A generative radio map collaborative estimation method for multi-signal source scenarios

Abstract

The application discloses a kind of generative radio map collaborative estimation methods for multi-signal source scene, comprising the following steps: S1. the received signal strength sample of multiple users on multiple frequencies is constructed into matrix form, and is decomposed using NMF matrix decomposition method, each received signal strength component decomposed corresponds to an independent signal source;S2. single signal source radio map inference model is constructed and trained;S3. for each received signal strength component obtained in step S1 decomposition, the corresponding is estimated using single signal source radio map inference model.The application decomposes the signal strength component corresponding to multiple signal sources using matrix decomposition theory, and trains the neural network using the learning algorithm based on GAN, improving the accuracy and generalization ability of the system. This method can also work effectively in the case of inaccurate geographic data or matrix decomposition results with errors.

Description

TECHNICAL FIELD

[0001] The present application relates to radio map estimation, in particular to a generative radio map collaborative estimation method for multi-signal source scenarios. BACKGROUND

[0002] In the 6G era, real-time monitoring and management of radio resources are crucial to support diversified wireless applications. Traditional radio map estimation (RME) methods, such as data-driven interpolation methods and model-based ray tracing methods, cannot simultaneously consider high accuracy and high speed, and cannot achieve accurate and fast radio map estimation. Existing deep learning-based methods are expected to solve the above problems, i.e. by training a deep neural network with high-precision data sets to enable the deep neural network to achieve fast and accurate radio map estimation.

[0003] However, current deep learning-based methods rely on signal source information to achieve accurate radio map estimation, i.e. signal source information is an indispensable part of neural network input information. However, in a wireless network, the signal source may be a base station or various mobile devices, and its information (especially real-time information) is often difficult to obtain. In addition, such estimation methods that rely on signal source information often cannot adapt to the case where multiple signal sources exist at the same time. This is because the number of signal sources corresponding to the training data is fixed or within a certain range, while the number of signal sources in the actual scene is often uncertain, so using the trained model for inference often causes large estimation bias. In addition, such methods have large errors and poor generalization ability when the geographical environment information is inaccurate. SUMMARY

[0004] The purpose of the present application is to overcome the shortcomings of the prior art and provide a generative radio map collaborative estimation method for multi-signal source scenarios, which uses matrix decomposition theory to decompose the signal strength components corresponding to multiple signal sources, and uses a GAN-based learning algorithm to train a neural network, improving the accuracy and generalization ability of the system. This method can also work effectively when the geographical data is inaccurate or the matrix decomposition result has errors.

[0005] The purpose of the present application is achieved by the following technical solution: a generative radio map collaborative estimation method for multi-signal source scenarios, comprising the following steps:

[0006] S1. Constructing the received signal strength samples of multiple users at multiple frequencies into a matrix form and decomposing it using the NMF matrix decomposition method, each received signal strength component decomposed corresponds to an independent signal source;

[0007] S2. Constructing and training a single-signal-source radio map inference model;

[0008] S3. For each received signal strength component obtained by the decomposition in step S1, estimate the corresponding single-signal-source radio map using the single-signal-source radio map inference model, and superimpose to obtain the radio map estimation result of the multi-signal-source scenario.

[0009] The present application has the following beneficial effects: The present application can estimate the radio map based on the received signal strength (RSS) information of multiple mobile users and the geographical environment map, without signal source information. The method uses matrix decomposition theory to decompose the signal strength components corresponding to multiple signal sources, and uses a learning algorithm based on generative adversarial networks (GAN) to train a neural network, thereby improving the accuracy and generalization ability of the system. The method can also work effectively in the case of inaccurate geographical data or matrix decomposition results with errors, automatically corrects some errors, and ensures the accuracy and robustness of the estimation. BRIEF DESCRIPTION OF DRAWINGS

[0010] Figure 1 Figure (300 users, NMSE = 0.0104) for the effect comparison between the radio map label of the multi-signal-source and the radio map recovered using signal sampling;

