A time-varying channel estimation method, device, and medium combining UKF and stacked ELM

Abstract

This invention discloses a time-varying channel estimation method, device, and medium combining UKF and stacked ELM. The method first characterizes the time-varying channel using a basis extension model. Then, based on historical channel basis coefficient estimates, it introduces the UKF method to predict an initial estimate of the basis coefficients at the current time. This initial estimate is then used to train and test a stacked ELM network to obtain a more accurate basis coefficient estimate. Finally, based on the transformation relationship between the basis coefficients and the frequency domain channel, the final frequency domain channel estimate is obtained. To address the parameter estimation problem under nonlinear state-space models, this invention, based on historical channel basis coefficient estimates, first introduces unscented Kalman filtering (UKF) to predict an initial estimate of the basis coefficients at the current time. This initial estimate is then used to train a stacked Extreme Learning Machine (ELM) network to extract deep channel features and further improve the accuracy of channel estimation. This invention offers higher estimation accuracy than existing methods and is suitable for acquiring time-varying channel information in high-speed mobile scenarios.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of wireless communication, and in particular to a time-varying channel estimation method combining UKF and stack ELM, a device and a storage medium. BACKGROUND

[0002] In recent years, with the continuous advancement of the research on the sixth generation of mobile communication technology, wireless communication in scenarios such as high-speed trains, low-orbit satellites, and Internet of Vehicles has attracted more and more attention from the academic and industrial circles. In these high-speed mobile environments, the existence of large Doppler frequency shift makes the channel present strong fast time-varying characteristics, which will bring severe challenges to the accurate acquisition of channel state information (CSI) in this scenario.

[0003] At present, deep learning has been widely used in the field of wireless communication, and has achieved some remarkable results in time-varying channel estimation. Among them, WANG T Q et al. (WANG T Q, WEN C K, JIN S, et al. Deeplearning-based CSI feedback approach for time-varing massive MIMO channels) use long short-term memory (LSTM) network to learn the time correlation between training samples, and realize the real-time feedback of CSI in frequency division duplex network. However, due to the complexity of the gating mechanism and memory unit in the LSTM network, a large number of weight parameters need to be trained, so the network training speed is slow. LI J et al. (LI J, ZHANG Z C, WANG Y K, et al. Deep learning-assisted OFDM channel estimation and signal detection technology) apply self-normalizing network (SNN) to channel estimation. Due to the introduction of self-normalization characteristics by using Selu activation function, the learning is more robust, and can effectively prevent gradient explosion and gradient disappearance. LIAO Y et al. (LIAO Y, HUA Y X, DAI X W, et al. ChanEstNet: a deep learning based channel estimation for high-speed scenarios) give a channel estimation method combined with convolutional neural network (CNN) and recurrent neural network (RNN). This method uses CNN to extract pilot sequence features, and then uses RNN to estimate the channel according to the forward and backward pilot information. However, the combination of the two neural networks makes this method have high computational complexity. SOLTANI M et al. (SOLTANI M, POURAHMADIV, MIRZAEI A, et al. Deep learning-based channel estimation) regard the CSI of pilot symbols as a low-resolution image, and input it into a super-resolution network and a denoising neural network to obtain a high-resolution image (i.e. high-precision CSI estimation). However, this method also has high computational complexity.LI L J et al. (LI L J, CHEN H, CHANG H H et al., "Deep residual learning meets OFDM channel estimation") gives a channel estimation method based on residual neural network, which first uses neural network to obtain higher-precision channel estimation of pilot symbols, and then uses linear interpolation to obtain channel estimation of data symbols, the network structure adopted by the method is compact and flexible, and the calculation complexity is low. In order to improve the generalization performance of the method, Song Gang (Song Gang, "Deep learning algorithm based on ELM and its application research in time series prediction") gives a channel estimation method based on extreme learning machine (ELM) network, however, the method adopts a shallow structure, which cannot extract deeper channel features from complex original data, which will limit the estimation accuracy of the method.

