A signal detection method and system for large-scale MIMO system

Abstract

The present application provides a signal detection method and system for a large-scale MIMO system, which determines a first damping coefficient and a second damping coefficient according to the number of antennas at the transmitting end and the number of antennas at the receiving end; determines the initial mean of the posterior probability distribution of the transmitted signal according to the received signal, the channel matrix, the average symbol energy and the first damping coefficient; uses the SORI method to perform SORI iteration on the mean of the posterior probability distribution of the transmitted signal, and determines the initial mean of the Gaussian distribution obeyed by the cavity edge probability according to the mean vector after the SORI iteration; uses the EPA iteration method to iteratively update the mean of the Gaussian distribution obeyed by the cavity edge probability, and uses the mean of the Gaussian distribution obeyed by the cavity edge probability that reaches the preset number of EPA iterations as the estimated value of the transmitted signal. SORI and EPA algorithms are combined, and SORI is used to approximate the matrix inversion of the initialization part of the EPA algorithm, thereby greatly improving the convergence speed.

Description

Technical Field

[0001] The present application relates to the field of mobile communication technology, and in particular to a signal detection method and system for a large-scale MIMO system. Background Art

[0002] As a solution for multi-user communication, massive multiple-input multiple-output (MIMO) technology has become one of the key technologies of 5G communication systems. Compared with traditional small-scale MIMO systems, massive antenna systems have great advantages in spectrum efficiency, energy efficiency and connection reliability due to the increase in the number of antennas at the receiving and transmitting ends. However, the large increase in the number of antennas also makes signal detection at the receiving end a huge challenge.

[0003] Current signal detection methods are mainly divided into linear detection and nonlinear detection. Traditional linear detection requires matrix inversion operations, which is very complex. In order to reduce the complexity, the prior art discloses some methods based on iterative approximation of matrix inversion, such as the second-order Richardson iteration (SORI) and the Newman series approximation (NSA). However, with the increase in the number of antennas, the performance of linear detection and their approximate algorithms has reached saturation. Nonlinear detection algorithms can often achieve better detection performance than linear detection, especially the expected propagation (EP) algorithm can achieve near-optimal detection performance.

[0004] Each iteration of the traditional EP algorithm requires a matrix inversion operation, and the extremely high complexity makes it difficult to implement in hardware. Prior art 1 (X. Tan, Y.-L. Ueng, Z. Zhang, X. You and C. Zhang, “A low complexity massive MIMO detection based on approximate expectation propagation,” IEEE Trans. Veh. Technol., vol. 68, no. 8, pp. 7260–7272, Aug. 2019) proposed an approximate EP algorithm (EPA), which simplifies the EP algorithm by reconstruction and approximation, while eliminating the matrix inversion in each iteration, but retains an initial matrix inversion operation. Prior art 2 (X.Tan et al:, “Approximate expectation propagation massive MIMO detector with weighted Neumann-series,” IEEE Trans.CircuitsSyst.II:Exp.Briefs,vol.68,no.2,pp.662–666,Feb.2021.) proposed an EPA-wNSA algorithm based on EPA, which approximated the initial matrix inversion through iterative operations of weighted Neumann sequence (wNSA), and further optimized the iterative process of EPA, thereby reducing the complexity.

[0005] However, the EPA-wNSA algorithm still has some room for improvement. Its slow convergence speed is not conducive to high-throughput hardware implementation, and the initialization part of the algorithm and the required division operations also consume a lot of computing resources. Summary of the invention

[0006] The present application provides a signal detection method and system for a large-scale MIMO system to achieve faster convergence speed and lower complexity signal detection.

[0007] The first aspect of the present application provides a signal detection method for a large-scale MIMO system, comprising:

[0008] Obtain the number of antennas at the transmitting end, the number of antennas at the receiving end, the received signal, the channel matrix, the average symbol energy and the noise variance;

[0009] Determining a first damping coefficient and a second damping coefficient of a signal detector at a receiving end according to the number of antennas at the transmitting end and the number of antennas at the receiving end;

[0010] Determine an initial mean of a posterior probability distribution of a transmitted signal, an inverse covariance matrix, and a matched filtering output of the received signal according to the received signal, the channel matrix, the noise variance, the average symbol energy, and the first damping coefficient;

[0011] According to the initial mean of the posterior probability distribution of the transmitted signal, the matched filter output, the inverse covariance matrix, the first damping coefficient and the second damping coefficient, SORI iteration is performed on the mean of the posterior probability distribution of the transmitted signal to obtain a mean vector of the posterior probability distribution of the transmitted signal after the SORI iteration is completed, and, according to the mean vector of the posterior probability distribution of the transmitted signal after the SORI iteration is completed, an initial mean of the Gaussian distribution obeyed by the cavity edge probability is determined;

[0012] The variance of the posterior probability distribution of the transmitted signal is obtained according to the inverse covariance matrix, and based on the variance of the posterior probability distribution and the initial mean of the Gaussian distribution obeyed by the cavity edge probability, the EPA iterative method is adopted to iteratively update the mean of the Gaussian distribution obeyed by the cavity edge probability, and the mean of the Gaussian distribution obeyed by the cavity edge probability that reaches a preset number of EPA iterations is used as the estimated value of the transmitted signal.

