A method for correcting positioning error of ultra-wideband TDoA based on a transformer network

Abstract

The application belongs to the technical field of radio positioning, and specifically discloses a super wide band TDoA positioning error correction method based on a Transformer network. First, a model M based on the Transformer network is constructed. Second, training data of the model M is collected and generated, and the model M is trained. Finally, a label card is moved, time stamps of UWB signals arriving at each base station are obtained, and time differences are calculated in pairs. The position of the label card is estimated by using a classic TDoA algorithm and the time differences. The position estimation value and CIR information of each base station are input into the trained model M. Finally, the final label position after error correction is output. The method is suitable for closed and seriously shielded scenes such as indoor and underground coal mines. When the UWB signal under the NLOS condition and almost no direct distance exists, the positioning error can be stably corrected. The ranging information with poor quality but potential useful data does not need to be discarded, and the precision and robustness of TDoA positioning under serious shielding are improved.

Description

Technical Field

[0001] This invention belongs to the field of radio positioning technology, and specifically discloses an ultra-wideband TDoA positioning error correction method based on Transformer networks. Background Technology

[0002] Robust and accurate indoor positioning systems (IPS) are playing an increasingly important role in mining, medical monitoring, smart logistics, and other positioning services. Meanwhile, ultra-wideband (UWB) positioning systems stand out among IPS systems due to their centimeter-level positioning accuracy, higher temporal resolution, and better multipath resilience.

[0003] Based on the precise temporal resolution of UWB, researchers have proposed many positioning algorithms, among which the Time Difference of Arrival (TDoA) method is one. TDoA calculates multiple distance differences (DDoA) by determining the time difference of arrival of the same signal between multiple base station pairs, and then uses this DDoA value to estimate the location of the target. Each DDoA value represents a candidate location of the target on a pair of hyperbolas / planes. However, in practical applications, UWB positioning still faces the challenge of non-line-of-sight (NLOS) multipath effects. Under NLOS conditions, the direct path of the UWB signal is blocked, meaning the range estimate is the true distance value after positive offset, leading to a decrease in positioning accuracy.

[0004] To address the multipath effect problem in Non-Non-Standing Operating Systems (NLOS), numerous studies have focused on high-precision positioning under NLOS conditions, such as Two-Way Ranging. In the TWR (Transient Flow) direction, many machine learning-based NLOS judgment and error correction algorithms exist. However, current TDoA (Transient DoA) algorithms merely assess channel communication quality using features like Channel Impulse Response (CIR) or other PHYs to filter out unreliable link signals, thereby further improving positioning accuracy. Furthermore, research indicates that filtering out too much unstable UWB measurement data may lead to increased positioning errors, suggesting that discarded low-quality UWB measurement data may contain potentially valuable information, such as multipath components. Therefore, it is necessary to design a TDoA algorithm that can achieve high-precision positioning under NLOS conditions without discarding poor-quality UWB signal data. Summary of the Invention

[0005] To address the aforementioned issues, this invention discloses an ultra-wideband TDoA positioning error correction method based on Transformer networks. This method utilizes Transformer networks and existing TDoA positioning technologies, and is applicable to enclosed and heavily obstructed environments such as indoor spaces and underground coal mines. Under NLOS conditions and when UWB signals at near-no direct range are almost nonexistent, it can stably correct positioning errors without discarding ranging information of poor quality but potentially useful data, thus improving the accuracy and robustness of TDoA positioning under severe obstruction.

[0006] To achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0007] A method for correcting positioning errors in ultra-wideband TDoA based on Transformer networks includes the following steps:

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

[0009] S1. Under NLOS conditions in open environments and indoor environments with severe obstruction, deploy multiple UWB base stations respectively, collect and generate training data for model M to form a training set, and use the training data in the training set to train model M to form a validation set.

[0010] S2. Move the tag card to obtain the timestamps of UWB signals arriving at each base station and pair them up to calculate the time difference. Using the classic TDoA algorithm and Tag location estimation is performed by inputting the estimated tag location along with the channel impulse response (CIR) information of each base station into the trained model M, and outputting the final tag location after error correction.

