A Blind Identification Method and System for Modulation Modes in a Large-Scale Multiple-Input Multiple-Output Communication System

Abstract

The present invention relates to the field of wireless communication technology, and proposes a method and system for blindly identifying modulation modes in a large-scale multi-antenna communication system, which comprises the following steps: T Based on the MDL algorithm, the number of independent UEs is estimated for each signal sample, and the estimated value of the number of independent UEs corresponding to any signal sample is obtained. #imgabs0# Based on the estimated value of the number of independent UEs, the N received by the base station BS is estimated based on the CFICA algorithm. T Perform aliasing signal separation on signal samples to obtain the corresponding UE transmit signal matrix #imgabs1# of any signal sample. Perform cyclostationary feature extraction on any UE transmit signal matrix #imgabs2# to construct N T The CF matrix #imgabs3# is input into the preset CDBN model for training to obtain the CDBN model; based on the trained CDBN model, the CF matrix is calculated for the test data, the obtained CF matrix is input into the trained CDBN model, and the blind recognition result of the modulation mode is output.

Description

Technical Field

[0001] The present invention relates to the technical field of wireless communication, and more particularly, to a method and system for blind recognition of modulation modes in a large-scale multi-antenna communication system. Background Art

[0002] The sharp increase in the number of user equipment (UE) in communication networks has further exacerbated the tension of spectrum resources. Due to its characteristic of improving spectrum efficiency, non-orthogonal multiple access (NOMA) technology has received extensive attention currently. At the same time, to further improve system flexibility and reduce pilot resource consumption, currently, blind modulation recognition (BMR) technology has received extensive attention due to its ability to further improve spectrum utilization and communication reliability.

[0003] Considering that deploying a classifier of a decision-theoretic based BMR (DT-BMR) mechanism, such as a maximum-likelihood based classifier, in an M-MIMO system will bring extremely high computational complexity and fail to fully consider the signal aliasing problem in the M-MIMO system, DT-BMR is not suitable for scenarios with multiple UEs. Compared with DT-BMR, in pattern recognition based BMR (PR-BMR), the main calculations are concentrated in the offline process and can be flexibly deployed in an actual system.

[0004] In the NOMA scenario of the M-MIMO system, different modulation methods are adopted by multiple UEs, resulting in the problem of multi-user interference (MUI). Considering the complexity of estimating the channel matrix in the M-MIMO system, due to its ability to reduce noise sensitivity without channel estimation, independent component analysis (ICA) has the problem of blind signal separation in the M-MIMO system. Existing research has proposed a blind signal separation scheme based on complex fast independent component analysis (CFICA). However, this scheme requires the number of independent UEs in the system as prior information, which can only be obtained after the UEs complete their respective random access processes, and the number of independent UEs may change rapidly over time in the IoT scenario. Therefore, the number of independent UEs is somewhat unknown. Currently, a multi-user shared access (MUSA) scheme has been proposed to solve the multi-user interference (MUI) problem in the case of an unknown number of UEs in the NOMA scenario. However, this scheme only considers the design of non-orthogonal complex spreading codes and fails to consider the problem of signal aliasing of multiple modulation methods. After separating the overlapping signals, PR-BMR can be performed for each UE signal based on the separated signals. Among them, PR-BMR is divided into two steps: feature extraction and classification. In modulation classification, it is generally done manually. However, in some special cases, especially in complex situations, relying on manual operators for modulation classification has the problem of low accuracy in terms of speed, accuracy, and stability. Summary of the Invention

[0005] To overcome the above defects in the BMR technology of the MIMO system, such as high computational complexity, excessive dependence on the number of independent UEs, failure to consider signal aliasing of multiple modulation methods, and low accuracy of modulation classification, the present invention provides a method and system for blind identification of modulation methods in a large-scale multi-antenna communication system.

[0006] To solve the above technical problems, the technical solution of the present invention is as follows:

[0007] A method for blind identification of modulation methods in a large-scale multi-antenna communication system, applied to an M-MIMO system, where the M-MIMO system includes N U user equipment UEs with transmit antennas and a base station BS equipped with M antennas.

[0008] The method includes the following steps:

[0009] S1. For the N received by the base station BS TIndependent UE number estimation is performed on N signal samples based on the MDL algorithm to obtain the estimated value of the independent UE number corresponding to any signal sample where t = 1, 2, ..., N T ;

[0010] S2. According to the estimated value of the independent UE number, perform mixed signal separation on the N signal samples received by the base station BS based on the CFICA algorithm to obtain the UE transmission signal matrix corresponding to any signal sample T where

[0011] S3. Perform cyclic stationary feature extraction on any of the UE transmission signal matrices to construct N CF matrices T where

[0012] S4. Input the CF matrix into a preset CDBN model for training to obtain the CDBN model; wherein, the CDBN model includes a plurality of maximally pooled CRBM blocks arranged in a stack and a softmax layer for generating an output; each CRBM block includes a visible layer, a detection layer, and a pooling layer

[0013] S5. Based on the CDBN model completed in training in S4, perform operations S1 - S3 on the signal data to be blindly identified, input the obtained CF matrix into the CDBN model completed in training, and output the blind recognition result of the modulation mode

[0014] Furthermore, the present invention also proposes a blind recognition system for the modulation mode of a large-scale multi-antenna communication system, which applies the blind recognition method for the modulation mode of the large-scale multi-antenna communication system proposed in the above technical solution

