Large-scale MIMO low-complexity hybrid precoding method based on HD-CNN

Abstract

The invention discloses a large-scale MIMO low-complexity hybrid precoding method based on an HD-CNN, and belongs to the technical field of HPAM attention mechanisms. According to the technical scheme, the method comprises the following steps: constructing a single-user downlink large-scale MIMO hybrid precoding communication system model, and carrying out accurate modeling by adopting a Saleh-Valenzuela channel model; the characteristics of a digital precoder and an analog precoder at the transmitting end are comprehensively considered, and the digital precoder and the analog precoder work cooperatively to form a final hybrid precoding system; determining a total target of the system according to constant modulus constraint and total transmitting power constraint conditions; a deep learning algorithm is constructed, an attention mechanism is introduced into the deep learning algorithm, and the learning efficiency of the model is improved by accurately focusing key features; and based on the constructed system model, generating experimental data for training, and optimizing hyper-parameters in combination with labels. The method has the beneficial effects that the calculation complexity and the hardware cost are remarkably reduced, the spectrum efficiency and the communication quality are greatly improved, the adaptability and the robustness of the system are enhanced, high-real-time and low-energy-consumption deployment is realized, and the method has the advantages of compatibility and expandability.

Description

Technical Field

[0001] The present invention belongs to the technical field of HPAM attention mechanism, and relates to a hybrid precoding method, and in particular to a large-scale MIMO low-complexity hybrid precoding method based on the addition of an attention mechanism and an expanded convolutional neural network strategy. Background Art

[0002] The introduction of 5G technology has brought faster communication speeds, lower latency, and higher connection density to various industries. Millimeter-wave communication, as a representative example of high-frequency wireless technology, has become one of the core enabling technologies for 5G's high-speed communications due to its higher data rates and spectral efficiency. However, in practical applications, millimeter-wave communication faces a number of significant challenges. Due to its millimeter-scale wavelength, millimeter-wave propagation is highly susceptible to obstacles and weather factors. In urban environments, obstacles such as buildings and vegetation can significantly attenuate millimeter-wave signals. Even weather conditions such as rain and fog can interfere with signal propagation, leading to unstable channels and reduced signal transmission quality. Furthermore, millimeter-wave communication has a relatively limited transmission range, with signal strength rapidly attenuating with increasing distance. This, to a certain extent, limits its coverage and application expansion. To effectively address the path loss issue in millimeter-wave communication, massive multiple-input, multiple-output (MIMO) technology has become a key solution. This technology achieves high directional gain by deploying large-scale arrays consisting of dozens or even hundreds of antennas at the transceiver end, combined with advanced precoding techniques, to compensate for signal loss along the transmission path in the millimeter wave band. In actual systems, precoding technology precisely adjusts the phase and amplitude of the antenna's transmitted and received signals to concentrate signal energy in a specific direction, not only improving signal coverage but also enhancing the system's anti-interference capabilities. In this way, the system can achieve higher spectral efficiency within limited spectrum resources, thereby meeting the stringent communication quality and capacity requirements of the 5G era.

[0003] In traditional multi-antenna MIMO systems, transceivers are typically equipped with a relatively small number of antennas, making full digital precoding a common choice. To fully realize the potential capacity gains of massive MIMO systems, each antenna typically requires a dedicated radio frequency (RF) chain. While full digital precoding offers advantages in transmission performance, the increasing number of antennas and the resulting large number of RF chains significantly increase system hardware cost and power consumption, making it difficult to implement on a large scale. To address this, the industry has proposed hybrid precoding (HP) technology, which reduces the number of RF chains to reduce system hardware complexity and power consumption, thereby improving overall system energy efficiency. Traditional hybrid precoding methods, such as those based on error minimization criteria, genetic algorithms, and alternating minimization algorithms, focus on optimizing system performance by combining digital and analog precoding. These methods are based on solid optimization theory and significantly reduce hardware cost and power consumption compared to full digital precoding. However, these methods often rely on linear or low-dimensional optimization objectives and constraints, making it difficult to fully exploit the nonlinear characteristics of complex communication environments. They also require numerous iterative operations, resulting in high computational complexity and difficulties in real-time implementation.

