CN117062002A - 5G NR indoor positioning method and system based on lightweight TRANSFORMER - Google Patents

Abstract

A5G NR indoor positioning method and system based on lightweight TRANSFORMER relates to the technical field of computers, in particular to the technical field of machine learning in computers. The method solves the problems of low positioning precision, low efficiency and difficult deployment caused by the limited subcarrier sensing range of the existing indoor fingerprint positioning method. The method comprises the following steps: processing the CSI input data by adopting a rectangular image block operation module with independent channels to obtain an output matrix; converting the three-dimensional matrix of the output matrix into a two-dimensional matrix compatible with an attention mechanism by adopting a Reshape () function; inputting the two-dimensional matrix compatible with the attention mechanism into a multichannel attention mechanism module based on ReLU to obtain final output data; and carrying out position location on the final output data by adopting a position mapping module to obtain a mapping position. The invention is suitable for a 5G NR indoor positioning method based on lightweight TRANSFORMER in a computer.

Description

A 5G NR indoor positioning method and system based on lightweight TRANSFORMER

Technical field

The present invention relates to the field of computer technology, and in particular to the field of machine learning technology in computers.

Background technique

Indoor fingerprint positioning based on Channel State Information (CSI) has been widely used in indoor navigation, personnel tracking and other fields. In addition, the rapid development of 5G technology has brought wider channel bandwidth, richer spatiotemporal channel description, and larger data volume. Therefore, 5G CSI can provide more channel information, which brings new characteristics and challenges to CSI-based fingerprint positioning.

At present, indoor fingerprint positioning methods used in engineering practice can be mainly divided into the following two categories:

(1) Methods based on machine learning, such as the patent document published on March 27, 2020: CN110933604A, which discloses the KNN indoor positioning method based on the timing characteristics of location fingerprints. This method calculates the distance between the current location and the historical location fingerprint. Similarity, get the top K similar position fingerprints, then calculate the distance between each position fingerprint in the top K similar position fingerprints and the positioning result at the previous moment, get the estimated distance, and then calculate the distance of the top K similar position fingerprints respectively The offset and weight are weighted and summed to the position coordinates of the first K similar position fingerprints to obtain the position coordinates at the current moment and complete the positioning. However, this method cannot effectively utilize the higher information richness and feature resolution provided by 5G CSI, resulting in low accuracy and efficiency of indoor positioning.

(2) Methods based on deep learning, such as the patent document published on March 15, 2022: CN114189809A, which discloses an indoor positioning method based on convolutional neural networks and high-dimensional 5G observation features. This method builds an offline image fingerprint database, The convolutional neural network position classification model is obtained by training with the fingerprint database. After that, the 5G observation values collected by the target device at the test point are processed and input into the convolutional neural network position classification model. The positioning coordinates of the test point are obtained through the probability weighted centroid method to complete the positioning. This method can better handle large-scale, high-dimensional data, and can automatically learn more abstract and advanced feature representations, thereby improving the performance and accuracy of the model. However, it has the problems of being difficult to deploy on the device and the vanishing gradient of the recurrent neural network. and explosion issues.

Contents of the invention

The present invention solves the problems of low positioning accuracy, low efficiency and difficulty in deployment caused by the limited subcarrier sensing range of existing indoor fingerprint positioning methods.

In order to achieve the above objects, the present invention provides the following solutions:

The present invention provides a 5G NR indoor positioning method based on lightweight TRANSFORMER. The method is:

S1. Use the channel-independent rectangular image block operation module to process the CSI input data and obtain the output matrix;

S2. Use the Reshape() function to convert the three-dimensional matrix of the output matrix into a two-dimensional matrix that is compatible with the attention mechanism;

S3. Input the two-dimensional matrix compatible with the attention mechanism into the multi-channel attention mechanism module based on ReLU to obtain the final output data;

S4. Use the position mapping module to position the final output data to obtain the mapping position.

Furthermore, there is another preferred embodiment. The processing flow of the channel-independent rectangular image block operation module in step S1 is as follows:

S11. Divide the CSI input data according to different channels to obtain _Hi data;

S12. Use convolution-based image blocks to perform image block operations on the _Hi data to obtain an output matrix.

