Wi-Fi angle of arrival estimation and positioning system and method based on end-to-end neural network

Abstract

The invention discloses a Wi-Fi (Wireless Fidelity) arrival angle estimation and positioning system and method based on an end-to-end neural network, in an arrival angle estimation stage, the neural network is adopted for analysis for screening of visual path observation points and arrival angle estimation, and in a positioning stage, an image filtering and probability aggregation method is adopted. The method comprises the following specific steps of: 1, estimating channel state information by using a Wi-Fi terminal receiving signal and a known pilot frequency; 2, carrying out dimension conversion and virtual and real number conversion on the channel state information to adapt to a neural network; 3, judging whether the visual path exists or not by using the visual path existence judgment neural network, if so, retaining the data on the observation point, and if not, discarding the data; 4, estimating the angle of arrival of the observation point with the visual path by using an angle of arrival estimation neural network; 5, carrying out the steps on each observation point on the track to obtain a plurality of pieces of arrival angle estimation information; and 6, determining a target position by adopting a method based on image filtering and probability aggregation in combination with measurement results on a plurality of observation points in consideration of errors of the measured values of the angles of arrival.

Description

Technical Field

[0001] The present invention relates to a Wi-Fi angle of arrival estimation and positioning system and method based on an end-to-end neural network, belonging to the technical field of wireless communication sensing. Background Art

[0002] With the rapid development of wireless communication technology, wireless positioning technology has shown broad application prospects and important research value in many fields. From intelligent transportation systems, Internet of Things device tracking to military target monitoring, wireless positioning technology has become an indispensable means to obtain accurate position information. Among them, angle of arrival (AOA) estimation, as a key technology in wireless positioning, is particularly crucial in positioning tasks in complex environments due to its high accuracy and anti-interference ability.

[0003] Different from satellite positioning systems, positioning technologies based on wireless communication systems need to utilize communication signals generated by existing or self-designed wireless communication systems to estimate and obtain the position of mobile terminals (such as mobile phones, laptops, wearable devices, wireless sensors, etc.) in outdoor or indoor environments. Wireless positioning technology mainly relies on the analysis of the propagation characteristics of wireless signals. By measuring signal parameters such as time of arrival (TOA), time difference of arrival (TDOA), received signal strength indicator (RSSI), and angle of arrival (AOA), etc., to estimate the position information of the target. Among them, AOA estimation can directly provide the direction information of the target by measuring the incident angle of the signal relative to the reference direction when it arrives at the receiving antenna array. Combining with other positioning parameters can significantly improve the accuracy and robustness of the positioning system.

[0004] In recent years, with the improvement of antenna technology, signal processing algorithms, and computing power, AOA estimation methods have continuously made breakthroughs. From traditional array signal processing based on the phase difference method to modern intelligent estimation algorithms based on advanced technologies such as compressive sensing and deep learning, the accuracy and efficiency of AOA estimation have been significantly improved. These advancements have not only promoted the development of wireless positioning technology but also provided strong support for solving high-precision positioning problems in complex environments.

[0005] However, AOA estimation still faces many challenges in practical applications, such as multipath effects, non-line-of-sight propagation, antenna array layout limitations, etc. Therefore, in-depth research on key technologies of wireless positioning, especially optimization algorithms and practical application strategies for AOA estimation, is of great significance for improving the overall performance of wireless positioning systems.

[0006] Indoor positioning is an important part of the integrated application of the Internet of Things and communication perception. Although GPS can provide accurate outdoor position tracking, its effectiveness will be reduced indoors due to obstacles blocking the line-of-sight path of satellites. Wi-Fi, with its widespread access point deployment, has become a reliable and cost-effective indoor positioning method. Traditional Wi-Fi positioning methods often face high requirements for the measurement accuracy of channel state information, the positioning accuracy depends on the line-of-sight propagation conditions, or a large amount of resources are consumed for fingerprint measurement. The present invention uses a deep learning algorithm to perform feature learning on Wi-Fi-based CSI, designs a high-precision, low-complexity, lightweight positioning method by exploring the relationship between CSI and the target position, and verifies the positioning effect of the proposed method through simulation and actual measurement.

