User positioning method based on channel state information fingerprint data

Abstract

The invention provides a user positioning method based on channel state information fingerprint data. The method comprises the following steps: collecting CSI data on a sampling point, and constructing a CSI fingerprint database; performing feature extraction on the CSI fingerprint data in the CSI fingerprint database by using a wavelet decomposition method to obtain CSI features; training the BLS through nonlinear mapping by using the CSI features to obtain a trained BLS, and training the KAN through an LBFGS optimization algorithm by using the CSI features to obtain a trained KAN; and inputting the CSI fingerprint data of the current position of the user into the trained KAN and BLS, respectively outputting the position estimation information of the user by the trained KAN and BLS, and carrying out temperature coefficient fusion on the position estimation information of the user respectively output by the trained KAN and BLS to obtain a final positioning result of the user. According to the method, noise is suppressed by designing a hybrid wavelet denoising strategy, efficient feature extraction is performed by using the KAN, and a temperature-driven BLS-KAN dynamic fusion mechanism is introduced, so that high-precision, high-adaptability and efficient-calculation indoor positioning is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of indoor positioning, and in particular to a user positioning method based on Channel State Information (CSI) fingerprint data. Background Art

[0002] Indoor positioning technology plays an important role in smart buildings, industrial Internet of Things, and Integrated Sensing and Communication (ISAC). The accuracy and reliability of indoor positioning directly affect the user experience and the overall performance of these systems. Among many positioning solutions, fingerprint positioning technology based on WiFi signals has become a research focus in academia and industry, mainly because it can utilize existing infrastructure without the need for dedicated hardware. Fingerprint positioning technology based on WiFi signals usually includes two key stages: the offline stage, where signal measurements are collected through grid sampling in the target area to build a comprehensive fingerprint database; the online stage, where real-time user data is matched with the pre-collected database to estimate the user's location. With the rapid development of ISAC technology, fingerprint-based positioning technology has attracted more and more attention from researchers due to its potential in providing accurate, efficient, and scalable indoor positioning solutions.

[0003] With the continuous evolution of indoor positioning technology, CSI has gradually become a more effective alternative to Received Signal Strength (RSS) in fingerprint positioning. Early positioning methods mainly relied on RSS, but its single-dimensional scalar feature is extremely vulnerable to multipath effects and dynamic occlusion, often resulting in large positioning errors. CSI can provide more refined features by analyzing the amplitude and phase data of Orthogonal Frequency Division Multiplexing (OFDM) subcarriers, providing a more reliable solution for high-precision indoor positioning.

[0004] In the practical application of indoor positioning technology, the selection of positioning algorithms is crucial, which is directly related to the final positioning performance. The CSI-based fingerprint positioning technology mainly includes fingerprint matching based on traditional machine learning and end-to-end methods based on deep learning. However, these methods still face significant challenges. Machine learning-based fingerprint positioning usually relies on shallow models trained offline for location classification. But there are usually problems such as high computational complexity, being less practical in large-scale databases, and strong dependence on the environment. Deep learning models require a large amount of computing resources and are difficult to deploy in actual application scenarios. Traditional methods often require manual parameter adjustment, which limits their adaptability in dynamic environments. Moreover, existing methods have difficulties in dealing with spatio-temporal frequency correlations and lack effective noise filtering mechanisms.

[0005] Existing technologies lack effective noise suppression means when processing CSI data, which affects the positioning accuracy. Traditional machine learning algorithms have drawbacks such as insufficient feature extraction, low positioning accuracy, and difficulty in applying to large-scale scenarios; while deep learning methods not only require a large amount of sample data, computing resources, and time costs during the training process, but also are prone to overfitting. Summary of the Invention

[0006] The present invention provides a user positioning method based on CSI fingerprint data to effectively perform indoor user positioning.

[0007] To achieve the above object, the present invention adopts the following technical solutions.

[0008] A user positioning method based on channel state information fingerprint data, comprising:

[0009] Collect CSI data at sampling points and construct a CSI fingerprint database;

[0010] Use the wavelet decomposition method to denoise the CSI fingerprint data in the CSI fingerprint database, and use KAN to extract features from the denoised CSI fingerprint data to obtain CSI features;

[0011] Use the CSI features to train BLS through nonlinear mapping to obtain a trained BLS, and use the CSI features to train KAN through the LBFGS optimization algorithm to obtain a trained KAN;

[0012] Input the CSI fingerprint data of the user's current location into the trained KAN. The trained KAN outputs the user's location estimation information. Input the CSI fingerprint data of the user's current location into the trained BLS. The trained BLS outputs the user's location estimation information. The location estimation information of the user output by the trained KAN and BLS respectively is fused by the temperature coefficient to obtain the final user positioning result.

