Low-bit quantization large-scale MIMO scattering channel estimation method based on deep learning

Abstract

The invention discloses a low-bit quantization large-scale MIMO scattering channel estimation method based on deep learning, and belongs to the technical field of signal detection. Deep learning is introduced into low-bit quantization large-scale MIMO scattering channel estimation, and a least square channel estimation result is used as input of a neural network, so that the convergence rate of the network is improved; the network structure designed by the invention can effectively process a one-bit quantization scene and a two-bit quantization scene with significant difference without structural adjustment, and meanwhile, key information related to a real channel state is extracted from highly-distorted low-bit quantization data by utilizing a multi-layer neural network, so that high-precision channel estimation is finally realized. According to the invention, the method of offline training and online testing is adopted, so that the calculation complexity of the algorithm is reduced. A simulation result shows that the method has obvious advantages compared with an existing algorithm.

Description

Technical Field

[0001] This invention relates to the field of signal detection technology for wireless communication systems, and more specifically to a low-bit quantization large-scale MIMO (Multiple-in Multiple-out) scattering channel estimation method based on deep learning. Background Technology

[0002] Tropospheric scattering communication is a type of beyond-line-of-sight wireless communication that utilizes the scattering of radio waves by the inhomogeneities of the troposphere. Tropospheric scattering transmission systems designed using the tropospheric scattering propagation mechanism can achieve beyond-line-of-sight transmission while possessing moderate transmission capacity, performance, and reliability, as well as strong resistance to nuclear explosions and ionospheric disturbances. However, the inherently high transmission loss characteristics of scattering channels limit the support of high-speed services for scattering communication systems. To further improve the spectral efficiency of existing scattering communication systems, massive MIMO technology is introduced into scattering communication, leveraging the multiplexing gain of massive MIMO systems to increase the capacity of scattering communication systems.

[0003] Large-scale MIMO systems employ high-bit (greater than 8-bit) analog-to-digital converters (ADCs) to improve system performance. However, the power loss of ADCs increases exponentially with the number of quantization bits, leading to a sharp increase in hardware costs. Low-bit ADCs, on the other hand, can effectively reduce system cost and power consumption and are easy to implement, making them a viable solution. Therefore, researching large-scale MIMO scattering channel estimation methods based on low-bit quantization to mitigate the impact of low-bit quantization and improve channel estimation accuracy is of great significance.

[0004] Traditional channel estimation algorithms do not consider the existence of quantization error, but low-bit quantization will bring a large quantization error, which will significantly degrade the performance of channel estimation algorithms such as least squares (LS) and gradient descent (GD), making them difficult to use in practical systems. Summary of the Invention

[0005] To address the performance degradation caused by low-bit quantization in existing channel estimation algorithms, this invention proposes a deep learning-based method for estimating large-scale MIMO scattering channels with low bit quantization. This invention utilizes deep neural networks to eliminate the impact of low quantization on system performance, thereby improving the accuracy of scattering channel estimation. Simultaneously, it employs offline training and online detection to reduce computational complexity.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A deep learning-based method for estimating low-bit quantization large-scale MIMO scattering channels includes the following steps:

[0008] Step 1: The transmitting end sends a pilot signal x, which reaches the signal receiving end through the scattering channel H. The scattering channel H is randomly generated using the tropospheric scattering communication 7-path channel model, and the noise n in the communication process is additive complex white Gaussian noise.

[0009] Step 2: At the receiving end, perform low-bit quantization on the received signal y, including one-bit quantization and two-bit quantization.

[0010] Step 3: Perform LS channel estimation on the received signals after one-bit quantization and two-bit quantization, and separate the real and imaginary parts of the channel estimation results to create training and test datasets for one-bit quantization and two-bit quantization.

[0011] Step 4: Construct a deep learning-based low-bit quantization large-scale MIMO scattering channel estimation neural network model, which includes the first to third convolutional layers, the first to third residual structures, and the fourth to sixth convolutional layers. The first to third convolutional layers are used to extract data features from the input data. The extracted data features are then input into the first to third residual structures to further extract different data features. Finally, the fourth to sixth convolutional layers are used for further feature extraction to match the output dimension of the data. The kernel size of each convolutional layer is 3×3.

[0012] Step 5: Input the training dataset and test dataset obtained in Step 3 into the neural network model constructed in Step 4 for training, and update the network weights and offsets using backpropagation gradient descent.

[0013] Step 6: Place the trained neural network model at the receiver of the low-bit massive MIMO scattering communication system. After the pilot received signal is quantized with low bits, channel estimation is performed. The channel estimation result is then separated into virtual and real signals and input into the neural network model for channel estimation. Finally, the channel estimation result is obtained.