Furthermore, there is a preferred embodiment. The formula of the convolution-based image block in the above step S12 is expressed as:

H ^p =Concat(P(H _0,: ),P(H _1,: ),...,P(H _C,: ));

in, is the output obtained, _Hi is the data of the i-th channel, C is the number of input channels, C' is the number of output channels, W is the input width, L is the input length, and K _w is the volume used for the block-based The width of the convolution kernel of the product, K _l is the length of the convolution kernel used for block-based convolution, /> is the output width obtained by dividing the input width by the kernel width, /> is the output length obtained by passing the input through the length of the input kernel.

Furthermore, there is a preferred embodiment, the setting step of the convolution-based image block in the above step S12 is:

The padding element is set to 0, K _w is set to 8, and K _l is set to 13;

Set the stride in the width direction equal to the width of the convolution kernel, and set the stride in the length direction equal to the length of the convolution kernel.

Furthermore, there is another preferred embodiment. The processing flow of the ReLU-based multi-channel attention mechanism module in the above step S3 is:

S31. Convert the format of the two-dimensional matrix that is compatible with the attention mechanism to obtain the converted two-dimensional matrix;

S32. Perform dimension transformation on the converted two-dimensional matrix to obtain the transformed two-dimensional matrix;

S33. Use the formula based on the ReLU attention mechanism to perform query processing on the transformed two-dimensional matrix, and obtain the parameter matrix W _q , the key generation parameter matrix W _k and the value generation parameter matrix W _v ;

S34. Use the score formula of the attention mechanism based on ReLU to calculate the parameter matrix W _q , the key generation parameter matrix W _k and the value generation parameter matrix W _v to obtain score data;

S35. Normalize the score data to obtain the final score data;

S36. Use the ReLU activation function to process the final score data, and generate the parameter matrix W _v from the processed result point multiplication value and accumulate it to obtain the final result X matrix;

S37. Input the final result X matrix into the attention mechanism module, and process it through the normalization layer, attention layer and normalization layer in sequence to obtain the output data X _output ;

S38. Input the output data X _output into the multi-layer perceptron layer to obtain the final output data X _output .

Furthermore, there is a preferred embodiment. The formula of the ReLU-based attention mechanism in the above step S33 is expressed as:

X _output =Att(LN(Non-Trans(X)))+X;

Among them, Non-Trans is the unchanged dimension, LN is layer norm regularization, and Att is the attention mechanism based on ReLU.

Furthermore, there is a preferred embodiment. The score formula of the ReLU attention mechanism in the above step S34 is expressed as:

score=ReLU(W _q X)(ReLU(W _k X) ^· ReLU(W _v X)).

The 5GNR indoor positioning method based on lightweight TRANSFORMER described in the present invention can all be implemented using computer software. Therefore, correspondingly, the present invention also provides a 5GNR indoor positioning system based on lightweight TRANSFORMER. The system is :

A storage device used to process CSI input data using a channel-independent rectangular image block operation module to obtain an output matrix;

A storage device for converting the three-dimensional matrix of the output matrix into a two-dimensional matrix compatible with the attention mechanism;

A storage device used to input the two-dimensional matrix compatible with the attention mechanism into the multi-channel attention mechanism module based on ReLU using the Reshape() function to obtain the final output data;

A storage device for positioning the final output data using a position mapping module to obtain a mapping position.

The present invention also provides a computer-readable storage medium. A computer program is stored on the computer-readable storage medium. When the computer program is run by a processor, the computer program executes any one of the above-mentioned lightweight TRANSFORMER-based 5GNR indoor Positioning method.

The present invention also provides a computer device. The device includes a memory and a processor. A computer program is stored in the memory. When the processor runs the computer program stored in the memory, the processor executes any of the above. The described 5G NR indoor positioning method based on lightweight TRANSFORMER.

The beneficial effects of the present invention are:

1. The present invention provides a 5G NR indoor positioning method based on lightweight TRANSFORMER. In the channel-independent rectangular image block operation module, the CSI data is transformed into an image block that can adapt to attention by using a convolution-based image block operation on the CSI data. force mechanism input to facilitate the subsequent attention calculation process; in the multi-channel attention mechanism module based on ReLU, the modeling dimension is concentrated on the channel dimension of the vector, enabling the interaction between subcarriers from multiple base stations to be analyzed Modeling is carried out and a functionally interactive attention mechanism is proposed to achieve faster and better performance. Existing indoor fingerprint positioning methods have limited subcarrier sensing range, resulting in low positioning accuracy and low efficiency.

