Ultra-wideband TDoA positioning error correction method based on Transform network

Abstract

The invention belongs to the technical field of radio positioning, and particularly discloses an ultra-wideband TDoA positioning error correction method based on a Transform network, which comprises the following steps: firstly, constructing a model M based on the Transform network, secondly, collecting and generating training data of the model M, training the model M, and finally, moving a tag card to obtain timestamps when UWB signals arrive at each base station, and pairwise pairing to calculate time difference, so as to correct the positioning error of the UWB signals. The label card position estimation is carried out by using a classical TDoA algorithm and a time difference, a position estimation value and CIR information of each base station are input into a trained model M, and finally a final label position after error correction is output, the method is suitable for indoor and underground coal mine closed and seriously shielded scenes, and when UWB signals in a direct distance almost do not exist under an NLOS condition, the label card position estimation method has the advantages that the label card position estimation is more accurate. The positioning error can be stably corrected, the ranging information which is poor in quality and has potential useful data does not need to be discarded, and the precision and robustness of TDoA positioning under the condition of serious shielding are improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of radio positioning, and particularly discloses a super-wideband TDoA positioning error correction method based on a Transformer network. BACKGROUND

[0002] Robust and accurate indoor positioning systems (IPS) are increasingly important in the fields of mine, medical monitoring, intelligent logistics and other positioning services. At the same time, the ultra-wideband (UWB) positioning system stands out among a large number of IPS systems due to its centimeter-level positioning accuracy, higher time resolution and better tolerance to multipath effects.

[0003] Based on the accurate time resolution of UWB, scholars have proposed many positioning algorithms, and the time difference of arrival (TDoA) is one of them. TDoA calculates the distance difference of arrival (DDoA) by calculating the time difference of arrival of the same signal between multiple base station pairs, and then calculates the position estimate of the positioning target. Each DDoA value means that the position candidate of the positioning target is on a pair of hyperbolic curves / surfaces. However, in practical applications, UWB positioning still faces the challenge of non-line-of-sight (NLOS) conditions and multipath effects. Under NLOS conditions, the direct path of the UWB signal is blocked, i.e. the ranging estimate value is the distance true value after the positive bias, which leads to a decrease in positioning accuracy.

[0004] To solve the problem of NLOS multipath effect, many research works have been devoted to high-precision positioning under NLOS conditions, such as two-way ranging (TWR) direction. There are many NLOS judgment and error correction algorithms based on machine learning, while the existing TDoA direction algorithm only evaluates the channel communication quality by channel impulse response or other PHY features to screen out unreliable link connection signals, thereby further improving the positioning accuracy. In addition, studies have shown that screening too many unstable UWB measurement data may lead to larger positioning errors, indicating that discarded low-quality UWB measurement data may contain potential useful information such as multipath components. Therefore, it is necessary to design a TDoA algorithm that does not need to discard poor-quality UWB signal data under NLOS conditions and can achieve high-precision positioning. SUMMARY

[0005] To solve the above problems, the application discloses a UWB TDoA positioning error correction method based on a Transformer network, which is based on a Transformer network and an existing TDoA positioning technology, is suitable for closed and seriously shielded scenes such as indoor and underground coal mines, and can stably correct positioning errors under NLOS conditions and almost no direct distance UWB signals, without discarding ranging information with poor quality but potential useful data, and improving the precision and robustness of TDoA positioning under serious shielding.

[0006] In order to achieve the above-mentioned application purposes, the application adopts the following technical solutions:

[0007] A UWB TDoA positioning error correction method based on a Transformer network comprises the following steps:

[0008] S0. Construct a model M based on a Transformer network;

[0009] S1. Deploy multiple UWB base stations under NLOS conditions in open environments and indoor environments with serious shielding, collect and generate training data of the model M to form a training set, and train the model M using the training data in the training set to form a verification set;

[0010] S2. Move the tag card, get the timestamps of UWB signals arriving at each base station, and calculate the time difference Δt by pairing, use the classic TDoA algorithm and Δt to estimate the position of the tag card, input the position estimation value of the tag card position and the channel impulse response CIR information of each base station into the trained model M, and output the final tag position after error correction.

