Semi-supervised learning positioning method based on channel mapping

Abstract

The invention discloses a semi-supervised learning positioning method based on channel mapping, belongs to the field of signal processing, and is applied to the problem of wireless positioning of user equipment in a communication and sensing integrated network. According to the method, mass unlabeled channel data which is easy to obtain is fully utilized, and the implicit relationship between the channel and the position is learned from the unlabeled data and the labeled data together. The semi-supervised learning scheme provided by the invention can be divided into two stages: firstly, pre-training an unsupervised dual-branch channel mapping twin network by using an unlabeled data set, and then performing supervised training fine tuning on the dual-branch channel mapping twin network and an overall positioning network of a positioning head module by using a labeled data set. According to the method, only a small part of labeled data is used, and robustness and positioning performance superior to a supervised learning algorithm and other baseline algorithms are obtained.

Description

Technical Field

[0001] The present invention relates to the field of signal processing technology, and in particular to a semi-supervised learning positioning method based on channel mapping. Background Art

[0002] Spatial location awareness has become an indispensable foundational service in many important areas, including smart cities, autonomous driving, and emergency rescue. As mobile communication networks evolve toward the integration of communication and awareness functions to achieve higher spectrum, hardware, and information processing efficiency, it is becoming possible to utilize communication networks to provide ubiquitous positioning services, potentially overcoming the limitations of traditional dedicated positioning networks in terms of coverage and application scenarios. However, complex channel environments, such as dense urban areas and cluttered indoor environments, with their multipath and non-line-of-sight propagation characteristics, pose significant challenges to traditional communication network positioning algorithms. Traditional methods suffer from severe performance degradation or require high sampling costs. In artificial intelligence (AI) positioning applications, supervised learning methods rely on large amounts of labeled data to construct high-precision positioning models. However, in real-world scenarios, the acquisition and location annotation of channel data are costly, and channel fluctuations in dynamic environments lead to significant data imbalance. Therefore, developing a semi-supervised learning positioning algorithm that utilizes limited labeled samples and massive amounts of unlabeled channel data to reduce the demand for labeled data and improve positioning accuracy and robustness is of great practical significance. Summary of the Invention

[0003] The present invention provides a semi-supervised learning positioning method based on channel mapping, which obtains the mapping relationship between channel state information and geometric position in a semi-supervised learning manner, realizes efficient and accurate position prediction of user equipment, and greatly reduces the demand for label data in the deployment process of AI positioning system.

[0004] An embodiment of the present invention provides a semi-supervised learning positioning method based on channel mapping, comprising the following steps:

[0005] Step 1: Multi-base station channel data acquisition and preprocessing: Channel state information is obtained through the OFDM reference signal transmitted by the user equipment. The base station performs channel estimation and inverse Fourier fast transform on the received baseband signal to obtain a time domain channel impulse response vector. The time domain channel impulse response vectors obtained by all base stations are matrixed and modulo-valued to generate a time domain channel impulse response vector modulus matrix. The time domain channel impulse response vector modulus matrix is ​​used to construct a semi-supervised training dataset containing labeled data and unlabeled data.

[0006] Step 2, unsupervised pre-training of a two-branch channel mapping twin network: Calculate the relative distance matrix between samples of unlabeled data in the semi-supervised training dataset, construct a two-branch channel mapping twin network, use the time domain channel impulse response vector modulus matrix of a pair of samples in the unlabeled data as the network input, use the network output and the relative distance matrix between samples to calculate the loss function, and backpropagate to update the network parameters;

[0007] Step 3: Supervised fine-tuning of the overall localization network: The pre-trained two-branch channel mapping twin network is connected in series with the localization head module to construct an overall localization network, and the labeled data in the semi-supervised training dataset is used to perform supervised training fine-tuning on the overall localization network;

[0008] Step 4, online positioning inference: By deploying and loading the trained overall positioning network, inputting the time domain channel impulse response vector modulus matrix generated by the user to be located, and outputting the coordinate prediction value of the user device to be located.

[0009] Optionally, in one embodiment of the present invention, step 1 specifically includes:

[0010] Assume that the user equipment transmits predefined reference signal data S=diag(s1, s2, ...s M ), it reaches base station i through P paths, and the arrival time of each path p is The channel attenuation coefficient of each path p is The received baseband signal in the frequency domain is expressed as:

[0011]

[0012] in, M is the number of subcarriers, Δf is the subcarrier spacing, w∈C M×1 is the frequency domain noise;

[0013] The frequency domain channel vector obtained by the base station through channel estimation of the received baseband signal is:

[0014]

[0015] Where w′ is the frequency domain channel estimation noise;

[0016] Perform inverse Fourier fast transform on the frequency domain channel vector to obtain the time domain channel impulse response vector: Where F is the Fourier transform matrix;

[0017] The time domain channel impulse response vectors obtained by all base stations are matrixed and modulo-valued to obtain the time domain channel impulse response vector modulus matrix Where B is the number of base stations and T is the number of sampling points in the time domain window.

