A low pilot overhead channel estimation method based on user-side channel knowledge map

Abstract

The present invention discloses a low pilot overhead channel estimation method based on a user-side channel knowledge map, which belongs to the field of channel estimation. The method includes: the user queries the user-side channel knowledge map locally according to its own location information; the user negotiates with the base station on the number of pilots used for the current communication based on the channel knowledge map; based on the large-scale downlink wireless millimeter wave channel model, the transmission azimuth angle domain is gridded and a suitable alternative function is found to replace the maximum likelihood function; the noise coefficient, gain matrix variance and angle offset are repeatedly iteratively optimized and solved using the prior information provided by the user-side channel knowledge map; the gain channel matrix is ​​accurately estimated based on the prior information provided by the locally queried user-side channel knowledge map and the received pilot data. The actual number of pilots required for transmission and reception in the present invention is relatively small, and the channel estimation result is comparable to the accuracy of traditional channel estimation.

Description

Technical Field

[0001] The present invention relates to the technical field of channel estimation, and in particular to a low pilot overhead channel estimation method based on a user-side channel knowledge map. Background Art

[0002] With the rapid development of the communications industry and mobile communications, the lower frequencies of the radio spectrum are becoming saturated. Even with the adoption of various multiple access technologies to expand the capacity of communication systems and improve spectrum utilization, they are unable to meet the needs of future communications development. Therefore, achieving high-speed, broadband wireless communications inevitably requires the development of new spectrum resources in the high-frequency microwave band. Millimeter waves, with their short wavelength and wide bandwidth, can effectively address many of the challenges facing high-speed, broadband wireless access, and therefore have broad application prospects in short-range wireless communications. Millimeter wave (mmWave) communication refers to a communication method that uses millimeter waves as a communication carrier. Millimeter waves refer to radio frequency bands with wavelengths between 1 and 10 mm, corresponding to a frequency range of 30 to 300 GHz. Compared with traditional communication frequency bands, millimeter waves have higher frequencies and greater bandwidths, enabling faster data transmission and more stable network connections.

[0003] With the rapid development of wireless positioning and sensing technologies, leveraging user location information, and even geolocation-based databases, in wireless communication systems has received increasing attention. In particular, the concept of a Channel Knowledge Map (CKM) has been proposed in recent years. Its goal is to enable context-aware communication by providing location-specific information about the inherent radio channel (rather than the traditional coarse, site-specific information), thereby facilitating or even avoiding complex real-time Camera Serial Interface (CSI) acquisition. Therefore, CKM-based context-aware communication is expected to achieve significant performance improvements compared to general context-aware communication that relies solely on real-time channel training, especially when real-time channel training is expensive or even impossible. The hardware cost, power consumption, and signal processing complexity of millimeter-wave massive Multiple-Input Multiple-Output (MIMO) systems make traditional all-digital baseband processing (such as transmitter / receiver precoding / combining) challenging, requiring each antenna element to be connected to an expensive radio frequency (RF) chain. Therefore, reducing pilot overhead in channel estimation is particularly important.

[0004] The current CKM generally refers to the CKM that only uses the base station side, that is, all users share the CKM in the same channel environment. However, due to the wide range of individual user RF channel performance, especially in multi-antenna scenarios, changes in the pitch angle of the receiving antenna will have different impacts on the actual channel estimation. Aggregating the channels of each user on the base station side to form a CKM that is common to each user often requires a complex calibration mechanism, which will undoubtedly greatly increase the cost. It is advocated that each user collects his own channel information and forms his own user-side CKM with the user's location, and the user-side CKMs are not shared with each other and will not affect each other. This method of using the user's own CKM information to assist in downlink user channel estimation can undoubtedly save more pilot overhead.

[0005] Sparse Bayesian Learning (SBL) is a method for processing high-dimensional data within the Bayesian statistical framework. High-dimensional data often contain a large number of features, but only a small fraction of them are truly useful for solving the problem. Sparse Bayesian learning aims to improve the generalization, interpretability, and efficiency of models by introducing sparsity, that is, selecting only a few important features. In traditional Bayesian statistics, the posterior distribution of a parameter is calculated using the Bayesian formula, which contains the probability distribution of all possible parameter values. However, calculating such a distribution in high-dimensional data is very difficult and can easily lead to overfitting. Sparse Bayesian learning introduces the sparsity assumption in the prior distribution, ensuring that only a subset of parameters are non-zero for the data generation. This reduces the search space for the parameter space and improves model efficiency. Sparse Bayesian methods have applications in fields such as feature selection, dimensionality reduction, and signal recovery. They can be used to build efficient models, reduce overfitting, and in some cases provide information about feature importance within the model. The SBL model requires only a small number of parameters, significantly improving model training efficiency and resulting in more reliable and stable results (robustness). Furthermore, the weights estimated by the SBL model can measure the importance of a feature in regression, similar to the significance of parameters in regression estimation, improving the interpretability of each feature. Due to its excellent classification and regression capabilities, SBL has been widely applied in many research fields. Furthermore, SBL has been demonstrated to have excellent performance in sparse signal recovery, sparse representation, and compressed sensing. Summary of the Invention

[0006] The present invention provides a low-pilot-overhead channel estimation method based on a user-side channel knowledge map. The user queries the local user-side channel knowledge map according to his or her own location, negotiates the number of pilots with the base station, and puts forward communication requirements. Based on the prior information stored in this map and a sparse Bayesian learning algorithm, an accurate estimation of the channel at the user's actual location is obtained while greatly reducing the overhead of sending and receiving pilots.

