A MIMO joint channel estimation and signal detection method based on annealing and structured covariance modeling
Patent Information
- Application Number
- CN202611030013.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-10
- Publication Date
- 2026-10-13
AI Technical Summary
传统OAMP-Net2在JCD架构下仍面临两大核心挑战：其一，信道协方差矩阵估计依赖固定结构与迹运算，不仅计算复杂度高，且难以保证矩阵正定性，导致相关信道下信道估计精度下降，进而放大检测误差；其二，JCD 迭代过程中缺乏信道侧的误差约束，易引发误差传播，同时缺少对估计可靠性的量化引导
[0055]与现有技术相比，本发明的有益效果是：针对传统 OAMP-Net2 在相关信道、高条件数信道下误差正交性假设失效、协方差估计鲁棒性差以及误差累积严重等问题。本申请在信号检测与信道估计双端引入退火离散决策。通过检测端与信道端可学习逆温度的协同调度，实现了从软决策到硬决策的数据驱动动态过渡，有效抑制了跨层干扰累积与误差传播。同时通过GS分解实现低复杂度自适应Toeplitz 协方差估计，解决相关信道下传统固定协方差矩阵建模精度不足，计算复杂度高且无法保证正定性的问题。辅以置信度加权与自适应阻尼优化，提升迭代稳定性。仿真结果表明，在联合信道估计与信号检测方面，该方法在i.i.d 瑞利MIMO、相关信道上性能均显著优于基线模型，且计算复杂度与参数量级与原模型相当，兼顾性能与工程实用性。
Smart Images
Figure CN122909072A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of MIMO joint channel estimation and signal detection, and in particular to a MIMO joint channel estimation and signal detection method based on annealing and structured covariance modeling. Background Technology
[0002] To overcome the limitations of single-input single-output (SISO) systems, multiple-input multiple-output (MIMO) technology has been developed. MIMO systems, by configuring multiple antennas at both the transmitter and receiver, can significantly improve spectral efficiency and overall communication performance through spatial multiplexing and diversity. These advantages have made MIMO one of the main research topics in the field of wireless communication and have led to its widespread application in recent years.
[0003] MIMO detectors have been continuously developed to balance performance improvements with reduced complexity. Maximum Likelihood (ML) detectors offer the best performance, but their complexity grows exponentially. Sphere Decoding (SD) detectors can reduce complexity by limiting the search scope, but their complexity remains high in high-dimensional scenarios. On the other hand, Minimum Mean Square Error (MMSE) detectors cannot withstand strong interference. These detectors have relatively low complexity but limited performance.
[0004] Approximate Message Passing (AMP) has been introduced into the field of MIMO detection. However, this algorithm can only be established under the condition that the channel matrix is independent and identically distributed (i.i.d.) Gaussian random matrix, which is not satisfied by real-world channels. Orthogonal Approximate Message Passing (OAMP) was proposed to overcome this shortcoming of AMP. It incorporates the principle of orthogonality to ensure that the errors in the iteration process are uncorrelated, thereby improving stability and robustness. Therefore, OAMP shows better adaptability than AMP when dealing with general non-i.i.d channels.
[0005] In recent years, deep learning (DL) has been introduced as an efficient tool into numerous learning tasks, such as machine learning, biology, and image recognition. In the field of MIMO detection, the earliest and most representative iterative unfolding algorithms are DetNet and MMNet. DetNet unfolds gradient projection descent and learns parameters. On the other hand, MMNet iteratively unfolds the Approximate Message Passing (AMP) algorithm to the network layer, solving the core problem of generalization of deep learning models in the communication domain.
[0006] Based on the above research, the Orthogonal Approximate Message Passing Network (OAMP-Net) and its evolution OAMP-Net2 subsequently emerged. This network introduces learnable parameters in the iterative process of the Orthogonal Approximate Message Passing (OAMP) algorithm and unfolds the algorithm into a neural network. It retains the advantages of the solid theoretical foundation of traditional algorithms and has the ability to learn from data to optimize detection performance
[14] . This architecture significantly improves the convergence speed and robustness to channel non-ideal problems. Nevertheless, in real-world scenarios, it is difficult to accurately calculate the noise variance, resulting in either over-denoising or under-denoising.