[0018] Optionally, in one embodiment of the present invention, in the semi-supervised training data set of step 1, the labeled data is the time domain channel impulse response vector modulus matrix and its real coordinate label randomly collected by a user device equipped with a high-precision positioning device in the area to be positioned; the unlabeled data is the time domain channel impulse response vector modulus matrix randomly collected by a user device not equipped with a positioning device in the area to be positioned.

[0019] Optionally, in one embodiment of the present invention, step 2 specifically includes:

[0020] Calculate the relative distance between two samples in unlabeled data. The data used include the time domain channel impulse response vector modulus matrix and continuous sampling timestamps. A fused distance metric is used for backpropagation in the training phase of the dual-branch channel mapping twin network, which is defined as:

[0021] d(H i ,H j ,t i ,t j )=min{d CSI (H i ,H j ),λd T (t i ,t j )} (3)

[0022] Among them, d(H i ,H j ,t i ,t j ) is the relative distance between samples i and j, d CSI (H i ,H j ) is the channel state information (CSI) distance between samples i and j, λ is the scaling factor, and d T (t i ,t j ) is the relative time distance between samples i and j, H i ,H j are the time domain channel impulse response vector modulus matrices of samples i and j, respectively, i ,t j are the sampling times of samples i and j respectively;

[0023] A distance metric that is a fusion of channel state information distance and time distance. The channel state information distance considers the small-scale continuity of multipath distribution differences in spatial positions, while the timestamp-based distance considers the temporal continuity of the target position.

[0024] dCSI The Manhattan distance is calculated based on the time domain channel impulse response vector modulus matrix of the input sample pair, and the expression is:

[0025]

[0026] Among them, K is the number of time domain points, N is the number of base stations, and H i (n,k) and H j (n, k) are the amplitude values ​​of the time domain channel impulse response vector modulus matrix of sample i corresponding to the nth base station and the kth time domain point, and the amplitude values ​​of the time domain channel impulse response vector modulus matrix of sample j corresponding to the nth base station and the kth time domain point, respectively. T The distance is calculated based on the time difference of the input sample pair, and its expression is:

[0027] d T (t i ,t j )=|t i -t j | (5)

[0028] The above metric calculates the local distance between two close sample points, and then uses the global distance calculation method based on the idea of ​​equal metric mapping to construct an adjacency graph according to the local distance, and takes the sum of the local distances on the shortest path between the two sample points on the adjacency graph as the global distance D ij ;

[0029] After obtaining the relative distances between all samples on the time-domain channel impulse response vector modulus matrix of the unlabeled data, unsupervised pre-training of the dual-branch channel mapping twin network is performed; the dual-branch channel mapping twin network consists of two twin sub-network branches with shared weights. In the prediction phase, only one sub-network branch is activated, inputs the time-domain channel impulse response vector, and outputs pseudo-coordinates in the L-dimensional latent space; in the training phase, both sub-network branches are in an activated state, and the inputs are the time-domain channel impulse response vectors H at two different positions randomly sampled from the unlabeled data. i ,H j , output two L-dimensional pseudo coordinates and calculate the Euclidean distance between the pseudo coordinates, called the predicted distance. The predicted distance and the relative distance of the samples output by the distance calculation module are used to calculate the training loss, perform backpropagation of the dual-branch channel mapping twin network, and update the network parameters;

[0030] The calculated sample relative distance is used as supervision for the training phase of the two-branch channel mapping twin network. The unsupervised pre-training loss function of the two-branch channel mapping twin network is:

[0031]

[0032] Among them, NB represents the batch size, Ψ Θ (·) represents a sub-network output of the channel mapping twin network, D ij represents the relative distance between samples i and j, and ||·||2 is the Euclidean distance.

[0033] Optionally, in one embodiment of the present invention, step 3 specifically includes:

[0034] The output of the dual-branch channel mapping twin network is not real physical coordinates, but pseudo-coordinates in the latent space. These coordinates need to be transformed into a two-dimensional physical coordinate space. A neural network positioning head is designed to learn the mapping from pseudo-coordinates to real coordinates. The structure consists of a fully connected layer activated by ReLU, and the output layer outputs two-dimensional coordinates. The positioning head and the pre-trained dual-branch channel mapping twin network together form the overall positioning network. The overall positioning network is fine-tuned through supervised training using labeled data from the semi-supervised training dataset.

