Low-dimensional channel fingerprint tracking method based on average flow model

Abstract

The invention relates to the technical field of wireless positioning, in particular to a low-dimensional channel fingerprint tracking method based on an average flow model, and the method comprises the steps: firstly obtaining a received signal, and estimating a channel parameter; secondly, mapping channel parameters to a manifold space, and constructing a vector field reflecting dynamic characteristics of a channel; calculating an average value of the vector field in the time window to obtain an average flow field; extracting average flow field features including intensity, direction, divergence and rotation; constructing a Laplacian operator on the channel parameter manifold to perform spectral decomposition; selecting the most significant feature, and calculating the projection coefficient of the average flow field on the feature function; constructing a spectral flow fingerprint vector SFPM; finally, the SFPM is matched with reference fingerprints in a database, and the target position is determined. According to the method, through the average flow model and manifold spectrum analysis, the fingerprint dimension is greatly reduced, the positioning precision is remarkably improved, the environmental adaptability is enhanced, the calculation efficiency is improved, and the anti-noise performance is enhanced.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of wireless positioning technology, in particular to a low-dimensional channel fingerprint tracking method based on an average flow model, which is suitable for high-precision indoor positioning applications in a large-scale MIMO system. BACKGROUND

[0002] With the popularization of 5G technology and the rapid development of the Internet of Things, the demand for indoor high-precision positioning is increasing. Traditional base station positioning technology mainly relies on parameters such as time of arrival (TOA) and angle of arrival (AOA) of signals, and needs a direct path, so the positioning accuracy is limited in a multipath environment.

[0003] In recent years, fingerprint-based positioning methods have received widespread attention. This type of method uses the channel characteristics at each location in the environment as a unique identifier of the location, and achieves positioning by matching the currently measured channel characteristics with a pre-established fingerprint database. In a large-scale MIMO system, the rich spatial information makes high-precision positioning possible.

[0004] However, the existing fingerprint-based positioning methods mainly have the following problems: first, the fingerprint dimension is too high, usually using channel state information (CSI) as the fingerprint, which can easily lead to "dimension disaster" in a large-scale MIMO system, increasing the storage burden and computational complexity; second, the environmental dynamic adaptability is poor, when the environment changes, the fingerprint characteristics often fail, and training data needs to be re-collected; third, the noise sensitivity problem, the positioning accuracy decreases sharply in a low signal-to-noise ratio environment. SUMMARY

[0005] The purpose of the present application is to provide a low-dimensional channel fingerprint tracking method based on an average flow model, which aims to solve the problems of high fingerprint dimension, poor environmental adaptability, and insufficient noise resistance in the prior art.

[0006] The present application provides a low-dimensional channel fingerprint tracking method based on an average flow model, which comprises:

[0007] Obtaining the received signal of a large-scale MIMO-OFDM system, estimating the channel parameters from the received signal, the channel parameters including gain, delay and angle of arrival information;

[0008] Mapping the channel parameters onto a channel parameter manifold, constructing an average flow model, the construction of the average flow model comprising:

[0009] Defining a vector field representing the dynamic changes of the channel parameters based on the channel parameter manifold;

[0010] Calculating the average value of the vector field within a time window to obtain an average flow field;

[0011] extracting features of the average flow field to form average flow features;

[0012] performing spectral dimension reduction on the average flow features to construct a low-dimensional fingerprint, the performing spectral dimension reduction on the average flow features comprising:

[0013] constructing a Laplacian operator on the channel parameter manifold;

[0014] calculating eigenfunctions and eigenvalues of the Laplacian operator;

[0015] calculating projection coefficients of the average flow field on the eigenfunctions;

[0016] constructing a spectral flowprint vector SFPM based on the projection coefficients and the eigenvalues;

[0017] matching the SFPM with reference fingerprints in a fingerprint database to determine a target location.

[0018] As a preference, the mapping the channel parameters onto a channel parameter manifold comprises:

[0019] constructing a Riemannian metric tensor reflecting correlations and sensitivities among channel parameters;

[0020] establishing a local coordinate system based on the Riemannian metric tensor so that similar channel features are closer in distance on the manifold;

[0021] optimizing the Riemannian metric tensor according to channel physical properties to maintain periodicity of angular parameters.

