Downlink channel knowledge map construction method and channel estimation method based on diffusion model

Abstract

The application discloses a downlink channel knowledge map (CKM) construction method and a channel estimation method based on a diffusion model, and belongs to the technical field of wireless communication. The method divides the service area of a base station into a grid, combines pilot data and position tags of a receiving device to construct an offline training data set, and uses a diffusion model to learn channel statistical information of different positions to generate a CKM. In the channel estimation stage, based on the CKM and a small amount of pilot data, noise is gradually removed through the reverse process of the diffusion model to realize efficient channel estimation. Compared with traditional methods, the application significantly reduces the number of pilot requirements (only 1-2 pilots are needed), improves the estimation accuracy (the NMSE is reduced by more than 40%), and is suitable for high-density MIMO systems and is applicable to 6G environment perception communication scenarios.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of channel estimation technology, in particular to a downlink channel knowledge map construction method and a channel estimation method based on a diffusion model. BACKGROUND

[0002] In recent years, various emerging applications, such as autonomous driving, networked robots, virtual reality, have higher and higher requirements for the latency and throughput of wireless communication systems, and the fifth generation mobile communication technology (5G) has been unable to meet all the needs. The application of millimeter wave and large-scale MIMO provides a possibility to solve these problems. Due to the increasing density of access devices and the increasing size of antennas, the energy consumption and cost of radio frequency modules cannot be ignored.

[0003] The above new challenges and opportunities bring a promising paradigm shift from traditional environment-unaware communication to new environment-aware communication, which fully utilizes prior knowledge of the actual local environment to design and optimize the communication network. Environment-aware communication can be seen as a further development of traditional location-aware communication, because it not only utilizes location information, but also utilizes local actual environment information where communication occurs. On the one hand, environment-aware communication is a promising technology to solve the new challenges faced by the sixth generation mobile communication technology (6G), such as the difficulty of obtaining real-time CSI (Channel State Information) as the network density and channel size increase. On the other hand, due to the accessibility of high-quality location information, the rich source of location-specific data, and the stronger data mining capability to understand the environment, environment-aware communication has become more feasible than ever.

[0004] CKM (Channel Knowledge Map) in environment-aware communication has recently received extensive attention. Unlike a physical environment map, CKM aims to directly reflect the inherent wireless channel properties of the local environment. Specifically, CKM is a database indexed by locations, marked with the locations of transmitters and / or receivers, which provides channel knowledge for specific locations, helps to improve environment awareness, and promotes or even avoids complex real-time CSI acquisition. The abstract term "knowledge" is deliberately chosen to emphasize that any information related to the local wireless channel can be the expected query output of CKM, including but not limited to channel modeling parameters for specific areas, such as path loss, delay, Doppler, and angular spread, etc.; and channel knowledge for specific locations, such as channel gains, delays, Dopplers, angle of arrivals (AoAs) and angle of departures (AoDs) of multi-paths, etc., and even the impulse response of the entire channel.

[0005] CKM only needs to be calculated offline according to the channel information obtained in advance within a certain range. Once the CKM is constructed, the communication between the user and the base station only needs a few pilots to obtain relatively accurate CSI. In fact, with the increasing computing resources, both the user and the base station can calculate the local CKM to help the uplink or downlink communication. The traditional CKM construction usually obtains some channel parameters at the current location using some channel estimation method, and stores the channel information of each path locally, which will consume a lot of storage space. Secondly, different channel estimation methods do not completely require the same information in the process of using CKM, and there may be incompatibility between different CKM construction and use methods.

[0006] With the development of artificial intelligence (AI), convolutional neural networks, adversarial neural networks and Transformer network architectures have had a major impact in different fields. In recent years, diffusion models have been continuously developed in the field of image generation and have become one of the mainstream models in the field of image generation. In addition, diffusion models can also solve the problem of solving linear equations with noise in a statistical sense, so they can be applied to channel estimation problems. During the training of the diffusion model, the position information of the channel is added as a condition to control the sampling process of the diffusion model, and only a small number of pilots are needed to recover the channel coefficients that meet the requirements of channel estimation in the inference stage. SUMMARY

[0007] The present application aims to provide a downlink channel knowledge map construction method and channel estimation method based on a diffusion model. By adding position information as a label during model training, the model can learn the channel statistics at different positions and use it as a CKM to help complete channel estimation. After the CKM is constructed, the diffusion model can generate position-based channel prior information in the constructed CKM and complete channel estimation with the help of a few pilots.

