A double-RIS-aided wireless positioning method based on CNN and improved KNN algorithm

Abstract

The application belongs to the field of communication positioning and relates to a double-RIS assisted wireless positioning method based on a CNN and an improved KNN algorithm; the method comprises selecting an optimal RIS through an RIS selection criterion. The load impedance value of the reflection unit of the optimal RIS is adjusted to obtain multiple RIS configurations, RSSI measurement values are used as positioning features to obtain an RSSI fingerprint dataset; the collected RSSI fingerprint data are input into a CNN network for training; the RSSI fingerprint data of a to-be-measured point are input into the trained CNN network, and the position of the to-be-measured target is estimated on the basis of the cosine similarity relationship between the to-be-measured point and a reference point and the KNN algorithm. The application can avoid the problems of path loss and time delay difference existing in single-RIS assisted user positioning, improve positioning performance, improve positioning efficiency, and make positioning more accurate.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of communication positioning, and particularly relates to a dual-RIS assisted wireless positioning method based on a CNN and an improved KNN algorithm. BACKGROUND

[0002] In recent years, with the increasing development of networks, high-precision positioning related services have attracted much attention. The services provided by the fifth generation mobile communication system will greatly depend on positioning information, which can be used for network optimization and design. For different signal measurement procedures, the positioning techniques used in cellular networks include: 1) trilateration, measuring the time of arrival (ToA), time difference of arrival (TDoA) or received signal strength (RSS). 2) triangulation, measuring the angle of arrival (AoA) and angle of departure (AoD). 3) proximity, detecting the nearest cell ID, and 4) fingerprinting methods. These traditional techniques usually require the participation of multiple base stations.

[0003] In the 5G system, positioning using millimeter wave signals and massive multiple-input multiple-output systems has been widely studied, but user equipment usually needs to be equipped with multiple antennas and perform heavy calculations, such as high-dimensional array processing and accurate angle estimation, which is not feasible for mobile devices with light weight and limited energy. Reconfigurable intelligent surface (RIS), also known as large intelligent surface, has attracted great attention due to its flexibility and high efficiency in communication services. RIS is a metasurface composed of integrated circuits that can modify the incoming electromagnetic field through programming. It can be regarded as a close integration of many micro antenna elements and reconfigurable processing networks, that is, RIS can redirect signal propagation by adjusting the phase shift of the signal reflection element with passive elements, and does not introduce additional thermal noise. In addition, RIS can be easily deployed in the surrounding environment and has low cost, so that the entire environment is electromagnetically controllable. Therefore, RIS can significantly improve the signal quality of the receiver by intentionally and deterministically changing the propagation environment.

[0004] One of the most prominent applications of RIS is to establish high-speed communication links through the reflected propagation path when the direct link is blocked. Another application is to counteract local signal dead zones, in which case the signal quality is not good enough in the urban or indoor harsh propagation environment, and RIS is both economical and energy-saving compared to deploying expensive base stations or repeaters. Since RIS has the ability to change the radio propagation environment, it is widely studied by many researchers as a major technology for post-5G positioning. Many studies have shown that a RIS can be used to assist wireless positioning, but there are still some problems in practical applications. When the UE is far away from the RIS, there is a large path loss, which causes the signal to be constantly weakened, thereby reducing the positioning accuracy. In addition, when the UE is on or near the straight line between the BS and the RIS, the delay difference and the angle between the direct path and the reflected path will be very small, which is easy to cause a large positioning error. SUMMARY

[0005] In view of the above problems, the present application provides a double RIS assisted wireless positioning method based on CNN and improved KNN algorithm. In the offline stage of the method, the RSSI fingerprint map collected is trained by using the learning ability of the CNN network, and in the test stage, the RSSI data of the test point is used to realize the prediction of the position of the test user.