[0013] Optionally, the first damping coefficient and the second damping coefficient of the signal detector at the receiving end are determined according to the number of antennas at the transmitting end and the number of antennas at the receiving end, and the following model is adopted:

[0014]

[0015] δ=2 / (λ1+λ2);

[0016]

[0017] Where K and N are the number of antennas at the transmitting end and the receiving end respectively, δ is the first damping coefficient, and ω is the second damping coefficient.

[0018] Optionally, determining an initial mean of a posterior probability distribution of a transmitted signal, an inverse covariance matrix, and a matched filtering output of a received signal according to the received signal, the channel matrix, the noise variance, the average symbol energy, and the first damping coefficient comprises:

[0019] The process of determining the inverse covariance matrix based on the channel matrix and noise variance adopts the following model:

[0020] W=H H H+diag′(Λ i );

[0021] Where W is the inverse covariance matrix, H is the channel matrix, diag′(·) represents a diagonal matrix consisting of the diagonal elements of a vector or a matrix, and Λ i To set the parameters, i represents the signal dimension of the transmitted signal, i.e. the serial number of the transmitting antenna. is the noise variance, E s is the average symbol energy;

[0022] The process of determining the matched filter output of the received signal based on the channel matrix and the received signal adopts the following model:

[0023]

[0024] Where B is the matched filter output, and y is the received signal vector;

[0025] The mean of the posterior probability distribution is initialized according to the matched filter output and the first damping coefficient to obtain the initial mean of the posterior probability distribution, which includes the mean of the posterior probability distribution of the first iteration and the mean of the posterior probability distribution of the second iteration, and is expressed as:

[0026] μ 0 =0,μ 1 = δB;

[0027] Among them, μ 0 is the mean of the posterior probability distribution of the first iteration, μ 1 is the mean of the posterior probability distribution of the second iteration.

[0028] Optionally, according to the initial mean of the posterior probability distribution of the transmitted signal, the matched filter output, the inverse covariance matrix, the first damping coefficient and the second damping coefficient, SORI iteration is performed on the mean of the posterior probability distribution of the transmitted signal to obtain a mean vector of the posterior probability distribution of the transmitted signal after the SORI iteration is completed, and the following model is adopted:

[0029]

[0030] In the formula, the preset number of SORI iterations is L1-1, then the mean vector of the posterior probability distribution of the transmitted signal after completing the SORI iteration is

[0031] Optionally, the initial mean of the Gaussian distribution obeyed by the cavity edge probability is determined according to the mean vector of the posterior probability distribution of the transmitted signal after completing the SORI iteration, using the following model:

[0032]

[0033] Where, t 0is the initial mean of the Gaussian distribution obeyed by the cavity edge probability.

[0034] Optionally, the EPA iteration method is used based on the variance of the posterior probability distribution and the initial mean of the Gaussian distribution obeyed by the cavity edge probability to iteratively update the mean of the Gaussian distribution obeyed by the cavity edge probability, and the mean of the Gaussian distribution obeyed by the cavity edge probability that reaches a preset number of EPA iterations is used as the estimated value of the transmitted signal, including:

[0035] The mean of the Gaussian distribution of the cavity edge probability in the i-th dimension Make a hard decision and get the estimated symbol vector of the current iteration The following model is used:

[0036]

[0037] In the formula, l2 represents the current number of iterations, is a constellation point in the constellation set Θ, which is determined by the modulation order and the modulation mode, Indicates that in the l2th iteration, The smallest hour The value of is the mean of the Gaussian distribution of the cavity edge probability updated in the last iteration;

[0038] Update the mean of the Gaussian distribution of the cavity edge probability of the i-th dimension The following model is used:

[0039]

[0040] In the formula, Gram matrix A = H H H, is the variance of the posterior probability distribution of the i-th dimension, B i The output of the matched filter The element of the i-th row of ;

[0041] Use the scaling factor β (0<β<1) to Scaling is performed using the following model:

[0042]

[0043] The number of iterations is preset, and the mean of the Gaussian distribution obeyed by the cavity edge probability reaching the preset number of iterations is obtained, and the mean of the Gaussian distribution obeyed by the cavity edge probability reaching the preset number of iterations is used as the estimated value of the transmitted signal.

[0044] The second aspect of the present application provides a signal detection system for a large-scale MIMO system, which is used to perform the signal detection method provided in the first aspect of the present application, including an input unit, a preprocessing unit, a SORI unit and an EPA unit;

[0045] The input unit is used to obtain the number of antennas at the transmitting end, the number of antennas at the receiving end, the received signal, the channel matrix, the average symbol energy and the noise variance, determine the first damping coefficient and the second damping coefficient of the signal detector at the receiving end, and input the received signal, the channel matrix, the average symbol energy and the noise variance, the first damping coefficient and the second damping coefficient to the preprocessing unit, the SORI unit and the EPA unit;

[0046] The preprocessing unit is used to determine the initial mean and inverse covariance matrix of the posterior probability distribution of the transmitted signal according to the received signal, the channel matrix, the noise variance, the average symbol energy and the first damping coefficient;

[0047] The SORI unit is used to obtain a mean vector of the posterior probability distribution of the transmitted signal after the SORI iteration is completed based on the SORI iteration method, and determine the initial mean of the Gaussian distribution obeyed by the cavity edge probability according to the mean vector of the approximate posterior probability distribution of the transmitted signal after the SORI iteration is completed;

[0048] The EPA unit is used to iteratively update the mean of the Gaussian distribution obeyed by the cavity edge probability based on the EPA iteration method and the initial mean of the Gaussian distribution obeyed by the cavity edge probability, and use the mean of the Gaussian distribution obeyed by the cavity edge probability that reaches a preset number of EPA iterations as the estimated value of the transmitted signal.

