A deep learning-assisted massive MIMO channel estimation method

Abstract

The present invention discloses a deep learning-assisted large-scale MIMO channel estimation method, which belongs to the field of channel estimation. The present invention first uses a traditional pilot-based channel estimation method to estimate a roughly estimated channel state matrix at the pilot, and then inputs the rough two-dimensional channel matrix into a neural network to obtain a refined channel estimate. The present invention can effectively reduce algorithm complexity and system overhead, while obtaining more accurate channel estimates and higher channel estimation performance. In low signal-to-noise ratio environments, the method exhibits excellent performance far exceeding that of traditional LS and MMSE channel estimation algorithms.

Description

Technical Field

[0001] The present invention belongs to the field of communication technology and relates to a channel estimation method for a wireless communication system, in particular to a deep learning-assisted large-scale multiple-input multiple-output (MIMO) communication system channel estimation method. Background Art

[0002] Massive MIMO technology is one of the emerging technologies in wireless communications in recent years. With advantages such as large aperture, high resolution, and beamforming, it can effectively improve system capacity, spectrum efficiency, and user experience rate, reduce latency and inter-user interference, enhance full-dimensional coverage, and save energy. Therefore, Massive MIMO is considered a core technology for 5G mobile communications and plays an indispensable role in emerging technologies such as the Internet of Vehicles, autonomous driving, telemedicine, and smart cities.

[0003] However, the development and application of Massive MIMO technology also faces numerous challenges. For example, because Massive MIMO technology utilizes massive antenna arrays, which differ from the antenna scale of traditional MIMO systems, the complexity of traditional channel estimation algorithms increases exponentially. Furthermore, operations such as receiver signal detection, transmitter precoding, and power matching in Massive MIMO systems all require channel estimation results. Therefore, research on channel estimation techniques in Massive MIMO systems is crucial.

[0004] The goal of channel estimation is to improve the accuracy of received signals by estimating the channel's time or frequency domain response, correcting and recovering received data. Channel estimation methods can be categorized into three types: pilot-based, blind, and semi-blind. Pilot-based methods are widely used due to their lower computational complexity and improved channel estimation performance. Classic channel estimation methods include least squares (LS) and minimum mean square error (MMSE). While the LS algorithm is simple and easy to implement, it ignores the effects of noise and has low accuracy, especially in low signal-to-noise ratio (SNR) channels. The MMSE algorithm takes noise into account, significantly improving estimation accuracy compared to the LS algorithm. However, the MMSE algorithm's algorithmic complexity increases significantly with the number of sampling points, resulting in high system overhead and difficulty in effective deployment in practical engineering applications.

[0005] At the same time, artificial intelligence (AI) technology has achieved unprecedented success in fields such as natural language processing, computer vision, speech recognition, and machine translation, naturally attracting the attention of researchers in the field of wireless communications. Researchers hope to integrate AI, particularly deep learning, with all aspects of wireless communications to create intelligent communications for future communication systems, and see this as the mainstream development direction in the post-5G era. Summary of the Invention

[0006] In order to overcome the shortcomings and defects of the existing technology, the present invention proposes a deep learning-assisted large-scale MIMO channel estimation method, which can effectively reduce the algorithm complexity and reduce system overhead, while obtaining more accurate channel estimation values ​​and higher channel estimation performance.

[0007] In order to achieve the above object, the technical solution adopted by the present invention is:

[0008] A deep learning-assisted massive MIMO channel estimation method comprises the following steps:

[0009] Step 1: Generate training data set;

[0010] Step 2: Build a MACE-Net neural network model. The model extracts noise information features and outputs a noise information matrix.

[0011] Step 3: Add Gaussian white noise to the training dataset constructed in step 1, and then input it into the MACE-Net neural network model constructed in step 2 for offline training. Observe the loss function curve to ensure that the model converges well, and save the model weight parameters.

