Single base station error suppression and positioning method and system based on unsupervised learning

Abstract

The invention provides a single base station error suppression and positioning method based on unsupervised learning, which automatically models to reach error distribution of a phase difference without label data through a deep learning network, and does not need manual labeling. The system extracts signal features, and non-line-of-sight propagation errors are effectively identified and corrected by using unsupervised learning. Compared with a traditional multi-base-station method, high-precision positioning can be achieved in a complex indoor environment only through a single base station. The method combines the feature extraction capability of deep learning and the distribution modeling of unsupervised learning, significantly reduces the deployment cost, improves the positioning precision and robustness, is especially suitable for scenes with significant dynamic and multipath effects, provides an innovative solution for ultra-wideband positioning, and has a wide application prospect.

Description

Technical Field

[0001] The invention relates to the field of radio signal processing, and in particular to a single base station error suppression and positioning method and system based on unsupervised learning. Background Art

[0002] Ultra-wideband (UWB) positioning technology has received extensive attention in recent years due to its advantages such as high precision, low power consumption, and strong anti-interference ability. In complex indoor environments, UWB positioning technology can effectively overcome the performance degradation of traditional wireless positioning technology under multipath propagation and noise interference, and its positioning accuracy can usually reach centimeter level. This makes UWB technology have broad application prospects in industrial automation, intelligent logistics, medical health monitoring, smart home and other fields.

[0003] In the UWB single-base station positioning scenario, the positioning method based on the phase difference of arrival (PDoA) is particularly important. PDoA receives the phase information of the signal through the array antenna. Only one base station is needed to calculate the relative angle and distance information of the target signal, thereby estimating the target position. However, the positioning accuracy of PDoA is easily affected by non-line-of-sight (NLOS) propagation and noise interference. In indoor environments, the signal propagation path is easily blocked by walls, furniture, and dynamic obstacles, causing the signal to experience multiple reflections, refractions, and diffractions, thereby introducing multipath effects. The NLOS effect will significantly affect the phase consistency of the signal, thereby reducing the accuracy of PDoA measurement. In addition, electronic equipment, RF interference sources and thermal noise of the system hardware in the environment will have a negative impact on the signal reception quality, causing the PDoA measurement value to deviate from the true value. Existing methods usually use supervised learning machine learning and deep learning methods to solve NLOS identification and regression prediction problems. For example, the "A UWB-based Indoor Fingerprint Positioning Method" (publication number: CN 117056818) applied by Chongqing Jiaotong University uses a slight NLOS ranging error regression model to estimate the ranging error of the slight NLOS link and correct the slight NLOS ranging. The "UWB NLOS Identification and Ranging Error Compensation Method Based on Feature Similarity" (publication number: CN 119004133) applied by Southeast University uses a dual-input attention neural network model for NLOS identification and error compensation to achieve UWB ranging error mitigation. However, the above-mentioned traditional error suppression methods often rely on a large amount of labeled data or empirical formulas, and these methods have the following limitations in practical applications. For example, the acquisition cost of labeled data is high and it is difficult to cover all complex scenarios. The empirical formulas are difficult to adapt to dynamic environmental changes and have limited generalization capabilities. In addition, traditional methods perform poorly when faced with high-dimensional nonlinear error distributions.

[0004] Therefore, it is urgent to improve the existing technology. Summary of the invention

[0005] The ultra-wideband single-base station PDoA positioning technology is easily affected by NLOS propagation and noise interference, which leads to a significant decrease in the phase difference and flight time measurement accuracy, and then causes the problem of positioning result drift. At the same time, the existing PDoA error suppression method usually relies on a large amount of labeled data or empirical formulas, and has limitations such as insufficient adaptability in practical applications. In order to solve the above problems, the present invention proposes a single-base station error suppression and positioning method based on unsupervised learning. The method uses a deep learning network to automatically learn the potential distribution law of PDoA errors based on unlabeled data, and can achieve error modeling and correction without relying on labeled data. In addition, the method only needs to deploy one base station to achieve high positioning accuracy, has significant cost advantages and extremely low deployment complexity, and is suitable for a variety of complex positioning scenarios.