[0049] Optionally, the pre-processing unit is configured with a diagonal-based systolic array, the systolic array includes K(K+1) / 2 processing units PE, and the processing unit PE processes multiple groups of data in each clock cycle;

[0050] The systolic array is used to receive input data including a channel matrix H and a received signal vector y, and output H H H and H H y; the processing unit PE includes a PE-A unit for calculating conjugation and a PE-A top Unit, calculation H H The PE-B unit and PE-B in the diagonal part of H top Unit, calculation H H y's PE-D unit and PE-B Top Unit and calculation H H PE-C units and PE-D units in the non-diagonal part of H;

[0051] The systolic array uses reverse data flow, and the input data is input from the long side of the systolic array to the PE-A unit and the PE-A top Unit, get H H ; The channel matrix H flows along the direction of the array from the long side of the systolic array to the side of the systolic array, and the H H Flows from the long side of the systolic array to the other side of the systolic array.

[0052] Optionally, the pre-processing unit uses a systolic array to calculate matrix-matrix or matrix-vector multiplication;

[0053] The systolic array is configured with an advance addition module for pre-processing a complex multiplication operation, wherein the complex multiplication operation is a multiplication of a first multiplier and a second multiplier, the first multiplier is represented by a+jb, and the second multiplier is represented by c+jd;

[0054] The advance addition module includes a first addition unit and a second addition unit, the first addition unit is used to calculate a+b, and the second addition unit is used to calculate c+d; the advance addition module is used to transfer the output of the first addition unit and the output of the second addition unit to the processing unit PE.

[0055] Optionally, the pre-processing unit uses a systolic array to calculate matrix-matrix or matrix-vector multiplication; the systolic array is configured with a rounding module, and the rounding module is used to round off input data to reduce the bit width.

[0056] It can be seen from the above technical solutions that the signal detection method and system for large-scale MIMO systems provided by the present application combine the SORI and EPA algorithms, and use SORI to approximate the matrix inversion of the initialization part of the EPA algorithm, thereby greatly improving the convergence speed, and can reduce the number of iterations by about 50% to 67% compared to EPA-wNSA. Secondly, the complexity of the initialization part is reduced by algorithm conversion, and the division operation between two iterations is eliminated. In addition, the signal detection system provided by the present application adopts a high-throughput hardware architecture, which includes a diagonal-based systolic array, which is used to calculate matrix-matrix multiplication and matrix-vector multiplication, and contains three improved methods that can be used optionally, including the use of reverse data flow, which can significantly reduce the operation time compared to the systolic array used in other MIMO detection; configure an advance addition unit, which can shorten the critical path of the systolic array when performing complex multiplication operations; configure a rounding unit, which can reduce the bit width of the input data and thus reduce hardware resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 A schematic diagram of the structure of a signal detection system provided in an embodiment of the present application;

[0058] Figure 2 A schematic diagram of the structure of a systolic array used in a pre-processing unit provided in an embodiment of the present application;

[0059] Figure 3 A schematic diagram of the timing arrangement of the signal detection system provided in an embodiment of the present application. DETAILED DESCRIPTION

[0060] The technical solution of the present application will be described in detail below with reference to the accompanying drawings.

[0061] See also Figure 1 The present application provides a signal detection method for a large-scale MIMO system, comprising steps S1 to S5.

[0062] S1. Obtain the number of antennas at the transmitting end, the number of antennas at the receiving end, the received signal, the channel matrix, the average symbol energy and the noise variance.

[0063] In the uplink of the massive MIMO system, the number of antennas at the transmitting and receiving ends is set to K and N respectively, and the transmitted signal adopts M-order quadrature amplitude modulation (M-QAM). Assume that the transmitted signal is represented by a vector, that is, the transmitted signal vector x is taken from the constellation set Θ, and the average symbol energy is E s , the vector representation of the received signal, that is, the received signal vector is y, then the MIMO system model can be expressed as y = Hx + n, where H is the Rayleigh flat fading channel matrix, n represents the mean is 0, and the variance is The MIMO detection task is to estimate the transmitted signal x based on the received signal y by the MIMO signal detector in the receiving base station.

[0064] S2. Determine a first damping coefficient and a second damping coefficient of a signal detector at a receiving end according to the number of antennas at the transmitting end and the number of antennas at the receiving end.

[0065] At the signal receiving end, the damping coefficient required by the receiving end signal detector is calculated in advance according to the number of antennas at the transmitting end and the number of antennas at the receiving end. The calculation method is:

[0066]

[0067] δ=2 / (λ1+λ2);

[0068]

[0069] Where K and N are the number of antennas at the transmitting end and the receiving end respectively, δ is the first damping coefficient, and ω is the second damping coefficient.