[0004] Compared with the traditional channel estimation method, the estimation method based on neural network is more suitable for high-speed mobile scene, however, most of these existing deep learning-based channel estimation methods adopt complex network structure, which needs a large number of training samples to train the network and the network training speed is slow. Therefore, it is necessary to study a practical channel estimation method with higher estimation accuracy and faster training speed suitable for high-speed mobile scene. SUMMARY

[0005] The time-varying channel estimation method, device and storage medium jointly using UKF and stacked ELM proposed by the application can at least solve one of the technical problems in the background art.

[0006] To achieve the above-mentioned purpose, the application adopts the following technical scheme:

[0007] A time-varying channel estimation method jointly using UKF and stacked ELM, comprising the following steps,

[0008] First, the time-varying channel is characterized by a basis expansion model, then based on the channel basis coefficient estimation of the historical time, the initial estimation of the basis coefficient of the current time is obtained by introducing the UKF method, and then the stacked ELM network is trained and tested by using the initial estimation, so as to obtain higher-precision basis coefficient estimation; finally, based on the conversion relationship between the basis coefficient and the frequency domain channel, the final frequency domain channel estimation is obtained.

[0009] Further, the time-varying channel is characterized by a basis expansion model, comprising,

[0010] Step 1: Constructing a state space model

[0011]

[0012] where c i-1 and c i are the base coefficients of the (i-1)th and ith OFDM symbol periods, respectively, T represents the correlation matrix of the base coefficients over two adjacent OFDM symbol periods, v i is the channel transition process noise vector of the ith OFDM symbol, and the covariance matrix of v I QL is a QLxQL dimensional identity matrix; the observation matrix where is the predicted value of the data symbol obtained by means of decision feedback.

[0013] Further, based on the channel base coefficient estimation at the historical time, the UKF method is introduced to predict the initial estimation of the base coefficients at the current time, including,

[0014] Step 2: Based on the existing channel estimation method, the channel base coefficient estimation at the (i-1)th OFDM symbol period is obtained and 2D+1 sigma points are generated, i.e.

[0015]

[0016] where D=QL, is the jth sigma point, P i-1 is the posterior covariance matrix of the base coefficients at the (i-1)th OFDM symbol period; ξ is a proportional coefficient, and ξ=α 2 (D+β)-D, and α and β determine the distribution of the sigma points around the mean value;

[0017] Step 3: The sigma points in Step 2 are substituted into the state prediction equation to obtain the predicted value of the base coefficients at the ith OFDM symbol period

[0018]

[0019] Step 4: Based on the predicted value of the base coefficients obtained in Step 3, the weighted average is performed to obtain the prior estimation of the base coefficients at the ith OFDM symbol period and the covariance matrix P i∣i-1 is

[0020]

[0021] where and are the weighted coefficients of the mean value and variance of the sigma point predicted value, respectively, and the specific values are

[0022]

[0023] Step 5: Based on the prior estimation of basis coefficients obtained in Step 4, 2D+1 sigma points are generated, i.e.

[0024]

[0025] Step 6: Substitute the sigma points in Step 5 into the state measurement equation to obtain the observation value of the i-th OFDM received symbol, i.e.

[0026]

[0027] Step 7: Based on the observation value of the i-th OFDM received symbol obtained in Step 6, the weighted average of the observation value is obtained, i.e. i , the covariance matrix Ω i and the cross-covariance matrix Θ i , i.e.

[0028]

[0029] In the formula, Q e =E i ;

[0030] Step 8: According to the Bayesian filtering principle, the posterior estimation of the channel basis coefficients and the covariance matrix P i are calculated, i.e.