[0006] The technical solution adopted by the present invention to solve its technical problem is:

[0007] A single base station error suppression and positioning method based on unsupervised learning comprises the following steps:

[0008] S1: UWB single base station data reading, first receive the original features of the PDoA signal, including phase difference, amplitude information, signal strength and signal-to-noise ratio, and collect environmental context features, that is, two-way channel impulse response data r(t):

[0009]

[0010] Where η c and α c is the fading coefficient and time delay of the cth path, C is the total number of signal propagation paths, and n(t) is the variance Additive Gaussian white noise;

[0011] S2: Data preprocessing, standardizing all features, unifying the input range, and converting the two-channel channel impulse response data into a time-frequency graph S through short-time Fourier transform 1 (t,f) and S 2 (t,f);

[0012] Then, calculate the cross power spectrum between the two:

[0013]

[0014] in, YesS 2 conjugation of;

[0015] Finally, the PDoA original signal feature vector X is obtained 1 The two-dimensional matrix X of the cross power spectrum 2 Represents as input X;

[0016] S3: Construct the unsupervised deep learning network encoder and decoder. First, use the deep learning network as the encoder Encoder(·) to extract the latent space representation Z of the input feature X. latent :

[0017] Z latent =Encoder(X) (3)

[0018] Get the latent space representation Z of the input feature latent After that, we construct a self-supervised target to learn the internal pattern of signal features. We first use the decoder Decoder(·) to reconstruct the input features:

[0019]

[0020] Then, through feature contrast learning, the ability to distinguish signal features is strengthened;

[0021] S4: construct an unsupervised deep learning network error regression module, and use latent space features to directly predict the PDoA error value. The error regression module includes two fully connected layers and a regression layer. A batch normalization layer, a linear rectification activation function and a dropout layer are added between the two fully connected layers. The number of hidden layers in the first fully connected layer is much larger than the number of hidden layers in the second fully connected layer. The linear rectification function is after the batch normalization layer, and the dropout layer is after the linear rectification function.

[0022] The latent space after feature extraction is represented by Z latent First input the first fully connected layer and get the output:

[0023] F 1 =W 1 F BAM +b 1 (5)

[0024] Among them, W 1 is the weight of the first fully connected layer, b 1 is the bias of the first fully connected layer;

[0025] Next, input the batch normalization layer BN(·) for batch normalization:

[0026] F BN =BN(F 1 ) (6)

[0027] Apply the linear rectification activation function ReLU(·) to the batch normalized features again:

[0028] F ReLU =ReLU(F BN ) (7)

[0029] Random neuron masking is performed through the dropout layer Dropout(·):

[0030] F Dropout = Dropout(F ReLU ) (8)

[0031] F Dropout Input the second fully connected layer, prediction error:

[0032] Δ PDOA =W 2 F Dropout +b 2 (9)

[0033] Among them, W 2 is the weight of the second fully connected layer, b 2 is the bias of the second fully connected layer;

[0034] S5: Set the loss function for unsupervised learning, first reconstruction loss, so that the decoder accurately reconstructs the input signal:

[0035]

[0036] Secondly, set the contrast loss to strengthen the consistency between similar signal features:

[0037]

[0038] Among them, Sim is the feature similarity measure;

[0039] Set the regression loss again, design a pseudo-label generation mechanism based on distribution matching, and use pseudo-labels to guide error regression:

[0040]

[0041] in, is the pseudo-label error generated from the distribution estimate, is the error prediction value output by the regression network;

[0042] Then set the regularization error to constrain the sparsity and continuity of the latent space:

[0043] L reg_latent =||Z latent || 1 (13)

[0044] The total loss function is obtained as:

[0045] L=λ 1 L rec +λ 2 L contrast +λ3 L reg +λ 4 L reg_latent (14)

[0046] Among them, λ 1 ,λ 2 ,λ 3 and λ 4 To adjust the hyperparameters of each loss weight;

[0047] S6: Non-line-of-sight error value predicted by unsupervised learning Directly used to correct the raw PDoA measurement data:

[0048] First, use the predicted error value to correct the phase difference:

[0049]

[0050] Calculate the final position based on the corrected phase difference and base station coordinates:

[0051] P final =f(P base ,φ corrected ,d ac ) (16)

[0052] Among them, P base is the base station location, f is the positioning solution function, d ac is the flight time distance between the transmitting antenna A and the receiving antenna on the tag;

[0053] Furthermore, the encoder can be divided into three steps: feature extraction, feature concatenation, and key feature extraction;

[0054] The feature extraction can be further divided into two independent branches: PDoA feature extraction and environmental feature extraction;