[0007] Despite the success of OAMP-Net2 in signal detection, practical systems still need to address the channel estimation problem simultaneously. Traditional receivers typically perform channel estimation and signal detection in a disjoint manner, leading to severe error propagation and performance degradation under imperfect CSI. To overcome this problem, the Joint Channel Estimation and Signal Detection (JCD) framework has emerged, feeding back the detection output to iteratively optimize the channel estimation. Traditional OAMP-Net2 under the JCD architecture still faces two major challenges: First, the channel covariance matrix estimation relies on a fixed structure and trace operations, resulting in high computational complexity and difficulty in guaranteeing matrix positive definiteness, leading to decreased channel estimation accuracy under correlated channels and amplifying detection errors; second, the lack of channel-side error constraints during JCD iteration easily triggers error propagation, and there is a lack of quantitative guidance for estimation reliability. Summary of the Invention
[0008] The purpose of this application is to provide a MIMO joint channel estimation and signal detection method based on annealing and structured covariance modeling. By coordinating the learnable inverse temperature between the detection end and the channel end, a data-driven dynamic transition from soft decision to hard decision is realized, effectively suppressing cross-layer interference accumulation and error propagation.
[0009] To achieve the above objectives, this application provides the following technical solution:
[0010] This application provides a MIMO joint channel estimation and signal detection method based on annealing and structured covariance modeling, including the following steps:
[0011] Step 1: Construct a MIMO system with M transmit antennas and N receive antennas, determine the system input-output relationship, decompose the received signal model of the pilot band and the data band, and complete the system initialization;
[0012] Step 2: Build the DA-OAMP-Net basic network architecture for joint channel estimation and signal detection under the system in Step 1, set up the signal detection module and the channel estimation module respectively, and establish the iterative feedback loop of the two modules;
[0013] Step 3: Introduce annealing discrete decision and damping factor at the signal detection end under the architecture of Step 2 to suppress the accumulation of iteration error;
[0014] Step 4: In the channel estimation end under the architecture of Step 2, an adaptive Toeplitz structured covariance estimation based on GS decomposition is introduced to construct the channel prior covariance matrix. At the same time, a channel-side annealing discrete decision module is introduced. The channel estimation end uses the detection results of Step 3 as virtual pilots and combines the pilot signals with the structured covariance to complete the LMMSE channel soft estimation.
[0015] Step 5: Feed the channel estimation results from Step 4 back to the signal detection module, introduce a confidence-weighted optimization iteration process, and output the final signal detection results and channel estimation results after multiple iterations.
[0016] The construction of a MIMO system with M transmit antennas and N receive antennas is described, and the input-output relationship of the MIMO system is expressed as follows:
[0017] (1)
[0018] The transmitted signal vector x is a column vector. Each element x j Let x represent the complex baseband signal transmitted by the j-th transmitting antenna, and x be taken from the discrete constellation diagram. Furthermore, the symbols sent must come from a finite number of complex constellation diagrams. The received signal is The channel matrix is This describes the channel response from all transmitters to all receivers, antenna to antenna, where the (i, j)th term describes the response from transmitter j to receiver i, and the additive noise vector. This represents the additive white Gaussian noise introduced at the receiver.
[0019] The core objective of a MIMO system detector is to estimate the original transmitted symbol vector from the received signal. The optimal detection criterion is the maximum likelihood (ML) criterion.
[0020] (2)
[0021] in H is the modulation constellation, H is the channel matrix, and y is the received signal.