[0004] With the rapid rise of deep learning, data-driven deep learning hybrid precoding technology has become a research hotspot. Hybrid precoding schemes based on deep neural networks (DNNs) were once highly favored. By constructing fully connected neural networks, they can fully learn the complex mapping relationship between channel state information and precoding matrices. However, with the deepening of research and the expansion of practical application scenarios, this scheme has exposed significant problems. Fully connected neural networks contain a large number of parameters. While the extensive connections between neurons enhance the model's expressive power, they also lead to an exponential increase in the number of parameters. This directly results in excessive computational overhead, requiring significant computing resources, such as high-performance graphics processing units (GPUs), and significantly increasing training time and inference latency. To reduce the number of neural network parameters, the application of convolutional neural networks in hybrid precoding has attracted widespread attention. CNNs, through their unique convolutional layer design, utilize local connections and weight sharing mechanisms, can significantly reduce the number of parameters while extracting data features. In hybrid precoding applications, CNNs can efficiently extract key features from massive amounts of channel data and use these compact feature representations to optimize the precoding matrix, significantly reducing computational complexity while ensuring communication performance. This makes the CNN-based hybrid precoding scheme more feasible and has higher application potential in actual communication systems, and provides new ideas and directions for solving the difficulties faced by current deep learning hybrid precoding technology.

[0005] However, although these methods have overcome the shortcomings of traditional optimization algorithms to a certain extent, there are still some challenges and limitations in practical applications, which are specifically manifested in the following aspects:

[0006] (1) Computational complexity is still high: Although the CNN-based precoding method can reduce computational overhead to a certain extent, its computational complexity problem is still prominent, which is mainly caused by the traditional convolutional layer. In the traditional CNN model, the convolution layer performs convolution operations by sliding the convolution kernel on the input data. This process involves a large number of multiplication and addition operations. Especially in large-scale MIMO systems, the input data has high dimensions and large data volume, which makes the computational complexity of the traditional convolutional layer grow exponentially. In order to train the model, a large number of samples are required for multiple iterations, which leads to extremely high time and computational costs for model training, seriously affecting the real-time performance and deployment efficiency of the system. Even with the use of accelerated hardware, the increased computational complexity brought by the traditional convolutional layer is still a key bottleneck limiting its practical application.

[0007] (2) System spectrum efficiency still needs to be improved: Although the CNN-based hybrid precoding method has improved the system spectrum efficiency compared with the traditional method and the DNN-based method, the existing CNN models of the current CNN-based precoding method usually use general convolution kernels and network structures, and often do not fully consider the particularity of the features in the channel state information (CSI), and lack targeted design for the particularity of CSI features. CSI contains rich information, such as the amplitude and phase of the signal, which have different physical meanings and statistical characteristics. Because the network model fails to effectively and comprehensively extract the key features in the CSI, the generated precoding matrix cannot adapt well to channel changes, thereby limiting the further improvement of the overall system spectrum efficiency. Summary of the Invention

[0008] In order to solve the technical problems existing in the prior art, the hybrid precoding method of the present invention aims to achieve low complexity, high real-time performance and maximization of spectrum efficiency in millimeter wave massive MIMO systems, and overcome the shortcomings of existing artificial intelligence technology in computational efficiency; the main purpose of the present invention is to overcome the defects of the prior art and propose a hybrid precoding scheme based on heterogeneous pooling attention and dilated convolutional networks. The system constructed by the scheme can greatly reduce the system complexity and significantly enhance the model's ability to express channel characteristics, ultimately improving spectrum efficiency.

[0009] To achieve the above object, the present invention provides the following technical solutions:

[0010] A low-complexity hybrid precoding method for massive MIMO based on HD-CNN, with the following steps:

[0011] S1. Construct a single-user downlink massive MIMO hybrid precoding communication system model and use the Saleh-Valenzuela channel model for accurate modeling.

[0012] S2. Taking into account the characteristics of the digital precoder and analog precoder at the transmitting end, the two work together to form the final hybrid precoding system;

[0013] S3. Determine the overall system goal based on the constant modulus constraint and the total transmit power constraint;

[0014] S4. Build a deep learning algorithm and introduce an attention mechanism into it to improve the learning efficiency of the model by accurately focusing on key features;

[0015] S5. Based on the constructed system model, generate experimental data for training, and optimize hyperparameters in combination with labels to improve the algorithm's adaptability and performance in complex environments.

[0016] Furthermore, in step S1, the channel model uses the Saleh-Valenzuela channel model to generate a channel matrix as channel state information; when both the receiver and the transmitter are linear antenna arrays, the channel matrix Expressed as:

[0017]

[0018] Where P represents the number of scattering paths, and are the direction vectors of the base station end and the user end respectively; and are the departure angle and arrival angle respectively; N t Indicates the number of transmitting antennas, N r represents the number of receiving antennas, α0 represents the channel gain coefficient of the main path, a t (·) represents the array response vector of the transmitting antenna array, a r (·) represents the array response vector of the receiving antenna array, α p represents the channel gain coefficient of the scattering path, and Expressed as:

[0019]

[0020] Where d is the antenna spacing, λ is the carrier wavelength, and T represents the matrix transpose.

