Self-adaptive reconfigurable intelligent surface positioning method based on long short-term memory network prediction

Abstract

The invention relates to the technical field of wireless communication, and discloses a self-adaptive reconfigurable intelligent surface positioning method based on long short-term memory network prediction, which comprises the following steps: step 1, constructing a self-adaptive reconfigurable intelligent surface positioning system model consisting of three entities including a reconfigurable intelligent surface RIS, a sensing target and an environment clutter; 2, explaining a passive reflection mechanism of the adaptive reconfigurable intelligent surface positioning system model from channel modeling and off-line and on-line training, and inferring angular domain information of a sensing target near the reconfigurable intelligent surface RIS by analyzing echo signals received on a direct link and a reconfigurable intelligent surface RIS reflection echo link; step 3, adaptively updating the reflection coefficient of the reconfigurable intelligent surface RIS in a codebook-free mode based on historical measurement data by using a learning positioning technology of a long short-term memory LSTM network; according to the invention, the product distance path loss in a reflection echo link is greatly reduced.

Description

Technical Field

[0001] The present invention relates to the field of wireless communication technology, and in particular to an adaptive reconfigurable intelligent surface positioning method and device based on long short-term memory network prediction. Background Art

[0002] In the context of the rapid development of wireless communication technology, reconfigurable smart surface technology has become a research field with great potential due to its unique capabilities; RIS technology uses an artificial structure composed of a large number of programmable reflection units to improve the wireless channel environment and enhance communication performance. RIS has the characteristics of dynamic regulation. By adjusting the reflection coefficient, phase, frequency and polarization parameters of the reflection unit, it can realize intelligent regulation of the incident electromagnetic wave. This flexibility enables RIS to adapt to different communication scenarios and optimize signal propagation and reception. Its key advantages include the ability of passive reflection beamforming. By reasonably designing the reflection coefficient, RIS can passively adjust the direction of the reflected wave, thereby constructing a more favorable electromagnetic transmission environment. RIS-assisted communication technology has become one of the most promising technologies for improving the performance of communication systems with its low hardware complexity, high energy efficiency, high spectrum efficiency and high security. The emergence of this technology has not only opened up a new path for improving the performance of wireless communications, but also inspired the exploration of its new applications in wireless networks. Summary of the invention

[0003] The purpose of the present invention is to solve the above problems and to design an adaptive reconfigurable intelligent surface positioning method based on long short-term memory network prediction.

[0004] The present invention provides an adaptive reconfigurable intelligent surface positioning method based on long short-term memory network prediction, and the adaptive reconfigurable intelligent surface positioning method based on long short-term memory network prediction comprises the following steps:

[0005] Step 1: Build an adaptive reconfigurable intelligent surface positioning system model, which consists of three entities, including reconfigurable intelligent surface RIS, sensing target and environmental clutter;

[0006] Step 2: The passive reflection mechanism of the adaptive reconfigurable intelligent surface positioning system model is explained from the perspective of channel modeling and offline and online training, and the angular domain information of the perceived target near the reconfigurable intelligent surface RIS is inferred by analyzing the echo signals received on the direct link and the reconfigurable intelligent surface RIS reflection echo link;

[0007] Step 3: Using the learning positioning technology of the long short-term memory (LSTM) network, the reflection coefficient of the reconfigurable smart surface (RIS) is adaptively updated in a codebook-free manner based on historical measurement data.

[0008] Optionally, in the first implementation of the present invention, the reconfigurable smart surface RIS in step 1 is composed of N=N h ×N v The block reflection unit is composed of a unit spacing of d I , where N h and N h are the number of RIS units in the horizontal and vertical directions, respectively;

[0009] The reflective unit is illuminated by the detection signal emitted by the reconfigurable intelligent surface RIS controller using an omnidirectional antenna within a short distance, and forms a reflected beam to sense the target in its vicinity;

[0010] M sensor receivers are also installed on the reconfigurable intelligent surface RIS to receive the echo signals of the perceived target, and then obtain the azimuth and pitch angle of the target through signal processing. The block sensing receiving units are placed in the horizontal and vertical directions, with a unit spacing of d. s , wherein the echo signal includes two source paths, namely a single-hop direct echo link and a double-hop RIS reflection echo link.

[0011] Optionally, in a second implementation of the present invention, a single-hop direct echo link is RIS controller→sensing target→RIS sensing unit, and a double-hop RIS reflection echo link is RIS controller→RIS reflection unit→sensing target→RIS sensing unit.

