Method, apparatus and terminal device for transmitting signal estimation

Abstract

This invention provides a method, apparatus, and terminal device for estimating transmitted signals, belonging to the field of wireless communication technology. The method includes determining a received signal vector, a channel matrix, and a noise variance; inputting the received signal vector, channel matrix, and noise variance into a signal estimation model to obtain an estimated transmitted signal value output by the signal estimation model; wherein, the signal estimation model is obtained by optimizing a minimum mean square error detector based on a Fourier transform algorithm and a pre-determined weighting factor, and the weighting factor is used to adjust the convergence step size. This invention uses the Fourier transform algorithm to replace matrix inversion in the minimum mean square error detector, avoiding explicit matrix operations, reducing computational complexity, and can be used in M-MIMO systems with large-scale antennas. Furthermore, adjusting the convergence step size based on the weighting factor is beneficial for stable convergence and prevents severe degradation of bit error rate performance under high-dimensional and high-order modulation conditions.

Description

Technical Field

[0001] This invention relates to the field of wireless communication technology, and in particular to a method, apparatus and terminal device for estimating transmitted signals. Background Technology

[0002] Massive multiple-input multiple-output (M-MIMO) technology is one of the key supporting technologies for Beyond Fifth Generation Mobile Communication Technology / Sixth Generation Mobile Communication Technology (B5G / 6G) wireless communication systems. By deploying hundreds of antennas at the base station, it can simultaneously provide services to multiple user terminals on the same time-frequency resources, thereby significantly improving the system's spectral efficiency, energy efficiency, and overall capacity. With the continuous expansion of system scale, baseband signal processing faces enormous challenges, among which signal estimation is widely recognized as one of the key bottlenecks restricting system performance and complexity.

[0003] Currently, the Minimum Mean Square Error (MMSE) detector or its improved iterative detectors are commonly used for signal estimation. However, existing MMSE detectors and their improved iterative detectors are not suitable for M-MIMO systems with high dimensionality, large-scale antennas, and high-order modulation. For example, the MMSE detector relies on matrix inversion, which has a computational complexity of cubic order. When the number of antennas in the system reaches hundreds, the computational burden of matrix inversion is too large, severely limiting its application in M-MIMO systems. For example, although the zero-forcing detector based on Fourier transform avoids matrix calculation and has a faster convergence speed, its bit error rate performance is severely degraded under high dimensionality and high-order modulation conditions.

[0004] Therefore, it is necessary to propose a signal estimation scheme suitable for M-MIMO systems with high dimensionality, large-scale antennas, and high-order modulation. Summary of the Invention

[0005] This invention provides a method, apparatus, and terminal device for estimating transmitted signals, applicable to M-MIMO systems with high-dimensional, large-scale antennas and high-order modulation, with low computational complexity and robust bit error rate performance.

[0006] This invention provides a method for estimating transmitted signals, comprising: Determine the received signal vector, channel matrix, and noise variance; The received signal vector, the channel matrix, and the noise variance are input into the signal estimation model to obtain the transmitted signal estimate output by the signal estimation model. The signal estimation model is obtained by optimizing the minimum mean square error detector based on the Fourier transform algorithm and a pre-determined weighting factor, wherein the weighting factor is used to adjust the convergence step size.

[0007] As an example, the signal estimation model is determined based on the following method: Based on the Fourier transform algorithm, a multi-order approximate expression for the inverse matrix of the noise regularization matrix in the minimum mean square error detector is determined to obtain the first detector. The multi-order approximation expression is weighted based on the weighting factor to optimize the first detector, thereby obtaining the second detector; The signal estimation model is obtained based on the second detector.

[0008] As one embodiment, obtaining the signal estimation model based on the second detector includes: The second detector is used as the signal estimation model, or a third detector is obtained by adjusting the multi-order approximation expression in the second detector based on a deep neural network, and the third detector is used as the signal estimation model, wherein the deep neural network is used to determine the weighting factor.

[0009] As one embodiment, determining the multi-order approximate expression of the inverse matrix of the noise regularization matrix in the minimum mean square error detector based on the Fourier transform algorithm to obtain the first detector includes: The noise regularization matrix is ​​defined as the Gram matrix of the channel matrix. Based on the Gram matrix and the discrete Fourier transform matrix, a multi-order approximation expression for the inverse matrix is ​​determined. The product of the multi-order approximation of the inverse matrix and the matched filter signal is used as the expression for the first detector; The matched filter signal is used to characterize the signal obtained after matching filtering the received signal.

[0010] As an example, the weighting factor is determined based on the following method: The deep neural network is constructed based on the multi-order approximation expression of the inverse matrix; The deep neural network is trained offline based on transmitted signal samples, received signal samples, channel matrix, and noise variance to obtain the weighting factor.

