Intelligent reflecting surface assisted environment perception and communication optimization method based on deep learning

Abstract

The application discloses an environment perception and communication optimization method assisted by an intelligent reflecting surface based on deep learning; in the method, a plurality of users UE send pilot signals to an access point AP through reflecting surface assistance in a plurality of observation time frames. The AP constructs a deep learning network based on a residual network, a long short-term memory network LSTM and a multi-layer perception machine MLP, including an adaptive reflecting surface optimization network that optimizes reflecting surface coefficients frame by frame, an environment perception network that obtains the point cloud distribution of a perception space from received signals, and a communication optimization network that outputs downlink communication beamforming and a reflecting surface reflection matrix. The application can effectively improve the environment perception accuracy of the AP and the system downlink transmission rate.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of wireless communication, in particular to a method for environment perception and communication optimization assisted by intelligent reflecting surface based on deep learning. BACKGROUND

[0002] Integrated sensing and communication (ISAC) combines communication and sensing functions on the same hardware and spectrum resources, enabling not only data transmission but also environment perception and monitoring. This technology can maintain efficient communication while precisely positioning, imaging, and tracking targets using reflected, scattered, or diffracted signals, widely applied in intelligent transportation, autonomous driving, augmented reality (AR), and smart cities. Intelligent reflecting surface (RIS) further enhances the performance of wireless communication networks. RIS consists of a large array of controllable reflecting elements that reshape the wireless signal propagation path by controlling the phase, amplitude, and direction of incident electromagnetic waves, thereby optimizing signal coverage, stability, and transmission rate. Unlike traditional relays or base stations, RIS does not actively transmit signals but achieves signal enhancement through intelligent reflection. This technology not only significantly reduces communication network energy consumption and deployment costs but also dynamically addresses issues such as multipath effects and signal fading in complex environments. The introduction of neural networks enables integrated sensing and communication and intelligent reflecting surface systems to have strong adaptive and self-optimizing capabilities. Through deep learning and machine learning algorithms, neural networks can extract complex patterns and features from large amounts of real-time communication and sensing data and make real-time adjustments based on environmental changes. For example, neural networks can predict the optimal reflection path, phase adjustment, and spectrum resource allocation by analyzing historical data and current network status, thereby improving network transmission efficiency and perception accuracy. Neural networks can also effectively handle interference and noise in nonlinear and complex environments, enhancing the system's adaptability and robustness to dynamic environments. SUMMARY

[0003] The present application aims to use neural networks to process received signal sequences in ISAC systems to effectively achieve accurate environment perception and communication optimization. The present application deploys intelligent reflecting surfaces between users and receiving points, establishes a communication system model, uses neural networks to process the received sequences of each observation point and obtains the reflecting surface coefficients of the next observation point. At the last observation point, the neural network processes all received sequences to obtain accurate position information and downlink communication rate.

[0004] To achieve the above application purposes, the present application adopts the following technical solutions:

[0005] A method for environment perception and communication optimization assisted by intelligent reflecting surface based on deep learning, comprising the following steps:

[0006] Step 1: Multiple users send pilot signals s to the AP at the same time. The sensing process is divided into F frames, and the received signal y of all frames of the AP, the pilot signal s, and the channel matrix H generate a neural network data set.

[0007] Step 2: An adaptive reflector optimization network model is constructed, and the reflector coefficient is optimized frame by frame.

[0008] Step 3: An environment sensing network and a downlink communication optimization network are constructed, the environment sensing network outputs the scatterer point cloud estimation value, and the downlink communication optimization network outputs the downlink beamforming matrix and the reflector coefficient.

[0009] Step 4: In the downlink phase of the intelligent reflector assisted environment communication sensing model, the downlink communication parameters output by the downlink communication optimization network model are used to set the base station and the reflector beamforming, and the downlink communication is performed.

[0010] Further, the step 1 specifically includes:

[0011] Step 1.1: There is a multi-antenna AP, K single-antenna users (UEs), I intelligent reflectors (IRSs), and random scatterers (several different shaped scatterers) in the space. The transmitted signal experiences both direct and reflected paths from the UE to the AP through the scatterer and the IRS; the third and higher order reflection paths are ignored.