[0012] Optionally, in a third implementation of the present invention, in step 2, modeling is performed for the single-hop direct echo link and the double-hop RIS reflection echo link respectively, and the steering vector function Defined as:

[0013]

[0014] in, represents the constant phase difference between two blocks of adjacent unit observations, (assumed to be an even number) represents the ULA size;

[0015] The single-hop direct echo link is modeled as:

[0016] H d =α d r(θ RT,h ,θ RT,v )

[0017] In the formula, α d =β d G d is the direct path complex gain, β d ∽CN(0,1) is the negative gain of the small-scale channel, represents the signal attenuation during a single hop, λ is the carrier wavelength, κ represents the radar cross-section, and d CT represents the distance from the RIS controller to the sensing target, d RT Indicates the distance from RIS to the perceived target;

[0018] r(·) represents the receiving response vector of the RIS sensor and is defined as:

[0019]

[0020] in, θ RT,h and θ RT,v Horizontal azimuth and vertical pitch angle from RIS to the perceived target;

[0021] The channel of the double-hop RIS reflection echo link at the snapshot can be expressed as:

[0022]

[0023] Assume that p RIS =[x RIS ,y RIS ,z RIS ] represents the position of RIS, and p = [x, y, z] represents the position of the perceived target. The position relationship satisfies the following relationship:

[0024] sin(θ RT,h )cos(θ RT,v )=(yy RIS ) / d RT

[0025] sin(θ RT,v )=(zz RIS ) / d RT .

[0026] Optionally, in a fourth implementation of the present invention, since the distance between the target and the RIS under far-field conditions is much greater than the wavelength and the size of the reflecting surface, in this case, the electromagnetic wave is a plane wave, the arrival angle AOA from the target to the RIS is the same as the departure angle AOD from the RIS to the target, and the echo link is modeled as:

[0027] H RTS =α r r(θ RT,h ,θ RT,v ) T (θ RT,h ,θ RT,v )

[0028] Among them, α r =β r G r , a(·) is the response vector function of RIS, defined as:

[0029]

[0030] The channel between the RIS controller and the RIS reflection unit can be modeled as:

[0031] H CR =α CR a(θ CT,h ,θ CT,v )

[0032] in, is the path complex gain from the controller to the reflector, d CR is the distance between the controller and the reflection unit, θ CT,h ∈[-π / 2,π / 2] and θ CT,v ∈[-π / 2,π / 2] represents the physical azimuth and vertical arrival angle at RIS, respectively;

[0033] The reflection coefficient of RIS is controlled by the RIS controller. Assume that the reflection coefficient of RIS is:

[0034]

[0035] in, Represents each snapshot The RIS reflection vector at , each coefficient as the phase shift of the nth element of RIS.

[0036] Optionally, in a fifth implementation of the present invention, during the offline online training in step 2, it is assumed that the environmental state remains stable on a time scale much larger than the real-time perception operation, and there is no perception target in the environment, and the signal vector received by the RIS sensor is:

[0037]

[0038] Among them, h r,l (t) and h d,l denote the reflected echo link and the direct echo link associated with each clutter l∈1,2,…,L at each snapshot t, respectively, and H CS represents the channel between the RIS controller and the RIS sensor. When x=1, all potential impacts of environmental clutter on the received signal are obtained.

[0039] Optionally, in a sixth implementation of the present invention, during the online evaluation in step 2, for each time snapshot t, the RIS reflection vector Based on the measured offline channel information, the unwanted interference from the RIS controller and all environmental clutter are removed from the received signal to obtain the received signal:

[0040]

[0041] Where x(t) is the signal sent by the RIS controller at snapshot t, assuming x(t) = 1, w(t) ∽ CN(0,2σ 2 I) represents the zero-mean additive white Gaussian noise vector at snapshot t, with normalized noise power of 2σ 2 .

[0042] Optionally, in a seventh implementation of the present invention, in each time frame t in step 3, an LSTM unit is used, and the hidden state vector s is updated with the new feature as input. t and the cell state vector c t , according to the constraints of the receiver detection hardware new features π t It can be the received signal strength |y(t)| 2 , or the real and imaginary parts of the received pilot [R(y(t)),I(y(t))], using To represent the product between elements, the update rule of the state vector is as follows:

[0043]

[0044] f t =sigmoid(A f π t +U f s t-1 +b f )

[0045] i t =sigmoid(A i π t +U i s t-1 +b i )

[0046] o t =sigmoid(A o π t +U o s t-1 +b o )

[0047] Among them, {A f ,A i ,A o ,A c ,U f ,U i ,Uo ,U c} and {b f ,b i ,b o ,b c} are the trainable weights and bias values ​​in the LSTM network respectively;

[0048] Initialize the LSTM network and set s -1 =c -1 =0, and y(0)=1, the hidden state vector s t Used as input to a fully connected neural network consisting of L layers, the RIS configuration for the next time frame t+1 is:

[0049]

[0050] where β l , l∈{1,…,L} is the activation function Relu(·) of the lth layer, and is a set of trainable weights and bias values;

[0051] Contains the real and imaginary parts of the RIS reflection coefficient:

[0052]

[0053] To enforce the unit norm constraint on each RIS cell, the output is element-wise normalized:

[0054]

[0055] After T time frames, the final estimated position of the perceived target is obtained through a fully connected neural network based on the final unit state:

[0056]

[0057] Among them, {A′, b′} are the sets of trainable weights and bias values, respectively, and σ r is the activation function of the rth layer. For the first R-1 layers, the activation function is set to the ReLU function, that is, σ r =relu(·), r=1,…,R-1.