[0013] Preferably, collecting CSI data at the sampling points to construct a CSI fingerprint database includes:

[0014] Divide the user's indoor positioning area into grids. Use the device installed with the CSI tool to carry out CSI data collection work in the user's indoor positioning area. Take the user as the sampling point. The CSI tool randomly collects the original CSI data that conforms to the normal distribution in the mobile hotspot WiFi signals of the sampling points in the grid area.

[0015] Design each layer of the DeepBLS network into a width learning block. Adopt the training method with the inter-layer residual scaled by the learning rate as the target output. Each layer is trained only once. Use the DeepBLS network to initially construct the original CSI fingerprint database with the original CSI data. Then, further extract features from the original CSI data with the help of the generative adversarial network and update the original CSI fingerprint database by secondary prediction. Select the DeepBLS confidence coefficient as 0.6 and the GAN confidence coefficient as 0.4. Construct the final CSI fingerprint database in the way of confidence weighting.

[0016] Preferably, the method of using wavelet decomposition to denoise the CSI fingerprint data in the CSI fingerprint database includes:

[0017] Select Daubechies4 wavelet as the basis function to perform multi-layer decomposition on the CSI fingerprint data in the final CSI fingerprint database. Transform the CSI data into the time-frequency domain. Perform four-level wavelet decomposition on each CSI channel to obtain the low-frequency approximation coefficient A L and the high-frequency detail coefficient D L , formula:

[0018]

[0019] where A L is the low-frequency approximation component, and D l contains the high-frequency detail coefficients of different decomposition levels;

[0020] Adopt the method based on VisuShrink to calculate the adaptive threshold τ. The calculation formula of the adaptive threshold τ is:

[0021]

[0022] where σ is the median absolute deviation of the detail coefficients, and N is the length of the coefficient vector;

[0023] Apply soft thresholding to the high-frequency detail coefficients D l The formula is:

[0024] D′ l = sign(D l )·max(|D l | - τ, 0) (3)

[0025] Place A L at the beginning of the coefficient list, and arrange D' l in sequence. Use the pywt.waverec function to perform inverse wavelet transform to reconstruct the signal with the coefficient list and the Daubechies4 wavelet basis as parameters. If the length of the reconstructed signal is inconsistent with the original CSI signal, perform truncation or padding processing to obtain the denoised CSI signal X′.

[0026] Preferably, the method for using KAN to extract features from the denoised CSI fingerprint data to obtain CSI features includes:

[0027] According to KAT, represent the complex continuous function as a combination of a series of simple functions. The formula is:

[0028]

[0029] where x = (x1,…,x n ) is the preprocessed CSI data feature vector, f(x) represents the continuous mapping function from CSI features to spatial positions, n is the input feature dimension, q is the summation term index, φ q (·) and are both univariate continuous functions, which are used to combine intermediate results and act on the input feature x p respectively;

[0030] Use KAN to process the CSI data X′. KAN converts the input activation x l to the output activation x l+1 through sum function operations, that is:

[0031]

[0032] where φ l,j,i is an adaptive non-linear function, parameterized by B-spline as:

[0033] φ(x) = w b b(x) + w s ∑ i c iB i (x) (6)

[0034] wherein, B i (x) is a basis spline function, c i is a trainable coefficient, w b and w s are weight parameters respectively;

[0035] b(x) = silu(x) = x / (1 + e -x ) (7)

[0036] When using KAN for feature extraction, the preprocessed CSI data X′ is used as the input of KAN. The dimension n of KAN corresponds to the number of CSI features. Based on formula (4), the nested combination of 2n + 1 univariate functions is used to approximate the non-linear relationship between the CSI data and the corresponding position labels, which is implemented by the inter-layer calculation formula (5). The KAN network is trained, and the result of the first hidden layer of the trained KAN is output as the extracted CSI features.

[0037] Preferably, training the BLS by non-linear mapping using the CSI features to obtain the trained BLS includes:

[0038] The BLS framework includes three main stages: feature mapping, enhancement layer construction, and weight calculation. In the feature mapping layer, the CSI features extracted by KAN are mapped to a higher-dimensional space through a non-linear mapping function to form multiple groups of feature nodes;

[0039] The calculation formula of the non-linear mapping function is as follows:

[0040] Z i = φ i (XW ei + β ei )(i = 1, ···, n) (9)

[0041] wherein, W ei and β ei are randomly initialized weights and biases, and X is the output CSI data after feature extraction;

[0042] The activation function φ i (·) is the Sigmoid function. The combined feature nodes form the feature set Z n , and the enhancement layer includes multiple enhancement node groups. Each feature set Z n is connected to a separate enhancement node group. The calculation formula of the enhancement node is:

[0043]

[0044] wherein, W hjand β hj They are also randomly initialized weights and biases.