[0007] An embodiment of the present invention provides a low pilot overhead channel estimation method based on a user-side channel knowledge map, comprising the following steps:

[0008] Step 1: Query the local user side channel knowledge map based on the user's actual location;

[0009] Step 2: Negotiate with the base station based on the user's actual location to obtain the required number of pilot signals for transmission and reception;

[0010] Step 3: Divide the emission angle domain into multiple parts evenly and construct a surrogate function to replace the maximum likelihood function, thereby changing the optimization objective to the constructed surrogate function;

[0011] Step 4: Based on the actual user location and using the prior information provided by the user-side channel knowledge map, the noise coefficient, gain matrix variance, and angle offset are iteratively optimized and solved.

[0012] Step 5: Accurately estimate the gain matrix based on the prior information provided by the local user-side channel knowledge map, and use the transmitted and received pilot signals to estimate the channel information at the user's location.

[0013] Optionally, in one embodiment of the present invention, in step 1, a normalized narrowband millimeter wave channel having N transmitting antennas, a single receiving antenna, and L scattering paths is modeled as a classic Saleh-Valenzuela model, and the channel matrix of the modeled downlink millimeter wave channel is:

[0014]

[0015] Among them, α l Represents the fading coefficient of each path, that is, the path loss and phase shift of the lth path, a r (θ l ) represents the direction vector of the receiving antenna of the lth path. Assuming that both the transmitting and receiving antennas are uniform linear arrays, then: is the array vector of the N-dimensional antenna array, in the downlink millimeter wave channel θ l Represents the departure angle of the lth path, modeled as obeying After the downlink millimeter wave channel model is established, the user queries the local user-side channel knowledge map according to his actual location to obtain the channel parameters of the user at each location through sparse Bayesian learning.

[0016] Optionally, in one embodiment of the present invention, in step 2, assuming that the antenna array of the base station includes N antennas, the user negotiates with the base station in the uplink channel and sends S pilots X. The rule for obtaining the number of pilots S is as follows: when the user position is within the position range of the user-side channel knowledge map queried locally, the corresponding table is searched according to L in the user-side channel knowledge map to obtain the number of sent pilots S, and the table is obtained based on actual measurement or channel environment simulation; when the user position is not within the position range of the local user-side channel knowledge map, the 5G / 4G default pilot number is used to obtain the received signal y=Xh+n at each user end, where n represents complex Gaussian white noise with a variance of σ 2 , h = A(β)w is the channel matrix to be estimated.

[0017] Optionally, in one embodiment of the present invention, step 2 specifically includes:

[0018] The actual launch azimuth angle AoD is expressed as: {θ l ,l=1,2,...,L}, where L=NcNs represents the total number of paths, Nc and Ns represent the number of scatterers and the number of paths contained in each scatterer respectively, each path corresponds to an AoD, so there are a total of L=NcNs AoDs to be estimated; the angle domain Divide evenly The grid angles after gridding are recorded as When the true angle θ l Just fell here When the channel model is written as h = Aw, a(θ) is the array vector, Is a sparse vector, the non-zero elements represent the real vector at the non-zero element on the grid vector, that is, if the first elements are non-zero and the corresponding real vector is θ l , then at this time there is An off-grid model is used to model the situation where the true angle is not exactly on the grid: if and is the distance θ l The nearest grid point is written as Represents the angle offset, that is, the offset of the true value relative to the nearest grid, that is:

[0019]

[0020] The array vector at this time is written as Correspondingly, the channel matrix h at this time is h = A(β)w, where The signal received by the receiving end is expressed as y=XA(β)w+n=Φ(β)w+n, where

[0021] Optionally, in one embodiment of the present invention, in step 3, the emission angle domain Divide evenly and construct an alternative function To replace the maximum likelihood function lnp(y,α,γ,β), thereby changing the optimization objective to an alternative function Specifically include:

[0022] The construction method of the substitution function is as follows: Under the assumption of complex Gaussian noise, where α = σ -2 Represents noise, which is usually unknown. The noise is modeled as a hyper-prior probability of the Gamma distribution, that is, p(α) = Γ(α; 1 + x, y). The general form of the Gamma distribution is Γ(α|x, y) = Γ(x) -1 y x α x-1 e -yα ,in represents the Gamma function, and substituting the super prior of α into the above formula, we get p(α)=Γ(1+x) -1 y 1+x α x e -yα , so far the prior probability of α is modeled; the prior probability of β is assumed to obey a uniform distribution without information: According to the user side channel knowledge map obtained locally and the sparse Bayesian learning method, each element in w is modeled as obeying the mean The complex Gaussian distribution of , with variance represented by γ:

[0023]

[0024] where γ=[γ1,γ2,...,γ L ] T , each element γ in γ i Modeled as independent Gamma distribution, that is The optimal solution to the estimation problem is as follows:

[0025]

[0026] is equivalent to:

[0027]