[0058] Compared with the prior art, the present invention adopts the above technical solution, and its significant advantages are:

[0059] (1) The designed adaptive RIS perception system, since the detection signal is sent by the RIS controller (rather than the BS) located near the RIS reflection unit, this configuration greatly reduces the product distance path loss in the reflection echo link compared to the traditional fully passive or semi-passive RIS assisted perception system.

[0060] (2) The LSTM network architecture used can capture the potential temporal correlation between the measurement data and automatically construct a scalable information vector (i.e., hidden state) based on the measurement. These state vectors can then be mapped to the RIS configuration of the next time frame by a deep neural network in a codebook-free form. After all the measurement data are collected, the final deep neural network is responsible for mapping the state of the LSTM unit to the estimated perceived target position. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the following detailed description of the preferred embodiment.The drawings are only for the purpose of illustrating the preferred embodiments and are not to be construed as limiting the invention.

[0062] Figure 1 The adaptive reconfigurable intelligent surface positioning system model provided by the present invention;

[0063] Figure 2 A neural network architecture diagram provided by the present invention;

[0064] Figure 3 A relationship diagram between positioning accuracy and signal-to-noise ratio when the input provided by the embodiment of the present invention is a pilot symbol;

[0065] Figure 4 A graph of the relationship between positioning accuracy and signal-to-noise ratio when the input provided for the embodiment of the present invention is the received signal strength;

[0066] Figure 5 A diagram showing the effect of the size of the RIS on the positioning accuracy provided by an embodiment of the present invention;

[0067] Figure 6 A received signal strength distribution diagram provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0068] The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, device, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0069] See also Figure 1-2 As shown, the adaptive reconfigurable intelligent surface positioning method based on long short-term memory network prediction specifically includes the following steps:

[0070] Step 1: Build an adaptive reconfigurable intelligent surface positioning system model, which consists of three entities, including reconfigurable intelligent surface RIS, sensing target and environmental clutter;

[0071] In this step, the reconfigurable smart surface RIS is composed of N=N h ×N v The block reflection unit is composed of a unit spacing of d I , where N h and N h are the number of RIS units in the horizontal and vertical directions respectively. The reflective unit is illuminated by the detection signal emitted by the RIS controller using the omnidirectional antenna within a short distance, and forms a reflected beam to sense the target nearby. In addition, M sensor receivers are installed on the RIS to receive the echo signals of the sensed target, and then the azimuth and pitch angle of the target can be accurately estimated through advanced signal processing technology, thereby achieving accurate positioning of the target position. The block sensing receiving units are placed in the horizontal and vertical directions, with a unit spacing of d. s The received echo signal mainly comes from two paths: 1) single-hop direct echo link, i.e. RIS controller → sensing target → RIS sensor unit; 2) double-hop RIS reflection echo link, i.e. RIS controller → RIS reflection unit → sensing target → RIS sensor unit.

[0072] Step 2: The passive reflection mechanism of the adaptive reconfigurable intelligent surface positioning system model is explained from the perspective of channel modeling and offline and online training, and the angular domain information of the perceived target near the reconfigurable intelligent surface RIS is inferred by analyzing the echo signals received on the direct link and the reconfigurable intelligent surface RIS reflection echo link;

[0073] In this step, (1) RIS dual-path channel modeling

[0074] In the designed RIS adaptive system, the RIS controller continuously sends the detection signal on snapshots. The block-fading model is used to describe the channel characteristics. The model assumes that the channel remains unchanged over multiple time frames within a coherent period, while the channel changes independently in the next coherent period. During the transmission of each snapshot, the sensor can not only receive the direct signal from the RIS controller, but also capture the signal reflected from the target object and other environmental clutter. This process is not affected by whether the signal passes through RIS reflection. Specifically, the target reflection signal is mainly received through the following two paths: 1) a single-hop direct echo link, i.e., RIS controller → sensing target → RIS sensing unit; 2) a double-hop RIS reflection echo link, i.e., RIS controller → RIS reflection unit → sensing target → RIS sensing unit. These two paths are modeled as follows.