[0021] Furthermore, in step S2, the transmitting end adopts a hybrid precoding architecture: the base station adopts a fully connected hybrid precoding architecture, is the input signal at the transmitter, and its normalized power is At the downlink base station end, the input signal s first passes through the digital precoder Precoding is performed, and then s is passed through N RF The root RF chain is converted to an analog phase shifter through an analog precoder Transmit baseband signal; the signal vector of the transmitting antenna array can be expressed by the mathematical formula X = Fs = F RF F BB s, the received signal vector y is y=HX+n=HF RF F BB s+n; where n represents additive white Gaussian noise, which conforms to the complex Gaussian distribution, has a mean of 0 and a variance of

[0022] F RF All elements of the matrix are subject to constant modulus constraints, i.e. |[F RF ] i,j | 2 =1, j, where |[F RF ] i,j |Corresponding to the matrix F RF The size of the (i, j)th element in ; In addition, by normalizing F BB To implement the total transmit power constraint

[0023] The analog precoder and the digital precoder are calculated in an iterative manner.

[0024] Furthermore, the nonlinear modeling capability of the neural network is used to effectively solve the analog precoding matrix. After obtaining the analog precoding matrix, the digital precoder F is solved by the least square method with low complexity. BB , specifically expressed as follows:

[0025] F BB =(F RF H F RF ) -1 F RF H F opt (4)

[0026] Among them F opt is the fully digital optimal precoding matrix, obtained by SVD decomposition, and will be used as the training label in the neural network later. RF H represents the transpose of the analog precoding matrix, F RF Represents the analog precoder; as shown in formula (3), by the conjugate transpose V of the diagonal matrix Σ containing singular values and the right singular vector matrix H Obtain the fully digital optimal precoding matrix F opt ,

[0027] H=UΣV H (5)

[0028] F opt =VΣ1 / 2 (6)

[0029] Where H represents the matrix transpose operation, U represents the left singular vector matrix, and V represents the right singular vector matrix.

[0030] Furthermore, the hybrid precoding design problem is expressed as follows:

[0031]

[0032] Where ρ is the signal-to-noise ratio (SNR) value, represents the unit data flow matrix, N s represents the data flow matrix, H represents the channel matrix, F RF represents the analog precoding matrix, F BB represents the digital precoding matrix, represents the transpose of the digital precoding matrix, represents the transpose of the analog precoding matrix, H H represents the transpose of the channel matrix, Tr() represents the trace of the matrix, i represents the row of the matrix, and j represents the column of the matrix.

[0033] Furthermore, in step S4, an HD-CNN neural network is constructed, and the channel matrix is used as a neural input to adaptively generate a feasible solution of the simulated precoding matrix.

[0034] Furthermore, the HD-CNN neural network consists of an input layer, a convolutional block, a flattening layer, a heterogeneous pooling attention mechanism, and a fully connected layer.

[0035] Furthermore, first, the real and imaginary parts of the complex-valued channel matrix are separated through the input layer and reorganized into a three-dimensional real-valued matrix of channel 2 to provide a unified input representation for subsequent processing; after being processed by the input layer, the channel matrix enters the convolution block for feature extraction; the dilated convolution block includes a dilated convolution layer, an activation layer, and a pooling layer; the input first enters the dilated convolution layer, which uses 8 convolution kernels of size 3×3, dilation rate 2, and stride 1 in both processing channels to extract features from the real and imaginary parts of the input channel matrix respectively; after convolution processing, the output result is nonlinearly transformed by the Swish activation function, which is expressed as:

[0036] Swish(x)=x·σ(x) (8)

[0037] Where x is the input and σ(x) is the Sigmoid function, i.e.:

[0038]

[0039] After processing by the convolutional blocks, the flattening layer maps the multi-dimensional features into a one-dimensional feature vector. The first fully connected layer further projects the input feature vector into a high-dimensional space to extract complex channel dependencies. The output is then passed to the heterogeneous pooling attention mechanism for higher-level feature extraction.

