A CSI combined with CDPM indoor positioning method and system

Abstract

The present invention discloses a method and system for indoor positioning using CSI combined with CDPM. In this method, the actual collected CSI data is first converted into a multi-channel amplitude feature map and combined with its corresponding reference positions to form an original multi-channel amplitude feature map dataset P. The original dataset P is then input into the CDPM, and a conditional noise predictor is trained through the CDPM's forward diffusion process. A large dataset Q conforming to a standard Gaussian distribution is then iteratively denoised using the trained conditional noise predictor and the CDPM model's reverse denoising process to generate a simulated multi-channel amplitude feature map dataset #imgabs0#. The original dataset P is then merged with the simulated dataset #imgabs1# to train a residual-structured neural network regressor for indoor positioning. Using this CSI dataset expansion method, indoor positioning accuracy can be improved by only collecting limited CSI data, significantly reducing the manpower and time required to collect CSI data.

Description

Technical Field

[0001] The present invention relates to an indoor positioning method combining CSI (Channel State Information) and CDPM (Conditional Diffusion Probabilistic Model), belonging to the technical field of wireless positioning based on deep learning. Background Art

[0002] In Wi-Fi-based indoor positioning systems, two main types of channel information are RSSI (Received Signal Strength Indicator) and CSI (CSI). RSSI reflects received signal strength and is susceptible to multipath and environmental factors. Even in a stable environment, the RSSI received by a single device can fluctuate by approximately ±5dB. While RSSI is easy to obtain, it contains less information and is not suitable for high-precision indoor positioning. In contrast, CSI, thanks to OFDM and MIMO technologies, encompasses channel information from different subcarriers on different antennas, better reflecting channel multipath and making it more suitable for high-precision indoor positioning.

[0003] Mainstream indoor positioning algorithms mainly include two categories: positioning methods based on channel attenuation models and positioning methods based on fingerprint databases. Positioning methods based on channel attenuation models often require setting channel hyperparameters related to the environment in advance, converting channel information into distance, and then using algorithms such as three-sided positioning for position estimation. The positioning error is large. The positioning method based on fingerprint databases does not require knowing the channel attenuation model in advance. It can use the matching degree between the extracted features and the data in the fingerprint database to obtain the estimated position, and is more robust.

[0004] The quality of collected fingerprint datasets is directly related to positioning accuracy. The less fingerprint data collected, the sparser the distribution of reference locations, and the lower the positioning accuracy. To expand fingerprint databases, many methods based on generative models have emerged in recent years. These methods can be mainly divided into two types: those based on variational autoencoders (VAE) and those based on generative adversarial networks (GAN). However, the loss function of VAEs consists of two parts: reconstruction loss and potential loss. Due to the trade-off, they often generate relatively blurry data. GANs, due to the game mechanism between the generator and discriminator, are difficult to train and prone to mode collapse, which greatly reduces the diversity of generated data.

[0005] Therefore, how to improve indoor positioning accuracy while generating diverse and high-quality CSI data has important application value. Summary of the Invention

[0006] The technical problem to be solved by the present invention is how to improve indoor positioning accuracy while generating diverse and high-quality CSI data.

[0007] To solve the above technical problems, the present invention provides an indoor positioning method combining CSI with CDPM, which includes an offline training phase and an online positioning phase;

[0008] The offline training phase includes the following steps:

[0009] Step 1. Extract the amplitude data from the actual CSI sampling sequence, segment the continuous CSI amplitude data using the sliding window method, and convert the resulting window into a multi-channel amplitude feature map to form the original data set;

[0010] Step 2. Use the original dataset P in step 1 to train a CDPM model, and use the trained CDPM model to generate more simulated multi-channel amplitude feature maps to form a simulated multi-channel amplitude feature map dataset; merge the dataset P and the simulated multi-channel amplitude feature map dataset to obtain the expanded multi-channel amplitude feature map dataset ;

[0011] Step 3. The expanded multi-channel amplitude feature map dataset obtained in step 2 Input a neural network regressor R based on a residual structure for training. The loss function of the regressor is the mean square error function. The network parameters of the regressor R are continuously updated using the gradient descent method until convergence.