[0011] As one embodiment, the weighting factor includes a sequence of geometrically decreasing parameters corresponding to the multi-order approximation expression of the inverse matrix, and the deep neural network includes... n Rank neural network unit, n The weighting factor, which has the same order as the multi-order approximation expression of the inverse matrix, is obtained by offline training of the deep neural network based on transmitted signal samples, received signal samples, the channel matrix, and noise variance, including: The received signal sample, the channel matrix, and the noise variance are input into the deep neural network to determine the matched filter signal, the diagonal matrix, the discrete Fourier transform matrix, the first-order weighting factor, and... n Order-order approximate cumulative matrix; Based on the matched filter signal, diagonal matrix, discrete Fourier transform matrix, first-order weighting factor, and n An approximate cumulative matrix of order is used to determine the predicted value of the transmitted signal; The loss function value of the deep neural network is determined based on the predicted value of the transmitted signal and the mean square error of the transmitted signal sample. If the loss function value meets a preset threshold or reaches a preset number of training iterations, offline training of the deep neural network is completed. The first-order weighting factor is substituted into the geometrically decreasing parameter model to obtain the weighting factor. Otherwise, the first-order weighting factor is adjusted and the process returns to the step of inputting the received signal sample, the channel matrix, and the noise variance into the first layer of the deep neural network.

[0012] The present invention also provides a transmission signal estimation apparatus, comprising: The determination module is used to determine the received signal vector, channel matrix, and noise variance. The estimation module is used to input the received signal vector, the channel matrix and the noise variance into the signal estimation model to obtain the transmitted signal estimate output by the signal estimation model; The signal estimation model is obtained by optimizing the minimum mean square error detector based on the Fourier transform algorithm and a pre-determined weighting factor, wherein the weighting factor is used to adjust the convergence step size.

[0013] The present invention also provides a terminal device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the transmission signal estimation method as described above.

[0014] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the transmission signal estimation method as described above.

[0015] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the transmission signal estimation method as described above.

[0016] The present invention provides a method, apparatus, and terminal device for estimating transmitted signals. The method determines a received signal vector, a channel matrix, and a noise variance. The received signal vector, the channel matrix, and the noise variance are input into a signal estimation model to obtain an estimated transmitted signal value output by the signal estimation model. The signal estimation model is obtained by optimizing a minimum mean square error detector (MMIMO) based on a Fourier transform algorithm and a pre-determined weighting factor. The weighting factor is used to adjust the convergence step size. This invention replaces matrix inversion in the MMIMO with a Fourier transform algorithm, avoiding explicit matrix operations, reducing computational complexity, and making it applicable to M-MIMO systems with large-scale antennas. Furthermore, adjusting the convergence step size based on the weighting factor promotes stable convergence and prevents severe degradation of bit error rate performance under high-dimensional and high-order modulation conditions. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating the signal estimation method provided by the present invention.

[0019] Figure 2 This is a schematic diagram of the structure of the deep neural network provided by the present invention.

[0020] Figure 3 This is a schematic diagram of the algorithm operation flow of the third detector provided by the present invention.

[0021] Figure 4 This is a schematic diagram showing the relationship between the complexity and signal-to-noise ratio of the present invention and existing technologies under different antenna ratios and modulation orders.

[0022] Figure 5 This is one of the schematic diagrams showing the BER variation curves of the present invention and existing technologies under different antenna ratios and modulation orders.

[0023] Figure 6 This is the second schematic diagram of the BER variation curves of the present invention and existing technologies under different antenna ratios and modulation orders.

[0024] Figure 7 This is a schematic diagram of the BER variation curves under different weighting factors according to the present invention.

[0025] Figure 8 This is a schematic diagram illustrating the robustness of the present invention.

[0026] Figure 9 This is a schematic diagram of the transmission signal estimation device provided by the present invention.

[0027] Figure 10 This is a schematic diagram of the structure of the terminal device provided by the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0029] Uplink refers to communication between multiple user terminals (such as mobile phones) and a base station (BS). Assuming the base station in a massive MIMO uplink communication system has... One receiving antenna and One transmitting antenna, of which The uplink massive MIMO communication system uses a constellation point modulation scheme. The Q-QAM modulation scheme has the following system model: (1) in, and These are the received signal vector and the transmitted signal vector, respectively. This is the channel matrix. In an independent and identically distributed Rayleigh channel with a mean of 0 and a variance of 1, It is an Additive White Gaussian Noise (AWGN) vector, each element satisfy , This represents the noise variance, and its value is equal to the noise power.

[0030] According to the received signal vector and channel matrix Existing linear detectors, such as the Zero Forcing (ZF) detector and the MMSE detector, estimate the transmitted signal vector as shown in the following equation: (2) in, This represents the transmitted signal vector estimated by the ZF detector. This represents the transmitted signal vector estimated by the MMSE detector.

[0031] As can be seen from equation (2), the matrix inversion operation relied upon by the ZF detector and the MMSE detector introduces a significant complexity burden in the M-MIMO system. Taking the MMSE detector as an example, its complexity mainly consists of two parts: one is the computation... complexity Secondly, the complexity of matrix inversion. Therefore, the overall complexity of the MMSE detector is... This has become the main bottleneck for its application in M-MIMO systems.

[0032] The zero-forcing detector based on Fourier transform transforms the matrix inversion calculation in the ZF detector to avoid explicit matrix inversion calculation. However, due to errors in the transformation process, the bit error rate performance of the M-MIMO system is severely degraded under high-dimensional and high-order modulation conditions.