[0012] Step 1.2: The sensing space where the scatterer is located is called ROI, and the ROI will be evenly divided into , [H x ,W y ,L z ] are the height, width, and length of the target space, respectively, and [h x ,w y ,l z ] are the height, width, and length of the small cube, respectively; the entire ROI space is represented as x=[x1,x2,…,x n ] T ; wherein x n ∈[0,1] represents whether the nth pixel block contains a scatterer; x n =0 means that there is no scatterer in the nth pixel block; the final ROI space x n can be represented as the point cloud distribution in the specific area where the scatterer is located, and the specific distribution position and state of the scatterer are determined accordingly.

[0013] Step 1.3: The uplink communication process is divided into F frames, and the channel matrix H is divided into h scatter through the scatterer and h N-scatter without passing through the scatterer, h scatter and h N-scatterBoth are divided into direct path and scattering path. Since in the intelligent reflecting surface assisted environment communication perception model h N-scatter It is known that the uplink communication process only considers h scatter . h scatter Both the direct path and the scattering path can be represented by the cascade mode of the following formula, for example, the direct path h LOS-s The formula is h LOS-s =(h U→S ⊙v U→S )diag(x)(h S→A ⊙v S→A ), h U→S and h S→A are the path loss matrices of UE to scatterer and scatterer to AP, v U→S and v S→A are the corresponding occlusion matrices, x is the scattering coefficient vector, and ⊙ is the Hadamard product.

[0014] The formula of each frame received signal is y=h scatter s+n, s is the user transmission signal of each frame, and n is the Gaussian white noise. Each frame y, s, is spliced as and used as the features of the training set, the validation set and the test set, and the scatterer distribution point cloud and the downlink maximum rate as the label.

[0015] Further, the step 2 specifically comprises:

[0016] An adaptive reflecting surface optimization network is constructed. The network is composed of a forward LSTM sequence and a reverse LSTM sequence and an L-layer DNN. In the f frame, all y0:y f-1 , s0:s f-1 , are spliced as and input into the adaptive reflecting surface optimization network, into the forward and reverse LSTM sequences, and then the hidden layers of the two are spliced and input into the L-layer DNN, outputting the reflecting surface coefficient to achieve frame-by-frame optimization of the reflecting surface coefficient.

[0017] Further, the step 3 specifically comprises:

[0018] Step 3.1: An environment perception network based on the RES-LSTM architecture is constructed. The RES-LSTM unit is based on the LSTM unit, and the output of the LSTM unit hidden layer is adjusted to the same dimension as the input through linear transformation (Linear layer), and then added to the input as the final unit output. In0:In F obtained by the environment perception network, sequentially passes through multiple layers of RES-LSTM, bidirectional LSTM, multiple layers of LSTM, and the cell state layer Cup The DNN input as the R layer obtains the network output scatterer point cloud estimation value In the training, the mean square error is taken as the loss function, and the formula is

[0019] Step 3.2: Constructing a downlink communication optimization network composed of the same single-layer RES-LSTM and single-layer LSTM as step 4.1. The downlink communication rate network obtains In0:In F , and the cell state layer C is obtained through the single-layer RES-LSTM and single-layer LSTM down , the cell state C down The DNN input as the U layer obtains the network output downlink beamforming matrix υ and downlink reflector coefficient ψ. The optimization goal is to maximize the downlink communication rate R max , and the formula is Where P is the transmission power. P max is the AP transmission power threshold, the ReLU function is the power penalty function, and λ is a hyperparameter, is the maximum sum rate function, and the formula is Where h(ψ) i is the downlink channel matrix of the i-th user, υ i is the beamforming matrix of the i-th user, is the beamforming matrix of the i-th user, σ 2 is the Gaussian white noise, [·] * is the conjugate transpose operation. The downlink communication optimization network takes the negative function as the loss function, and the formula is L down =-R max .

[0020] The design idea of the application is that multiple users UE send pilot signals to an access point AP through a reflector in multiple observation time frames; the AP constructs a deep learning network composed of a residual network, a long short-term memory network LSTM and a multi-layer perception MLP, including an adaptive reflector optimization network that optimizes the reflector coefficient frame by frame, an environment perception network that obtains the point cloud distribution of the perception space from the received signal, and a communication optimization network that outputs the downlink communication beamforming and the reflector reflection matrix; the environment perception accuracy of the AP and the system downlink transmission rate can be effectively improved.