[0006] The double RIS assisted wireless positioning method based on CNN and improved KNN algorithm of the present application comprises:

[0007] Select the best RIS from the RIS array through the RIS selection criterion;

[0008] Determine a plurality of RIS configurations by adjusting the load impedance value of the reflection unit of the best RIS;

[0009] Under each RIS configuration, collect the RSSI fingerprint data set according to the RSSI value received by the reference point;

[0010] The collected RSSI data set is input into the CNN network for training as a feature;

[0011] The RSSI fingerprint data of the test point is input into the trained CNN network, and the weighted KNN algorithm is used to estimate the positioning position of the test point.

[0012] The present application has the following advantages:

[0013] 1) In the online positioning stage of the present application, a convolutional neural network is introduced, which can better learn the features between the RSSI fingerprint data and more accurately predict the user position coordinates due to its nonlinear learning ability.

[0014] 2) The present application does not use the Euclidean distance between the reference point and the test point as the weight to calculate the position of the test point, but uses the cosine similarity between the two as the weight to calculate the position, which can avoid or even eliminate the positioning error caused by environmental factors and different device terminals. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 is a communication system model schematic diagram of an embodiment of the present application;

[0016] Figure 2 is a double-RIS assisted wireless positioning block diagram based on CNN and improved KNN algorithm of an embodiment of the present application;

[0017] Figure 3 is a RIS assisted wireless positioning method flow chart based on deep learning of an embodiment of the present application;

[0018] Figure 4 is a CNN model structure diagram used in an embodiment of the present application;

[0019] Figure 5 is a CNN network training flow chart of an embodiment of the present application;

[0020] Figure 6 is a positioning error principle diagram of different models of terminals of an embodiment of the present application. DETAILED DESCRIPTION

[0021] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0022] Figure 1 is a communication system model schematic diagram of an embodiment of the present application; as Figure 1 shown, a two-dimensional scene of a square area is considered, the positioning area includes a base station equipped with N antennas, a single antenna user randomly distributed in the area, two RISs composed of M passive reflecting antenna units, both located at the boundary of the area, each RIS can be connected to the base station through a programmable controller, so the base station can control the phase shift of each RIS. The positions of BS, two RISs and UE are b=(b x ,b y ), s1=(s 1x ,s 1y ), s2=(s 2x ,s 2y ), u=(u x ,uy ), for convenience of description, the above communication system model will be introduced in the case of one RIS. We consider a point-to-point uplink transmission system, where the reference signal for positioning is transmitted by the UE to the BS.

[0023] The channel of the direct path between the BS and the UE is constructed using the scatterer model as as shown in equation (1).

[0024]

[0025] where, is the channel coefficient of the pth path, which satisfies independent and identically distributed, the path loss and fading affect the value of. is the antenna steering vector of the pth path, θ UB,p is the AoA of the path, which takes values a(θ UB,p ) is the ith term of:

[0026]

[0027] where λ is the wavelength and d is the antenna spacing.

[0028] The channel between the UE and the RIS is

[0029]

[0030] where P2 is the number of arrival paths from the UE to the RIS, which satisfies independent and identically distributed. Similarly, the channel from the RIS to the BS is modeled as:

[0031]

[0032] where P3 is the number of propagation paths from the RIS to the BS, is the AoD of the pth path between the RIS and the BS.

[0033] The AoA and AoD of each propagation path are expressed as:

[0034]

[0035] Assuming that the UE transmits a positioning reference signal x with a duration of t0to the BS, the received signal y d on the direct path between the UE and the BS is expressed as:

[0036]

[0037] where ρ is the transmission power of the positioning reference signal, t d,UB =||u-b||2 / c is the time delay of the direct path from the UE to the BS, c is the speed of light, and ||·||2 denotes the Euclidean norm.

[0038] Assume that the phase shift matrix on the RIS is where β1, β2, … β M denote the phases of the M reflecting elements. Then, the received signal y r of the reflected path is:

[0039]

[0040] where t d,US =||u-s||2 / c and t d,SB =||s-b||2 / c denote the time delays from the UE to the RIS and from the RIS to the BS, respectively.