[0011] Figure 2 Figure for the collaborative estimation framework of the radio map based on deep learning;

[0012] Figure 3 Figure for the model training process based on the GAN framework;

[0013] Figure 4 Figure for the estimation accuracy of the proposed scheme when the physical environment map is error-free: (left) label, (middle) estimation effect based on 300 RSS sampling points, (right) estimation effect based on 1000 RSS sampling points;

[0014] Figure 5 Figure for the estimation accuracy of the proposed scheme when the physical environment map has errors (part of the buildings are missing, the black part is missing): (left) label, (middle) estimation effect based on 300 RSS sampling points, (right) estimation effect based on 1000 RSS sampling points;

[0015] Figure 6 Figure for the method flowchart of the present application. DETAILED DESCRIPTION

[0016] The technical solutions of the present application will be described in further detail below in conjunction with the accompanying drawings, but the scope of protection of the present application is not limited to the following description.

[0017] The technical scheme of the present application can adopt a dataset constructed based on a ray tracing method for deep neural network training, which is different from existing work in that, first, we utilize the received signal strength information of the user end without directly using the signal source information, which is a more practical solution.Second, in the case of unknown signal source quantity, we use a matrix decomposition method to decompose the multi-signal source superimposed received signal strength collected by the user, and each received signal strength component decomposed corresponds to an independent signal source.Further, for a single signal source scene, first, the received signal strength data is added with noise to simulate the signal strength component after decomposition, and then a training method based on a generative adversarial network (GAN) is designed, UNet is used as a generator, and another CNN neural network model is introduced as a discriminator, and the two networks are collaboratively trained to greatly enhance the performance of the two networks.The performance-enhanced UNet network can realize accurate radio map estimation, and the accuracy difference between the training set data is very small, which can be considered close to the accuracy of the ray tracing method.

[0018] As shown in Figure 6 In an embodiment of the present application, a generative radio map collaborative estimation method for a multi-signal source scene includes the following steps:

[0019] S1. Construct the received signal strength samples of multiple users at multiple frequencies into a matrix form, and decompose using the NMF matrix decomposition method, each received signal strength component decomposed corresponds to an independent signal source;

[0020] The present application constructs the received signal strength samples of multiple users at multiple frequencies into a matrix form, and decomposes using the NMF matrix decomposition method, under the premise of not exceeding the matrix rank, requires selecting an appropriate number of signal sources, so that the sum of the energies of all decomposed received signal strength components accounts for more than 95% of the energy of the original matrix.

[0021] S101. Suppose the user end is in frequency and position for multi-signal source signal reception, wherein the signals received at position and frequency are:

[0022]

[0023] wherein, is the signal superposition of all signal sources at position and frequency , is the signal noise at position and frequency , For the first i A signal source at frequency k The signal on For the first i A signal source to the location j The channel gain coefficient, , , , K Indicates the number of frequencies. This indicates the number of locations the user is in. Indicates the number of signal sources;

[0024] By collecting all frequency and In each position The matrix formed by the received signals is obtained. :

[0025]

[0026] in, , The noise matrix is ​​represented by... frequency and In each position constitute, The matrix representing the channel gain coefficients is composed of... T A signal source to J each position constitute, It is a matrix composed of transmitted signals, consisting of... T A signal source in K At each frequency constitute;

[0027] Construct an NMF problem, and then... Decomposed into two nonnegative matrices and , where NMF refers to nonnegative matrix factorization;

[0028] The NMF problem is formulated as follows:

[0029]

[0030] The NMF problem is a bounded optimization problem with an optimal solution. and Respectively represent and The matrix formed by the channel gain coefficients corresponding to each signal source and the matrix formed by the transmitted signal are given. Since the transmitted signal satisfies an average energy of 1, therefore... Characterizes the channel gain of each signal source at different locations;

[0031] for Decomposition results in Calculate the energy separately, where the i-th decomposition result is... The energy is ;

[0032] By setting a threshold, decomposition results with energy less than the threshold are considered noise, and only decomposition results with energy greater than the threshold are retained. Each decomposition result corresponds to a signal source, thus obtaining the channel gain of each signal source at different locations, which is also the received signal strength of the user at different locations.