[0055] The PDoA feature extraction uses a multi-layer fully connected network FCN (·) to extract the nonlinear feature representation of the original PDoA signal:

[0056] F PDOA =FCN(X 1 ) (17)

[0057] The environmental feature extraction uses a deep residual network ResNet(·) to extract the spatial features of the time-frequency graph, and uses a bidirectional long short-term memory network BiLSTM(·) to capture the temporal dynamic characteristics of the signal:

[0058] F env =BiLSTM(ResNet(X 2 )) (18)

[0059] After completing PDoA feature extraction and environmental feature extraction, the features of the two branches are spliced:

[0060] F fusion =Concat(F PDOA ,F env ) (19)

[0061] The key feature extraction is achieved through the bottleneck attention mechanism, which focuses on the feature map F through the spatial dimension and the channel dimension respectively. fusion The key area in the output is the extracted key feature F BAM , that is, input feature X 1 and X 2 The latent space representation Z latent ;

[0062] Furthermore, the positioning solution function f first converts the phase difference φ after the mitigation corrected Converted into distance difference Δd:

[0063] Δd=φ corrected ×c / 2 (20)

[0064] Where c is the propagation speed of radio in the current environment;

[0065] Get the final position x of the label final and final coordinate:

[0066]

[0067] Furthermore, the pseudo label generation mechanism is designed based on distribution matching, and pseudo labels are generated by modeling the probability distribution of observed data. The specific implementation steps are as follows:

[0068] S51: First input the unlabeled data X f ={x 1 ,x 2 ,...,x n}, each data point x i It is an observed feature. The true label Y corresponding to the data is not available, so a pseudo label needs to be generated.

[0069] S52: Use the Gaussian mixture model to fit the distribution of unlabeled data to obtain the probability distribution of the data. The Gaussian mixture model assumes that the data obeys a weighted combination of K Gaussian distributions:

[0070]

[0071] Where K is the number of mixture components, π k is the mixing coefficient of the kth Gaussian component, satisfying N(x|μ k ,δ k ) is the kth Gaussian component with mean μ k , with covariance δ k ;

[0072] S53: Training the Gaussian mixture model using the maximum expectation algorithm, including the following steps:

[0073] Step E: Calculate the posterior probability that each data point belongs to each mixture component:

[0074]

[0075] Among them, γ i,k Represents data point x i The probability of belonging to the kth component;

[0076] Step M: Update Gaussian mixture model parameters π k , μ k ,δ k ;

[0077] S54: For each data point x i , according to the posterior probability γ i,k , select the component with the largest probability as the pseudo label:

[0078]

[0079] in, is the data point x i Pseudo labels of

[0080] S55: Distribution matching, the generated pseudo labels are used for downstream tasks, and the distribution regularization of the pseudo labels is introduced to ensure that the model prediction is consistent with the distribution learned by the Gaussian mixture model:

[0081]

[0082] where KL(·) is the KL divergence function, is the label distribution predicted by the model, and p(x) is the distribution learned by the Gaussian mixture model;

[0083] Furthermore, the bottleneck attention includes two branches: channel attention and spatial attention;

[0084] The channel attention is encoded into a one-dimensional feature vector using global average pooling GAP(·); then, the one-dimensional feature vector is reduced in dimension by the fully connected layer FC1(·), and nonlinearly processed by the rectified linear activation function ReLU(·); then, the fully connected layer FC2(·) is used for dimensionality increase, and finally, the channel weight is obtained by batch normalization BN(·):

[0085] W channel =BN(FC2(ReLU(BN(FC1(GAP(F fusion )))))) (27)

[0086] The spatial attention performs multiple convolution operations Conv2D(·) on the input features to generate a spatial weight map:

[0087] W spatial =Conv2D(F fusion ) (28)

[0088] Finally, update the output feature of the bottleneck attention mechanism, that is, the input feature X 1 and X 2 The latent space representation of is:

[0089]

[0090] Among them, ⊙ represents element-by-element multiplication, Indicates broadcast extension;

[0091] In addition, the present invention also provides a single base station error suppression and positioning system based on unsupervised learning, including a signal receiving module, a data processing and feature extraction module, an unsupervised error suppression module and a positioning solution module;

[0092] The signal receiving module collects the original signal through the array antenna, and the intervals between the antenna arrays are the same, and the intervals are d aten Less than half of the ultra-wideband radio wave wavelength λ, i.e. d aten <λ / 2;