[0012] In the online positioning stage, the amplitude data in the collected CSI data is extracted, and the CSI amplitude data is converted into a multi-channel amplitude feature map using the method in step 1 in the offline training stage. The map is input into the trained regressor R to obtain the estimated position of the target to be measured.

[0013] A CSI combined with CDPM indoor positioning system, characterized by comprising an offline training module and an online positioning module;

[0014] The offline training module includes the following functional modules:

[0015] The original data set generation module extracts the amplitude data from the actual CSI sampling sequence, segments the continuous CSI amplitude data using the sliding window method, and converts the resulting windows into multi-channel amplitude feature maps to form the original data set.

[0016] Simulation dataset module: Use the original dataset P in the original dataset generation module to train a CDPM model, use the trained CDPM model to generate more simulated multi-channel amplitude feature maps to form a simulated multi-channel amplitude feature map dataset; merge the dataset P and the simulated multi-channel amplitude feature map dataset to obtain the expanded multi-channel amplitude feature map dataset ;

[0017] Neural network training module: The expanded multi-channel amplitude feature map dataset obtained in the simulation dataset module A neural network regressor R based on a residual structure is input for training. The loss function of the regressor is the mean square error function. The network parameters of the regressor R are continuously updated using the gradient descent method until convergence.

[0018] In the online positioning module, the amplitude data in the collected CSI data is extracted, and the CSI amplitude data is converted into a multi-channel amplitude feature map using the method of the original data set generation module in the offline training module. The map is then input into the trained neural network regressor R to obtain the estimated position of the target to be measured.

[0019] The present invention achieves the following beneficial effects: using CDPM to simulate multidimensional amplitude feature maps reduces the manpower and time required to collect CSI data, improving positioning accuracy by only collecting limited CSI data. Compared with traditional generative models, CDPM avoids the adversarial training issue and is therefore easier to train. The present invention utilizes a residual-structured neural network classifier and regressor to increase network depth while avoiding gradient vanishing and gradient exploding, preventing network degradation, and enhancing the network's fitting capabilities, resulting in improved classification accuracy and positioning precision. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 This is the overall system architecture diagram of the present invention;

[0021] Figure 2 Schematic diagram of CDPM forward and reverse process;

[0022] Figure 3 Schematic diagram of training the conditional noise predictor for the CDPM forward process;

[0023] Figure 4 This is the conditional predictor network structure in CDPM;

[0024] Figure 5 Schematic diagram of the CDPM reverse process;

[0025] Figure 6 Generate a resulting diagram for CDPM. DETAILED DESCRIPTION

[0026] The technical solution of the present invention is further described in detail below with reference to the accompanying drawings and embodiments.

[0027] Example 1

[0028] like Figure 1 As shown, a CSI combined with CDPM indoor positioning method of the present invention includes an offline training phase and an online positioning phase;

[0029] The offline training phase includes the following steps:

[0030] Step 1. Extract the amplitude data from the actual CSI sampling sequence, segment the continuous CSI amplitude data using the sliding window method, and convert the resulting window into a multi-channel amplitude feature map to form the original data set;

[0031] The specific steps include:

[0032] Assuming that indoor positioning adopts transmitting antennas, receiving antennas, subcarriers, each reference position collects M sampling cycles, transmit antennas and The receiving antennas can be combined into pairs ( × ) For the antenna, extract the amplitude data from the CSI sampling sequence actually collected at L different reference positions, and use the sliding window method to segment the continuous CSI amplitude data. The size of the sliding window is set to W sampling periods, and the sliding step is set to S sampling periods. Each reference position obtains M / S windows, and each window contains × × ×W amplitude values;