[0007] This method combines the characteristics of neural networks with the advantages of Wi-Fi communication perception, and designs a Wi-Fi angle of arrival estimation and positioning system based on an end-to-end neural network. Summary of the Invention

[0008] Object of the Invention: The present invention aims to use a deep learning algorithm to perform feature learning on Wi-Fi-based CSI, estimate the angle of arrival of the signal by exploring the relationship between CSI and the target position, and then combine the measurement of the IMU trajectory of the object itself to provide a high-precision, low-complexity, lightweight positioning method for Wi-Fi access points and the position of the object itself.

[0009] Technical Solution: To achieve the above object, the present invention adopts the following solution: A Wi-Fi angle of arrival estimation and positioning system based on an end-to-end neural network, which is characterized in that it includes a Wi-Fi receiving antenna, a Wi-Fi access point, a Wi-Fi signal analyzer, a calculation module, and an inertial measurement unit. The Wi-Fi receiving antenna receives the signal sent by the Wi-Fi access point, and the signal is analyzed by the Wi-Fi signal analyzer to obtain a channel state information matrix. The channel state information matrix is then used by the calculation module to obtain the estimated angle of arrival, and combined with the trajectory information measured by the inertial measurement unit, the estimated position of the access point is obtained.

[0010] A Wi-Fi angle of arrival estimation and positioning method based on an end-to-end neural network, the method comprising the following steps:

[0011] Step 1, use a Wi-Fi terminal to receive signals and known pilots to estimate channel state information;

[0012] Step 2, perform dimensionality conversion and imaginary-real number conversion on the channel state information to adapt to the neural network;

[0013] Step 3, use a line-of-sight path existence discrimination neural network to determine whether there is a line-of-sight path. If there is, retain the data at this observation point; if not, discard it;

[0014] Step 4: For observation points with sight paths, use the arrival angle estimation neural network to estimate the arrival angle;

[0015] Step 5: Perform the above steps at each observation point on the trajectory to obtain multiple arrival angle estimation information;

[0016] Step 6: Considering the errors in the arrival angle measurements, a method based on image filtering and probability aggregation is used to combine the measurement results at multiple observation points to determine the target position.

[0017] Among them, a fully connected neural network is used to determine whether the observation point has a view path. Step 3 is specifically as follows: the channel state information after dimension conversion and real-imaginary number conversion obtained in step 2 is input into the fully connected neural network to obtain a two-channel output. The two element values of the two-channel output are used to determine whether the observation point has a view path. If so, the data is retained for the next step, otherwise it is discarded.

[0018] Among them, a fully connected neural network is used for regression analysis to estimate the arrival angle. Step 4 is as follows: the channel state information with line of sight obtained after screening in step 3 is input into the fully connected neural network to obtain a single-channel output. The angle of this single-channel output is the estimated arrival angle.

[0019] Among them, step 6 is specifically as follows:

[0020] A method from arrival angle to probability image based on intersection statistics is used to obtain a probability cloud map, and then a probability cloud map processing method based on image filtering is used to filter the probability cloud map to obtain a processed image, and then a clustered probability cloud map processing method is used to obtain the final target position estimate.

[0021] Among them, according to the measured position and arrival angle of each observation point, a number of straight lines equal to the number of observation points are drawn, and these straight lines intersect each other to form a number of intersections. Then, an image is constructed using a limited part of the plane in which they are located. The plane is divided into a number of small blocks as pixels of the image, and the number of intersections in each pixel area or its monotonic function is used as the brightness value of the pixel to obtain a possibility cloud map. The brightness of the pixel is monotonically related to the possibility of the target position in the pixel area.

[0022] Among them, the whole image is convolved, and the part with low convolution result and low probability is eliminated, that is, the brightness of the pixels in this part is set to the brightness of the case where the probability is 0. This can avoid the influence of some points with large errors and far away from the clustered points on the results.