[0015] The blind recognition system for the modulation mode of the large-scale multi-antenna communication system proposed by the present invention is applied to an M-MIMO system, and the M-MIMO system includes N U user equipment UEs with transmitting antennas and a base station BS equipped with M antennas. The system includes:

[0016] An independent UE number estimation module for performing independent UE number estimation on N signal samples received by the base station BS based on the MDL algorithm to obtain the estimated value of the independent UE number corresponding to any signal sample T where t = 1, 2, ..., N T ;

[0017] A mixed signal separation module for performing mixed signal separation on N signal samples received by the base station BS based on the CFICA algorithm T ​​Perform aliased signal separation on a signal sample to obtain the UE transmission signal matrix corresponding to any signal sample Among them,

[0018] The cyclic stationary feature extraction module is used to perform cyclic stationary feature extraction on any of the UE transmission signal matrices to construct N T CF matrices

[0019] The blind recognition module, which includes a trained CDBN model, is used to perform blind recognition of the modulation method on the input CF matrix and output the blind recognition result of the modulation method;

[0020] Among them, the CDBN model includes several stacked CRBM blocks with max pooling, and a softmax layer for generating the output; each CRBM block includes a visible layer, a detection layer, and a pooling layer.

[0021] Compared with the prior art, the beneficial effects of the technical solution of the present invention are:

[0022] (1) By using the MDL algorithm to assist complex fast independent component analysis, the present invention can estimate the number of antennas of the transmitting device without significant performance loss, realizing true "blind separation";

[0023] (2) Considering the complex characteristics of the signal and the requirement of low system delay, the present invention performs aliased signal separation on the signal sample based on the CFICA algorithm without the need for channel matrix estimation, significantly improving the adaptability to the actual environment;

[0024] (3) The present invention uses the CDBN model based on the CF matrix for BMR, thereby improving the recognition accuracy, enhancing the robustness of the CCMM-BMR scheme, and effectively reducing the computational complexity of feature transformation in the CDBN model. Description of the Drawings

[0025] Figure 1 It is the architecture diagram of the large-scale multi-antenna communication system in Embodiment 1.

[0026] Figure 2 It is the flowchart of the modulation method blind recognition method for the large-scale multi-antenna communication system in Embodiment 1.

[0027] Figure 3 It is the architecture diagram of the CSDN model in Embodiment 1.

[0028] Figure 4 It is the schematic diagram of the CRBM block in Embodiment 1.

[0029] Figure 5 Schematic diagram of the MDL-CFICA performance for the simulation experiment of Example 2.

[0030] Figure 6 Schematic diagram of the MTER performance for the simulation experiment of Example 2.

[0031] Figure 7 Schematic diagram of the confusion matrix performance for the simulation experiment of Example 2.

[0032] Figure 8 Architecture diagram of the modulation mode blind recognition system for the large-scale multi-antenna communication system of Example 3. Detailed implementation manner

[0033] The accompanying drawings are only for illustrative purposes and should not be construed as limitations on this patent;

[0034] To better illustrate this embodiment, some components in the accompanying drawings will be omitted, enlarged or reduced, which do not represent the dimensions of the actual product;

[0035] For those skilled in the art, it is understandable that some well-known structures and their descriptions in the accompanying drawings may be omitted.

[0036] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0037] This embodiment shows a system model for the multi-UE scenario of an M-MIMO system, which includes N U user equipments UE with transmitting antennas, and a base station BS equipped with M antennas.

[0038] As Figure 1 shown, it is the architecture diagram of the M-MIMO system of this embodiment.

[0039] Considering the uplink transmission process, let N T define the number of collected signal samples. For the t-th signal sample, t = 1,..., N0, considering scenarios such as multi-UE and unscheduled access (commonly found in CRN), it is assumed that in the t-th sampling, only N (t) UEs send signals to the BS simultaneously, N (t) < N U << M.

[0040] For the i-th independent UE signal in the t-th signal sample Assume that the elements in the signal vector are independent and identically distributed, and have the same average transmit power, and use to represent the transmit signal corresponding to the i-th UE in the t-th signal sample adopting the modulation mode. Note that, The modulation methods all belong to N M sets of linear memoryless modulation methods, that is In addition, define to have a length of 1×S, then the t-th received signal sample r (t) can be expressed as:

[0041] r (t) = G (t) s (t) + n (t) (1)

[0042] where represents a transmission signal matrix of size N (t) ×S, is a complex Gaussian white noise matrix of size M×S, whose mean and variance are 0 and G (t) is an M×N (t) matrix, which represents the flat fast fading channel response.

[0043] Embodiment 1

[0044] This embodiment proposes a blind identification method for the modulation method of a large-scale multi-antenna communication system, as Figure 2 shown, which is the flowchart of the blind identification method for the modulation method of the large-scale multi-antenna communication system in this embodiment.

[0045] The blind identification method for the modulation method of the large-scale multi-antenna communication system proposed in this embodiment includes the following steps:

[0046] S1. Estimate the number of independent UEs for the N T signal samples received by the base station BS based on the MDL algorithm, and obtain the estimated value of the number of independent UEs corresponding to any signal sample where t = 1, 2,..., N T ;

[0047] S2. According to the estimated value of the number of independent UEs, separate the aliased signals for the N T signal samples received by the base station BS based on the CFICA algorithm, and obtain the UE transmission signal matrix corresponding to any signal sample where

[0048] S3. Extract the cyclic stationary characteristics of any of the UE transmission signal matrices and construct N T CF matrices

[0049] S4. The CF matrix Input the preset CDBN model for training to obtain the CDBN model;

[0050] S5. Based on the CDBN model completed in training in S4, perform operations S1 - S3 on the signal data to be blindly identified, input the obtained CF matrix into the CDBN model completed in training, and output the blind recognition result of the modulation mode.

