A positioning method based on channel state information

Abstract

The present invention relates to a positioning solution based on channel state information, which includes: preparation stage: dividing the training data set into N areas, the i-th area is Are i ; Coarse positioning stage: process the channel state information data into image data to obtain the coarse position #imgabs0# and the positioning error #imgabs1##imgabs2# at the same time; Fine positioning stage and positioning output stage: the output positioning result is the sum of the coarse position and the positioning error, and the deviation is corrected, that is, #imgabs3# The present invention processes the received channel state information data into image data, and then uses a deep neural network for positioning research, which can significantly reduce the system positioning error and thus improve positioning accuracy.

Description

Technical Field

[0001] The present invention relates to the field of positioning technology, and in particular to a positioning method based on channel state information. Background Art

[0002] High-precision positioning has become a fundamental requirement for smart cities and smart factories, and location-based services (LBS) are playing an increasingly important role in the next generation of emerging mobile applications. In outdoor environments, mobile device positioning typically uses the Global Positioning System (GPS), which can provide meter-level positioning accuracy during normal operation. However, standalone GPS is unreliable indoors because indoor radio propagation is accompanied by multipath effects, shadow fading, and delay distortion, resulting in weak GPS signals and inadequate performance in indoor environments.

[0003] Currently, several different approaches have been proposed for indoor positioning, such as WiFi, Bluetooth, and ultra-wideband (UWB). These approaches can be further categorized as range-based and range-free indoor positioning methods. Range-based positioning systems use geometric mapping techniques, such as trilateration, triangulation, and multi-dimensional measurement, to locate the target object. They rely on parameter estimation of received signals, such as time difference of arrival (TDOA), angle of arrival (AoA), and received signal strength indicator (RSSI). These techniques require multiple access points (APs) to calculate the mobile unit (MU) position and are sensitive to multipath effects. Furthermore, most of these methods are designed for Loss of Surface (LOS) channels and perform poorly in scenarios where NLOS is dominant. Range-free systems use fingerprints to determine the location of an object. Each location is represented by a unique channel fingerprint. Channel fingerprints are derived by processing RSSI and channel state information (CSI) to construct a fingerprint library. During the online testing phase, the corresponding positioning algorithm is used to compare the newly acquired fingerprint with the data in the fingerprint library to obtain the positioning result. However, this method requires a significant amount of time to build a fingerprint database, the precision of which directly affects positioning accuracy, and the database is difficult to maintain. Furthermore, this method has limited coverage and increased deployment costs, making it difficult to achieve wide-area coverage.

[0004] Due to the existence of spatially consistent scene changes caused by the transition of scattering points, there will be system positioning errors when using positioning technology for positioning. However, there is currently no solution to this problem in the positioning system. Summary of the Invention

[0005] In order to solve the problem of systematic errors in positioning results caused by spatial consistency transformation due to scattering point transition in positioning scenarios, the present invention proposes a positioning method based on channel state information, which can reduce system positioning errors and achieve higher positioning accuracy.

[0006] To achieve the above object, the technical solution adopted by the present invention is:

[0007] A positioning solution based on channel state information, the method comprising:

[0008] Preparation stage: Divide the training data set into N regions, with the i-th region being Are i ;

[0009] Coarse position determination: Process the channel state information data into image data and input it into the feature extraction module to obtain the coarse position And at the same time get the positioning error Where Pos is the real position coordinate;

[0010] Fine positioning stage: use selector P ro According to the rough position Selection bias network Ld i , training bias network The positioning error obtained in the first step when training the network For labels, place the And the features of the feature extraction network after the first step are spliced ​​as the i-th deviation network Ld i The input of the bias correction network is trained, and the positioning error selected by the error selector is obtained as

[0011] Positioning output stage: The output positioning result is the sum of the rough position and the positioning error, and the deviation is corrected, that is,

[0012] The coarse positioning stage includes the following steps:

[0013] Step S11: Data processing, processing the received channel state information data into image data;

[0014] Step S12: constructing a positioning network for extracting information from the processed channel state information image data, reflecting the relationship mapping between the channel state information and the two-dimensional position coordinates;

[0015] Step S13: training the positioning network, training the positioning network proposed in step S12, using the processed channel state information data as the input of the positioning network, and using the real position of each position point as the label to train the positioning network; obtaining the rough position The positioning network is represented by function f, and this process is expressed by the following formula:

[0016]

[0017] Among them, the loss function is:

[0018]

[0019] in, Indicates the rough positioning of the i-th position, Pos i is the accurate positioning of the i-th position, N sp is the number of sample points; is L2 regularization, θ 2 is the penalty term, and the parameter η controls the strength of the penalty term;

[0020] Step S14: Calculate the difference between the rough position obtained in step S13 and the true position, i.e., the positioning error: Where Pos is the real position coordinate.