[0011] Specifically, in step S0, the model M based on the Transformer network is constructed, the model M only uses a Transformer network encoder structure Encoder, and the original decoder structure Decoder corresponding to the Transformer network is replaced by a feedforward neural network FNN to realize a regression task, so as to output error-corrected positioning coordinate information, and the specific steps are as follows:

[0012] S00. Assuming that the number of UWB base stations deployed in the measurement environment is N, the timestamp of the same UWB signal received by each base station is t n (n=1,2,...,N), and the corresponding CIR signal is CIR n (t):

[0013]

[0014] Wherein: S is the number of multipath components, α s,n and τ s,nrespectively, are the amplitude and time delay of the s-th path component in the CIR signal of base station n, δ(·) is the Dirac impulse function, n n (t) is the inherent channel noise of base station n.

[0015] For the convenience of inputting the model M, the CIR n (t) is intercepted at the marker point t n with a time interval of 1 ns and converted into a discrete amplitude vector CIR n , and the corresponding dimension is denoted as D CIR .

[0016] S01. The input information of the model M is the tag card position estimate p TDoA calculated by the classical TDoA algorithm. n

[0017] S02. The sorted N CIR n s are input into the learnable embedding layer Embedding, i.e., a linear layer, which linearly maps the input single CIR n to a unit vector Token with a dimension of d model .

[0018] S03. The N Tokens generated in step S02 are subjected to spatial position coding processing.

[0019] S04. A classification unit CLS is randomly initialized and input into the Encoder together with the N Tokens subjected to the position coding processing in step S03 to obtain the coding result EncOutput, wherein the input and output dimensions of the Encoder should be consistent.

[0020] S05. The first vector unit of the EncOutput, i.e., the coding output result of the CLS, is input into the feedforward neural network FNN together with the tag card position estimate p TDoA to obtain the final positioning coordinates after error correction.

[0021] Specifically, the spatial position coding method in step S03 is to convert the input position information by using the sine and cosine functions in combination with the logarithmic interval frequency bands, and to fuse the coordinate information of the base station n into each Token to provide additional relative or absolute position information for the model M, which is specifically as follows:

[0022] S030. Assuming that the coordinate system dimension is dim, the base station coordinate is p, and the maximum coordinate value in the positioning scene is P, the base station coordinate is normalized to obtain p' = (diag(P)) -1 p; define F frequency band logarithmic distributions, and the formula is as follows: ​

[0023]

[0024] wherein: ω f is the angular frequency of the f-th frequency band, ω min is the minimum angular frequency, ω max is the maximum angular frequency;

[0025] S031. The calculation function of the base station position information encoding is as follows:

[0026]

[0027] wherein: PE spatial The vector dimension of (p') is dimxF, which needs to satisfy (dimxF)≤d model , the maximum value of F is determined, and zero padding is right filled to make the dimension of the spatial position encoding vector d model , concat(·) is a concatenation function.

[0028] Specifically, the block structure of the Encoder in step S04 is consistent with the Transformer, and both are composed of a stack of Num identical layers EncLayer, and the number of layers of EncLayer is the same as that in the Transformer encoder.

[0029] Specifically, the feedforward neural network FNN in step S05 uses a multi-layer perceptron network structure, and the FNN includes a series of fully connected layers each having 256, 128, 64, or dim neurons.

[0030] Specifically, in step S1, when collecting the training set and validation set data of the model M, the information is collected in terms of a single position coordinate point, and the information is the corresponding millimeter-level position coordinate information, the UWB signal timestamps t n to each base station, and the CIR n (t) signal of each base station.

[0031] Specifically, in step S1, the sample number ratio of the training set and the validation set should be between 2:1 and 3:1; the training set data adopts a more systematic straight line trajectory data, and the validation set data adopts a more random trajectory data, so as to ensure the ability of the trained model M to generalize and interpolate in the area not covered by the training, and the performance evaluation standard of the trained model M adopts two indexes of mean absolute error MAE and circular error probability CEP.

[0032] Specifically, the ultra-wideband TDoA positioning error correction method introduces the self-attention mechanism of the Transformer network and the efficiency of processing long sequence input, only uses the Transformer encoder structure, the input and output sequence dimensions are fixed, and there is no need for autoregression in the reasoning process, and the unconstrained bidirectional information flow is realized.