[0033] Compared with the ZF detector, the MMSE detector has better performance and a wider range of applications. Therefore, this invention optimizes the MMSE detector based on the Fourier transform algorithm and a predetermined weighting factor to obtain a signal estimation model. The signal estimation model is then used to perform the transmission signal estimation step to obtain the transmission signal estimate. The invention will be described in detail below with reference to the accompanying drawings.

[0034] Figure 1 This is a flowchart illustrating the transmission signal estimation method provided by the present invention, as shown below. Figure 1 As shown, the present invention provides a method for estimating transmitted signals, applicable to uplink massive MIMO communication systems, and the method includes the following steps.

[0035] Step S100: Determine the received signal vector, channel matrix, and noise variance.

[0036] Step S200: Input the received signal vector, the channel matrix, and the noise variance into the signal estimation model to obtain the transmitted signal estimate output by the signal estimation model.

[0037] The signal estimation model is obtained by optimizing the minimum mean square error detector based on the Fourier transform algorithm and a pre-determined weighting factor, wherein the weighting factor is used to adjust the convergence step size.

[0038] Before step S100, the base station samples all receiving antennas at the same time and combines the sampled values ​​of all antennas as the received signal vector. The channel matrix is ​​determined by the base station through the channel estimation process, and the noise variance is measured by the base station during deployment or idle periods, which is a priori known system parameter.

[0039] Understandably, this invention uses the Fourier transform algorithm to replace the matrix inversion in the minimum mean square error detector, avoiding explicit matrix operations, reducing computational complexity, and can be used in M-MIMO systems with large-scale antennas. It also adjusts the convergence step size based on weighting factors, which is beneficial for stable convergence and avoids severe degradation of bit error rate performance under high-dimensional and high-order modulation conditions.

[0040] As an example, the optimization of the MMSE detector based on the Fourier transform algorithm and a predetermined weighting factor includes optimizing the MMSE detector based on the Fourier transform algorithm, and then optimizing the optimized MMSE detector based on the weighting factor.

[0041] Optionally, the MMSE detector can be optimized based on the Fourier transform algorithm, including converting the matrix multiplication operation in the MMSE detector into a discrete Fourier transform algorithm. n Multiplication of stage matrices and vectors.

[0042] Optionally, the optimization of the MMSE detector based on the Fourier transform algorithm also includes: implementing the multiplication operation of matrices and vectors of various orders based on the fast Fourier transform algorithm to reduce complexity.

[0043] Optionally, the optimized MMSE detector can be further optimized based on weighting factors, including introducing weighting factors into the multiplication operations of matrices and vectors of each order, and dynamically adjusting the convergence step size during the iterative operation to reduce the error generated by the matrix multiplication operation.

[0044] The weighting factor used to adjust the convergence step size can be understood as adjusting the convergence step size of the multi-order approximation expression of the inverse matrix of the noise regularization matrix in the minimum mean square error detector.

[0045] Optionally, the signal estimation model includes an optimized MMSE detector, with inputs being the received signal vector, the channel matrix, and the noise variance, and outputting an estimate of the transmitted signal.

[0046] Understandably, this application converts the matrix multiplication operation in the MMSE detector into a discrete Fourier transform algorithm. nThe method avoids matrix inversion by performing multiplication operations on matrices and vectors of different orders using the Fast Fourier Transform algorithm, thus reducing complexity. Weighting factors are introduced into the multiplication operations of matrices and vectors of different orders, and the convergence step size is dynamically adjusted during the iterative operation to reduce the error caused by the transformation of matrix multiplication operations.

[0047] As an example, the signal estimation model is determined based on the following method: Based on the Fourier transform algorithm, a multi-order approximate expression for the inverse matrix of the noise regularization matrix in the minimum mean square error detector is determined to obtain the first detector. The multi-order approximation expression is weighted based on the weighting factor to optimize the first detector, thereby obtaining the second detector; The signal estimation model is obtained based on the second detector.

[0048] Understandably, this invention uses the Fourier transform algorithm to determine a multi-order approximation expression of the inverse matrix of the noise regularization matrix in the minimum mean square error detector to obtain the first detector; it then uses a weighting factor to weight the multi-order approximation expression to optimize the first detector, resulting in the second detector; based on the second detector, a signal estimation model is obtained, and a weighted Fourier approximation method is used to replace matrix inversion in the existing MMSE detector, avoiding explicit matrix operations and significantly reducing computational complexity. Furthermore, the weighting factor is used to dynamically adjust the convergence step size in different iterations, maintaining robust bit error rate performance while reducing computational complexity.

[0049] As one embodiment, determining the multi-order approximate expression of the inverse matrix of the noise regularization matrix in the minimum mean square error detector based on the Fourier transform algorithm to obtain the first detector includes: The noise regularization matrix is ​​defined as the Gram matrix of the channel matrix. Based on the Gram matrix and the discrete Fourier transform matrix, a multi-order approximation expression for the inverse matrix is ​​determined. The product of the multi-order approximation of the inverse matrix and the matched filter signal is used as the expression for the first detector; The matched filter signal is used to characterize the signal obtained after matching filtering the received signal.

[0050] Understandably, this application defines the noise regularization matrix as the Gram matrix of the channel matrix, which allows the first detector to be calculated recursively. With the participation of the matched filter signal, matrix-level multiplication can be replaced by matrix-vector multiplication, effectively reducing the computational complexity burden.