[0022] To suppress the accumulation of iterative errors, annealed discrete decision-making is introduced at both the signal detection and channel estimation ends. Specifically, an annealed discrete noise denoiser (ADD) is synchronously introduced at both the detection and channel estimation ends to achieve soft-decision scheduling at both ends. The core idea of the annealed discrete noise denoiser is to introduce inverse temperature... This is used to adjust the "softness" or "hardness" of the noise reduction process. In a MIMO system, let the transmitted symbol vector be... ,in Representing a discrete symbol set, for QPSK, The receiving end obtains the observed values. The goal of ADD is to estimate x based on r, and its expression is as follows:
[0023] (3)
[0024] The output of the annealing discrete denoiser is determined by the inverse temperature. The weighted posterior expectation will reduce the original error variance. Replace with temperature inverse As a learnable parameter,
[0025] The OAMP-Net2-ADD network consists of T layers, each with the same structure. After inputting data, it iterates through these layers multiple times, ultimately outputting the desired estimate. The core iteration process of the network is as follows:
[0026] (4)
[0027] (5)
[0028] (6)
[0029] W t This represents the trace-normalized linear minimum mean square error (LMMSE) filter matrix in the t-th iteration.
[0030] In the nonlinear estimation module, use ADD replacement. The following expression is obtained for the nonlinear denoiser:
[0031] (7)
[0032] here Let ADD be the noise denoiser for the t-th layer, where inverse temperature... It is a learnable parameter. The original learnable parameters control the scaling ratios of the denoised output and linear compensation respectively, and both participate in the optimization. The internal extension of ADD is:
[0033] (8)
[0034] In the error variance update module (6), Indicates the expected power of the signal. This represents the expected power of the denoiser output. This method simplifies the propagation of variance and allows for the use of... Instead of complex matrix tracing operations, it reduces computational complexity.
[0035] Based on the introduction of ADD at the signal detection end, the channel estimation module synchronously introduces channel-level ADD to achieve dynamic scheduling of soft decision at both the detection and channel ends. The ADD at the channel estimation end performs annealing scheduling for the soft estimation of the channel matrix, and its core formula is:
[0036] (9)
[0037] in, The channel vectorization estimate for the t-th iteration is input to ADD by the LCR module. Let be the inverse temperature parameter that can be learned at the channel end, and be the feasible region of the channel response. For the prior distribution of the channel,
[0038] A low-complexity regularized LCR adaptive Toeplitz covariance estimation method based on Gohberg-Semencul decomposition is introduced into the channel estimation module.
[0039] (1) Gohberg-Semencul Toeplitz covariance decomposition,
[0040] The traditional diagonal covariance is replaced by a Toeplitz structure precision matrix based on GS decomposition, with the following decomposition form:
[0041] (10)
[0042] in, The matrix is a lower triangular Toeplitz matrix, and its element in the k-th row and l-th column is... k >= l, the rest are 0, from the vector Constructed according to the first column, Z tIt is a strictly lower triangular Toeplitz matrix, and its element in the k-th row and l-th column is... k > l, the rest are 0, from vector Constructed according to the first column, α i,t Here, p represents the learnable parameters of a low-order autoregressive AR model, and p is the order of the AR model. This decomposition strictly guarantees... This solves the positive definite problem of traditional covariance estimation.
[0043] (2) LCR covariance-enhanced LMMSE channel estimation,
[0044] Substituting the LCR adaptive Toeplitz covariance into the LMMSE channel estimation, we obtain the enhanced channel update formula:
[0045] (11)
[0046] (12)
[0047] in Let be the equivalent observation matrix for the t-th iteration. The noise variance is given by formula (6), where y is the received signal vector combining the pilot and data segments. The original soft channel estimate is output to the channel-side ADD.
[0048] The final channel estimate is obtained after calculation using formula (9):
[0049] (13)
[0050] (14)
[0051] in, For channel-side learnable inverse temperature, The final channel matrix for the t-th iteration is directly fed back to the signal detection end in the next round as the input channel for detection.
[0052] To further enhance iterative stability, a confidence mechanism was introduced, using the posterior probability distribution to calculate the reliability of each symbol. In subsequent channel updates, symbols with higher confidence levels will be used to construct the "virtual pilot" matrix. Signs with higher confidence levels are given greater weight, while those with lower confidence levels are given less weight by the damping factor. weaken:
[0053] (15)
[0054] This confidence-driven feedback logic effectively isolates the interference of initial detection errors on the channel refinement process, preventing detection errors from contaminating the channel estimation.
