UWB positioning method fusing CIR interval constraint and TOA ranging optimization

Abstract

The invention provides a UWB positioning method fusing CIR interval constraint and ToA ranging optimization. The UWB positioning method is characterized in that a CIR positioning interval estimation module, a ToA ranging error module and a fusion position prediction module are included; the CIR positioning interval estimation module comprises a CNN (Convolutional Neural Network) model and a Copula model; the ToA ranging error module comprises a GMM model; the positioning method comprises the following steps: a CIR positioning interval estimation module processes CIR data of a positioning target to obtain a first positioning error interval and a second positioning error interval, and a ToA ranging error module processes ToA data of the positioning target to obtain an optimized target function and a weighted cost function; and the fusion position prediction module adopts a PSO algorithm, performs particle optimization in a first positioning error interval according to an optimization objective function, and then performs weighted average operation processing on particles in a second positioning error interval according to a weighted cost function to obtain a positioning prediction value of a positioning target. By adopting the positioning method, the positioning precision and robustness of the positioning target in a complex indoor environment can be greatly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of wireless positioning, and particularly to a UWB positioning method that integrates CIR interval constraint and TOA ranging optimization. Background Art

[0002] In recent years, with the wide application of smart phones and other wireless communication devices, the demand for positioning technology has grown rapidly in multiple industrial fields. In outdoor environments, the Global Navigation Satellite System (GNSS) has become the standard solution for positioning technology. However, in indoor environments, due to the obstruction of obstacles such as buildings, walls, and furniture, the propagation of wireless signals is severely interfered, making it difficult for GNSS technology to provide stable and reliable positioning services. Therefore, researchers have actively explored various indoor positioning technologies, among which Ultra Wide Band (UWB) technology has become a research hotspot due to its strong penetration ability and high positioning accuracy. The positioning accuracy of UWB technology mainly depends on the ranging accuracy of the Time of Arrival (ToA) of the signal. However, Non-Line-of-Sight (NLoS) propagation phenomena, such as penetration, reflection, and diffraction, will cause ranging errors and significantly reduce the positioning accuracy. Therefore, how to effectively solve the positioning accuracy problem in the NLoS environment has become the key challenge for the further development of UWB technology.

[0003] Currently, the mainstream UWB indoor positioning methods are mainly positioning technologies based on ToA ranging. The ToA ranging-based method calculates the distance by measuring the signal propagation time. However, in the NLoS environment, the signal generates errors due to phenomena such as reflection and diffraction, resulting in a decrease in positioning accuracy. Therefore, the prior art has proposed NLoS identification and mitigation methods, such as distinguishing Line-of-Sight (LoS) and NLoS signals, or attempting to suppress NLoS errors. However, these methods have limited effects when LoS signals are scarce or it is difficult to completely eliminate the ranging errors in the NLoS environment. On the other hand, the positioning method based on Channel Impulse Response (CIR) fingerprint performs well in the NLoS environment, but its accuracy is usually lower than that of the ToA ranging-based method in the LoS scenario. The prior art has attempted to combine ranging and fingerprint positioning, but most are limited to selecting one of the methods according to conditions, and the positioning accuracy in complex indoor environments is not ideal. Summary of the Invention

[0004] Aiming at the problems in the background art, the present invention provides a UWB positioning method that integrates CIR interval constraint and TOA ranging optimization to solve the problems of low positioning accuracy and poor robustness in complex indoor environments.