[0070] S3. Determine an initial mean and an inverse covariance matrix of a posterior probability distribution of the transmitted signal according to the received signal, the channel matrix, the noise variance, the average symbol energy and the first damping coefficient.

[0071] Since in a large-scale MIMO system, directly solving the posterior probability of each transmitted signal vector will bring great computational complexity, the embodiment of the present application approximates the posterior probability distribution of the transmitted signal vector to a Gaussian distribution. First, a preprocessing operation is performed to determine the initial mean μ and inverse covariance matrix W of the approximate Gaussian distribution of the posterior probability, so as to provide the required variables for subsequent SORI iterations and EPA iterations.

[0072] The inverse covariance matrix used in the EPA-wNSA method of prior art 2 is The parameter Λ is set i The value of i represents the signal dimension of the transmitted signal, i.e., the serial number of the transmitting antenna, and the received signal vector y is subjected to matched filtering to obtain the matched filtering output:

[0073] In order to reduce the multiplication operation, the embodiment of the present application enlarges the inverse covariance matrix W in the prior art 2. Amplify the matched filter output B It is expressed as:

[0074]

[0075]

[0076]

[0077] Wherein, diag′(·) represents a diagonal matrix consisting of the diagonal elements of a vector or a matrix.

[0078] The embodiment of the present application can reduce 4K through the above conversion 2 The multiplication operation.

[0079] The mean μ of the posterior probability distribution is initialized according to the matched filter output B and the first damping coefficient δ. The mean μ is iteratively updated in subsequent SORI iterations and tends to be optimal. The initial mean includes the mean of the posterior probability distribution of the first iteration and the mean of the posterior probability distribution of the second iteration, expressed as:

[0080] μ 0 =0,μ 1 = δB;

[0081] Among them, μ 0 is the mean of the posterior probability distribution of the first iteration, μ 1 is the mean of the posterior probability distribution of the second iteration.

[0082] After obtaining the inverse covariance matrix, the variance σ of the approximate Gaussian distribution of the posterior probability is determined by the inverse covariance matrix. 2 It should be noted that the step of determining the variance of the posterior probability can be performed after obtaining the inverse covariance matrix, or can be performed at the beginning of the EPA iteration process. In the signal detection method of the embodiment of the present application, the variance can be calculated once. The diagonal matrix D formed by the diagonal of the inverse covariance matrix W can be used to obtain the variance of the posterior probability distribution:

[0083] σ 2 =diag(D -1 );

[0084] S4. Obtain the mean vector of the posterior probability distribution of the transmitted signal after the SORI iteration based on the SORI iteration method, and determine the initial mean of the Gaussian distribution obeyed by the cavity edge probability according to the mean vector of the approximate posterior probability distribution of the transmitted signal after the SORI iteration.

[0085] The EPA-wNSA method of prior art 2 adopts wNSA iteration. The embodiment of the present application adopts SORI iteration to replace wNSA iteration. First, according to the initial mean of the posterior probability distribution of the transmitted signal, the matched filter output, the inverse covariance matrix, the first damping coefficient and the second damping coefficient, the mean of the posterior probability distribution of the transmitted signal is subjected to SORI iteration to obtain the mean vector of the posterior probability distribution of the transmitted signal after the SORI iteration is completed. The following model is adopted:

[0086]

[0087] The number of SORI iterations is preset to L1-1, and the iteration result is

[0088] After obtaining the mean vector of the posterior probability distribution of the transmitted signal after the SORI iteration, the initial mean of the Gaussian distribution obeyed by the cavity edge probability is determined. In the EPA-wNSA method of the prior art 2, the mean t of the Gaussian distribution obeyed by the cavity edge probability is initialized as (diag(·) represents a vector composed of diagonal elements of a matrix). In the embodiment of the present application, by improving step S2, it can be obtained Due to the variance of the noise diag(D -1 )→1 / N, we can deduce With E s The increase of SNR approaches 0 infinitely. Therefore, the embodiment of the present application proposes to set t 0 Approximately It is expressed as:

[0089]

[0090] Through the above operation, 2K divisions, 2K multiplications and 2K addition operations can be reduced.

[0091] S5. Based on the EPA iteration method and the initial mean of the Gaussian distribution obeyed by the cavity edge probability, the mean of the Gaussian distribution obeyed by the cavity edge probability is iteratively updated, and the mean of the Gaussian distribution obeyed by the cavity edge probability that reaches the preset number of EPA iterations is used as the estimated value of the transmitted signal.

[0092] The number of iterations of EPA is preset to be L2. For K transmitting end antenna numbers i=1, ..., K, the following iteration process can be performed simultaneously or in batches:

[0093] S501, the mean of the Gaussian distribution obeyed by the cavity edge probability of the i-th dimension Make a hard decision and get the estimated symbol vector of the current iteration

[0094]

[0095] Among them, l2 represents the current number of iterations, is a constellation point in the constellation set Θ, which is determined by the modulation order and the modulation mode, Indicates that in the l2th iteration, The smallest hour The value of is the mean of the Gaussian distribution of the cavity edge probability updated in the last iteration.