[0055] Compared with existing technologies, the advantages of this invention are as follows: It addresses the problems of traditional OAMP-Net2, such as the failure of error orthogonality assumptions, poor robustness of covariance estimation, and severe error accumulation in correlated and high condition number channels. This application introduces annealed discrete decision-making at both the signal detection and channel estimation ends. Through the cooperative scheduling of learnable inverse temperature at the detection and channel ends, a data-driven dynamic transition from soft to hard decision-making is achieved, effectively suppressing cross-layer interference accumulation and error propagation. Simultaneously, low-complexity adaptive Toeplitz covariance estimation is achieved through GS decomposition, solving the problems of insufficient accuracy, high computational complexity, and inability to guarantee positive definiteness in traditional fixed covariance matrix modeling under correlated channels. Confidence weighting and adaptive damping optimization are further enhanced to improve iterative stability. Simulation results show that, in terms of joint channel estimation and signal detection, this method significantly outperforms the baseline model on iid Rayleigh MIMO and correlated channels, while maintaining comparable computational complexity and parameter magnitude, thus balancing performance and engineering practicality. Attached Figure Description
[0056] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.
[0057] Figure 1 This is a schematic diagram of the OAMP-Net2-ADD architecture provided in an embodiment of the present invention.
[0058] Figure 2 This is a flowchart of DA-OAMP-Net provided in an embodiment of the present invention.
[0059] Figure 3 This is a performance comparison chart between the iid Rayleigh fading channel DA-OAMP-Net architecture and the traditional OAMP-Net2-JCD architecture provided in this embodiment of the invention.
[0060] Figure 4 This is a comparison chart of the BER performance between the DA-OAMP-Net architecture provided in this embodiment of the invention and the traditional OAMP-Net2-JCD architecture. Detailed Implementation
[0061] The technical solutions of the embodiments of this application will now be described with reference to the accompanying drawings. It should be noted that similar reference numerals and letters in the following drawings indicate similar items; therefore, once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings.
[0062] The terms “comprising,” “including,” or any other variations thereof are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase “comprising one…” does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0063] The terms “first,” “second,” etc., are used only to distinguish one entity or operation from another, and should not be construed as indicating or implying relative importance, nor as requiring or implying any such actual relationship or order between these entities or operations.
[0064] This embodiment is achieved through the following technical solutions, such as... Figure 1 and Figure 2 As shown, a joint MIMO channel estimation and signal detection method based on annealing and structured covariance modeling includes the following steps:
[0065] Construct a MIMO system with M transmit antennas and N receive antennas;
[0066] For the enhanced network architecture DA-OAMP-Net used for joint channel estimation and detection of MIMO, annealing discrete decision is introduced at both the signal detection and channel estimation ends to suppress the accumulation of iteration error.
[0067] The construction of a MIMO system with M transmit antennas and N receive antennas is described, and the input-output relationship of the MIMO system is expressed as follows:
[0068] (1)
[0069] The transmitted signal vector x is a column vector. Each element x j Let x represent the complex baseband signal transmitted by the j-th transmitting antenna, and x be taken from the discrete constellation diagram. Furthermore, the symbols sent must come from a finite number of complex constellation diagrams. The received signal is The channel matrix is This describes the channel response from all transmitters to all receivers, antenna to antenna, where the (i, j)th term describes the response from transmitter j to receiver i, and the additive noise vector. This represents the additive white Gaussian noise introduced at the receiver.
[0070] The core objective of a MIMO system detector is to estimate the original transmitted symbol vector from the received signal. The optimal detection criterion is the maximum likelihood (ML) criterion.
[0071] (2)
[0072] in H is the modulation constellation, H is the channel matrix, and y is the received signal.