[0011] Specifically, in step S0, a model M based on a Transformer network is constructed. Model M only uses the Transformer network encoder structure (Encoder), while the original decoder structure (Decoder) corresponding to the Transformer network is replaced by a feedforward neural network (FNN) to implement the regression task, outputting the error-corrected positioning coordinate information. The specific steps are as follows:

[0012] S00. Assume the number of UWB base stations deployed in the measurement environment is... The timestamp of each base station receiving the same UWB signal is recorded as follows: And the corresponding CIR signal is denoted as :

[0013] ;

[0014] in: The number of multipath components. and Base stations The CIR signal of the first The amplitude and delay time of the path component. Let be the Dirac impulse function. For base stations The inherent channel noise;

[0015] To facilitate inputting model M, for by The marked points are truncated and IQ sampled at 1ns intervals to convert them into discrete amplitude vectors. The corresponding dimension is denoted as ;

[0016] S01. The input information of model M is the estimated tag position value calculated by the classic TDoA algorithm. and according to Sort in ascending order ;

[0017] S02. After sorting... indivual The input can be learned embedding layer, which is a linear layer that integrates a single input. Linear mapping to a dimension of Token, a unit vector;

[0018] S03. For the product generated in step S02 Each token undergoes spatial location encoding.

[0019] S04. Randomly initialize a classification unit CLS, and compare it with the position-encoded unit CLS from step S03. All tokens are input into the Encoder to obtain the encoded result EncOutput, where the input and output dimensions of the Encoder are consistent.

[0020] The encoded output of the first vector unit of S05.EncOutput, namely CLS, is compared with the estimated tag position value. The data is fed into a feedforward neural network (FNN) to obtain the final positioning coordinates after error correction.

[0021] Specifically, the spatial location coding method in step S03 involves using sine and cosine functions combined with logarithmic interval frequency bands to transform the input location information, thereby converting the base station... The coordinate information is integrated into each token, providing model M with additional relative or absolute position information, specifically:

[0022] S030. Assume the coordinate system has dimensions of... The base station coordinates are The maximum coordinate value in the location scene is Normalizing the base station coordinates yields ;definition The frequency bands are logarithmicly distributed, and the formula is as follows:

[0023] ;

[0024] in: For the first angular frequency of each frequency band The minimum angular frequency, The maximum angular frequency;

[0025] S031. The calculation function for base station location information encoding is as follows:

[0026] ;

[0027] in: The vector dimension is ,satisfy ,Sure The maximum value is then padded with zeros to make the spatial location encoding vector have a dimension of 1. , It is a concatenation function.

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

[0029] Specifically, the feedforward neural network (FNN) in step S05 uses a multilayer perceptron network structure. The FNN includes a series of layers with 256, 128, 64, or... A fully connected layer of neurons.

[0030] Specifically, in step S1, when collecting training and validation set data for model M, the information collected, based on a single location coordinate point, includes the location coordinate information with millimeter-level precision and the timestamps of UWB signals arriving at each base station. and each base station Signal.

[0031] Specifically, in step S1, the ratio of the number of samples in the training set to the number of samples in the validation set should be between 2:1 and 3:1. The training set data uses relatively systematic straight-line trajectory data, while the validation set data uses more random trajectory data to ensure effective evaluation of the ability of the training model M to generalize and interpolate in areas not covered by training. The performance evaluation criteria for the training model M use two indicators: mean absolute error (MAE) and circular probability error (CEP).

[0032] Specifically, the ultra-wideband TDoA positioning error correction method introduces the self-attention mechanism of the Transformer network and its high efficiency in processing long sequence inputs. It only uses the Transformer encoder structure, with fixed input and output sequence dimensions. No autoregression is required during inference, thus achieving unconstrained bidirectional information flow.

[0033] Specifically, the ultra-wideband TDoA positioning error correction method corrects the error received and sampled by a single base station. The signal is input into model M as a vector sequence subunit.

[0034] Specifically, the ultra-wideband TDoA localization error correction method only inputs 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 solution, the beneficial effects of the present invention are:

[0036] This invention discloses an ultra-wideband TDoA positioning error correction method based on Transformer networks. First, a Transformer network-based model M is constructed. Second, training data for model M is collected and generated, and the model M is trained. Finally, the tag is moved to obtain the timestamps of 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's position. The estimated position value and the CIR information of each base station are input into the trained model M. Finally, the corrected tag position is output. This method is suitable for enclosed and heavily obstructed scenarios such as indoors and underground coal mines. Under NLOS conditions and when there are almost no UWB signals at direct range, it can stably correct positioning errors without discarding ranging information with poor quality but potentially useful data. It also improves the accuracy and robustness of TDoA positioning under severe obstruction conditions. This is a TDoA positioning error correction method suitable for NLOS conditions, with high robustness, strong generalization, and scalability.