[0045] ξ j (·) and φ i (·) are the same to form the enhanced set H m ;

[0046] In the weight calculation stage, BLS uses ridge regression to calculate the output weights, and the formula is:

[0047] W m =(λI + [Z|H] T [Z|H]) -1 [Z|H] T Y (11)

[0048] where λ is the regularization parameter used to prevent overfitting; I is the identity matrix; Z is the feature set output by the feature mapping layer calculated by formula (9), which is composed of n groups of feature nodes Z i combined; H is the enhanced set output by the enhanced layer calculated by formula (10), which is composed of m groups of enhanced nodes H j combined; Y is the actual position label;

[0049] Calculate the weight matrix W of BLS m , and update the parameters of the weight matrix W of BLS through the pseudo-inverse solution to obtain the trained BLS. m

[0050] Preferably, training the KAN using the CSI features through the LBFGS optimization algorithm to obtain the trained KAN includes:

[0051] Before training, divide the CSI feature data into a training set and a validation set, and standardize it. Train the KAN using the CSI features through the LBFGS optimization algorithm. In each round of training, calculate the composite loss through forward propagation. The mean square error term measures the model accuracy by calculating the square difference between the predicted value and the true value, and is used to represent the importance probability of the nodes in the probability distribution. Promote the smoothness of the parameter distribution by maximizing the entropy of the probability distribution to obtain the trained KAN.

[0052] Preferably, inputting the CSI fingerprint data of the user's current position into the trained KAN, the trained KAN outputs the position estimation information of the user, and inputting the CSI fingerprint data of the user's current position into the trained BLS, the trained BLS outputs the position estimation information of the user; fusing the position estimation information of the user output by the trained KAN and BLS respectively through the temperature coefficient to obtain the final positioning result of the user, including:

[0053] ​Input the CSI fingerprint data of the user's current location into the trained KAN, and calculate the location estimate through forward propagation. The specific formula is:

[0054]

[0055] where x’ is the preprocessed CSI fingerprint feature input in real time, and y KAN is the two-dimensional coordinate estimate output by the KAN;

[0056] Input the CSI fingerprint data of the user's current location into the trained BLS, and obtain the two-dimensional coordinate estimate y output by the BLS BLS ;

[0057] Calculate the positioning errors of the BLS and the KAN respectively through the temperature-controlled weight assignment function:

[0058]

[0059] where ε BLS represents the positioning error of the BLS on the validation set; εKAN represents the positioning error of the KAN on the validation set, μ is the weight assigned to the BLS, 1 - μ corresponds to the weight of the KAN, and the parameter τ adjusts the weight distribution. Combine the prediction results of the BLS and the KAN according to the dynamically assigned weights to obtain the final positioning result y of the user final :

[0060] y final = μ · y BLS + (1 ― μ) · y kan (13)

[0061] where y BLS represents the two-dimensional coordinate positioning result output by the BLS after processing the real-time CSI data; y KAN represents the two-dimensional coordinate positioning result of the user output by the KAN after processing the real-time CSI data.

[0062] It can be seen from the technical solutions provided by the embodiments of the present invention described above that the present invention suppresses noise by designing a hybrid wavelet denoising strategy, uses the Kolmogorov - Arnold network (KAN) for efficient feature extraction, introduces a temperature-driven Broad Learning System (BLS)-KAN dynamic fusion mechanism, adaptively adjusts the model weights, improves the positioning accuracy and robustness, and realizes high-precision, high-adaptability and computationally efficient indoor positioning.

[0063] Additional aspects and advantages of the present invention will be given in part in the following description, and these will become obvious from the following description, or can be understood through the practice of the present invention. Brief Description of the Drawings

[0064] To more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the accompanying drawings required for the description of the embodiments. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other accompanying drawings can be obtained based on these drawings.

[0065] Figure 1 It is a schematic diagram of the implementation principle of a user positioning method based on CSI fingerprint data provided by an embodiment of the present invention.

[0066] Figure 2 It is a processing flowchart of a user positioning method based on CSI fingerprint data provided by an embodiment of the present invention. Specific Embodiments

[0067] The following will describe in detail the embodiments of the present invention. The examples of the embodiments are shown in the accompanying drawings, where the same or similar reference numerals represent the same or similar elements or elements with the same or similar functions from beginning to end. The embodiments described below by referring to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be construed as a limitation to the present invention.

[0068] Those skilled in the art of the present technology can understand that unless specifically stated otherwise, the singular forms "a", "an", "the", and "said" used here may also include the plural forms. It should be further understood that the term "including" used in the description of the present invention means the presence of the described features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or their groups. It should be understood that when we say an element is "connected" or "coupled" to another element, it can be directly connected or coupled to other elements, or there may also be intermediate elements. In addition, the "connection" or "coupling" used here may include wireless connection or coupling. The phrase "and / or" used here includes any unit and all combinations of one or more related listed items.