[0033] Specifically, the ultra-wideband TDoA positioning error correction method inputs the CIR signal received and sampled by a single base station as a vector sequence subunit into the model M. n Signal.

[0034] Specifically, the ultra-wideband TDoA positioning error correction method only transmits the EncOutput encoding vector corresponding to the classification unit CLS into the FNN to reduce the complexity of the regression task.

[0035] Due to the adoption of the above technical solutions, the present application has the following advantages:

[0036] The ultra-wideband TDoA positioning error correction method based on the Transformer network disclosed in the present application firstly constructs a model M based on the Transformer network, secondly collects and generates training data of the model M, trains the model M, finally moves the label card, obtains the timestamps of the UWB signals arriving at each base station, and calculates the time difference by pairing, estimates the position of the label card by using the classic TDoA algorithm and the time difference, inputs the position estimation value and the CIR information of each base station into the trained model M, and finally outputs the final label position after error correction, which is suitable for closed and severely occluded scenes such as indoor and underground coal mines, and can stably correct the positioning error under NLOS conditions and almost no direct distance UWB signal, without discarding the ranging information with poor quality but potential useful data, and improving the precision and robustness of TDoA positioning under severe occlusion, which is a TDoA positioning error correction method suitable for NLOS conditions, high robustness, strong generalization, and scalability.

[0037] The ultra-wideband TDoA positioning error correction method based on the Transformer network disclosed in the present application introduces a spatial position encoding strategy and combines CIR signals, which can be applied to larger and more complex positioning scenes, unlike the traditional position encoding which only describes the order of data sequences, the present spatial position encoding method transmits the physical geometric position information of the base station to the model M, and combines the CIR n Signal directly corrects the TDoA positioning estimation value, so that the model M is closer to the theoretical logic of the traditional TDoA algorithm, and unlike the traditional method which relies on discarding poor CIR signal links, it can still maintain effective error correction even in the case of poor CIR signal quality. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 This is a diagram showing the overall algorithm flow and network structure of the M-model of this invention. Detailed Implementation

[0039] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments:

[0040] like Figure 1 As shown, an ultra-wideband TDoA positioning error correction method based on Transformer networks is proposed. First, a Transformer network-based model M is constructed. Multiple UWB base stations are deployed under NLOS conditions, including open environments and indoor environments with severe obstruction. Training data for model M is collected and generated to form a training set. The training data in the training set is used to train model M to form a validation set. Finally, the tag is moved to obtain the timestamps of the UWB signals arriving at each base station, and the time difference is calculated by pairing them. The classic TDoA algorithm and the time difference are used to estimate the tag position. The estimated tag position value and the CIR information of each base station are input into the trained model M, and the final tag position after error correction is output.

[0041] Taking UWB 3D positioning as an example, where the coordinate system dimension dim = 3, a positioning environment with severe occlusion is selected, and a corner point is set as the origin (0,0,0). The maximum coordinate value within this environment is P = [XYZ]. T N base stations are deployed under both open and NLOS conditions. In this embodiment, N is taken as 15 and is not limited. The timestamp of each base station receiving the same UWB signal is recorded as t. n (n = 1, 2, ..., N) and their corresponding CIR signals are denoted as CIR. n (t), CIR n The discrete vector (t) after sampling by IQ over a 1ns time interval is denoted as CIR. n The vector dimension is denoted as D. CIR In this embodiment, D CIR The value is 150, in t n For each marker point, continuous data within the time interval from the first 50 ns to the last 100 ns is sampled, i.e., t n For the first 50 and the last 100 times, use t n The estimated tag location p is calculated using the coordinates of each base station and the classic TDoA algorithm. TDoA The specific process of this algorithm includes the following steps:

[0042] Step one: Start to build the model M based on the Transformer network, the model M only uses the Transformer network encoder structure Encoder, and the original decoder structure Decoder corresponding to the Transformer network is replaced by a feedforward neural network FNN to realize the regression task, so as to output the positioning coordinate information after error correction;

[0043] Step two: CIR n is processed in ascending order according to t n to obtain and the label card position estimation value p TDoA as the input of the model M;