[0005] To achieve the object of the present invention, the present invention provides a UWB positioning method that integrates CIR interval constraint and ToA ranging optimization. The innovation lies in: including a CIR positioning interval estimation module, a ToA ranging error module, and a fusion position prediction module; the CIR positioning interval estimation module includes a CNN model and a Copula model; the ToA ranging error module includes a GMM model; the positioning method includes:

[0006] I) Offline stage:

[0007] Collect the CIR training data set, and use the CIR training data set to train the CNN model and the Copula model; the CNN model can output an initial positioning prediction value, and the Copula model can output a first positioning error interval and a second positioning error interval according to the initial positioning prediction value; among them, the significance level parameter ε1 required to obtain the first positioning error interval and the significance level parameter ε2 required to obtain the second positioning error interval are both preset values, and ε1 < ε2;

[0008] Collect the ToA training data set, and use the ToA training data set to train the GMM model; the GMM model can output an optimization objective function and a weighted cost function;

[0009] II) Online stage:

[0010] 1) Input the CIR data of the positioning target into the CIR positioning interval estimation module, and after processing, output the first positioning error interval and the second positioning error interval of the positioning target. Then, the CIR positioning interval estimation module transmits the first positioning error interval and the second positioning error interval to the fusion position prediction module; input the ToA data of the positioning target into the ToA ranging error module, and after processing, output the optimization objective function and the weighted cost function of the positioning target. Then, the ToA ranging error module transmits the optimization objective function and the weighted cost function to the fusion position prediction module;

[0011] 2) The fusion position prediction module randomly generates multiple PSO particles within the range of the first positioning error interval;

[0012] 3) The fusion position prediction module uses the PSO algorithm to iteratively optimize the generated multiple PSO particles according to the optimization objective function to obtain multiple optimized particles;

[0013] 4) The fusion position prediction module performs a weighted average operation on the multiple optimized particles within the range of the second positioning error interval according to the weighted cost function to obtain the positioning prediction value of the positioning target.

[0014] Further, the CNN model integrates CBAM and RES.

[0015] Further, ε1 is 0.01 and ε2 is 0.1.

[0016] The principle of the present invention is as follows:

[0017] On the one hand, although the traditional ToA ranging-based positioning method has high positioning accuracy in LoS environments, it often has difficulty effectively coping with multipath, reflection, and diffraction phenomena in NLoS environments. These factors lead to a significant increase in ranging error, thus affecting positioning accuracy. On the other hand, although the CIR fingerprint positioning method performs well in NLoS environments, its accuracy is usually not as good as that of the ToA ranging-based positioning method in LoS scenarios. On the other hand, although the deep learning-based NLoS suppression method can improve positioning accuracy, its high demand for computing resources and time (high cost) limits its popularization in practical applications.

[0018] To solve these problems, the present invention proposes an innovative UWB positioning method that combines CIR fingerprint interval constraints and ToA ranging optimization. This method constructs a multi-modal collaborative positioning framework by deeply integrating CIR fingerprint features and ToA ranging information.

[0019] First, a CNN convolutional neural network model is used to perform multi-scale feature extraction and adaptive enhancement on the CIR signal. As an optimization, CBAM (channel and spatial attention mechanism) and residual (RES) connections are integrated in the CNN to achieve high-dimensional representation of environmental features and fine fingerprint modeling. This model dynamically focuses on key scattering features through the attention mechanism and effectively alleviates the gradient vanishing problem by combining residual learning, significantly improving the stability of feature extraction. Then, a multi-objective regression conformal prediction method based on the Copula function is introduced to establish a probability distribution model of fingerprint positioning error. By constructing a joint distribution function of multi-dimensional error variables, a prediction interval with statistical confidence is generated, providing a reliable spatial constraint boundary (the first positioning error interval and the second positioning error interval) for the subsequent optimization process.

[0020] Secondly, to address the NLoS bias problem in ToA ranging, a GMM is used to perform non-parametric modeling of the error distribution, accurately characterizing the error characteristics caused by the multipath effect, and outputting an optimization objective function and a weighted cost function.

[0021] Furthermore, combined with the PSO algorithm, global optimization is performed under the constraint of the first positioning error interval according to the optimization objective function, and the optimal approximation of position calculation is achieved through swarm intelligence search. Finally, to avoid the possible local optimum problem in PSO optimization, invalid particle positions are eliminated according to the second positioning error interval, and the remaining particle positions are weighted and fused according to the weighted cost function. By constructing a dynamic weight matrix, the contribution degrees of different particles are balanced, further improving the positioning accuracy while maintaining global convergence.