[0040] In the heterogeneous pooling attention mechanism, the real part of the complex channel features is processed by global average pooling to extract the overall characteristics of the channel; while the imaginary part adopts the maximum pooling strategy, focusing on extracting local phase features to help the model retain significant phase changes in the signal; the input embedding vector is first converted into query, key and value through a linear layer, and then split into multiple heads to parallelly calculate the complex dot product between the query and the key to obtain the attention scores of the real and imaginary parts respectively, and normalized; then, the attention weight value vector is weighted using the attention weight to calculate the attention output; after the multi-head attention results are merged, the attention output is processed by a convolution layer with a convolution kernel size of 1 and a stride of 1; finally, the real part is pooled by average pooling, and the imaginary part is pooled by maximum pooling, and the pooled real and imaginary parts are merged to obtain the final output.

[0041] Furthermore, the second fully connected layer maps the features extracted by the attention mechanism into an approximate solution of the precoding matrix. Finally, a constant modulus constraint is introduced through a custom layer to ensure that the analog precoder output meets the actual constant modulus constraint of the communication system. This constraint transformation can be expressed in the following mathematical form:

[0042]

[0043] Among them, F RF [:,i] represents the matrix F RF In the i-th column of , ||·|| is the modulus of the vector, and c is the target amplitude constant, which is set to 1.0;

[0044] In the offline training phase, the network parameters are optimized in batches. For each training batch, the model parameters are updated by minimizing the loss function. Huber loss is selected as the loss function of the network. The expression of Huber loss function is:

[0045]

[0046] Among them, y is the true value, is the label value, and δ is a hyperparameter that controls the transition point of the loss function under large errors and is set to 1.0.

[0047] Furthermore, after the offline training of the neural network is completed, all the parameters of the network are determined and saved as a model; then in the online testing phase, the output simulated precoding matrix F is obtained from the input test channel matrix. RFAfter obtaining the analog precoding matrix, the digital precoder F is obtained by the least square method. BB .

[0048] Beneficial effects of the present invention:

[0049] Compared with the existing technology, the HD-CNN-based large-scale MIMO low-complexity hybrid precoding method described in the present invention has the following technical features and beneficial effects:

[0050] (1) Significantly reduce computational complexity and hardware costs: By introducing the dilated convolutional network (HD-CNN), the feature extraction receptive field is expanded while the number of parameters is reduced, avoiding the exponential growth of computational complexity of traditional convolutional layers under high-dimensional channel data. Combined with the heterogeneous pooling attention mechanism, the differential processing of the real part (global average pooling) and imaginary part (max pooling) of the channel further optimizes feature extraction efficiency and reduces redundant calculations. This significantly reduces the hardware resource consumption of model training and inference, shortens processing time, and improves the real-time performance of the system.

[0051] (2) Significantly improve spectrum efficiency and communication quality: By using the attention mechanism to dynamically focus on key features (such as signal amplitude and phase) in the channel state information (CSI), combined with dilated convolution to deeply explore complex channel characteristics, the generated precoding matrix is more consistent with actual channel changes. In addition, through the joint optimization of the constant modulus constraint layer and the Huber loss function, the model accurately approximates the theoretical optimal precoding scheme while ensuring signal power balance, thereby significantly improving the system's spectrum efficiency and anti-interference capabilities.

[0052] (3) Enhanced system adaptability and robustness: The heterogeneous pooling strategy differentially processes the real and imaginary channel features, fully capturing the global statistical characteristics and local phase details of the channel matrix, improving the model's adaptability to complex wireless environments (such as multipath fading and obstruction). Combined with the smooth gradient characteristics of the Swish activation function, the model training converges faster and has stronger generalization performance under different channel conditions, ensuring that the system remains stable and efficient in dynamic scenarios.

[0053] (4) Achieve high real-time performance and low energy consumption deployment: Through the decoupling design of offline training and online inference, the model parameters can be directly applied to the actual communication system after solidification. Combined with a lightweight network architecture (such as dilated convolution and pooling layer feature dimension compression), online computing latency and power consumption are significantly reduced. At the same time, the least squares method is used to quickly solve the digital precoding matrix, further simplifying the real-time calculation process of hybrid precoding, meeting the stringent requirements of 5G / 6G communications for low latency and high energy efficiency.

[0054] (5) Compatibility and scalability advantages: The system framework based on the Saleh-Valenzuela (SV) channel model can flexibly adapt to antenna arrays of different sizes and frequency band configurations, providing a scalable algorithm foundation for the future evolution of large-scale MIMO technology (such as terahertz communication and smart metasurfaces). BRIEF DESCRIPTION OF THE DRAWINGS

[0055] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the present invention will be described in detail below in combination with the accompanying drawings and detailed embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0056] in:

[0057] Figure 1 The present invention is a schematic diagram of the structure of a large-scale MIMO hybrid precoding communication system;

[0058] Figure 2 This is a schematic diagram of the structure of the massive MIMO hybrid precoding communication system of the present invention. DETAILED DESCRIPTION

[0059] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. Figure 1-2 The low-complexity hybrid precoding method for massive MIMO based on HD-CNN is further explained.