[0093] The data processing and feature extraction module receives the collected original signal, completes the signal preprocessing and feature construction, and expands the low-dimensional observation signal into multi-dimensional feature data;

[0094] After receiving the multi-dimensional feature data after feature extraction, the unsupervised error suppression module uses a distribution-based modeling method to perform feature learning and error mode on the observed signal; by modeling the feature distribution of the signal, the error mode caused by non-line-of-sight or environmental noise is captured, an error correction value is generated, and the error is adaptively suppressed;

[0095] The positioning solution module combines the phase difference value after error correction to estimate the position through an ultra-wideband single base station positioning algorithm;

[0096] The beneficial effects of the present invention are mainly manifested in:

[0097] (1) Through unsupervised learning, the received signal features are distributed and modeled, and the non-line-of-sight propagation mode hidden in the signal features is deeply explored, and errors can be effectively identified and corrected without annotating data. Compared with traditional empirical formulas or rule design methods, this technology can adapt to complex environmental changes in a variety of dynamic indoor scenes, solve the multipath effect and signal obstruction problems caused by non-line-of-sight signals, and improve the robustness and adaptability of single-base station UWB positioning accuracy;

[0098] (2) Using convolutional neural networks and bottleneck attention mechanisms in deep learning, the system can extract multi-dimensional features from signal data, focus on important information channels, and filter out noise and irrelevant signals. Compared with traditional low-dimensional feature processing methods, this technology significantly enhances the ability to characterize high-dimensional signal features, and can fully suppress the impact of environmental noise, dynamic interference sources, and system hardware errors on PDoA measurement, thereby improving the efficiency and reliability of signal processing;

[0099] (3) Using single-base station UWB positioning and unsupervised models to generate pseudo-error labels, the system can complete error correction without considering the configuration optimization problem of different base station arrangements, nor relying on expensive and difficult-to-obtain large amounts of labeled data. Compared with traditional supervised learning methods, the pseudo-labeling mechanism of unsupervised learning greatly reduces the cost of data labeling and demonstrates superior generalization ability in mining the distribution of unlabeled data. BRIEF DESCRIPTION OF THE DRAWINGS

[0100] Figure 1 It is a flow chart of a single base station error suppression and positioning method based on unsupervised learning of the present invention;

[0101] Figure 2 It is an operation flow chart of the encoder in the single base station error suppression and positioning method of the present invention;

[0102] Figure 3 It is the composition and specific network structure diagram of the encoder of the present invention;

[0103] Figure 4 It is a schematic diagram of the principle of positioning using the phase difference of arrival (PDoA) of a single base station;

[0104] Figure 5 It is an operation flow chart of the pseudo label generation mechanism designed based on distribution matching of the present invention;

[0105] Figure 6 This is a schematic diagram of the principle of bottleneck attention;

[0106] Figure 7 It is a schematic diagram of the composition of a single base station error suppression and positioning system based on unsupervised learning. DETAILED DESCRIPTION

[0107] In order to facilitate the understanding of the present invention, the present invention will be described more comprehensively and meticulously below in conjunction with the accompanying drawings and preferred embodiments of the present invention, but the protection scope of the present invention is not limited to the following specific embodiments.

[0108] like Figure 1 As shown, a single base station error suppression and positioning method based on unsupervised learning includes the following steps:

[0109] S1: UWB single base station data reading, first receive the original features of the PDoA signal, including phase difference, amplitude information, signal strength and signal-to-noise ratio, and collect environmental context features, that is, two-way channel impulse response data r(t):

[0110]

[0111] Where η c and α c is the fading coefficient and time delay of the cth path, C is the total number of signal propagation paths, and n(t) is the variance Additive Gaussian white noise;

[0112] S2: Data preprocessing, standardizing all features, unifying the input range, and converting the two-channel channel impulse response data into a time-frequency graph S through short-time Fourier transform 1 (t,f) and S 2 (t,f);

[0113] Then, calculate the cross power spectrum between the two:

[0114]

[0115] in, YesS 2 conjugation of;

[0116] Finally, the PDoA original signal feature vector X is obtained 1 The two-dimensional matrix X of the cross power spectrum 2 Represents as input X;

[0117] S3: Construct the unsupervised deep learning network encoder and decoder. First, use the deep learning network as the encoder Encoder(·) to extract the latent space representation Z of the input feature X. latent :

[0118] Z latent =Encoder(X) (3)