[0028] Iteratively construct a continuous surrogate function for the objective function lnp(y,α,γ,β), and then alternately maximize the constructed surrogate function with respect to α, γ, and β. The choice of the surrogate function is to ensure that each variable in the alternating maximization has a closed and simple solution. Specifically, let As a fixed point The constructed substitution function satisfies the following properties:

[0029]

[0030]

[0031] After constructing the substitution function, update α, γ, and β using the following formula:

[0032]

[0033] in(·) (i) Indicates the i-th iteration. The overall process of the iterative algorithm is to initialize α (0) , γ (0) and β (0) After that, the three variables α, γ and β are iteratively updated in sequence through the above three formulas until convergence. According to the idea of ​​Expectation Maximization algorithm, the corresponding lower bound function is selected as the replacement function ——For any fixed point Construct the replacement function as:

[0034]

[0035] Optionally, in one embodiment of the present invention, in step 4, the noise coefficient α, the gain matrix variance γ, and the angle offset β are iteratively optimized and solved according to the actual position of the user and using prior information provided by the user-side channel knowledge map, specifically including:

[0036] Since p(w|y,α,γ,β) satisfies the complex Gaussian distribution:

[0037]

[0038] The mean and variance are:

[0039]

[0040] Σ(α,γ,β)=(αΦ H (β)Φ(β)+diag(γ)) -1

[0041] Therefore, there is a closed solution to the maximization problem of α:

[0042]

[0043] in:

[0044]

[0045] The maximization problem of γ has a closed-form solution of the following form:

[0046]

[0047] in,

[0048] Since the maximization problem with respect to β is non-convex and it is difficult to find its optimal solution, we apply gradient update to the objective function of the maximization problem of β and obtain a simple one-step update of β. The derivative of the objective function with respect to β is calculated as:

[0049]

[0050] in,

[0051] and Represent μ(α (i+1) , γ (i+1) , β (i) ) and Σ(α (i+1) ,γ (i+1) ,β (i) )’s (j,l)th element; Obviously, the optimal solution of β is difficult to obtain, but due to the convergence of the iterative algorithm, it is only necessary to find a suboptimal solution so that the value of the objective function gradually increases. The most commonly used numerical method is to update β in the derivative direction. l In order to reduce the computational complexity, we choose to use a fixed step size to update β:

[0052]

[0053] Where, represents the grid interval, sign(·) represents the signum function;

[0054] The above iterations are repeated until convergence to obtain the optimal estimated parameter (α ★ ,γ ★ ,β ★ ).

[0055] Optionally, in one embodiment of the present invention, in step 5, the optimal estimation parameter (α ★ ,γ ★ ,β ★) and then use the following formula to accurately estimate the gain matrix:

[0056]

[0057] Therefore, the channel matrix of the user at the current actual position is h e =A(β ★ )w new .

[0058] The low pilot overhead channel estimation method based on the user-side channel knowledge map according to the embodiment of the present invention has the following beneficial effects:

[0059] 1. This invention uses the user's location to query the local user-side channel knowledge map, combining the user-side channel information with the base station's CKM as the user-side CKM. User-side CKMs are not shared and do not affect each other. Simultaneously, the base station and the user's own radio frequency are utilized, reducing the overhead of transmitting and receiving pilot signals, saving significant costs in actual channel estimation.

[0060] 2. The iterative algorithm of this invention is based on sparse Bayesian learning, which can guarantee convergence to a large extent. At the same time, due to the sparsity of the grid matrix, the algorithm converges quickly, requires fewer iterations, and has a low runtime complexity.

[0061] 3. The final channel estimation result of the method proposed in the present invention has a high accuracy. At the same time, since it utilizes the prior information in the user-side CKM, it has a certain robustness against interference in the actual user channel.

[0062] Additional aspects and advantages of the present invention will be set forth in part in the description which follows and, in part, will be obvious from the description which follows, or may be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0064] Figure 1 A flowchart of a low pilot overhead channel estimation method based on a user-side channel knowledge map according to an embodiment of the present invention;

[0065] Figure 2 A schematic diagram of a channel model established for an embodiment of the present invention;

[0066] Figure 3 This is a flow chart of a low pilot overhead channel estimation method based on a user-side channel knowledge map according to a specific embodiment of the present invention;

[0067] Figure 4This is a convergence diagram of the distance between the channel estimate and the true value as the iteration increases in different channel environments according to an embodiment of the present invention;

[0068] Figure 5 This is a comparison chart of pilot overhead when the user-side CKM is used as prior information and when the user-side CKM is not used in an embodiment of the present invention. DETAILED DESCRIPTION

[0069] The following describes embodiments of the present invention in detail, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and are not to be construed as limiting the present invention.

[0070] Figure 1 The present invention provides a flowchart of a low pilot overhead channel estimation method based on a user-side channel knowledge map according to an embodiment of the present invention.

[0071] like Figure 1 As shown, the low pilot overhead channel estimation method based on the user side channel knowledge map includes the following steps:

[0072] Step 1: Query the local user side channel knowledge map based on the user's actual location.