[0014] In step 2, the low-bit quantized signal y_Q is:

[0015] y_Q=Q(y)=Q(Hx+n)

[0016] In the formula, Q(·) represents the low-bit quantization operator that operates on each element in the vector, including one-bit quantization Q1(·) and two-bit quantization Q2(·);

[0017] One bit is quantized as:

[0018] Q1(x)=sgn(Re{x})+jsgn(Im{x})

[0019] In the formula, Re{x} and Im{x} represent the real and imaginary parts of x, respectively, and sgn(·) represents the sign function, defined as:

[0020]

[0021] Two-bit quantization is:

[0022]

[0023] The present invention has the following beneficial effects:

[0024] (1) Compared with traditional low-bit quantization large-scale MIMO scattering channel estimation algorithms, this invention adopts an offline training and online detection method, which reduces the computational complexity of the algorithm and improves the channel estimation performance.

[0025] (2) Compared with existing deep learning-based low-bit quantization large-scale MIMO scattering channel estimation algorithms, the neural network model proposed in this invention is applicable not only to one-bit quantization large-scale MIMO scattering channel estimation, but also to two-bit quantization large-scale MIMO scattering channel estimation, and guarantees the channel estimation accuracy. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the process of a low-bit quantization large-scale MIMO scattering channel estimation method based on deep learning proposed in this invention.

[0027] Figure 2 This is a schematic diagram of the neural network model structure of a low-bit quantization large-scale MIMO scattering channel estimation method based on deep learning proposed in this invention.

[0028] Figure 3 This is a curve showing the relationship between NMSE and SNR for different algorithms under one-bit quantization.

[0029] Figure 4 These are the NMSE and SNR curves for different algorithms under two-bit quantization. Detailed Implementation

[0030] The present invention will now be described in detail with reference to the accompanying drawings.

[0031] A deep learning-based method for estimating large-scale MIMO scattering channels with low bit quantization is proposed. First, a large amount of simulated scattering channel data is generated based on existing scattering channel models. The received signal is then low-bit quantized. Next, the channel estimation is performed using a traditional least squares channel estimation algorithm. Then, the channel estimation value is used as input, and the real and imaginary parts are separated and fed into a neural network. The neural network is used to eliminate the impact of low quantization on system performance and improve the accuracy of channel estimation.

[0032] The specific steps of this method are as follows:

[0033] Step 1: First, construct a large-scale MIMO scattering communication system architecture based on low-bit quantization. The transmitter sends a pilot signal x, which reaches the receiver through the scattering channel H. The scattering channel H is randomly generated using a tropospheric scattering communication 7-path channel model, and the noise n in the communication process is additive complex white Gaussian noise.

[0034] Step 2: At the receiving end, perform low-bit quantization on the received signal y. The low-bit quantized signal y_Q is:

[0035] y_Q=Q(y)=Q(Hx+n)

[0036] In the formula, Q(·) represents the low-bit quantization operator that operates on each element in the vector, including one-bit quantization Q1(·) and two-bit quantization Q2(·);

[0037] One bit is quantized as:

[0038] Q1(x)=sgn(Re{x})+jsgn(Im{x})

[0039] In the formula, Re{x} and Im{x} represent the real and imaginary parts of x, respectively, and sgn(·) represents the sign function, defined as:

[0040]

[0041] Two-bit quantization is:

[0042]

[0043] Step 3: Perform LS channel estimation on the received signals after one-bit quantization and two-bit quantization, and separate the real and imaginary parts of the channel estimation results to create training and test datasets H for one-bit quantization and two-bit quantization. ls :

[0044]

[0045] In the formula, h(n) represents the nth channel estimation result, and Re(x) and Im(x) represent the real and imaginary parts of h(n), respectively;

[0046] Step 4: Construct a deep learning-based neural network model for estimating low-bit quantization large-scale MIMO scattering channels. The neural network model is as follows: Figure 2 As shown, the model sequentially includes first to third convolutional layers, first to third residual structures, and fourth to sixth convolutional layers. The residual structures consist of four convolutional layers. After dimension matching, the input data is fed into the neural network model. The data first passes through the first to third convolutional layers for feature extraction. Then, the extracted features are input into the first to third residual structures to further extract different features for channel reconstruction, improving learning efficiency and mitigating the vanishing gradient problem. Finally, the fourth to sixth fully convolutional layers further extract features, improving the network's denoising capability. Ultimately, the output dimension of the data is matched, eliminating the impact of low quantization on system performance and improving channel estimation accuracy.

[0047] Step 5: Input the training and test datasets obtained in Step 3 into the neural network model constructed in Step 4 for training, and update the weights and biases using backpropagation gradient descent. The loss function used in updating the weights and biases using backpropagation gradient descent is represented by the mean squared error between the true channel and the estimated channel; the optimizer is the Adam optimizer, and the activation function is ReLU; the algorithm performance is measured by the normalized mean squared error, defined as:

[0048]

[0049] In the formula, h and These represent the actual channel and the estimated channel, respectively.