[0060] Example 1

[0061] This application proposes a hybrid precoding scheme based on heterogeneous pooling attention and dilated convolutional networks. The specific technical solutions adopted are as follows:

[0062] (1) A single-user downlink massive MIMO hybrid precoding communication system model is constructed. The Saleh-Valenzuela (SV) channel model is used for precise modeling. The characteristics of the digital precoder and analog precoder at the transmitter are comprehensively considered, and the two work together to form the final hybrid precoding system. At the same time, the overall system goal is determined based on the constant modulus constraint and the total transmit power constraint, aiming to maximize the system's spectral efficiency. Through this modeling approach, the performance of the hybrid precoding system under different channel conditions and different precoding parameter settings can be deeply analyzed, providing a theoretical basis for optimizing the precoding algorithm and thus improving the communication performance of the massive MIMO system.

[0063] (2) Build a deep learning algorithm and introduce an attention mechanism into it to improve the learning efficiency of the model by accurately focusing on key features. Based on the constructed system model, generate experimental data for training, and optimize hyperparameters based on labels to improve the algorithm's adaptability and performance in complex environments.

[0064] The final optimized precoding scheme is obtained based on the deep learning algorithm after the above training.

[0065] The above technical solution adopted by the present invention has the following advantages over the existing technology: the present invention combines deep learning, dilated convolutional networks, attention mechanisms, and Huber loss functions to construct a large-scale MIMO hybrid precoding scheme. This scheme can reduce the complexity of neural networks and enhance the ability to express neural network features, thereby improving the system's real-time performance and spectral efficiency. In addition, the present invention integrates dilated convolution and heterogeneous pooling attention mechanisms to effectively enhance the algorithm's adaptability, enabling the system to maintain high performance in complex channel environments and significantly reducing system processing time.

[0066] Example 2

[0067] The present invention provides a large-scale MIMO hybrid precoding scheme based on heterogeneous pooling attention and dilated convolutional networks:

[0068] Step 1: Establish the basic framework of the communication system and define the precoding environment and conditions. Consider the characteristics of the digital and analog precoders at the transmitter to ensure they can work together. This model provides the foundation for subsequent CNN neural network design and precoding solutions.

[0069] (1) Channel modeling: The channel model uses the Saleh-Valenzuela (SV) channel model to generate a channel matrix as the channel state information. Consider that when both the receiver and the transmitter are linear antenna arrays, the channel matrix Expressed as:

[0070]

[0071] Where P represents the number of scattering paths, and are the direction vectors of the base station end and the user end respectively. and They are departure angle and arrival angle respectively. and It can be expressed as:

[0072]

[0073] Where d is the antenna spacing and λ is the carrier wavelength.

[0074] (2) Hybrid precoding architecture at the transmitter: Considering the fully connected hybrid precoding architecture at the base station, is the input signal at the transmitter, and its normalized power is At the downlink base station end, the input signal s first passes through the digital precoder Precoding is performed, and then s is passed through N RF The root RF chain is converted to an analog phase shifter through an analog precoder Transmit baseband signal. The signal vector of the transmitting antenna array can be expressed by the mathematical formula X = Fs = F RF F BB s, the received signal vector y is y=HX+n=HF RF F BB s+n. Where n represents additive white Gaussian noise, which conforms to the complex Gaussian distribution, has a mean of 0 and a variance of

[0075] Since the analog precoder at the transmitting end is generated by an analog phase shifter, F RF All elements of the matrix should be subject to constant modulus constraints, i.e. |[F RF ] i,j | 2 =1, j, where |[F RF ] i,j |Corresponding to the matrix F RF In addition, by normalizing F BB To implement the total transmit power constraint

[0076] Analog precoders and digital precoders are usually calculated in an iterative manner, but due to the analog precoder F RF Its nonlinear characteristics and high dimensionality lead to high computational complexity. The nonlinear modeling capability of neural networks can be used to effectively solve the analog precoding matrix. After obtaining the analog precoding matrix, the digital precoder F is solved by the least square method with low complexity. BB , specifically expressed as follows:

[0077] F BB =(F RF H F RF ) -1 F RF H F opt (4)

[0078] Among them F optis the fully digital optimal precoding matrix, which is obtained by SVD decomposition and will be used as the training label in the neural network. As shown in formula (3), the conjugate transpose V of the diagonal matrix containing singular values Σ and the right singular vector matrix is obtained. H Obtain the fully digital optimal precoding matrix F opt .