[0073] In one embodiment of the present invention, a normalized narrowband millimeter wave channel with N transmitting antennas, a single receiving antenna, and L scattering paths is modeled as a classic Saleh-Valenzuela model (taking a linear uniform array as an example). The user's current actual location is obtained based on positioning technologies such as GPS and the user-side channel knowledge map is queried. In this embodiment of the present invention, the L, α, γ, β,

[0074] In step 1, taking the downlink linear antenna array millimeter wave channel as an example, the channel matrix modeled is:

[0075]

[0076] where α l Represents the fading coefficient of each path, that is, the path loss and phase shift of the lth path. r (θ l ) represents the direction vector of the receiving antenna at the lth path. Assuming that both the transmitting and receiving antennas are uniform linear arrays (ULA), then

[0077]

[0078] is the array response of the N-dimensional antenna array. In the downlink millimeter wave channel, θ l Represents the Angle of Departure (AoD) of the lth path, which is generally modeled as obeying After the channel model is established, the user queries the local user-side channel knowledge map based on their own location. That is, sparse Bayesian learning is used to obtain the channel parameters of the user at each location, especially the angle offset β, the gain matrix variance γ, and the gain matrix w.

[0079] Furthermore, the downlink millimeter wave channel modeling of the user-side CKM in step 1 is as follows:

[0080] h=A(θ)w

[0081] in represents an array vector, Represents the gain matrix. Based on the local user-side CKM previously queried from the user's own position, the channel matrix and correlation matrix coefficients of each position in the environment can be obtained, especially the transmission azimuth And the channel gain matrix w. Then the dimension will be constructed based on w in the user side CKM In the actual channel estimation The mean of the distribution The transmission azimuth angle stored in the user-side CKM is used as the initial value of the actual channel estimation iteration to reduce pilot and time overhead.

[0082] Step 2: Negotiate with the base station based on the actual location of the user to obtain the required number of transmitting and receiving pilot frequencies.

[0083] According to the modeled downlink millimeter wave channel, it is assumed that the antenna array ULA of the base station BS contains N antennas. The user negotiates with the base station in the uplink channel and sends S pilots X, where X is a complex matrix with a dimension of S×N. The rules for obtaining the number of pilots S are as follows: When the user position is within the user-side CKM position range queried locally, the corresponding table is searched according to L in the user-side CKM to obtain the number of sent pilots S, and the table is obtained based on actual measurements or channel environment simulation; when the user position is not within the local user-side CKM position range, the 5G / 4G default pilot number is used. For each user, each user terminal obtains the received signal y=Xh+n, where y is a complex matrix with a dimension of S×1. Represents complex Gaussian white noise with mean 0 and variance σ 2 , h = A(β)w is the channel matrix to be estimated.

[0084] In step 2, the actual launch azimuth angle AoD is expressed as: {θ l,l=1,2,...,L}, where L=NcNs represents the total number of paths, Nc and Ns represent the number of scatterers and the number of paths contained in each scatterer respectively. Each path corresponds to an AoD, so there are a total of L=NcNs AoDs to be estimated. Divide evenly The grid angles after gridding are recorded as When the true angle θ l Just fell here When the channel model is written as h = Aw, a(θ) is the array vector mentioned above. Obviously Is a sparse vector, the non-zero elements represent the real vector at that location on the grid vector, that is, if the first elements are non-zero and the corresponding real vector is θ l , then at this time there is This is an ideal situation, but in practice the actual angle is often not exactly on the grid, that is, it does not match the grid, and there will be an offset on the basis of the grid. Therefore, the off-grid model is used to model this more general situation: if and is the distance θ l The nearest grid point is written as Represents the angle offset, that is, the offset of the true value relative to the nearest grid, that is:

[0085]

[0086] The array vector at this time is written as Correspondingly, the channel matrix h at this time is h = A(β)w. The signal received by the receiving end can be expressed as y = XA(β)w+n = Φ(β)w+n, where

[0087] Step 3: Divide the emission angle domain into multiple parts evenly and construct a surrogate function to replace the maximum likelihood function, thereby changing the optimization objective to the constructed surrogate function.

[0088] Specifically, the emission angle domain Divide evenly and construct an alternative function To replace the maximum likelihood function lnp(y,α,γ,β), thereby changing the optimization objective to an alternative function

[0089] The construction method of the substitution function in step 3 is as follows: Under the assumption of complex Gaussian noise, we have where α = σ-2 represents noise. Since it is usually unknown, it is modeled as a hyper-prior probability of the Gamma distribution, that is, p(α) = μ(α; 1 + x, y). The general form of the Gamma distribution is Γ(α|x, y) = Γ(x) -1 y x α x-1 e -yα ,in represents the Gamma function. Substituting the super prior of α into the above formula, we can get p(α)=Γ(1+x) -1 y 1+x α x e -yα , so far the prior probability of α is modeled. Assume that the prior probability of β follows a uniform distribution without information: According to the user side channel knowledge map obtained locally and the sparse Bayesian learning method, each element in w is modeled as obeying the mean The complex Gaussian distribution of , with variance represented by γ:

[0090]

[0091] where γ=[γ1,γ2,...,γ L ] T , similarly, each element γ in γ i Modeled as independent Gamma distribution, that is So far, the optimal solution to the estimation problem is as follows:

[0092]

[0093] Or, equivalently,

[0094]