[0051] As one embodiment, obtaining the signal estimation model based on the second detector includes: Use the second detector as the signal estimation model, or A third detector is obtained by adjusting the multi-order approximation expression in the second detector based on a deep neural network, and the third detector is used as the signal estimation model, wherein the deep neural network is used to determine the weighting factor.

[0052] Optionally, adjusting the multi-order approximation expression in the second detector based on a deep neural network includes introducing a multi-order approximation cumulative matrix related to the weighting factor into the multi-order approximation expression in the second detector based on a deep neural network.

[0053] It is understood that this invention can directly use the second detector as the signal estimation model, with the weighting factor set manually, thus reducing computational complexity while maintaining robust bit error rate performance. Alternatively, this invention can determine the weighting factor based on a deep neural network, adjust the multi-order approximation expression in the second detector based on the deep neural network, obtain a third detector, and use the third detector as the signal estimation model. This eliminates the need for manual adjustment of the weighting factor, accelerating convergence and achieving better bit error rate (BER) performance. The design problem of the weighting factor is transformed into an end-to-end training and optimization process through a deep unfolded network, avoiding the complexity of manual design. This invention is suitable for high-order, high-dimensional modulation, and very large approximation orders (e.g., M-MIMO communication system (=30).

[0054] As an example, the weighting factor is determined based on the following method: The deep neural network is constructed based on the multi-order approximation expression of the inverse matrix; The deep neural network is trained offline based on transmitted signal samples, received signal samples, channel matrix, and noise variance to obtain the weighting factor.

[0055] Optionally, constructing the deep neural network based on the multi-order approximation expression of the inverse matrix includes: determining the number of layers of the deep neural network based on the order of the multi-order approximation expression of the inverse matrix, and determining the parameters to be calculated for each layer of the deep neural network based on the formula of the multi-order approximation expression of the inverse matrix.

[0056] Optionally, the transmitted signal samples, received signal samples, channel matrix, and noise variance are generated at random signal-to-noise ratio (SNR) points within an acceptable range, rather than at fixed SNR points, which helps improve the generalization ability of the third detector.

[0057] It is understood that the present invention constructs the deep neural network based on the multi-order approximation expression of the inverse matrix, so that the weighting factor obtained by the deep neural network training has a higher degree of agreement with the multi-order approximation expression of the inverse matrix. The optimal weighting factor is automatically obtained through offline training, and only the fixed weighting factor needs to be called when estimating the transmitted signal online, avoiding additional training overhead, significantly reducing the complexity of online estimation, and enabling it to run efficiently in the hardware system.

[0058] As one embodiment, the weighting factor includes a sequence of geometrically decreasing parameters corresponding to the multi-order approximation expression of the inverse matrix, and the deep neural network includes... n Rank neural network unit, n The weighting factor, which has the same order as the multi-order approximation expression of the inverse matrix, is obtained by offline training of the deep neural network based on transmitted signal samples, received signal samples, the channel matrix, and noise variance, including: The received signal sample, the channel matrix, and the noise variance are input into the deep neural network to determine the matched filter signal, the diagonal matrix, the discrete Fourier transform matrix, the first-order weighting factor, and... n An approximate cumulative matrix of order 1; specifically, the received signal sample, the channel matrix, and the noise variance are input to the first layer of the deep neural network, and the first layer of the neural network transmits the noise regularization matrix and the diagonal matrix to the second layer of the neural network, respectively. n The first layer of the neural network unit outputs a first-order approximation matrix and a first-order approximation cumulative matrix to the second layer of the neural network unit, and outputs the matched filter signal, diagonal matrix, discrete Fourier transform matrix, and first-order weighting factor to the second layer of the neural network unit. n Layer neural network unit; i Layered neural network units are used for output i Approximate matrix of order and i The approximate cumulative matrix of order is passed to the next layer of the neural network unit. i The range of values ​​is 2-n； Based on the matched filter signal, diagonal matrix, discrete Fourier transform matrix, first-order weighting factor, and n An approximate cumulative matrix of order is used to determine the predicted value of the transmitted signal; The loss function value of the deep neural network is determined based on the predicted value of the transmitted signal and the mean square error of the transmitted signal sample. If the loss function value meets a preset threshold or reaches a preset number of training iterations, offline training of the deep neural network is completed. The first-order weighting factor is substituted into the geometrically decreasing parameter model to obtain the weighting factor. Otherwise, the first-order weighting factor is adjusted and the process returns to the step of inputting the received signal sample, the channel matrix, and the noise variance into the first layer of the deep neural network.

[0059] Understandably, this invention transforms the design problem of weighting factors into an end-to-end training and optimization process through deep network expansion, avoiding the complexity of manual design. Optimal weighting factors are automatically obtained through offline training, and online detection only requires calling fixed weighting factors, avoiding additional training overhead. This invention also proposes a geometrically decreasing parameter model and provides a feasible range. Unlike completely free parameter design, this provides structural constraints for hardware implementation, reducing parameter storage and computational paths, and decreasing the use of storage units and multiply-accumulate units. This not only contributes to stable convergence during training but also provides good structured characteristics for chip-level implementation.