[0119] Get the latent space representation Z of the input feature latent After that, we construct a self-supervised target to learn the internal pattern of signal features. We first use the decoder Decoder(·) to reconstruct the input features:

[0120]

[0121] Then, through feature contrast learning, the ability to distinguish signal features is strengthened;

[0122] S4: construct an unsupervised deep learning network error regression module, and use latent space features to directly predict the PDoA error value. The error regression module includes two fully connected layers and a regression layer. A batch normalization layer, a linear rectification activation function and a dropout layer are added between the two fully connected layers. The number of hidden layers in the first fully connected layer is much larger than the number of hidden layers in the second fully connected layer. The linear rectification function is after the batch normalization layer, and the dropout layer is after the linear rectification function.

[0123] The latent space after feature extraction is represented by Z latent First input the first fully connected layer and get the output:

[0124] F 1 =W 1 F BAM +b 1 (5)

[0125] Among them, W 1 is the weight of the first fully connected layer, b 1 is the bias of the first fully connected layer;

[0126] Next, input the batch normalization layer BN(·) for batch normalization:

[0127] F BN =BN(F 1 ) (6)

[0128] Apply the linear rectification activation function ReLU(·) to the batch normalized features again:

[0129] F ReLU =ReLU(F BN ) (7)

[0130] Random neuron masking is performed through the dropout layer Dropout(·):

[0131] F Dropout = Dropout(F ReLU ) (8)

[0132] F DropoutInput the second fully connected layer, prediction error:

[0133] Δ PDOA =W 2 F Dropout +b 2 (9)

[0134] Among them, W 2 is the weight of the second fully connected layer, b 2 is the bias of the second fully connected layer;

[0135] S5: Set the loss function for unsupervised learning, first reconstruction loss, so that the decoder accurately reconstructs the input signal:

[0136]

[0137] Secondly, set the contrast loss to strengthen the consistency between similar signal features:

[0138]

[0139] Among them, Sim is the feature similarity measure;

[0140] Set the regression loss again, design a pseudo-label generation mechanism based on distribution matching, and use pseudo-labels to guide error regression:

[0141]

[0142] in, is the pseudo-label error generated from the distribution estimate, is the error prediction value output by the regression network;

[0143] Then set the regularization error to constrain the sparsity and continuity of the latent space:

[0144] L reg_latent =||Z latent || 1 (13)

[0145] The total loss function is obtained as:

[0146] L=λ 1 L rec +λ 2 L contrast +λ 3 L reg +λ 4 L reg_latent (14)

[0147] Among them, λ 1 ,λ 2 ,λ 3 and λ 4To adjust the hyperparameters of each loss weight;

[0148] S6: Non-line-of-sight error value predicted by unsupervised learning Directly used to correct the raw PDoA measurement data:

[0149] First, use the predicted error value to correct the phase difference:

[0150]

[0151] Calculate the final position based on the corrected phase difference and base station coordinates:

[0152] P final =f(P base ,φ corrected ,d ac ) (16)

[0153] Among them, P base is the base station location, f is the positioning solution function, d ac is the flight time distance between the transmitting antenna A and the receiving antenna on the tag;

[0154] like Figure 2 As shown, the encoder can be divided into three steps: feature extraction, feature concatenation and key feature extraction;

[0155] like Figure 3 As shown, the feature extraction can be further divided into two independent branches: PDoA feature extraction and environmental feature extraction;

[0156] The PDoA feature extraction uses a multi-layer fully connected network FCN (·) to extract the nonlinear feature representation of the original PDoA signal:

[0157] F PDOA =FCN(X 1 ) (17)

[0158] The environmental feature extraction uses a deep residual network ResNet(·) to extract the spatial features of the time-frequency graph, and uses a bidirectional long short-term memory network BiLSTM(·) to capture the temporal dynamic characteristics of the signal:

[0159] F env =BiLSTM(ResNet(X 2 )) (18)

[0160] After completing PDoA feature extraction and environmental feature extraction, the features of the two branches are spliced:

[0161] F fusion =Concat(F PDOA ,F env) (19)

[0162] The key feature extraction is achieved through the bottleneck attention mechanism, which focuses on the feature map F through the spatial dimension and the channel dimension respectively. fusion The key area in the output is the extracted key feature F BAM , that is, input feature X 1 and X 2 The latent space representation Z latent ;

[0163] like Figure 4 As shown, the positioning solution function f first converts the phase difference φ after the alleviation corrected Converted into distance difference Δd:

[0164] Δd=φ corrected ×c / 2 (20)

[0165] Where c is the propagation speed of radio in the current environment;

[0166] Get the final position x of the label final and final coordinate:

[0167]

[0168]

[0169] like Figure 5 As shown, the pseudo label generation mechanism is designed based on distribution matching, and pseudo labels are generated by modeling the probability distribution of observed data. The specific implementation steps are as follows:

[0170] S51: First input the unlabeled data X f ={x 1 ,x 2 ,...,x n}, each data point x i It is an observed feature. The true label Y corresponding to the data is not available, so a pseudo label needs to be generated.