[0041] At the base station, the total received signal y is the combination of the direct path and reflected path received signal vectors, denoted as:

[0042] y(t) = W|y d (t) + y r (t) | + Wn(t) (8)

[0043] where, is the combination matrix of the base station, denotes the additive white Gaussian noise, satisfying the normal distribution.

[0044] Based on the above analysis, Figure 2 is the dual-RIS assisted wireless positioning block diagram based on CNN and improved KNN algorithm of the embodiments of the present application, as Figure 2 shown, the method mainly includes two stages: in the first stage, the phase shift of each reflecting element of the RIS is designed first, then the user sends the positioning reference signal to the base station, and the delay difference corresponding to each RIS of the direct and reflected paths is calculated through the cross-correlation of the signals on the two paths, and finally the best RIS is selected through the RIS selection criterion. In the second stage, using the selected RIS, by adjusting the load impedance value of the reflecting unit, a plurality of RIS configurations are obtained, under different RIS configurations, the RSSI values received by each reference point are different, therefore, the RSSI measurement values can be used as the positioning features to obtain the RSSI fingerprint data set; the collected RSSI fingerprint data is input into the convolutional neural network for training to obtain the offline fingerprint database; the RSSI fingerprint data of the to-be-measured point is input into the deep learning model after training, and the cosine similarity relationship between the to-be-measured point and the reference point is used to estimate the position of the to-be-measured target based on the KNN algorithm.

[0045] Specifically, a RIS-aided wireless positioning method based on CNN and improved KNN algorithm, as shown in Figure 3 The method comprises the following steps:

[0046] 101. Selecting the optimal RIS from the RIS array through the RIS selection criterion;

[0047] In the embodiment of the present application, the step of selecting the optimal RIS from the RIS array through the RIS selection criterion comprises the following steps: controlling the phase shift of the reflecting unit of each RIS by the base station, sending the positioning reference signal by the reference node to the base station, calculating the time delay difference between the direct path and the reflected path of each RIS by the cross-correlation of the reference signal, determining the peak intensity ratio of the noise and the time delay between the reflected path and the direct path by the time delay difference, and selecting the RIS with the maximum peak intensity ratio as the optimal RIS through the RIS selection criterion.

[0048] The step of controlling the phase shift of the reflecting unit of each RIS by the base station comprises the following steps: calculating the total channel gain of the reflected propagation path by using the transmission channel between each RIS and the base station, the transmission channel between each RIS and the reference node, and the phase shift matrix on the RIS, controlling the phase shift of the reflecting unit of each RIS by maximizing the total channel gain of the reflected path of each RIS; wherein the phase shift of the reflecting unit of each RIS is related to the maximization of the total channel gain of the reflected propagation path.

[0049]

[0050] s.t.0≤β1,β2,…β M <2π

[0051] The step of selecting the optimal RIS through the RIS selection criterion comprises the following steps:

[0052]

[0053] Wherein, ξ id is the peak intensity ratio of the noise and the time delay between the reflected path and the direct path, the index id represents the RIS with the same id, and is represented as:

[0054]

[0055] Wherein, ρ represents the transmission power of the positioning reference signal, represents the noise variance, M represents the number of passive reflecting antenna units of each RIS, and α 1,id represents the channel coefficient of the LoS channel between the reference node UE and the base station BS, and α 2,id represents the channel coefficient of the LoS channel between the UE and the RIS, and α 3,idchannel coefficient representing a LoS channel between the RIS and the BS, a(θ 1,id ) represents an antenna steering vector between the UE and the BS, a(θ 3,id ) represents an antenna steering vector between the RIS and the BS, and the superscript H represents a Hermitian matrix.