[0096] S502, update the mean of the Gaussian distribution obeyed by the cavity edge probability of the i-th dimension

[0097]

[0098] Among them, Gram matrix A = H H H, is the variance of the posterior probability distribution of the i-th dimension, B i The output of the matched filter The element of the i-th row of .

[0099] S503, using a scaling factor β (0<β<1) to Scaling is performed to speed up convergence.

[0100]

[0101] After L2 iterations, we get This is the final estimate of the soft information of the transmitted signal vector x.

[0102] An embodiment of the present application also provides a signal detection system for a large-scale MIMO system, which is used to execute the signal detection method provided in the embodiment of the present application. The signal detection system includes an input unit, a preprocessing unit, a SORI unit and an EPA unit.

[0103] The input unit is used to obtain the number of antennas at the transmitting end, the number of antennas at the receiving end, the received signal, the channel matrix, the average symbol energy and the noise variance, determine the first damping coefficient and the second damping coefficient of the signal detector at the receiving end, and input the received signal, the channel matrix, the average symbol energy and the noise variance, the first damping coefficient and the second damping coefficient to other units of the signal detection system.

[0104] The preprocessing unit is used to determine the initial mean and inverse covariance matrix of the posterior probability distribution of the transmitted signal according to the received signal, the channel matrix, the noise variance, the average symbol energy and the first damping coefficient.

[0105] The SORI unit is used to obtain the mean vector of the posterior probability distribution of the transmitted signal after completing the SORI iteration based on the SORI iteration method, and to determine the initial mean of the Gaussian distribution obeyed by the cavity edge probability according to the mean vector of the approximate posterior probability distribution of the transmitted signal after completing the SORI iteration.

[0106] The EPA unit is used to iteratively update the mean of the Gaussian distribution obeyed by the cavity edge probability based on the EPA iteration method and the initial mean of the Gaussian distribution obeyed by the cavity edge probability, and use the mean of the Gaussian distribution obeyed by the cavity edge probability that reaches a preset number of EPA iterations as the estimated value of the transmitted signal.

[0107] In some preferred embodiments, the preprocessing unit uses a diagonal-based systolic array to calculate H H H and H H The present application proposes three improved methods for diagonal-based systolic arrays, which are used to calculate matrix-matrix multiplication and matrix-vector multiplication, and can be used optionally or in combination.

[0108] (1) The first improvement method is to use a reverse data flow in the systolic array, which can significantly reduce the computing time compared to other systolic arrays used in MIMO detection.

[0109] The systolic array processes data in parallel and includes K(K+1) / 2 processing units PE. The processing unit PE processes multiple groups of data in each clock cycle. Assuming the degree of parallelism is P, each processing unit PE processes P groups of data in each clock cycle.

[0110] The systolic array is used to receive input data including a channel matrix H and a received signal vector y, and output H H H and H H y; the processing unit PE includes a PE-A unit for calculating conjugation and a PE-A top Unit, calculation H H The PE-B unit and PE-B in the diagonal part of H top Unit, calculation H H y's PE-D unit and PE-B top Unit and calculation H H The PE-C unit and the PE-D unit of the non-diagonal part of H; the systolic array adopts the data flow direction opposite to the systolic array used in other MIMO detection, and the input data is input from the long side of the systolic array to the PE-A unit and the PE-A top Unit, get H H ; The channel matrix H flows along the direction of the array from the long side of the systolic array to the side of the systolic array, and the H H Flows from the long side of the systolic array to the other side of the systolic array.

[0111] In this way, the operation time of matrix-matrix multiplication is reduced to N / P+K+1, compared with the processing time of up to N / P+2K-1 required by other MIMO detection applications using systolic arrays.

[0112] (2) The second improved method is to configure the systolic array with an advance addition unit, which can shorten the critical path of the systolic array when performing complex multiplication operations.

[0113] The preprocessing unit uses a systolic array to calculate matrix-matrix or matrix-vector multiplication, and the systolic array is configured with an advance addition module for preprocessing a complex multiplication operation, wherein the complex multiplication operation is a multiplication of a first multiplier and a second multiplier, the first multiplier is represented by a+jb, and the second multiplier is represented by c+jd;

[0114] The advance addition module includes a first addition unit and a second addition unit, the first addition unit is used to calculate a+b, and the second addition unit is used to calculate c+d; the advance addition module is used to transfer the output of the first addition unit and the output of the second addition unit to the processing unit PE.

[0115] Since the complex multiplication of the first multiplier and the second multiplier can be transformed into:

[0116] (a+jb)(c+jd)=ac-bd+j[(a+b)(c+d)-(ac+bd)];

[0117] In order to shorten the critical path, the first adding unit and the second adding unit are configured, and the calculation result is transmitted to the processing unit PE together with a, b, c, and d. This method can reduce the critical path of the systolic array by one addition, while only increasing the hardware area negligibly.

[0118] (3) The third improvement method is to configure a rounding unit in the systolic array, which can reduce the bit width of the input data and thus reduce hardware resources.