2. The present invention provides a 5G NR indoor positioning method based on lightweight TRANSFORMER. By using the Relu-based attention mechanism to replace the hardware-unfriendly Softmax attention mechanism, the present invention can be used even on devices with limited resources. Deployment is easier, thereby solving the difficult deployment problem of existing indoor fingerprint positioning methods.

The present invention is suitable for 5G NR indoor positioning method based on lightweight TRANSFORMER in computers.

Description of the drawings

Figure 1 is a flow chart of a 5G NR indoor positioning method based on lightweight TRANSFORMER described in the first embodiment;

Figure 2 is a processing flow chart of the channel-independent rectangular image block operation module described in the second embodiment;

Figure 3 is a processing flow chart of the ReLU-based multi-channel attention mechanism module described in Embodiment 5;

Figure 4(a) is a comparison chart of the results of different methods under the SNR10 data set described in Embodiment 11;

Figure 4(b) is a comparison chart of the results of different methods under the SNR20 data set described in Embodiment 11;

Figure 4(c) is a comparison chart of the results of different methods under the SNR50 data set described in Embodiment 11;

Figure 5 is a comparison diagram of the error size between the 5G NR indoor positioning method based on lightweight TRANSFORMER and the existing positioning method described in the first embodiment.

Among them, SNR is the signal-to-noise ratio.

Detailed ways

The specific implementation modes of the present invention will be further described in detail below with reference to the accompanying drawings and examples. The following examples will help those skilled in the art to further understand the present invention, but do not limit the present invention in any form. It should be noted that, for those of ordinary skill in the art, several changes and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention.

Embodiment 1. Refer to Figure 1 to illustrate this implementation. This implementation provides a 5G NR indoor positioning method based on lightweight TRANSFORMER. The method is:

S1. Use the channel-independent rectangular image block operation module to process the CSI input data and obtain the output matrix;

S2. Use the Reshape() function to convert the three-dimensional matrix of the output matrix into a two-dimensional matrix that is compatible with the attention mechanism;

S3. Input the two-dimensional matrix compatible with the attention mechanism into the multi-channel attention mechanism module based on ReLU to obtain the final output data;

S4. Use the position mapping module to position the final output data to obtain the mapping position.

In practical application, this implementation mode includes a channel-independent rectangular image block operation module, a ReLU-based multi-channel attention mechanism module and a position mapping module. First, after frequency division multiplexing demodulation and time offset correction of the received signal, the user equipment can obtain the receiving resource grid. The received resource grid is defined as rxGrid, and the known pilot is defined as refGrid, so each antenna can be represented by a k-dimensional vector as H _raw = (rxGrid) _k×1 / (refGrid) _k×1 , where k represents the number of subcarriers, and H _raw represents the channel impulse response.

The subcarriers of a single antenna can be expressed as h _i ∈ R ^{1 × k} , while the overall subcarriers of a 5G base station with multiple input and multiple output can be expressed as H' ∈ R ^{n × k} , where n represents the number of antennas. In a typical indoor positioning task, multiple base stations are involved, so the overall CSI data can be expressed as H∈R ^b×n×k , where b represents the number of base stations.

The overall framework of fingerprint-based indoor positioning consists of two stages, namely offline and online, that is, training and testing. The goal of this framework is to use the training data set C _train = {(csi ₁ , loc ₁ ), (csi ₂ , loc ₂ ),..., (csi _n ,loc _n )} so that the test data C _query = Achieve the best positioning performance on {(csi _n+1 ,loc _n+1 ),(csi _n+2 ,loc _n+2 ),...,(csi _m ,loc _m )}, thereby learning a good general Performance mapping method. The training data and test data are both composed of a pair of data. In a pair of data, csi _i represents CSI data and loc _i represents location data.

In the offline stage, a deep learning-based mapping method is utilized to minimize the loss function, where the predicted position is represented by the output of the mapping method and the actual position is represented as the ground truth. Deep learning-based mapping methods have a small number of parameters and exhibit high localization accuracy on test data.

When applying, the input data is divided into different data to be operated: for a given CSI input data H, H is divided according to different channels of H to obtain _Hi , where the subscript i of _Hi represents the i-th of H aisle.

Perform a channel-independent rectangular image block operation on the data to be operated: perform a channel-independent rectangular image block operation on the preprocessed data to be operated on and a given convolution-based image block operation P. The specific formula is as follows.