[0171] S52: Use the Gaussian mixture model to fit the distribution of unlabeled data to obtain the probability distribution of the data. The Gaussian mixture model assumes that the data obeys a weighted combination of K Gaussian distributions:

[0172]

[0173] Where K is the number of mixture components, π k is the mixing coefficient of the kth Gaussian component, satisfying N(x|μ k ,δ k) is the kth Gaussian component with mean μ k , with covariance δ k ;

[0174] S53: Training the Gaussian mixture model using the maximum expectation algorithm, including the following steps:

[0175] Step E: Calculate the posterior probability that each data point belongs to each mixture component:

[0176]

[0177] Among them, γ i,k Represents data point x i The probability of belonging to the kth component;

[0178] Step M: Update Gaussian mixture model parameters π k , μ k ,δ k ;

[0179] S54: For each data point x i , according to the posterior probability γ i,k , select the component with the largest probability as the pseudo label:

[0180]

[0181] in, is the data point x i Pseudo labels of

[0182] S55: Distribution matching, the generated pseudo labels are used for downstream tasks, and the distribution regularization of the pseudo labels is introduced to ensure that the model prediction is consistent with the distribution learned by the Gaussian mixture model:

[0183]

[0184] where KL(·) is the KL divergence function, is the label distribution predicted by the model, and p(x) is the distribution learned by the Gaussian mixture model;

[0185] like Figure 6 As shown, the method includes two branches: channel attention and spatial attention;

[0186] The channel attention is encoded into a one-dimensional feature vector using global average pooling GAP(·); then, the one-dimensional feature vector is reduced in dimension by the fully connected layer FC1(·), and nonlinearly processed by the rectified linear activation function ReLU(·); then, the fully connected layer FC2(·) is used for dimensionality increase, and finally, the channel weight is obtained by batch normalization BN(·):

[0187] Wchannel =BN(FC2(ReLU(BN(FC1(GAP(F fusion ))))) (27)

[0188] The spatial attention performs multiple convolution operations Conv2D(·) on the input features to generate a spatial weight map:

[0189] W spatial =Conv2D(F fusion ) (28)

[0190] Finally, update the output feature of the bottleneck attention mechanism, that is, the input feature X 1 and X 2 The latent space representation of is:

[0191]

[0192] Among them, ⊙ represents element-by-element multiplication, Indicates broadcast extension;

[0193] like Figure 7 As shown, the present invention also provides a single base station error suppression and positioning system based on unsupervised learning, including a signal receiving module, a data processing and feature extraction module, an unsupervised error suppression module and a positioning solution module;

[0194] The signal receiving module collects the original signal through the array antenna, and the intervals between the antenna arrays are the same, and the intervals are d aten Less than half of the ultra-wideband radio wave wavelength λ, i.e. d aten <λ / 2;

[0195] The data processing and feature extraction module receives the collected original signal, completes the signal preprocessing and feature construction, and expands the low-dimensional observation signal into multi-dimensional feature data;

[0196] After receiving the multi-dimensional feature data after feature extraction, the unsupervised error suppression module uses a distribution-based modeling method to perform feature learning and error mode on the observed signal; by modeling the feature distribution of the signal, the error mode caused by non-line-of-sight or environmental noise is captured, an error correction value is generated, and the error is adaptively suppressed;

[0197] The positioning solution module combines the phase difference value after error correction to estimate the position through an ultra-wideband single base station positioning algorithm;

[0198] With the aid of the teachings presented in the foregoing description and the associated drawings, many modifications and other embodiments of the present invention will occur to those skilled in the art to which the present invention pertains. Therefore, it is to be understood that the present invention is not limited to the specific embodiments disclosed, and modifications and other embodiments are considered to be included within the scope of the appended claims. Although specific terms are used herein, they are used in a generic and descriptive sense only and not for limitation.