[0073] To suppress the accumulation of iterative errors, annealed discrete decision-making is introduced at both the signal detection and channel estimation ends. Specifically, an annealed discrete noise denoiser (ADD) is synchronously introduced at both the detection and channel estimation ends to achieve soft-decision scheduling at both ends. The core idea of the annealed discrete noise denoiser is to introduce inverse temperature... This is used to adjust the "softness" or "hardness" of the noise reduction process. In a MIMO system, let the transmitted symbol vector be... ,in Representing a discrete symbol set, for QPSK, The receiving end obtains the observed values. The goal of ADD is to estimate x based on r, and its expression is as follows:
[0074] (3)
[0075] The output of the annealing discrete denoiser is determined by the inverse temperature. The weighted posterior expectation will reduce the original error variance. Replace with temperature inverse As a learnable parameter,
[0076] The OAMP-Net2-ADD network consists of T layers, each with the same structure. After inputting data, it iterates through these layers multiple times, ultimately outputting the desired estimate. The core iteration process of the network is as follows:
[0077] (4)
[0078] (5)
[0079] (6)
[0080] W t This represents the linear minimum mean square error (LMMSE) matrix.
[0081] In the nonlinear estimation module, use ADD replacement. The following expression is obtained for the nonlinear denoiser:
[0082] (7)
[0083] here Let ADD be the noise denoiser for the t-th layer, where inverse temperature... It is a learnable parameter. The original learnable parameters control the scaling ratios of the denoised output and linear compensation respectively, and both participate in the optimization. The internal extension of ADD is:
[0084] (8)
[0085] In the error variance update module (6), Indicates the expected power of the signal. This represents the expected power of the denoiser output. This method simplifies the propagation of variance and allows for the use of... Instead of complex matrix tracing operations, it reduces computational complexity.
[0086] This network integrates ADD without changing the iterative logic of the linear estimator. We replaced the original learnable parameters in OAMP-Net2 with inverse temperature. It is trained together with other learnable parameters. This eliminates the computationally intensive and matrix mismatch-sensitive tracking operation in OAMP-Net2, which not only significantly reduces the computational complexity of the model, but also enhances the model's robustness under complex channels. After training, the strength and variance propagation of nonlinear denoising can be controlled simultaneously, enabling the ADD framework to more accurately approximate the MAP performance of discrete signal detection, thus typically outperforming the traditional OAMP-Net2 in terms of bit error rate (BER). This method significantly improves the efficiency, flexibility, and final detection performance of the nonlinear denoising module.
[0087] Based on the introduction of ADD at the signal detection end, the channel estimation module synchronously introduces channel-level ADD to achieve dynamic scheduling of soft decision at both the detection and channel ends. The ADD at the channel estimation end performs annealing scheduling for the soft estimation of the channel matrix, and its core formula is:
[0088] (9)
[0089] in, The channel vectorization estimate for the t-th iteration is input to ADD by the LCR module. Let be the inverse temperature parameter that can be learned at the channel end, and be the feasible region of the channel response. For the prior distribution of the channel,
[0090] A low-complexity regularized LCR adaptive Toeplitz covariance estimation method based on Gohberg-Semencul decomposition is introduced into the channel estimation module.
[0091] (1) Gohberg-Semencul Toeplitz covariance decomposition,
[0092] The traditional diagonal covariance is replaced by a Toeplitz structure precision matrix based on GS decomposition, with the following decomposition form:
[0093] (10)
[0094] in, The matrix is a lower triangular Toeplitz matrix, and its element in the k-th row and l-th column is... k >= l, the rest are 0, from the vector Constructed according to the first column, Z t It is a strictly lower triangular Toeplitz matrix, and its element in the k-th row and l-th column is... k > l, the rest are 0, from vector Constructed according to the first column, α i,t Here, p represents the learnable parameters of a low-order autoregressive AR model, and p is the order of the AR model. This decomposition strictly guarantees... This solves the positive definite problem of traditional covariance estimation.
[0095] (2) LCR covariance-enhanced LMMSE channel estimation,
[0096] Substituting the LCR adaptive Toeplitz covariance into the LMMSE channel estimation, we obtain the enhanced channel update formula:
[0097] (11)
[0098] (12)
[0099] in Let be the equivalent observation matrix for the t-th iteration. The noise variance is given by formula (6), where y is the received signal vector combining the pilot and data segments. The original soft channel estimate is output to the channel-side ADD.
[0100] The final channel estimate is obtained after calculation using formula (9):
[0101] (13)
[0102] (14)
[0103] in, For channel-side learnable inverse temperature, The final channel matrix for the t-th iteration is directly fed back to the signal detection end in the next round as the input channel for detection.