[0023] Among them, for the method of processing the probability cloud map of aggregation, the entire image is convolved to find the convolution region with the largest convolution result as the target region. Then, within the target region, using the brightness of the pixels as weights, the x-coordinate and y-coordinate are weighted and averaged respectively to obtain the estimated x-coordinate and y-coordinate. This estimated coordinate is the final estimated value of the target position.

[0024] In the arrival angle estimation stage, a neural network is adopted for the screening of the line-of-sight observation points and the estimation of the arrival angle. In the positioning stage, a method of image filtering and probability aggregation is adopted. Specific steps: 1. According to the channel state information obtained through acquisition and processing, use a fully connected neural network to analyze whether it has a line of sight; 2. For the observation points with a line of sight, use a fully connected neural network to calculate their arrival angles; remove the data of the observation points that do not contain a line of sight; 3. Use the method of image filtering and probability aggregation to perform positioning based on the coordinates of each point on the trajectory and its arrival angle to obtain the position of the Wi-Fi access point.

[0025] The specific positioning method is carried out according to the following steps:

[0026] a. Use the Wi-Fi terminal to receive the Wi-Fi signal and estimate the channel state information in combination with the known pilot.

[0027] b. Split the real and imaginary parts of the obtained channel state information and convert it into a tensor to adapt to the subsequent neural network.

[0028] c. The line-of-sight discrimination neural network can include several layers and adopt several forms, but it is required that the input dimension conforms to the dimension of the channel state information and the output satisfies a two-channel output. Before use, the already trained network model provided by the inventor can be used, or the user's own model can be used for retraining.

[0029] d. The output result of the previous step is a tensor with one row and two columns, which respectively represent the degree of tending to have a line of sight and tending to have no line of sight. Based on this, it is possible to judge whether there is a line of sight by comparing the magnitudes of the two column elements: if the element in the first column is greater than the second column, it is judged that there is a line of sight; otherwise, it is judged that there is no line of sight. For the case of having a line of sight, keep the data; for the case of having no line of sight, discard the data.

[0030] e. For the case of having a line of sight obtained in the previous step, perform arrival angle estimation. Adopt an arrival angle estimation neural network, the input dimension matches the dimension of the channel state information or the dimension of the channel state information after dimension conversion, and the output is an angle normalized to the range of [0, 1], and the arrival angle represented by the angle value can be obtained by multiplying by 180.

[0031] f. Multiple observation points can be sampled on a trajectory. By performing the above-mentioned visible path existence discrimination and angle of arrival estimation for each observation point, the angle information at several observation points can be obtained. Combining this angle information with the position information of the observation points measured by the inertial measurement method, at each observation point, several straight lines can be drawn according to the slope obtained from the coordinates and angle information of the observation point. The intersection of every two straight lines may be the position of the observation point.

[0032] g. Mark these intersection points on the graph. Divide the intersection point graph into several small squares with a size of 0.1m * 0.1m. For each small square, let its brightness value be the number of intersection points within the range of this square. For this image, first perform convolution processing on the entire image using a small-sized all-ones convolutional kernel to filter out pixel points with smaller values. This can reduce the influence of points with larger errors on the calculation results. Further, perform convolution processing on the image using a larger-sized all-ones convolutional kernel, sort the corresponding values of each pixel point, retain the maximum value, and record the area where it is located. For each pixel point in the above-mentioned area, perform weighted calculation according to its coordinates and brightness, that is, calculate the average value of the coordinates by weighting with brightness. The obtained coordinates are the coordinates of the estimated target to be measured.

[0033] Beneficial effects: A Wi-Fi angle of arrival estimation and positioning system based on an end-to-end neural network provided by the present invention has the following advantages:

[0034] 1. Existing Wi-Fi-based indoor geometric positioning technologies often only consider single-point positioning or two-point positioning methods, without considering the method of combining channel sensing and geometric positioning. This method proposes a method of combining wireless positioning and movement trajectories, improving the accuracy and precision of positioning. In some scenarios, the positioning accuracy can reach the decimeter level.