[0012] Step 4: Use LS or MMSE channel estimation method to obtain a roughly estimated channel information matrix at the pilot;

[0013] In step 5, the roughly estimated channel information matrix obtained in step 4 is input into the trained MACE-Net neural network model to obtain the noise information matrix. The noise information matrix is ​​then subtracted from the input channel information matrix to obtain the accurate channel information estimation matrix.

[0014] Furthermore, the specific method of step 1 is:

[0015] Build a Rayleigh fading MIMO channel in MATLAB simulation software and save the channel matrix as a two-dimensional matrix as training data;

[0016] Multiple training data are combined into a training data set.

[0017] Furthermore, the MACE-Net neural network model constructed in step 2 has a total of 52 layers, the first layer is the input layer, the middle 50 layers are hidden layers, and the last layer is the output layer; the input of the network model is a two-dimensional channel matrix that has been artificially noised.

[0018] Furthermore, the input layer of the MACE-Net neural network model is a convolutional layer, which uses a convolution kernel of size 3×3 and a stride of 1. The ReLU nonlinear activation function is used after the convolution operation.

[0019] The middle 50 hidden layers are all convolutional layers, and each uses a cascade structure of two convolution kernels of size 1×3. After the convolution operation, a ReLU nonlinear activation function and a batch normalization operation are added, and then connected to the CBAM attention module;

[0020] The output layer structure is the same as the input layer, containing a convolution kernel of size 3×3 and a stride of 1, and using the ReLU nonlinear activation function after the convolution layer to output the result.

[0021] Furthermore, the loss function described in step 3 is the mean square error loss function, which is calculated by calculating the predicted value f(x i ) and the target value y i The mean of the sum of squared differences between the two is used to constrain the model training convergence, and its expression is as follows:

[0022]

[0023] Among them, f(x i ) is the predicted value, y i is the target value.

[0024] Furthermore, the model weight parameters described in step 3 include the weight and bias of the fully connected layer and the convolution kernel size, step size, and bias of the convolution layer.

[0025] Furthermore, the CBAM attention module consists of two parts: a channel attention module and a spatial attention module, which are sequentially cascaded. Given an intermediate feature map, the CBAM attention module infers the attention map in sequence according to the two independent dimensions of channel and space, and then multiplies the attention map into the input feature map for adaptive feature refinement. During this process, the channel attention module uses the feature relationship between channels to generate a channel attention feature map, which is responsible for finding the key input features that the network needs to pay special attention to. The spatial attention module generates a spatial attention feature map through the spatial relationship of the feature space, which is responsible for finding the location of important information features that need to be extracted.

[0026] The beneficial effects of the present invention are:

[0027] (1) Compared with the traditional pilot-based channel estimation algorithm, the present invention adopts an algorithm based on the MACE-Net neural network model, which concentrates the complex calculation process in the offline training process, greatly reducing the computational complexity of the algorithm in actual engineering line use and reducing system overhead.

[0028] (2) The present invention adopts a massive MIMO channel estimation algorithm assisted by a MACE-Net neural network. Compared with the traditional channel estimation method based on LS or MMSE algorithm, the NMSE performance is better and the channel estimation accuracy is higher.

[0029] (3) Compared with the current deep learning-based algorithms, the MACE-Net neural network model algorithm of the present invention has stronger feature extraction and denoising capabilities, a higher degree of channel matrix information recovery, and more accurate channel estimation results. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 This is a flowchart of a deep learning-assisted large-scale MIMO channel estimation method in an embodiment of the present invention.

[0031] Figure 2 It is a schematic diagram of the principle of the MACE-Net network model algorithm proposed in the present invention.

[0032] Figure 3 This is a schematic diagram of the principle of the CBAM attention module.

[0033] Figure 4 2 is a performance simulation result diagram of the method according to the embodiment of the present invention. DETAILED DESCRIPTION

[0034] The present invention will be described in detail below with reference to the accompanying drawings.