[0095] Next, we iteratively construct a continuous surrogate function (lower bound) for the objective function lnp(y,α,γ,β), and then alternately maximize the surrogate function with respect to α, γ, and β. The surrogate function is chosen so that each variable in the alternating maximization has a closed-form and simple solution. Specifically, let As a fixed point Construct a substitution function that satisfies the following properties:

[0096]

[0097] After constructing the substitution function, update α, γ, and β using the following formula:

[0098]

[0099] in(·) (i) Represents the i-th iteration. The overall process of the iterative algorithm is to initialize α (0) , γ (0) and β (0) After that, the three variables α, γ and β are iteratively updated in sequence through the above three formulas until convergence. According to the idea of ​​Expectation Maximization (EM) algorithm, the corresponding lower bound function is selected as the replacement function ——For any fixed point Construct the replacement function as:

[0100]

[0101] And it can be proved that it meets all the requirements for the substitution function.

[0102] Step 4: Based on the user’s actual location and using the prior information provided by the user’s side channel knowledge map The noise coefficient α, gain matrix variance γ and angle offset β are optimized and solved by repeated iterations.

[0103] In step 4, since p(w|y,α,γ,β) satisfies the complex Gaussian distribution:

[0104]

[0105] Its mean and variance are:

[0106]

[0107] Σ(α,γ,β)=(αΦ H (β)Φ(β)+diag(γ)) -1

[0108] Thus, there is a simple closed-form solution to the maximization problem of α:

[0109]

[0110] in:

[0111]

[0112] The maximization problem of γ has a closed-form solution of the following form:

[0113]

[0114] Among them Since the maximization problem with respect to β is non-convex and it is difficult to find its optimal solution, we apply gradient update to its objective function and obtain a simple one-step update of β. The derivative of its objective function with respect to β can be calculated as:

[0115]

[0116] in and

[0117] Represent μ(α (i+1) ,γ (i+1) ,β (i) ) and Σ(α (i+1) ,γ (i+1) ,β (i) ). Obviously, the optimal solution of β is difficult to obtain, but due to the convergence of the iterative algorithm, it is only necessary to find a suboptimal solution so that the value of the objective function gradually increases. The most commonly used numerical method is to update β in the direction of the derivative. l In order to reduce the computational complexity, we choose to use a fixed step size to update β:

[0118]

[0119] Where, represents the grid interval, sign(·) represents the signum function, whose computational complexity can be ignored.

[0120] Since a small difference between the obtained angle and the true angle will not have much impact on the channel estimation performance, Ensure that the final gap is less than r θ 0.5% of , and in the worst case, the (approximate) true value can be obtained within 200 iterations.

[0121] The optimal estimated parameters (α ★ ,γ ★ ,β ★ ).

[0122] Furthermore, in step 4,

[0123]

[0124] Observing the above equation, the left side of the equation is the probability about w, while the denominator on the right side of the equation is the probability about y. Therefore, the denominator p(y|α,γ,β) can be understood as a coefficient that adjusts the size. It only plays the role of making the right side of the equation a probability. From its equivalent expression ∫p(y|w,α,γ,β)p(w|α,γ,β)dw, it can also be seen that it is independent of w, so it can be directly regarded as a constant. Observing the numerator, it is found that since p(y|w,α,γ,β) and p(w|α,γ,β) are both complex Gaussian distributions, it can be proved that p(w|y,α,γ,β) is also a complex Gaussian distribution, and its mean and variance are determined by the numerator of the above equation on the right. Therefore, we have:

[0125]

[0126] Where C, C', and C" represent different constants. From the above formula, we can see that p(w|y,α,γ,β) satisfies the following complex Gaussian distribution: The mean and variance are:

[0127]

[0128] Furthermore, the objective function of α in step 4 is:

[0129]

[0130] After neglecting the constant term, we get:

[0131]

[0132] And because:

[0133]

[0134] Ignoring -Tlnπ, we get:

[0135]

[0136] The integral derivation process in the above formula is as follows:

[0137]

[0138] Furthermore, the objective function of γ in step 4 is:

[0139]

[0140] After neglecting the constant term, we get:

[0141]

[0142] The above formula is used to obtain the partial derivative of each γ:

[0143]

[0144] Finally, let the partial derivative be 0, solve for γ, and get the only optimal solution:

[0145]

[0146] in

[0147] The integrals in the above formula are explained here:

[0148] because:

[0149]

[0150] Accordingly,

[0151]

[0152] For the second integral, we have:

[0153]

[0154] Furthermore, the objective function related to β in step 4 can be rewritten as:

[0155]

[0156] Note that p(β) in the above formula is different from the two variables p(α) and p(γ) mentioned above. It is a constant that is independent of β because it is a Therefore, ignoring the constant term in the above formula that is independent of β, we get:

[0157]

[0158] It is worth noting that μ(α (i+1) ,γ (i+1) ,β (i) ) and Σ(α (i+1) ,γ (i+1) ,β (i) ) refers to p(w|y,α (i+1) ,γ (i+1) ,β (i) ), the mean and variance of , both α and β are fixed, so they can be regarded as constants. This is very important, otherwise the problem will become quite complicated.

