension N t ×(LN t ), (·) 1 / 2 Indicates taking square roots of matrix elements; 4.b) R T Perform eigenvalue decomposition and get in σ t,i R T The i-th eigenvalue of t , and meet U T It is R T The eigenvalue matrix of 4.c) R H Perform eigenvalue decomposition and get in σ h,j R H The j-th eigenvalue of r , and satisfy U h It is R H The eigenvalue matrix of 4.d) Let the right singular value matrix of S be equal to R T The eigenvalue matrix of V s =U T , then the perception mutual information is updated as 4.e) Using the Lagrange multiplier method, the optimal communication receiving beam that maximizes the communication rate is obtained as The communication rate is updated to in Indicates h k conjugation of; 4.f) According to the perceived signal transmission power limit, let Among them, P s is the maximum transmit power of the sensing signal; 4.g) Find the maximum achievable communication rate M c And the maximum achievable perceptual mutual information M s ; 4.h) Build and train a deep learning network to find the overall goal of the system, which is the weighted sum of communication rate and perceptual mutual information. The solution with the largest value σ s,i , i=1,…,N t , that is, the perception transmission waveform S and communication receiving beam w are obtained k , where α is the communication weight and 1-α is the perception weight.

2. A communication and perception fusion method based on deep learning according to claim 1, characterized in that: In step 3), the calculation method of communication rate and perceptual mutual information is: 3.a) The formula for calculating the communication rate is in is the signal to interference and noise ratio of the signal transmitted by the kth mobile terminal received by the base station, where p k is the power of the signal transmitted by the kth mobile terminal, is the variance of Gaussian white noise, ||·|| represents the 2-norm, (·) T represents transpose, (·) H represents the conjugate transpose; 3.b) The calculation formula of perceptual mutual information is: in The diagonal elements are p1,…,p K The K-dimensional diagonal matrix, (·) -1 represents the inverse of the matrix, represents the Kronecker product, Indicates dimension is L×N r where det(·) represents the determinant and diag(x) represents the diagonal matrix whose diagonal elements are the vector x.

3. The communication and perception fusion method based on deep learning according to claim 1, characterized in that: In step 4.g), the maximum achievable communication rate is obtained by using the Lagrange multiplier method. Using water injection algorithm to obtain the maximum reachable mutual information in max(x,y) means taking the maximum value of x and y.

4. The communication and perception fusion method based on deep learning according to claim 1, characterized in that: The deep learning network constructed in step 4.h) is as follows: The network input consists of the communication channel state information h1,...,h K The real and imaginary parts of the matrix are related to the eigenvalues ​​of the perceptual emission correlation matrix The dimension of the composition is 2N r K+N t The real-valued feature vector of the network structure consists of four fully connected layers, and the number of neurons is 16N t , 8N t , 4N t , N t , the activation function of the first three layers is Relu function, the activation function of the fourth layer is Sigmoid function, and a batch normalization layer is added before each fully connected layer; the network output is Lambda layer to meet the transmission power constraint. The function of Lambda function is to normalize the input vector and multiply it by P s , then sort the elements of the vector in descending order; the loss function of the network is Where N is the number of training samples, is the weighted sum of the communication rate and the perceptual mutual information of the qth sample.

5. The communication and perception fusion method based on deep learning according to claim 4, characterized in that: During the training phase of the deep learning network, the number of training samples was set to 10,000, the Adam optimizer was used with an initial learning rate of 0.001, the maximum epoch was set to 1,000, and the early-stopping tool with a patience of 20 was used to improve training efficiency. In addition, to accelerate convergence, the learning rate was decayed by a factor of 0.33 as long as the verification loss did not decrease in 10 consecutive epochs.

Images