[0079] H=UΣV H (5)

[0080] F opt =VΣ 1 / 2 (6)

[0081] (3) Problem Modeling: The hybrid architecture system studied in this paper aims to maximize the system's spectrum efficiency (SE) under the constraints of constant modulus and total transmit power. The hybrid precoding design problem can be expressed as follows:

[0082]

[0083] Here, ρ is the signal-to-noise ratio (SNR) value.

[0084] Since the constant modulus constraint of the phase shifter is non-convex, it is difficult to solve this problem. Therefore, based on the convolutional neural network, this paper proposes a new neural network architecture to further optimize the hybrid precoding problem.

[0085] Step 2: The present invention constructs an HD-CNN neural network, which uses the channel matrix as the neural input and adaptively generates a feasible solution to the analog precoding matrix, thereby improving the overall performance of the system. The neural network consists of an input layer, a convolution block, a flat layer, a heterogeneous pooling attention mechanism, and a fully connected layer. First, the real and imaginary parts of the complex-valued channel matrix are separated by the input layer and reorganized into a three-dimensional real-valued matrix of channel 2, providing a unified input representation for subsequent processing. After being processed by the input layer, the channel matrix enters the convolution block for feature extraction. The dilated convolution block includes an dilated convolution layer, an activation layer, and a pooling layer. The input first enters the dilated convolution layer, which uses 8 convolution kernels of size 3×3, a dilation rate of 2, and a step size of 1 in both processing channels to extract features from the real and imaginary parts of the input channel matrix respectively. After convolution processing, the output result is nonlinearly transformed by the Swish activation function, which can smooth the gradient and enhance the flow of information, thereby accelerating the convergence of the network. The Swish activation function is expressed as:

[0086] Swish(x)=x·σ(x) (8)

[0087] Where x is the input and σ(x) is the Sigmoid function, i.e.:

[0088]

[0089] The pooling layer reduces the size of the feature matrix, reduces computational complexity, retains key features, and enhances the model's robustness to input transformations, thereby helping the model maintain good performance in the face of different environments or changes.

[0090] After processing by the convolutional blocks, the flattening layer maps the multidimensional features into a one-dimensional feature vector. The first fully connected layer further projects the input feature vector into a high-dimensional space to extract complex channel dependencies. The output is then passed to the heterogeneous pooling attention mechanism for higher-level feature extraction.

[0091] In the heterogeneous pooling attention mechanism, the real part of the complex-valued channel features is processed using global average pooling (GAP), which effectively aggregates global channel statistics and extracts overall channel features. The imaginary part, on the other hand, uses a maximum pooling (MP) strategy, focusing on extracting local phase features and helping the model preserve significant phase variations in the signal. This heterogeneous pooling approach allows the model to independently optimize the real and imaginary features, more accurately fitting complex channel models and fully mining and extracting latent features in the channel matrix. The input embedding vector is first converted into a query, key, and value using a linear layer. It is then split into multiple heads, where the complex dot product between the query and key is computed in parallel. Attention scores for the real and imaginary parts are obtained and normalized. Next, the attention output is calculated using the weighted value vector using the attention weights. After the multi-head attention results are combined, the attention output is processed through a convolutional layer with a kernel size of 1 and a stride of 1. Finally, the real part is pooled using average pooling, and the imaginary part is pooled using max pooling. The pooled real and imaginary parts are then combined to form the final output.

[0092] Subsequently, the second fully connected layer maps the features extracted by the attention mechanism into an approximate solution of the precoding matrix. Finally, a custom layer introduces a constant modulus constraint to ensure that the analog precoder output meets the actual constant modulus constraint of the communication system. This constraint transformation can be expressed in the following mathematical form:

[0093]

[0094] Among them, F RF [:,i] represents the matrix F RF The i-th column of , ||·|| is the modulus (amplitude) of the vector, and c is the target amplitude constant, typically set to 1.0. This normalization process ensures that the output of the Lambda layer naturally satisfies the constant modulus constraint, ensuring that the transmission power of each column signal in the channel is balanced.