[0008] To solve the above technical problems, the specific technical solutions of the present application are as follows:

[0009] A downlink channel knowledge map construction method based on a diffusion model, comprising the following steps:

[0010] Step S1, dividing the base station service area into an n x n grid, receiving equipment moving in the grid and collecting pilot data to generate an original data set with position labels;

[0011] Step S2: based on the data set of step S1, using an offline channel estimation algorithm to obtain the channel parameters of each position, and constructing training data;

[0012] Step S3: build a diffusion model, the forward process of which adds noise step by step according to a Markov chain, and the backward process of which predicts noise through a U-Net network, and the loss function is:

[0013]

[0014] wherein β t is a preset noise parameter, x0 is original input data, ∈ is a standard Gaussian process random sampling, ∈ θ is an output function of the U-Net network;

[0015] Step S4: after training, the diffusion model generates a position-based CKM, and stores channel statistical information.

[0016] Preferably, the grid division density in step S1 satisfies n≥50, and the pilot data includes multipath delay, angle of arrival (AoA) and complex channel coefficient.

[0017] Preferably, the input of the U-Net network is a two-channel matrix obtained by splicing the real part and the imaginary part of the complex channel data, and the position code and the diffusion step number are embedded.

[0018] The application also provides a channel estimation method based on the CKM, comprising:

[0019] Step S5: a user receives pilot data Y=HX+N, extracts a current position p, wherein H is a channel matrix, N is additive noise, X is a pilot matrix, and Y is a received signal;

[0020] Step S6: input p into the CKM to obtain prior channel information and initialize an estimated channel;

[0021] Step S7: in the diffusion backward process, the following is iteratively executed:

[0022]

[0023] wherein, H t is an estimated channel obtained in the tth iteration, ∈ θ represents a noise prediction network trained in advance, α t , β t represent preset noise parameters, is a cumulative noise attenuation coefficient, α t =1-β t ;

[0024] Step S8: combine the likelihood gradient to correct the estimation, and finally output an estimated channel.

[0025] Preferably, the likelihood gradient calculation is as follows:

[0026]

[0027] σ 2 is the noise variance, I is the identity matrix.

[0028] Preferably, the number of iterations T in step S7 is less than or equal to 100, and the final NMSE is less than -20 dB.

[0029] Preferably, the method supports a large-scale MIMO system, the number of base station antennas N t ≥ 32, and the number of users K ≥ 16.

[0030] Preferably, the diffusion model uses an Adam optimizer during training, with a learning rate lr = 10 -4 and a batch size batch = 64.

[0031] Preferably, the CKM is adapted to the millimeter wave frequency band (28 GHz and above), and supports a dynamic update mechanism that periodically fuses newly collected channel data.

[0032] Preferably, the method is integrated into a 6G base station device.

[0033] Beneficial effects:

[0034] 1. The diffusion model is applied to the construction of a downlink channel knowledge map (CKM). By dividing the base station service area into a grid, combining pilot data collected by receiving devices and location tags to construct an offline training data set, and letting the diffusion model learn the channel statistical information at different locations to generate a CKM.

[0035] Compared to traditional CKM construction methods that consume a large amount of storage space, this method uses the diffusion model to learn the characteristics of data distribution, more efficiently stores and represents channel statistical information. At the same time, the introduction of the diffusion model makes the CKM construction and use more compatible, avoiding the incompatibility problem between different methods.

[0036] 2. In the channel estimation stage, based on the constructed CKM and a small amount of pilot data (only 1-2 pilots), the diffusion model is used to gradually remove noise and achieve channel estimation.

[0037] This method significantly reduces the dependence on the number of pilots, reduces pilot data transmission, improves data transmission rate, and alleviates the problem of data transmission rate being affected by the increase in the number of pilots, especially suitable for 6G communication scenarios with extremely high data transmission rate requirements.