[0104] To further enhance iterative stability, a confidence mechanism was introduced, using the posterior probability distribution to calculate the reliability of each symbol. In subsequent channel updates, symbols with higher confidence levels will be used to construct the "virtual pilot" matrix. Signs with higher confidence levels are given greater weight, while those with lower confidence levels are given less weight by the damping factor. weaken:
[0105] (15)
[0106] This confidence-driven feedback logic effectively isolates the interference of initial detection errors on the channel refinement process, preventing detection errors from contaminating the channel estimation.
[0107] The overall computational complexity of DA-OAMP-Net in this application is: The core overhead comes from the matrix inversion operation of each LMMSE matrix. In the signal detection module, the complexity of matrix inversion is O(n log n). Learnable inverse temperature This replaces the variance estimation module in OAMP-Net2 that relies on matrix tracking operations, reducing the actual computational cost per layer by approximately O(TN). t This reduces hardware resource consumption and computational latency. The newly added adaptive damping term only introduces O(T) level scalar computation (which will be mentioned in the experimental section). In the channel estimation module, LCR adaptive covariance estimation reduces the covariance update complexity of each layer from... Down to The actual computational cost is reduced by approximately Two-level ADD and confidence-weighted ADD only introduce... The extra computations at this level have a negligible impact on the overall complexity.
[0108] In terms of parameter scale, the OAMP-Net2-ADD detector has only 4T trainable parameters (each layer includes...). ), channel-side ADD AR parameters of LCR covariance The total parameter scale is 6T+(p+1), completely independent of the number of antennas N. t and N rThe number of antennas is determined solely by the number of iteration layers T and the AR order p. This characteristic allows it to operate in large-scale MIMO scenarios (such as 16 × 16 and 32 × 32 configurations) without requiring additional parameter or training data expansion due to the increase in the number of antennas.
[0109] We implemented the DA-OAMP-Net architecture on a 4×4 MIMO system using TensorFlow, with 6 layers (T), a QPSK modulated signal as input (x), and 1000 iterations. Training data was randomly generated, and parameters were updated using a gradient descent optimizer during training. The learning rate was set to 0.0005 for low SNR and 0.0001 for high SNR (30dB-35B). Furthermore, to improve iterative stability and convergence smoothness, an adaptive damping mechanism was introduced during signal detection iteration. This mechanism uses historical iteration results to weight and smooth the current detected symbol, effectively suppressing error propagation and numerical oscillations, ensuring the model's training and inference stability in low SNR and correlated channel scenarios.
[0110] The loss function adopts a multi-task joint loss function (JointLoss), which optimizes both channel estimation accuracy and signal detection accuracy, achieving synergistic enhancement of the two tasks.
[0111]
[0112] in Let be the mean square error of the channel estimation at layer t. For the L2 loss of signal detection, These are configurable hierarchical weight coefficients.
[0113] The system signal-to-noise ratio (SNR) is defined as:
[0114]
[0115] in, Let H represent the mathematical expectation, and let H be the channel matrix, x d Let n be the transmitted signal vector. d The numerator represents the average power of the effective signal after transmission through the channel, and the denominator is the average power of the noise. The larger the SNR value, the weaker the signal is affected by noise, and the better the system's transmission and detection performance.
[0116] (1) iidMIMO Rayleigh fading channel
[0117] Figure 3This paper presents a performance comparison between our proposed DA-OAMP-Net architecture and the traditional OAMP-Net2-JCD architecture under an iid MIMO Rayleigh fading channel. It can be seen that our network achieves a significant performance improvement in 4×4 MIMO systems under this channel condition. Specifically, when our target is 5×10... -2 When the BER was reduced, the SNR improved by approximately 3 dB.
[0118] (2) Related channels
[0119] Figure 4 This paper presents a comparison of the BER performance of our proposed DA-OAMP-Net architecture with the traditional OAMP-Net2-JCD architecture in a correlation channel when c=0.5. We employ the Kronecker correlation model. With a high correlation coefficient, a strong positive correlation exists between channels, which is more common in real-world scenarios. Compared to the previously mentioned iid MIMO Rayleigh fading channel, the DA-OAMP-Net architecture demonstrates superior performance in the correlation channel. Specifically, when our target is 5×10... -2 When the BER was reduced, the SNR increased by approximately 5 dB.