[0022] As a preference, the defining a vector field representing dynamic changes of channel parameters comprises:

[0023] collecting multiple sets of channel parameters over a time series;

[0024] calculating channel parameter change rates at adjacent time points;

[0025] mapping the change rates to tangent spaces of the channel parameter manifold to form a vector field;

[0026] analyzing divergence and curl of the vector field to capture diffusion and rotation properties of channel changes.

[0027] As a preference, the calculating an average value of the vector field within a time window comprises:

[0028] setting a time window length T;

[0029] collecting vector field samples within the time window;

[0030] performing time accumulation on all samples;

[0031] The normalization processing is performed to obtain an average flow field;

[0032] The length of the time window T is determined according to the dynamic characteristics of the environment, and the value range is 1 to 5 seconds.

[0033] As a preferred, the feature extraction of the average flow field includes:

[0034] The average flow field intensity is calculated to represent the degree of channel change;

[0035] The average flow field direction is calculated to represent the main change trend;

[0036] The average flow field divergence is calculated to represent the signal energy diffusion characteristics;

[0037] The average flow field curl is calculated to represent the channel phase rotation characteristics;

[0038] The above features are combined to form an average flow feature set.

[0039] As a preferred, the construction of the Laplace operator on the channel parameter manifold includes:

[0040] The weight function is constructed to reflect the importance of different regions;

[0041] Based on the Riemann metric tensor of the channel parameter manifold, the Laplace-Beltrami operator is constructed;

[0042] The discrete approximation method is used to convert the continuous operator into a discrete matrix form;

[0043] The matrix structure is optimized to ensure the calculation efficiency and numerical stability.

[0044] As a preferred, the calculation of the eigenfunctions and eigenvalues of the Laplace operator includes:

[0045] The characteristic equation is solved to obtain the eigenfunction set and the corresponding eigenvalues;

[0046] The eigenfunctions are sorted according to the eigenvalue size;

[0047] The eigenvalue decay curve is analyzed to determine the truncation threshold;

[0048] The first k most significant eigenfunctions and eigenvalues are selected, where k is in the range of 10 to 20.

[0049] As a preferred, the construction of the spectral flow fingerprint vector SFPM based on the projection coefficient and the eigenvalue includes:

[0050] For each eigenfunction, the projection coefficient of the average flow field is calculated;

[0051] The weight coefficient is designed according to the eigenvalue, and important features are emphasized.

[0052] The weighted projection coefficients are combined to form the SFPM fingerprint vector.

[0053] The SFPM fingerprint vector is normalized to enhance stability.

[0054] The SFPM fingerprint vector is quantized and encoded to reduce storage requirements.

[0055] As preferred, the matching of the SFPM with reference fingerprints in the fingerprint database to determine the target position comprises:

[0056] The distances between the SFPM to be positioned and the SFPMs of each reference point in the database are calculated.

[0057] The n closest reference points are selected as candidates, where n is in the range of 3 to 5.

[0058] The weights of each candidate point are calculated based on the SFPM distance.

[0059] The weighted average is performed to obtain the target position estimate.

[0060] The trajectory smoothing algorithm is applied to improve the stability of the position estimate.

[0061] As preferred, it also includes a dynamic environment adaptation mechanism:

[0062] The change in the distribution of SFPM features is monitored.

[0063] The positioning error trend is analyzed.

[0064] The area that needs to be updated is identified.

[0065] Resampling is performed on the changed area to update the SFPM of the affected reference points.

[0066] The spectral filter parameters and feature weights are dynamically adjusted.

[0067] The matching algorithm parameters are optimized to adapt to environmental changes.

[0068] By representing the channel parameters as points in a manifold space, the present application introduces an average flow model to capture the dynamic characteristics of the channel, and combines manifold spectral analysis for effective dimension reduction, to construct a robust, low-dimensional spectral flow fingerprint vector (SFPM). Compared with the prior art, the present application has the following beneficial effects:

[0069] 1. The dimension of the fingerprint is greatly reduced: the SFPM fingerprint dimension of the present application is fixed at 10-20 dimensions, and does not increase with the number of multi-paths, reducing storage requirements by 70-85% compared with traditional methods.