[0119] The preprocessing unit uses a systolic array to calculate matrix-matrix or matrix-vector multiplication. The systolic array is configured with a rounding module, which is used to round off input data to reduce the bit width, which can significantly shorten the critical path of the systolic array and reduce complexity while still maintaining good detection performance.

[0120] In another preferred embodiment, the matrix-vector multiplication operation in the SORI unit and the EPA unit can be performed in a partially parallel manner, where the F columns of the matrix are multiplied by the F elements of the vector in each clock cycle, and the entire operation can be completed by repeating this multiple times.

[0121] In another preferred embodiment, a simplified hard decision method is used in the operation of the EPA unit, which has the characteristics of low complexity and low latency and is mainly applied to the hard decision of unnormalized symbols. As shown in Table 1, the simplified hard decision method is: Assume (Binary representation) is the sign bit and integer bit of the real or imaginary part of the symbol to be determined. and the highest integer digit To perform overflow protection, and When they are 0 and 1 respectively, the judgment result is When the two are 1 and 0 respectively, the judgment result is When the lowest integer bit b0 is 0, the decision result is b+1, and when it is 1, the decision result is b. This simplified hard decision method only needs to decide three bits, compared with the traditional hard decision method. An addition, Absolute value operation and Level comparison operation.

[0122] Table 1 Simplified hard decision method

[0123]

[0124] The following examples will further illustrate the present application.

[0125] Example 1

[0126] See also Figure 1In the embodiment of the present application, the architecture of the signal detection system is implemented by taking K=32 and N=128 as an example. The architecture can support multiple modulations from QPSK to 256-QAM.

[0127] The signal detection system provided in the embodiment of the present application includes an input unit, a preprocessing unit, a SORI unit and an EPA unit. Figure 2 In the embodiment of the present application, the preprocessing unit uses a diagonal systolic array to calculate H H H and H H y. The diagonal systolic array uses a two-parallel approach, with each processing unit PE processing two sets of data in each clock cycle. top The function of calculating conjugate is configured in , and H can be input into it to calculate its conjugate transpose H H . Processing units PE-B and PE-B top To calculate H H The diagonal part of H, H H y is processed by processing units PE-D and PE-B top Calculated. H The non-diagonal part of H is obtained by processing units PE-C and PE-D.

[0128] The systolic array uses the opposite data flow direction to the systolic array used in other MIMO signal detectors. H and y are input from the long side of the systolic array to the pre-processing units PE-A and PE-A. top After being processed by the preprocessing unit, H H , then H flows from top to bottom, H H Flows from right to left. In this way, H H H.H H The computation time of y is only N / 2+K+1=97, compared to N / 2+2K-1=127 for the systolic arrays used in other MIMO signal detectors.

[0129] In addition, by adding an advance addition module in the preprocessing unit to shorten the critical path of the systolic array, the calculation of complex multiplication can be transformed into:

[0130] (a+jb)(c+jd)=ac-bd+j[(a+b)(c+d)-(ac+bd)];

[0131] In order to reduce the critical path, a+b and c+d are calculated in the preprocessing unit, that is, H, y and H are first H The real part and the imaginary part are added, and then the calculation result and the input data are passed to the next operation unit. This method can reduce one addition in the critical path.

[0132] In order to further reduce the hardware complexity, a rounding module is added to the preprocessing unit. The specific operation is: rounding the input data to reduce the bit width. By rounding, H and H H The bit width of H is reduced from 9 bits to 8 bits, and the bit width of y is reduced from 11 bits to 10 bits. In addition, in order to reduce the performance loss caused by the bit width reduction, H A compensation term of 0.0625 is added to the calculation result of H. The above method can significantly shorten the critical path of the systolic array and reduce the complexity while still maintaining good detection performance.

[0133] The matrix-vector multiplication operations in the iterations of the SORI unit and the EPA unit are partially parallelized. In the SORI unit, the three columns of the inverse covariance matrix W are multiplied by the estimate of the mean of the posterior probability distribution every clock cycle. Multiply the three elements of , so It can be obtained by calculation through 11 clock cycles. In the EPA unit, two columns of the Gram matrix A are combined with the symbol vector after hard decision in each clock cycle. After 16 clock cycles of calculation, we can get

[0134] Further, the is an unnormalized symbol, so a simplified hard decision method is also used in the iteration of EPA, specifically: Assume (binary representation) is The sign bit and integer bit of the real or imaginary part are determined by and the highest integer digit To perform overflow protection, and When they are 0 and 1 respectively, the judgment result is When the two are 1 and 0 respectively, the judgment result is When the lowest integer bit b0 is 0, the judgment result is b+1, and when it is 1, the judgment result is b.

[0135] Furthermore, in order to fully utilize hardware resources and improve throughput, the architecture of the embodiment of the present application adopts an overlap structure to continuously input data into the systolic array. Figure 3 , when one frame of data is input to the systolic array, the next frame of data is input immediately. At 256-QAM, the SORI and EPA units require 4 and 3 iterations respectively. With this structure, the calculation result of one frame is obtained every N / P=64 clock cycles.