[0021] The beneficial effects of the application are as follows:

[0022] This method enables efficient coordination of environmental perception and communication in single-base station, multi-reflector scenarios. Multiple intelligent reflectors provide additional perspectives to obtain more comprehensive channel information. In the environmental perception section, higher-quality communication signals are achieved by continuously optimizing the phase of the intelligent reflectors, and the powerful fitting capabilities of deep learning are leveraged to recover scatterers in the perceived environment from complex communication signals. In the communication section, downlink communication parameters are directly optimized using high-quality, complete received signals, resulting in superior communication performance. Furthermore, deep learning effectively solves traditional mathematical optimization problems in the communication process. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of a scenario using an intelligent reflective surface-assisted environmental communication perception method; the space consists of one access point (AP), two intelligent reflective surfaces (RIS), four user UEs, and the ROI to be perceived.

[0024] Figure 2 Neural network structure diagram of the intelligent reflective surface-assisted environmental perception and communication optimization method;

[0025] Figure 3 The present invention compares the perceptual error map of the scatterer point cloud distribution with the comparative scheme under different signal-to-noise ratios. The first comparative scheme is a simplified scheme of the present invention, which uses the environmental perception network of the present invention for the random IRS reflection coefficient. The second comparative scheme is a generalized approximate message passing method.

[0026] Figure 4 The present invention compares the downlink maximum communication rate diagram with the comparative scheme under different signal-to-noise ratios; the comparative scheme one uses a deep learning method for random IRS reflection coefficients; the comparative scheme two uses randomly generated AP downlink beamforming matrix and IRS reflection coefficient matrix. Detailed Implementation

[0027] The present invention will be further described below with reference to the accompanying drawings.

[0028] The signal transmission process of the deep learning-based environment perception and communication optimization method assisted by IRS is as follows: Figure 1 As shown. Before localization, the area where the scatterer is located is first divided into small cubes and then into 8 time periods. The received signal y, pilot signal s, and channel matrix H of the first frame are acquired to generate a neural network dataset. The data is input into the neural network to adaptively adjust the reflector coefficient, and the point cloud distribution, downlink reflector coefficient, and beamforming matrix of the specified area are output in the last time period, as shown. Figure 2 As shown. The specific implementation process is as follows:

[0029] Step 1) Multiple users send pilot signals s to the AP simultaneously; divide the sensing process into F frames, obtain the received signal y of all frames of the AP, the pilot signal s, and the channel matrix H to generate a neural network data set;

[0030] 1.1) There is one multi-antenna AP, four single-antenna users (UEs), two intelligent reflecting surfaces (IRSs), and random scatterers in the space. The transmitted signal experiences both direct and reflected paths from the UE to the AP through the scatterer and the IRS. Paths with three or more reflections are ignored.

[0031] 1.2) The sensing space where the scatterer is located is called ROI, which will be evenly divided into [H x ,W y ,L z ] are the height, width, and length of the target space, respectively, and [h x ,w y ,l z ] are the height, width, and length of the small cube, respectively. The entire ROI space is represented as x = [x1, x2,..., x n ] T ; where x n ∈ [0, 1] represents whether the nth pixel block contains a scatterer or not; x n = 0 means that there is no scatterer in the pixel block. The final ROI space x n can be represented as the point cloud distribution in the specific area where the scatterer is located, and the specific distribution position and state of the scatterer are determined accordingly.

[0032] 1.3) The uplink communication process is divided into F frames, and the channel matrix H is divided into h scatter that passes through the scatterer and h N-scatter that does not pass through the scatterer. Both h scatter and h N-scatter are divided into direct and scattering paths. Since h N-scatter is known in the intelligent reflecting surface assisted environment communication sensing model, the uplink communication process only considers h scatter . h scatter contains both direct and scattering paths and can be represented by the following cascade,

[0033] The direct path h LOS-s formula is h LOS-s = (h U→S ⊙v U→S )diag(x)(h S→A ⊙v S→A ), h U→S and h S→A are the path loss matrices from the UE to the scatterer and from the scatterer to the AP, v U→S and vS→A is the Hadamard product.

[0034] The scattering path is:

[0035]

[0036] h S→I,m represents the channel matrix from the scatterer to the mth IRS, h I,m→A represents the channel matrix from the IRS to the AP, h U→I,m represents the channel matrix from the UE to the IRS, h I,m→S represents the channel matrix from the scatterer to the UE, v S→I,m represents the occlusion matrix from the mth IRS to the scatterer, v I,m→S represents the occlusion matrix from the scatterer to the mth IRS, Θ m is the reflection phase coefficient matrix.

[0037] The formula of each frame received signal is y = h scatter s + n, s is the user transmission signal of each frame, and n is Gaussian white noise. Each frame y, s, is spliced into and used as the features of the training set, the validation set and the test set, the scatterer distribution point cloud and the downlink maximum rate as labels.