[0035] A deep learning-assisted massive MIMO channel estimation method uses a Rayleigh fading channel as the channel model. This method first uses traditional pilot-based channel estimation methods to obtain a rough estimate of the channel state matrix at the pilot. This rough two-dimensional channel matrix is ​​then input into a neural network to obtain a refined channel estimate. This neural network is a trained deep neural network that identifies and removes noise and bias information contained in the channel matrix, thereby refining the channel estimate.

[0036] A deep learning-assisted massive MIMO channel estimation method, characterized by comprising the following steps:

[0037] Step 1: Generate training data set;

[0038] Step 2: Build a MACE-Net neural network model. The model extracts noise information features and outputs a noise information matrix.

[0039] Step 3: Add Gaussian white noise to the training data set constructed in step 1, and then input it into the MACE-Net neural network model constructed in step 2 for training. Observe the loss function curve to ensure that the model converges well, and save the model weight parameters.

[0040] Step 4: Use the traditional LS or MMSE channel estimation method to obtain a roughly estimated channel information matrix at the pilot.

[0041] In step 5, the roughly estimated channel information matrix obtained in step 4 is input into the trained MACE-Net neural network model to obtain the noise information matrix. The noise information matrix is ​​then subtracted from the input channel information matrix to obtain the accurate channel information estimation matrix.

[0042] This method, based on deep learning theory, constructs a 52-layer neural network model. The first layer is the input layer, the middle 50 layers are hidden layers, and the final layer is the output layer. The network input is a two-dimensional channel matrix that has been artificially noised. The network employs residual learning, where the convolutional layers extract noise and bias features rather than the original information. Therefore, at the end of the network, the noise and bias features extracted by the network are subtracted from the input noisy channel matrix features. This yields a channel matrix with less noise and bias, thereby improving the accuracy of channel estimation.

[0043] The following is the simulation implementation process of the method of the present invention:

[0044] like Figure 1 As shown, this method specifically includes the following steps:

[0045] Step 1: Construct a massive MIMO channel with Rayleigh fading.

[0046] Step 2: Send the pilot signal {x1, x2, ..., x T} is used to realize channel estimation, where And||x t ||2=1,t=1,...,T,N t is the number of transmitting antennas, T is the number of pilot signals, and T=N t .

[0047] Step 3: In the i-th time frame, the T pilot signals {x1, x2, ..., x T} respectively through wireless channels Get the received signal:

[0048] yi,t =H i x t +w i,t ,t=1,2,...,T (5)

[0049] where N r is the number of antennas at the receiving end, and N r =N t , represents additive white Gaussian noise, Represents the noise power.

[0050] Step 4: All received signals {y i,1 ,y i,2 ,...,y i,T} to form a matrix, we get:

[0051] Y i =[y i,1 ,...,y i,T ]=H i [x1,...,x T ]+[w i,1 ,...,w i,T ]=H i X+W i (6)

[0052] Step 5: Based on the received signal obtained in step 4, use the LS channel estimation algorithm to estimate the channel information matrix:

[0053]

[0054] Step 6: Generate training channels based on the system model and channel model, and construct a training data set. This patent constructs a training data set containing 10,000 channel matrices.

[0055] Step 7: The method for constructing a training data set described in step 6 is specifically to build a Rayleigh fading MIMO channel, and then save the channel information in the form of a two-dimensional matrix to form training data, and then add Gaussian white noise to serve as the input of the neural network.

[0056] Step 8: Design and build the MACE-Net neural network model algorithm. This network model has a total of 52 layers: the first layer is the network input layer, the middle 50 layers are the network hidden layers, and the last layer is the network output layer. This neural network model uses the principle of residual learning. The network learns to estimate the noise and bias characteristics of the input channel matrix. The noise and bias characteristics of the network output are subtracted from the original input channel matrix characteristics to obtain an accurate channel matrix with minimal noise and bias.