[0075] Steering vector function Defined as:

[0076]

[0077] in, represents the constant phase difference between two blocks of adjacent unit observations, (assumed to be an even number) represents the ULA size. Then, according to the far-field line-of-sight channel model, the single-hop direct echo link (i.e., RIS controller → sensing target → RIS sensing unit) is modeled as:

[0078] H d =α d r(θ RT,h ,θ RT,v )

[0079] In the formula, α d =β d G d is the direct path complex gain, β d ∽CN(0,1) is the negative gain of the small-scale channel, represents the signal attenuation during a single hop. Where λ is the carrier wavelength, κ represents the radar cross section (RCS), and d CT represents the distance from the RIS controller to the sensing target, d RT represents the distance from RIS to the perceived target. r(·) represents the receiving response vector of the RIS sensor, which is defined as:

[0080]

[0081] in, θ RT,h and θ RT,v The horizontal azimuth and vertical pitch angles from the RIS to the sensing target. Since the distance between the target and the RIS under far-field conditions is much larger than the wavelength and the size of the reflecting surface, the electromagnetic wave can be approximated as a plane wave. Therefore, the angle of arrival (AOA) from the target to the RIS is the same as the angle of departure (AOD) from the RIS to the target. Therefore, the echo link (RIS reflection unit → sensing target → RIS sensing unit) can be modeled as

[0082] H RTS =α r r(θ RT,h ,θ RT,v ) T (θ RT,h ,θ RT,v )

[0083] where α r =β r G r , a(·) is the response vector function of RIS, defined as

[0084]

[0085] On the other hand, the channel between the RIS controller and the RIS reflection unit can be modeled as

[0086] H CR =α CR a(θ CT,h ,θ CT,v )

[0087] in, is the path complex gain from the controller to the reflector, d CR is the distance between the controller and the reflection unit, θ CT,h ∈[-π / 2,π / 2] and θ CT,v ∈[-π / 2,π / 2] represents the physical azimuth and vertical arrival angle at RIS, respectively.

[0088] The reflection coefficient of RIS is controlled by the RIS controller. Assume that the reflection coefficient of RIS is:

[0089]

[0090] in, Represents each snapshot The RIS reflection vector at as the phase shift of the nth element of RIS.

[0091] In summary, the channel of the double-hop RIS reflection echo link (RIS controller → RIS reflection unit → sensing target → RIS sensing unit) at the snapshot can be expressed as:

[0092]

[0093] Assume that p RIS =[x RIS ,y RIS ,z RIS ] represents the position of RIS, and p = [x, y, z] represents the position of the perceived target, then the position relationship satisfies the following relationship:

[0094] sin(θ RT,h )cos(θ RT,v )=(yy RIS ) / d RT

[0095] sin(θ RT,v )=(zz RIS ) / d RT

[0096] (2) Offline training

[0097] Given that the clutter in the environment will introduce unnecessary information into the received signal, it is necessary to measure the impact of environmental clutter on the system during the offline stage. Specifically, during offline training, it is assumed that the environmental state remains stable on a time scale much larger than the real-time perception operation, and there is no perception target in the environment. During this period, the signal vector received by the RIS sensor is

[0098]

[0099] Among them, h r,l (t) and h d,l denote the reflected echo link and the direct echo link associated with each clutter l∈1,2,…,L at each snapshot t, respectively, and H CS represents the channel from the RIS controller to the RIS sensor. When x=1, all potential effects of environmental clutter on the received signal can be obtained. It is worth mentioning that the detection signal sent by the RIS controller will also be directly received by the RIS sensor, which is undesirable. Therefore, this part of the impact is also measured in the offline stage.

[0100] (3) Online evaluation

[0101] During the online evaluation, for each RIS reflection vector in time snapshot t Based on the pre-measured offline channel information, the unwanted interference from the RIS controller and all environmental clutter are removed from the received signal, so the received signal can be obtained:

[0102]

[0103] Where x(t) is the signal sent by the RIS controller at snapshot t. To simplify the calculation, x(t) = 1 is assumed here. w(t)∽CN(0,2σ 2 I) represents the zero-mean additive white Gaussian noise vector at snapshot t, where the normalized noise power is 2σ 2 .

[0104] Step 3: Using the learning positioning technology of the long short-term memory (LSTM) network, the reflection coefficient of the reconfigurable smart surface (RIS) is adaptively updated in a codebook-free manner based on historical measurement data.

[0105] In this step, the active RIS configuration scheme proposed in this way can significantly reduce the positioning error.

[0106] (1) Problem Construction

[0107] In t time frames, the RIS controller designs the RIS reflection coefficient configuration in the next time frame t+1 based on the observation data collected so far:

[0108]

[0109] It is worth noting that before any measurement begins (i.e., t < 0), since there are no prior observations available, the first set of RIS configurations is through The function receives an empty collection. As input, no historical observation data is available at the initial time, and then a fixed initial RIS reflection coefficient configuration is generated.