[0035] 2. Existing neural network-based Wi-Fi indoor positioning technologies often do not consider the cleaning of the dataset. This method uses a visible path existence discrimination neural network to discriminate whether there is a visible path, discards the part without a visible path that will bring larger errors, further improves the accuracy and precision of positioning, and can also reduce the number of observation points that need to be referenced, reducing the computational burden.

[0036] 3. In the process of estimating the target position through the angle of arrival at each observation point, this method uses the method of image filtering to filter out points with larger deviations. Through data statistics, the stability and accuracy of this method are higher than those of the traditional least squares method. Description of the Drawings

[0037] Figure 1 is a flowchart of a Wi-Fi angle of arrival estimation and positioning system based on an end-to-end neural network.

[0038] Figure 2It is the architecture diagram of a Wi-Fi angle of arrival estimation and positioning system based on an end-to-end neural network in an embodiment of the present invention. Detailed implementation manners

[0039] In the following description, specific details such as specific system architectures are presented for the purpose of illustration rather than limitation, so as to thoroughly understand the present invention. However, those skilled in the art should clearly understand that the present invention can also be implemented in other embodiments without these specific details.

[0040] Embodiment: In the solutions described herein, the angle of arrival estimation method can be used in various Wi-Fi wireless positioning methods, and the methods based on image filtering and likelihood aggregation can be used in various positioning methods based on angle of arrival estimation.

[0041] In order to make the objectives, technical solutions, and advantages of the present invention clearer, the following will be described in detail with reference to the accompanying drawings. The Wi-Fi angle of arrival estimation and positioning system based on an end-to-end neural network provided in the embodiments of the present invention, as Figure 1 shown, the method includes:

[0042] Step 101: Use a Wi-Fi terminal to receive signals and known pilots to estimate the channel state information.

[0043] It should be noted that for the channel state information matrix obtained by processing the received matrix, the number of rows is the number of receiving antennas, and the number of columns is the number of subcarriers of orthogonal frequency division multiplexing modulation. In order to measure the angle information, the number of receiving antennas is at least 2.

[0044] Step 102: Perform dimension conversion and real / imaginary number conversion on the channel state information to adapt to the neural network.

[0045] It should be noted that for the channel state information matrix obtained by processing the received matrix, the number of rows is the number of receiving antennas, and the number of columns is the number of subcarriers of orthogonal frequency division multiplexing modulation, but this may not match the input dimension of the subsequent neural network, and dimension transformation is required to match the input dimension of the neural network. For example, if the number of receiving antennas is 2, the number of orthogonal frequency division multiplexing subcarriers is 3276, and the elements in the channel state information matrix are imaginary numbers, and the subsequent neural network used is a fully connected neural network, the channel state information matrix can be converted into a 1*13104 matrix for subsequent processing. The following gives a conversion method: for the two elements in each column of the channel state information matrix, split them into their respective real and imaginary parts, and then form a vector in the order that the real part of the element with the original row number 1 is the first, the imaginary part of the element with the original row number 1 is the second, the real part of the element with the original row number 2 is the third, and the imaginary part of the element with the original row number 2 is the fourth; combine the vectors corresponding to each column in the original order to obtain a 1*13104 tensor as the feature input of the neural network.

[0046] Step 103: Use the visual diameter existence discrimination neural network to determine whether there is a visual diameter. If there is, retain the data at this observation point; if not, discard it.

[0047] It should be noted that the discrimination neural network can adopt the network provided by this specific solution. If other neural networks are used, they need to be trained independently, change the corresponding parameters, and train with the input data set. The output result is a tensor with one row and two columns, representing the degree of tending to have a visual diameter and tending to have no visual diameter respectively. Therefore, at this time, the discrimination of the presence or absence of a visual diameter based on this can be judged based on a simple size comparison, that is, if the tensor element in the first column is greater than the tensor element in the second column, it is considered that there is a visual diameter; otherwise, it is considered that there is no visual diameter. For the case of having a visual diameter, retain it; for the case of having no visual diameter, discard the data.