[0035] In the supervised training fine-tuning stage, the loss function of the overall positioning network CCSN-LH is defined as:

[0036]

[0037] Among them, N l is the number of training samples with labeled data, y n represents the true location label, Λ Ω (·) represents the overall positioning network output of the dual-branch channel mapping twin network and the positioning head module, is the square of the Euclidean distance.

[0038] The channel mapping-based semi-supervised learning positioning method of an embodiment of the present invention is applied to the wireless positioning problem of user equipment in a synaesthesia network. This method makes full use of easily accessible unlabeled channel data, and jointly learns the implicit relationship between channel and position from unlabeled and labeled data. The semi-supervised learning training scheme of the model includes two stages. First, an unsupervised two-branch channel mapping twin network is pre-trained using an unlabeled dataset, and then a labeled dataset is used to supervise the training and fine-tune the overall positioning network of the two-branch channel mapping twin network and the positioning head module. The method of the present invention uses only a small amount of labeled data, and obtains robustness and positioning performance that is superior to supervised learning algorithms and other baseline algorithms.

[0039] Additional aspects and advantages of the present invention will be set forth in part in the description which follows and, in part, will be obvious from the description which follows, or may be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0041] Figure 1 A flowchart of a semi-supervised learning positioning method based on channel mapping provided according to an embodiment of the present invention;

[0042] Figure 2 This is the overall architecture of the semi-supervised learning positioning solution of an embodiment of the present invention;

[0043] Figure 3 A schematic diagram of setting up an experimental scenario for an embodiment of the present invention;

[0044] Figure 4 The change of CT with the ratio of nearest neighbor nodes for the algorithm and unsupervised channel mapping of the embodiment of the present invention;

[0045] Figure 5 The change of TW with the ratio of nearest neighbor nodes for the algorithm and unsupervised channel mapping in the embodiment of the present invention;

[0046] Figure 6 CDF curves of positioning errors of the algorithm in the embodiment of the present invention and other baseline algorithms in different propagation scenarios;

[0047] Figure 7 The positioning performance indicators of the algorithm and supervised learning method of the embodiment of the present invention with different labeled data set sizes;

[0048] Figure 8 This is the CDF curve of the position prediction error of the algorithm of the embodiment of the present invention and the supervised learning method with different labeled data set sizes. DETAILED DESCRIPTION

[0049] The following describes embodiments of the present invention in detail, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and are not to be construed as limiting the present invention.

[0050] like Figure 1 and Figure 2 As shown, the semi-supervised learning positioning method based on channel mapping includes the following steps:

[0051] Step 1: Multi-base station channel data acquisition and preprocessing: Channel state information is obtained through the OFDM reference signal transmitted by the user equipment (UE). The base station performs channel estimation and inverse Fourier fast transform on the received baseband signal to obtain the time domain channel impulse response vector (CIR). The time domain CIR vectors obtained by all base stations are matrixed and modulo-valued to generate a time domain CIR vector modulus matrix. The time domain CIR vector modulus matrix is ​​used to construct a semi-supervised training dataset containing labeled data and unlabeled data.

[0052] Specifically, in the semi-supervised training data set of step 1, the labeled data is the time domain channel impulse response vector modulus matrix and its real coordinate labels randomly collected by user equipment equipped with a high-precision positioning device in the area to be positioned; the unlabeled data is the time domain channel impulse response vector modulus matrix randomly collected by a user equipment not equipped with a positioning device in the area to be positioned.

[0053] Optionally, in one embodiment of the present invention, step 1 specifically includes:

[0054] Assume that the user equipment transmits predefined reference signal data S=diag(s1, s2, ...s M ), it reaches base station i through P paths, and the arrival time of each path p is The channel attenuation coefficient of each path p is The received baseband signal in the frequency domain is expressed as:

[0055]

[0056] in, M is the number of subcarriers, Δf is the subcarrier spacing, w∈C M×1 is the frequency domain noise;

[0057] The frequency domain channel vector obtained by the base station through channel estimation of the received baseband signal is:

[0058]

[0059] Where w′ is the frequency domain channel estimation noise; the subscript f here represents It is frequency domain data, not specific variables. The t subscript in the following formula is similar, indicating It is time domain data.

[0060] Furthermore, the frequency domain channel vector is subjected to an inverse Fourier fast transform to obtain the time domain channel impulse response vector: Where F is the Fourier transform matrix; the time domain channel impulse response vectors obtained by all base stations are matrixed and modulo-valued to obtain the time domain channel impulse response vector modulus matrix Where B is the number of base stations and T is the number of sampling points in the time domain window.