[0159] For the convenience of notation, μ is used in the following derivation. (i) and Σ (i) To represent μ(α (i+1) , γ (i+1) , β(i) ) and Σ(α (i+1) ,γ (i+1) ,β (i) ). Calculate each term in the above formula with respect to β l The derivative of , due to:

[0160]

[0161] Let the above formula be l Find the partial derivative and pay attention to β l Represents the angle, which is a real number, so there is no need to conjugate the transpose when taking the partial derivative.

[0162]

[0163] So the first item has:

[0164]

[0165] For the second item:

[0166]

[0167] At this point, the noise coefficient α, gain matrix variance γ, and angle offset β are all updated.

[0168] Step 5: Based on the prior information provided by the local user side channel knowledge map Accurately estimate the gain matrix w and use the received and transmitted pilots to estimate the channel information h at the user's location e .

[0169] In step 5, we get the optimal (α ★ ,γ ★ ,β ★ ) and then use the following formula to accurately estimate the gain matrix:

[0170]

[0171] Therefore, the channel matrix of the user at this position is h e =A(β ★ )w new .

[0172] Furthermore, the update of w in step 5 is the solution to the following optimization problem:

[0173]

[0174] From a geometric and practical perspective, this optimization problem requires that the new w needs to be updated based on the new pilot and the new received signal y (the first term of the objective function), while also not deviating too far from the old w during the update (the second term of the objective function). In other words, the update of w is actually achieved by fine-tuning the initial value provided by the user-side CKM based on the new pilot and received signal.

[0175] Let’s derive the closed-form solution of the above equation. Let the objective function be w * Taking partial derivatives we get:

[0176]

[0177] Let the above partial derivatives be equal to 0, so we have:

[0178]

[0179] Finally, the estimated channel matrix of the user at this position is h e =A(β ★ )w new .

[0180] The following describes in detail the low pilot overhead channel estimation method based on the user-side channel knowledge map of the present invention through specific embodiments.

[0181] like Figure 2 The millimeter wave channel model shown in the figure adopts the millimeter wave channel estimation method proposed in the present invention. It can query the local user-side channel knowledge map according to the user's own location and use the prior information provided by the user-side channel knowledge map to accurately estimate the channel matrix of the user's actual location, such as Figure 3 As shown, the specific steps are:

[0182] Step 1: Use the channel simulation software SEU_PML_6GPCS to obtain the channel information of the user under a certain trajectory. SEU_PML_6GPCS is a 6G universal channel simulator independently developed by the Purple Mountain Laboratory of Southeast University. It can set various parameters of the millimeter wave channel, such as the initial number of clusters, the number of sub-paths within a cluster, etc., to obtain the channel information of the user under a certain trajectory.

[0183] The operation effect of this method is demonstrated by taking the number of transmitting antennas as 100, the number of receiving antennas as 1, the trajectory type as "circle", the speed as 5m / s, the movement duration as 100s, and the sampling frequency as 104Hz as an example. Through this simulator, the channel matrix and various initialization parameters of the user's sampling position on the trajectory can be obtained. Then, the corresponding user-side channel knowledge map is queried according to the channel information of the user at different positions, and the various parameters of the channel model of the user at different positions are stored. After obtaining the channel matrix, it is used as the objective function of the SBL algorithm and it is fully iterated until it is fully converged, and the various parameters of the SBL algorithm corresponding to each user position are finally stored, such as the noise coefficient α, the gain matrix variance γ, and the angle offset β. In the process, a gain matrix w with the same dimension as the number of scattering paths will be obtained, and a dimension of w is constructed. In the actual channel estimation The mean of the distribution The specific method is: assign a small near-zero value to the non-sparse part corresponding to the γ parameter, and assign the value of the gain matrix w with the same dimension as the scattering path according to the corresponding position in the non-sparse part corresponding to the γ parameter, thereby constructing a grid with the same dimension as the grid number.

[0184] In step 2, the user negotiates with the base station for the number of pilots based on the channel knowledge map and specifies communication requirements. The specific requirements are as follows: When the user's location is within the local user-side CKM location range, the user searches the corresponding table based on L in the user-side CKM to determine the number of pilots S to be sent. The table is derived from actual measurements or channel environment simulations. When the user's location is not within the local user-side CKM location range, the default number of 5G / 4G pilots is used.

[0185] Step 3: After querying the local user side channel knowledge map based on the user's own location, the channel stored in the nearest user side CKM is used in the actual channel estimation. The iterative solution is performed as the variance of the distribution obeyed by the gain matrix w estimate.

[0186] Step 4: In the iteration, other parameters (here specifically the noise coefficient α, the gain matrix variance γ and the angle offset β) are used as the initial values ​​of the corresponding parameters of the iterative algorithm and the iterative estimation is started.

[0187] Step 5: Use the initial values ​​of the noise coefficient α, gain matrix variance γ, and angle offset β stored in the user-side CKM and the variance of the distribution obeyed by w estimation The noise coefficient α, gain matrix variance γ, and angle offset β are iteratively solved according to the following three formulas:

[0188]

[0189] in:

[0190]

[0191]

[0192] in

[0193]

[0194] in

[0195] All the means and variances in the above formula are given by the following formula:

[0196]

[0197] Σ(α,γ,β)=(αΦ H (β)Φ(β)+diag(γ)) -1

[0198] When iteratively solving the noise coefficient α, the gain matrix variance γ, and the angle offset β, if γ converges or reaches the maximum number of iterations, the iteration is exited and the final estimated values ​​of each parameter are obtained.