[0070] 2. Significantly improve positioning accuracy: By using the average flow model and manifold spectrum analysis, the essential characteristics of the channel are effectively captured, and the positioning reliability at one-meter accuracy is improved from 90% to 98%.

[0071] 3. Enhance environmental adaptability: The average flow field can extract the statistical characteristics of channel changes and has strong robustness to environmental changes. It does not need to be retrained under 80% of environmental change scenarios.

[0072] 4. Improve computational efficiency: After dimensionality reduction, the computational complexity is significantly reduced, the positioning delay is reduced by about 75%, and the energy consumption is reduced by about 70%.

[0073] 5. Enhance noise resistance: The filtering mechanism based on manifold spectrum analysis can effectively suppress the influence of noise, and the performance degradation is controlled within 20% under low SNR environment, and the working SNR range is expanded by about 8dB. BRIEF DESCRIPTION OF DRAWINGS

[0074] Figure 1 The overall flowchart of the method of the present application.

[0075] Figure 2 The flowchart for constructing the average flow model.

[0076] Figure 3 The schematic diagram for dimensionality reduction and SFPM fingerprint construction. DETAILED DESCRIPTION

[0077] Please refer to Figures 1-3 , the present application will be further described in detail below in combination with the drawings and specific embodiments.

[0078] As Figure 1 shown, the low-dimensional channel fingerprint tracking method based on the average flow model mainly includes five core steps: channel parameter estimation, channel parameter manifold representation, average flow model construction, dimensionality reduction and fingerprint construction, and position matching and tracking. Through this series of steps, the complete processing flow from signal reception to position determination is realized.

[0079] In one embodiment of the present application, the received signal of the large-scale MIMO-OFDM system is first obtained, and the channel parameters including gain, delay and angle of arrival information are estimated from the received signal.

[0080] In the large-scale MIMO-OFDM system, the base station is equipped with N antennas, the user equipment is equipped with a single antenna, the system bandwidth is 20MHz, and the number of subcarriers is 2048. The received signal model can be represented as:

[0081] ,

[0082] wherein represents the received signal matrix, represents a channel frequency response matrix, represents a transmitted pilot matrix, represents a noise matrix, is the number of antennas, is the number of subcarriers.

[0083] Based on the received signal, the channel parameters are estimated using the SAGE (Space Alternating Generalized Expectation-Maximization) algorithm. The SAGE algorithm is an iterative algorithm that can effectively handle the problem of multi-path resolution and is suitable for use in large-scale MIMO systems. After processing by the SAGE algorithm, a set of multi-path channel parameters can be obtained:

[0084]

[0085] wherein represents the complex gain of the th path, represents the time delay, represents the angle of arrival, represents the number of multi-paths. In practical applications, is usually taken to be 10-20, depending on the complexity of the environment.

[0086] As shown in Figure 2 , the present application maps the channel parameters onto a channel parameter manifold. The conventional method simply regards the channel parameters as vectors in Euclidean space, ignoring the geometric relationship between the parameters. The present application models the channel parameter space as a manifold structure from a geometric perspective.

[0087] Preferably, in the channel parameter manifold representation, a Riemannian metric tensor is constructed, which reflects the correlation and sensitivity between the channel parameters. The Riemannian metric tensor defines the distance metric on the manifold, for two points and on the manifold, the geodesic distance between them is:

[0088]

[0089] wherein represents the path connecting and , represents the tangent vector of the path, represents the metric tensor at the point , represents the minimum value among all paths connecting and ,​​ represents the integral from parameter to .

[0090] The construction of the Riemannian metric tensor is based on the physical properties of channel parameters. For example, for the angle parameter, its periodic property is considered; for the delay parameter, the linear relationship with distance is considered; for the gain parameter, the inverse square relationship with distance is considered. Based on a large amount of data analysis, the typical construction of the Riemannian metric tensor is:

[0091] ,

[0092] where represents the weight coefficient of the gain parameter, represents the weight coefficient of the delay parameter, represents the weight coefficient of the angle parameter, which is determined according to the actual environmental characteristics. For example, in a typical indoor environment, , , .