[0038] 3、Through the cooperation of the diffusion model and CKM, the normalized mean square error (NMSE) of channel estimation is reduced by more than 40%. In complex communication environments such as 6G macro cell scenarios, more accurate channel estimation can be achieved, effectively improving communication quality and reliability, ensuring stable signal transmission, and reducing data transmission errors and losses.

[0039] 4、The method supports large-scale MIMO systems, with the number of base station antennas Nt≥32 and the number of users K≥16. It meets the communication needs as the access device density and antenna size continue to increase, adapts to the development trend of future communication networks, and provides an effective channel estimation solution for the application of high-density MIMO systems in 6G and subsequent communication technologies.

[0040] 5、CKM is adapted to millimeter wave frequency bands (28GHz and above), and supports a dynamic update mechanism that can periodically integrate newly collected channel data. It meets the application needs of 6G in millimeter wave frequency bands, and the dynamic update mechanism can make CKM adapt to changes in the communication environment, maintain accurate representation of channel information, and further improve the accuracy and adaptability of channel estimation.

[0041] 6、The channel estimation method is integrated into 6G base station equipment, making it easy to apply directly in actual 6G communication systems and accelerating the transformation of the technology from theoretical research to practical deployment, promoting the development and application of 6G communication technology. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 Schematic diagram of the area served by the base station.

[0043] Figure 2 Schematic diagram of the U-Net structure.

[0044] Figure 3 Channel estimation process diagram.

[0045] Figure 4 NMSE trend chart with increasing number of pilots.

[0046] Figure 5 NMSE trend chart with SNR. DETAILED DESCRIPTION

[0047] In the 6G macro cell scenario, due to the presence of scatterers such as buildings and trees, there are many NLos paths. In the case of a large number of antennas, real-time and accurate channel estimation is a challenging problem. Although the base station-to-user channel changes in real time, the scatterers on the signal propagation path are usually relatively stable in size and relatively stable in position, bringing the characteristic of relatively stable statistical information of the channel. Therefore, a channel knowledge map can be established in advance to assist in channel estimation using relatively stable statistical information.

[0048] As shown in Figure 1 , it is assumed that G areas are divided in the area served by the base station, and each area is expressed, where and the base station side can know that the served user is located at a certain position, so that more accurate estimation can be achieved.

[0049] The application discloses a downlink channel knowledge map (CKM) construction method and a channel estimation method based on a diffusion model, which comprises the following steps:

[0050] I. Offline CKM construction

[0051] Step 1: The service range of the base station is divided into a plurality of grid points. In order to obtain the information required for constructing the CKM, the receiving device is moved randomly in the service range of the base station in advance, and the pilot data transmitted by the base station is received at different positions and is labeled with position tags.

[0052] Step 2: The pilot data, received signals and position information in step 1 are used to construct an offline training data set. Since the offline computing power is sufficient, a channel estimation method with high complexity can be used to obtain estimated channel data.

[0053] Step 3: A diffusion model is built, and the constructed data set is used for training. In the training process, noise is randomly added to the forward process of the tth step, and the network prediction noise is used in the backward process. The network convergence process means that the construction of the CKM is completed.

[0054] II. Real-time application of CKM

[0055] Step 1: The user receives the pilot data transmitted by the base station at any position in the service range of the base station, and the user side knows the pilot data and the current position.

[0056] Step 2: In each step of the backward process, the estimated channel obtained in the last step, the tth step and the current position are input into the network, and the network outputs the prediction noise of the current step.

[0057] Step 3: The prediction noise output by the network is combined with the gradient calculated from the received data and the pilot to obtain the estimated channel of the current step, until the t reaches the preset step number, and the final estimated channel is obtained.

[0058] The application will be further explained and described in combination with the embodiments.

[0059] Embodiment 1: Offline CKM construction

[0060] The diffusion model includes a forward process and a backward process. The forward process can be seen as a Markovian process with noise added step by step, which is represented as follows:

[0061]

[0062] where 0 < β1< β2< … < β T <1 is a preset noise parameter, which is used to add noise to the original data step by step to make Therefore, according to the properties of Gaussian distribution and Markov process, the distribution can be obtained by one step:

[0063]

[0064] where α t = 1-β t Therefore, the distribution of x t in the forward process satisfies:

[0065]

[0066] where

[0067] If the backward process is defined as a parameterized variational Markov chain, we can get:

[0068]

[0069] where ∈ θ represents a noise prediction network trained in advance, which can fit the noise of each step according to the input. According to the following equation:

[0070]

[0071] and Jensen's inequality, we get:

[0072]

[0073] So the minimum of the evidence lower bound (ELBO) is:

[0074]

[0075] The right side of the above equation can be simplified as:

[0076]

[0077] where the first term is independent of the network, and the third term is only related to the one-step noise process, so we focus on the second term.