H ^p =Concat(P(H _0,: ),P(H _1,: ),...,P(H _C,: ));

in represents the obtained output, the subscript i of H _i represents the i-th channel of H, C and C' represent the number of input and output channels respectively, W and L represent the input width and length respectively, K _w and K _l represent respectively The width and length of the convolution kernel used for block-based convolution, /> and/> represent the output width and length obtained by dividing the input width and length by the kernel width and length, respectively.

Set related parameters: set the padding element to 0, set K _w to 8, set K _l to 13, set the stride in the width direction to be equal to the width of the convolution kernel, and set the stride in the length direction is equal to the length of the convolution kernel.

Perform a reshape operation on the output matrix: Use the Reshape operation on the obtained output H ^p to convert H ^p from a three-dimensional matrix into a two-dimensional matrix that is compatible with the attention mechanism.

Multi-channel attention mechanism operation based on ReLU: Dimension transformation of input X:

Let X=W _i , Dimension transformation is performed on X. The specific description formula is as follows.

For example, define X as a matrix:

According to the above transformation formula, the dimension transformation of X can be obtained:

Through this operation, the input X can be adapted to the attention mechanism.

Calculation of the self-attention layer: first calculate the query generation parameter matrix W _q , key generation parameter matrix W _k , and value generation parameter matrix W _v of X. Then calculate the score based on the three generation parameter matrices of X. The specific calculation formula is as follows Show.

score=ReLU(W _q X)(ReLU(W _k X) ^· ReLU(W _v X));

In order to stabilize the gradient, the calculated score is regularized to obtain the final score. The specific regularization formula is as follows.

Among them, α is the regularization constant, which is set to 1×10 ^-15 .

Use the ReLU activation function for the score to make the score a positive number or 0. Using the ReLU activation function is faster than using the softmax activation function normally.

Then the obtained result is multiplied by the W _v value to obtain the weighted score V of each input vector, and after addition, the final output result matrix X is obtained.

Feature fusion based on multi-layer self-attention layer:

By building a multi-layer self-attention layer, the above process is repeated several times, the normalization operation is performed through the normalization layer, and the output and input after each attention calculation are superimposed. The specific process description formula is as follows.

X _output =Att(LN(Non-Trans(X)))+X;

Position mapping: The output obtained from the attention mechanism module is input to the position mapping module composed of a fully connected layer for position positioning. The position mapping module consists of a fully connected layer. The specific description formula is as follows:

(X,Y,Z)=Linear(LN(X _output ));

Where LN represents the layer norm and Linear represents a simple linear layer.

This implementation provides a 5GNR indoor positioning method based on lightweight TRANSFORMER. In the channel-independent rectangular image block operation module, the CSI data is transformed into an image block that can adapt to the attention mechanism by using a convolution-based image block operation on the CSI data. input to facilitate the subsequent attention calculation process; in the multi-channel attention mechanism module based on ReLU, the modeling dimension is concentrated on the channel dimension of the vector, enabling the interaction between subcarriers from multiple base stations to be constructed. model, and also proposes a functional interactive attention mechanism to achieve faster and better performance. Existing indoor fingerprint positioning methods have limited subcarrier sensing range, resulting in low positioning accuracy, low efficiency, and difficulty in deployment.

Embodiment 2. Refer to Figure 2 to illustrate this embodiment. This embodiment is the processing of the channel-independent rectangular image block operation module in step S1 in the lightweight TRANSFORMER-based 5G NR indoor positioning method described in Embodiment 1. The process is given as an example. The processing flow is:

S11. Divide the CSI input data according to different channels to obtain _Hi data;

S12. Use convolution-based image blocks to perform image block operations on the _Hi data to obtain an output matrix.

In practical application of this embodiment, as shown in Figure 2, for the data _Hi of the i-th channel of different given input H, the formula H ^p =Concat(P(H _0,: ),P(H _{1 ,:} ),...,P(H _C,: )) perform image block operations.

in represents the obtained output, the subscript i of H _i represents the i-th channel of H, C and C' represent the number of input and output channels respectively, W and L represent the input width and length respectively, K _w and K _l represent respectively The width and length of the convolution kernel used for block-based convolution, /> and/> represent the output width and length obtained by dividing the input width and length by the kernel width and length, respectively.