[0110] Estimated perceived target position It can also be written as a function of all T historical measurements:

[0111]

[0112] In summary, assuming there is a distortion metric E(·), the localization parameter estimation problem can be formulated as:

[0113]

[0114] It is difficult to solve this optimization problem because it is a function expression and F(·). In order to make this problem easier to solve, a common approach is to select the RIS configuration from a predefined set of RIS patterns (i.e., codebook) based on a heuristic method. However, this method has certain limitations because the configuration of the RIS reflection coefficient designed in this way is subject to certain constraints. Therefore, we choose to use a neural network to parameterize and F(·), thereby realizing the adaptive configuration of the RIS reflection coefficient.

[0115] (2) Solution based on neural network architecture

[0116] Since the adaptive sensing function G t The input dimension of (·) increases with time t (if no preprocessing is performed), so it is difficult to design the RIS reflection coefficient configuration at t+1 using the entire historical measurement value. To solve this problem, the traditional approach is to abstract the historical observations into a low-dimensional state variable, such as the posterior distribution p, which can be regarded as a sufficient statistic for the estimation problem. Based on the posterior distribution, the RIS configuration at time t can be derived and the final estimate at time T However, computing the posterior distribution requires knowledge of the system model and parameter distribution, information that may not always be readily available in real-world problems. Even if this information is available, efficient computation of the posterior distribution is usually only applicable to a specific class of system models. In addition, when the system parameters are continuous variables, the posterior distribution is presented in the form of a probability distribution function, which requires discretization and results in inaccurate calculations.

[0117] Therefore, the key issue in neural network design is how to extract useful information from historical measurements and map it to an information vector of fixed dimension.

[0118] A LSTM network is used, which can automatically construct a fixed-dimensional information vector from the existing measurement data and extract temporal features and long-term dependencies from the time series input. More specifically, since the observations are continuous and temporally correlated, an RNN based on the LSTM architecture is used to capture the temporal correlation between measurements. At each time step, the LSTM also generates a hidden state variable for designing the corresponding RIS reflection coefficient. In the final stage, the position parameters are estimated from the final unit state variables.

[0119] A Long Short-Term Memory (LSTM) network is used, which is well suited for tracking long-term dependencies in input sequences. In each time frame t, an LSTM unit is used, which takes the new features as input to update the hidden state vector s t and the cell state vector c t . New feature π tIt can be the received signal strength |y(t)| 2 , or the real and imaginary parts of the received pilot [R(y(t)),I(y(t))], depending on the constraints of the receiver detection hardware. To represent the product between elements, the update rule of the state vector is as follows:

[0120]

[0121] f t =sigmoid(A f π t +U f s t-1 +b f )

[0122] i t =sigmoid(A i π t +U i s t-1 +b i )

[0123] o t =sigmoid(A o π t +U o s t-1 +b o )

[0124] Where {A f ,A i ,A o ,A c ,U f ,U i ,U o ,U c} and {b f ,b i ,b o ,b c} are the trainable weights and bias values ​​in the LSTM network respectively. In addition, the activation vector f of the forget gate t , the activation vector i of the input / update gate t and the activation vector o of the output gate t is the intermediate vector generated inside the LSTM unit and used to update s t and c t According to the conventional initialization of LSTM network, set s -1 =c -1 =0, and y(0)=1.

[0125] The design will be a hidden state vector s tis used as input into a fully connected neural network consisting of L layers to design the RIS configuration for the next time frame t+1:

[0126]

[0127] where β l , l∈{1,…,L} is the activation function Relu(·) of the lth layer, and is a set of trainable weights and bias values. Contains the real and imaginary parts of the RIS reflection coefficient:

[0128]

[0129] To ensure the output has the correct dimensions, i.e. The weight of the last layer is A L and bias b L The dimensions of need to be designed to match accordingly. To enforce the unit modulus constraint on each RIS unit, the output is element-wise normalized:

[0130]

[0131] After T time frames, the final estimated position of the perceived target is obtained through a fully connected neural network based on the final unit state:

[0132]

[0133] where {A′, b′} are sets of trainable weights and bias values, respectively, and σ r is the activation function of the rth layer. For the first R-1 layers, the activation function is set to the ReLU function, that is, σ r =relu(·), r=1,…,R-1, and for the last layer, an appropriate normalized activation function is chosen to ensure that the final output satisfies any potential constraints.