[0078] According to Bayes' theorem where

[0079]

[0080] Thus we can get

[0081]

[0082] The mean and variance of which are

[0083] If we assume and the following is true:

[0084]

[0085] The ELBO lower bound can be further simplified as:

[0086]

[0087] To further simplify, we use x t to express:

[0088]

[0089] Therefore θ (x t , t) must have a similar form, assuming

[0090]

[0091] Bring both means into the ELBO derived earlier:

[0092]

[0093] At this point, the loss function for training the diffusion model has been derived, and it can be seen that the diffusion model is learning the distribution of the data x0. Therefore, for the channel estimation problem, we can use the channel information of multiple states as the data set to learn the statistical information of the current channel. The network used to predict noise in the diffusion model is usually a U-Net structure. U-Net can better preserve data information than traditional autoencoder structures. The U-Net structure adopted in this experiment is shown in Figure 2 The real and imaginary parts of the complex channel data are combined together as two-channel images. After convolution, down-sampling and self-attention modules, the data dimensions are as shown in the figure. The position information g and the diffusion model step t are combined together and fed into the network data.

[0094] For constructing CKM, channel information of different locations needs to be collected and data sets with location information are created for training. At this time, there are G data sets H g For model training, the diffusion model will learn the statistical information p of different data sets according to the label g θ (H g ), so the construction of CKM is completed.

[0095] Embodiment 2: Method for using CKM

[0096] In a mobile communication downlink transmission system, as the user moves, the environment around the user is constantly changing, which causes the scatterer between the communication links to change, which has a great influence on the channel model parameters such as multipath time delay and Doppler frequency shift, so in general, the base station needs to send pilot data to estimate the channel data in real time and ensure the normal progress of the communication process. However, the increase of pilot data will significantly affect the data transmission rate, so a channel estimation method based on CKM of diffusion model is proposed.

[0097] Consider a single-cell downlink FDD system, the base station side is equipped with a uniform linear array of N t antennas, so the frequency domain response between the base station and each user is:

[0098]

[0099] Where ξ l represents a standard complex Gaussian coefficient, ρ l represents the complex gain of the lth sub-path, φ l represents the angle of arrival, a(φ l ) represents the frequency domain array response vector, which has the following form:

[0100]

[0101] When there are N r users in a certain area receiving N T pilots, the transmitted data can be expressed as The received signal can be expressed as

[0102] Y=HX+N

[0103] Where N represents a zero-mean additive complex Gaussian noise with variance σ 2 I, the channel matrix can be expressed as Assuming that the DFT matrix is transmitted, it satisfies tr(X H X)=N t N T .

[0104] According to the above model, the channel estimation problem can be described as estimating the channel matrix at the current time according to the received signal. According to the maximum posteriori criterion, the following problem needs to be solved:

[0105] max H p(H|Y)∝p(H)p(Y|H)

[0106] The step of the diffusion model reverse process is to subtract the prediction noise from a Gaussian noise sample continuously, so the noise of each step can be regarded as a gradient, i.e.:

[0107]

[0108] In the problem with posterior information, the above gradient should be changed to According to the Bayes formula, we can get:

[0109]

[0110] Through the diffusion model, the distribution can be learned Therefore, the calculation of the above formula is still missing the likelihood probability of the second term, because p θ (Y|H t )=∫p θ (Y|H0)p θ (H0|H t )dH0, wherein p θ (H0|H t ) cannot be accurately calculated, so it is assumed that Therefore, it satisfies:

[0111]

[0112] According to the CKM constructed before, the statistical channel information of each position has been obtained, i.e. p(H) in the above formula, so the following formula is true:

[0113]

[0114] wherein represents a standard complex Gaussian distribution, and the above formula is brought into the formula of the received signal to obtain

[0115]

[0116] Therefore, the likelihood probability can be calculated as

[0117]

[0118] Therefore, the gradient of the likelihood probability is

[0119]

[0120] According to the assumptions in the CKM construction procedure:

[0121]

[0122] Substitute the posterior probability gradient with the following:

[0123]

[0124] Thus the complete channel estimation procedure is obtained, as shown in Figure 3 .