[0051] In this embodiment, the CDBN model includes several stacked CRBM blocks with max - pooling, and a softmax layer for generating the output; each CRBM block includes a visible layer, a detection layer, and a pooling layer.

[0052] In this embodiment, in the traditional ICA - based signal separation algorithm, the number of independent UEs (number of independent signal sources), as a basic variable, is usually assumed to be a known parameter. In a multi - UE scenario, considering cases such as UE unscheduled access, it is very difficult to obtain the number of UEs in advance. To eliminate the influence of the unrealistic condition of "assuming the number of UEs is known" on signal separation, and thus improve the flexibility of scheme deployment. This embodiment estimates the number of independent UEs for the N T signal samples received by the base station BS based on the Minimum Description Length (MDL) algorithm of information theory.

[0053] In an alternative embodiment, the steps for estimating the number of independent UEs for the N T signal samples received by the base station BS include:

[0054] S1.1. For the t - th signal sample, calculate the autocorrelation matrix R (t) of the received signal r A ; its expression is as follows:

[0055] R A = E{(r (t) ) H} (2)

[0056] where (·) H represents the Hermitian transformation of the matrix, and E{·} represents taking the expectation;

[0057] S1.2. Perform eigenvalue decomposition on the autocorrelation matrix R A , and arrange its M eigenvalues in descending order as:

[0058] λ = [λ1,..., λ M (3)

[0059] S1.3. Use the MDL algorithm to estimate the number of independent UEs to obtain the estimated value of the number of independent UEs, and its expression is as follows:

[0060]

[0061] wherein, n ∈ {1, …, M - 1}, λ i represents the i-th eigenvalue in Equation (3), and S is the signal length.

[0062] Based on Equation (4), the estimated value of the number of independent UEs estimated by the MDL algorithm can be obtained, and then the aliased M-MIMO signals can be separated.

[0063] In this embodiment, considering the linear complex-valued characteristics of the independent UE source and the aliased received signal, based on the Complex Fast Independent Component Analysis (CFICA) algorithm, N T signal samples received by the base station BS are separated from the aliased signals.

[0064] Further, in an optional embodiment, in the step S2, it further includes whitening preprocessing the received signal for further aliased signal separation. The specific steps include:

[0065] S2.1.1. Select the first eigenvalues from the sorted eigenvalue set λ to construct the whitening matrix V W ; its expression is as follows:

[0066]

[0067] wherein, D W is a diagonal matrix, and its diagonal elements are composed of the first eigenvalues in the eigenvalue set λ; the matrix F W is composed of the eigenvectors of size M×1 corresponding to the first eigenvalues; is the noise variance, which is estimated by the mean of the remaining eigenvalues in the eigenvalue set λ; I is the identity matrix of size .

[0068] S2.1.2. According to the whitening matrix V W whiten the received signal r (t) ; its expression is as follows:

[0069]

[0070] wherein, is the received signal after whitening preprocessing, with a size of

[0071] In this embodiment, the objectives of whitening preprocessing of the aliased signals include removing the correlation between independent signal components and normalizing the variance; and reducing the complexity of the CFICA algorithm by extracting the main components of r (t) The main components of

[0072] Since Whitening preprocessing of the signals can effectively reduce the complexity of the subsequent CFICA aliased signal separation process.

[0073] Further, the steps of separating the aliased signals from the N T signal samples received by the base station BS include:

[0074] S2.2.1. For the t-th signal sample, construct a separation matrix with a size of

[0075] where, for w C,i is the i-th column complex weight vector of the matrix W C with a size of

[0076] S2.2.2. Through iterative updating of the separation matrix W C using a single-unit algorithm, it is used to separate the aliased signals from the whitened preprocessed received signals to obtain the transmitted signal matrix of the UEs, with a size of

[0077] The transmitted signal matrix is expressed as follows:

[0078]

[0079] Thus, the signal separation problem can be transformed into the problem of finding the optimal W C In this embodiment, the separation matrix W C is updated using a single-unit algorithm, then there is:

[0080]

[0081] In the formula, (·) + represents the pseudo-inverse, (·) * represents the complex conjugate, G' C (·) and G" C "(·) are respectively G C (y) =

[0082] The first and second derivatives of log2(a + y), where a is the parameter of the smoothing function G C (y), y is the independent variable of the function, and ‖·‖ represents the Euclidean norm.

[0083] Through the iterative processes of Equation (8) and Equation (9), the aliased signals can be separated using Equation (7).

[0084] In step S3 of this embodiment, considering the influence of noise and other interferences in the actual channel, the received signals at the receiving end are generally non-stationary signals, but their correlation functions exhibit cyclostationarity (i.e., cyclic stationarity). Compared with the spectrum and power spectrum, the cyclic spectrum has certain potential in BMR because it can reflect more signal characteristic information. Specifically:

[0085] 1) For most signals, the differences in the cyclic spectrum characteristics of different types of signals are relatively obvious;

[0086] 2) Different from the modulated signals, the noise in the channel usually does not have the characteristic of cyclic stationarity, which makes the cyclic spectrum have strong robustness in anti-noise.

[0087] Therefore, this embodiment utilizes the characteristics of the cyclic spectrum and further completes BMR by constructing a CF matrix as the input of the CDBN.