[0060] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0061] Based on the channel-hardening phenomenon, the approximate expression for the noise regularization matrix is ​​as follows: (3) in, Represents the noise regularization matrix. This represents the inverse of the noise regularization matrix.

[0062] The reasoning process for determining the multi-order approximate expression of the inverse matrix of the noise regularization matrix in the minimum mean square error detector based on the Fourier transform algorithm includes the reasoning process from equation (4) to equation (10).

[0063] The approximate expression for the noise regularization matrix is ​​further approximated as a circulant matrix. Circular matrix The expression is as follows: (4) Among them, the optimal By minimizing and The optimal Frobenius norm error between them is obtained. The calculation formula is as follows: (5) Solving for the optimal solution using the Lagrangian multipliers method The solution formula is shown below: (6) in , represents a vector consisting entirely of "1"s.

[0064] Based on equations (4) to (6), the update of the cyclic matrix can be derived. The expression is as follows: (7) in, The normalized Discrete Fourier Transform (DFT) matrix. , For matrix The first column vector of is expressed as follows: (8) Using the Cayley-Hamilton theorem, we can obtain the following equation: (9) According to equations (3)-(6), the inverse matrix can be determined. of The approximate expression for order is as follows: (10) in, , where represents the matched filtered signal, F represents the discrete Fourier transform matrix, and n represents the order of the multi-order approximation expression of the inverse matrix. Represents a diagonal matrix. Let I represent the noise variance, and let I represent the identity matrix.

[0065] Since equation (10) requires matrix-level multiplication in each iteration, resulting in a high computational burden, this embodiment of the invention defines... Correspondingly, the expression for the first detector is as follows: (11) in, The first detector can be calculated recursively. With the participation of [unclear], matrix-level multiplication can be replaced by matrix-vector multiplication, effectively reducing the computational complexity burden.

[0066] Furthermore, since the optimization principle of the first detector is based on the Fourier transform to approximate the inverse matrix of the noise regularization matrix into an nth-order approximate expression, the first detector can also be represented as a Fourier transform-based approximate MMSE (FTMMSE) detector.

[0067] However, due to the approximation error in the multi-order approximation expression of the inverse matrix, the first detector suffers performance degradation in high-order modulation and high-dimensional MIMO scenarios. Therefore, this invention introduces a weighting factor (denoted as ) into the first detector. This is used to dynamically adjust the convergence step size during different iterations.

[0068] After introducing weighting factors, the matrix of The order approximation can be expressed as follows: (12) According to equation (12), the expression for the second detector is as follows: (13) in, The second detector is based on the first detector and introduces a weighting factor. Therefore, the second detector can also be represented as a weighted FTMMSE detector, i.e., WeFTMM (weighted FTMMSE).

[0069] In the second detector mentioned above, the weighting factors are manually set. To accelerate convergence and improve accuracy, a fully expanded deep neural network can be used to optimize the second detector, automatically optimizing the weighting factors corresponding to each layer to obtain the third detector. The third detector is obtained by optimizing the second detector based on the deep neural network, and therefore can be represented as a WeFTMM detector optimized based on a deep neural network, namely the WeFTNet (DNN-aided weighted FT-based approximateMMSE) detector.

[0070] The optimization of the second detector based on a deep neural network includes adjusting the multi-order approximation expression in the second detector based on the deep neural network.

[0071] Specifically, based on the deep neural network, the multi-order approximation expression in the second detector is adjusted to obtain the adjusted multi-order approximation expression as follows: (14) in, For corresponding of The weighting factor for the order approximation. yes An approximate cumulative matrix of order 1. To train the weighting factors via a DNN, Expand as The order of recursion is given by the first order, where the first order is the first order. The item involves weighting factors .

[0072] The expression for the multi-order approximate cumulative matrix is ​​shown below: (15) in, for The order approximation matrix, for Accumulation and satisfy .

[0073] Figure 2 This is a schematic diagram of the deep neural network structure provided by the present invention, as shown below. Figure 2 As shown, the deep neural network provided by this invention is a fully unfolded deep neural network (DNN) architecture, with the first layer of neural network units serving as a preprocessing layer for receiving input signals. and calculate The following Layer usage. From the second layer to the third layer. Layers, matrices and ( The process is recursively updated according to equation (15). Finally, the result is obtained by calculating using equation (14). of This approximation yields an estimate of the transmitted signal vector. . It is not used as the final detection result of the entire detection algorithm, but as the output and transmitted signal vector sample of the deep neural network. The loss function of the deep neural network is calculated and its performance is continuously optimized to obtain the optimal weight factors, resulting in the final detection result. ,in As shown in equation (14), These are the trained weight factors.

[0074] For example, the offline training of the deep neural network is based on the widely used PyTorch framework, generating 50 training epochs. In each epoch, 14,000 and 1,400 sample data are generated as the training set and validation set, respectively, according to the antenna configuration of the M-MIMO scene to be detected. These samples include transmitted signal samples. Received signal samples Channel matrix samples and noise variance samples .