[0061] Step 2, unsupervised pre-training of a two-branch channel charting Siamese network (CCSN): Calculate the relative distance matrix between samples of unlabeled data in the semi-supervised training dataset, construct a two-branch channel charting Siamese network, use the time domain channel impulse response vector modulus matrix of a pair of samples in the unlabeled data as the network input, use the network output and the relative distance matrix between samples to calculate the loss function, and backpropagate to update the network parameters.

[0062] Optionally, in one embodiment of the present invention, step 2 specifically includes:

[0063] Calculate the relative distance between two samples in unlabeled data. The data used include the time domain channel impulse response vector modulus matrix and continuous sampling timestamps. A fused distance metric is used for backpropagation in the training phase of the dual-branch channel mapping twin network, which is defined as:

[0064] d(H i ,H j ,t i ,t j )=min{d CSI (H i ,H j ),λd T (t i ,t j )} (3)

[0065] Among them, d(H i ,H j ,t i ,t j ) is the relative distance between samples i and j, d CSI (H i ,H j ) is the channel state information distance between samples i and j, λ is the scaling factor, d T (t i ,t j ) is the relative time distance between samples i and j, H i ,H j are the time domain channel impulse response vector modulus matrices of samples i and j, respectively, i ,t j are the sampling times of samples i and j respectively;

[0066] A distance metric that is a fusion of channel state information distance and time distance. The channel state information distance considers the small-scale continuity of multipath distribution differences in spatial positions, while the timestamp-based distance considers the temporal continuity of the target position.

[0067] d CSI The Manhattan distance is calculated based on the time domain channel impulse response vector modulus matrix of the input sample pair, and the expression is:

[0068]

[0069] Among them, K is the number of time domain points, N is the number of base stations, and H i (n,k) and H j (n, k) are the amplitude values ​​of the time domain channel impulse response vector modulus matrix of sample i corresponding to the nth base station and the kth time domain point, and the amplitude values ​​of the time domain channel impulse response vector modulus matrix of sample j corresponding to the nth base station and the kth time domain point, respectively. T The distance is calculated based on the time difference of the input sample pair, and its expression is:

[0070] d T (t i ,t j )=|t i -t j | (5)

[0071] The above metric calculates the local distance between two close sample points, and then uses the global distance calculation method based on the idea of ​​equal metric mapping to construct an adjacency graph according to the local distance, and takes the sum of the local distances on the shortest path between the two sample points on the adjacency graph as the global distance D ij ;

[0072] After obtaining the relative distances between all samples on the time-domain channel impulse response vector modulus matrix of the unlabeled data, unsupervised pre-training of the dual-branch channel mapping twin network is performed; the dual-branch channel mapping twin network consists of two twin sub-network branches with shared weights. In the prediction phase, only one sub-network branch is activated, inputs the time-domain channel impulse response vector, and outputs pseudo-coordinates in the L-dimensional latent space; in the training phase, both sub-network branches are in an activated state, and the inputs are the time-domain channel impulse response vectors H at two different positions randomly sampled from the unlabeled data. i ,H j , output two L-dimensional pseudo coordinates and calculate the Euclidean distance between the pseudo coordinates, called the predicted distance. The predicted distance and the relative distance of the samples output by the distance calculation module are used to calculate the training loss, perform backpropagation of the dual-branch channel mapping twin network, and update the network parameters;

[0073] The calculated sample relative distance is used as supervision for the training phase of the two-branch channel mapping twin network. The unsupervised pre-training loss function of the two-branch channel mapping twin network is:

[0074]

[0075] Among them, N B represents the batch size, Ψ Θ (·) represents a sub-network output of the channel mapping twin network, D ij represents the relative distance between samples i and j, and ||·||2 is the Euclidean distance.

[0076] In the embodiment of the present invention, a sample refers to a sampling result at a certain position, including a time-domain channel impulse response vector modulus matrix, and if it is labeled data, also includes the coordinates corresponding to the position.

[0077] Step 3: Supervised fine-tuning of the overall localization network: The pre-trained two-branch channel mapping Siamese network is connected in series with the localization head (LH) module to construct an overall localization network, and the labeled data in the semi-supervised training dataset is used to perform supervised training fine-tuning on the overall localization network.

[0078] Optionally, in one embodiment of the present invention, step 3 specifically includes:

[0079] The output of CCSN is not real physical coordinates, but pseudo-coordinates in latent space, which require transformation to a two-dimensional physical coordinate space. This paper designs a neural network localization head to learn the mapping from pseudo-coordinates to real coordinates. Its structure consists of fully connected layers activated by Reluctant Units (ReLUs), and its output layer outputs two-dimensional coordinates. This localization head, together with the pre-trained CCSN, forms the overall localization network. This network is fine-tuned through supervised training using labeled data from the semi-supervised training dataset to achieve real physical coordinate prediction.