[0033] One advantage of this method is that we do not need to consider the number of signal sources. Accurate information. Specifically, through the analysis of... Setting a threshold can separate signal components from noise, removing some components and thus adapting to... This ensures that the final number of retained signals will not exceed the actual number of signal sources.

[0034] Based on the above description, by using the NMF method, this decomposition effectively simplifies the radio map estimation problem in a multi-source environment into a problem in a single-source scenario.

[0035] S2. Construct and train a radio map inference model for a single signal source;

[0036] The overall approach is as follows:

[0037] Training data preprocessing: 1) Geographic environment map is converted into image form as data feature; 2) Received signal strength and location information (RSS information) of multiple users are noise-added and converted into image form as data feature; 3) Radio map obtained based on ray tracing method is used as label to construct training dataset.

[0038] Generator design: The UNet model is used as the generator. The input is a graphical geographic environment map and RSS information, and the output is a graphical radio map estimation result.

[0039] Discriminator design: A convolutional neural network (CNN) model is used to evaluate the “realism” of the radio map output by the generator.

[0040] GAN learning algorithm: Through adversarial training, the generator and discriminator optimize each other. The generator generates higher quality radio maps, and the discriminator improves its ability to distinguish the generated results.

[0041] Model training and inference: During the training phase, the generator and discriminator are jointly optimized; during the inference phase, only the generator is used for radio map estimation.

[0042] In the embodiments of the present application, the specific process is as follows:

[0043] Step 1: A new dataset is constructed using the dataset RadioMapSeer (literature [1]) based on ray tracing method, denoted as , where for each data sample, the data feature vector is , where represents the matrix of user (mobile device) position and received signal strength, that is, the optimal solution of the channel matrix corresponding to the aforementioned signal source decomposition The energy of a certain frequency component in it, for example ; is the matrix representing the physical environment information, and the matrix represents the corresponding radio map, which is calculated by ray tracing method. Assuming that a 256m*256m area is considered, and the radio map resolution is 1m*1m, the dimensions of the above three matrices are all 256*256.

[0044] Step 2: Select a 17-layer UNet network as the generator, denoted as , whose parameter matrix is ; select a 5-layer CNN network as the discriminator, denoted as , whose parameter matrix is . As shown in Figure 2 .

[0045] Step 3: For any data sample , the received signal strength part of the data feature is added with noise, that is, to construct , where the elements in the matrix are all Gaussian random noise with variance , and the matrix only retains random noise elements at positions where is not 0, and the rest of the positions are all zero. After adding noise, the data feature is input to the UNet generator to obtain the "false label" . For any data sample, a noise matrix needs to be randomly generated again in each training round. (The noise adding operation here is to simulate the phenomenon of natural noise after data decomposition in the multi-signal source scenario, which is different from the standard noise-free scheme)

[0046] Step 4: For any data sample , construct the "restricted label" , is a random number between (0, 1).

[0047] ​Step 5: Pair the noisy data features with the true labels and input them into the CNN discriminator to calculate the score .

[0048] Step 6: Pair the noisy data features with "fake labels" and input them into the CNN discriminator to calculate the score .

[0049] Step 7: Pair the noisy data features with "restricted labels" and input them into the CNN discriminator to calculate the score .

[0050] Step 8: According to the above scores and , construct the loss function (Loss Function): for the CNN discriminator, the loss function is ; for the UNet generator, the loss function is ; where , is an adjustable weight coefficient.

[0051] Step 9: Repeat steps 3-8, accumulate the loss function for all samples: ,

[0052] Step 10: According to the above loss function, train the two networks using gradient descent method: , .

[0053] Step 11: Repeat steps 9-10, repeatedly train the two networks until convergence. At this time, the UNet generator can accurately estimate the radio map according to 1) user received signal strength and location and 2) geographic environment map information; and when the geographic environment map information has certain errors, it can automatically correct part of the errors and still ensure high estimation accuracy. The specific effect is shown in Figure 4 , Figure 5 .