[0031]

[0032] In the formula, Ψ i represents the UKF gain, and

[0033] Further, the initial estimation is used to train and test the stacked ELM network to obtain a higher precision basis coefficient estimation, including,

[0034] Step 9: Based on the initial basis coefficient estimation in the i-th OFDM symbol period obtained in Step 8 , the training sample set is constructed as

[0035]

[0036] In the formula, U is the number of samples, represents the initial basis coefficient estimation in the i-th OFDM symbol period in the u-th training sample, which is the input sample; is the ideal basis coefficient value in the i-th OFDM symbol period in the u-th training sample, which is the target of the stacked ELM network training approximation; Γ(·) is a complex-to-real operation, Γ(·) = [Re(·), Im(·)], wherein Re(·) and Im(·) are real part and imaginary part operations, respectively;

[0037] Step 10: training the stacked ELM network with the training sample set;

[0038] Step 11: based on the trained stacked ELM network with output weights in Step 10, taking the initial estimation of the base coefficients of the current OFDM symbol period obtained based on the UKF as the input, the high-precision base coefficient estimation of the current symbol can be obtained

[0039] Further, based on the conversion relationship between the base coefficients and the frequency domain channel, the final frequency domain channel estimation is obtained, including,

[0040] Step 12: based on the high-precision base coefficient estimation of the current symbol obtained in Step 11, the high-precision frequency domain channel estimation of the current OFDM symbol can be obtained according to the conversion relationship between the base coefficients and the frequency domain channel

[0041]

[0042] In the formula,

[0043] Further, the stacked ELM network structure is adopted in Step 10, each ELM module contains an input layer, a hidden layer and an output layer, and the number of neurons of the input layer and the output layer is S, and the number of neurons of the hidden layer is Z;

[0044] After the training sample set is sent into the first ELM module, the ELM will randomly generate the weights and biases from the input layer to the hidden layer, and calculate the hidden layer output matrix, and then obtain the weights between the hidden layer and the output layer through the least square method, that is,

[0045]

[0046] In the formula, represents the output weight matrix of the ELM network, c i is the ideal base coefficient of the channel on the ith OFDM symbol period, represents the pseudo-inverse of the matrix; A is the hidden layer output matrix, that is,

[0047]

[0048] In the formula, represents the hidden layer feature mapping function of the zth node, is the weight vector between the input layer neurons and the hidden layer neurons, ε z is the bias of the zth neuron in the hidden layer, and G(·) is the hidden layer activation function;

[0049] The output of the first ELM module is

[0050]

[0051] Based on the output in the above formula, the original feature matrix C1 is randomly offset, then a new feature matrix C2 is obtained through a kernel function, and then it is taken as the input of the second ELM module, and the subsequent process is similar, then

[0052] C j+1 = G(C1+ λO j θ j )

[0053] In the formula, O j is the output of the jth ELM network, the projection matrix θ j ∈R S×S is randomly sampled from the normal distribution N(0, 1), and λ is a weight parameter for controlling the degree of random offset.

[0054] In another aspect, the application also discloses a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to make the processor execute the steps of the above method.

[0055] In another aspect, the application also discloses a computer device, which comprises a memory and a processor, and the memory stores a computer program, and the computer program is executed by the processor to make the processor execute the steps of the above method.

[0056] According to the above technical solution, the time-varying channel estimation method combining UKF and stacked ELM of the application first characterizes the time-varying channel through a basis expansion model, then predicts the initial estimation of the basis coefficient at the current time based on the basis coefficient estimation at the historical time, and then trains and tests the stacked ELM network by using the initial estimation, so as to obtain a higher-precision basis coefficient estimation. Finally, based on the conversion relationship between the basis coefficient and the frequency-domain channel, the final frequency-domain channel estimation is obtained.

[0057] Specifically, for a high-speed mobile orthogonal frequency division multiplexing system, the application first uses a basis expansion model to fit the fast time-varying channel, and converts the channel estimation into the estimation of the channel basis coefficient with limited parameters, which greatly reduces the calculation complexity. In order to solve the parameter estimation problem under the nonlinear state space model, the application first introduces the unscented Kalman filter (UKF) to predict the initial estimation of the basis coefficient at the current time based on the basis coefficient estimation at the historical time, and then trains the stacked extreme learning machine (ELM) network by using the initial estimation, so as to extract the deep features of the channel and further improve the estimation accuracy of the channel. The application has higher estimation accuracy than the existing method, and is suitable for obtaining time-varying channel information in a high-speed mobile scene.