[0038] Step 2) Construct an adaptive reflecting surface optimization network model to optimize the reflecting surface coefficient frame by frame;

[0039] The network is composed of a forward LSTM sequence and a reverse LSTM sequence and an L-layer DNN. In the fth frame, all y0:y f-1 , s0:s f-1 , is spliced into is input into the forward and reverse LSTM sequences, and then the hidden layers of the two are spliced and input into the L-layer DNN, and the reflecting surface coefficient is output, so as to optimize the reflecting surface coefficient frame by frame.

[0040] Step 3) Construct an environment perception network model and a downlink communication optimization network model. The environment perception network model outputs the scatterer point cloud estimation value, and the downlink communication optimization network model outputs the downlink beamforming matrix and the reflecting surface coefficient;

[0041] 3.1) Constructing the environment perception network based on RES-LSTM architecture, the RES-LSTM unit is based on the LSTM unit, the LSTM unit hidden layer output is adjusted to the same dimension as the input through linear transformation (Linear layer), and then added to the input as the final unit output. The environment perception network obtains In0:In F , sequentially passing through four layers of RES-LSTM, bidirectional LSTM, four layers of LSTM, and multi-layer LSTM output cell state layer C up The DNN input of R layer obtains the network output scatterer point cloud estimation value In the training, the mean square error is taken as the loss function, and the formula is

[0042] 3.2) Constructing a downlink communication optimization network, which is composed of the same single-layer RES-LSTM and single-layer LSTM as step 3.1. The downlink communication rate network obtains In0:In F , and the cell state layer C is obtained by passing through the single-layer RES-LSTM and the single-layer LSTM down , the cell state C down The DNN input of U layer obtains the network output downlink beamforming matrix υ and downlink reflector coefficient ψ. The optimization goal is to maximize the downlink communication rate R max , the formula is Where P is the transmit power. P max is the AP transmit power threshold, the ReLU function is the power penalty function, and λ is a hyperparameter, is the maximum sum rate function, and the formula is Where h(ψ) i is the downlink channel matrix of the i-th user, υ i is the beamforming matrix of the i-th user, is the beamforming matrix of the i-th user, σ 2 is the Gaussian white noise, [·] * is the conjugate transpose operation. The downlink communication optimization network takes the negative function as the loss function, and the formula is L down =-R max .

[0043] As shown in Figure 3 , as the signal-to-noise ratio increases, the neural network method used in the application has a significant performance improvement over the traditional generalized approximate message passing of scheme two, and at high signal-to-noise ratio, the application can achieve lower error and restore more accurate scatterers than the random phase deep learning method of scheme one and the generalized approximate message passing method of scheme two. As shown in Figure 4As shown, under different signal-to-noise ratios, the neural network method used in the application has a significant performance improvement compared with the traditional random generation method, and the first scheme is a simple scheme of the application, and in the low signal-to-noise ratio, the application improves the channel information quality by optimizing the phase of the intelligent reflecting surface, thereby further optimizing the downlink reflecting surface phase and the beamforming matrix, and realizing higher communication performance; under the condition of high signal-to-noise ratio, the neural network receives high-quality signals, and compared with the random generation method of the second scheme, the communication performance can be significantly improved.