[0093] Based on the Riemannian metric tensor, a local coordinate system is established, so that similar channel characteristics are close in distance on the manifold. At each point of the manifold, a tangent space is defined, which contains all possible directions of change of channel parameters. The vector on the tangent space represents a small change in channel parameters.

[0094] Preferably, the Riemannian metric tensor is optimized according to the physical properties of the channel, maintaining the periodicity of the angle parameter. The angle parameter has periodicity, and the distance between and is very far in Euclidean space, but in fact they represent similar directions. By introducing appropriate metric structures on the manifold, this periodicity can be correctly handled.

[0095] In practical applications, the dimension of the channel parameter manifold is , where is the number of multipaths, and the additional 2 dimensions represent the two-dimensional position coordinates of the user equipment.

[0096] As shown in Figure 2 , based on the channel parameter manifold, the present application defines a vector field representing the dynamic change of channel parameters. The vector field maps each point on the manifold to a vector in its tangent space, i.e. , for each point , .

[0097] In embodiments of the application, the vector field is constructed based on time series observations. Specifically, multiple sets of channel parameters are collected over a time series, the rate of change of channel parameters at adjacent time points is calculated, the rate of change is mapped to the tangent space of the channel parameter manifold, and a vector field is formed.

[0098] For time points and , the change of channel parameters can be represented as:

[0099] ,

[0100] where represents the amount of change of channel parameters in time interval , represents the channel parameter at time , represents the channel parameter at time .

[0101] Mapping to the tangent space, the value of the vector field at point is obtained:

[0102] ,

[0103] where is a function mapping channel parameters to the manifold, represents the derivative of channel parameter change after mapping to the manifold, i.e., the rate of change.

[0104] Preferably, the divergence and curl of the vector field are analyzed to capture the diffusion and rotation characteristics of channel changes. The divergence of the vector field represents the diffusion characteristics of signal energy, and the curl represents the rotation characteristics of channel phase. In complex multipath environments, these characteristics are of great value for distinguishing channel characteristics at different locations.

[0105] Channel characteristics can have random fluctuations in a short time, but exhibit certain statistical regularities in space. The present application captures the statistical characteristics of channel changes by calculating the average value of the vector field within a time window to obtain an average flow field.

[0106] Specifically, the length of the time window T is set, the vector field samples are collected within the time window, the time accumulation is performed on all samples, and the normalization processing is performed to obtain the average flow field AF:

[0107] ,

[0108] where represents the average flow field value at point , represents the point at time vector field value, denotes the length of time window, denotes the integral from time 0 to denotes the integral from time 0 to denotes the normalization factor. In practical applications, the length of time window is determined according to the dynamic characteristics of the environment, and the value range is 1 to 5 seconds. For a stable environment, a longer time window can be selected; for a rapidly changing environment, a shorter time window should be selected.

[0109] After obtaining the average flow field, its features are extracted to form an average flow feature set. Mainly including:

[0110] 1) Calculate the average flow field intensity, which represents the degree of channel change:

[0111] ,

[0112] wherein denotes the average flow field intensity at point , denotes the Euclidean norm of the average flow field vector.

[0113] 2) Calculate the average flow field direction, which represents the main change trend:

[0114] ,

[0115] wherein denotes the average flow field direction at point , denotes the average flow field vector at point , denotes the Euclidean norm of the vector, denotes the normalized unit direction vector.

[0116] 3) Calculate the average flow field divergence, which represents the diffusion characteristics of signal energy:

[0117] ,

[0118] wherein, denotes the average flow field divergence at point , denotes the divergence operator, denotes the average flow field vector at point .

[0119] 4) Calculate the average flow field curl, which represents the rotation characteristics of the channel phase:

[0120] ,

[0121] wherein denotes the average flow field curl at point The average flow curl at that location, Represents the curl operator, Point The average flow field vector at that location.

[0122] 5) Combining the above features forms the average flow feature set:

[0123] ,

[0124] in, Point The average flow characteristic set at the location, Indicates the average flow field intensity. Indicates the direction of the average flow field. Represents the average flow field divergence. This represents the average flow field curl.

[0125] Average flow characteristics can effectively capture the stable characteristics of the channel as its spatial distribution changes, and have strong robustness to environmental changes.