[0069] Those skilled in the art of the present technology can understand that unless otherwise defined, all terms (including technical terms and scientific terms) used here have the same meaning as the general understanding of those of ordinary skill in the art to which the present invention belongs. It should also be understood that terms such as those defined in a general dictionary should be understood to have a meaning consistent with the meaning in the context of the prior art, and will not be interpreted with an idealized or overly formal meaning unless defined as here.

[0070] For the convenience of understanding the embodiments of the present invention, the following will further explain and illustrate with several specific embodiments in conjunction with the accompanying drawings, and each embodiment does not constitute a limitation to the embodiments of the present invention.

[0071] The implementation principle of a user positioning method based on CSI fingerprint data provided by the embodiments of the present invention is as Figure 1 shown. In the offline stage, first, collect some original CSI fingerprint data and construct a CSI fingerprint database. Then, use the wavelet decomposition method to reduce the noise of the CSI fingerprint data and improve the positioning accuracy. Next, use KAN to extract features from the CSI fingerprint database after noise reduction, accurately capture the hidden rules in the CSI fingerprint data, and lay a solid foundation for subsequent model training. Finally, use the processed CSI fingerprint data to train KAN and BLS respectively to achieve a high-precision user positioning function. In the online stage, use the trained KAN and BLS to process the new CSI fingerprint data, output the positioning results respectively, and obtain the final result through temperature coefficient fusion to estimate the user's position.

[0072] The processing flow of a user positioning method based on CSI fingerprint data provided by the embodiments of the present invention is as Figure 2 shown, including the following processing steps:

[0073] Step S10: Combine the CSI data collected at the sampling points and the CSI data collected on the optimal path to obtain the original CSI data.

[0074] Divide the indoor positioning area of the user into grids, use a device equipped with a Linux system and an installed modified driver, and carry out CSI data collection work through the 802.11n CSI tool. The 802.11n CSI tool supports extracting fine-grained CSI information containing 30 subcarriers from the WiFi signal, and this CSI information can describe the transmission characteristics of the wireless signal in detail. Then, randomly collect the original CSI data of a small number of sampling points that conform to the normal distribution in the grid area, and use these data to fit a multivariate Gaussian model.

[0075] Design each layer of the DeepBLS network into a width learning block, adopt a training method with the layer-by-layer residual scaled by the learning rate as the target output, and each layer is trained only once.

[0076] The original CSI fingerprint database is initially constructed using the DeepBLS network with the original CSI data. Then, with the help of the Generative Adversarial Networks (GAN), features are further extracted from the original CSI data, and the original CSI fingerprint database is updated by secondary prediction. After a large number of our previous practices, the DeepBLS confidence coefficient is selected as 0.6, and the GAN confidence coefficient is 0.4. The final CSI fingerprint database is constructed in the way of confidence weighting.

[0077] Step S20: Process the above final CSI fingerprint database using a wavelet-based denoising method to obtain the CSI fingerprint data after denoising processing.

[0078] Select the Daubechies4 (Db4) wavelet as the basis function to perform multi-level decomposition on the CSI fingerprint data in the above final CSI fingerprint database, transform the CSI data into the time-frequency domain, and perform four-level wavelet decomposition on each CSI channel to obtain the low-frequency approximation coefficient A L and the high-frequency detail coefficient D L , formula

[0079]

[0080] where A L is the low-frequency approximation component, and D l contains the high-frequency detail coefficients of different decomposition levels. Calculation method: With the help of the PyWavelets library in Python, use the wavedec function to perform four-level decomposition on the CSI fingerprint data with the Db4 wavelet. The first element in the coefficient list returned by this function is A L , and the remaining elements are D1 to D4 in sequence.

[0081] To further suppress noise and retain key structural information, calculate the adaptive threshold τ using the VisuShrink method. The calculation formula for the adaptive threshold τ is:

[0082]

[0083] where σ is the median absolute deviation of the detail coefficients, and N is the length of the coefficient vector.

[0084] Apply soft threshold processing to the high-frequency detail coefficient D l , and the formula is:

[0085] D′ l = sign(D l )·max(|D l |-τ, 0) (3)

[0086] Finally, first AL Place it at the beginning of the coefficient list and arrange D' in sequence l .

[0087] Use the pywt.waverec function to perform inverse wavelet transform to reconstruct the signal with the above coefficient list and db4 wavelet basis as parameters. If the length of the reconstructed signal is inconsistent with that of the original CSI signal, perform truncation or padding processing to obtain the denoised CSI signal X', effectively improving the quality of the CSI fingerprint database.

[0088] Step S30: Use KAN to extract features from the denoised CSI fingerprint data to obtain CSI features.