Set the padding element to 0, in addition to setting the stride in the width direction equal to the width of the convolution kernel, and setting the stride in the length direction equal to the length of the convolution kernel. Set K _w to 8 and K _l to 13. Because the special aspect ratio of CSI requires such a design to effectively extract features. Also a smaller value of 8 corresponds to a smaller antenna size, while a larger value of 13 corresponds to a larger subcarrier size.

Use the formula for the obtained output matrix H _p Perform matrix dimension conversion to obtain a two-dimensional matrix output H _p that is compatible with the attention mechanism.

will get the output It is left to the ReLU-based multi-channel attention mechanism module to model the interactions between sub-carriers from multiple base stations.

Embodiment 3. This embodiment is an example of the formula of the convolution-based image block in step S12 of the lightweight TRANSFORMER-based 5GNR indoor positioning method described in Embodiment 2. The convolution-based image The formula for the block is expressed as:

H ^p =Concat(P(H _0,: ),P(H _1,: ),...,P(H _C,: ));

in, is the output obtained, _Hi is the data of the i-th channel, C is the number of input channels, C' is the number of output channels, W is the input width, L is the input length, and K _w is the volume used for the block-based The width of the convolution kernel of the product, K _l is the length of the convolution kernel used for block-based convolution, /> is the output width obtained by dividing the input width by the kernel width, /> is the output length obtained by passing the input through the length of the input kernel.

Embodiment 4. This embodiment is an example of the setting steps of the convolution-based image blocks in step S12 of the lightweight TRANSFORMER-based 5G NR indoor positioning method described in Embodiment 2. The setting steps are: :

The padding element is set to 0, K _w is set to 8, and K _l is set to 13;

Set the stride in the width direction equal to the width of the convolution kernel, and set the stride in the length direction equal to the length of the convolution kernel.

In actual application, the padding element is set to 0, the stride in the width direction is set equal to the width of the convolution kernel, and the stride in the length direction is set equal to the length of the convolution kernel. Set K _w to 8 and K _l to 13. Because the special aspect ratio of CSI requires such a design to effectively extract features. Also a smaller value of 8 corresponds to a smaller antenna size, while a larger value of 13 corresponds to a larger subcarrier size.

Embodiment 5. Refer to Figure 3 to illustrate this embodiment. This embodiment is a modification of the ReLU-based multi-channel attention mechanism module in step S3 of the lightweight TRANSFORMER-based 5G NR indoor positioning method described in Embodiment 1. The processing flow is given as an example. The processing flow is as follows:

S31. Convert the format of the two-dimensional matrix that is compatible with the attention mechanism to obtain the converted two-dimensional matrix;

S32. Perform dimension transformation on the converted two-dimensional matrix to obtain the transformed two-dimensional matrix;

S33. Use the formula based on the ReLU attention mechanism to perform query processing on the transformed two-dimensional matrix, and obtain the parameter matrix W _q , the key generation parameter matrix W _k and the value generation parameter matrix W _v ;

S34. Use the score formula of the attention mechanism based on ReLU to calculate the parameter matrix W _q , the key generation parameter matrix W _k and the value generation parameter matrix W _v to obtain score data;

S35. Normalize the score data to obtain the final score data;

S36. Use the ReLU activation function to process the final score data, and generate the parameter matrix W _v from the processed result point multiplication value and accumulate it to obtain the final result X matrix;

S37. Input the final result X matrix into the attention mechanism module, and process it through the normalization layer, attention layer and normalization layer in sequence to obtain the output data X _output ;

S38. Input the output data X _output into the multi-layer perceptron layer to obtain the final output data X _output .

When this implementation is actually applied, as shown in Figure 3, the characteristic information obtained by the channel-independent rectangular image block operation module A ReLU-based attention mechanism is adopted to capture the global interactions between subcarriers while maintaining equal or even better time efficiency. The output X _output obtained after performing the attention calculation is input into the multi-layer perceptron to increase the richness of the features. The obtained output is then handed over to the position mapping module for three-dimensional coordinate positioning to complete the positioning task.

The entire formula based on the ReLU attention mechanism is:

X _output =Att(LN(Non-Trans(X)))+X;

Among them, Non-Trans represents the unchanged dimension, LN represents layer norm regularization, and Att represents the attention mechanism based on ReLU.