[0125] Simulation results:

[0126] In this downlink channel estimation, the minimum mean square error (MMSE), sparse Bayesian learning (SBL), and orthogonal matching pursuit algorithm (OMP) are selected as comparison methods. The base station is equipped with a uniform linear array of 32 antennas, and serves 32 users in a region. The number of multipaths is 21. In a square region of 50m*50m, the region is divided into d segments in the horizontal and vertical directions, and there are total regions. The pilot is sent as the column of the DFT matrix, and the standard normalized mean square error (NMSE) is used as the evaluation index. The trend of NMSE with the increase of the number of pilots is shown in Figure 4 . First, it can be seen that with the increasing of the region division density , the NMSE becomes smaller and smaller. Compared with the method without channel prior information, the method based on the diffusion model can achieve good performance with only 1-2 pilots. Secondly, from the trend of the curve, it can be seen that the channel estimation assisted by the diffusion model CKM is close to convergence with 4 pilots, and the performance is better than that of SBL with 32 pilots. In addition Figure 5 , the trend of NMSE with the change of SNR is shown, and it can also be seen that with the increasing of the division density , the NMSE becomes smaller and smaller, representing that the estimation performance is better and better. Compared with the estimation method without CKM assistance, the estimation accuracy is obviously higher.​

Claims

1. A method for constructing a downlink channel knowledge map based on a diffusion model, characterized in that: The steps include: Step S1: Divide the base station service area into an n×n grid. The receiving device moves within the grid and collects pilot data to generate an original data set with location tags. Step S2: Based on the data set in step S1, an offline channel estimation algorithm is used to obtain the channel parameters of each position and construct training data; Step S3: Build a diffusion model. The forward process gradually adds noise according to the Markov chain, and the reverse process predicts the noise through the U-Net network. The loss function is: in α t , β t is the noise parameter set in advance, is the cumulative noise attenuation coefficient, α t =1-β t , x0 is the original input data, ∈ is the standard Gaussian process random sampling, ∈ θ is the output function of the U-Net network, and t represents the tth iteration; Step S4: After training is completed, the diffusion model generates a location-based CKM and stores channel statistics.

2. The method for constructing a downlink channel knowledge map based on a diffusion model according to claim 1, characterized in that: In step S1, the grid division density satisfies n≥50, and the pilot data includes multipath delay, arrival angle and complex channel coefficient.

3. The method for constructing a downlink channel knowledge map based on a diffusion model according to claim 1, characterized in that: The U-Net network input is a two-channel matrix of the real and imaginary parts of the complex channel data, and is embedded with position encoding and diffusion steps.

4. A channel estimation method based on any one of claims 1 to 3, characterized in that: include: Step S5: The user receives pilot data Y=HX+N and extracts the current position p, where H is the channel matrix, N is the additive noise, X is the pilot matrix, and Y is the received signal; Step S6: Input p into CKM to obtain prior channel information and initialize the estimated channel; Step S7: In the diffusion reverse process, iteratively execute: Among them, H t is the estimated channel obtained at the tth iteration; Step S8: Combine likelihood gradients Correct the estimate and finally output the estimated channel.

5. The method according to claim 4, characterized in that The likelihood gradient is calculated as: σ 2 is the noise variance, and I is the identity matrix.

6. The method according to claim 4, characterized in that The number of iterations T in step S7 is ≤ 100, and the final NMSE is lower than -20 dB.

7. The method according to claim 4, characterized in that The method supports large-scale MIMO systems with N base station antennas. t ≥32, the number of users K ≥16.

8. The method according to claim 1, characterized in that The diffusion model is trained using the Adam optimizer with a learning rate of lr = 10 -4 , batch size batch = 64.

9. The method according to claim 1, characterized in that The CKM is adapted to the millimeter wave frequency band and supports a dynamic update mechanism to periodically integrate newly collected channel data.

10. The method according to claim 4, characterized in that The method is integrated into 6G base station equipment.

Images