[0089] KAN is based on the Kolmogorov - Arnold theorem (KAT) and has strong non - linear learning ability, which can effectively capture complex data patterns. According to KAT, any complex continuous function can be expressed as a combination of a series of simple functions, and the formula is:

[0090]

[0091] where x=(x1,…,x n ) is the pre - processed CSI data feature vector (with dimension n), f(x) represents the continuous mapping function from CSI features to spatial positions; n is the input feature dimension, q is the summation term index (range is [1, 2n + 1]), φ q (·) and are both univariate continuous functions, which are used to combine intermediate results and act on the input feature x p . This formula shows that any continuous function can be approximated by the nested combination of 2n + 1 univariate functions, so as to capture the non - linear relationship between CSI features and positions, avoiding the high computational cost of traditional matrix multiplication.

[0092] Use KAN to process the CSI data X'. KAN converts the input activation x l to the output activation x l+1 through sum - function operations instead of traditional matrix multiplication, that is:

[0093]

[0094] where φ l,j,i is an adaptive non - linear function, parameterized by B - splines as:

[0095] φ(x)=w b b(x)+w s ∑ i c i B i (x) (6)

[0096] Among them, B i B(x) is a basis spline function, and c i is a trainable coefficient. w b and w s are weight parameters respectively. The former controls the contribution degree of the fixed activation function b(x), and the latter controls the contribution degree of the linear combination part of the basis spline function.

[0097] In most cases,

[0098] b(x) = silu(x) = x / (1 + e -x )(7)

[0099] B i B(x) and c i enable KAN to dynamically adjust the function complexity, efficiently extract CSI features and maintain interpretability. KAN performs well in function approximation and generalization, and the error bound of its approximation function is:

[0100]

[0101] Among them, G represents the grid resolution, k represents the spline order, and m is the differentiable order. This error bound ensures that KAN avoids the curse of dimensionality and achieves more efficient expansion than multi-layer perceptrons (MLPs).

[0102] The test loss is scaled by G -(k+1) compared with the multi-layer perceptron (MLP) scaled by N -1 / d (d is the intrinsic dimension of the data), KAN can avoid the curse of dimensionality and has higher learning efficiency. Moreover, KAN can optimize the learning function through grid expansion, improve the resolution of the B-spline function without retraining, effectively learn the non-linear mapping between CSI signals and spatial positions, and improve the positioning accuracy.

[0103] When using KAN for feature extraction, first use the preprocessed CSI data X′ as the input, and its dimension n corresponds to the number of CSI features. Based on formula (4), use the nested combination of 2n + 1 univariate functions to approximate the non-linear relationship between CSI data and the corresponding position labels, and the inter-layer calculation formula (5) is implemented to train the KAN network. Output the result of the first hidden layer of the trained KAN as the extracted CSI feature. Using this feature extraction method, the CSI data dimension is reduced by 50%, greatly shortening the subsequent model training time while the error increase is <10%.

[0104] Step S40: In the offline stage, use the CSI data after feature extraction in step S30 to train BLS through non-linear mapping to obtain the trained BLS. In the online stage, by inputting the preprocessed real-time input CSI data into BLS, the two-dimensional coordinate position estimation of the user can be output.

[0105] Traditional deep learning models have a long training time and high computational resource requirements. When adjusting the network structure, they need to be retrained, which is not suitable for real-time positioning. BLS adopts a flat architecture, avoiding deep stacking, improving efficiency, and can effectively process the CSI features extracted by KAN and perform positioning prediction quickly and accurately. The BLS framework consists of three main stages: feature mapping, enhancement layer construction, and weight calculation. In the feature mapping layer, the CSI features extracted by KAN are mapped to a higher-dimensional space through a non-linear mapping function to form multiple groups of feature nodes.

[0106] The calculation formula of the above non-linear mapping function is as follows:

[0107] Z i =φ i (XW ei +β ei )(i=1,···,n) (9)

[0108] Where, W ei and β ei are randomly initialized weights and biases, and X is the CSI data after feature extraction as the output.

[0109] The activation function φ i (·) is the Sigmoid function, and the combined feature nodes form the feature set Z n . The enhancement layer includes multiple enhancement node groups. Each feature set Z n is connected to a separate enhancement node group. The calculation formula of the enhancement node is:

[0110]

[0111] Where, W hj and β hj are also randomly initialized weights and biases.

[0112] ξ j (·) and φ i (·) are the same, forming the enhancement set H m .

[0113] In the weight calculation stage, BLS uses ridge regression to calculate the output weights. The formula is:

[0114] W m =(λI+[Z|H] T [Z|H]) -1 [Z|H] T Y (11)

[0115] Among them, λ is the regularization parameter for preventing overfitting; I is the identity matrix; Z is the feature set output by the feature mapping layer calculated by formula (9), which is composed of n groups of feature nodes Z i combined; H is the enhancement set output by the enhancement layer calculated by formula (10), which is composed of m groups of enhancement nodes H j combined; Y is the actual position label.