[0134] To train the LSTM network, the Adam optimizer is used to minimize the estimated position The average mean square error (MSE) between the true position and the original position, so the distortion metric E(·) can be written as:

[0135]

[0136] Since the distortion metric function used focuses on evaluating the positioning error at the end of the entire measurement cycle (i.e., the final time frame T) rather than evaluating the error at each individual time point, it allows the neural network to have more flexibility to adjust and optimize the RIS reflection coefficient configuration throughout the entire measurement cycle (all time points from the beginning to the end), rather than being forced to minimize the error as much as possible at each time point.

[0137] Example 1

[0138] The embodiments of the present invention are described in detail as follows. The system simulation adopts the Pytorch framework of Python, and the LSTM network of the adaptive RIS perception system is implemented using the parameters in the table.

[0139]

[0140] L=4 is set, and the proposed deep active learning framework is implemented on TensorFlow and Keras platforms. The model samples a total of 2,048,000 training data in 2,000 training cycles. The positioning performance of the proposed algorithm is compared with the following baseline schemes:

[0141] Deep Neural Network (DNN) based on random or learned RIS configuration: In this scheme, the RIS configuration is non-adaptive. There are two possible configurations: i) the RIS configuration is randomly selected, or ii) the RIS configuration is learned from the training data (i.e., channel statistics) but remains unchanged during the test phase. For implementation in Tensorflow, the reflection coefficient is set as a trainable parameter during the DNN training phase, so that the reflection coefficient can be dynamically adjusted according to the feedback during the training process to adapt to the changes in the channel and optimize the localization performance. The structural dimension of the DNN is set to [200, 200, 200, 3].

[0142] Fingerprint positioning based on random RIS configuration: In this scheme, the RIS configuration is non-adaptive and random. In the 30m×70m area, each 1m×1m cell block is associated with a received signal strength (RSS) vector as its unique fingerprint in the form of [|y(1)| 2 ,|y(2)| 2 ,…,|y(T)| 2 ] and are stored in an offline database. The weighted nearest neighbor (wKNN) classifier is used in the matching process between online and offline measurements, and κ is set here to 5. This method determines the user's location by comparing the RSS fingerprint of the online measurement with the RSS fingerprint stored in the database.

[0143] Cramer-Rao Bound (CRLB) based on Gradient Descent (GD): This is an adaptive positioning method that does not rely on machine learning techniques and is based on estimation theory. The core idea is to use the measurement data received at each time step to minimize the CRLB, thereby designing an effective active perception strategy. In each time frame, the maximum a posteriori (MAP) estimator is first used to estimate the position of the perceived target based on the received measurement data. Then, this estimated position is used to update the CRLB of the position error, that is, an approximate CRLB value is calculated by replacing the true position with the estimated position. Finally, the GD algorithm calculates the gradient of the objective function (i.e., CRLB) with respect to the RIS reflection coefficient and updates the reflection coefficient in the direction of this gradient to gradually reduce the CRLB.

[0144] Figure 3 and Figure 4 is a relationship diagram between positioning performance (with mean square error as an indicator) and signal-to-noise ratio (SNR) in an embodiment of the present invention. Whether based on the real and imaginary parts of the received pilot signal [R(y(t)), I(y(t))], or based on the received signal strength (RSS) value |y(t)| 2 As input, the proposed adaptive RIS design shows better positioning accuracy under various signal-to-noise ratio conditions compared to other non-adaptive RIS design baselines. This result verifies that the proposed framework can effectively combine current and historical measurement data to design a RIS configuration that is more suitable for future time frames, thereby minimizing positioning errors.

[0145] exist Figure 3 In , although the gradient descent method was used to optimize CRLB, this CRLB-based method did not become the optimal RIS design strategy for reducing MSE. This is mainly because the CRLB-based method adopts a greedy strategy when designing the problem, that is, it tries to optimize the performance by independently designing in each time frame. This step-by-step optimization method ignores the overall optimal configuration across time frames, resulting in the inability to achieve the global optimal positioning accuracy. Secondly, in the case of low SNR or limited amount of observation data, CRLB may only be a loose lower bound of MSE. In addition, the estimated perceived target position is used when calculating CRLB. Since these estimates themselves may not be accurate enough, the effectiveness of CRLB optimization is affected. Finally, the optimization problem of CRLB is non-convex, and local optimization algorithms such as gradient descent may only find local optimal solutions, rather than global optimal solutions.

[0146] exist Figure 4The poor performance of the fingerprint recognition method combined with the weighted K nearest neighbor (wKNN) classifier is due to the fact that the randomness in the NLOS Rice channel model has a negative impact on the accuracy of fingerprint matching. This randomness causes fluctuations in signal strength, which in turn affects the reliability and accuracy of the positioning system based on RSS fingerprints.