[0021] In step S11, the format of the channel state information data is: number of base stations * number of antennas of each base station * number of sampling points * real part and imaginary part of each CSI data; then the specific method of processing the channel state information data into image data is:

[0022] The number of base stations * the number of antennas of each base station is taken as the height H of the image, the real and imaginary parts of each channel status information data are taken as the width W of the image, and the number of sampling points is taken as the number of channels C of the image.

[0023] The fine positioning stage specifically includes the following steps:

[0024] Step S31: Constructing a deviation network. Step S0 divides the positioning space into N regions. Each region constructs a deep neural network that can be used for deviation correction. Assume that the i-th deviation correction network is g i ;

[0025] Step S32: Construct a deviation network selector based on the rough position obtained in step S13 Make a selection of bias network;

[0026] Step S33: Train the bias network and combine the CSI features with the rough position obtained in step S13 Data splicing, here represented as location data Position data As the input of the deviation compensation network, the positioning error obtained in step S4 is As a label, train the bias network selected by the selector; the selected bias network uses the function g i This process is expressed as follows:

[0027]

[0028] Its loss function is expressed as:

[0029]

[0030] The output positioning result is

[0031]

[0032] After adopting the above scheme, the present invention uses a machine learning method to conduct indoor positioning research, which has the advantages of not requiring complex geometric modeling and not requiring the storage of a channel state information fingerprint library. In addition, when constructing a deep neural network, the present invention proposes a deviation correction network, which can reduce the system positioning error caused by the spatial consistency scene change caused by the transition of scattering points, thereby achieving high-precision positioning. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 This is a flow chart of the deviation correction solution proposed by the present invention;

[0034] Figure 2 This is a display diagram of processing channel state information data into an image proposed by the present invention;

[0035] Figure 3 FIG. 1 is a diagram of an experimental environment in one embodiment of the present invention;

[0036] Figure 4 This is a diagram of the positioning network structure of an embodiment of the deviation network proposed in the present invention. DETAILED DESCRIPTION

[0037] like Figure 1 As shown, the present invention discloses a positioning method based on channel state information, which includes: a preparation stage, a coarse positioning stage, a fine positioning stage and a positioning output stage.

[0038] Preparation phase: Prepare the training data set and divide it into regions. Specifically, it includes:

[0039] Step S0: Divide the training data into N regions, with the i-th region being Are i .

[0040] The coarse positioning stage is used to process the received channel state information data into image data, input the feature extraction module, and obtain the coarse position And at the same time get the positioning error Where Pos is the real position coordinate.

[0041] In this embodiment, the experimental environment is a 60*120m indoor factory. The channel state information data received by the receiver is the channel impulse response (CIR) data in the time domain. Its data format is: number of base stations * number of antennas per base station * number of sampling points * real and imaginary parts of each CIR data. In this embodiment, the data is 18*4*256*2. The coarse positioning stage specifically includes:

[0042] Step S11: Data processing, processing the received channel state information data into an image format. Specifically, a matrix transposition method is used: the data in the original embodiment is 18*4*256*2, and is processed into the H×W×C format in the image. Figure 2 As shown, (the number of base stations * the number of antennas of each base station) is used as the height of the image, which is 18*4 in the embodiment, that is, the height of the channel state information image is 72; the two-dimensional data of the real part and the imaginary part of each channel impulse response data is used as the width of the image, which is 2 in the embodiment, that is, the width of the channel state information image is 2; the number of sampling points at each position is used as the number of channels of the channel state information image, and the number of sampling points in the embodiment is 256, so the number of channels here is 256.