The score formula of the attention mechanism based on ReLU is described as follows:

score=ReLU(W _q X)(ReLU(W _k X) ^· ReLU(W _v X));

where W _q , W _k and W _v represent the generating parameter matrices of query, key and value respectively. through regularization Obtain the final score, where α is the regularization constant, set it to 1×10 ^-15 . Since the matrix multiplication operation of ReLU-based attention first calculates keys and values, the attention calculation is not centered on C'.

The form of the feature matrix before input _Wi is the output obtained by operating on channel-independent rectangular image patches. It is necessary to perform dimension transformation on X. The specific description formula of the dimension transformation matrix is as follows:

Among them, _xi represents different subcarrier characteristics from the same channel, and x, y and z represent different channels from X. For ease of explanation, the weight subscript can be omitted and represented as w, which represents the weight parameter in a broad sense. In the result of matrix multiplication on the left, each value has the form (wx ₁ +wx ₂ +wx ₃ ). On the contrary, in the result of matrix multiplication on the right, each value is of the form (wx ₁ +wy ₁ +wz ₁ ), which means that there is interaction between subcarriers from different channels, and in the following attention This feature will be enhanced in the calculation of the mechanism. The matrix multiplication on the left is in calculated on the format.

Embodiment 6. This embodiment is an example of the ReLU attention mechanism formula in step S33 of the lightweight TRANSFORMER-based 5G NR indoor positioning method described in Embodiment 5. The ReLU attention-based The mechanism formula is expressed as:

X _output =Att(LN(Non-Trans(X)))+X;

Among them, Non-Trans is the unchanged dimension, LN is layer norm regularization, and Att is the attention mechanism based on ReLU.

Embodiment 7. This embodiment is an example of the scoring formula of the ReLU attention mechanism in step S34 of the lightweight TRANSFORMER-based 5G NR indoor positioning method described in Embodiment 5. The attention of the ReLU The score formula of the force mechanism is expressed as:

score=ReLU(W _q X)(ReLU(W _k X) ^· ReLU(W _v X)).

Embodiment 8. This embodiment provides a 5G NR indoor positioning system based on lightweight TRANSFORMER. The system is:

A storage device used to process CSI input data using a channel-independent rectangular image block operation module to obtain an output matrix;

A storage device for converting the three-dimensional matrix of the output matrix into a two-dimensional matrix compatible with the attention mechanism;

A storage device used to input the two-dimensional matrix compatible with the attention mechanism into the multi-channel attention mechanism module based on ReLU using the Reshape() function to obtain the final output data;

A storage device for positioning the final output data using a position mapping module to obtain a mapping position.

Embodiment 9. This embodiment provides a computer-readable storage medium. A computer program is stored on the computer-readable storage medium. When the computer program is run by a processor, it executes one of the steps described in any one of Embodiments 1 to 7. A 5G NR indoor positioning method based on lightweight TRANSFORMER.

Embodiment 10. This embodiment provides a computer device. A computer program is stored in the memory. When the processor runs the computer program stored in the memory, the processor executes any one of Embodiments 1 to 7. A 5G NR indoor positioning method based on lightweight TRANSFORMER described in the item.

Embodiment 11. Refer to Figure 4 and Figure 5 to illustrate this implementation. This implementation is a verification explanation of a lightweight TRANSFORMER-based 5G NR indoor positioning method described in any one of Embodiment 1 to Embodiment 7:

Three real 5G scene data sets collected in the indoor space of the new experimental building of the Chinese Academy of Sciences in Beijing were used as experimental data. The size of the entire indoor space is 20 meters × 60 meters × 4 meters. It is a large room suitable for applications such as factories and museums. . To obtain CSI, five 5G base stations were deployed using integrated sensing and communications at 3.5GHz with a bandwidth of 100MHz and a power of 40W. The base stations were mounted on plastic brackets at a height of 2.4 meters above the ground, and a random floating height of 0.1 meters was introduced during the simulation to prevent coplanarity. The user equipment acts as a receiver and is placed on a marked lift truck at a height of 1.2 meters above the ground, simulating a person holding a mobile phone at a height of 1.8 meters. The obtained dataset includes 4816 positioning samples, with three datasets corresponding to different representations of CSI at SNR (Signal to Noise Ratio) 10, SNR20 and SNR50. In order to divide the data set into training set, validation set and test set, an approximate ratio of 6:2:2 is used according to different signal-to-noise ratios, resulting in 2888, 964 and 964 samples respectively. The size of a single CSI matrix is 5×16×193, indicating the presence of 5 base stations, each with 16 antennas and each antenna with 193 subcarriers.