Claims

1. A single base station error suppression and positioning method based on unsupervised learning, characterized in that: The following steps are involved: S1: UWB single base station data reading, first receive the original features of the PDoA signal, including phase difference, amplitude information, signal strength and signal-to-noise ratio, and collect environmental context features, that is, two-way channel impulse response data r(t): Where η c and α c is the fading coefficient and time delay of the cth path, C is the total number of signal propagation paths, and n(t) is the variance Additive Gaussian white noise; S2: Data preprocessing, standardizing all features, unifying the input range, and converting the two-channel channel impulse response data into time-frequency diagrams S1(t,f) and S2(t,f) through short-time Fourier transform; Then, calculate the cross power spectrum between the two: in, is the conjugate of S2; Finally, the PDoA original signal feature vector X1 and the two-dimensional matrix X2 of the cross power spectrum are obtained as input X; S3: Construct the unsupervised deep learning network encoder and decoder. First, use the deep learning network as the encoder Encoder(·) to extract the latent space representation Z of the input feature X. latent : Z latent =Encoder(X) (3) Get the latent space representation Z of the input feature latent After that, we construct a self-supervised target to learn the internal pattern of signal features. We first use the decoder Decoder(·) to reconstruct the input features: Then, through feature contrast learning, the ability to distinguish signal features is strengthened; S4: construct an unsupervised deep learning network error regression module, and use latent space features to directly predict the PDoA error value. The error regression module includes two fully connected layers and a regression layer. A batch normalization layer, a linear rectification activation function and a dropout layer are added between the two fully connected layers. The number of hidden layers in the first fully connected layer is much larger than the number of hidden layers in the second fully connected layer. The linear rectification function is after the batch normalization layer, and the dropout layer is after the linear rectification function. The latent space after feature extraction is represented by Z latent First input the first fully connected layer and get the output: F1=W1F BAM +b1 (5) Where W1 is the weight of the first fully connected layer, b1 is the bias of the first fully connected layer; Next, input the batch normalization layer BN(·) for batch normalization: F BN =BN(F1) (6) Apply the linear rectification activation function ReLU(·) to the batch normalized features again: F ReLU =ReLU(F BN ) (7) Random neuron masking is performed through the dropout layer Dropout(·): F Dropout =Dropout(F ReLU ) (8) F Dropout Input the second fully connected layer, prediction error: Δ PDOA =W2F Dropout +b2 (9) Where W2 is the weight of the second fully connected layer, and b2 is the bias of the second fully connected layer; S5: Set the loss function for unsupervised learning, first reconstruction loss, so that the decoder accurately reconstructs the input signal: Secondly, set the contrast loss to strengthen the consistency between similar signal features: Among them, Sim is the feature similarity measure; Set the regression loss again, design a pseudo-label generation mechanism based on distribution matching, and use pseudo-labels to guide error regression: in, is the pseudo-label error generated from the distribution estimate, is the error prediction value output by the regression network; Then set the regularization error to constrain the sparsity and continuity of the latent space: L reg_latent =||Z latent ||1 (13) The total loss function is obtained as: L=λ1L rec +λ2L contrast +λ3L reg +λ4L reg_latent (14) Among them, λ1, λ2, λ3 and λ4 are hyperparameters for adjusting the weights of each loss; S6: Non-line-of-sight error value predicted by unsupervised learning Directly used to correct the raw PDoA measurement data: First, use the predicted error value to correct the phase difference: Calculate the final position based on the corrected phase difference and base station coordinates: P final =f(P base ,φ corrected ,d ac ) (16) Among them, P base is the base station location, f is the positioning solution function, d ac is the flight time distance between the transmitting antenna A and the receiving antenna on the tag.

2. The single base station error suppression and positioning method based on unsupervised learning as claimed in claim 1, characterized in that: The encoder can be divided into three steps: feature extraction, feature concatenation and key feature extraction; The feature extraction can be further divided into two independent branches: PDoA feature extraction and environmental feature extraction; The PDoA feature extraction uses a multi-layer fully connected network FCN (·) to extract the nonlinear feature representation of the original PDoA signal: F PDOA =FCN(X1) (17) The environmental feature extraction uses a deep residual network ResNet(·) to extract the spatial features of the time-frequency graph, and uses a bidirectional long short-term memory network BiLSTM(·) to capture the temporal dynamic characteristics of the signal: F env =BiLSTM(ResNet(X2)) (18) After completing PDoA feature extraction and environmental feature extraction, the features of the two branches are spliced: F fusion =Concat(F PDOA ,F env ) (19) The key feature extraction is achieved through the bottleneck attention mechanism, which focuses on the feature map F through the spatial dimension and the channel dimension respectively. fusion The key area in the output is the extracted key feature F BAM , that is, the latent space representation Z of the input features X1 and X2 latent .