[0075] To improve training efficiency, this invention employs a batch training strategy, which tracks the loss and updates trainable parameters in each batch. Furthermore, mean squared error (MSE) is used as the loss function to measure the output of the deep neural network. With transmitted signal sample The difference between them is expressed as follows: (16) in, Set the batch size to 32. Indicates the first in this batch The estimation results for each training sample The spectral norm of a vector or matrix.

[0076] Furthermore, to ensure better convergence performance, the Adam optimizer can be used for training, with an initial learning rate of 0.1, which decays at a rate of 0.8 every 10 epochs. It should be noted that since this is offline training, it does not incur the complexity of online detection.

[0077] Optionally, this invention parameterizes the weighting factor as a geometrically decreasing sequence, and the geometrically decreasing parameter model expression is as follows: (17) in, Indicates the first Order weighting factor, This indicates a decreasing parameter. Only one pair of parameters needs to be learned. Substituting into equation (17), we can obtain all the weighting factors. Larger weights are used in the early iterations to achieve fast convergence, while smaller weights are used in the later iterations to refine the results.

[0078] Furthermore, by calculating the approximation error This invention analyzes the convergence of the algorithm and proves that the range of values ​​for the weight factors is... From this, it can be deduced that Therefore, in each layer of the DNN, the parameters obtained through training... They will all be projected onto Within this range. By leveraging this geometrically decreasing model and value constraints, the WeFTNet proposed in this invention not only alleviates instability during convergence but also possesses good generalization performance. Finally, based on the parameters obtained during training... The weight factors of each layer can be calculated using equation (17). .

[0079] Figure 3 This is a schematic diagram of the algorithm operation flow of the third detector provided by the present invention, as follows: Figure 3 As shown, the estimated transmitted signal vector is calculated based on the third detector. In the process, the multiplication operation of matrices and vectors of various orders is realized based on the Fast Fourier Transform algorithm.

[0080] Specifically, the DFT matrix is ​​performed using the Fast Fourier Transform (FFT / IFFT). Multiplication with vectors, so as to convert the original... The complexity is reduced to .

[0081] The calculation uses an iterative approach, involving only matrix-vector multiplication and eliminating the need for matrix-level multiplication. Specifically, in obtaining... During the iteration process, left multiplication can be performed first in each calculation step. Multiply by the left again This avoids explicitly constructing the matrix. This greatly reduces computational complexity.

[0082] It is understood that this invention implements multiplication operations of matrices and vectors of various orders based on the Fast Fourier Transform, and can directly utilize mature modules on digital signal processor (DSP), field-programmable gate array (FPGA), and application-specific integrated circuit (ASIC) platforms, making it suitable for hardware acceleration.

[0083] The simulation performance comparison between the present invention and the prior art is described in detail below with reference to the accompanying drawings.

[0084] Figure 4 The diagram illustrates the complexity and signal-to-noise ratio relationship between the present invention and existing technologies under different antenna ratios and modulation orders in a Rayleigh channel. The blue pentagram represents the existing MMSE detector, × represents the third detector provided by the present invention, the purple box represents the second detector provided by the present invention, and the green circle represents the weighted Neumann-series approximation (wNSA) method. The wNSA method is a linear iterative detection algorithm that can reduce complexity and achieve high throughput on the hardware platform, but it requires the calculation of the Gram matrix (i.e., the product of the conjugate transpose of the channel matrix and itself), and the computational burden is still heavy in large-scale antenna systems.

[0085] Figure 5 The diagram shows the BER variation curves of the present invention and existing technologies under large-scale antenna ratio and modulation order in a Rayleigh channel. FTZF is a Fourier Transform-based zero-forcing detector.

[0086] Figure 6 The diagram shows the BER variation curves of the present invention and the prior art under medium-scale antenna ratio and modulation order in a Rayleigh channel.

[0087] Figure 7 The BER performance of the WeFTNet detector of the present invention is shown when equipped with different weighting factor designs, where Geom represents the geometrically decreasing weight factor model and Indep represents the independently optimized weight factor.

[0088] Figure 8 This demonstrates the robustness of the WeFTNet detector of this invention. The WeFTNet detector is trained in a 32×12 MIMO system and applied to a 32×16 256-QAM MIMO-OFDM system employing linear state (LS, least square) channel estimation. OFDM stands for Orthogonal Frequency Division Multiplexing.

[0089] Simulation results demonstrate that the proposed WeFTNet detector exhibits excellent performance in both large-scale and medium-scale MIMO systems. In 128×64 and 256×128 256-QAM modulation systems, WeFTNet not only achieves detection performance close to MMSE, but also gains up to approximately 3.6 dB of performance compared to the traditional wNSA detector, and approximately 2.1 dB compared to WeFTMM, while reducing complexity by 20%–64%. In a 64×32 medium-scale MIMO system, the WeFTNet detector maintains a performance advantage of approximately 2.2 dB compared to linear iterative detectors such as wNSA. Furthermore, training the WeFTNet detector in a 32×12 MIMO system and applying it to a 32×16 256-QAM MIMO-OFDM system employing linear state (LS, least square) channel estimation still maintains stable detection performance, improving by approximately 3.0 dB and 5.0 dB compared to wNSA and WeFTMM, respectively, fully validating its good generalization ability.