[0147] Figure 5 This is a diagram showing the effect of RIS size on positioning accuracy in an embodiment of the present invention. It can be seen from the figure that the larger the size of the RIS, the higher the corresponding positioning accuracy. Since a large-sized RIS can provide more angular information about the position of the sensing device, the system positioning performance will be improved. In order to adapt to RIS of different sizes, the output dimension of the last layer of the neural network needs to be adjusted.

[0148] Figure 6 The received signal strength distribution diagram in the embodiment of the present invention is used to optimize the beamforming mode by adaptively adjusting the RIS coefficient in different time frames, so as to achieve effective positioning of the sensing target at any position. In the experiment, the received signal strength distribution diagram, i.e., the radio heat map, is drawn to intuitively display this process. Specifically, the system records the current RIS configuration in each time frame and draws the RSS value in units of 1m×1m on the xy plane of the area of ​​interest. The sensing target position is marked with a red dot. It can be clearly seen from the observation results of Figure (a) that the system uses a wider beam for user search in the first few time frames, and then gradually focuses the beam on the real position of the sensing target as time goes on. This phenomenon verifies that the proposed LSTM network can intelligently design an effective RIS configuration based on real-time measurement data. It is worth mentioning that at t=2, although the beam is basically aligned with the user position, in the subsequent time frame, the system actively expands the beam range. This is not a design defect, but a strategic decision of the LSTM network, which is intended to avoid excessive focus on a certain direction too early and ignore other possible sensing target positions. This design strategy reflects the non-greedy nature of the LSTM network, which consciously explores other possible directions to ensure that accurate target position estimation is ultimately obtained by jointly designing the RIS pattern sequence throughout the measurement cycle. Compared to Figure (b), which shows a static beamforming strategy that only sets the RIS configuration scheme based on past channel statistics, this scheme fails to show a beam pattern that gradually focuses over time, highlighting the superiority of the LSTM network-based method in dynamic environments.

[0149] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. The above embodiments and descriptions are only preferred examples of the present invention and are not intended to limit the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. An adaptive reconfigurable intelligent surface positioning method based on long short-term memory network prediction, characterized in that: The adaptive reconfigurable intelligent surface positioning method based on long short-term memory network prediction comprises the following steps: Step 1: Build an adaptive reconfigurable intelligent surface positioning system model, which consists of three entities, including reconfigurable intelligent surface RIS, sensing target and environmental clutter; Step 2: The passive reflection mechanism of the adaptive reconfigurable intelligent surface positioning system model is explained from the perspective of channel modeling and offline and online training, and the angular domain information of the perceived target near the reconfigurable intelligent surface RIS is inferred by analyzing the echo signals received on the direct link and the reconfigurable intelligent surface RIS reflection echo link; Step 3: Using the learning positioning technology of the long short-term memory (LSTM) network, the reflection coefficient of the reconfigurable smart surface (RIS) is adaptively updated in a codebook-free manner based on historical measurement data.

2. The adaptive reconfigurable intelligent surface positioning method based on long short-term memory network prediction according to claim 1 is characterized in that: In step 1, the reconfigurable smart surface RIS is composed of N = N h ×N v The block reflection unit is composed of a unit spacing of d I , where N h and N h are the number of RIS units in the horizontal and vertical directions, respectively; The reflective unit is illuminated by the detection signal emitted by the reconfigurable intelligent surface RIS controller using an omnidirectional antenna within a short distance, and forms a reflected beam to sense the target in its vicinity; M sensor receivers are also installed on the reconfigurable intelligent surface RIS to receive the echo signals of the perceived target, and then obtain the azimuth and pitch angle of the target through signal processing. The block sensing receiving units are placed in the horizontal and vertical directions, with a unit spacing of d. s , wherein the echo signal includes two source paths, namely a single-hop direct echo link and a double-hop RIS reflection echo link.

3. The adaptive reconfigurable intelligent surface positioning method based on long short-term memory network prediction according to claim 2 is characterized in that: The single-hop direct echo link is RIS controller→sensing target→RIS sensor unit, and the double-hop RIS reflection echo link is RIS controller→RIS reflection unit→sensing target→RIS sensor unit.