[0022] It can be seen that the present invention gives full play to the positioning advantages of both CIR fingerprint positioning and ToA ranging positioning, fuses the positioning data of the two through the PSO algorithm and performs in-depth optimization, and is particularly suitable for high-precision and robust positioning of positioning targets in complex NLoS environments.

[0023] Thus, the present invention has the following beneficial effects: By using the positioning method of the present invention, the positioning accuracy and robustness of the positioning target in a complex indoor environment can be greatly improved, and the positioning cost is also reduced. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] The drawings of the present invention are described as follows.

[0025] Attached Figure 1 is a schematic structural diagram of the module related to the present invention;

[0026] Attached Figure 2 is a schematic structural diagram of the optimized CNN model; DETAILED DESCRIPTION OF THE EMBODIMENTS

[0027] The present invention will be further described below in conjunction with embodiments.

[0028] The UWB positioning method that fuses CIR interval constraint and ToA ranging optimization according to the present invention is characterized in that it includes a CIR positioning interval estimation module, a ToA ranging error module, and a fusion position prediction module; the CIR positioning interval estimation module includes a CNN (Convolutional Neural Networks) model and a Copula model. In this embodiment, the CNN model incorporates a CBAM (Convolutional Block Attention Module) and a RES (Residual Neural Network); the ToA ranging error module includes a GMM (Gaussian Mixture Model) model; the positioning method is as shown in the attached Figure 1 and includes:

[0029] I) Offline stage:

[0030] Collect the CIR training dataset and use the CIR training dataset to train the CNN model and the Copula model; the CNN model can output an initial positioning prediction value, and the Copula model can output a first positioning error interval and a second positioning error interval according to the initial positioning prediction value; among them, the significance level parameter ε1 required to obtain the first positioning error interval and the significance level parameter ε2 required to obtain the second positioning error interval are both preset values, and ε1 < ε2; in this embodiment, after experimentation, the ε1 is 0.01 and the ε2 is 0.1; the above-mentioned CIR fingerprint interval conformal prediction process can adopt the existing technology, and the specific introduction is as follows:

[0031] In the UWB positioning system, the CIR can provide rich information on the multi-path propagation of signals through direct, reflected, and diffracted paths, fully reflecting the internal relationship between signal propagation and the surrounding environment. Based on this characteristic, the CIR can be implicitly modeled by using a set of recorded training data to achieve fingerprint positioning. However, the UWB signal is easily interfered by the NLoS environment during the propagation process, resulting in instability of the CIR data, and further causing a certain estimation error in the fingerprint positioning model. Therefore, based on the position prediction, the present invention further quantifies the error range, which not only improves the reliability of fingerprint positioning but also provides a rough interval constraint for subsequent optimization.

[0032] 1) Fingerprint positioning model

[0033] The present invention designs a fingerprint positioning model based on CBAM-RES-CNN, and its structure is as Figure 2 shown. The model takes the convolutional neural network (CNN) as the core framework, and integrates the channel and spatial attention mechanism (CBAM) and the residual (RES) connection. The input of the model is an N×152×2 CIR data matrix, representing the amplitude and phase information of 152 sampling points of N anchor nodes. The model first performs preliminary feature extraction through two convolutional layers, and uses batch normalization to accelerate the training process. The residual module enhances the feature learning ability through two layers of convolution and skip connection, avoiding the problem of gradient disappearance. The CBAM module weights the features through the channel and spatial attention mechanisms, improving the ability to capture key features. Finally, the model reduces the dimension through global average pooling, and the fully connected layer further extracts features and outputs two-dimensional coordinates to complete the positioning.