[0033] The amplitude window obtained after segmentation is converted into a multi-channel amplitude feature map. The number of channels of the multi-channel amplitude feature map is ( × ), the length of the multi-channel amplitude feature map is , the width is W, and the multi-channel amplitude feature map is recorded as ,in, , Indicates the Amplitude data of W sampling periods corresponding to the subcarriers on the transmitting and receiving antennas;

[0034] All multi-channel amplitude feature maps and their corresponding reference positions are combined into the original multi-channel amplitude feature map dataset, denoted as ,in represents the kth multi-channel amplitude feature map in the original data set P, Multi-channel amplitude feature map The corresponding reference position, K is the total number of multi-channel amplitude feature maps in the original dataset P, and the number is M / S×L.

[0035] Step 2. Use the original dataset P in step 1 to train a CDPM model, and use the trained CDPM model to generate more simulated multi-channel amplitude feature maps to form a simulated multi-channel amplitude feature map dataset; merge the dataset P and the simulated multi-channel amplitude feature map dataset to obtain the expanded multi-channel amplitude feature map dataset .

[0036] The CDPM model includes a forward diffusion process, a conditional noise predictor and a reverse denoising process;

[0037] The forward diffusion process specifically comprises the following steps:

[0038] Take a multi-channel amplitude feature map in the original data set P As input, the multi-channel amplitude feature map is continuously fed into the T-step diffusion in sequence. Add specific and multi-channel amplitude feature maps Gaussian noise of the same dimension , t=1,2,…,T, obtain the noisy multi-channel amplitude feature map under different diffusion steps t , after T-step diffusion, a clear multi-channel amplitude feature map becomes isotropic standard Gaussian distribution noise, where Gaussian noise is added Refers to the pixel value of the amplitude feature map of each channel in the multi-channel amplitude feature map and Gaussian noise Add the corresponding values.

[0039] In the forward diffusion process, the specific Gaussian noise added at different diffusion steps t Is a multi-channel amplitude feature map The mean of the same dimension is , the variance is The matrix of the Gaussian distribution of , , , , , is the noise coefficient that changes with the diffusion step, through The cumulative multiplication calculation is obtained. is the Gaussian distribution parameter of the current diffusion step, through and Calculated, is pi, cos is a trigonometric function, Indicates that the multi-channel amplitude feature map Each pixel value of the amplitude feature map of each channel is multiplied by .

[0040] The conditional noise predictor of the CDPM model is a neural network, which is composed of an embedding transformation part, a downsampling part, an upsampling part, and a restoration part in sequence. The embedding transformation part includes a two-dimensional convolutional layer and an embedding transformation module. The embedding transformation module includes a Sinusoidal embedding transformation layer and a multi-layer perceptron for extracting feature information of diffusion steps and reference positions, respectively. The downsampling part uses four downsampling modules, each of which includes a residual module and a downsampling layer based on maximum pooling, and each downsampling layer downsamples by one time. The upsampling part uses four upsampling modules, each of which includes a residual module and an upsampling layer based on nearest neighbor interpolation, and each upsampling layer upsamples by one time. When calculating each upsampling layer, the current result is parallelized with the downsampling information of the corresponding resolution. The restoration part includes two residual modules and a two-dimensional convolutional layer to restore the resolution of the upsampling part result to the resolution of the input image. The above-mentioned residual module includes two convolution modules connected in series, each of which includes a two-dimensional convolutional layer, a regularization layer, and a SiLU activation function.