[0126] like Figure 3 As shown, to further reduce the fingerprint dimensionality, this invention constructs a Laplace operator on the channel parameter manifold. The Laplace-Beltrami operator is a generalization of the differential operator on the manifold and plays an important role in function analysis and spectral geometry.

[0127] Preferably, a weighting function is constructed to reflect the importance of different regions. The amount of information in different regions may vary across the channel parameter manifold. For example, strong multipath regions typically contain more location-related information. By introducing a weighting function, the characteristics of important regions can be emphasized.

[0128] Based on the Riemannian metric tensor of the channel parameter manifold, the Laplace-Beltrami operator can be expressed as:

[0129] ,

[0130] in, Representation function In manifold The result of the Laplace-Beltrami operator on the above, This represents the Riemannian metric tensor. Represent its determinant, Tensor The elements of the inverse matrix, Describe a function on a manifold. and Represents coordinates and The partial derivatives, This indicates all coordinate indicesThe sum of and .

[0131] In practical applications, a discrete approximation method is used to convert the continuous operator into a discrete matrix form. Commonly used discretization methods include the finite element method and the graph Laplacian method. The present application preferably uses the graph Laplacian method to construct a weight graph where the vertex set corresponds to the sampling points, the edge set represents the connection relationship between points, and the weight matrix represents the weight of the edge.

[0132] The discrete Laplacian matrix can be represented as:

[0133] ,

[0134] where represents the Laplacian matrix, represents the degree matrix, represents the weight matrix, represents the degree of the vertex , i.e., the sum of the weights of all edges connected to the vertex . To ensure numerical stability, the normalized form is usually used:

[0135] ,

[0136] where represents the normalized Laplacian matrix, represents the negative two-thirds power of the degree matrix , and represents the identity matrix.

[0137] Based on the Laplacian operator, spectral decomposition is performed to extract the main components of the channel features. The characteristic equation is solved:

[0138] ,

[0139] where represents the eigenvalue, represents the corresponding eigenfunction.

[0140] The eigenfunction set and the corresponding eigenvalue are obtained.

[0141] In the discrete case, the characteristic equation becomes:

[0142] ,

[0143] where represents the eigenvector, and the corresponding eigenfunction discrete sampling values.

[0144] In the feature selection strategy, the feature functions are sorted according to the feature values, the decay curve of the feature values is analyzed, and the truncation threshold is determined. In practical applications, the feature values usually show a rapid decay trend, and the top most significant feature functions and feature values can capture the main information. Preferably, The value range is 10 to 20.

[0145] In the feature weight design, the weight coefficient is designed based on the feature value:

[0146] ,

[0147] wherein w represents the weight of the i-th feature, λ represents the i-th feature value, μ represents the adjustment parameter, and exp represents the exponential decay weight based on the feature value. The adjustment parameter is usually 0.1-0.5. This design makes the feature function corresponding to a smaller feature value (low-frequency information) obtain a larger weight, emphasizing the global structural features. The projection coefficient of the average flow field on the feature function is calculated:

[0148] ,

[0149] ,

[0150] wherein φ represents the projection coefficient of the average flow field on the feature function , represents the inner product operation, represents the integration of the volume element on the manifold , represents the dot product of the vector and the function value . In the discrete case, the projection coefficient calculation is simplified as:

[0151]

[0152] ,

[0153] wherein x represents the sampling point, w represents the weight of the sampling point , and represents the summation of all sampling points. In the projection coefficient optimization aspect, the present application adopts the following strategies:

[0154] ​​​​

[0155] 1) Remove redundant projection components, calculate the correlation between projection coefficients, and delete height-related terms.

[0156] 2) Standardization processing: The projection coefficients are normalized to the [0,1] interval to enhance stability.

[0157] 3) Quantization encoding: The projection coefficients are quantized according to the accuracy requirements. Each coefficient is usually represented by 8 bits to reduce storage requirements.

[0158] Constructing the spectral flow fingerprint vector SFPM based on projection coefficients and eigenvalues:

[0159] ,

[0160] in, Point Spectral flow fingerprint vector at the location, Indicates the first The weights of each feature, The mean flow field is represented in the first... Projection coefficients on each characteristic function This represents the weighted projection coefficients. This indicates a vector concatenation operation.