[0034] 2) Single-object conformal prediction

[0035] Single-object regression conformal prediction first divides the dataset into a training set and a calibration set, where the training set is used to train the basic regression model f(x) to obtain the prediction value In the calibration set, calculate the error score of each sample:

[0036]

[0037] to measure the prediction deviation, score and sort these errors, and select the quantile threshold a that satisfies:

[0038] so as to ensure that at least 1 - ε of the calibration samples are covered by the prediction interval for the test sample x s , for the test sample x, the predicted value k the predicted interval of is:

[0039]

[0040] wherein, the standard method assumes that the interval widths of all samples are the same. However, considering the different prediction accuracies of different samples, the scale factor σ j = exp(μ j ) can be introduced to assist the model estimation, for example, σ j = exp(μ j ), where Based on this, the predicted interval is adjusted to:

[0041]

[0042] so as to dynamically adjust the interval width to adapt to the uncertainties of different samples and improve the reliability and flexibility of the prediction.

[0043] 3) Fingerprint positioning error quantification

[0044] The goal of fingerprint positioning error quantification is to provide an interval estimate of the predicted position for the positioning model, rather than a single variable prediction result. This process includes two key points: the center of the interval is determined by the model prediction point, and the interval width is determined by the Copula-based multi-objective regression conformal prediction method to reflect the error range. The determination of the error interval needs to meet two conditions: effectiveness requires that the true value coverage rate is not less than 1 - ε (ε is the preset significance level), and efficiency requires that the interval width is as narrow as possible to improve the accuracy. For two-dimensional fingerprint positioning, the present invention uses the Copula conformal prediction method to quantify the error and generate the predicted intervals of the x coordinate and the y coordinate to ensure that the true value coverage rate meets the significance level requirement of 1 - ε.

[0045] Similar to single-target regression conformal prediction, first, a normalized error scoring function is calculated using a calibration set to measure the deviation between the predicted value and the true value, and to ensure that the error metrics for different target variables (x-coordinate and y-coordinate) are on the same scale for subsequent analysis. All samples in the calibration set form an error scoring matrix α, where rows correspond to samples and columns correspond to target variables. Subsequently, a Copula is combined to model the error scoring matrix to construct an error interval that meets the preset significance level ε. First, the error scores for each target variable are sorted according to the calibration set data to obtain the marginal distribution F j (α j ), and then an empirical Copula is used to model its joint distribution, which has the form:

[0046]

[0047] where u1 and u2 are the marginal distribution values of the target variables; 1 is an indicator function that takes the value 1 when the condition holds and 0 otherwise. To generate an error interval that meets the significance level ε, the equation needs to be solved:

[0048] C(F1(α s ),F2(α s ))=1 - ε(6)

[0049] where F1(α s ) and F2(α s ) represent the mapped values of the error quantile threshold a s in the marginal distribution. Finally, the error interval is adjusted through formula (4) to make it applicable to each target variable, realizing the quantification of fingerprint positioning error.

[0050] 4) ToA ranging error modeling

[0051] Collect a ToA training data set and use the ToA training data set to train a GMM model; the GMM model can output an optimized objective function and a weighted cost function; the establishment of the GMM model is introduced according to the existing technology as follows:

[0052] In actual ToA ranging applications, measurement errors often exhibit complex distribution characteristics due to factors such as multipath, occlusion, and NLoS. A single Gaussian distribution often cannot well fit such multi-modal or skewed error distributions. Therefore, the present invention uses a GMM to model the ToA ranging error. A GMM is a probability model commonly used to describe complex data distributions, and it fits the actual probability distribution of data through a weighted combination of multiple Gaussian distributions. For the nth anchor point among N anchor points, its ToA ranging error e n as a one-dimensional random variable, its probability density function (PDF) can be modeled by a GMM as:

[0053]

[0054] Among them, K n is the number of mixture components; π n,k is the mixture weight, satisfying μ n,k and are the mean and variance of the k-th Gaussian component respectively. For the ranging error sample set of the n-th anchor point The GMM parameter set can be obtained by maximizing the log-likelihood estimation, and its expression is:

[0055]

[0056] The solution of maximizing the likelihood usually uses the EM algorithm, where the E-step calculates the posterior responsibility:

[0057]

[0058] The M-step updates the GMM parameters:

[0059]

[0060] Continuously iterate the E-step and the M-step until the log-likelihood converges or reaches the preset iteration upper limit, and the GMM parameter estimation can be obtained. The key problem is how to determine the optimal number of components. If K n is too small, the model is difficult to fit the complex error distribution; if K n is too large, overfitting is likely to occur. Therefore, the Bayesian Information Criterion (BIC) is usually used to evaluate the model, and in the preset candidate set K n = 1, 2, …, K max select the K n with the smallest BIC as the optimal number of components, and its calculation formula is:

[0061]

[0062] Among them, M is the total number of samples, p n is the number of model parameters, is the likelihood value under the maximum likelihood estimation. By calculating the BIC value to screen the optimal K n , the GMM model applicable to the ToA ranging error can be constructed.

[0063] II) Online stage:

[0064] 1) Input the CIR data of the positioning target into the CIR positioning interval estimation module. After processing, output the first positioning error interval and the second positioning error interval of the positioning target. Then, the CIR positioning interval estimation module transmits the first positioning error interval and the second positioning error interval to the fusion position prediction module. Input the ToA data of the positioning target into the ToA ranging error module. After processing, output the optimization objective function and the weighted cost function of the positioning target. Then, the ToA ranging error module transmits the optimization objective function and the weighted cost function to the fusion position prediction module;

[0065] 2) The fusion position prediction module randomly generates multiple PSO particles within the range of the first positioning error interval;

[0066] 3) The fusion position prediction module uses the PSO algorithm to iteratively optimize the generated multiple PSO particles according to the optimization objective function to obtain multiple optimized particles;

[0067] 4) The fusion position prediction module performs a weighted average operation on the multiple optimized particles within the range of the second positioning error interval according to the weighted cost function to obtain the positioning prediction value of the positioning target.

[0068] The optimization process of the above PSO algorithm is introduced as follows according to the prior art:

[0069] A) Construct the objective function

[0070] Assume that there are N anchor points in the system, and the coordinates of the nth anchor point are a n =(x n , y n ), and its ToA ranging to the tag is When assuming that the tag is located at x=(x, y), its geometric distance to the anchor point is:

[0071]

[0072] The measurement residual is defined as the difference between the measured value and the theoretical value:

[0073]

[0074] Combined with the method in 4) ToA ranging error modeling in the offline stage, perform GMM modeling on the ranging error e n (x) of each anchor point. To reduce the complexity of the objective function, take a smaller K max to improve the PSO optimization efficiency.

[0075] To maximize the joint likelihood of all anchor point measurement residuals, usually minimize the negative log-likelihood function as the cost function. Assuming that the ranging errors of each anchor point are independent, the joint log-likelihood function is:

[0076]

[0077] Among them, p n (e n (x)) represents the error probability density function of the nth anchor point. Finally, through optimization algorithms such as PSO, the optimal position estimation is solved in the search space :

[0078]

[0079] That is, the optimal label position x under the maximum likelihood sense * .

[0080] B) PSO optimization process

[0081] PSO is a global optimization algorithm based on swarm intelligence, which simulates swarm cooperation to search for the optimal solution. Taking two-dimensional positioning as an example, the PSO process is briefly introduced in the present invention. Each particle is represented by a position x p and a velocity v p , and is initialized as:

[0082] x p (0) = x low + r()(x up - x low ), v p (0) = α(r() - 0.5)(16)

[0083] Among them, r() is a random number in [0, 1], and x up and x low are determined by the interval with a preset ε of 0.01 in the content 3) Fingerprint positioning error quantization in the offline stage, and α controls the initial velocity value.