[0054] In the embodiments of the present application, the multi-signal source radio map label, and the estimation result after signal sampling, decomposition, collaborative estimation and integration are shown in Figure 1 . Among them, the verification is in the case of 2 signal sources in the environment.

[0055] The idea and structure of using user 1) to estimate the radio map according to user received signal strength and location and 2) geographic environment map information are shown in Figure 2 . The specific training process is shown in Figure 3 . ​​

[0056] S3. For each received signal strength component obtained by decomposing in step S1, construct an input sample together with the geographical environment information, input it into the trained UNet generator to estimate the corresponding single-signal-source radio map, and superimpose to obtain the radio map estimation result of the multi-signal-source scene: the received signal strength at any position is the sum of the received signal strengths of all single-signal-source radio maps at the position.

[0057] In summary, in the scenario where the number of signal sources is unknown, the present application uses matrix decomposition technology to separate the signal sampling components corresponding to each signal source based on multi-point signal sampling. The GAN training architecture and [UNet, CNN] model are used, and in the case of sufficient user received signal strength (RSS) sample quantity, the estimation accuracy of the proposed technical solution is better than that of the existing method. (Signal noise in training: when training in a single-signal-source scene, the signal sampling is added with noise, which can simulate the signal sampling effect after the original signal decomposition in a multi-signal-source scene. After adding the noise setting in the training process, the estimation accuracy of the trained model can be effectively improved. If the noise is not added according to the standard method, the estimation accuracy of the trained model is low. The present application introduces a gradient penalty mechanism to improve the training stability: referring to the WGAN theory, the gradient penalty setting is adopted to avoid the gradient descent or gradient guarantee problem, ensuring that the GAN architecture can be trained smoothly. In addition, the loss function adopts the setting of BCELoss, which is different from the MSELoss in the paper, and has certain improvement effect on the accuracy in actual measurement. The present application also has error correction capability: the GAN training architecture is adopted and the distributed user received signal strength information is used as input. Even if the geographical map is not accurate, the proposed technical solution can still correct errors through the deep features in the user received signal strength (RSS) sample. Based on the deep neural network (i.e. UNet generator), fast reasoning is ensured, which is suitable for real-time applications. Since the signal source information is not used, but the distributed user received signal strength information is used as input, the distributed user received signal strength (RSS) sample and the pre-downloaded high-resolution geographical map are used to reduce the data acquisition cost.

[0058] The above describes the preferred embodiments of the present application. It should be understood that the present application is not limited to the forms disclosed herein, should not be regarded as excluding other embodiments, and can be used in other combinations, modifications and environments, and can be modified by the above teachings or related art or knowledge within the scope of the concepts described herein. Any modification and change made by those skilled in the art without departing from the spirit and scope of the present application shall be within the protection scope of the appended claims of the present application.

Claims

1. A generative radio map cooperative estimation method for multi-signal source scenarios, characterized in that: Includes the following steps: S1. The received signal strength samples from multiple users at multiple frequencies are constructed into a matrix, and the NMF matrix decomposition method is used for decomposition. Each received signal strength component obtained by decomposition corresponds to an independent signal source. S2. Construct and train a radio map inference model for a single signal source; S3. For each received signal strength component obtained in step S1, estimate the corresponding single-source radio map using the single-source radio map inference model, and then overlay them to obtain the radio map estimation result for the multi-source scenario. Step S2 includes: S201. Using the received signal strength information and geographical environment information of multiple users as data features, and radio maps as labels, a training dataset is constructed, where the RSS information includes location information: The training dataset is denoted as The training dataset contains L data samples. For the Lth... One data sample: Data feature vector is ,in A matrix representing the received signal strength information and RSS information of all users. A matrix representing geographic environmental information, a matrix express The corresponding radio map, i.e., the first The true labels of each data sample, among which ; S202. The UNet model is used as the UNet generator. The input is a visualized geographic environment map and RSS information, and the output is a visualized radio map estimation result. The UNet generator is represented as follows: Its parameter matrix is ; A CNN model is used as the CNN discriminator to evaluate the realism of the radio map output by the generator; the CNN model refers to a convolutional neural network model; the CNN discriminator is represented as... Its parameter matrix is ; S203. Based on the constructed training dataset, jointly optimize the UNet generator and the CNN discriminator.