[0043] Step S12: Build a positioning network. Figure 1 As shown in the deviation correction scheme flow chart, the channel state information data is first input into the positioning network CEAP to obtain the rough position At the same time, the positioning error is calculated Pos is the real location.

[0044] The positioning grid is composed of three CEAP blocks, each of which consists of a convolutional module and an EN block. Convolutional neural networks, due to their weight sharing and translation invariance, offer significant advantages in image feature extraction. Therefore, when designing the positioning network, we consider that channel state information data can be treated as image data. Transposed channel state information image data is first fed into the convolutional module for preliminary feature extraction, yielding coarse features. This can be expressed as follows:

[0045]

[0046] Where x is the transposed channel state information image data. w is the corresponding convolution filter, b is the bias term, f represents the activation function, and max represents the maximum pooling operation. n represents the number of CNN modules used, which is 1 in this embodiment. LocalFeature is the extracted coarse feature. Assume that the convolution stride is 1 and the convolution kernel size is 3.

[0047] After the initial feature extraction is done by the convolution module, part of the data is input into the EN module, and the other part directly skips the EN module. In order to make the module stackable and not forget the previous information, the module is connected in the form of residuals. Figure 4 As shown in the figure, an EN module consists of Encoder1 and Encoder2. Encoder1 consists of ConvBN and activation (ReLU) layers, and Encoder2 consists of ConvBN. A ConvBN consists of a convolution (Conv) layer and a batch normalization (BN) layer. The process of further feature extraction through the EN module using the features of the convolution module can be expressed as follows:

[0048] DeepFeature=ENBlock(LocalFeature) (7)

[0049] The CNN and EN modules above form a CEAP module. Because its input and output dimensions remain constant, this module can be stacked multiple times. In this embodiment, three stacking steps are used, meaning that feature extraction is performed through three CEAP modules. The final extracted features are fed into a fully connected layer for feature-to-position mapping.

[0050] Step S13: Train the positioning network to obtain a rough position. The channel state information data processed in step S1 is used as the input of the positioning network, and the real position coordinates Pos are used as labels to train the positioning network to obtain a rough position. The positioning network can be represented by the function f, and the process is expressed by the following formula:

[0051]

[0052] In this embodiment, the positioning network adopts CEAP, and formula (1) can be expressed as follows:

[0053]

[0054] Its loss function is shown in Formula 2:

[0055]

[0056] in, Indicates the rough positioning of the i-th position, Pos i is the accurate positioning of the i-th position, N sp is the number of sample points. is L2 regularization, θ 2 is the penalty term, and the parameter η controls the strength of the penalty term.

[0057] Step S14: Calculate the positioning error. After step S13, the rough position can be obtained. At the same time, the real position label Pos is known, which can be obtained by the formula The position error is calculated.

[0058] Fine positioning stage: use selector P ro According to the rough position Selection bias network Ld i , training bias network The positioning error obtained in the coarse positioning stage when training the network For labels, place the And the features of the feature extraction network after the first step are spliced ​​as the i-th deviation network Ld i The fine positioning phase includes the following steps:

[0059] Step S21: Constructing a deviation network. Step S0 divides the training data set into N regions, and constructs a deep neural network for deviation correction in each region. Assume that the i-th deviation correction network is g i In this embodiment, i fully connected layers are used as i deviation compensation networks.

[0060] Step S22: Construct a deviation network selector based on the rough position obtained in step S3 Select the bias network. Assuming the rough position is obtained After that, the probability that the true position coordinates fall into the adjacent area follows a normal distribution. The bias network selector can be described by a normal distribution, and its probability density function can be expressed as the following formula:

[0061]

[0062] Step S23: Construct the input data of the deviation network. The channel state information data in the image format after the transposition process in step S11 is input into a feature extraction network for feature extraction. In this embodiment, the positioning network CEAP constructed in step S12 is used as the feature extraction network after removing the fully connected layer. Assume that the extracted features can be used as CSI. f Indicates that the rough position Combined with the features extracted here, we get:

[0063]

[0064] Step S24: training the bias network. Divided into N regions, each represented by As the input of the deviation network, the positioning error of the i-th region obtained in the coarse positioning stage is As the label of the i-th bias network, train the bias network. Here, the fully connected layer is used as the bias network, and this process can be expressed by the following formula:

[0065]

[0066] After further substitution, we get:

[0067]

[0068] Its loss function is shown in Formula 4:

[0069]

[0070] Positioning output stage:

[0071] Step S31: deviation correction, positioning output. After the rough positioning stage, the rough position is obtained. And the positioning error obtained after the fine positioning stage After that, the fine positioning result is output as the coarse position and positioning error The sum of , deviation correction is performed; then the output positioning result is It can be expressed as follows:

[0072]

[0073] After specific substitution, we get:

[0074]

[0075] The performance indicators used in this invention are root mean square error (RMSE) and mean error (MeanErr), which are defined as follows:

[0076]

[0077]

[0078] in, represents the estimated position of the i-th UE, S i is the actual position of the i-th UE, N UE is the number of UEs, and ||·||2 is the Euclidean distance.

[0079] In summary, the present invention adopts a machine learning method to study indoor positioning, which has the advantages of not requiring complex geometric modeling and not requiring the storage of a channel state information fingerprint library. In addition, when constructing a deep neural network, the present invention proposes a deviation correction network, which can reduce the system positioning error caused by the spatial consistency scene change caused by the transition of scattering points, thereby achieving high-precision positioning.

[0080] The above description is merely an embodiment of the present invention and does not limit the technical scope of the present invention. Therefore, any minor modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention are still within the scope of the technical solution of the present invention.

Claims

1. A positioning method based on channel state information, characterized by: The method comprises: Preparation phase: Divide the training data set into N regions. The area is ; Coarse positioning stage: Process the channel state information data into image data, input it into the feature extraction module, and obtain the coarse position , and at the same time obtain the positioning error , where Pos is the real position coordinate; Fine positioning stage: using selectors According to the rough position Selection biased networks , training bias network Ld: ; The positioning error obtained in the first step when training the network For labels, place the And the features of the feature extraction network after the first step are spliced ​​as the first Bias Network The input of the bias correction network is trained, and the positioning error selected by the error selector is obtained as ; Positioning output stage: The output positioning result is the sum of the rough position and the positioning error, and the deviation is corrected, that is, ; The coarse positioning stage comprises the following steps: Step S11: Data processing, processing the received channel state information data into image data; Step S12: constructing a positioning network for extracting information from the processed channel state information image data, reflecting the relationship mapping between the channel state information and the two-dimensional position coordinates; Step S13: training the positioning network, training the positioning network proposed in step S12, using the processed channel state information data as the input of the positioning network, and using the real position of each position point as the label to train the positioning network; obtaining the rough position ; Positioning network function To express it, this process can be expressed by the following formula: Among them, the loss function is: in, Indicates the Coarse positioning of the position, For the Accurate positioning of the location, is the number of sample points; is L2 regularization, is the penalty term, parameter Control the strength of the penalty term; Step S14: Calculate the difference between the rough position obtained in step S13 and the true position, i.e., the positioning error: , where Pos is the real position coordinate; The fine positioning stage specifically includes the following steps: Step S31: Constructing a deviation network. Step S0 divides the positioning space into N regions. Each region constructs a deep neural network that can be used for deviation correction. Assume that the i-th deviation correction network is ; Step S32: Construct a deviation network selector based on the rough position obtained in step S13 Make a selection of bias network; Step S33: Train the bias network and combine the CSI features with the rough position obtained in step S13 Data splicing, here represented as location data ; Position data As the input of the deviation compensation network, the positioning error obtained in step S4 is As a label, train the bias network selected by the selector; the selected bias network uses the function This process is expressed as follows: Its loss function is expressed as: 。 2. The positioning method based on channel state information according to claim 1, characterized in that: In step S11, the format of the channel state information data is: number of base stations * number of antennas of each base station * number of sampling points * real part and imaginary part of each CSI data; then the specific method of processing the channel state information data into image data is: The number of base stations * the number of antennas of each base station is taken as the height H of the image, the real and imaginary parts of each channel status information data are taken as the width W of the image, and the number of sampling points is taken as the number of channels C of the image.

3. The positioning method based on channel state information according to claim 1, characterized in that: The output positioning result is 。

Images