[0037] This invention discloses an ultra-wideband TDoA positioning error correction method based on Transformer networks. It introduces a spatial location coding strategy and combines it with CIR signals, enabling its application to larger and more complex positioning scenarios. Unlike traditional location coding, which only describes the order of data sequences, this spatial location coding method transmits the physical geometric location information of the base station to the model M and combines it with… The signal directly corrects the TDoA positioning estimate, making model M closer to the theoretical logic of the traditional TDoA algorithm. Unlike traditional methods that rely on discarding poor CIR signal links, this method can still maintain effective error correction even when all CIR signals are of poor quality. Attached Figure Description

[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, that is, coordinate system dimensions Select a positioning environment with severe occlusion and set a corner point as the origin. The maximum coordinate value within this environment Base stations were deployed in open and NLOS conditions respectively. In this embodiment The value can be 15 and is not limited. The timestamp of each base station receiving the same UWB signal is recorded as follows. And the corresponding CIR signal is denoted as , The discrete vector after IQ sampling over a 1ns time interval is denoted as The vector dimension is denoted as In this embodiment The value is 150. Continuous data within the time interval from the first 50 ns to the last 100 ns is sampled for the marker point. The first 50 and the last 100 moments are used The estimated location of the tag is calculated by combining the coordinates of each base station with the classic TDoA algorithm. The specific process of this algorithm includes the following steps:

[0042] Step 1: Start building a model M based on the Transformer network. 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 implement the regression task and output the error-corrected positioning coordinate information.

[0043] Step Two: according to Obtain by ascending order and estimated tag location As input to model M;

[0044] Step 3: The input is a learnable embedding layer, which is a linear layer that takes a single input... Linear mapping to a dimension of Token, a unit vector;

[0045] Step 4: Process the generated data. Each token undergoes spatial location encoding, and the base station coordinates corresponding to each token are incorporated. First of all Normalization yields and define The frequency bands are logarithmicly distributed, and the formula is as follows:

[0046] ;

[0047] in: For the first angular frequency of each frequency band The minimum angular frequency, The maximum angular frequency;

[0048] The base station location information encoding calculation function is as follows: It utilizes sine and cosine functions combined with logarithmic interval frequency bands to transform the input location information:

[0049] ;

[0050] in: The vector dimension is , must meet ,Sure The maximum value is then padded with zeros to make the spatial location encoding vector have a dimension of 1. ; It is a concatenation function;

[0051] Step 5: Randomly initialize a classification unit CLS, and compare it with the position-encoded unit from Step 4. All tokens are input into the Encoder to obtain the encoded result EncOutput, where the input and output dimensions of the Encoder should be consistent;

[0052] Step Six: The encoded output of the first vector unit of EncOutput, i.e., CLS, is compared with the estimated tag position value. The input is a feedforward neural network (FNN), where the FNN uses an MLP network. The FNN includes components with 256, 128, 64, or... A fully connected layer of 1 neuron, in this embodiment The value is 3;

[0053] Step 7: The FNN outputs the final positioning coordinates after error correction;

[0054] Step 8: Collect and generate training data for model M to form a training set. Use the training data in the training set to train model M to form a validation set. The training set and validation set are collected at a single location coordinate point, and the collected information is location coordinate information with millimeter-level precision and the timestamp of UWB signal arrival at each base station. and each base station Signal; In this embodiment, the ratio of the number of samples in the training set to the number of samples in the validation set is approximately 2:1; The training set data is selected from relatively systematic straight-line trajectory data, while the validation set data uses more random trajectory data; The performance evaluation standard for the training model M uses two indicators: mean absolute error (MAE) and circular error probabilities (CEP).

[0055] Step 9: Move the tag card to get... Timestamps of each base station receiving UWB signals and the corresponding CIR signal The estimated position of the tag card is calculated using the classic TDoA algorithm. Then proceed to steps two through seven, and finally model M outputs the position coordinates of the label card after error correction.

[0056] This invention, based on radio positioning technology, incorporates the self-attention mechanism and high efficiency of Transformer networks for handling long input sequences. It combines the TDoA positioning algorithm with Transformer networks to propose an ultra-wideband TDoA positioning error correction method based on Transformer networks. This invention is applicable to high-precision ultra-wideband TDoA positioning under NLOS occlusion conditions. Unlike traditional methods that rely on discarding poor-quality CIR signal links, it maintains effective error correction even when all CIR signals are of poor quality, and can be extended to larger and more complex positioning scenarios.