[0041] The forward diffusion process trains the conditional noise predictor in the CDPM model. The training steps are:

[0042] 1) Take a set of multi-channel amplitude feature maps and their reference positions in the original data set P in step 1 and a random diffusion step t input conditional noise predictor, where t is a random integer in the integer set {1, 2, ..., T};

[0043] 2) Using the above forward diffusion process, calculate The image that becomes noisy in the t-th diffusion step ;

[0044] 3) Using the embedded transformation part of the conditional noise predictor to transform the diffusion step t and the reference position Perform embedding transformation, where empty labels are used Replace the reference position with a probability of 20% ;

[0045] 4) Apply the following loss function and update the neural network parameters of the conditional noise predictor using gradient descent:

[0046]

[0047] In the formula represents the L2 normal form, The original multi-channel amplitude feature map Same-dimensional obedience The standard Gaussian matrix of Conditional noise predictor for the original multi-channel amplitude feature map when taking an empty label for the reference position The predicted value of the added noise at the t-th diffusion step in the forward diffusion process, The reference position is Time-conditional noise predictor for the original multi-channel amplitude feature map The predicted value of the added noise at the t-th diffusion step in the forward diffusion process, is the weighted prediction noise, which is obtained by weighting the above two prediction noises;

[0048] 5) Repeat steps 1) to 4), iterate the k value, update the neural network parameters, and use the updated neural network parameters as the neural network parameters of the conditional noise predictor in the next cycle until the loss function converges and the conditional noise predictor training is completed.

[0049] In the reverse denoising process of the CDPM model, the number of denoising steps is the same as the diffusion steps in the forward diffusion process. In the reverse denoising process, T denoising steps are performed, and the processing order is opposite to that of the diffusion process, that is, t = T, T-1, ..., 1. is a standard Gaussian noise matrix, for Iterate denoising to the noisy feature map at step t. If you want to generate a reference position The multi-dimensional amplitude feature map of the noise feature map and reference position Input the reverse denoising process, using the noise predicted by the conditional noise predictor For noisy feature maps Perform iterative denoising, and the denoising formula is , where , , and Obtained from the forward diffusion process, To obey The matrix of the noisy feature map is continuously The corresponding reference position without noise obtained by iterative denoising is Simulated multi-channel amplitude characteristic diagram of .

[0050] The reference position is generated by using the trained conditional noise predictor and the reverse denoising process as The specific steps of simulating multi-dimensional amplitude characteristic map are as follows:

[0051] 1) Sample from the standard Gaussian distribution to obtain the original multi-channel amplitude feature map Standard Gaussian noise matrix of the same dimension ;

[0052] 2) Repeat step 1) to convert a large number of standard Gaussian noise matrices and the reference position to be generated The set composed of , I represents The number of standard Gaussian noise matrices, For collection The i-th standard Gaussian noise matrix in;

[0053] 3) Using the above CDPM reverse denoising process, the standard Gaussian noise matrix Iterative denoising to obtain simulated multi-channel amplitude feature map ;

[0054] 4) Repeat 3), iterating i until the set All matrices in the denoising are obtained ={ }, The corresponding reference position The simulated multi-channel amplitude feature map dataset;

[0055] 5) Iteration , get the simulated multi-channel amplitude feature map dataset of all reference positions , , L represents the set of all reference positions.

[0056] Merge dataset P and simulated multi-channel amplitude feature map dataset Get the expanded multi-channel amplitude feature map dataset .

[0057] Step 3. The expanded multi-channel amplitude feature map dataset obtained in step 2 A neural network regressor R based on a residual structure is input for training. The loss function of the regressor is the mean square error function. The gradient descent method is used to continuously update the network parameters of the regressor R until convergence.

[0058] The regressor R is a neural network comprising an initial convolutional part, a residual part and a fully connected part. The initial convolutional part comprises a two-dimensional convolutional layer and a maximum pooling layer. The residual part comprises four connected residual modules with dimensions of 64, 128, 256, and 512, respectively. The fully connected part comprises an average pooling layer and a fully connected layer, and the output dimension of the fully connected layer is 2.

[0059] In the online positioning stage, the amplitude data in the collected CSI data is extracted, and the CSI amplitude data is converted into a multi-channel amplitude feature map using the method in step 1 in the offline training stage. The map is input into the trained regressor R to obtain the estimated position of the target to be measured.