[0116] Thus, the weight matrix W of BLS can be calculated m . The efficiency of weight calculation is ensured by the pseudo-inverse solution, which is applicable to the real-time positioning application of large-scale CSI fingerprint databases.

[0117] During the online positioning phase, when the CSI fingerprint data of the user's current position is input into BLS, the current position estimate of the user can be output.

[0118] Step S50: Use the CSI data after feature extraction in step S30 to train KAN to obtain the trained KAN. During the online phase, by inputting the preprocessed real-time input CSI data into KAN, the two-dimensional coordinate position estimate of the user can be output.

[0119] The input data during the training process of KAN is the CSI data after feature extraction in S30. Before training, the data is divided into a training set (80%) and a validation set (20%), and standardized:

[0120]

[0121] Among them, N is the total number of samples, is the number of samples in the validation set, and U(0,1) represents standardization to the [0,1] interval.

[0122] The network parameters are initialized with He:

[0123]

[0124] Among them, din is the input dimension, dgrid = 5 is the default number of grids, and the noise scale coefficient is 0.1.

[0125] Then, the distribution of spline nodes is dynamically adjusted every 5 epochs:

[0126]

[0127] Among them, dgrid = 5 is the number of grids, and xmin / xmax is the minimum / maximum value of the input of the current layer.

[0128] The composite loss is calculated through forward propagation in each round of training:

[0129]

[0130] Among them, B is the number of samples in each round of training; the mean squared error (MSE) term measures the model accuracy by calculating the squared difference between the predicted value ŷ b and the true value y b . represents the sum of the absolute values of all spline functions. The probability distribution represents the importance probability of the nodes, and promotes the smoothness of the parameter distribution by maximizing the entropy of the probability distribution p ij .

[0131] After multiple rounds of training, the KAN training is completed and enters the online positioning stage. The CSI fingerprint data of the user's current location is input into the trained KAN, and the position estimation is calculated through forward propagation. The specific formula is:

[0132]

[0133] where x' is the preprocessed CSI fingerprint feature input in real time, and the output is the two-dimensional coordinate estimation.

[0134] Step S60: Temperature coefficient-driven dynamic weight fusion positioning. BLS has high computational efficiency but low accuracy, while KAN has strong feature learning ability but large computational overhead. To give full play to the complementary advantages of the two, a dynamic weighted fusion strategy is proposed. First, calculate the positioning errors of BLS and KAN respectively through the weight allocation function controlled by temperature:

[0135]

[0136] where ε BLS represents the positioning error of BLS on the validation set; εKAN represents the positioning error of KAN on the validation set.

[0137] Dynamically determine the contributions of each model. μ is the weight assigned to BLS, and 1 - μ corresponds to the weight of KAN. The parameter τ adjusts the weight distribution. Finally, combine the prediction results of BLS and KAN according to the dynamically assigned weights:

[0138] y final = μ · y BLS + (1 - μ) · y kan (13)

[0139] where y BLS represents the two-dimensional coordinate positioning result output by BLS after processing the real-time CSI data; y KAN represents the two-dimensional coordinate positioning result of the user output by KAN after processing the real-time CSI data.

[0140] This adaptive fusion mechanism enables KFBK to adapt to CSI signal changes and environmental dynamics in real time. When the signal is stable, BLS ensures positioning efficiency and reliability at low cost; in interference scenarios, KAN captures complex CSI patterns to improve accuracy. Through dynamic optimization of the model, the overall positioning robustness is enhanced.

[0141] The present invention transmits and receives data by means of the monitoring mode of the Linux 802.11n CSI tool, and collects a small amount of raw CSI data. In this way, the data packet transmission rate can be accurately regulated, thereby shortening the time-consuming of the offline stage. A laptop dedicated to data collection is used for data reception, and the remaining algorithms are executed on a local server. This configuration greatly reduces the device's requirements for resources and computing power, and is suitable for large-scale localization scenarios. The algorithm is based on Python and is implemented using the PyTorch, Sklearn, and TensorFlow libraries, and runs on a Lenovo laptop equipped with an i5-1240P CPU and 16GB of RAM.

[0142] In summary, the embodiment of the present invention designs an innovative KFBK positioning framework to achieve high-precision, highly adaptable, and computationally efficient indoor positioning. First, in the data preprocessing stage, the wavelet decomposition technology is used to process the CSI fingerprint data. By retaining the key CSI time-frequency features, the data quality is greatly improved, providing a reliable data basis for subsequent processing. Subsequently, KAN is used for feature extraction. It effectively captures the complex non-linear spatio-temporal dependence relationships in the high-dimensional CSI data, reduces information loss while reducing the data dimension, avoids the curse of dimensionality, and improves the feature representation ability. Then, BLS and KAN are used to train independent positioning models respectively. Finally, a dynamic weight fusion mechanism based on the temperature coefficient adaptively adjusts the weights of BLS and KAN in the final positioning result according to their real-time positioning errors. This mechanism overcomes the problem that it is difficult for existing methods to balance speed and accuracy, enables the system to achieve high-precision positioning in different environments, and improves the overall positioning robustness and environmental adaptability.