[0088] In an optional embodiment, in step S3, for any one of the UE transmitted signal matrices The steps for extracting the cyclic stationary characteristics include:

[0089] Construct the corresponding CF matrix according to the UE transmitted signal matrix corresponding to the t-th signal sample Among them, the size of the CF matrix is N V ×N V , and the parameters are {α, f}. The parameters α and f represent the cyclic frequency and digital frequency respectively, and correspond to the rows and columns of the CF matrix ; based on the elements in the UE transmitted signal matrix The elements in the CF matrix are expressed as:

[0090]

[0091] Among them, for each CF matrix corresponds to an independent modulation method

[0092] It should be noted that each Each corresponds to an independent modulation method Therefore, for the training dataset can be used as the input of the CDBN for training the CDBN model.

[0093] In step S4 of this embodiment, the CDBN model is used to solve the BMR problem.

[0094] Among them, CDBN is a popular DL technology in the field of image recognition and processing, which is specifically used to analyze the correlation of pixels in images. And the correlation characteristics shown based on the CF matrix seem to be similar to the correlation characteristics shown by the pixels in the image. In addition, different from other classic networks such as CNN and recurrent neural networks, the advantage of the CDBN model is that even if there is a lack of labeled data during the training process, it can still be trained through an unsupervised pre-training step. This means that in extreme cases, if the labeled data is unavailable or insufficient, CDBN can still be initially trained.

[0095] As Figure 3 shown, it is the architecture diagram of the CDBN model of this embodiment. The CDBN model of this embodiment can be regarded as a network stacked by N R max-pooling CRBM blocks. Each CRBM block includes a visible layer V μ 、a detection layer D μ and a pooling layer P μ , μ = 1…, N R , N R is the number of CRBM blocks.

[0096] Among them, the input V of the CDBN model is an N V ×N V matrix which is also the input of the first CRBM block. In addition, the output layer O of the CDBN is a vector, and each element corresponds to a unit. Among them, the i-th unit represents the probability of being classified into the i-th class.

[0097] In an alternative embodiment, as Figure 4 shown, it is the schematic diagram of the CRBM block of this embodiment.

[0098] In the CDBN model, for the visible layer V μ , detection layer D μ and pooling layer P μ in the μ-th CRBM block:

[0099] The visible layer V μ contains U μ units of size N V,μ×N V,μ matrix

[0100] The detection layer D μ contains K μ matrices of size N D,μ ×N D,μ matrix

[0101] The pooling layer P μ contains K μ matrices of size N P,μ ×N P,μ matrix

[0102] Among them, the matrix and the matrix are expressed as:

[0103]

[0104] In the formula, is the k-th weight matrix of size N connecting and W,μ ×N W,μ , N W,μ = N V,μ - N D,μ + 1; is to horizontally and vertically flip the matrix ; is the bias matrix of size N μ in the detection layer D D,μ ×N D,μ , with elements ; c μ is the bias matrix of size N μ in the visual layer V V,μ ×N V,μ , with elements c μ ; * represents the convolution operation.

[0105] In this embodiment, all units in the matrix share the same and All units in the visual layer V μ share the same c μ . The weight matrix bias matrix and the bias matrix c μ are network parameters to be optimized.

[0106] It can be seen from equations (12) and (13) that and cμ Depending on multiple coupled variables and Therefore, it is necessary to and decouple them, and then calculate and c μ the optimization results of

[0107] By means of equations (12) and (13), and are convolved to generate each unit of, until the complete detection layer D is constructed μ .

[0108] Then, the matrix is divided into N B,μ non - overlapping blocks of size N C,μ ×N C,μ . Select the α - th block from the matrix and each is only connected to one unit in the pooling layer group for extracting high - level feature information; where

[0109] When the pooling layer P μ of the μ - th CRBM block is generated, the pooling layer P μ is used for the visible layer V μ+1 of the (μ + 1)-th CRBM block, that is:

[0110]

[0111] where u = k, U μ+1 = K μ and μ = 1,…, N R - 1

[0112] Secondly, as Figure 4 shown, in the CRBM block of this embodiment, by splicing the elements of each column in K μ matrices , the N R pooling layers P R in the N μ CRBM blocks are rearranged into a vector R of size L R ×1,

[0113]

[0114] Then, through a weight matrix of size N M ×L R ​Map the vector R to the output O; generate the output of the CDBN model from the output O through a softmax layer.

[0115] For the above optimizable network parameter weight matrix Bias matrix and bias matrix c μ , in this embodiment, the CDBN model is optionally pre-trained.

[0116] For convenience of description, in this embodiment, the target training parameters of the CDBN model are defined as:

[0117]

[0118] where μ = 1,..., N R , k = 1,..., K μ , u = 1,..., U μ .

[0119] In an alternative embodiment, in step S4, the CF matrix is input into a preset CDBN model for training to obtain a CDBN model. The specific steps include:

[0120] S4.1.1. Compose the CF matrix constructed from signal samples of several known modulation modes and their corresponding modulation modes into a training data set and input it into the CDBN model for training; introduce an unsupervised pre-training process for all CRBM blocks, and optimize the weight matrix bias matrix and bias matrix c μ .

[0121] Specifically, to describe the pre-training process, this embodiment defines the following functions and operation operations:

[0122] 1) The Sigmoid function is defined as:

[0123] 2) The summation function f sum (x) represents the sum of all elements of the given vector / matrix x.

[0124] 3) The Gibbs sampling function is defined as:

[0125]

[0126] where ∈ ∈ (0, 1) is a random number, x i and y i are the i-th elements of vectors x and y respectively.