[0060] A CSI combined with CDPM indoor positioning system, characterized by comprising an offline training module and an online positioning module;

[0061] The offline training module includes the following functional modules:

[0062] The original data set generation module extracts the amplitude data from the actual CSI sampling sequence, segments the continuous CSI amplitude data using the sliding window method, and converts the resulting windows into multi-channel amplitude feature maps to form the original data set.

[0063] Simulation dataset module: Use the original dataset P in the original dataset generation module to train a CDPM model, use the trained CDPM model to generate more simulated multi-channel amplitude feature maps to form a simulated multi-channel amplitude feature map dataset; merge the dataset P and the simulated multi-channel amplitude feature map dataset to obtain the expanded multi-channel amplitude feature map dataset ;

[0064] Neural network training module: The expanded multi-channel amplitude feature map dataset obtained in the simulation dataset module Input a neural network regressor R based on a residual structure for training. The loss function of the regressor is the mean square error function. The network parameters of the regressor R are continuously updated using the gradient descent method until convergence.

[0065] In the online positioning module, the amplitude data in the collected CSI data is extracted, and the CSI amplitude data is converted into a multi-channel amplitude feature map using the method of the original data set generation module in the offline training module. The map is then input into the trained neural network regressor R to obtain the estimated position of the target to be measured.

[0066] The working process of each functional module in the indoor positioning system of CSI combined with CDPM is the same as that in the indoor positioning method of CSI combined with CDPM.

[0067] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0068] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0069] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0070] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0071] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A CSI combined with CDPM indoor positioning method, characterized in that: Includes offline training phase and online positioning phase; The offline training phase includes the following steps: Step 1. Extract the amplitude data from the actual CSI sampling sequence, segment the continuous CSI amplitude data using the sliding window method, and convert the resulting window into a multi-channel amplitude feature map to form the original data set; Step 2. Use the original dataset P in step 1 to train a CDPM model, and use the trained CDPM model to generate more simulated multi-channel amplitude feature maps to form a simulated multi-channel amplitude feature map dataset; merge the dataset P and the simulated multi-channel amplitude feature map dataset to obtain the expanded multi-channel amplitude feature map dataset Step 3. The expanded multi-channel amplitude feature map dataset obtained in step 2 A neural network regressor R based on a residual structure is input for training. The loss function of the regressor is the mean square error function. The network parameters of the regressor R are continuously updated using the gradient descent method until convergence. In the online positioning stage, the amplitude data in the collected CSI data is extracted, and the CSI amplitude data is converted into a multi-channel amplitude feature map using the method in step 1 in the offline training stage. The map is input into the trained regressor R to obtain the estimated position of the target to be measured.

2. The indoor positioning method combining CSI with CDPM according to claim 1, characterized in that: In step 1, the following steps are specifically included: Assume that indoor positioning adopts N d Transmit antennas, N r receiving antennas, N s subcarriers, each reference position collects M sampling cycles, N d transmit antennas and N r The receiving antennas can be composed of two pairs (N d ×N r ) For the antenna, extract the amplitude data from the CSI sampling sequence actually collected at L different reference positions, and use the sliding window method to segment the continuous CSI amplitude data. The size of the sliding window is set to W sampling periods, and the sliding step is set to S sampling periods. Each reference position obtains M / S windows, and each window contains N d ×N r ×N s ×W amplitude values; The amplitude window obtained after segmentation is converted into a multi-channel amplitude feature map. The number of channels of the multi-channel amplitude feature map is (N d ×N r ), the length of the multi-channel amplitude feature map is N s , the width is W, and the multi-channel amplitude feature map is recorded as in, Indicates the Amplitude data of W sampling periods corresponding to the subcarriers on the transmitting and receiving antennas; All multi-channel amplitude feature maps and their corresponding reference positions are combined into the original multi-channel amplitude feature map dataset, denoted as Among them C k represents the kth multi-channel amplitude feature map in the original dataset P, l k is the multi-channel amplitude feature map C k The corresponding reference position, K is the total number of multi-channel amplitude feature maps in the original dataset P, and the number is M / S×L.