[0143] To verify the effectiveness of the present invention, a large number of experiments were carried out in non-line-of-sight propagation environments and line-of-sight propagation environments. Compared with several existing methods, the present invention shows excellent performance in updating the CSI fingerprint database, and the positioning accuracy is increased by an average of 39.9%.

[0144] Those of ordinary skill in the art can understand that the drawings are only schematic diagrams of an embodiment, and the modules or processes in the drawings are not necessarily essential for implementing the present invention.

[0145] As can be seen from the description of the above embodiments, those skilled in the art can clearly understand that the present invention can be implemented by means of software plus a necessary general hardware platform. Based on such an understanding, the technical solution of the present invention, in essence, or the part that makes a contribution to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in various embodiments or some parts of the embodiments of the present invention.

[0146] Each embodiment in this specification is described in a progressive manner. For the same or similar parts among the embodiments, reference can be made to each other. The key point of each embodiment is to illustrate the differences from other embodiments. In particular, for the device or system embodiments, since they are basically similar to the method embodiments, the description is relatively simple. For the relevant parts, reference can be made to the partial description of the method embodiments. The device and system embodiments described above are only illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. Those of ordinary skill in the art can understand and implement it without creative efforts.

[0147] As mentioned above, the above are only the preferred specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present invention should be covered by the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims

1. A user positioning method based on channel state information fingerprint data, characterized in that Including: Collect CSI data at sampling points and construct a CSI fingerprint database; Use the wavelet decomposition method to denoise the CSI fingerprint data in the CSI fingerprint database, and use KAN to extract features from the denoised CSI fingerprint data to obtain CSI features; Use the CSI features to train BLS through nonlinear mapping to obtain a trained BLS, and use the CSI features to train KAN through the LBFGS optimization algorithm to obtain a trained KAN; Input the CSI fingerprint data of the user's current location into the trained KAN, and the trained KAN outputs the user's position estimation information. Input the CSI fingerprint data of the user's current location into the trained BLS, and the trained BLS outputs the user's position estimation information. The position estimation information of the user respectively output by the trained KAN and BLS is fused through the temperature coefficient to obtain the final user positioning result.

2. The method according to claim 1, wherein The collecting CSI data at sampling points and constructing a CSI fingerprint database includes: Divide the indoor positioning area of the user into grids, and use the device installed with the CSI tool to carry out CSI data collection work in the indoor positioning area of the user. Take the user as the sampling point, and the CSI tool randomly collects the original CSI data conforming to the normal distribution in the mobile hotspot WiFi signal of the sampling point in the grid area; Design each layer of the DeepBLS network into a width learning block, adopt the training method with the interlayer residual scaled by the learning rate as the target output, and each layer is trained only once. Use the DeepBLS network to initially construct the original CSI fingerprint database using the original CSI data, and then further extract features from the original CSI data with the help of the generative adversarial network and update the original CSI fingerprint database by secondary prediction. Select the DeepBLS confidence coefficient as 0.6 and the GAN confidence coefficient as 0.4, and construct the final CSI fingerprint database in the way of confidence weighting.

3. The method according to claim 2, characterized in that, The denoising the CSI fingerprint data in the CSI fingerprint database by using the wavelet decomposition method includes: Select the Daubechies4 wavelet as the basis function to perform multi-level decomposition on the CSI fingerprint data in the final CSI fingerprint database, transform the CSI data into the time-frequency domain, and perform four-level wavelet decomposition on each CSI channel to obtain the low-frequency approximation coefficient A L and the high-frequency detail coefficient D L , formula: Among them, A L is the low-frequency approximation component, and D l contains high-frequency detail coefficients at different decomposition levels; Calculate the adaptive threshold τ using the VisuShrink method, and the calculation formula of the adaptive threshold τ is: where σ is the median absolute deviation of the detail coefficient and N is the length of the coefficient vector; For the high-frequency detail coefficient D l Apply soft thresholding, and the formula is: D′ l = sign(D l )·max(|D l | - τ, 0) (3) Put A L at the beginning of the coefficient list, and arrange D' in sequence l , use the pywt.waverec function to perform inverse wavelet transform to reconstruct the signal with the coefficient list and the Daubechies4 wavelet basis as parameters. If the length of the reconstructed signal is inconsistent with the original CSI signal, perform truncation or padding processing to obtain the denoised CSI signal X'.