[0095] During the offline training phase, the network parameters are optimized in batches. For each training batch, the model parameters are updated by minimizing the loss function. The present invention selects Huber loss as the loss function of the network. Huber loss combines the advantages of mean square error (MSE) and absolute error (MAE) to avoid the problem of excessive error amplification when facing large errors. The expression of Huber loss function is:

[0096]

[0097] Where y is the true value, i.e., the output value of the network proposed in this invention; is the label value, i.e., the optimal simulated precoding matrix for a given channel sample; δ is a hyperparameter that controls the transition point of the loss function under large errors and is usually set to 1.0.

[0098] The implementation process of this algorithm is shown in Algorithm 1.

[0099]

[0100]

[0101] Step 3: After the offline training of the neural network is completed, all the parameters of the network (including weights and biases, etc.) are determined and saved as a model. Then, in the online testing phase, the output simulated precoding matrix F is obtained from the input test channel matrix. RF After obtaining the analog precoding matrix, the digital precoder F is solved by the least square method. BB , the specific formula is shown in formula (4).

[0102] Through the above steps, the present invention constructs a low-complexity, spectrally efficient, massive MIMO hybrid precoding system. This system can achieve high real-time communication in complex wireless communication scenarios while effectively ensuring the spectrum utilization efficiency during the communication process.

[0103] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A low-complexity hybrid precoding method for massive MIMO based on HD-CNN, characterized in that: Here are the steps: S1. Construct a single-user downlink massive MIMO hybrid precoding communication system model and use the Saleh-Valenzuela channel model for accurate modeling. S2. Taking into account the characteristics of the digital precoder and analog precoder at the transmitting end, the two work together to form the final hybrid precoding system; S3. Determine the overall system goal based on the constant modulus constraint and the total transmit power constraint; S4. Build a deep learning algorithm and introduce an attention mechanism into it to improve the learning efficiency of the model by accurately focusing on key features; S5. Based on the constructed system model, generate experimental data for training, and optimize hyperparameters in combination with labels to improve the algorithm's adaptability and performance in complex environments.

2. The low-complexity hybrid precoding method for massive MIMO based on HD-CNN according to claim 1, wherein: In step S1, the channel model uses the Saleh-Valenzuela channel model to generate a channel matrix as channel state information; when both the receiver and the transmitter are linear antenna arrays, the channel matrix Expressed as: Where P represents the number of scattering paths, and are the direction vectors of the base station end and the user end respectively; and are the departure angle and arrival angle respectively; N t Indicates the number of transmitting antennas, N r represents the number of receiving antennas, α0 represents the channel gain coefficient of the main path, a t (·) represents the array response vector of the transmitting antenna array, a r (·) represents the array response vector of the receiving antenna array, α p represents the channel gain coefficient of the scattering path, and Expressed as: Where d is the antenna spacing, λ is the carrier wavelength, and T represents the matrix transpose.

3. The low-complexity hybrid precoding method for massive MIMO based on HD-CNN according to claim 2, wherein: In step S2, the transmitting end adopts a hybrid precoding architecture: the base station adopts a fully connected hybrid precoding architecture, is the input signal at the transmitter, and its normalized power is At the downlink base station end, the input signal s first passes through the digital precoder Precoding is performed, and then s is passed through N RF The root RF chain is converted to an analog phase shifter through an analog precoder Transmit baseband signal; the signal vector of the transmitting antenna array can be expressed by the mathematical formula X = Fs = F RF F BB s, the received signal vector y is y=HX+n=HF RF F BB s+n; where n represents additive white Gaussian noise, which conforms to the complex Gaussian distribution, has a mean of 0 and a variance of F RF All elements of the matrix are subject to constant modulus constraints, i.e. where |[F RF ] i,j |Corresponding to the matrix F RF The size of the (i, j)th element in ; In addition, by normalizing F BB To implement the total transmit power constraint The analog precoder and the digital precoder are calculated in an iterative manner.

4. The low-complexity hybrid precoding method for massive MIMO based on HD-CNN according to claim 3, wherein: The nonlinear modeling capability of the neural network is used to effectively solve the analog precoding matrix. After obtaining the analog precoding matrix, the digital precoder F is solved by the least square method with low complexity. BB , specifically expressed as follows: F BB =(F RF H F RF ) -1 F RF H F opt (4) Among them F opt is the fully digital optimal precoding matrix, obtained by SVD decomposition, and will be used as the training label in the neural network later. RF H represents the transpose of the analog precoding matrix, F RF Represents the analog precoder; as shown in formula (3), by the conjugate transpose V of the diagonal matrix Σ containing singular values and the right singular vector matrix H Obtain the fully digital optimal precoding matrix F opt , H=UΣV H (5) F opt =VΣ 1 / 2 (6) Where H represents the matrix transpose operation, U represents the left singular vector matrix, and V represents the right singular vector matrix.