[0136] In order to evaluate the feasibility and performance of the design of Example 1, the hardware design was implemented using SystemVerilog, synthesized using Design Compiler under TSMC 28nm process, and placed and routed using Integrated Circuit (IC) compiler, and the power consumption was evaluated by PrimeTime PX in average mode. Finally, the EPA-SORI detector proposed in the embodiment of the present application achieved higher throughput, area efficiency, and energy efficiency than the previous optimal detector EPA-wNSA, with a throughput of 2.2 times and an area efficiency of 1.36 times.

[0137] For other specific details, please refer to the signal detection method described above in the embodiment of the present application, which will not be repeated here.

[0138] It can be seen from the above technical solutions that the embodiment of the present application provides a signal detection method and system for a large-scale MIMO system, which combines the existing SORI method with the EPA-wNSA method, and uses SORI to approximate the matrix inversion of the initialization part of the EPA algorithm, thereby reducing the number of iterations required for MIMO signal detection while ensuring good detection performance; converting the variables of the preprocessing part into a new calculation method, that is, omitting some of the multiplication operations therein, thereby reducing 4K 2 The multiplication operation reduces the complexity; in the data transmission process between SORI iteration and EPA iteration, the algorithm derives t 0 Approximately Some division, multiplication and addition operations are eliminated, further reducing the algorithm complexity.

[0139] Furthermore, the embodiment of the present application also uses a systolic array in the preprocessing unit of the signal detection system to calculate matrix-matrix multiplication and matrix-vector multiplication. The systolic array can use a reverse data flow, which can significantly reduce the operation time compared to other systolic arrays used in MIMO detection. It can also be configured with an advance addition unit to shorten the critical path of the systolic array when performing complex multiplication operations. It can also be configured with a rounding unit to reduce the bit width of the input data and thus reduce hardware resources.

[0140] Furthermore, the embodiment of the present application also proposes a simplified hard decision method in the EPA iteration unit, which has the characteristics of low complexity and low latency.

[0141] In summary, compared with the previous EPA-wNSA method, the signal detection method and system proposed in the embodiment of the present application have fewer iterations, are more conducive to high-throughput hardware implementation, and can achieve the same or even better performance. At the same time, the hardware architecture of the signal detection system enables higher throughput, higher area efficiency and higher energy efficiency.

[0142] The above-described embodiments of the present application do not constitute a limitation on the protection scope of the present application.

Claims

1. A signal detection method for a massive MIMO system, characterized in that: include: Obtain the number of antennas at the transmitting end, the number of antennas at the receiving end, the received signal, the channel matrix, the noise variance, and the average symbol energy; Determining a first damping coefficient and a second damping coefficient of a signal detector at a receiving end according to the number of antennas at the transmitting end and the number of antennas at the receiving end; Determining an initial mean of a posterior probability distribution of a transmitted signal, an inverse covariance matrix, and a matched filtering output of the received signal according to the received signal, the channel matrix, the noise variance, the average symbol energy, and the first damping coefficient, comprising: The process of determining the inverse covariance matrix based on the channel matrix and noise variance adopts the following model: W=H H H+diag'(Λ i ); Where W is the inverse covariance matrix, H is the channel matrix, diag'(·) represents a diagonal matrix consisting of the diagonal elements of a vector or a matrix, and Λ i To set the parameters, i represents the signal dimension of the transmitted signal, i.e. the serial number of the transmitting antenna. is the noise variance, E s is the average symbol energy; The process of determining the matched filter output of the received signal based on the channel matrix and the received signal adopts the following model: Where B is the matched filter output, and y is the received signal vector; The mean of the posterior probability distribution is initialized according to the matched filter output and the first damping coefficient to obtain the initial mean of the posterior probability distribution, which is expressed as: m 0 =0, μ 1 =δB; Where, δ is the first damping coefficient; According to the initial mean of the posterior probability distribution of the transmitted signal, the matched filter output, the inverse covariance matrix, the first damping coefficient and the second damping coefficient, SORI iteration is performed on the mean of the posterior probability distribution of the transmitted signal to obtain a mean vector of the posterior probability distribution of the transmitted signal after the SORI iteration is completed, and, according to the mean vector of the posterior probability distribution of the transmitted signal after the SORI iteration is completed, an initial mean of the Gaussian distribution obeyed by the cavity edge probability is determined, wherein the SORI iteration adopts the following model: In the formula, the preset number of SORI iterations is L-1, then the mean vector of the posterior probability distribution of the transmitted signal after completing L-1 SORI iterations is μ L , ω is the second damping coefficient; The variance of the posterior probability distribution of the transmitted signal is obtained according to the inverse covariance matrix, and based on the variance of the posterior probability distribution and the initial mean of the Gaussian distribution obeyed by the cavity edge probability, the EPA iterative method is adopted to iteratively update the mean of the Gaussian distribution obeyed by the cavity edge probability, and the mean of the Gaussian distribution obeyed by the cavity edge probability that reaches a preset number of EPA iterations is used as the estimated value of the transmitted signal.

2. The signal detection method for a massive MIMO system according to claim 1, characterized in that: The first damping coefficient and the second damping coefficient of the signal detector at the receiving end are determined according to the number of antennas at the transmitting end and the number of antennas at the receiving end, and the following model is adopted: Where K and N are the number of antennas at the transmitting and receiving ends respectively.