4. The method according to claim 3, wherein The extracting features from the denoised CSI fingerprint data by using KAN to obtain CSI features includes: According to KAT, represent the complex continuous function as a combination of a series of simple functions, and the formula is: where x = (x1, …, x n ) is the CSI data feature vector after preprocessing, f(x) represents the continuous mapping function from CSI features to spatial positions, n is the input feature dimension, q is the summation term index, φ q (·) and are both univariate continuous functions, used to combine intermediate results and act on the input feature x p ; Processing CSI data X′ using KAN, where KAN converts the input activation x l to the output activation x l+1 , i.e.: where φ l,j,i is an adaptive non-linear function, parametrically represented by B-splines as: φ(x) = w b b(x) + w s ∑ i c i B i (x) (6) Among them, B i (x) is a basis spline function, and c i is a trainable coefficient, and w b and w s are weight parameters respectively; b(x) = silu(x) = x / (1 + e -x ) (7) When using KAN for feature extraction, take the preprocessed CSI data X′ as the input of KAN. The dimension n of KAN corresponds to the number of CSI features. Based on formula (4), use the nested combination of 2n + 1 univariate functions to approximate the nonlinear relationship between the CSI data and the corresponding position label, and implement it through the interlayer calculation formula (5). Train the KAN network, and output the result of the first hidden layer of the trained KAN as the extracted CSI features.

5. The method according to claim 4, wherein The training BLS through nonlinear mapping by using the CSI features to obtain a trained BLS includes: The BLS framework consists of three main stages: feature mapping, enhancement layer construction, and weight calculation. In the feature mapping layer, the CSI features extracted by KAN are mapped to a higher-dimensional space through a non-linear mapping function to form multiple groups of feature nodes; The calculation formula of the non-linear mapping function is as follows: Z i = φ i (XW ei + β ei )(i = 1, ···, n) (9) Among them, W ei and β ei are randomly initialized weights and biases, and X is the CSI data after feature extraction of the output; The activation function φ i (·) is a Sigmoid function, and the combined feature nodes form a feature set Z n , the enhancement layer includes multiple enhancement node groups, and each feature set Z n is connected to a separate enhancement node group, and the calculation formula for the enhancement nodes is: Among them, W hj and β hj are also randomly initialized weights and biases. ξ j (·) and φ i (·) are the same, forming the enhanced set H m ; In the weight calculation stage, BLS uses ridge regression to calculate the output weights, and the formula is: W m = (λI + [Z|H] T [Z|H]) -1 [Z|H] T Y (11) Among them, λ is a regularization parameter used to prevent overfitting; I is the identity matrix; Z is the feature set output by the feature mapping layer calculated by formula (9), which is composed of n groups of feature nodes Z i combined; H is the enhancement set output by the enhancement layer calculated by formula (10), which is composed of m groups of enhancement nodes H j combined; Y is the actual position label; Calculate the weight matrix W of BLS m , update the weight matrix W of BLS by the pseudo-inverse solution m parameters of, and obtain the trained BLS 6. The method according to claim 4, wherein Training KAN using the CSI features through the LBFGS optimization algorithm to obtain the trained KAN includes: Before training, the CSI feature data is divided into a training set and a validation set and standardized. Training KAN using the CSI features through the LBFGS optimization algorithm. In each round of training, the composite loss is calculated through forward propagation. The mean square error term measures the model accuracy by calculating the square difference between the predicted value and the true value. It is used to represent the importance probability of nodes in the probability distribution, and the smoothness of the parameter distribution is promoted by maximizing the entropy of the probability distribution to obtain the trained KAN.

7. The method according to claim 4 or 5, characterized in that Inputting the CSI fingerprint data of the user's current location into the trained KAN, the trained KAN outputs the location estimation information of the user. Inputting the CSI fingerprint data of the user's current location into the trained BLS, the trained BLS outputs the location estimation information of the user; Fusing the location estimation information of the user output by the trained KAN and BLS respectively through the temperature coefficient to obtain the final user location result, including: Inputting the CSI fingerprint data of the user's current location into the trained KAN, and calculating the location estimation through forward propagation. The specific formula is: where x’ is the preprocessed CSI fingerprint feature input in real time, and y KAN is the two-dimensional coordinate estimation output by KAN; Input the CSI fingerprint data at the user's current location into the trained BLS to obtain the two-dimensional coordinate estimate y output by the BLS BLS ; Calculating the positioning errors of BLS and KAN respectively through the weight allocation function controlled by temperature: where ε BLS represents the positioning error of BLS on the validation set; εKAN represents the positioning error of KAN on the validation set, μ is the weight assigned to BLS, 1 - μ corresponds to the weight of KAN, the parameter τ adjusts the weight distribution, and the prediction results of BLS and KAN are combined according to the dynamically assigned weights to obtain the final positioning result y of the user final : y final = μ·y BLS +(1 – μ)·y kan (13) Among them, y BLS represents the two-dimensional coordinate positioning result output by BLS after processing real-time CSI data; y KAN represents the two-dimensional coordinate positioning result of the user output by KAN after processing real-time CSI data.

Images