[0090] Furthermore, this invention effectively solves the convergence instability problem by using geometrically decreasing weighting factors, achieving performance close to MMSE. Combined with an FFT-based computational structure, this invention significantly reduces hardware implementation complexity and has promising engineering application prospects.

[0091] In summary, the present invention is able to and The invention achieves a performance gain of 2.0–3.6 dB in 256-QAM modulated M-MIMO systems, while reducing complexity by 26%–64% while maintaining near-MMSE detection performance. Furthermore, after training, the invention can be transferred to systems with different antenna sizes while maintaining stable performance. Its robustness reduces the need for repeated adjustments to the detector structure due to system configuration changes, thus lowering hardware development and maintenance costs.

[0092] The transmission signal estimation apparatus provided by the present invention is described below. The transmission signal estimation apparatus described below can be referred to in correspondence with the transmission signal estimation method described above.

[0093] Figure 9 This is a schematic diagram of the transmission signal estimation device provided by the present invention, as shown below. Figure 9 As shown, the present invention also provides a signal estimation device, which includes the following modules.

[0094] Module 10 is used to determine the received signal vector, channel matrix, and noise variance; The estimation module 20 is used to input the received signal vector, the channel matrix and the noise variance into the signal estimation model to obtain the transmitted signal estimate value output by the signal estimation model; The signal estimation model is obtained by optimizing the minimum mean square error detector based on the Fourier transform algorithm and a pre-determined weighting factor, wherein the weighting factor is used to adjust the convergence step size.

[0095] As one embodiment, the system further includes a model determination module, which is used to: Based on the Fourier transform algorithm, a multi-order approximate expression for the inverse of the noise regularization matrix in the minimum mean square error detector is determined to obtain the first detector. The multi-order approximation expression is weighted based on the weighting factor to optimize the first detector, thereby obtaining the second detector; The signal estimation model is obtained based on the second detector.

[0096] As one embodiment, the model determination module is further configured to: Use the second detector as the signal estimation model, or A third detector is obtained by adjusting the multi-order approximation expression in the second detector based on a deep neural network, and the third detector is used as the signal estimation model, wherein the deep neural network is used to determine the weighting factor.

[0097] As one embodiment, the model determination module is further configured to: The noise regularization matrix is ​​defined as the Gram matrix of the channel matrix. Based on the Gram matrix and the discrete Fourier transform matrix, a multi-order approximation expression for the inverse matrix is ​​determined. The product of the multi-order approximation of the inverse matrix and the matched filter signal is used as the expression for the first detector; The matched filter signal is used to characterize the signal obtained after matching filtering the received signal.

[0098] As one embodiment, the model determination module is further configured to: The deep neural network is constructed based on the multi-order approximation expression of the inverse matrix; The deep neural network is trained offline based on transmitted signal samples, received signal samples, channel matrix, and noise variance to obtain the weighting factor.

[0099] As one embodiment, the weighting factor includes a sequence of geometrically decreasing parameters corresponding to the multi-order approximation expression of the inverse matrix, and the deep neural network includes... n Rank neural network unit, n The model determination module, having the same order as the multi-order approximation expression of the inverse matrix, is further configured to: The received signal sample, the channel matrix, and the noise variance are input into the deep neural network to determine the matched filter signal, the diagonal matrix, the discrete Fourier transform matrix, the first-order weighting factor, and... n Order-order approximate cumulative matrix; Based on the matched filter signal, diagonal matrix, discrete Fourier transform matrix, first-order weighting factor, and n An approximate cumulative matrix of order is used to determine the predicted value of the transmitted signal; The loss function value of the deep neural network is determined based on the predicted value of the transmitted signal and the mean square error of the transmitted signal sample. If the loss function value meets a preset threshold or reaches a preset number of training iterations, offline training of the deep neural network is completed. The first-order weighting factor is substituted into the geometrically decreasing parameter model to obtain the weighting factor. Otherwise, the first-order weighting factor is adjusted and the process returns to the step of inputting the received signal sample, the channel matrix, and the noise variance into the first layer of the deep neural network.

[0100] The transmission signal estimation apparatus provided by the present invention is used to execute the transmission signal estimation method described in any of the above embodiments, and has the technical effects corresponding to the transmission signal estimation method, which will not be described in detail here.

[0101] Figure 10 An example is a schematic diagram of the physical structure of a terminal device, such as... Figure 10 As shown, the terminal device may include: a processor 1010, a communications interface 1020, a memory 1030, and a communication bus 1040, wherein the processor 1010, the communications interface 1020, and the memory 1030 communicate with each other through the communication bus 1040. The processor 1010 can call logical instructions in the memory 1030 to execute the transmitted signal estimation method, the method including: Determine the received signal vector, channel matrix, and noise variance; The received signal vector, the channel matrix, and the noise variance are input into the signal estimation model to obtain the transmitted signal estimate output by the signal estimation model. The signal estimation model is obtained by optimizing the minimum mean square error detector based on the Fourier transform algorithm and a pre-determined weighting factor, wherein the weighting factor is used to adjust the convergence step size.