[0058] Compared with the prior art, the technical scheme of the present application is a time-varying channel estimation method combining UKF and stacked ELM. Firstly, a basis expansion model is used to fit the fast time-varying channel, and the channel estimation is converted into the estimation of the channel basis coefficient with limited parameters, thereby reducing the calculation complexity. Then, based on the channel basis coefficient estimation at the historical time, UKF is introduced to obtain the initial estimation of the basis coefficient at the current time. Finally, the initial estimation is used to train the stacked ELM network to extract the deep features of the channel and further improve the accuracy of the channel estimation. The method has high estimation performance, fast training speed and low complexity, and therefore has certain practical value. BRIEF DESCRIPTION OF DRAWINGS

[0059] Figure 1 The flowchart of the embodiment of the present application is shown in the figure.

[0060] Figure 2 The structure diagram of the stacked ELM network used in the embodiment of the present application is shown in the figure.

[0061] Figure 3 The MSE performance comparison diagram of the embodiment of the present application and other channel estimation methods under different signal-to-noise ratios is shown in the figure.

[0062] Figure 4 The MSE performance comparison diagram of the embodiment of the present application and other channel estimation methods under different normalized Doppler frequency shifts is shown in the figure. DETAILED DESCRIPTION

[0063] To make the purpose, technical scheme and advantages of the embodiment of the present application clearer, the technical scheme in the embodiment of the present application will be described clearly and completely below in combination with the figures in the embodiment of the present application. Obviously, the described embodiment is a part of the embodiments of the present application, rather than all the embodiments of the present application.

[0064] As Figure 1The time-varying channel estimation method combining UKF and stack ELM described in the embodiment, namely a time-varying channel estimation method combining UKF and stack ELM applied to a high-speed mobile OFDM system, aims to improve the estimation accuracy of the time-varying channel. The method first uses a GCE-BEM to represent the time-varying channel, then predicts the initial estimation of the base coefficient at the current time based on UKF and historical base coefficient estimation, extracts the deep features of the channel through the stack ELM network, thereby obtains the high-precision base coefficient estimation, and finally obtains the final frequency domain channel estimation by using the conversion relationship between the derived base coefficient and the frequency domain channel.

[0065] Specifically, the technical scheme adopted by the present application comprises the following steps:

[0066] Step 1: Constructing a state space model

[0067]

[0068] In the formula, c i-1 and c i are the base coefficients of the i-1th and i-th OFDM symbol periods, T represents the correlation matrix of the base coefficients on adjacent two OFDM symbol periods, v i is the channel transition process noise vector of the i-th OFDM symbol, and the covariance matrix is I QL is a QLxQL-dimensional unit matrix. The observation matrix wherein is the predicted value of the data symbol obtained by the decision feedback.

[0069] Step 2: Based on the existing channel estimation method, the base coefficient estimation of the i-1th OFDM symbol period is obtained and 2D+1 sigma points are generated, namely

[0070]

[0071] In the formula, D=QL, is the j-th sigma point, P i-1 is the posterior covariance matrix of the base coefficient on the i-1th OFDM symbol period. ξ is a proportional coefficient, and ξ=α 2(D+β)-D, and β determine the distribution of the sigma points around the mean.

[0072] Step 3: Substitute the sigma points in Step 2 into the state prediction equation to obtain the predicted value of the base coefficients in the i-th OFDM symbol period

[0073]

[0074] Step 4: Based on the predicted value of the base coefficients obtained in Step 3, perform weighted averaging to obtain the prior estimation of the base coefficients in the i-th OFDM symbol period and the covariance matrix P i∣i-1 is

[0075]

[0076] In the formula, and are the weighted coefficients of the mean and variance of the sigma point predicted value, and the specific values are

[0077]

[0078] Step 5: Based on the prior estimation of the base coefficients obtained in Step 4, generate 2D+1 sigma points, that is,

[0079]

[0080] Step 6: Substitute the sigma points in Step 5 into the state measurement equation to obtain the observation value of the i-th OFDM received symbol

[0081]

[0082] Step 7: Based on the observation value of the i-th OFDM received symbol obtained in Step 6, perform weighted averaging to obtain the mean μ i of the observation value, the covariance matrix Ω i and the cross-covariance matrix Θ i , that is

[0083]

[0084] In the formula, Q e = E i .