Claims

1. A method for environmental perception and communication optimization based on deep learning and assisted by intelligent reflective surfaces, characterized in that, Includes the following steps: Step 1: Multiple users simultaneously send pilot signals s to the AP; the sensing process is divided into F frames, and the received signals y of all frames of the AP, pilot signals s, and channel matrix H are obtained to generate a neural network dataset; Step 2: Construct an adaptive reflector optimization network model and optimize the reflector coefficients frame by frame; Step 3: Construct an environmental perception network model and a downlink communication optimization network model. The environmental perception network model outputs the estimated point cloud value of the scatterer, and the downlink communication optimization network model outputs the downlink beamforming matrix and reflector coefficients. Step 4: In the downlink phase of the intelligent reflector-assisted environmental communication sensing model, downlink communication parameters output by the downlink communication optimization network model are set to enable downlink communication by configuring the base station and reflector beamforming. Step 1 specifically includes: Step 1.1: There is a multi-antenna AP, K single-antenna user UEs, I intelligent reflector IRS and several scatterers of different shapes in space; the transmitted signal travels from the UE through two paths, direct connection and reflection, through the scatterers and IRS to reach the AP, where reflection paths of three or more times are ignored; Step 1.2: The sensing space where the scatterer is located is called the Region of Interest (ROI). The ROI will be uniformly divided into... One, [H] x W y ,L z [h] represents the height, width, and length of the target space, respectively. x ,w y ,l z [x1, x2, ..., xn] represents the height, width, and length of the small cube, respectively; the entire ROI space is represented as x = [x1, x2, ..., xn]. n ] T ;where x n ∈[0,1] represents whether the nth pixel block contains a scatterer; x n =0 indicates that there is no scatterer in the nth pixel block; the final ROI space x n This is represented as the point cloud distribution in a specific region where the scatterer is located, and the specific distribution location and state of the scatterer are determined accordingly. Step 1.3: Divide the uplink communication process into F frames, and divide the channel matrix H into h frames that have passed through the scatterer. scatter and h without passing through the scatterer N-scatter Composition, where h scatter and h N-scatter Both are divided into direct connection paths and scattering paths; Because in the intelligent reflector-assisted environmental communication sensing model h N-scatter Given that h is known, the uplink communication process only needs to consider h. scatter h scatter Both direct paths and scattering paths can be represented by the following cascade method, where the direct path h LOS-s The formula is: h LOS-s =(h U→S ⊙v U→S )diag(x)(h S→A ⊙v S→A ) h U→S and h S→A Let v be the path loss matrix from UE to scatterer and from scatterer to AP. U→S and v S→A Here is the corresponding occlusion matrix, x is the scattering coefficient vector, and ⊙ is the Hadamard product; The scattering path is: h S→I,m h represents the channel matrix from the scatterer to the m-th IRS. I,m→A h represents the channel matrix from IRS to AP. U→I,m h represents the channel matrix from UE to IRS. I,m→S The channel matrix from the scatterer to the UE, v S→I,m v represents the blocking matrix from the m-th IRS to the scatterer. I,m→S Θ represents the shading matrix from the scatterer to the m-th IRS. m This is the reflection phase coefficient matrix; The formula for the received signal in each frame is y = h scatter s+n, where s is the user-transmitted signal for each frame and n is Gaussian white noise; for each frame y, s, spliced ​​as Features, scatterer distribution point cloud, downlink maximization, and rate are used as labels and are used as training, validation, and test sets.

2. The method for environmental perception and communication optimization based on deep learning and assisted by intelligent reflective surfaces according to claim 1, characterized in that, Step 2 specifically includes: An adaptive reflector optimization network is constructed, which consists of a forward LSTM sequence, an inverse LSTM sequence, and an L-layer DNN. In the f-th frame, all the previous f-1 frames are... spliced ​​as The input to the adaptive reflector optimization network is fed into forward and inverse LSTM sequences, and then the hidden layers of these sequences are concatenated and fed into an L-layer DNN to output the reflector coefficients. Achieve frame-by-frame optimization of the reflective surface coefficient.

3. The method for environmental perception and communication optimization based on deep learning and assisted by intelligent reflective surfaces according to claim 1, characterized in that, Step 3 specifically includes: Step 3.1: Construct an environment perception network based on the RES-LSTM architecture. The RES-LSTM unit is based on the LSTM unit, and its LSTM unit hidden layer output is linearly transformed to be the same dimension as the input, and then added to the input as the final unit output; the environment perception network obtains In0:In F The cell state layer C is sequentially processed through multiple layers of RES-LSTM, bidirectional LSTM, multiple layers of LSTM, and multiple layers of LSTM. up The DNN input to the R layer is used to obtain the network output scatterer point cloud estimate. During training, the mean squared error is used as the loss function, and the formula is: Step 3.2 Construct a downlink communication optimization network; this network consists of a single-layer RES-LSTM and a single-layer LSTM; the downlink communication rate network obtains In0:In F Cell state layer C was obtained by passing through a single layer of RES-LSTM and a single layer of LSTM. down Cell state C down The DNN inputs to the U layer yield the downlink beamforming matrix υ and the downlink reflector coefficient ψ of the network output; The optimization objective is to maximize the downlink communication rate R. max The formula is: Where P is the transmission power, P max The AP transmit power threshold is defined by the ReLU function, which is the power penalty function, and λ is a hyperparameter. The maximum sum rate function is given by the following formula: Where h(ψ) i Let υ be the downlink channel matrix of the i-th user. i Let be the beamforming matrix for the i-th user. For the beamforming matrix excluding the i-th user, σ 2 It is Gaussian white noise, [·] * This is the conjugate transpose operation; The downlink communication optimization network uses the negative function as the loss function, as shown in the formula: L down =-R max 。

Images