[0050] Step 6: Place the trained neural network model at the receiver of the low-bit massive MIMO scattering communication system. After low-bit quantization, the pilot received signal is input into the LS channel estimator for channel estimation. The channel estimation result is then separated into virtual and real signals and input into the neural network model for further channel estimation. The neural network is used to eliminate the impact of low quantization on system performance, improve the channel estimation accuracy, and finally obtain a high-precision channel estimation result. The algorithm flow is as follows: Figure 1 As shown.

[0051] To better demonstrate the performance of the method proposed in this patent, the relationship between NMSE and SNR for different algorithms was tested under one-bit quantization and two-bit quantization. Figure 3 These are the NMSE and SNR curves for different algorithms under one-bit quantization. Figure 4 These are the NMSE and SNR curves for different algorithms under two-bit quantization. From Figure 3 and Figure 4As can be seen, the method proposed in this invention, under one-bit quantization, improves the signal-to-noise ratio (SNR) by 2.5 dB compared to existing least squares, gradient descent, and denoised convolutional neural network channel estimation algorithms when NMSE = 0.06, which is far superior to the least squares and denoised convolutional neural network channel estimation algorithms. Under two-bit quantization, the method proposed in this invention improves the SNR by 5 dB compared to existing gradient descent channel estimation algorithms when NMSE = 0.02, which is also far superior to the least squares and denoised convolutional neural network channel estimation algorithms.

[0052] In summary, addressing the issues of high computational complexity and difficulty in eliminating the impact of low-bit quantization on system performance in existing low-bit quantization large-scale MIMO scattering channel estimation algorithms, this invention proposes a deep learning-based method for low-bit quantization large-scale MIMO scattering channel estimation. Deep learning is introduced into this method, using the least-squares channel estimation result as the input to the neural network to improve its convergence rate. The network structure designed in this invention can effectively handle two significantly different quantization scenarios (one-bit and two-bit) without structural adjustments. Simultaneously, the use of multi-layer convolutional networks effectively learns channel characteristics, extracting key information related to the true channel state from highly distorted low-bit quantization data, ultimately achieving high-precision channel estimation. This invention employs offline training and online testing, reducing the computational complexity of the algorithm. Simulation results demonstrate that this invention has significant advantages over existing algorithms.

Claims

1. A method for estimating low-bit quantization large-scale MIMO scattering channels based on deep learning, characterized in that, Includes the following steps: Step 1: The transmitting end sends a pilot signal x, which reaches the signal receiving end through the scattering channel H. The scattering channel H is randomly generated using the tropospheric scattering communication 7-path channel model, and the noise n in the communication process is additive complex white Gaussian noise. Step 2: At the receiving end, perform low-bit quantization on the received signal y, including one-bit quantization and two-bit quantization. Step 3: Perform LS channel estimation on the received signals after one-bit quantization and two-bit quantization, and separate the real and imaginary parts of the channel estimation results to create training and test datasets for one-bit quantization and two-bit quantization. Step 4: Construct a deep learning-based low-bit quantization large-scale MIMO scattering channel estimation neural network model, which includes the first to third convolutional layers, the first to third residual structures, and the fourth to sixth convolutional layers. The first to third convolutional layers are used to extract data features from the input data. The extracted data features are then input into the first to third residual structures to further extract different data features. Finally, the fourth to sixth convolutional layers are used for further feature extraction to match the output dimension of the data. The kernel size of each convolutional layer is 3×3. Step 5: Input the training dataset and test dataset obtained in Step 3 into the neural network model constructed in Step 4 for training, and update the network weights and offsets using backpropagation gradient descent. Step 6: Place the trained neural network model at the receiver of the low-bit massive MIMO scattering communication system. After the pilot received signal is quantized with low bits, channel estimation is performed. The channel estimation result is then separated into virtual and real signals and input into the neural network model for channel estimation. Finally, the channel estimation result is obtained.

2. The method for estimating low-bit quantization large-scale MIMO scattering channels based on deep learning according to claim 1, characterized in that, The low-bit quantized signal y_Q in step 2 is: y_Q=Q(y)=Q(Hx+n) In the formula, Q(·) represents the low-bit quantization operator that operates on each element in the vector, including one-bit quantization Q1(·) and two-bit quantization Q2(·); One bit is quantized as: Q1(x)=sgn(Re{x})+jsgn(Im{x}) In the formula, Re{x} and Im{x} represent the real and imaginary parts of x, respectively, and sgn(·) represents the sign function, defined as: Two-bit quantization is:

Images