[0056] It can be understood that in the actual application of the traditional single-RIS-aided positioning, there are two problems, first, when the UE is far away from the RIS, the strength of the reflected signal is reduced in the propagation process, resulting in a large positioning error; second, when the UE is located on or near the straight line between the BS and the RIS, the delay difference and the angle between the direct path and the reflected path are small, thereby reducing the positioning accuracy. In view of the above problems, the double-RIS-aided positioning is considered, in the above embodiment of the present application, first, the phase shift of the reflecting element on each RIS is designed, then the UE sends a positioning reference signal to the BS, and the BS calculates the delay difference of the direct and reflected paths corresponding to each RIS through the cross-correlation of the signals on the direct and reflected paths. Then, the best RIS is selected according to the RIS selection criterion, that is, the peak intensity ratio of the noise to the delay between the reflected path and the direct path is determined by the time delay, the best RIS with the maximum peak intensity ratio is selected, at this time, the best RIS can include most of the configurations as possible, ensuring the integrity and diversity of the RIS fingerprint training samples.

[0057] 102. A plurality of RIS configurations are determined by adjusting the load impedance values of the reflecting units of the best RIS;

[0058] In the embodiment of the present application, a plurality of RIS configurations are obtained by adjusting the load impedance values of the reflecting units of the selected RIS, under different RIS configurations, the RSSI values received by each reference point are different, therefore, the RSSI measurement values can be used as the positioning features to obtain the RSSI fingerprint data set.

[0059] 103. The RSSI fingerprint data set is collected according to the RSSI values received by the reference points under each RIS configuration;

[0060] In the embodiment of the present application, each RSSI fingerprint data in the RSSI fingerprint dataset is composed of an RSSI vector; the position coordinates of the RSSI fingerprint map of the measured point can be marked by those skilled in the art, and in some embodiments, the sampling area can also be gridded, and part of the RSSI fingerprint map data in each grid area is collected, and the grid code is taken as a label; for example, the positioning area is divided into 100 grids, and the real label used is 1-100, each number represents a fixed grid area, and the label number of the grid area can be taken as the known positioning position of the measured point. The position coordinates of the RSSI fingerprint map of the to-be-measured point are unknown and need to be predicted and estimated through subsequent steps to realize the prediction of the positioning position of the to-be-measured point.

[0061] 104. inputting the collected RSSI dataset as a feature into the CNN network for training;

[0062] In the embodiment of the present application, a deep learning model based on a convolutional neural network (CNN) model is selected to quickly locate the initial positioning position of the to-be-measured point user, which lays a foundation for online positioning, and the structure is as shown in Figure 4 Therefore, the input of the neural network is the data of the RSSI fingerprint map, and the output is the serial number of the measured point user, that is, the position of the measured point.

[0063] Specifically, in the present embodiment, the RSSI dataset is used to train the convolutional neural network model for multiple rounds to obtain a reliable convolutional neural network model, as shown in Figure 5 First, a single-channel RSSI fingerprint grayscale image is input, then the CNN network weight is initialized, the connection weight of all neurons is set to a random value between 0 and 1, and the sample error threshold and network learning rate are set. The network is trained by using the back propagation algorithm, if the loss function between adjacent iterations decreases below the threshold or the number of iterations is met, then the network is stable, the parameters can be saved and the training is exited, otherwise, the input is reselected from the training data for continuous training.

[0064] In order to prevent gradient disappearance and non-linear problems between data, a ReLU activation function is used between each layer of hidden layers. Compared with many traditional neural network activation functions such as tanh function and logic function, the ReLU function is a one-sided inhibition function with sparse activation, which not only effectively solves the problems of gradient disappearance and gradient explosion in the network training process, but also simplifies the calculation process and reduces the calculation cost.

[0065] In addition, to avoid overfitting or underfitting during the entire training process, we adopted the dropout technique. The principle is that in a neural network, some neurons are disabled with a certain probability. Thus, the neurons used for training are different each time, and the resulting subnetwork is obtained randomly according to probability, which enhances the robustness of the network.

[0066] 105. Input the RSSI fingerprint data of the point to be tested into the trained CNN network, and use the weighted KNN algorithm to estimate the location of the point to be tested.