3. The indoor positioning method combining CSI with CDPM according to claim 1, characterized in that: In step 2, the CDPM model includes a forward diffusion process, a conditional noise predictor, and a reverse denoising process; The forward diffusion process specifically comprises the following steps: Take a multi-channel amplitude feature map C in the original data set P k As input, it continuously transmits the multi-channel amplitude feature map C to the multi-channel amplitude feature map C in the T-step diffusion process. k Add specific and multi-channel amplitude feature maps C k Gaussian noise ε of the same dimension t , t=1,2,…,T, obtain the noisy multi-channel amplitude feature map under different diffusion steps t After T-step diffusion, a clear multi-channel amplitude feature map C k becomes isotropic standard Gaussian distribution noise, where Gaussian noise ε is added t Refers to the pixel value of the amplitude feature map of each channel in the multi-channel amplitude feature map and the Gaussian noise ε t Add the corresponding values.

4. The indoor positioning method combining CSI with CDPM according to claim 3, characterized in that: In the forward diffusion process, the specific Gaussian noise ε added at different diffusion steps t t Is a multi-channel amplitude feature map C k The mean of the same dimension is The variance is 1-α t The matrix of the Gaussian distribution, where α t =1-β t , is the noise coefficient that changes with the diffusion step, through α t Obtained by cumulative multiplication, β t is the Gaussian distribution parameter of the current diffusion step, through γ t and f(t), π is pi, cos is the trigonometric function, Indicates that each pixel value of the amplitude feature map of each channel of the multi-channel amplitude feature map C is multiplied by 5. The indoor positioning method combining CSI with CDPM according to claim 4, characterized in that: The conditional noise predictor of the CDPM model is a neural network, which is composed of an embedding transformation part, a downsampling part, an upsampling part and a restoration part in sequence; the embedding transformation part includes a two-dimensional convolutional layer and an embedding transformation module, and the embedding transformation module includes a sinusoidal embedding transformation layer and a multi-layer perceptron for extracting feature information of diffusion steps and reference positions respectively; the downsampling part adopts four downsampling modules, each of which includes a residual module and a downsampling layer based on maximum pooling, and each downsampling layer downsamples by one time; the upsampling part adopts four upsampling modules, each of which includes a residual module and an upsampling layer based on nearest neighbor interpolation, and each upsampling layer upsamples by one time, and when calculating each upsampling layer, the current result is paralleled with the downsampling information of the corresponding resolution; The restoration part includes two residual modules and a two-dimensional convolution layer for restoring the resolution of the upsampling part result to the resolution of the input image. The residual module includes two convolution modules connected in series, and each convolution module includes a two-dimensional convolution layer, a regularization layer and a SiLU activation function.

6. The indoor positioning method combining CSI with CDPM according to claim 4, characterized in that: The forward diffusion process trains the conditional noise predictor in the CDPM model. The training steps are: 1) Take a set of multi-channel amplitude feature maps in the original data set P in step 1 and its reference position (C k ,l k ) and a random diffusion step t input conditional noise predictor, t is a random integer in the integer set {1,2,…,T}; 2) Using the forward diffusion process above, calculate C k The image that becomes noisy in the t-th diffusion step 3) Using the embedded transformation part of the conditional noise predictor to transform the diffusion step t and the reference position l k Perform embedding transformation; 4) Apply the following loss function and update the neural network parameters of the conditional noise predictor using gradient descent: In the formula 2 represents the L2 norm, ε t is the original multi-channel amplitude feature map C k The same-dimensional standard Gaussian matrix obeys N(0,I), The conditional noise predictor is used to calculate the original multi-channel amplitude feature map C when taking an empty label for the reference position. k The predicted value of the added noise at the t-th diffusion step in the forward diffusion process, The reference position is l k The time-conditional noise predictor is used to calculate the original multi-channel amplitude feature map C k The predicted value of the added noise at the t-th diffusion step in the forward diffusion process, is the weighted prediction noise, which is obtained by weighting the above two prediction noises; 5) Repeat steps 1) to 4), iterate the k value, update the neural network parameters, and use the updated neural network parameters as the neural network parameters of the conditional noise predictor in the next cycle until the loss function converges and the conditional noise predictor training is completed.