[0085] Step 8: According to the Bayesian filtering principle, calculate the posterior estimation of the channel base coefficients and the covariance matrix P i is

[0086]

[0087] In the formula, Ψi Represents UKF gain, and

[0088] Step 9: Estimation of the initial basis coefficients for the i-th OFDM symbol period obtained in Step 8 Construct the training sample set as

[0089]

[0090] In the formula, U is the number of samples. This represents the initial basis coefficient estimate for the i-th OFDM symbol period in the u-th training sample, which is the input sample. Let be the ideal basis coefficient value in the i-th OFDM symbol period of the u-th training sample, which is the target approximation for training the stacked ELM network. Γ(·) is the complex-to-real number conversion operation, Γ(·)=[Re(·),Im(·)], where Re(·) and Im(·) are the operations of extracting the real and imaginary parts, respectively.

[0091] Step 10: Train a stacked ELM network using the training sample set;

[0092] Step 11: Based on the stacked ELM network with output weights trained in Step 10, the initial estimate of the basis coefficients for the current OFDM symbol period obtained based on UKF is used as input to obtain a high-precision basis coefficient estimate for the current symbol.

[0093] Step 12: Based on the high-precision basis coefficient estimate of the current symbol obtained in Step 11, the high-precision frequency domain channel estimate of the current OFDM symbol can be obtained according to the conversion relationship between the basis coefficients and the frequency domain channel.

[0094]

[0095] In the formula,

[0096] The following are examples:

[0097] Consider a single-transmit, single-receive OFDM system with N subcarriers. At the transmitter, assume that the symbol X transmitted on the k-th subcarrier is the i-th OFDM symbol. i (k), after IFFT, the time-domain signal at time n is obtained as follows:

[0098]

[0099] To mitigate inter-symbol interference, a cyclic prefix (CP) is inserted. After the signal passes through a time-varying channel, assuming ideal synchronization, the received i-th OFDM symbol can be represented as...

[0100]

[0101] where e i (n) is an additive white Gaussian noise(AWGN) with mean 0 and variance h i (n,l) is the channel gain at the n-th time instant on the l-th path of the i-th OFDM symbol period, L is the number of paths of the channel, τ l is the normalized delay of the l-th path.

[0102] The received i-th OFDM symbol can be further expressed in matrix form as

[0103] y i = g i x i + e i

[0104] where x i = [x i (0), x i (1),..., x i (N-1)] T , e i = [e i (0), e i (1),..., e i (N-1)] T , g i ∈ C N ×N is the time-domain channel matrix on the i-th OFDM symbol period, i.e.

[0105]

[0106] Since the basis expansion model(BEM) can compress the time-varying channel into a low-dimensional space through a series of basis functions, it greatly reduces the number of parameters to be estimated in time-domain channel estimation, thereby reducing the complexity of estimation. Therefore, this paper will use BEM to model the time-domain channel, which converts the channel into the superposition of the weights of a number of mutually orthogonal basis functions, i.e.

[0107]

[0108] where Q denotes the number of BEM basis functions, b i,q (n) is the n-th element of the q-th basis function of the i-th OFDM symbol period, c i,l (q) denotes the basis coefficient corresponding to the q-th basis function of the l-th path of the i-th OFDM symbol period, ζi,l (n) is the BEM modeling error.

[0109] The time-domain channel can be expressed in the form of a matrix as

[0110] h i,l = B i c i,l + ζ i,l

[0111] where c i,l = [c i,l (0), c i,l (1), …, c i,l (Q-1)] T , and ζ i,l = [ζ i,l (0), ζ i,l (1), …, ζ i,l (N-1)] T are neglected. B i is an N x Q-dimensional basis function matrix, and B i = [b i,0 , b i,1 , …, b i,Q-1 ], b i,q = [b q ((N+N cp )i+N cp ), b q ((N+N cp )i+N cp +1), …, b q ((N+N cp )(i+1)-1)] T , N cp is the length of the CP in the OFDM symbol. In this paper, the complex exponential basis expansion model is considered to have spectrum leakage and Gibbs effect, and therefore a generalized complex exponential basis expansion model (GCE-BEM) is used for channel modeling, and then

[0112] b q (n) = e j2πn(q-Q / 2) / (ηN)

[0113] where η = 1 / (f max T sym ), f max is the maximum Doppler shift, and T sym is the symbol duration.