[0161] SFPM fingerprints have the following characteristics:

[0162] 1) Low dimensionality: k is typically 10-20, much lower than traditional methods;

[0163] 2) High discriminative power: Captures both global and local structural features of the channel;

[0164] 3) High robustness: It has a strong adaptability to noise and environmental changes;

[0165] 4) High storage efficiency: Each SFPM fingerprint occupies approximately 100-200 bytes;

[0166] To further improve the stability of SFPM, the SFPM fingerprint vector is standardized:

[0167] ,

[0168] in, This represents the standardized SFPM fingerprint vector. Represents the original SFPM fingerprint vector. Denotes the Euclidean norm of the original vector. This indicates a normalization operation.

[0169] Furthermore, to reduce storage requirements, quantization encoding is performed on the SFPM fingerprint vector. Depending on the actual accuracy requirements, each component can be quantized using 8 bits or 16 bits, reducing storage overhead while maintaining discriminability.

[0170] Reference point signals are collected at certain intervals (usually 1-2 meters) within the target area. A complete processing flow is performed on each reference point to construct an SFPM fingerprint, and the reference point location coordinates and corresponding SFPM fingerprints are stored in the database.

[0171] To accelerate the query process, this invention constructs an index structure based on a KD-tree. A KD-tree is a spatial partitioning tree suitable for nearest neighbor searches in multidimensional space, significantly improving matching efficiency.

[0172] For the user equipment to be located, the system matches the SFPM with a reference fingerprint in the fingerprint database to determine the target location. Specifically, this includes:

[0173] 1) Calculate the distance between the SFPM of the point to be located and the SFPM of each reference point in the database:

[0174] ,

[0175] in Indicates the point to be located With reference point SFPM distance between This represents the SFPM fingerprint of the point to be located. Indicates reference point SFPM fingerprint, This represents the Euclidean distance norm.

[0176] In practical applications, cosine similarity is the preferred metric.

[0177] ,

[0178] in Indicates the point to be located With reference point The cosine distance between them This represents the dot product of two SFPM vectors. and Let represent the Euclidean norms of the two vectors respectively. Represents cosine similarity. Represents the cosine distance.

[0179] 2) Select the n nearest reference points as candidates, where n ranges from 3 to 5:

[0180] ,

[0181] wherein, represents the set of candidate reference points,

[0182] represents the distance between the candidate reference point and the to-be-located point p,

[0183] 3) Calculate the weight of each candidate point based on SFPM distance:

[0184] ,

[0185] wherein represents the weight of the candidate reference point , represents the exponential decay function based on distance, represents the adjustment parameter, which controls the steepness of the weight distribution, represents the sum of all candidate point exponential decay values, used for normalization, represents the normalized weight. The adjustment parameter is usually taken as 2-5.

[0186] 4. Perform weighted average to get the target position estimate:

[0187] ,

[0188] wherein, represents the estimated target position, represents the weight of the candidate reference point , represents the position coordinates of the candidate reference point, represents the weighted average operation.

[0189] 5) Apply trajectory smoothing algorithm to improve the stability of position estimation. The present invention adopts an improved Kalman filter algorithm to combine the position estimation with the motion model, effectively suppressing random fluctuations.

[0190] To deal with the challenges brought by environmental changes, the present invention realizes a dynamic environment adaptation mechanism, including:

[0191] 1) Monitor the change of SFPM feature distribution: regularly analyze the distribution characteristics of SFPM fingerprint, and calculate the distribution offset index:

[0192] ,

[0193] wherein, represents the distribution offset measure, represents the current SFPM distribution and the reference SFPM distribution​​ KL divergence (Kullback-Leibler divergence) between two probability distributions to measure the difference between the two probability distributions.

[0194] 2) Analyze positioning error trends: Monitor system positioning errors, when the error exceeds the threshold (usually set to 1.5 meters) and lasts for a certain time (usually set to 10 minutes), trigger environment change detection.

[0195] 3) Identify areas that need to be updated: Based on the spatial distribution of positioning errors, identify areas of environmental change, and prioritize areas that are more affected.

[0196] 4) Perform resampling on the changed area, update the affected reference points SFPM, and keep the reference points in other areas unchanged, achieving incremental update.