4. The adaptive reconfigurable intelligent surface positioning method based on long short-term memory network prediction according to claim 1, characterized in that: In step 2, the single-hop direct echo link and the double-hop RIS reflection echo link are modeled respectively, and the steering vector function Defined as: in, represents the constant phase difference between two blocks of adjacent unit observations, (assumed to be an even number) represents the ULA size; The single-hop direct echo link is modeled as: H d =a d r(θ RT,h ,i RT,v ) In the formula, α d =β d G d is the direct path complex gain, β d ∽CN(0,1) is the negative gain of the small-scale channel, represents the signal attenuation during a single hop, λ is the carrier wavelength, κ represents the radar cross-section, and d CT represents the distance from the RIS controller to the sensing target, d RT Indicates the distance from RIS to the perceived target; r(·) represents the receiving response vector of the RIS sensor and is defined as: in, θ RT,h and θ RT,v Horizontal azimuth and vertical pitch angle from RIS to the perceived target; The channel of the double-hop RIS reflection echo link at the snapshot can be expressed as: Assume that p RIS =[x RIS ,y RIS ,z RIS ] represents the position of RIS, and p = [x, y, z] represents the position of the perceived target. The position relationship satisfies the following relationship: sin(θ RT,h )cos(θ RT,v )=(yy RIS ) / d RT sin(θ RT,v )=(zz RIS ) / d RT 。 5. The adaptive reconfigurable intelligent surface positioning method based on long short-term memory network prediction according to claim 4 is characterized in that: Since the distance between the target and RIS under far-field conditions is much larger than the wavelength and the size of the reflecting surface, in this case, the electromagnetic wave is a plane wave, and the arrival angle AOA from the target to the RIS is the same as the departure angle AOD from the RIS to the target. The echo link is modeled as: H RTS =a r r(θ RT,h ,i RT,v )a T (i RT,h ,i RT,v ) Among them, α r =β r G r , a(·) is the response vector function of RIS, defined as: The channel between the RIS controller and the RIS reflection unit can be modeled as: H CR =a CR a(θ CT,h ,i CT,v ) in, is the path complex gain from the controller to the reflector unit, d CR is the distance between the controller and the reflection unit, θ CT,h ∈[-π / 2,π / 2] and θ CT,v ∈[-π / 2,π / 2] represents the physical azimuth and vertical arrival angle at RIS, respectively; The reflection coefficient of RIS is controlled by the RIS controller. Assume that the reflection coefficient of RIS is: in, Represents each snapshot The RIS reflection vector at as the phase shift of the nth element of RIS.

6. The adaptive reconfigurable intelligent surface positioning method based on long short-term memory network prediction according to claim 1, characterized in that: During the offline training in step 2, it is assumed that the environment state remains stable on a time scale much longer than the real-time perception operation, and there is no perception target in the environment. The signal vector received by the RIS sensor is: Among them, h r,l (t) and h d,l denote the reflected echo link and the direct echo link associated with each clutter l∈1,2,…,L at each snapshot t, respectively, and H CS represents the channel between the RIS controller and the RIS sensor. When x=1, all potential impacts of environmental clutter on the received signal are obtained.

7. The adaptive reconfigurable intelligent surface positioning method based on long short-term memory network prediction according to claim 1, characterized in that: During the online evaluation in step 2, for each RIS reflection vector in time snapshot t, Based on the measured offline channel information, the unwanted interference from the RIS controller and all environmental clutter are removed from the received signal to obtain the received signal: Where x(t) is the signal sent by the RIS controller at snapshot t, assuming x(t) = 1, w(t) ∽ CN(0,2σ 2 I) represents the zero-mean additive white Gaussian noise vector at snapshot t, with normalized noise power of 2σ 2 .

8. The adaptive reconfigurable intelligent surface positioning method based on long short-term memory network prediction according to claim 1, characterized in that: In each time frame t in step 3, an LSTM unit is used and the hidden state vector s is updated with the new features as input t and the cell state vector c t , according to the constraints of the receiver detection hardware new features π t It can be the received signal strength |y(t)| 2 , or the real and imaginary parts of the received pilot [R(y(t)),I(y(t))], using To represent the product between elements, the update rule of the state vector is as follows: c t =f t °c t-1 +i t °tanh(A c π t +U c s t-1 +b c ) S t =o t °tanh(c t ) f t =sigmoid(A f π t +U f s t-1 +b f ) i t =sigmoid(A i π t +U i s t-1 +b i ) o t =sigmoid(A o π t +U o s t-1 +b o ) Among them, {A f ,A i ,A o ,A c ,U f ,U i ,U o ,U c } and {b f ,b i ,b o ,b c } are the trainable weights and bias values ​​in the LSTM network respectively; Initialize the LSTM network and set s -1 =c -1 =0, and y(0)=1, the hidden state vector s t Used as input to a fully connected neural network consisting of L layers, the RIS configuration for the next time frame t+1 is: where β l , l∈{1,…,L} is the activation function Relu(·) of the lth layer, and is a set of trainable weights and bias values; Contains the real and imaginary parts of the RIS reflection coefficient: To enforce the unit norm constraint on each RIS cell, the output is element-wise normalized: After T time frames, the final estimated position of the perceived target is obtained through a fully connected neural network based on the final unit state: Among them, {A′, b′} are the sets of trainable weights and bias values, respectively, and σ r is the activation function of the rth layer. For the first R-1 layers, the activation function is set to the ReLU function, that is, σ r =relu(·), r=1,…,R-1.

Images