[0199] Step 6: Estimate the optimal channel parameters (α ★ ,γ ★ ,β ★ ) obtains an estimate of the gain matrix and obtains the optimal estimate of the channel matrix w as follows:

[0200]

[0201] Finally, the estimated channel matrix of the user at this position is h e =A(β ★ )w new .

[0202] Figure 4 The proposed low-pilot-overhead channel estimation method based on a user-side channel knowledge map, presented in this paper, shows the convergence of the distance between the channel estimate and the true value as the number of iterations increases under different channel environments. Using the common case of 30-50 scattering paths as an example, it can be seen that the proposed scheme converges to the correct value more quickly with increasing iterations. Furthermore, as the number of paths increases, i.e., as the environment becomes more complex, the number of iterations also increases accordingly.

[0203] Figure 5A comparison chart of pilot overheads of the channel estimation method with low pilot overhead based on the user-side channel knowledge map proposed by the present invention is compared when using the user-side CKM as prior information and when not using the user-side CKM. As can be seen from the figure, when using the user-side CKM as prior information, the pilot overhead is generally less than when not using the user-side CKM. This is an important feature of the present invention, that is, using the channel information of the user's previous position to assist in estimating the channel information of the user's current position, thereby achieving the purpose of reducing the pilot. It can also be seen from the figure that when the channel environment is not complex (the number of scattering paths is small), the number of pilots required for using the user-side CKM is much smaller than when not using the user-side CKM. As the channel environment becomes more complex (the number of scattering paths is large), the number of pilots required for using the user-side CKM begins to increase to a level close to that of not using the user-side CKM, which means that when the channel environment becomes complex, the information provided by the channel at the user's previous position has little reference value.

[0204] The present invention proposes a low-pilot-overhead channel estimation method based on a user-side channel knowledge map. The method uses the user's own location to query the local user-side channel knowledge map. The user negotiates the number of pilots with the base station based on the channel knowledge map and puts forward communication requirements. Based on the prior information stored in this map and the sparse Bayesian learning algorithm, an accurate estimation of the channel at the user's actual location is obtained while greatly reducing the overhead of sending and receiving pilots. The algorithm converges quickly, the channel estimation result is highly accurate, and it has a certain degree of robustness against interference in the actual user channel.

[0205] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or N embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine different embodiments or examples described in this specification and the features of different embodiments or examples without contradiction.

[0206] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of technical features indicated. Thus, a feature specified as "first" or "second" may explicitly or implicitly include at least one such feature. In the description of the present invention, "N" means at least two, such as two, three, etc., unless otherwise specifically defined.

[0207] Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, segment or portion of code comprising one or N executable instructions for implementing a custom logical function or step of a process, and the scope of the preferred embodiments of the present invention includes alternative implementations in which functions may be performed out of the order shown or discussed, including performing functions in a substantially simultaneous manner or in the reverse order depending on the functions involved, which should be understood by those skilled in the art to which the embodiments of the present invention pertain.

Claims

1. A low pilot overhead channel estimation method based on user-side channel knowledge map, characterized in that: The following steps are involved: Step 1: Query the local user side channel knowledge map based on the user's actual location; Step 2: Negotiate with the base station based on the user's actual location to obtain the required number of pilot signals for transmission and reception; Step 3, divide the emission angle domain into multiple parts evenly, and construct a substitute function to replace the maximum likelihood function, so that the optimization objective becomes the constructed substitute function; in step 3, the emission angle domain is divided into multiple parts evenly, and a substitute function is constructed to replace the maximum likelihood function, so as to change the optimization objective into the constructed substitute function; Divide evenly and construct an alternative function To replace the maximum likelihood function lnp(y,α,γ,β), thereby changing the optimization objective to an alternative function Specifically include: The construction method of the substitution function is as follows: Under the assumption of complex Gaussian noise, in, Is a sparse vector, the non-zero elements represent the real vector at the non-zero element on the grid vector, α=σ -2 Represents noise, which is usually unknown. The noise is modeled as a hyper-prior probability of the Gamma distribution, that is, p(α) = Γ(α; 1 + x, y). The general form of the Gamma distribution is Γ(α|x, y) = Γ(x) -1 y x α x-1 e -yα ,in represents the Gamma function, and substituting the super prior of α into the above formula, we get p(α)=Γ(1+x) -1 y 1+x α x e -yα , so far the prior probability of α is modeled; the prior probability of β is assumed to obey a uniform distribution without information: According to the user side channel knowledge map obtained locally and the sparse Bayesian learning method, each element in w is modeled as obeying the mean The complex Gaussian distribution of , with variance represented by γ: in Each element γ in γ i Modeled as independent Gamma distribution, that is The optimal solution to the estimation problem is as follows: is equivalent to: Iteratively construct a continuous surrogate function for the objective function lnp(y, α, γ, β), and then alternately maximize the constructed surrogate function with respect to α, γ, and β. The surrogate function is chosen so that each variable in the alternating maximization has a closed-form and simple solution. Step 4: Based on the actual user location and using the prior information provided by the user-side channel knowledge map, the noise coefficient α, the gain matrix variance γ, and the angle offset β are iteratively optimized and solved. Step 5: Accurately estimate the gain matrix based on the prior information provided by the local user-side channel knowledge map, and use the transmitted and received pilot signals to estimate the channel information at the user's location.