[0048] Step 104: For the observation points with a visual diameter, use the arrival angle estimation neural network to estimate the arrival angle.

[0049] It should be noted that the arrival angle estimation neural network can adopt the network provided by this specific solution. If other neural networks are used, they need to be trained independently, change the corresponding parameters, and train with the input data set. The output result is a single-element tensor representing the normalized angle value, ranging from 0 to 1, and multiplying by 180 gives the angle value of the arrival angle.

[0050] Step 105: Perform the above steps at each observation point on the trajectory to obtain multiple arrival angle estimation information. Multiple observation points can be sampled on a single trajectory. By performing the above visual diameter existence discrimination and arrival angle estimation on each observation point, the angle information at several observation points can be obtained. For this angle information, combined with the position information of the observation point measured by the inertial measurement method, at each observation point, several straight lines can be drawn according to the slope obtained from the coordinates and angle information of the observation point. The intersection of every two straight lines may be the position of the observation point.

[0051] Considering an observation point, given the coordinates of the observation point (x1, y1) and the arrival angle θ, the straight line can be made as follows:

[0052] y = tan(θ)x - tan(θ)x1 + y1

[0053] Then, when all the measurement values are accurate, the target must be located on this straight line.

[0054] Step 106: Considering the errors in each arrival angle measurement value, use the method based on image filtering and likelihood aggregation to combine the measurement results at multiple observation points to determine the target position.

[0055] It should be noted that this part is divided into four steps: possible image generation, filtering of points with large deviations, determination of the central range, and weighted averaging. For possible image generation, the intersection points from the previous step are marked on the graph, and then the intersection point graph is divided into several small squares with a size of 0.1m * 0.1m. For each small square, its brightness value is set to the number of intersection points within the range of this square. For filtering of points with large deviations, a 5 * 5 matrix with all elements being 1 is used to perform convolution on the entire image. If the convolution at a certain pixel point is less than 3, it is considered that the estimated point has too large an offset from other points, and the gray value of this point is set to 0. The matrix size and the filtering threshold can be modified according to the actual situation. For determination of the central range, a 9 * 9 matrix with all elements being 1 is used to perform convolution on the entire image. The 9 * 9 region with the largest convolution is considered the central region, that is, the region with the highest possibility. For weighted averaging, for each pixel point in the central region, according to its brightness, the x and y coordinates are respectively weighted and averaged to obtain the estimated position of the target.

[0056] As Figure 2 shown, a Wi-Fi angle of arrival estimation and positioning system based on an end-to-end neural network disclosed in an embodiment of the present invention includes a Wi-Fi receiving antenna, a Wi-Fi access point, a Wi-Fi signal analyzer, a calculation module, and an inertial measurement unit.

[0057] The Wi-Fi receiving antenna is used to receive 2.4GHz Wi-Fi signals, and at least 2 are required.

[0058] The Wi-Fi access point, that is, the AP, is used to omnidirectionally transmit Wi-Fi signals.

[0059] The Wi-Fi signal analyzer can analyze the signal attenuation matrix according to the signals received by the antennas.

[0060] The calculation module can convert the calculation of the signal attenuation matrix into a channel state information matrix, with the number of rows being the number of receiving antennas and the number of columns being the number of subcarriers. It can support the calculation of the forward propagation of the neural network.

[0061] The inertial measurement unit can infer the position coordinates of an object in real time according to the motion state of the object.

[0062] The above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of changes or substitutions, which should all be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claimed rights.