[0114] At the receiving end, the i-th received signal in the frequency domain can be expressed as

[0115] Y i = Fgi F H X i +E i

[0116] =M i c i +E i

[0117] In the formula, F is the Fourier transform matrix, and X... i =[X i (0),X i (1),...,X i (N-1)] T , c i,q =[c i,q (0),c i,q (1),…,c i,q (L-1)] T E i This indicates that the mean is zero and the variance is... The frequency domain noise vector. M i It is an N×LQ dimensional observation matrix, and

[0118] M i =[M i,0 M i,1 ,…,M i,Q-1 ]

[0119] In the formula, M i,q =Fdiag(b i,q )F H diag(X i )F L F L It consists of the first L columns of matrix F.

[0120] This invention employs a stacked ELM network structure, such as... Figure 2 As shown, each ELM module contains one input layer, one hidden layer, and one output layer. The number of neurons in both the input and output layers is S, and the number of neurons in the hidden layer is Z. After the training sample set is fed into the first ELM module, the ELM randomly generates the weights and biases from the input layer to the hidden layer, calculates the hidden layer output matrix, and then obtains the weights between the hidden and output layers using the least squares method.

[0121]

[0122] In the formula, c represents the output weight matrix of the ELM network. i Let be the ideal basis coefficients of the channel during the i-th OFDM symbol period. denotes the pseudo-inverse of a matrix. A is the hidden layer output matrix, i.e.

[0123]

[0124] where, denotes the hidden layer feature mapping function of the z-th node, is the weight vector between input layer neurons and hidden layer neurons, ε z is the bias of the z-th neuron in the hidden layer, and G(·) is the hidden layer activation function.

[0125] The output of the first ELM module is

[0126]

[0127] Based on the output in the above equation, a random offset is made to the original feature matrix C1, and then a new feature matrix C2 is obtained by a kernel function, which is taken as the input of the second ELM module, and the subsequent process is similar, and then

[0128] C j+1 = G(C1+ λO j θ j )

[0129] where O j is the output of the j-th ELM network, the projection matrix θ j ∈ R S×S is randomly sampled from a normal distribution N(0, 1), and λ is a weight parameter for controlling the degree of random offset.

[0130] The simulation results of the embodiment are as follows:

[0131] This section will simulate and analyze the present technology to verify its estimation performance. The present technology considers a single-transmit-single-receive OFDM system, in which the FFT / IFFT length is 128, the cyclic prefix length is 16, the transmitted data block length is 2 OFDM symbols, the block pilot structure is used, the first symbol is a pilot symbol, and the second symbol is a data symbol. In the simulation, the channel adopts a 6-path Rayleigh channel, the power delay profile conforms to exponential decay exp(-l), where l is the channel tap index. The carrier frequency is considered to be 2.35 GHz, and the subcarrier spacing is 15 kHz. The number of input layer neurons and the number of output layer neurons of the network are both S = 48, the number of hidden layer neurons is Z = 40, and the activation function is a Sigmoid function. In order to compare the performance of the present technology, the simulation also gives a channel estimation method based on UKF and a channel estimation method based on BP neural network.

[0132] Figure 3MSE performance of different channel estimation methods at a moving speed of 300km / h is given. As can be seen from the figure, the estimation performance of various methods will be better and better with the increase of signal-to-noise ratio, wherein the channel estimation method based on UKF has the worst performance, because the method cannot well obtain the deep features of the time-varying channel. The channel estimation method based on BP neural network can better capture the nonlinear characteristics of the time-varying channel by learning the internal characteristics of the channel through training samples, and thus is superior to the traditional UKF estimation method based on model. The present technology has obvious performance advantages compared with other channel estimation methods, mainly because it adopts a deep ELM model, and extracts more useful deep features from the initial basis coefficient estimation, and thus has better estimation performance.