[0044] Step three: The input single CIR n is linearly mapped to a unit vector Token with a dimension of d model = 256 by passing through the learnable embedding layer Embedding, that is, a linear layer;

[0045] Step four: The generated N Tokens are processed by spatial position coding, and the base station coordinates corresponding to each Token are introduced into p = [x y z] T , First, normalize p to obtain and define the log distribution of F frequency bands as follows:

[0046]

[0047] Where: ω f is the angular frequency of the fth frequency band, ω min is the minimum angular frequency, and ω max is the maximum angular frequency;

[0048] The input position information is converted by using the sine and cosine functions combined with the log interval frequency band, and the encoding calculation function of the base station position information is as follows:

[0049]

[0050] Where: PE spatial (p') is a vector with a dimension of dim x F, which needs to satisfy (dim x F) ≤ d model , Determine the maximum value of F, and right-pad with zeros to make the dimension of the spatial position coding vector d model ; concat(·) is the concatenation function;

[0051] Step five: randomly initialize a classification unit CLS, input the N tokens after position encoding in step four into the encoder to obtain the encoding result EncOutput, wherein the input and output dimensions of the encoder should be consistent;

[0052] Step six: the encoding output result of the first vector unit of EncOutput, i.e., CLS, is combined with the label card position estimate p TDoA into a feedforward neural network FNN, wherein the FNN uses an MLP network, the FNN includes fully connected layers with 256, 128, 64 or dim neurons respectively, and in the embodiment, dim is 3;

[0053] Step seven: the FNN outputs the final positioning coordinates after error correction.

[0054] Step eight: collect and generate training data of the model M to form a training set, train the model M using the training data in the training set to form a validation set, the training set and the validation set are collected in units of single position coordinates, and the collected information is position coordinate information with millimeter-level precision, UWB signal timestamps t n and CIR of each base station n (t) signals; in the embodiment, the sample quantity ratio of the training set and the validation set is about 2:1; the training set data selects relatively systematic straight-line trajectory data, and the validation set data adopts more random trajectory data; two indexes, mean absolute error MAE and circular error probable CEP, are used as the performance evaluation standards of the training model M.

[0055] Step nine: move the label card to obtain the timestamps t n (n=1, 2,..., N) of the UWB signals received by the N base stations, and the corresponding CIR signals CIR n (t), and use the classic TDoA algorithm to solve the label card position estimate p TDoA Then, steps two to seven are performed, and finally the model M outputs the position coordinates of the label card after error correction.

[0056] The application is based on radio positioning technology, introduces the self-attention mechanism of the Transformer network and the efficiency of processing long sequence input, combines the TDoA positioning algorithm and the Transformer network, and proposes a UWB TDoA positioning error correction method based on the Transformer network. The application is suitable for high-precision UWB TDoA positioning under NLOS shielding conditions, and unlike the traditional method which relies on discarding poor CIR signal links, the application can still maintain effective error correction even in the case of all poor CIR signal quality, and can be applied to larger and more complex positioning scenarios.

Claims

1. A method for correcting positioning error of ultra-wideband TDoA based on a Transformer network, characterized in that, Comprising the following steps: S0. Constructing a model M based on a Transformer network; S1. Deploying multiple UWB base stations in an open environment and an indoor environment with severe occlusion under NLOS conditions, collecting and generating training data for the model M to form a training set, and training the model M using the training data in the training set to form a validation set; S2. Moving the tag card, obtaining the timestamps of the UWB signals arriving at each base station, and calculating the time difference Δt pairwise, using the classic TDoA algorithm and Δt to estimate the position of the tag card, inputting the estimated value of the tag card position and the channel impulse response information CIR of each base station into the trained model M, and outputting the final tag position after error correction.