7. The indoor positioning method combining CSI with CDPM according to claim 6, characterized in that: In the reverse denoising process of the CDPM model, the number of denoising steps is the same as the diffusion steps in the forward diffusion process. In the reverse denoising process, T denoising steps are performed, and the processing order is opposite to that of the diffusion process, that is, t = T, T-1, ..., 1. is a standard Gaussian noise matrix, for Iterate denoising to the noisy feature map at step t. If you want to generate a reference position The multi-dimensional amplitude feature map of the noise feature map and reference position Input the reverse denoising process, using the noise predicted by the conditional noise predictor For noisy feature maps Perform iterative denoising, and the denoising formula is Where t=T,T-1,...,1, and β t Obtained from the forward diffusion process, z is a matrix that obeys N(0,I), and the above formula is used to continuously calculate the noisy feature map The corresponding reference position without noise obtained by iterative denoising is Simulated multi-channel amplitude characteristic diagram of 8. The indoor positioning method combining CSI with CDPM according to claim 7, characterized in that: The reference position is generated by using the trained conditional noise predictor and the reverse denoising process as The specific steps of simulating multi-dimensional amplitude characteristic map are as follows: 1) Sampling from the standard Gaussian distribution to obtain the original multi-channel amplitude feature map C k The standard Gaussian noise matrix Y of the same dimension; 2) Repeat step 1) and add the standard Gaussian noise matrix Y and the reference position to be generated The set composed of I represents The number of standard Gaussian noise matrices, Y i For collection The i-th standard Gaussian noise matrix in; 3) Using the reverse denoising process of CDPM, the standard Gaussian noise matrix Y i Iterative denoising to obtain simulated multi-channel amplitude feature map 4) Repeat 3), iterating i until the set All matrices in the denoising are obtained The corresponding reference position The simulated multi-channel amplitude feature map dataset; 5) Iteration Get the simulated multi-channel amplitude feature map dataset of all reference positions L represents the set of all reference positions.

9. The indoor positioning method combining CSI with CDPM according to claim 1, characterized in that: In step 3, the regressor R is a neural network comprising an initial convolutional part, a residual part and a fully connected part. The initial convolutional part comprises a two-dimensional convolutional layer and a maximum pooling layer. The residual part comprises four connected residual modules with dimensions of 64, 128, 256, and 512, respectively. The fully connected part comprises an average pooling layer and a fully connected layer. The output dimension of the fully connected layer is 2.

10. An indoor positioning system combining CSI with CDPM, characterized in that: Includes offline training module and online positioning module; The offline training module includes the following functional modules: The original data set generation module extracts the amplitude data from the actual CSI sampling sequence, segments the continuous CSI amplitude data using the sliding window method, and converts the resulting windows into multi-channel amplitude feature maps to form the original data set. Simulation dataset module: Use the original dataset P in the original dataset generation module to train a CDPM model, use the trained CDPM model to generate more simulated multi-channel amplitude feature maps to form a simulated multi-channel amplitude feature map dataset; merge the dataset P and the simulated multi-channel amplitude feature map dataset to obtain the expanded multi-channel amplitude feature map dataset Neural network training module: The expanded multi-channel amplitude feature map dataset obtained in the simulation dataset module A neural network regressor R based on a residual structure is input for training. The loss function of the regressor is the mean square error function. The network parameters of the regressor R are continuously updated using the gradient descent method until convergence. In the online positioning module, the amplitude data in the collected CSI data is extracted, and the CSI amplitude data is converted into a multi-channel amplitude feature map using the method of the original data set generation module in the offline training module. The map is then input into the trained neural network regressor R to obtain the estimated position of the target to be measured.

Images