[0120] This application proposes an enhanced OAMP-Net2-ADD-JCD framework based on two-stage annealed discrete decision-making and LCR adaptive covariance estimation, targeting JCD scenarios. It addresses the problems of traditional OAMP-Net2, such as the failure of error orthogonality assumptions, poor robustness of covariance estimation, and severe error accumulation in correlated and high condition number channels. The proposed method introduces annealed discrete decision-making at both the signal detection and channel estimation ends. Through cooperative scheduling with learnable inverse temperatures at the detection and channel ends, a data-driven dynamic transition from soft to hard decision-making is achieved, effectively suppressing cross-layer interference accumulation and error propagation. Simultaneously, low-complexity adaptive Toeplitz covariance estimation is achieved through GS decomposition, solving the problems of insufficient accuracy, high computational complexity, and inability to guarantee positive definiteness in traditional fixed covariance matrix modeling under correlated channels. Confidence weighting and adaptive damping optimization are further employed to improve iterative stability. Simulation results show that, in terms of joint channel estimation and signal detection, this method significantly outperforms the baseline model on iid Rayleigh MIMO and related channels, while maintaining comparable computational complexity and parameter magnitude to the original model, thus balancing performance and engineering practicality.
[0121] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.
Claims
1. A joint MIMO channel estimation and signal detection method based on annealing and structured covariance modeling, characterized in that, Includes the following steps: Step 1: Construct a MIMO system with M transmit antennas and N receive antennas, determine the system input-output relationship, decompose the received signal model of the pilot band and the data band, and complete the system initialization; Step 2: Build the DA-OAMP-Net basic network architecture for joint channel estimation and signal detection under the system in Step 1, set up the signal detection module and the channel estimation module respectively, and establish the iterative feedback loop of the two modules; Step 3: Introduce annealing discrete decision and damping factor at the signal detection end under the architecture of Step 2 to suppress the accumulation of iteration error; Step 4: In the channel estimation end under the architecture of Step 2, an adaptive Toeplitz structured covariance estimation based on GS decomposition is introduced to construct the channel prior covariance matrix. At the same time, a channel-side annealing discrete decision module is introduced. The channel estimation end uses the detection results of Step 3 as virtual pilots and combines the pilot signals with the structured covariance to complete the LMMSE channel soft estimation. Step 5: Feed the channel estimation results from Step 4 back to the signal detection module, introduce a confidence-weighted optimization iteration process, and output the final signal detection results and channel estimation results after multiple iterations.
2. The MIMO joint channel estimation and signal detection method based on annealing and structured covariance modeling as described in claim 1, characterized in that, In step 1, the input-output relationship of the MIMO system is expressed as follows: （1） The transmitted signal vector x is a column vector. Each element x j Let x represent the complex baseband signal transmitted by the j-th transmitting antenna, and x be taken from the discrete constellation diagram. Furthermore, the symbols sent must come from a finite number of complex constellation diagrams. The received signal is The channel matrix is This describes the channel response from all transmitters to all receivers, antenna to antenna, where the (i, j)th term describes the response from transmitter j to receiver i, and the additive noise vector. This represents the additive white Gaussian noise introduced at the receiver.
3. The MIMO joint channel estimation and signal detection method based on annealing and structured covariance modeling according to claim 2, characterized in that, The core objective of a MIMO system detector is to estimate the original transmitted symbol vector from the received signal. The optimal detection criterion is the maximum likelihood (ML) criterion. （2） in H is the modulation constellation, H is the channel matrix, and y is the received signal.