[0133] Figure 4 MSE performance curves of the present technology and existing channel estimation methods under different normalized Doppler shifts are given. As can be seen from the figure, the MSE performance of various methods will be worse with the increase of Doppler shift, but the MSE performance of the present technology is always superior to other existing methods, because the deep ELM structure adopted by the present technology is more conducive to extracting deep features of data, and can better adapt to the time-varying channel state characteristics compared with the channel estimation method based on UKF and the channel estimation method based on BP neural network. Therefore, the present technology can better adapt to the change of Doppler shift, and has stronger applicability.

[0134] In another aspect, the present application also discloses a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to make the processor execute the steps of the above method.

[0135] In another aspect, the present application also discloses a computer device, which comprises a memory and a processor, and the memory stores a computer program, and the computer program is executed by the processor to make the processor execute the steps of the above method.

[0136] In another embodiment provided in the present application, a computer program product containing instructions is also provided, which makes a computer execute the time-varying channel estimation method of joint UKF and stack ELM in any of the above embodiments when running on the computer.

[0137] It can be understood that the system, device and storage medium provided by the embodiments of the present application correspond to the method provided by the embodiments of the present application, and the explanation, examples and beneficial effects of the related contents can be referred to the corresponding part in the above method.

[0138] In the embodiments described above, all or some of the steps can be implemented by hardware, software, firmware or any combination thereof. When implemented by software, all or some of the steps can be implemented in the form of one or more computer programs or program elements. The computer programs reside (at least temporarily) in a memory of a computer during execution. The memory can be a RAM memory, a flash memory, a ROM memory, an EPROM memory, or any other suitable memory. The memory can be integral to or separate from the computer. The computer programs can be written in any suitable programming language, such as C, C++, Java, Visual Basic, etc. The computer programs can be written in assembly or machine language, if desired. The computer programs can be distributed over network coupled file servers, or can be distributed by any other suitable means.

[0139] It is to be noted that the terms such as first and second, etc., are used herein merely to differentiate one entity or action from another, and do not necessarily require or imply any such actual relationship or order between such entities or actions. Also, the terms "comprising", "containing", or any other similar term are intended to encompass non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not necessarily contain only those elements, but can contain other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0140] Each of the embodiments in the present specification is described in a related manner, and the same or similar parts among the embodiments can be referred to each other. Each of the embodiments focuses on the difference from other embodiments. In particular, for the system embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the part of the description of the method embodiments.

[0141] The above examples are only used to illustrate the technical solutions of the present application, and are not intended to limit the present application; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that the technical solutions recorded in the foregoing examples can be modified, or some technical features can be replaced by equivalent features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A time-varying channel estimation method combining UKF and stacked ELM, characterized in that, The method comprises the following steps of: Firstly, the time-varying channel is characterized by a basis expansion model, and then based on the channel basis coefficient estimation of the historical time, the UKF method is introduced to predict the initial estimation of the basis coefficient at the current time, and then the initial estimation is used to train and test the stacked ELM network to obtain a higher precision basis coefficient estimation; finally, based on the conversion relationship between the basis coefficient and the frequency domain channel, the final frequency domain channel estimation is obtained; The time-varying channel is characterized by a basis expansion model, comprising, Step 1: constructing a state space model where and are the basis coefficients of the first and the first OFDM symbol periods, respectively, represents the correlation matrix of the basis coefficients over two adjacent OFDM symbol periods, is the channel transfer process noise vector of the first OFDM symbol, whose covariance matrix is , is an dimensional identity matrix; the observation matrix is given by , is the predicted value of the data symbol obtained by means of decision feedback. Then based on the channel basis coefficient estimation of the historical time, the UKF method is introduced to predict the initial estimation of the basis coefficient at the current time, comprising, Step 2: Obtain the channel basis coefficient estimates on the jth OFDM symbol period based on the existing channel estimation method, and generate Ns number of sigma points, i.e. hj= [hj,0, hj,1,..., hj,N-1]T and Step 2: Obtain the channel basis coefficient estimates on the jth OFDM symbol period based on the existing channel estimation method, and generate Ns number of sigma points wherein , is the th sigma point, is the th OFDM symbol period; is a scaling factor, and , and determines the distribution of the sigma points around the mean. Step 3: Substitute the sigma points in step 2 into the state prediction equation to obtain the predicted value of the base coefficients on the jth OFDM symbol period as ​ Step 4: Based on the predicted values of the basis coefficients obtained in Step 3, a weighted average is performed to obtain the a priori estimate of the basis coefficients on the first OFDM symbol period and the covariance matrix is wherein and are the weighted coefficients of the mean and variance of the sigma point prediction, respectively, and are given by Step 5: Based on the a priori estimate of the basis coefficients obtained in Step 4, generate sigma points, i.e. Step 6: Substitute the sigma points from Step 5 into the state measurement equation to obtain the observations for the nth OFDM receive symbol as ​ Step 7: Based on the observations of the first OFDM received symbol obtained in step 6, a weighted average is performed to obtain the mean , the covariance matrix and the cross-covariance matrix , i.e. In the formulae, ; Step 8: Compute the posterior estimates of the channel basis coefficients according to the Bayesian filtering principle and the covariance matrix are wherein represents the UKF gain, and .