[0127] Then, in this step, for the weight matrix Bias matrix and bias matrix c μ The specific steps for optimization include:

[0128] For the μ-th CRBM block, its visible layer is represented as:

[0129]

[0130] In the formula, the superscript (·) <j< represents the j-th stage in pre-training, and according to the two-step contrastive divergence algorithm, it is set that j ∈ {0, 1};

[0131] Calculate the probability that each unit in the detection layer D μ is activated Its expression is:

[0132]

[0133] In the formula, σ(·) is the Sigmoid function. Among them, the values of each unit in the detection layer D μ correspond to the elements in the detection matrix .

[0134] According to the probability reconstruct the visible layer in the second stage Its expression is:

[0135]

[0136] In the formula, f GS (·) is the Gibbs sampling function.

[0137] Through formula (18), let j = 1, and in this embodiment, the

[0138] in formula (17) can be calculated according to the probability and the reconstructed visible layer update the target parameters and c μ ; its expression is as follows:

[0139]

[0140] Among them, and c μ are initialized to 0, is initialized by the Gaussian distribution N(0, 0.01), and λ CDBN is the learning rate. In addition, in formula (20) is:

[0141]

[0142] Among them, ζ is the target sparsity.

[0143] According to the target parameters updated by equations (19)-(22) and c μ , calculate the probability that each unit in the α-th pooling layer P μ is activated and use the probability to construct the pooling layer matrix whose expression is as follows:

[0144]

[0145] Using the probability in equation (23) a matrix can be constructed

[0146]

[0147] where α = j P + N P,μ (i P - 1), i P = 1, …, N P,μ and j P = 1, …, N P,μ , which are respectively the row and column indices of the elements in indicating the operation of selecting the α-th block from the matrix . The values of each unit in the pooling layer P μ correspond to the elements in the pooling matrix .

[0148] Substituting equation (24) into (16), the pre-training process of the next CRBM can be carried out. Based on the pooling layer matrix optimize the target parameters and c μ in the next CRBM block until the target parameters and c μ in all CRBM blocks are optimized.

[0149] S4.1.2. Update all CRBM blocks according to the optimized weight matrix bias matrix and bias matrix c μ and further train the CDBN model using the limited-memory Broyden-Fletcher-Goldfarb-Shanno (L-BFGS) algorithm.

[0150] Specifically, the steps of training the CDBN model using the L-BFGS algorithm include:

[0151] For the t-th signal sample, the input to the CDBN model is to obtain the optimized and use N R pooling layer outputs Rearrange to obtain vector R (t) . Among them, each of the CF matrices corresponds to an independent modulation method

[0152] Use the softmax model to generate the output of the output layer of the CDBN, and then obtain the estimation of the modulation method:

[0153]

[0154] where Θ follows a Gaussian distribution and is initialized;

[0155] Furthermore, to improve the convergence speed of the training process, different from the quadratic loss function, this embodiment uses the following loss function:

[0156]

[0157] In the formula, f b (x) is a Boolean function, defined as:

[0158]

[0159] Update the parameter set χ of the CDBN model through the L-BFGS algorithm CDBN ; its expression is as follows:

[0160]

[0161] Subsequently, the updated Θ can be used to calculate Repeat the above steps until J CDBN is less than the preset threshold, and the training of the CDBN model is completed.

[0162] In this embodiment, by assisting complex fast independent component analysis based on the MDL algorithm, the MDL-CFICA algorithm is proposed to separate the signals of independent UEs from the received aliased signals. Specifically, with the help of the MDL algorithm, in the BMR (CDBN aided CF Analysis based Multi-User, Multi-Modulation-Type BMR, CCMM-BMR) scheme with multiple modulation methods, the number of transmitting device antennas can be estimated without significant performance loss, realizing true "blind separation". Since signal separation based on CFICA does not require channel matrix estimation, the adaptability of the CCMM-BMR scheme proposed in this embodiment to the actual environment can be significantly improved. This embodiment further uses CDBN based on the CF matrix for BMR, thereby improving the recognition accuracy and enhancing the robustness of the CCMM-BMR scheme.

[0163] In addition, this embodiment only selects the CF matrix based on SCF as the input for CDBN model training, thereby reducing the complexity of feature transformation and improving the adaptability of the CCMM-BMR scheme to low-latency B5G communication.

[0164] Embodiment 2

[0165] This embodiment applies the blind recognition method of modulation methods in a large-scale multi-antenna communication system proposed in Embodiment 1 to conduct simulation experiments.

[0166] Without loss of generality, this embodiment considers two common modulation methods: Phase Shift Keying (PSK) and Quadrature Amplitude Modulation (QAM).

[0167] This embodiment tests the recognition performance of the CCMM-BMR scheme for 7 different modulation methods, that is, {ξ1,…,ξ7} correspond to Binary Phase Shift Keying (BPSK), Quadrature Phase Shift Keying (QPSK), 8PSK, 8QAM, 16QAM, 32QAM, and 64QAM respectively. At the same time, this embodiment selects the following schemes for performance comparison:

[0168] 1) Solutions based on Deep Learning (DL): The DBN aided CF Analysis based Multi-User, Multi-Modulation-Type BMR (DCMM-BMR) solution and the BPNN aided CF Analysis based Multi-User, Multi-Modulation-Type BMR (BCMM-BMR) solution.

[0169] 2) Non-DL solutions: The HOC based Multi-User, Multi-Modulation-Type BMR (HCMM-BMR) solution.

[0170] In the simulation experiment, the simulation parameters are shown in Table 1 below.