[0197] 5) Dynamically adjust the spectral filter parameters and feature weights: Adjust system parameters according to environmental characteristics, such as enhancing filter strength in noisy environments and emphasizing high-frequency features in dynamic environments.

[0198] 6) Optimize matching algorithm parameters: Dynamically adjust the number of candidate points n and the weight calculation parameter γ according to the complexity of the environment.

[0199] Through the above mechanisms, the system can adapt to environmental changes and maintain long-term positioning accuracy. Experiments show that in 80% of the environmental change scenarios, the system does not need to be completely retrained, but only needs to be updated incrementally to restore positioning accuracy.

[0200] To verify the effectiveness of the present application, a large number of experimental tests were conducted in a typical indoor environment. The test scene is a 20m x 30m office area, using a large-scale MIMO system with 128 antennas, system bandwidth 20MHz, and working frequency 3.5GHz.

[0201] The positioning accuracy comparison results of the present application and the prior art show that the reliability of the present application method reaches 98% at one meter accuracy, which is significantly better than the 90% of the traditional method. The median positioning error is reduced from 0.5 meters to 0.25 meters, showing excellent positioning performance.

[0202] Compared with the traditional method, the SFPM fingerprint dimension of the present application is fixed at 10-20 dimensions, which does not grow with the number of multipaths L, while the traditional GDAPM fingerprint dimension is 3L. In a complex environment (L=20), the storage overhead is reduced by about 85%. For a test area of 20m x 30m, sampling at 1 meter intervals, the traditional method requires 12MB of storage space, while the present application only requires 1.8MB.

[0203] In terms of positioning calculation efficiency, the position matching delay of the application is about 1.2 milliseconds, which is significantly lower than 5 milliseconds of the traditional method. The main reason is that the fingerprint dimension of the application is low, and the efficient KD tree index structure is adopted. In addition, the energy consumption of the application is only 30% of the traditional method, which is particularly suitable for mobile device applications.

[0204] In the environmental change test, environmental changes are simulated by moving furniture, adding obstacles, etc. The experimental results show that the positioning accuracy of the traditional method decreases by more than 80% after the environmental change, and the training data needs to be completely re-collected. The positioning accuracy of the application method decreases within 30% after the environmental change, and through the dynamic environmental adaptation mechanism, only the incremental update of the changed area is required to restore the positioning accuracy.

[0205] By adding Gaussian white noise of different intensities, the performance of the system under different signal-to-noise ratios is tested. The experimental results show that in the low signal-to-noise ratio (SNR<0dB) scene, the positioning accuracy of the traditional method decreases by more than 60%, while the accuracy of the application method decreases within 20%, showing excellent anti-noise performance.

[0206] The application proposes a low-dimensional channel fingerprint tracking method based on the average flow model. Through channel parameter manifold representation, average flow model construction and spectral dimension reduction technology, a low-dimensional, efficient and robust position fingerprint representation and tracking is realized. This method greatly reduces the storage requirement and calculation complexity while maintaining high-precision positioning, and significantly improves the environmental adaptability and anti-noise performance. The experimental results show that the application is superior to the prior art in terms of positioning accuracy, storage efficiency, calculation overhead and environmental adaptability, and provides an effective solution for high-precision indoor positioning in large-scale MIMO systems.

[0207] The above-described embodiments only express the specific implementation of the application, and the description is more specific and detailed, but it cannot be understood as a limitation on the scope of the patent. It should be noted that for ordinary skilled in the art, without departing from the concept of the application, a number of modifications and improvements can be made, which are within the scope of protection of the application.

Claims

1. A low-dimensional channel fingerprinting method based on an average flow model, characterized in that, include: Acquire the received signal of a large-scale MIMO-OFDM system, and estimate channel parameters from the received signal, the channel parameters including gain, delay and angle of arrival information; Mapping the channel parameters onto a channel parameter manifold to construct an average flow model, wherein constructing the average flow model includes: Based on the channel parameter manifold, a vector field representing the dynamic changes of the channel parameters is defined; The average flow field is obtained by calculating the average value of the vector field within the time window; Extract the features of the average flow field to form average flow features; The average flow features are subjected to spectral dimensionality reduction to construct a low-dimensional fingerprint. The spectral dimensionality reduction of the average flow features includes: Construct a Laplace operator on the channel parameter manifold; Calculate the characteristic functions and eigenvalues ​​of the Laplacian operator; Calculate the projection coefficients of the average flow field onto the characteristic function; Construct a spectral flow fingerprint vector SFPM based on the projection coefficients and the eigenvalues; The SFPM is matched with a reference fingerprint in the fingerprint database to determine the target location.