[0080] In the supervised training fine-tuning stage, the loss function of the overall positioning network CCSN-LH is defined as:

[0081]

[0082] Among them, N l is the number of training samples with labeled data, y n represents the true location label, Λ Ω (·) represents the overall positioning network output of the dual-branch channel mapping twin network and the positioning head module, is the square of the Euclidean distance.

[0083] Step 4, online positioning inference: By deploying and loading the trained overall positioning network, inputting the time domain channel impulse response vector modulus matrix generated by the user to be located, and outputting the coordinate prediction value of the user device to be located.

[0084] The specific training algorithm steps of the semi-supervised positioning model of the present invention are as follows:

[0085]

[0086]

[0087] The CCSN-LH network structure of the embodiment of the present invention is shown in Table 1:

[0088] Table 1 CCSN-LH network structure

[0089] Layer Category Convolution kernel Activation Function Output scale Conv1 (3,3,8) ReLU (6,128,8) Conv2 (5,5,8) ReLU (6,128,8) Conv3 (7,7,16) ReLU (6,128,16) Conv4 (9,9,16) ReLU (6,128,16) AvgPool - - (1,6,128) FC1 - ReLU 500 FC2 - ReLU 2 FC3(LH) - ReLU 64 FC4(LH) - - 2

[0090] The method of the present invention is described in detail below through a specific embodiment to verify the performance advantages of the method of the present invention.

[0091] (1) Experimental settings and parameters

[0092] The simulation experiment considers an indoor scenario with multiple base stations, with parameter settings shown in Table 2. A wideband positioning signal in the millimeter-wave band is used to obtain a high-resolution power delay profile, with a carrier frequency of 30 GHz and a bandwidth of 200 MHz. Because the proposed solution does not involve angle information, the base station and UE only need to be equipped with basic omnidirectional antennas. Figure 3 The experimental scenario setup is shown in Figure 1. The UE is located in an 8m×10m rectangular area. Six base stations, each 8m high, are positioned around the positioning area. The channel propagation between each base station and the UE is individually configured as either LoS or NLoS scenarios to simulate possible obstruction conditions in real-world scenarios. Both scenarios account for multipath propagation: the LoS scenario channel has 8 multipath paths, and the NLoS scenario channel has 20 multipath paths. This section configures the channel propagation between each base station and the UE individually, and obtains channel datasets for scenarios with all LoS base stations and varying numbers of NLoS base stations to verify the robustness of the proposed solution in NLoS scenarios.

[0093] Table 2 Experimental parameter configuration

[0094] parameter Value Center frequency 30GHz bandwidth 200MHz Number of subcarriers 1024 Number of base stations 6 Base station height 8m Communication scenarios LoS / NLoS Multipath number 8 / 20 User height 1m Positioning area 8m×10m Dataset size 10000

[0095] For each propagation scenario with different numbers of NLoS base stations, a dataset of 10,000 instances was generated, with 80% used as training, 15% as testing, and 5% as validation. 2,000 instances in the training set were labeled with their locations for supervised training, while the remaining instances were unlabeled.

[0096] (2) Channel mapping performance

[0097] To verify the effectiveness of unsupervised pre-training, the quality of channel mapping is evaluated. Metrics used to evaluate the results of dimensionality reduction methods are used to assess the geometric consistency of the channel mapping. Specifically, the metrics used are continuity (CT) and trustworthiness (TW). The dimensionality reduction effect is evaluated by performing operations on point sets in the original space and the representation space. The original space is the physical space, and the representation space is the pseudo-coordinate space output by channel mapping. Channel mapping maps the data point set in the original space to the data point set in the low-dimensional representation space. Neighboring points in the original space may be far apart in the representation space. In this case, the representation space does not maintain the continuity of the original space point set. Continuity measures whether neighboring points in the original space are close in the representation space. Alternatively, the representation space may introduce neighboring points that are far apart in the original space, leading to erroneous sample point relationships and reduced trustworthiness.

[0098] (1) Continuity (CT): measures whether the low-dimensional space introduces the wrong nearest neighbor and imposes corresponding penalties based on the ranking of the nearest neighbor. The calculation method is as shown in formula (8). are the k nearest neighbors of sample point i in the low-dimensional space, and r(i, j) is the distance ranking of sample point j from i in the original space. CT values ​​range from 0 to 1. A high CT value indicates that most of the neighborhood relationships in the original space are preserved in the low-dimensional space, resulting in better continuity.

[0099] (2) Trustworthiness (TW): This measures whether there are different nearest neighbors in the original space from those modeled in the channel graph. The calculation method is as shown in Equation (9). The value range of TW is also [0, 1]. The closer it is to 1, the higher the trustworthiness.