2. The method of claim 1, wherein the method is a Transformer network-based ultra- wideband TDoA positioning error correction method, characterized in that, In step S0, the model M based on the Transformer network is constructed, and the model M only uses the Transformer network encoder structure Encoder. The original decoder structure Decoder corresponding to the Transformer network is replaced by a feedforward neural network FNN to realize the regression task. The specific steps are as follows: S00. Suppose the number of UWB base stations deployed in the measurement environment is N, and the time stamp of each base station receiving the same UWB signal is t n (n = 1, 2,..., N), and the corresponding CIR signal is denoted as CIR n (t): where S is the number of multipath components, α s,n and τ s,n are the amplitude and delay time of the s path component in the CIR signal of base station n, δ(·) is the Dirac impulse function, n n (t) is the inherent channel noise of base station n; CIR n (t) is intercepted at the marker point t n and converted into a discrete amplitude vector CIR n with the corresponding dimension D CIR ; S01. The input information of the model M is the tag card position estimate value p calculated by the classical TDoA algorithm TDoA and according to t n ascending order S02. Sort the N CIRs. n The input can be learned embedding layer, i.e., a linear layer, which embeds a single CIR of the input. n Linear mapping to a dimension d model Token, a unit vector; S03. Spatial position coding processing is performed on the N tokens generated in step S02; S04. Randomly initialize a classification unit CLS, and input it together with the N tokens after position coding processing in step S03 into the Encoder to obtain the coding result EncOutput, wherein the input and output dimensions of the Encoder should be consistent; S05. The encoding output result of the 1st vector unit of the EncOutput, i.e. CLS, is compared with the label card position estimation value p TDoA The incoming feedforward neural network FNN, to obtain the final positioning coordinates after error correction. 3.The method of claim 2, wherein the method is characterized in that, The spatial position coding method in step S03 is to use the sine and cosine functions and combine the logarithmic interval band to convert the input position information, and fuse the coordinate information of base station n into each Token, to provide additional relative or absolute position information for the model M. Specifically: S030.Assume the dimension of coordinate system is dim, the coordinate of base station is p, the largest coordinate value in positioning scenario is P, and the normalized coordinate of base station is p'=(diag(P)) -1p -1 p; define the logarithmic distribution of F frequency bands, and the formula is as follows: where: ω f is the angular frequency of the fth frequency band, ω min is the minimum angular frequency, ω max is the maximum angular frequency; S031. The calculation function of the base station position information coding is as follows: where: PE spatial The vector dimension of (p') is dimxF, which needs to satisfy (dimxF)≤d model The maximum value of F is determined, and the right padding is filled with zeros to make the dimension of the spatial position encoding vector d model concat(·) is the concatenation function.

4. The method of claim 2, wherein the method is characterized by, The block structure of the Encoder in step S04 is consistent with the Transformer, and both are composed of a stack of Num identical layers EncLayer, which has the same number of layers as the Transformer encoder.

5. The method of claim 2, wherein the method is characterized by, The feedforward neural network FNN in step S05 uses the Multilayer Perceptron network structure, and the FNN includes a series of fully connected layers with 256, 128, 64, or dim neurons respectively.

6. The method of claim 1, wherein the method is characterized by, In step S1, when collecting the training set and the verification set data of the model M, the information collected is the position coordinate information corresponding to the millimeter-level precision, the UWB signal timestamps t to each base station n and the CIR of each base station n (t) signal.

7. The method of claim 1, wherein the method is characterized by, In step S1, the sample number ratio of the training set and the validation set is between 2:1 and 3:1; the training set data uses more systematic straight trajectory data, and the validation set data uses more random trajectory data, to ensure the ability of the trained model M to generalize and interpolate in the untrained coverage area. The performance evaluation standard of the trained model M uses two indicators of mean absolute error MAE and circular error probability CEP.

8. The method of claim 2, wherein the method is based on a Transformer network. The super wideband TDoA positioning error correction method introduces the self-attention mechanism of the Transformer network and the high efficiency of processing long sequence input, only uses the Transformer encoder structure, the input and output sequence dimensions are fixed, and there is no need for autoregression in the reasoning process, and the unconstrained bidirectional information flow can be realized. 9.The method of claim 2, wherein the method is characterized in that, The ultra-wideband TDoA positioning error correction method receives and samples the CIR from a single base station n The signal is input into the model M as a sequence of vectors subunit.

10. The method of claim 2, wherein the method is characterized by, The super wideband TDoA positioning error correction method only transmits the EncOutput encoding vector corresponding to the classification unit CLS into the FNN, so as to reduce the complexity of the regression task.

Images