[0057] like Figure 2 As shown in the figure, the model's input layer is a convolutional layer using a 3×3 kernel with a stride of 1. A ReLU nonlinear activation function is added after the convolution operation. The middle 50 layers are also convolutional layers, each using a cascaded structure of two 1×3 kernels. The convolution operation is followed by a ReLU nonlinear activation function and batch normalization, and then connected to the CBAM attention module. The output layer, like the input layer, uses a 3×3 kernel with a stride of 1 and a ReLU nonlinear activation function after the convolution operation.

[0058] Among them, the CBAM attention module is essentially an efficient feedforward convolutional neural network, which consists of two parts: the channel attention module and the spatial attention module. Figure 3 As shown. Given an intermediate feature map, the module infers the attention map in two independent dimensions, channel and space, and then multiplies the attention map into the input feature map for adaptive feature refinement. In this process, channel attention uses the feature relationship between channels to generate a channel attention feature map, which is responsible for finding the key input features that the network needs to pay special attention to. Spatial attention generates a spatial attention feature map through the spatial relationship of the feature space, which is responsible for finding the location of the important information features that need to be extracted. Spatial attention is a supplement to channel attention. Based on this idea, given an intermediate feature map, the module infers the attention map in two independent dimensions, channel and space, and then multiplies the attention map into the input feature map for adaptive feature refinement. In this process, channel attention uses the feature relationship between channels to generate a channel attention feature map, which is responsible for finding the location of the important information features that need to be extracted. Spatial attention is a supplement to channel attention. As input, the CBAM attention module gradually predicts a one-dimensional channel attention map and a 2D spatial attention map The entire attention process can be summarized as:

[0059]

[0060] Here, f represents element-wise multiplication. During this multiplication, the channel attention values ​​are propagated and replicated along the spatial dimensions, and vice versa. F' is the refined result of the final output of the CBAM module. Introducing the CBAM attention mechanism module can effectively improve the feature extraction capability of the model algorithm.

[0061] Step 9: Perform offline training on the neural network model algorithm described in step 8 on the training data set constructed in step 6, determine the model convergence by observing the loss function curve, and save the trained model weight parameters.

[0062] Among them, the loss function is the Mean Square Error (MSE) loss function, and its expression is as follows:

[0063]

[0064] Among them, f(x i ) is the predicted value, y i is the target value.

[0065] Step 10: Load the model weight parameters described in step 9, and save the channel matrix estimated by the LS or MMSE algorithm described in step 5 as a two-dimensional matrix, input it into the trained network model, and obtain an accurate channel estimation value.

[0066] Furthermore, the normalized mean square error (NMSE) can be used as an evaluation indicator. 1000 channel matrices estimated by the LS or MMSE algorithm under different antenna conditions are input respectively. Then, the NMSE between the precise channel matrix output by the model and the original input channel matrix is ​​calculated to evaluate the channel estimation performance of the model algorithm.

[0067] Figure 4 The NMSE and SNR curves of different algorithms are compared. The comparison algorithms are LS channel estimation algorithm (denoted as "LS"), MMSE channel estimation algorithm (denoted as "MMSE"), DnCNN neural network assisted channel estimation algorithm (denoted as "LS+DnCNN") and MACE-Net neural network assisted channel estimation algorithm (denoted as "LS+MACE-Net" and "MMSE+MACE-Net"). Figure 4 It can be seen that the LS+MACE-Net channel estimation algorithm and MMSE+MACE-Net channel estimation algorithm proposed in this invention both improve performance by approximately 3dB compared to the corresponding single LS and MMSE channel estimation algorithms, while the LS+MACE-Net channel estimation algorithm improves performance by approximately 2dB compared to the MMSE channel estimation algorithm. Compared to the existing DnCNN deep learning method, the MACE-Net channel estimation algorithm proposed in this invention improves performance by approximately 1dB under low signal-to-noise ratio conditions. However, as the signal-to-noise ratio increases, the performance of the DnCNN neural network model algorithm reaches a bottleneck early. In particular, under high signal-to-noise ratio conditions, the MACE-Net neural network improves performance by more than 5dB compared to the DnCNN neural network.