3. The signal detection method for a massive MIMO system according to claim 1, characterized in that: According to the mean vector of the posterior probability distribution of the transmitted signal after completing the SORI iteration, the initial mean of the Gaussian distribution obeyed by the cavity edge probability is determined, and the following model is adopted: t 0 =μ L ; Where, t 0 is the initial mean of the Gaussian distribution obeyed by the cavity edge probability.

4. The signal detection method for a massive MIMO system according to claim 1, characterized in that: The method of iteratively updating the mean of the Gaussian distribution obeyed by the cavity edge probability using the EPA iteration method based on the variance of the posterior probability distribution and the initial mean of the Gaussian distribution obeyed by the cavity edge probability, and taking the mean of the Gaussian distribution obeyed by the cavity edge probability that reaches the preset number of EPA iterations as the estimated value of the transmitted signal includes: The mean of the Gaussian distribution of the cavity edge probability in the i-th dimension Make a hard decision and get the estimated symbol vector of the current iteration The following model is used: In the formula, l2 represents the current number of iterations, is a constellation point in the constellation set Θ, which is determined by the modulation order and the modulation mode, Indicates that in the l2th iteration, The smallest hour The value of is the mean of the Gaussian distribution of the cavity edge probability updated in the last iteration; Update the mean of the Gaussian distribution of the cavity edge probability of the i-th dimension The following model is used: In the formula, Gram matrix A = H H H, is the variance of the posterior probability distribution of the i-th dimension, B i The output of the matched filter The element of the i-th row of ; Use the scaling factor β (0<β<1) to Scaling is performed using the following model: The number of iterations is preset, and the mean of the Gaussian distribution obeyed by the cavity edge probability reaching the preset number of iterations is obtained, and the mean of the Gaussian distribution obeyed by the cavity edge probability reaching the preset number of iterations is used as the estimated value of the transmitted signal.

5. A signal detection system for a massive MIMO system, used to execute the signal detection method according to any one of claims 1 to 4, characterized in that: It includes input unit, pre-treatment unit, SORI unit and EPA unit; The input unit is used to obtain the number of antennas at the transmitting end, the number of antennas at the receiving end, the received signal, the channel matrix, the average symbol energy and the noise variance, determine the first damping coefficient and the second damping coefficient of the signal detector at the receiving end, and input the received signal, the channel matrix, the average symbol energy and the noise variance, the first damping coefficient and the second damping coefficient to the preprocessing unit, the SORI unit and the EPA unit; The preprocessing unit is used to determine the initial mean and inverse covariance matrix of the posterior probability distribution of the transmitted signal according to the received signal, the channel matrix, the noise variance, the average symbol energy and the first damping coefficient; The SORI unit is used to obtain a mean vector of the posterior probability distribution of the transmitted signal after the SORI iteration is completed based on the SORI iteration method, and determine the initial mean of the Gaussian distribution obeyed by the cavity edge probability according to the mean vector of the approximate posterior probability distribution of the transmitted signal after the SORI iteration is completed; The EPA unit is used to iteratively update the mean of the Gaussian distribution obeyed by the cavity edge probability based on the EPA iteration method and the initial mean of the Gaussian distribution obeyed by the cavity edge probability, and use the mean of the Gaussian distribution obeyed by the cavity edge probability that reaches a preset number of EPA iterations as the estimated value of the transmitted signal.

6. The signal detection system for a massive MIMO system according to claim 5, characterized in that: The pre-processing unit is configured with a diagonal-based systolic array, the systolic array includes K(K+1) / 2 processing units PE, and the processing unit PE processes multiple groups of data in each clock cycle; The systolic array is used to receive input data including a channel matrix H and a received signal vector y, and output H H H and H H y; the processing unit PE includes a PE-A unit for calculating conjugation and a PE-A top Unit, calculation H H The PE-B unit and PE-B in the diagonal part of H top Unit, calculation H H y's PE-D unit and PE-B top Unit and calculation H H PE-C units and PE-D units in the non-diagonal part of H; The systolic array uses reverse data flow, and the input data is input from the long side of the systolic array to the PE-A unit and the PE-A top Unit, get H H ; The channel matrix H flows along the direction of the array from the long side of the systolic array to the side of the systolic array, and the H H Flows from the long side of the systolic array to the other side of the systolic array.

7. The signal detection system for a massive MIMO system according to claim 5, characterized in that: The preprocessing unit uses a systolic array to calculate matrix-matrix or matrix-vector multiplication; The systolic array is configured with an advance addition module for pre-processing a complex multiplication operation, wherein the complex multiplication operation is a multiplication of a first multiplier and a second multiplier, the first multiplier is represented by a+jb, and the second multiplier is represented by c+jd; The advance addition module includes a first addition unit and a second addition unit, the first addition unit is used to calculate a+b, and the second addition unit is used to calculate c+d; the advance addition module is used to transfer the output of the first addition unit and the output of the second addition unit to the processing unit PE.

8. The signal detection system for a massive MIMO system according to claim 5, characterized in that: The pre-processing unit uses a systolic array to calculate matrix-matrix or matrix-vector multiplication; the systolic array is configured with a rounding module, and the rounding module is used to round input data to reduce the bit width.

Images