[0084] The particle simultaneously records its own historical best position and the global best position of the swarm . In each iteration, the particle updates its velocity and position according to the following rules:

[0085]

[0086] Among them, w is the inertia weight, c1 and c2 are the self-cognition factor and the social-cognition factor respectively, and r1 and r2 are random numbers. If the particle position exceeds the boundary, the velocity is truncated and reversed.

[0087] After each iteration, the individual and global optimal solutions are updated:

[0088]

[0089] Through iterative optimization, PSO finally converges to the optimal solution.

[0090] C) Weighted positioning with fingerprint interval constraints

[0091] After PSO optimization, local optimal problems may occur, and it only relies on ToA ranging. To improve the accuracy, the present invention combines fingerprint information to screen and weight the particles.

[0092] First, eliminate the particles that exceed the fingerprint error interval (determined by presetting ε = 0.1 in the fingerprint positioning error quantization in the offline stage content 3), and only retain the finite particle set {x p ∣p ∈ S}. Then, based on the ToA ranging error modeling in the offline stage content 4), take a larger K max Construct a more complex GMM to calculate the likelihood of the ranging residual, and calculate the cost value for each effective particle x p Calculate the cost value:

[0093]

[0094] Next, use Softmax normalization to calculate the particle weights:

[0095]

[0096] Finally, use weighted average to obtain the final positioning result:

[0097]

[0098] This method improves the positioning accuracy by eliminating invalid particles and assigning greater weights to low-cost (high-likelihood) particles.

[0099] UWB, CIR fingerprint positioning, ToA ranging, PSO, CNN, Copula, GMM, CBAM, RES, NloS, and LoS applied in the present invention are all very common processing means or calculation methods in the prior art. For related content, those skilled in the art can obtain it from the relevant literature of the prior art.

Claims

1. A UWB positioning method that integrates CIR interval constraints and ToA ranging optimization, characterized in that: It includes a CIR positioning interval estimation module, a ToA ranging error module, and a fusion position prediction module; the CIR positioning interval estimation module includes a CNN model and a Copula model; the ToA ranging error module includes a GMM model; the positioning method includes: I) Offline stage: Collect a CIR training data set and use the CIR training data set to train the CNN model and the Copula model; the CNN model can output an initial positioning prediction value, and the Copula model can output a first positioning error interval and a second positioning error interval according to the initial positioning prediction value; among them, the significance level parameter ε1 required to obtain the first positioning error interval and the significance level parameter ε2 required to obtain the second positioning error interval are both preset values, and ε1 < ε2; Collect a ToA training data set and use the ToA training data set to train the GMM model; the GMM model can output an optimization objective function and a weighted cost function; II) Online stage: 1) Input the CIR data of the positioning target into the CIR positioning interval estimation module, and after processing, output the first positioning error interval and the second positioning error interval of the positioning target. Then, the CIR positioning interval estimation module transmits the first positioning error interval and the second positioning error interval to the fusion position prediction module; input the ToA data of the positioning target into the ToA ranging error module, and after processing, output the optimization objective function and the weighted cost function of the positioning target. Then, the ToA ranging error module transmits the optimization objective function and the weighted cost function to the fusion position prediction module; 2) The fusion position prediction module randomly generates multiple PSO particles within the range of the first positioning error interval; 3) The fusion position prediction module uses the PSO algorithm to iteratively optimize the generated multiple PSO particles according to the optimization objective function to obtain multiple optimized particles; 4) The fusion position prediction module performs a weighted average operation on the multiple optimized particles within the range of the second positioning error interval according to the weighted cost function to obtain the positioning prediction value of the positioning target.

2. The UWB positioning method that integrates CIR interval constraint and ToA ranging optimization according to claim 1, characterized in that: The CNN model integrates CBAM and RES.

3. The UWB positioning method that integrates CIR interval constraint and ToA ranging optimization as claimed in claim 1, wherein: The ε1 is 0.01, and the ε2 is 0.1.

Images