4. The MIMO joint channel estimation and signal detection method based on annealing and structured covariance modeling according to claim 3, characterized in that, To suppress the accumulation of iteration errors, annealed discrete decision-making is introduced at both the signal detection and channel estimation ends. Specifically, an annealed discrete noise denoiser (ADD) is simultaneously introduced at both the detection and channel estimation ends to achieve soft-decision scheduling at both ends, resulting in the OAMP-Net2-ADD network. The core idea of the annealed discrete noise denoiser is to introduce inverse temperature... This is used to adjust the "softness" or "hardness" of the noise reduction process. In a MIMO system, let the transmitted symbol vector be... ,in Representing a discrete symbol set, for QPSK, The receiving end obtains the observed values. The goal of ADD is to estimate x based on r, and its expression is as follows: （3） The output of the annealing discrete denoiser is determined by the inverse temperature. The weighted posterior expectation will reduce the original error variance. Replace with temperature inverse As a learnable parameter, The OAMP-Net2-ADD network consists of T layers, each with the same structure. After inputting data, it iterates through these layers multiple times, ultimately outputting the desired estimate. The core iteration process of the network is as follows: （4） （5） （6） W t This represents the trace-normalized linear minimum mean square error (LMMSE) filter matrix in the t-th iteration. In the nonlinear estimation module, use ADD replacement. The following expression is obtained for the nonlinear denoiser: （7） here Let ADD be the noise denoiser for the t-th layer, where inverse temperature... It is a learnable parameter. The original learnable parameters control the scaling ratios of the denoised output and linear compensation respectively, and both participate in the optimization. The internal extension of ADD is: （8） In the error variance update module (6), Indicates the expected power of the signal. This represents the expected power of the denoiser output. This method simplifies the propagation of variance and allows for the use of... Instead of complex matrix tracing operations, it reduces computational complexity.
5. The MIMO joint channel estimation and signal detection method based on annealing and structured covariance modeling according to claim 4, characterized in that, Based on the introduction of ADD at the signal detection end, the channel estimation module synchronously introduces channel-level ADD to achieve dynamic scheduling of soft decision at both the detection and channel ends. The ADD at the channel estimation end performs annealing scheduling for the soft estimation of the channel matrix, and its core formula is: （9） in, The channel vectorization estimate for the t-th iteration is input to ADD by the LCR module. Let be the inverse temperature parameter that can be learned at the channel end, and be the feasible region of the channel response. For the prior distribution of the channel, A low-complexity regularized LCR adaptive Toeplitz covariance estimation method based on Gohberg-Semencul decomposition is introduced into the channel estimation module. (1) Gohberg-Semencul Toeplitz covariance decomposition, The traditional diagonal covariance is replaced by a Toeplitz structure precision matrix based on GS decomposition, with the following decomposition form: （10） in, The matrix is a lower triangular Toeplitz matrix, and its element in the k-th row and l-th column is... k >= l, the rest are 0, from the vector Constructed according to the first column, Z t It is a strictly lower triangular Toeplitz matrix, and its element in the k-th row and l-th column is... k > l, the rest are 0, from vector Constructed according to the first column, α i,t Here, p represents the learnable parameters of a low-order autoregressive AR model, and p is the order of the AR model. This decomposition strictly guarantees... This solves the positive definite problem of traditional covariance estimation. (2) LCR covariance-enhanced LMMSE channel estimation, Substituting the LCR adaptive Toeplitz covariance into the LMMSE channel estimation, we obtain the enhanced channel update formula: （11） （12） in Let be the equivalent observation matrix for the t-th iteration. The noise variance is given by formula (6), where y is the received signal vector combining the pilot and data segments. The original soft channel estimate is output to the channel-side ADD. The final channel estimate is obtained after calculation using formula (9): （13） （14） in, For channel-side learnable inverse temperature, The final channel matrix for the t-th iteration is directly fed back to the signal detection end in the next round as the input channel for detection.
6. The MIMO joint channel estimation and signal detection method based on annealing and structured covariance modeling according to claim 5, characterized in that, To further enhance iterative stability, a confidence mechanism was introduced, using the posterior probability distribution to calculate the reliability of each symbol. In subsequent channel updates, symbols with higher confidence levels will be used to construct the "virtual pilot" matrix. Signs with higher confidence levels are given greater weight, while those with lower confidence levels are given less weight by the damping factor. weaken: （15） This confidence-driven feedback logic effectively isolates the interference of initial detection errors on the channel refinement process, preventing detection errors from contaminating the channel estimation.