2. The method according to claim 1, characterized in that, The step of mapping the channel parameters onto the channel parameter manifold includes: Construct a Riemann metric tensor that reflects the correlation and sensitivity between channel parameters; A local coordinate system is established based on the Riemannian metric tensor, so that similar channel features are closer together on the manifold; The Riemann metric tensor is optimized based on the channel's physical characteristics, while preserving the periodicity of the angle parameters.

3. The method according to claim 1, characterized in that, The definition represents a vector field representing the dynamic changes of channel parameters, including: Multiple sets of channel parameters were collected over a time series. Calculate the rate of change of channel parameters at adjacent time points; The rate of change is mapped to the tangent space of the channel parameter manifold to form a vector field; The divergence and curl of the vector field are analyzed to capture the diffusion and rotation characteristics of channel variations.

4. The method according to claim 1, characterized in that, The calculation of the average value of the vector field within the time window includes: Set the time window length T; Vector field samples are acquired within the time window; Cumulative execution time for all samples; The average flow field is obtained by normalization. The time window length T is determined based on the dynamic characteristics of the environment, and its value ranges from 1 to 5 seconds.

5. The method according to claim 1, characterized in that, The features extracted from the average flow field include: Calculate the average flow field intensity, which represents the degree of drastic change in the channel. Calculate the direction of the average flow field to represent the main trend of change; Calculate the average flow field divergence, which represents the signal energy diffusion characteristics; Calculate the average flow field curl, representing the channel phase rotation characteristic; Combining the above features forms the average flow feature set.

6. The method according to claim 1, characterized in that, The construction of the Laplace operator on the channel parameter manifold includes: Construct a weighting function to reflect the importance of different regions; Based on the Riemann metric tensor of the channel parameter manifold, a Laplace-Beltramian operator is constructed. A discrete approximation method is used to convert continuous operators into discrete matrix form; Optimize the matrix structure to ensure computational efficiency and numerical stability.

7. The method according to claim 1, characterized in that, The calculation of the characteristic function and eigenvalues ​​of the Laplace operator includes: Solve the characteristic equation to obtain the set of characteristic functions and their corresponding eigenvalues; Sort the feature functions according to the size of their feature values; Analyze the eigenvalue decay curves to determine the cutoff threshold; Select the k most significant feature functions and feature values, where k ranges from 10 to 20.

8. The method according to claim 1, characterized in that, The construction of the spectral flow fingerprint vector SFPM based on the projection coefficients and the eigenvalues ​​includes: For each characteristic function, calculate the projection coefficient of the mean flow field; Weighting coefficients are designed based on eigenvalues, emphasizing important features; The weighted projection coefficients are combined to form the SFPM fingerprint vector; The SFPM fingerprint vector is standardized to enhance its stability; Quantization encoding is performed on the SFPM fingerprint vector to reduce storage requirements.

9. The method according to claim 1, characterized in that, The step of matching the SFPM with a reference fingerprint in the fingerprint database to determine the target location includes: Calculate the distance between the SFPM of the point to be located and the SFPM of each reference point in the database; Select the n nearest reference points as candidates, where n ranges from 3 to 5; The weights of each candidate point are calculated based on the SFPM distance. A weighted average is applied to obtain an estimate of the target location; A trajectory smoothing algorithm is applied to improve the stability of position estimation.

10. The method according to claim 1, characterized in that, It also includes a dynamic environmental adaptation mechanism: monitoring changes in SFPM feature distribution; analyzing positioning error trends; identifying areas that need updating; resampling the changed areas to update the affected reference point SFPM; dynamically adjusting spectral filter parameters and feature weights; and optimizing matching algorithm parameters to adapt to environmental changes.

Images