[0171] Table 1 Simulation Parameter Table

[0172]

[0173]

[0174] To evaluate the performance of MDL-CFICA, in this embodiment, the Normalized Deviation Rate of Signal-to-Interference-Noise Ratio (NDR-SINR) is defined as follows:

[0175]

[0176] where and are the actual SINR and the SINR after separation by MDL-CFICA respectively, and N tes1 is the number of test samples.

[0177] In addition, to evaluate the performance of the CCMM-BMR solution, we define the Modulation Type Estimation Error Rate (MTER) as:

[0178]

[0179] where can be obtained from Equation (26).

[0180] Regarding the modulation mode recognition performance of the present invention, based on Equation (29), this embodiment presents the performance of MDL-CFICA under different numbers of UEs and signal-to-noise ratios (SNRs), as Figure 5 shown, which is the schematic diagram of the MDL-CFICA performance of this embodiment. Among them Figure 5 (a) is the schematic diagram of the MDL-CFICA performance when γ = -10 dB, Figure 5 (b) is the schematic diagram of the MDL-CFICA performance when γ = 10 dB.

[0181] It can be seen from the figure that when M = 16 and 64, and SNR γ = -10 dB and 10 dB, due to the multi-antenna gain, the NDR-SINR of the MDL-CFICA scheme decreases as M increases. When γ increases, β decreases somewhat, which is because the estimation accuracy of the MDL module for N (t) increases as γ increases. When N (t) increases, the performance of MDL-CFICA decreases slightly, which is mainly due to the incorrect estimation of N (t) .

[0182] However, compared with CFICA that relies on the prior knowledge N (t) , the MDL-CFICA scheme proposed in this embodiment can achieve a separation performance similar to it when the prior knowledge N (t) is unknown, which illustrates the effectiveness of the MDL-CFICA scheme proposed in this embodiment in realizing signal blind separation.

[0183] Furthermore, this embodiment analyzes the performance of MTER when the number of UEs is N (t) = 4 and 10, and M = 16 and 64, as Figure 6 shown, which is the schematic diagram of the MTER performance of this embodiment, where Figure 6 (a) is the schematic diagram of the MTER performance when M = 16, Figure 6 (b) is the schematic diagram of the MTER performance when M = 64.

[0184] It can be seen from the figure that the proposed CCMM-BMR scheme can obtain higher BMR accuracy compared with other schemes. Compared with the case where N (t) is known, the CCMM-BMR scheme proposed in this embodiment in N (t)Similar MTER performance can be achieved in unknown situations, demonstrating that the CCMM-BMR scheme can eliminate the traditional dependence on prior information. In addition, the MTER performance decreases as γ increases because, in the case of a higher γ, the CFs corresponding to different modulation schemes are more distinguishable, thereby improving the classification accuracy of the CDBN. At the same time, due to the multi-antenna gain, P MTER decreases as M increases. When N (t) is low, P MTER is even lower because the separation effect of MDL-CFICA is better when N (t) is low.

[0185] Furthermore, this embodiment presents the system performance measured by a confusion matrix. As Figure 7 shown, it is a schematic diagram of the confusion matrix performance of this embodiment. Among them, Figure 7 (a) in it is a schematic diagram of the confusion matrix performance when M = 16, Figure 7 (b) in it is a schematic diagram of the confusion matrix performance when M = 64.

[0186] As can be seen from the figure, when N (t) = 4 and M = 64, it can be obtained from the figure that higher-order modulation schemes are more likely to have classification errors because the CFs corresponding to higher-order modulation schemes are more complex than those of lower-order modulation schemes. At the same time, due to the similarity of CFs, misclassification is more likely to occur between modulation schemes of the same order. Thus, it can be seen that the CCMM-BMR scheme proposed in this embodiment is superior to other comparison schemes in the recognition performance of different modulation schemes.

[0187] Embodiment 3

[0188] This embodiment proposes a blind recognition system for modulation schemes in a large-scale multi-antenna communication system, applying the blind recognition method for modulation schemes in a large-scale multi-antenna communication system proposed in Embodiment 1.

[0189] As Figure 8 shown, it is an architecture diagram of the blind recognition system for modulation schemes in a large-scale multi-antenna communication system of this embodiment. In the blind recognition system for modulation schemes in a large-scale multi-antenna communication system proposed in this embodiment, it includes:

[0190] An independent UE number estimation module, which is used to estimate the number of independent UEs for N T signal samples received by the base station BS based on the MDL algorithm, and obtain the estimated value of the number of independent UEs corresponding to any signal sample where t = 1, 2,..., N T ;

[0191] An aliased signal separation module, which is used to separate the N signals received by the base station BS based on the CFICA algorithmT Perform aliased signal separation on a signal sample to obtain a UE transmission signal matrix corresponding to any signal sample Among them,

[0192]

[0193] A cyclic stationary feature extraction module is used to perform cyclic stationary feature extraction on any of the UE transmission signal matrices to construct N T CF matrices

[0194] A blind recognition module, which includes a trained CDBN model, is used to perform blind recognition of the modulation method on the input CF matrix and output the blind recognition result of the modulation method.

[0195] Among them, the CDBN model includes a number of stacked CRBM blocks with max pooling, and a softmax layer for generating the output; each CRBM block includes a visible layer, a detection layer, and a pooling layer.

[0196] The same or similar reference numerals correspond to the same or similar components;

[0197] The terms describing the positional relationship in the drawings are only for illustrative purposes and should not be construed as a limitation of this patent;

[0198] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, rather than limitations on the implementation manners of the present invention. For those of ordinary skill in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to enumerate all the implementation manners here. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention should be included in the protection scope of the claims of the present invention.