2. The generative radio map cooperative estimation method for multi-signal source scenarios according to claim 1, characterized in that: Step S1 includes: S101. Assume users are in frequency and Signal reception from multiple signal sources is performed at various locations, including at location [location name missing]. and frequency The signal received at the location for: ， in, For all signal sources at position and frequency The signal superposition, For position and frequency Signal noise, Let i be the signal of the i-th signal source at frequency k. Let be the channel gain coefficient from the i-th signal source to position j. , , K represents the number of frequencies. This indicates the number of locations the user is in. Indicates the number of signal sources; By collecting all frequency and In each position The matrix formed by the received signals is obtained. : ， in, , The noise matrix is ​​represented by... frequency and In each position constitute, This represents a matrix consisting of channel gain coefficients, from T signal sources to J locations. constitute, It is a matrix composed of transmitted signals from T signal sources at K frequencies. constitute; Construct an NMF problem, and then... Decomposed into two nonnegative matrices and , where NMF refers to nonnegative matrix factorization; The NMF problem is formulated as follows: . The NMF problem is a bounded optimization problem with an optimal solution. and Respectively represent and The matrix formed by the channel gain coefficients corresponding to each signal source and the matrix formed by the transmitted signal are given. Since the transmitted signal satisfies an average energy of 1, therefore... Characterizes the channel gain of each signal source at different locations; for Decomposition results Calculate the energy separately, where the i-th decomposition result is... The energy is ; A threshold is set, and the decomposition results with energy less than the threshold are regarded as noise. Only the decomposition results with energy greater than the threshold are retained. Each decomposition result corresponds to a signal source, which means that the channel gain of each signal source at different locations is obtained, that is, the received signal strength of the user at different locations.

3. The generative radio map cooperative estimation method for multi-signal source scenarios according to claim 1, characterized in that: Step S203 includes: A1. For any data sample Data features The received signal strength portion is noise-added, i.e., a noise is constructed. , where the matrix Each element in the set has a variance of . Gaussian random noise, while the matrix Only Random noise elements are retained at positions with non-zero values, and all other positions are set to zero. A2. After adding noise, the data features The input is fed into the UNet generator to calculate the "fake labels". For any data sample, a noise matrix needs to be randomly regenerated in each training round. A3. For any data sample Construct "restriction labels" , A random number between (0, 1); A4. Pair the noisy data features with the real labels. The data is then fed into the CNN discriminator to calculate a score. ; A5. Pair the features of the noisy data with "false labels" as The data is then fed into the CNN discriminator to calculate a score. ; A6. Pair the noisy data features with the "restriction labels" as follows: The data is then fed into the CNN discriminator to calculate a score. ; A7. Based on the above scores , , Constructing the loss function: For the CNN discriminator, the loss function is: ； For the UNet generator, the loss function is: ； in , These are the weighting coefficients; A8. For all samples, repeat steps A1~A7, cumulative loss function: , ; A9. Training the UNet generator and CNN discriminator using gradient descent: ， ； A10. Repeat steps S203~S211 to perform the next round of training until convergence or the set number of training rounds is reached. Convergence means that the loss function is less than a given threshold on the evaluation dataset.

4. The generative radio map cooperative estimation method for multi-signal source scenarios according to claim 1, characterized in that: Step S3 includes: S301. For each received signal strength component obtained by decomposition in step S1, construct an input sample together with the geographic environment information, and input it into the trained UNet generator to estimate the corresponding single-signal source radio map. S302. Overlay to obtain the radio map estimation result for a multi-signal source scenario: For a single frequency, the received signal strength at any location is the sum of the received signal strengths of all single-signal source radio maps at that location.

Images