Claims

1. A method for correcting positioning errors in ultra-wideband TDoA based on Transformer networks, characterized in that, Includes the following steps: S0. Construct a model M based on a Transformer network. Model M only uses the Transformer network encoder structure (Encoder), while the original decoder structure (Decoder) is replaced by a feedforward neural network (FNN) to implement the regression task. The specific steps are as follows: S00. Assume the number of UWB base stations deployed in the measurement environment is... The timestamp of each base station receiving the same UWB signal is recorded as follows: And the corresponding CIR signal is denoted as : ； in: The number of multipath components. and Base stations The CIR signal of the first The amplitude and delay time of the path component. Let be the Dirac impulse function. For base stations The inherent channel noise; right by The marked points are truncated and IQ sampled at 1ns intervals to convert them into discrete amplitude vectors. The corresponding dimension is denoted as ; S01. The input information of model M is the estimated tag position value calculated by the classic TDoA algorithm. and according to Sort in ascending order ; S02. After sorting... indivual The input can be learned embedding layer, which is a linear layer that integrates a single input. Linear mapping to a dimension of Token, a unit vector; S03. For the product generated in step S02 Each token undergoes spatial location encoding. S04. Randomly initialize a classification unit CLS, and compare it with the position-encoded unit CLS obtained in step S03. All tokens are input into the Encoder to obtain the encoded result EncOutput, where the input and output dimensions of the Encoder are consistent. The encoded output of the first vector unit of S05.EncOutput, namely CLS, is compared with the estimated tag position value. The data is fed into a feedforward neural network (FNN) to obtain the final positioning coordinates after error correction. S1. Under NLOS conditions in open environments and indoor environments with severe obstruction, deploy multiple UWB base stations respectively, collect and generate training data for model M to form a training set, and use the training data in the training set to train model M to form a validation set. S2. Move the tag card to obtain the timestamps of UWB signals arriving at each base station and pair them up to calculate the time difference. Using the classic TDoA algorithm and Tag location estimation is performed by inputting the estimated tag location and the channel impulse response information (CIR) of each base station into the trained model M, and outputting the final tag location after error correction.

2. The ultra-wideband TDoA positioning error correction method based on Transformer networks according to claim 1, characterized in that, The spatial location coding method in step S03 involves using sine and cosine functions combined with logarithmic interval frequency bands to transform the input location information, thereby converting the base station... The coordinate information is integrated into each token, providing model M with additional relative or absolute position information, specifically: S030. Assume the coordinate system has dimensions of... The base station coordinates are The maximum coordinate value in the location scene is Normalizing the base station coordinates yields ;definition The frequency bands are logarithmicly distributed, and the formula is as follows: ； in: For the first angular frequency of each frequency band The minimum angular frequency, The maximum angular frequency; S031. The calculation function for base station location information encoding is as follows: ； in: The vector dimension is ,satisfy( ) ,Sure The maximum value is then padded with zeros to make the spatial location encoding vector have a dimension of 1. , It is a concatenation function.

3. The ultra-wideband TDoA positioning error correction method based on Transformer networks according to claim 1, characterized in that, In step S04, the block structure of the Encoder is consistent with that of the Transformer, both consisting of... It consists of a stack of identical EncLayer layers, with the same number of layers as in the Transformer encoder.

4. The ultra-wideband TDoA positioning error correction method based on Transformer networks according to claim 1, characterized in that, Step S05 uses a feedforward neural network (FNN) with a multilayer perceptron network structure. The FNN includes a series of structures with frequencies of 256, 128, 64, or... A fully connected layer of neurons.

5. The ultra-wideband TDoA positioning error correction method based on Transformer networks according to claim 1, characterized in that, In step S1, when collecting training and validation set data for model M, the information collected, calculated for a single location coordinate point, includes the location coordinate information with millimeter-level precision and the timestamps of UWB signals arriving at each base station. and each base station Signal.

6. The ultra-wideband TDoA positioning error correction method based on Transformer networks according to claim 1, characterized in that, In step S1, the ratio of the number of samples in the training set to the number of samples in the validation set is between 2:1 and 3:

1. The training set data uses relatively systematic straight-line trajectory data, while the validation set data uses more random trajectory data to ensure that the ability of the training model M to generalize and interpolate in areas not covered by training is effectively evaluated. The performance evaluation criteria for the training model M use two indicators: mean absolute error (MAE) and circular probability error (CEP).

7. The ultra-wideband TDoA positioning error correction method based on Transformer networks according to claim 2, characterized in that, The ultra-wideband TDoA positioning error correction method corrects the error received and sampled by a single base station. The signal is input into model M as a vector sequence subunit.

8. The ultra-wideband TDoA positioning error correction method based on Transformer networks according to claim 2, characterized in that, The ultra-wideband TDoA localization error correction method only inputs the EncOutput encoding vector corresponding to the classification unit CLS into the FNN to reduce the complexity of the regression task.

Images