[0100]

[0101] The performance of channel mapping directly impacts the performance of downstream localization tasks. The channel mapping performance of our semi-supervised learning method is evaluated using two standard metrics, CT and TW. To verify the improved channel mapping performance achieved by the proposed supervised training fine-tuning, this section uses an unsupervised channel mapping Siamese network as a baseline method. This network structure is identical to the CCSN used in the proposed solution, but it can only learn from unlabeled data.

[0102] Considering different channel conditions, the channel mapping performance of the proposed schemes was compared under three scenarios: LoS, mild NLoS, and severe NLoS. LoS occurs when there is no obstruction between all base stations and the UE. Mild NLoS occurs when there are only one or two NLoS base stations, and severe NLoS occurs when there are more than two. Figure 4 Figures (a) and (b) compare the continuity of channel mapping using the proposed semi-supervised method and the unsupervised Siamese network method. The horizontal axis represents the ratio of the number of nearest neighbors k selected when calculating CT to the total number of samples, and the vertical axis represents the value of CT. The solid line represents the proposed semi-supervised method, and the dashed line represents the unsupervised channel mapping Siamese network method.

[0103] Figure 4 (a) and Figure 4 (b) shows the variation of CT with the ratio of nearest neighbor nodes on the training and test data sets, respectively. As the number of nearest neighbors increases, the continuity metric of the unsupervised twin network method decreases, indicating that the larger the local area, the less physical proximity relationships are reflected in the channel graph. The CT value of the results generated by the proposed semi-supervised method remains stable with the number of nearest neighbors and is higher than that of the unsupervised twin network method under all three propagation settings.

[0104] Figure 5 (a) and Figure 5 (b) shows the results of TW on the training set and test set data as the ratio of nearest neighbor nodes changes. Similar to the CT results, the credibility of the unsupervised twin network method decreases as the number of nearest neighbors increases, indicating that when the local area is larger, there are more nodes with incorrect neighbor relationships in the channel graph. The TW value of the generated results of the proposed semi-supervised method remains stable with the number of nearest neighbors, and is higher than that of the unsupervised twin network method under all three propagation settings.

[0105] Experimental results demonstrate that the channel map generated by our semi-supervised approach exhibits superior continuity and credibility compared to the baseline unsupervised channel mapping method. This demonstrates that supervised training and fine-tuning effectively improve the quality of channel mapping, enhance its robustness to NLoS propagation, and enhance its advantages for downstream positioning tasks. Furthermore, the two approaches achieve similar results on both training and test data, demonstrating the good generalization of channel mapping.

[0106] (3) Positioning performance

[0107] This section evaluates the positioning accuracy of the proposed semi-supervised localization scheme and compares its performance with baseline algorithms to verify the effectiveness of the proposed design. The selected baseline algorithms include supervised learning localization, channel mapping combined with affine transformation, and a geometry-based ToA localization algorithm.

[0108] Table 3 shows the positioning performance indicators of the proposed semi-supervised method and other baseline methods on the same training set data, including MAE, RMSE and CE90, considering three propagation scenarios: full LoS, mild NLoS, and severe NLoS. Figure 6 (a), (b) and (c) show the CDF curves of the positioning error of the proposed semi-supervised method and other baseline methods on the same training set data. Figure 6 The CDF curves of positioning errors in different propagation scenarios are shown. Figure 6 (a), (b) and (c) are LoS, mild NLoS and severe NLoS respectively. As can be seen from the experimental results chart, the 90% errors of the proposed semi-supervised scheme in the three scenarios are 0.29m, 0.36m and 0.64m respectively, which are better than other baseline algorithms. The semi-supervised method of the present invention is better than the performance of the supervised learning method, which shows that in addition to learning from labeled data, the proposed semi-supervised method also learns the relative distance of samples from unlabeled data, verifying the effectiveness of combining unsupervised CCSN. ​​The method of the present invention is better than the method of channel mapping combined with affine transformation ("CC+AT"), which shows that affine transformation cannot fully utilize the information of labeled data, while the method of the present invention can use labeled data to learn nonlinear mapping. In addition, the method of the present invention is significantly better than the ToA method, especially in the severe NLoS scenario, where the positioning accuracy of the ToA method deteriorates significantly.

[0109] Table 3 Algorithm positioning performance indicators (m)

[0110]

[0111] In order to analyze the impact of the size of the labeled dataset on the positioning performance of the scheme, this section changes the number of labeled data included in the training set, trains the models separately, and compares the performance of the proposed semi-supervised learning scheme with the supervised learning scheme.