Claims

1. A blind identification method for modulation schemes in a large-scale multiple-input multiple-output (M-MIMO) communication system, which is applied to an M-MIMO system, characterized in that, The M-MIMO system includes N U user equipments UE with transmitting antennas, and a base station BS equipped with M antennas; The method includes the following steps: S1. Independently estimate the number of UEs for the N signal samples received by the base station BS based on the MDL algorithm, and obtain the estimated value of the number of independent UEs corresponding to any signal sample T wherein, t = 1, 2,..., N ; T ​ S2. Based on the estimated value of the number of independent UEs, perform mixed signal separation on the N signal samples received by the base station BS using the CFICA algorithm to obtain the UE transmission signal matrix corresponding to any signal sample T wherein the N signal samples S3. Perform cyclic stationary feature extraction on the signal matrix transmitted to any of the UEs, and construct N CF matrices. T S4. Input the CF matrix into a preset CDBN model for training to obtain the CDBN model. Among them, the CDBN model includes a number of stacked CRBM blocks with max pooling, and a softmax layer for generating the output. Each CRBM block includes a visible layer, a detection layer, and a pooling layer. S5. Based on the CDBN model completed in training in S4, perform operations S1 - S3 on the signal data to be blindly identified, input the obtained CF matrix into the CDBN model completed in training, and output the blind recognition result of the modulation mode; Among them, in the S2 step, the following steps are further included: For the t-th signal sample, the received signal r (t) is preprocessed by whitening and further used for aliased signal separation. The specific steps are as follows: S2.1.

1. Select the first eigenvalues from the sorted eigenvalue set λ to construct the whitening matrix V W ; its expression is as follows: where D W is a diagonal matrix, and its diagonal elements are composed of the first eigenvalues in the set of eigenvalues λ; the matrix F W is composed of the eigenvectors of size M×1 corresponding to the first eigenvalues; is the noise variance, which is estimated by the mean of the remaining eigenvalues in the set of eigenvalues λ; I is the identity matrix of size ; S2.1.

2. According to the whitening matrix V W perform whitening preprocessing on the received signal r (t) The expression is as follows: In the formula, is the received signal after whitening preprocessing, with a size of And, Steps for separating the aliased signals from N signal samples received by the base station BS include: T ​ S2.2.

1. For the t-th signal sample, construct a separation matrix of size where w is the i-th column complex weight vector of matrix W C,i with a size of C and the size of the separation matrix is S2.2.

2. By using a single-unit algorithm to iteratively update the separation matrix W C for separating the aliased signals from the received signals that have undergone whitening preprocessing to obtain the transmitted signal matrix of UEs after separation The expression is as follows: where (·) + denotes the pseudo-inverse, (·) * denotes the complex conjugate, G ′ C (·) and G " C ″ (·) are the first-order and second-order derivatives of G C (y) = log2(a + y) respectively, where a is the parameter of the smoothing function G C (y), y is the independent variable of the function, and ‖·‖ represents the Euclidean norm.

2. The blind identification method for modulation schemes in a large-scale multiple-input multiple-output (M-MIMO) communication system according to claim 1, characterized in that, In the step S1, the steps of estimating the number of independent UEs for N signal samples received by the base station BS include: T ​ S1.

1. For the t-th signal sample, calculate the autocorrelation matrix R (t) of the received signal r A ; The expression is as follows: R A = E{(r (t) ) H} where (·) H represents the Hermitian transformation of a matrix, and E{·} represents taking the expectation; S1.

2. Perform eigenvalue decomposition on the autocorrelation matrix R A and arrange its M eigenvalues in descending order as λ = [λ1, …, λ M ; S1.

3. Estimate the number of independent UEs using the MDL algorithm to obtain an estimated value of the number of independent UEs Its expression is as follows: where \(n\in\{1,\ldots,M - 1\}\), \(\lambda\) i represents the \(i\)-th eigenvalue, and \(S\) is the signal length.

3. The blind identification method for modulation schemes in a large-scale multiple-input multiple-output (M-MIMO) communication system according to claim 1, characterized in that, In the step S3, for any of the UE transmission signal matrices the steps of performing cyclic stationary feature extraction include: UE transmission signal matrix corresponding to the t-th signal sample Construct the corresponding CF matrix Among them, the CF matrix has a size of N V ×N V , the parameters are {α, f}, the parameters α and f represent the cyclic frequency and the digital frequency respectively, and correspond to the rows and columns of the CF matrix ; based on the elements in the UE transmission signal matrix The elements in the said CF matrix are expressed as: is expressed as: Among them, for t = 1, …, N T , each CF matrix corresponds to an independent modulation method 4. The blind identification method for modulation schemes in a large-scale multiple-input multiple-output (M-MIMO) communication system according to any one of claims 1 to 3, characterized in that, In the CDBN model, for the visible layer V in the μ-th CRBM block μ , the detection layer D μ and the pooling layer P μ : The visual layer V μ includes U μ matrices of size N V,μ ×N V,μ ​ The detection layer D μ contains K μ matrices of size N D,μ ×N D,μ ​ The pooling layer P μ includes K μ matrices of size N P,μ ×N P,μ ​ where μ = 1…, N R , N R is the number of CRBM blocks; the matrix and the matrix are expressed as: wherein, is the k-th weight matrix of size N and with dimensions N W,μ ×N W,μ , where N W,μ = N V,μ - N D,μ + 1; is to horizontally and vertically flip the matrix ; is the detection layer D μ with a bias matrix of size N D,μ ×N D,μ and elements ; c μ is the visible layer V μ with a bias matrix of size N V,μ ×N V,μ and elements c μ ; * represents the convolution operation; The matrix All cells in share the same and all cells in the visual layer V μ share the same c μ ; The weight matrix the bias matrix and the bias matrix c μ are network parameters to be optimized; Divide the matrix into N B,μ non - overlapping blocks of size N C,μ ×N C,μ , select the α - th block from the matrix and each is only connected to one cell in the pooling layer group ; where ​ When the pooling layer P of the μ-th CRBM block μ is generated, the pooling layer P μ is used as the visible layer V of the (μ + 1)-th CRBM block μ+1 , that is: where u = k, U μ+1 = K μ and μ = 1, …, N R - 1; In the CRBM block, by splicing K μ Matrix The elements of each column in N R N of the CRBM blocks R Pooling layer P μ Rearrange to size L R ×1 vector R, Then, through a size N M ×L R The weight matrix Map the vector R to the output O; pass the output O through the softmax layer to generate the output of the CDBN model.