For channel-independent rectangular image patches, K _w will be set to 8 and K _l will be set to 13. The dimension of hidden layer C' is 385. The dimensions of the hidden layers in the multilayer perceptron module and the normalization module are both 31. The final localizer uses a linear layer with input size 385 and output size 3.

The initial learning rate is 1×10 ^-4 . After 100 training epochs, the learning rate is halved every 25 epochs, the training round epoch is 300, and the batch size batch_size is 16.

The two evaluation criteria used during the experiment are Root Mean Square Error (RMSE),

Mean Absolute Error (MAE):

where _yi represents the predicted value, represents the true value.

Comparison methods of this implementation include: Clnet method: Complex input lightweight neural network designed for massive MIMO CSI feedback (Complex input lightweight neural network designed for massive MIMO CSI feedback), KNN method: Deep learning for massive MIMO CSI feedback (Deep learning of massive MIMO CSI feedback), MIMOnet method: MaMIMO CSI-Based Positioning usingCNNs: Peeking inside the Black Box (CNN-based MaMIMO CSI positioning: peeking inside the black box), Hiloc method: Hybrid Indoor Localization via Enhanced 5G NR CSI (Hybrid indoor positioning based on enhanced 5G NR CSI), SVM method: Toward 5G NR High-Precision Indoor Positioning via Channel Frequency Response: A New Paradigm and Dataset Generation Method (5G NR high-precision indoor positioning based on channel frequency response: A New Paradigm and Dataset Generation Method) and MPRI method: Toward 5G NR High-Precision Indoor Positioning via Channel FrequencyResponse: A New Paradigm and Dataset Generation Method (5G NR High-Precision Indoor Positioning via Channel Frequency Response: A New Paradigm and Dataset Generation Method) Dataset generation method).

The error of the above method is compared with the error of a 5G NR indoor positioning method based on lightweight TRANSFORMER described in this embodiment. The comparison results are shown in Figure 5. It can be seen from the figure that the error of this embodiment is Compared with CLnet, KNN, CSInet, MIMOnet, Hiloc, SVM, and MPRI methods, the described 5G NR indoor positioning method based on lightweight TRANSFORMER significantly reduces the error size and effectively improves the prediction accuracy.

At the same time, the cumulative distribution function distribution of the above method is compared with the 5G NR indoor positioning method based on lightweight TRANSFORMER described in this embodiment, as shown in Figure 4(a), Figure 4(b) and Figure 4(c) , it can be seen that the 5G NR indoor positioning method based on lightweight TRANSFORMER described in this embodiment has greatly improved positioning accuracy compared with CLnet, KNN, CSInet, MIMOnet, Hiloc, SVM, and MPRI methods.

In addition, the terms “first” and “second” are used for descriptive purposes only and cannot be understood as indicating or implying relative importance or implicitly indicating the quantity of indicated technical features. Therefore, features defined as "first" and "second" may explicitly or implicitly include at least one of these features. In the description of the present invention, "plurality" means at least two, such as two, three, etc., unless otherwise expressly and specifically limited.

In the description of this specification, reference to the terms "one embodiment," "some embodiments," "an example," "specific examples," or "some examples" or the like means that specific features are described in connection with the embodiment or example. , structures, materials or features are included in at least one embodiment or example of the invention. In this specification, the schematic expressions of the above terms are not necessarily directed to the same embodiment or example. Furthermore, the specific features, structures, materials or characteristics described may be combined in any suitable manner in any one or more embodiments or examples. Furthermore, those skilled in the art may combine and combine different embodiments or examples and features of different embodiments or examples described in this specification unless they are inconsistent with each other.

The above are only embodiments of the present invention, and are not limited to the present invention. For those skilled in the art, the present invention may have various modifications and changes. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention shall be included in the scope of the claims of the present invention.