[0068] In summary, this paper addresses the problems of high computational complexity, high system overhead, and low estimation accuracy in traditional channel estimation methods, as well as the poor performance of existing deep learning-based channel estimation methods. By designing and building a deep neural network algorithm model, MACE-Net, we further refine the estimation results based on traditional LS or MMSE algorithm estimation results, effectively improving the accuracy of channel estimation. At the same time, we adopt an offline training and online testing model, transferring most of the model training complexity to offline, effectively reducing the algorithm computational complexity in actual use.

[0069] The neural network model of this invention uses offline training and online testing, shifting computational effort and overhead to the offline training process, significantly reducing the feedback overhead in practical applications. Both theoretical analysis and simulation results demonstrate that this invention offers significant advantages over existing algorithms, particularly in low signal-to-noise ratio environments, far surpassing the performance of traditional LS and MMSE channel estimation algorithms.

Claims

1. A deep learning-assisted massive MIMO channel estimation method, characterized in that: The following steps are involved: Step 1: Generate training data set; Step 2: Construct a MACE-Net neural network model. The model extracts noise information features and outputs a noise information matrix. The MACE-Net neural network model has 52 layers, with the first layer being the input layer, the middle 50 layers being hidden layers, and the last layer being the output layer. The input of the network model is a two-dimensional channel matrix that has been artificially noised. The input layer of the MACE-Net neural network model is a convolutional layer, which uses a convolution kernel of size 3×3 and a stride of 1. The ReLU nonlinear activation function is used after the convolution operation. The middle 50 hidden layers are all convolutional layers, and each uses a cascade structure of two convolution kernels of size 1×3. After the convolution operation, a ReLU nonlinear activation function and a batch normalization operation are added, and then connected to the CBAM attention module; The CBAM attention module consists of two parts: a channel attention module and a spatial attention module in a sequential cascade. Given an intermediate feature map, the CBAM attention module sequentially infers attention maps along two independent dimensions, channel and spatial, and then multiplies the attention maps into the input feature map for adaptive feature refinement. During this process, the channel attention module uses the feature relationship between channels to generate a channel attention feature map, which is responsible for finding the key input features that the network needs to pay special attention to. The spatial attention module generates a spatial attention feature map based on the spatial relationship of the feature space, and is responsible for finding the location of the important information features that need to be extracted; The output layer structure is the same as the input layer, containing a convolution kernel of size 3×3 and stride 1, and using the ReLU nonlinear activation function after the convolution layer to output the result; Step 3: Add Gaussian white noise to the training data set constructed in step 1, and then input it into the MACE-Net neural network model constructed in step 2 for offline training. Observe the loss function curve to ensure that the model converges well, and save the model weight parameters. The loss function is the mean square error loss function. By calculating the predicted value f(x i ) and the target value y i The mean of the sum of squared differences between the two is used to constrain the model training convergence, and its expression is as follows: Among them, f(x i ) is the predicted value, y i is the target value; The model weight parameters include the weight and bias of the fully connected layer and the convolution kernel size, step size, and bias of the convolution layer; Step 4: Use LS or MMSE channel estimation method to obtain a roughly estimated channel information matrix at the pilot; In step 5, the roughly estimated channel information matrix obtained in step 4 is input into the trained MACE-Net neural network model to obtain the noise information matrix. The noise information matrix is ​​then subtracted from the input channel information matrix to obtain the accurate channel information estimation matrix.

2. The deep learning-assisted massive MIMO channel estimation method according to claim 1, characterized in that The specific method of step 1 is: Build a Rayleigh fading MIMO channel in MATLAB simulation software and save the channel matrix as a two-dimensional matrix as training data; Multiple training data are combined into a training data set.

Images