2. The method according to claim 1, characterized in that In step 1, a normalized narrowband millimeter-wave channel with N transmit antennas, a single receive antenna, and L scattering paths is modeled as the classic Saleh-Valenzuela model. The channel matrix of the modeled downlink millimeter-wave channel is: Among them, α l Represents the fading coefficient of each path, that is, the path loss and phase shift of the lth path, a r (θ l ) represents the direction vector of the receiving antenna of the lth path. Assuming that both the transmitting and receiving antennas are uniform linear arrays, then: is the array vector of the N-dimensional antenna array, in the downlink millimeter wave channel θ l Represents the departure angle of the lth path, modeled as obeying After the downlink millimeter wave channel model is established, the user queries the local user-side channel knowledge map according to his actual location to obtain the channel parameters of the user at each location through sparse Bayesian learning.

3. The method according to claim 1, characterized in that In step 2, assuming that the antenna array of the base station contains N antennas, the user negotiates with the base station in the uplink channel and sends S pilots X. The number of pilots S is obtained according to the following rules: when the user location is within the location range of the user-side channel knowledge map queried locally, the corresponding table is searched according to L in the user-side channel knowledge map to obtain the number of sent pilots S. The table is obtained based on actual measurement or channel environment simulation; when the user location is not within the location range of the local user-side channel knowledge map, the 5G / 4G default pilot number is used to obtain the received signal y = Xh + n at each user end, where n represents complex Gaussian white noise with a variance of σ 2 , h is the channel matrix to be estimated.

4. The method according to claim 3, characterized in that Step 2 specifically includes: The actual launch azimuth angle AoD is expressed as: {θ l ,l=1,2,...,L}, where L=NcNs represents the total number of paths, Nc and Ns represent the number of scatterers and the number of paths contained in each scatterer respectively, each path corresponds to an AoD, so there are a total of L=NcNs AoDs to be estimated; the angle domain Divide evenly The grid angles after gridding are recorded as When the true angle θ l Just fell here When the channel model is written as h = Aw, a(θ) is the array vector, Is a sparse vector, the non-zero elements represent the real vector at the non-zero element on the grid vector, that is, if the first elements are non-zero and the corresponding real vector is θ l , then at this time there is An off-grid model is used to model the situation where the true angle is not exactly on the grid: if and is the distance θ l The nearest grid point is written as Represents the angle offset, that is, the offset of the true value relative to the nearest grid, that is: The array vector at this time is written as Correspondingly, the channel matrix h at this time is h = A(β)w, where The signal received by the receiving end is expressed as y=XA(β)w+n=Φ(β)w+n, where 5. The method according to claim 1, characterized in that In step 3, the emission angle domain Divide evenly and construct an alternative function To replace the maximum likelihood function lnp(y,α,γ,β), thereby changing the optimization objective to an alternative function Specifically include: let As a fixed point The constructed substitution function satisfies the following properties: After constructing the substitution function, update α, γ, and β using the following formula: in(·) (i) Indicates the i-th iteration. The overall process of the iterative algorithm is to initialize α (0) , γ (0) and β (0) After that, the three variables α, γ and β are iteratively updated in sequence through the above three formulas until convergence. According to the idea of ​​ExpectationMaximization algorithm, the corresponding lower bound function is selected as the replacement function ——For any fixed point Construct the replacement function as:

6. The method according to claim 1, characterized in that In step 4, the noise coefficient α, gain matrix variance γ, and angle offset β are iteratively optimized based on the actual user location and the prior information provided by the user-side channel knowledge map. Specifically, the following steps are performed: Since p(w|y,α,γ,β) satisfies the complex Gaussian distribution: The mean and variance are: Σ(α,γ,β)=(αΦ H (β)Φ(β)+diag(γ)) -1 Therefore, there is a closed solution to the maximization problem of α: in: The maximization problem of γ has a closed-form solution of the following form: in, Since the maximization problem with respect to β is non-convex and it is difficult to find its optimal solution, we apply gradient update to the objective function of the maximization problem of β and obtain a simple one-step update of β. The derivative of the objective function with respect to β is calculated as: in, and and Represent μ(α (i+1) , γ (i+1) , β (i) )'s lth element and Σ(α (i+1) ,γ (i+1) ,β (i) )’s (j,l)th element; Obviously, the optimal solution of β is difficult to obtain, but due to the convergence of the iterative algorithm, it is only necessary to find a suboptimal solution so that the value of the objective function gradually increases. The most commonly used numerical method is to update β in the derivative direction. l In order to reduce the computational complexity, we choose to use a fixed step size to update β: Where, represents the grid interval, sign(·) represents the signum function; The above iterations are repeated until convergence to obtain the optimal estimated parameter (α ＊ ,γ ＊ ,β*).

7. The method according to claim 6, characterized in that In step 5, the optimal estimated parameters (α ＊ ,γ ＊ ,β*) and then use the following formula to accurately estimate the gain matrix: Therefore, the channel matrix of the user at the current actual position is h e =A(β ＊ )w new .

Images