[0112] For the semi-supervised scheme of the present invention, the overall number of training sets remains unchanged, and the number of labeled items in the training set increases from 1000 to 5000. Figure 7 (a), (b) and (c) are comparisons of test set positioning errors given different labeled dataset sizes. Figure 7 Subfigures (a), (b), and (c) show the comparison of the semi-supervised learning scheme with the supervised learning scheme using the MAE, RMSE, and CE90 metrics, respectively. As the number of labeled samples increases to over 3,000, the performance of both schemes stabilizes, with the error barely decreasing. The proposed semi-supervised scheme achieves lower errors than the supervised learning scheme, demonstrating that the unsupervised channel mapping module effectively utilizes unlabeled data.

[0113] Figure 8The CDF curves of the location prediction errors of the two schemes for different labeled dataset sizes are presented. The blue curve represents the prediction error of different models trained with the proposed semi-supervised scheme on the training set, while the red curve represents the prediction error of the supervised learning method. The CDF curves clearly show that, using the same amount of labeled data, the proposed scheme achieves better localization performance due to the unsupervised CCSN pre-training.

[0114] The semi-supervised learning positioning method based on channel mapping proposed in an embodiment of the present invention is applied to the positioning problem of user equipment in a synaesthesia network. This method makes full use of the easily accessible massive unlabeled channel data, and jointly learns the implicit relationship between channel and position from unlabeled and labeled data. The semi-supervised learning scheme of the present invention can be divided into two stages. First, an unsupervised two-branch channel mapping twin network is pre-trained using an unlabeled dataset, and then a labeled dataset is used to supervise the training and fine-tune the overall positioning network of the two-branch channel mapping twin network and the positioning head module. The proposed method only uses a small amount of labeled data, and the channel mapping performance is better than the unsupervised channel mapping baseline algorithm, and the positioning accuracy and robustness to NLoS scenarios are better than the supervised learning algorithm and other baseline algorithms.

[0115] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or N embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine different embodiments or examples described in this specification and the features of different embodiments or examples without contradiction.

[0116] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of technical features indicated. Thus, a feature specified as "first" or "second" may explicitly or implicitly include at least one such feature. In the description of the present invention, "N" means at least two, such as two, three, etc., unless otherwise specifically defined.

[0117] Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, segment or portion of code comprising one or N executable instructions for implementing a custom logical function or step of a process, and the scope of the preferred embodiments of the present invention includes alternative implementations in which functions may be performed out of the order shown or discussed, including performing functions in a substantially simultaneous manner or in the reverse order depending on the functions involved, which should be understood by those skilled in the art to which the embodiments of the present invention pertain.

Claims

1. A semi-supervised learning positioning method based on channel mapping, characterized in that: The following steps are involved: Step 1: Multi-base station channel data acquisition and preprocessing: Channel state information is obtained through the OFDM reference signal transmitted by the user equipment. The base station performs channel estimation and inverse Fourier fast transform on the received baseband signal to obtain a time domain channel impulse response vector. The time domain channel impulse response vectors obtained by all base stations are matrixed and modulo-valued to generate a time domain channel impulse response vector modulus matrix. The time domain channel impulse response vector modulus matrix is ​​used to construct a semi-supervised training dataset containing labeled data and unlabeled data. Step 2, unsupervised pre-training of a two-branch channel mapping twin network: Calculate the relative distance matrix between samples of unlabeled data in the semi-supervised training dataset, construct a two-branch channel mapping twin network, use the time domain channel impulse response vector modulus matrix of a pair of samples in the unlabeled data as the network input, use the network output and the relative distance matrix between samples to calculate the loss function, and backpropagate to update the network parameters; Step 3: Supervised fine-tuning of the overall localization network: The pre-trained two-branch channel mapping twin network is connected in series with the localization head module to construct an overall localization network, and the labeled data in the semi-supervised training dataset is used to perform supervised training fine-tuning on the overall localization network; Step 4, online positioning inference: By deploying and loading the trained overall positioning network, inputting the time domain channel impulse response vector modulus matrix generated by the user to be located, and outputting the coordinate prediction value of the user device to be located.

2. The method according to claim 1, characterized in that Step 1 specifically includes: Assume that the user equipment transmits predefined reference signal data S=diag(s1, s2, ...s M ), it reaches base station i through P paths, and the arrival time of each path p is The channel attenuation coefficient of each path p is The received baseband signal in the frequency domain is expressed as: in, M is the number of subcarriers, Δf is the subcarrier spacing, w∈C M×1 is the frequency domain noise; The frequency domain channel vector obtained by the base station through channel estimation of the received baseband signal is: Where w′ is the frequency domain channel estimation noise; Perform inverse Fourier fast transform on the frequency domain channel vector to obtain the time domain channel impulse response vector: Where F is the Fourier transform matrix; The time domain channel impulse response vectors obtained by all base stations are matrixed and modulo-valued to obtain the time domain channel impulse response vector modulus matrix Where B is the number of base stations and T is the number of sampling points in the time domain window.