Claims

1. A Wi-Fi angle of arrival estimation and positioning system based on an end-to-end neural network, characterized in that It includes a WiFi receiving antenna, a Wi-Fi access terminal, a Wi-Fi signal analyzer, a calculation module, and an inertial measurement unit. The Wi-Fi receiving antenna receives the signal emitted by the Wi-Fi access terminal. The channel state information matrix is obtained by analyzing this signal with the Wi-Fi signal analyzer. The estimated angle of arrival is obtained from the channel state information matrix through the calculation module. Combining with the trajectory information measured by the inertial measurement unit, the estimated access point position is obtained.

2. A Wi-Fi angle of arrival estimation and positioning method based on an end-to-end neural network, characterized in that, The method includes the following steps: Step 1: Use the Wi-Fi terminal to receive signals and estimate the channel state information with known pilots; Step 2: Perform dimension conversion and real-imaginary conversion on the channel state information to adapt to the neural network; Step 3: Use the line-of-sight existence discrimination neural network to determine whether there is a line of sight. If there is, retain the data at this observation point; if not, discard it; Step 4: For the observation points with a line of sight, use the angle-of-arrival estimation neural network to estimate the angle of arrival; Step 5: Perform the above steps at each observation point on the trajectory to obtain multiple angle-of-arrival estimation information; Step 6: Considering the errors in each angle-of-arrival measurement value, use a method based on image filtering and probability aggregation to combine the measurement results at multiple observation points to determine the target position.

3. The Wi-Fi angle-of-arrival estimation and positioning method based on an end-to-end neural network according to claim 2, wherein Use a fully connected neural network to determine whether there is a line of sight at this observation point. Step 3 is specifically as follows: Input the channel state information after dimension conversion and real-imaginary conversion obtained in Step 2 into the fully connected neural network to obtain a two-channel output. Determine whether there is a line of sight at this observation point according to the magnitudes of the two element values of this two-channel output. If there is, retain this data for the next step; if not, discard it.

4. A Wi-Fi angle-of-arrival estimation and positioning method based on an end-to-end neural network according to claim 2, characterized in that Use a fully connected neural network for regression analysis to estimate the angle of arrival. Step 4 is specifically as follows: Input the channel state information with a line of sight obtained after screening in Step 3 into the fully connected neural network to obtain a single-channel output. The angle of this single-channel output is the estimated angle of arrival.

5. A Wi-Fi angle-of-arrival estimation and positioning method based on an end-to-end neural network according to claim 2, characterized in that Step 6 is specifically as follows: Use a method based on intersection statistics from the angle of arrival to the probability image to obtain a probability cloud map. Then use a method for processing the probability cloud map based on image filtering to filter the probability cloud map to obtain a processed image. Then use an aggregated method for processing the probability cloud map to obtain the final estimated value of the target position.

6. The Wi-Fi angle-of-arrival estimation and positioning method based on an end-to-end neural network according to claim 5, characterized in that Based on the measured positions and angles of arrival of each observation point, draw a number of straight lines equal to the number of observation points. These straight lines intersect pairwise to form a number of intersection points. Then construct an image with the finite part of the plane where they are located. Divide this plane into several small blocks as the pixels of the image. Use the number of intersection points or its monotonic function within each pixel area as the brightness value of this pixel to obtain a likelihood cloud map. The brightness of the pixel has a monotonic relationship with the likelihood of the target position within this pixel area.

7. A Wi-Fi angle-of-arrival estimation and positioning method based on an end-to-end neural network according to claim 5, characterized in that, Perform convolution on the entire image to remove the parts with lower convolution results and lower likelihoods, that is, set the brightness of the pixels in this part to the brightness in the case where the likelihood is 0.

8. The Wi-Fi angle-of-arrival estimation and positioning method based on an end-to-end neural network according to claim 2, characterized in that The method for processing the aggregated probability cloud map convolves the entire image, finds the convolution region with the largest convolution result as the target region, and then within the target region, uses the brightness of the pixels as weights to perform weighted averaging on the x-coordinate and y-coordinate respectively to obtain the estimated x-coordinate and y-coordinate. This estimated coordinate is the final estimated value of the target position.

Images