[0102] Furthermore, the logical instructions in the aforementioned memory 1030 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0103] On the other hand, the present invention also provides a computer program product, the computer program product comprising a computer program that can be stored on a non-transitory computer-readable storage medium, wherein when the computer program is executed by a processor, the computer is capable of executing the transmission signal estimation method provided by the above methods, the method comprising: Determine the received signal vector, channel matrix, and noise variance; The received signal vector, the channel matrix, and the noise variance are input into the signal estimation model to obtain the transmitted signal estimate output by the signal estimation model. The signal estimation model is obtained by optimizing the minimum mean square error detector based on the Fourier transform algorithm and a pre-determined weighting factor, wherein the weighting factor is used to adjust the convergence step size.

[0104] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the transmitted signal estimation method provided by the methods described above, the method comprising: Determine the received signal vector, channel matrix, and noise variance; The received signal vector, the channel matrix, and the noise variance are input into the signal estimation model to obtain the transmitted signal estimate output by the signal estimation model. The signal estimation model is obtained by optimizing the minimum mean square error detector based on the Fourier transform algorithm and a pre-determined weighting factor, wherein the weighting factor is used to adjust the convergence step size.

[0105] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0106] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0107] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for estimating transmitted signals, characterized in that, include: Determine the received signal vector, channel matrix, and noise variance; The received signal vector, the channel matrix, and the noise variance are input into the signal estimation model to obtain the transmitted signal estimate output by the signal estimation model. The signal estimation model is obtained by optimizing the minimum mean square error detector based on the Fourier transform algorithm and a pre-determined weighting factor, wherein the weighting factor is used to adjust the convergence step size.

2. The method for estimating transmitted signals according to claim 1, characterized in that, The signal estimation model is determined based on the following method: Based on the Fourier transform algorithm, a multi-order approximate expression for the inverse matrix of the noise regularization matrix in the minimum mean square error detector is determined to obtain the first detector. The multi-order approximation expression is weighted based on the weighting factor to optimize the first detector, thereby obtaining the second detector; The signal estimation model is obtained based on the second detector.

3. The method for estimating transmitted signals according to claim 2, characterized in that, The process of obtaining the signal estimation model based on the second detector includes: Use the second detector as the signal estimation model, or A third detector is obtained by adjusting the multi-order approximation expression in the second detector based on a deep neural network, and the third detector is used as the signal estimation model, wherein the deep neural network is used to determine the weighting factor.

4. The method for estimating transmitted signals according to claim 2, characterized in that, The step of determining a multi-order approximate expression for the inverse of the noise regularization matrix in the minimum mean square error detector based on the Fourier transform algorithm to obtain the first detector includes: The noise regularization matrix is ​​defined as the Gram matrix of the channel matrix. Based on the Gram matrix and the discrete Fourier transform matrix, a multi-order approximation expression for the inverse matrix is ​​determined. The product of the multi-order approximation of the inverse matrix and the matched filter signal is used as the expression for the first detector; The matched filter signal is used to characterize the signal obtained after matching filtering the received signal.

5. The method for estimating transmitted signals according to claim 3, characterized in that, The weighting factors are determined based on the following method: The deep neural network is constructed based on the multi-order approximation expression of the inverse matrix; The deep neural network is trained offline based on transmitted signal samples, received signal samples, channel matrix, and noise variance to obtain the weighting factor.

6. The method for estimating transmitted signals according to claim 5, characterized in that, The weighting factors include a sequence of geometrically decreasing parameters corresponding to the multi-order approximation expression of the inverse matrix, and the deep neural network includes... n Rank neural network unit, n The weighting factor, which has the same order as the multi-order approximation expression of the inverse matrix, is obtained by offline training of the deep neural network based on transmitted signal samples, received signal samples, the channel matrix, and noise variance, including: The received signal sample, the channel matrix, and the noise variance are input into the deep neural network to determine the matched filter signal, the diagonal matrix, the discrete Fourier transform matrix, the first-order weighting factor, and... n Order-order approximate cumulative matrix; Based on the matched filter signal, diagonal matrix, discrete Fourier transform matrix, first-order weighting factor, and n An approximate cumulative matrix of order is used to determine the predicted value of the transmitted signal; The loss function value of the deep neural network is determined based on the predicted value of the transmitted signal and the mean square error of the transmitted signal sample. If the loss function value meets a preset threshold or reaches a preset number of training iterations, offline training of the deep neural network is completed. The first-order weighting factor is substituted into the geometrically decreasing parameter model to obtain the weighting factor. Otherwise, the first-order weighting factor is adjusted and the process returns to the step of inputting the received signal sample, the channel matrix, and the noise variance into the first layer of the deep neural network.

7. A signal estimation device, characterized in that, include: The determination module is used to determine the received signal vector, channel matrix, and noise variance. The estimation module is used to input the received signal vector, the channel matrix and the noise variance into the signal estimation model to obtain the transmitted signal estimate output by the signal estimation model; The signal estimation model is obtained by optimizing the minimum mean square error detector based on the Fourier transform algorithm and a pre-determined weighting factor, wherein the weighting factor is used to adjust the convergence step size.

8. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the transmission signal estimation method as described in any one of claims 1-6.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the transmission signal estimation method as described in any one of claims 1-6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the transmission signal estimation method as described in any one of claims 1-6.

Images