Claims (10)

1. A 5G NR indoor positioning method based on lightweight TRANSFORMER, characterized in that the method is: S1. Use the channel-independent rectangular image block operation module to process the CSI input data and obtain the output matrix; S2. Use the Reshape() function to convert the three-dimensional matrix of the output matrix into a two-dimensional matrix that is compatible with the attention mechanism; S3. Input the two-dimensional matrix compatible with the attention mechanism into the multi-channel attention mechanism module based on ReLU to obtain the final output data; S4. Use the position mapping module to position the final output data to obtain the mapping position. 2. A 5G NR indoor positioning method based on lightweight TRANSFORMER according to claim 1, characterized in that the processing flow of the channel-independent rectangular image block operation module in step S1 is: S11. Divide the CSI input data according to different channels to obtain _Hi data; S12. Use convolution-based image blocks to perform image block operations on the _Hi data to obtain an output matrix. 3. A 5G NR indoor positioning method based on lightweight TRANSFORMER according to claim 2, characterized in that the formula of the convolution-based image block in step S12 is expressed as: H ^p =Concat(P(H _0,: ),P(H _1,: ),...,P(H _C,: )); in, is the output obtained, _Hi is the data of the i-th channel, C is the number of input channels, C' is the number of output channels, W is the input width, L is the input length, and K _w is the volume used for the block-based The width of the convolution kernel of the product, K _l is the length of the convolution kernel used for block-based convolution, /> is the output width obtained by dividing the input width by the kernel width, /> is the output length obtained by passing the input through the length of the input kernel. 4. A 5G NR indoor positioning method based on lightweight TRANSFORMER according to claim 2, characterized in that the setting step of the convolution-based image block in step S12 is: The padding element is set to 0, K _w is set to 8, and K _l is set to 13; Set the stride in the width direction equal to the width of the convolution kernel, and set the stride in the length direction equal to the length of the convolution kernel. 5. A 5G NR indoor positioning method based on lightweight TRANSFORMER according to claim 1, characterized in that the processing flow of the ReLU-based multi-channel attention mechanism module in step S3 is: S31. Convert the format of the two-dimensional matrix that is compatible with the attention mechanism to obtain the converted two-dimensional matrix; S32. Perform dimension transformation on the converted two-dimensional matrix to obtain the transformed two-dimensional matrix; S33. Use the formula based on the ReLU attention mechanism to perform query processing on the transformed two-dimensional matrix, and obtain the parameter matrix W _q , the key generation parameter matrix W _k and the value generation parameter matrix W _v ; S34. Use the score formula of the attention mechanism based on ReLU to calculate the parameter matrix W _q , the key generation parameter matrix W _k and the value generation parameter matrix W _v to obtain score data; S35. Normalize the score data to obtain the final score data; S36. Use the ReLU activation function to process the final score data, and generate the parameter matrix W _v from the processed result point multiplication value and accumulate it to obtain the final result X matrix; S37. Input the final result X matrix into the attention mechanism module, and process it through the normalization layer, attention layer and normalization layer in sequence to obtain the output data X _output ; S38. Input the output data X _output into the multi-layer perceptron layer to obtain the final output data X _output . 6. A 5G NR indoor positioning method based on lightweight TRANSFORMER according to claim 5, characterized in that the ReLU attention mechanism formula in step S33 is expressed as: X _output =Att(LN(Non-Trans(X)))+X; Among them, Non-Trans is the unchanged dimension, LN is layer norm regularization, and Att is the attention mechanism based on ReLU. 7. A 5G NR indoor positioning method based on lightweight TRANSFORMER according to claim 5, characterized in that the score formula of the attention mechanism of ReLU in step S34 is expressed as: score=ReLU(W _q X)(ReLU(W _k X) ^· ReLU(W _v X)). 8. A 5G NR indoor positioning system based on lightweight TRANSFORMER, characterized in that the system is: A storage device used to process CSI input data using a channel-independent rectangular image block operation module to obtain an output matrix; A storage device for converting the three-dimensional matrix of the output matrix into a two-dimensional matrix compatible with the attention mechanism; A storage device used to input the two-dimensional matrix compatible with the attention mechanism into the multi-channel attention mechanism module based on ReLU using the Reshape() function to obtain the final output data; A storage device for positioning the final output data using a position mapping module to obtain a mapping position. 9. A computer-readable storage medium, characterized in that a computer program is stored on the computer-readable storage medium, and when the computer program is run by a processor, it executes the light-based method described in any one of claims 1-7. 5G NR indoor positioning method of magnitude TRANSFORMER. 10. A computer device, characterized in that: the device includes a memory and a processor, a computer program is stored in the memory, and when the processor runs the computer program stored in the memory, the processor executes the claim A 5G NR indoor positioning method based on lightweight TRANSFORMER as described in any one of 1-7.