[0067] In this embodiment of the invention, considering that the traditional weighted KNN algorithm is a commonly used localization algorithm, its principle is to calculate the Euclidean distance between the RSSI vector in the fingerprint database and the RSSI vector acquired online, select the K nearest reference points to the measurement point as localization labels, use the distance weight as the weight of the label, and then calculate the coordinates of the point to be measured. However, in practical applications, different environments and measurement terminals will affect the RSSI value. For example... Figure 6 As shown, OP1 and OP2 represent methods for calculating the position of the test point using the Euclidean distance between the reference point and the test point as weights. The lengths of the two vectors SP1 and SP2 are different, but their cosine similarity with the reference point is the same. Therefore, this invention proposes an improved KNN algorithm, which can avoid or even eliminate positioning errors caused by environmental factors and different device terminals when using this as a weight for position calculation. The similarity relationship between the reference point and the test point is compared as shown in formula (9).

[0068]

[0069] Among them, <RSS O RSS P > represents the inner product of the RSSI vectors at points O and P, |RSS O |·|RSS P | represents the product of the lengths of the RSSI vectors. Then, using the weighted centroid algorithm, a polygon composed of K different reference points is obtained, and the centroid of this polygon is the position of the point to be measured.

[0070] In this embodiment of the invention, considering that the measured point and the K reference points are relatively far apart, and that the final result will be affected by the interference of these reference points, causing the estimated position to be biased towards some reference points and resulting in unbalanced information provided by each reference point; therefore, this invention also selects the K closest reference points based on the density of reference points around the measured point; specifically as follows:

[0071] The local density of each reference point is calculated according to cosine similarity, whether the reference points can be used as clustering centers is determined according to the local density, if the reference points are not used as clustering centers, the non-clustering centers are classified according to the corresponding evaluation value; the distance weight is used as the weight of the label, and the coordinates of the to-be-tested point can be obtained after calculation.

[0072] The calculation formula of the local density of the reference point is represented as:

[0073]

[0074] In the formula, ρ i represents the local density of the i th reference point; d ij represents the distance between the i th reference point and the j th reference point, and the cosine similarity distance is taken as an example; the parameter d c is a truncation distance.

[0075] The calculation formula of the evaluation value is represented as:

[0076]

[0077] In the formula, μ (i) represents the evaluation value of the i th reference point; i ∈ C k , C k represents the k th cluster, k = 1, 2,..., m; α represents the first weight, β represents the second weight, the range of the two weights is 0-1, and the range of 1-α and 1-β is also 0-1; d ic represents the distance between the i th reference point and the cluster center point c; ρ i represents the local density of the i th reference point; ρ min represents the minimum value of the local density of the reference points in the cluster; ρ max represents the maximum value of the local density of the reference points in the cluster; R represents the cluster radius surrounding most reference points.

[0078] In the above evaluation value of the application, first, the closer the distance between the reference point and the cluster center point, the greater the evaluation value; second, the greater the local density of the reference point, the greater the evaluation value; in this way, more important reference points can be included in the positioning estimation; and the accuracy of wireless positioning is improved.

[0079] Those skilled in the art can understand that all or part of the steps in the above-mentioned embodiments can be completed by programs instructing related hardware, and the programs can be stored in a computer readable storage medium, and the storage medium can include ROM, RAM, magnetic disk or optical disk, etc.

[0080] While embodiments of the application have been shown and described, it is to be understood that the embodiments described are merely exemplary of the principles and application of the present application. Numerous modifications and adaptions can be effected without departing from the spirit and scope of the present application, which is not limited to the exact construction and arrangement described. It is intended, therefore, to cover all modifications and adaptions that fall within the scope of the claims and their equivalents.