5. The low-complexity hybrid precoding method for massive MIMO based on HD-CNN according to claim 4, wherein: The hybrid precoding design problem is expressed as follows: Where ρ is the signal-to-noise ratio (SNR) value, represents the unit data flow matrix, N s represents the data flow matrix, H represents the channel matrix, F RF represents the analog precoding matrix, F BB represents the digital precoding matrix, represents the transpose of the digital precoding matrix, represents the transpose of the analog precoding matrix, H H represents the transpose of the channel matrix, Tr() represents the trace of the matrix, i represents the row of the matrix, and j represents the column of the matrix.

6. The low-complexity hybrid precoding method for massive MIMO based on HD-CNN according to claim 1, wherein: In step S4, an HD-CNN neural network is constructed, and the channel matrix is used as the neural input to adaptively generate a feasible solution of the analog precoding matrix.

7. The low-complexity hybrid precoding method for massive MIMO based on HD-CNN according to claim 6, wherein: The HD-CNN neural network consists of an input layer, a convolutional block, a flattening layer, a heterogeneous pooling attention mechanism, and a fully connected layer.

8. The low-complexity hybrid precoding method for massive MIMO based on HD-CNN according to claim 7, wherein: First, the real and imaginary parts of the complex-valued channel matrix are separated through the input layer and reorganized into a three-dimensional real-valued matrix of channel 2, providing a unified input representation for subsequent processing; the channel matrix is processed by the input layer and then enters the convolution block for feature extraction; the dilated convolution block includes a dilated convolution layer, an activation layer, and a pooling layer; the input first enters the dilated convolution layer, which uses 8 convolution kernels of size 3×3, dilation rate 2, and stride 1 in both processing channels to extract features from the real and imaginary parts of the input channel matrix respectively; after convolution processing, the output result is nonlinearly transformed by the Swish activation function, which is expressed as: Swish(x)=x·σ(x) (8) Where x is the input and σ(x) is the Sigmoid function, i.e.: After processing by the convolutional blocks, the flattening layer maps the multi-dimensional features into a one-dimensional feature vector. The first fully connected layer further projects the input feature vector into a high-dimensional space to extract complex channel dependencies. The output is then passed to the heterogeneous pooling attention mechanism for higher-level feature extraction. In the heterogeneous pooling attention mechanism, the real part of the complex channel features is processed by global average pooling to extract the overall characteristics of the channel; while the imaginary part adopts the maximum pooling strategy, focusing on extracting local phase features to help the model retain significant phase changes in the signal; the input embedding vector is first converted into query, key and value through a linear layer, and then split into multiple heads to parallelly calculate the complex dot product between the query and the key to obtain the attention scores of the real and imaginary parts respectively, and normalized; then, the attention weight value vector is weighted using the attention weight to calculate the attention output; after the multi-head attention results are merged, the attention output is processed by a convolution layer with a convolution kernel size of 1 and a stride of 1; finally, the real part is pooled by average pooling, and the imaginary part is pooled by maximum pooling, and the pooled real and imaginary parts are merged to obtain the final output.

9. The low-complexity hybrid precoding method for massive MIMO based on HD-CNN according to claim 8, wherein: The second fully connected layer maps the features extracted by the attention mechanism into an approximate solution of the precoding matrix. Finally, a custom layer introduces a constant modulus constraint to ensure that the analog precoder output meets the actual constant modulus constraint of the communication system. This constraint transformation can be expressed in the following mathematical form: Among them, F RF [:,i] represents the matrix F RF In the i-th column of , ||·|| is the modulus of the vector, and c is the target amplitude constant, which is set to 1.0; In the offline training phase, the network parameters are optimized in batches. For each training batch, the model parameters are updated by minimizing the loss function. Huber loss is selected as the loss function of the network. The expression of Huber loss function is: Among them, y is the true value, is the label value, and δ is a hyperparameter that controls the transition point of the loss function under large errors and is set to 1.

0.

10. The low-complexity hybrid precoding method for massive MIMO based on HD-CNN according to any one of claims 1 to 9, wherein: After the offline training of the neural network is completed, all the parameters of the network are determined and saved as a model; then in the online testing phase, the output simulated precoding matrix F is obtained from the input test channel matrix RF After obtaining the analog precoding matrix, the digital precoder F is obtained by the least square method. BB .

Images