5. The blind recognition method for modulation mode of a large-scale multi-antenna communication system according to claim 4, wherein, In the step S4, the CF matrix is input into a preset CDBN model for training. The specific steps for obtaining the CDBN model include: S4.1.

1. Construct a CF matrix composed of signal samples of several known modulation methods and their corresponding modulation methods to form a training data set and input it into the CDBN model for training; introduce an unsupervised pre-training process for all CRBM blocks, and optimize the weight matrix bias matrix and bias matrix c μ for optimization; S4.1.

2. Update all CRBM blocks according to the optimized weight matrix bias matrix and bias matrix c μ and train the CDBN model using the limited-memory L-BFGS algorithm.

6. The blind recognition method for modulation mode of a large-scale multi-antenna communication system according to claim 5, wherein, In the step S4.1.1, for the weight matrix bias matrix and the bias matrix c μ The steps for optimization include: For the μ-th CRBM block, its visible layer is expressed as: Wherein, the superscript (·) <j>< / j> represents the j-th stage in pre-training, and j ∈ {0, 1}; is the input of the CDBN model; Calculate the probability that each unit in detection layer D μ is activated, and its expression is: In the formula, σ(·) is the Sigmoid function; According to probability Reconstruct the visual layer of the second stage Its expression is: where f GS (·) is the Gibbs sampling function; According to probability and the reconstructed visual layer update the target parameter and c μ The expression is as follows: wherein, and c μ are initialized to 0, is initialized by a Gaussian distribution λ CDBN is the learning rate; ζ is the target sparsity; f sum (·) is the summation function, representing the summation of all elements of a given vector / matrix; According to the updated target parameters and c μ , calculate the probability that each unit in the α-th pooling layer P μ is activated and use the probability to construct a pooling layer matrix whose expression is as follows: The expression is as follows: where α = j P + N P,μ (i P - 1), i P = 1, …, N P,μ and j P = 1, …, N P,μ are respectively the indices of the row and column to which the element in denotes the operation of selecting the α-th block from the matrix ; Based on the pooling layer matrix optimize the target parameters and c μ in the next CRBM block until the target parameters and c μ in all CRBM blocks are optimized.

7. The blind recognition method for modulation mode of a large-scale multi-antenna communication system according to claim 5, wherein, In the S4.1.2 step, the steps of training the CDBN model using the limited-storage L-BFGS algorithm include: For the t-th signal sample, the input to the CDBN model is Obtain the optimized And use N R Pooling layer outputs Rearrange to obtain the vector R (t) ; Among them, each of the said CF matrices corresponds to an independent modulation method Generate the output of the output layer of the CDBN using the softmax model, and then obtain the estimation of the modulation mode: where Θ is initialized from a Gaussian distribution Initialize; Calculate the loss function for the CDBN model, and its expression is as follows: where f b (x) is a Boolean function; Update the parameter set χ of the CDBN model by the L-BFGS algorithm CDBN ; Its expression is as follows: Repeat the above steps until J CDBN Less than the preset threshold value, and complete the training of the CDBN model.

8. A blind recognition system for modulation methods in a large-scale multi-antenna communication system, applying the blind recognition method for modulation methods in a large-scale multi-antenna communication system according to any one of claims 1 to 7, characterized in that: The system is applied to an M-MIMO system, characterized in that the M-MIMO system includes N U user equipments UE with transmitting antennas, and a base station BS equipped with M antennas; The system includes: An independent UE number estimation module, which is used to estimate the number of independent UEs for N T signal samples received by the base station BS based on the MDL algorithm, and obtain the estimated value of the number of independent UEs corresponding to any signal sample where t = 1, 2,..., N T ; An aliased signal separation module, which is used to separate the aliased signals of N signal samples received by the base station BS based on the CFICA algorithm, so as to obtain the UE transmission signal matrix corresponding to any signal sample T wherein, N is Among them, A cyclic stationary feature extraction module for performing cyclic stationary feature extraction on any of the UE transmission signal matrices to construct N T CF matrices Blind recognition module, including a trained CDBN model, for the input CF matrix to perform blind recognition of modulation mode and output the blind recognition result of modulation mode; Among them, the CDBN model includes a number of stacked CRBM blocks with max-pooling, and a softmax layer for generating the output; each CRBM block includes a visible layer, a detection layer, and a pooling layer.

Images