Claims

1. A dual-RIS-assisted wireless positioning method based on CNN and an improved KNN algorithm, characterized in that, The method includes: The optimal RIS is selected from the RIS array using RIS selection criteria. This includes: controlling the phase shift of the reflection unit of each RIS through the base station; the reference node sending a positioning reference signal to the base station; calculating the time delay difference between the direct path and the reflection path of each RIS using the cross-correlation of the reference signals; determining the peak intensity ratio of the noise to the time delay between the reflection path and the direct path using the time delay difference; and selecting the RIS with the largest peak intensity ratio as the optimal RIS. The step of controlling the phase shift of the reflection unit of each RIS via the base station includes: calculating the total channel gain of the reflection propagation path using the transmission channel between each RIS and the base station, the transmission channel between each RIS and the reference node, and the phase shift matrix on the RIS; and controlling the phase shift of the reflection unit of each RIS by maximizing the total channel gain of the reflection path of each RIS. The formula for selecting the optimal RIS is: ； in, The index represents the ratio of the peak intensity of the noise to the time delay between the reflection path and the direct path, respectively. Indicates that there are the same RIS; The ratio of the peak intensity of the noise to the time delay between the reflection path and the direct path is given by the following formula: ； in, This indicates the transmission power of the positioning reference signal. Indicates the noise variance. This indicates the number of passive reflective antenna elements in each RIS. This represents the channel coefficient of the LoS channel between the reference node and the base station. This represents the channel coefficient of the Loss Channel between the reference node and the RIS. This represents the channel coefficient of the LoS channel between the RIS and the base station. This represents the antenna steering vector between the reference node and the BS. This represents the antenna steering vector between the RIS and the base station, with the superscript H indicating the Hermitian matrix; By adjusting the load impedance value of the reflector unit of the optimal RIS, various RIS configurations can be determined; Under each RIS configuration, an RSSI fingerprint dataset is collected based on the RSSI values ​​received at the reference point; The collected RSSI dataset is used as a feature input to the CNN network for training; The RSSI fingerprint data of the point to be tested is input into the trained CNN network, and the localization location of the point to be tested is estimated by the weighted KNN algorithm. The step of estimating the location of the test point using the weighted KNN algorithm includes: determining the K reference points that are closest to the test point by cosine similarity distance, and using the weighted centroid algorithm to obtain a polygon composed of the K reference points, the centroid of which is the location of the test point. The K reference points closest to the cosine similarity distance of the point to be tested are determined. The formula for the cosine similarity between the reference points and the point to be tested is as follows: ； in, This represents the inner product of the RSSI vectors at the reference point O and the point to be measured P. This represents the product of the lengths of the RSSI vectors at the reference point O and the point to be measured P.

2. The dual RIS-assisted wireless positioning method based on CNN and improved KNN algorithm according to claim 1, characterized in that, The step of using the collected RSSI dataset as features to train a CNN network includes dividing the collected RSSI dataset into multiple fingerprint vectors, processing the fingerprint vectors into fingerprint matrices, converting each fingerprint vector in the RSSI dataset into a single-channel fingerprint grayscale image, inputting the fingerprint grayscale image into the CNN network, initializing the CNN network weights, setting the connection weights of all neurons in the CNN network to random values ​​between [0,1], and setting the sample error threshold and network learning rate; training the network using the backpropagation algorithm; if the loss function between adjacent iterations decreases below the error threshold or the required number of iterations is met, then the CNN network has completed training, the network weight parameters are saved, and training is terminated; otherwise, training continues.

3. The dual RIS-assisted wireless positioning method based on CNN and improved KNN algorithm according to claim 1, characterized in that, The process of determining the K reference points closest to the cosine similarity distance of the point to be tested includes calculating the local density of each reference point based on the cosine similarity distance, determining whether these reference points can serve as cluster centers based on their local density, and classifying these non-cluster centers according to their corresponding evaluation values ​​if they are non-cluster centers; the K cluster centers are then used as the K reference points; the local density of the reference points is calculated using the following formula: ； in, Indicates the first Local density at a reference point; Indicates the first The distance between the j-th reference point and the j-th reference point, parameter To cut off the distance; The evaluation value is calculated using the following formula: ； in, Indicates the first Evaluation values ​​for each reference point; , Indicates the first One cluster, ; Indicates the first weight. This represents the second weight; both weights are in the range of 0 to 1, and... and The range is also between 0 and 1; Indicates the first Reference points to the cluster center The distance; Indicates the first Local density at a reference point; The minimum local density of the reference point in this cluster; This represents the maximum local density of the reference point in the cluster; Indicates the cluster radius.

Images