3. The single base station error suppression and positioning method based on unsupervised learning according to claim 1, characterized in that: The positioning solution function f first converts the phase difference φ after the relaxation corrected Converted into distance difference Δd: Δd=φ corrected ×c / 2 (20) Where c is the propagation speed of radio in the current environment; Get the final position x of the label final and final coordinate:

4. The single base station error suppression and positioning method based on unsupervised learning according to claim 1, characterized in that: The pseudo label generation mechanism is designed based on distribution matching, and pseudo labels are generated by modeling the probability distribution of observed data. The specific implementation steps are as follows: S51: First input the unlabeled data X f ={x1,x2,...,x n }, each data point x i It is an observed feature. The true label Y corresponding to the data is not available, so a pseudo label needs to be generated. S52: Use the Gaussian mixture model to fit the distribution of unlabeled data to obtain the probability distribution of the data. The Gaussian mixture model assumes that the data obeys a weighted combination of K Gaussian distributions: Where K is the number of mixture components, π k is the mixing coefficient of the kth Gaussian component, satisfying N(x|μ k ,δ k ) is the kth Gaussian component with mean μ k , with covariance δ k ; S53: Training the Gaussian mixture model using the maximum expectation algorithm, including the following steps: Step E: Calculate the posterior probability that each data point belongs to each mixture component: Among them, γ i,k Represents data point x i The probability of belonging to the kth component; Step M: Update Gaussian mixture model parameters π k , μ k ,δ k ; S54: For each data point x i , according to the posterior probability γ i,k , select the component with the largest probability as the pseudo label: in, is the data point x i Pseudo labels of S55: Distribution matching, the generated pseudo labels are used for downstream tasks, and the distribution regularization of the pseudo labels is introduced to ensure that the model prediction is consistent with the distribution learned by the Gaussian mixture model: where KL(·) is the KL divergence function, is the label distribution predicted by the model, and p(x) is the distribution learned by the Gaussian mixture model.

5. The single base station error suppression and positioning method based on unsupervised learning according to claim 2, characterized in that: The bottleneck attention includes two branches: channel attention and spatial attention; The channel attention is encoded into a one-dimensional feature vector using global average pooling GAP(·); then, the one-dimensional feature vector is reduced in dimension by the fully connected layer FC1(·), and nonlinearly processed by the rectified linear activation function ReLU(·); then, the fully connected layer FC2(·) is used for dimensionality increase, and finally, the channel weight is obtained by batch normalization BN(·): W channel =BN(FC2(ReLU(BN(FC1(GAP(F fusion )))))) (27) The spatial attention performs multiple convolution operations Conv2D(·) on the input features to generate a spatial weight map: W spatial =Conv2D(F fusion ) (28) Finally, the output features of the bottleneck attention mechanism are updated, that is, the latent space representation of the input features X1 and X2: Among them, ⊙ represents element-by-element multiplication, Indicates a broadcast extension.

6. A single base station error suppression and positioning system based on unsupervised learning, using the single base station error suppression and positioning method according to any one of claims 1 to 5, characterized in that: It includes signal receiving module, data processing and feature extraction module, unsupervised error suppression module and positioning solution module: The signal receiving module collects the original signal through the array antenna, and the intervals between the antenna arrays are the same, and the intervals are d aten Less than half of the ultra-wideband radio wave wavelength λ, i.e. d aten <λ / 2; The data processing and feature extraction module receives the collected original signal, completes the signal preprocessing and feature construction, and expands the low-dimensional observation signal into multi-dimensional feature data; After receiving the multi-dimensional feature data after feature extraction, the unsupervised error suppression module uses a distribution-based modeling method to perform feature learning and error mode on the observed signal; by modeling the feature distribution of the signal, the error mode caused by non-line-of-sight or environmental noise is captured, an error correction value is generated, and the error is adaptively suppressed; The positioning solution module combines the phase difference value after error correction and estimates the position through an ultra-wideband single base station positioning algorithm.

Images