2. The time-varying channel estimation method of claim 1, wherein: Then the initial estimation is used to train and test the stacked ELM network to obtain a higher precision basis coefficient estimation, comprising, Step 9: Based on the initial basis coefficient estimates on the first OFDM symbol period obtained in Step 8 , the training sample set is constructed as wherein, is the number of samples, denotes the initial basis coefficient estimate on the th OFDM symbol period in the th training sample, which is the input sample; is the ideal basis coefficient value on the th OFDM symbol period in the th training sample, which is the target for the stacked ELM network training approximation; is the complex-to-real operation, wherein and are the take real and imaginary operations, respectively. Step 10: training the stacked ELM network by using the training sample set; Step 11: Based on the stack ELM network with output weights trained in step 10, the initial estimation of the basis coefficients of the current OFDM symbol period obtained based on the UKF is taken as the input, and the high-precision estimation of the basis coefficients of the current symbol is obtained .

3. The time-varying channel estimation method of claim 2, wherein: Finally, based on the conversion relationship between the basis coefficient and the frequency domain channel, the final frequency domain channel estimation is obtained, comprising, Step 12: based on the high-precision basis coefficient estimation of the current symbol obtained in step 11, the high-precision frequency domain channel estimation of the current OFDM symbol can be obtained according to the conversion relationship between the basis coefficient and the frequency domain channel, that is In the formulae, .

4. The time-varying channel estimation method of claim 3, wherein: The stack ELM network structure is adopted in step 10, each ELM module contains an input layer, a hidden layer and an output layer, and the neuron numbers of the input layer and the output layer are , and the neuron number of the hidden layer is ; After the training sample set is sent into the first ELM module, the ELM will randomly generate the weights and biases from the input layer to the hidden layer, and calculate the hidden layer output matrix, and then the weights between the hidden layer and the output layer are obtained by the least square method, that is wherein represents the ELM network output weight matrix, is the ideal basis coefficient of the channel on the th OFDM symbol period, denotes the pseudo-inverse of a matrix; is the hidden layer output matrix, i.e. wherein, represents the hidden layer feature mapping function of the th node, is the weight vector between the input layer neuron and the hidden layer neuron, is the bias of the th neuron of the hidden layer, is the hidden layer activation function; The output of the first ELM module is Based on the output in the above equation, the original feature matrix A random offset is made, and then a new feature matrix is obtained through a kernel function It is then taken as the input of the second ELM module, and the subsequent process is repeated in this way. wherein is the output of the th ELM network, the projection matrix is randomly sampled from a normal distribution , is a weight parameter that controls the degree of random offset. 5.A computer readable storage medium storing a computer program, wherein the computer program is executed by a processor to make the processor perform the steps of the method according to any one of claims 1 to 4. 6.A computer device comprising a memory and a processor, wherein the memory stores a computer program, and the computer program is executed by the processor to make the processor perform the steps of the method according to any one of claims 1 to 4.

Images