3. The method according to claim 1 or 2, characterized in that In the semi-supervised training data set of step 1, the labeled data is the time domain channel impulse response vector modulus matrix and its real coordinate labels randomly collected by user equipment equipped with a high-precision positioning device in the area to be positioned; the unlabeled data is the time domain channel impulse response vector modulus matrix randomly collected by a user equipment not equipped with a positioning device in the area to be positioned.

4. The method according to claim 2, characterized in that Step 2 specifically includes: Calculate the relative distance between two samples in unlabeled data. The data used include the time domain channel impulse response vector modulus matrix and continuous sampling timestamps. A fused distance metric is used for backpropagation in the training phase of the dual-branch channel mapping twin network, which is defined as: d(H i ,H j ,t i ,t j )=min{d CSI (H i ,H j ),λd T (t i ,t j )} (3) Among them, d(H i ,H j ,t i ,t j ) is the relative distance between samples i and j, d CSI (H i ,H j ) is the channel state information distance between samples i and j, λ is the scaling factor, d T (t i ,t j ) is the relative time distance between samples i and j, H i ,H j are the time domain channel impulse response vector modulus matrices of samples i and j, respectively, i ,t j are the sampling times of samples i and j respectively; A distance metric that is a fusion of channel state information distance and time distance. The channel state information distance considers the small-scale continuity of multipath distribution differences in spatial positions, while the timestamp-based distance considers the temporal continuity of the target position. d CSI The Manhattan distance is calculated based on the time domain channel impulse response vector modulus matrix of the input sample pair, and the expression is: Among them, K is the number of time domain points, N is the number of base stations, and H i (n,k) and H j (n, k) are the amplitude values ​​of the time domain channel impulse response vector modulus matrix of sample i corresponding to the nth base station and the kth time domain point, and the amplitude values ​​of the time domain channel impulse response vector modulus matrix of sample j corresponding to the nth base station and the kth time domain point, respectively. T The distance is calculated based on the time difference of the input sample pair, and its expression is: d T (t i ,t j )=|t i -t j | (5) The above metric calculates the local distance between two close sample points, and then uses the global distance calculation method based on the idea of ​​equal metric mapping to construct an adjacency graph according to the local distance, and takes the sum of the local distances on the shortest path between the two sample points on the adjacency graph as the global distance D ij ; After obtaining the relative distances between all samples on the time-domain channel impulse response vector modulus matrix of the unlabeled data, unsupervised pre-training of the dual-branch channel mapping twin network is performed; the dual-branch channel mapping twin network consists of two twin sub-network branches with shared weights. In the prediction phase, only one sub-network branch is activated, inputs the time-domain channel impulse response vector, and outputs pseudo-coordinates in the L-dimensional latent space; in the training phase, both sub-network branches are in an activated state, and the inputs are the time-domain channel impulse response vectors H at two different positions randomly sampled from the unlabeled data. i ,H j , output two L-dimensional pseudo coordinates and calculate the Euclidean distance between the pseudo coordinates, called the predicted distance. The predicted distance and the relative distance of the samples output by the distance calculation module are used to calculate the training loss, perform backpropagation of the dual-branch channel mapping twin network, and update the network parameters; The calculated sample relative distance is used as supervision for the training phase of the two-branch channel mapping twin network. The unsupervised pre-training loss function of the two-branch channel mapping twin network is: Among them, N B represents the batch size, Ψ Θ (·) represents a sub-network output of the channel mapping twin network, D ij represents the relative distance between samples i and j, and ||·||2 is the Euclidean distance.

5. The method according to claim 4, characterized in that Step 3 specifically includes: The output of the dual-branch channel mapping twin network is not real physical coordinates, but pseudo-coordinates in the latent space. These coordinates need to be transformed into a two-dimensional physical coordinate space. A neural network positioning head is designed to learn the mapping from pseudo-coordinates to real coordinates. The structure consists of a fully connected layer activated by ReLU, and the output layer outputs two-dimensional coordinates. The positioning head and the pre-trained dual-branch channel mapping twin network together form the overall positioning network. The overall positioning network is fine-tuned through supervised training using labeled data from the semi-supervised training dataset. In the supervised training fine-tuning stage, the loss function of the overall positioning network CCSN-LH is defined as: Among them, N l is the number of training samples with labeled data, y n represents the true location label, Λ Ω (·) represents the overall positioning network output of the dual-branch channel mapping twin network and the positioning head module, is the square of the Euclidean distance.

Images