CN114884775A - Deep learning-based large-scale MIMO system channel estimation method - Google Patents

Abstract

The invention discloses a deep learning-based large-scale MIMO system channel estimation method, which comprises the following steps: the pilot signal is transmitted through an MIMO communication system, a receiving end receives the communication signal, preliminary channel estimation is carried out by utilizing a least square algorithm, and preliminary estimation channel state information is input into a pre-trained deep neural network to determine channel estimation. The invention utilizes the point convolution layer, the grouping convolution layer and the depth separable convolution layer to construct the depth neural network, the network reduces the number of parameters to be stored and the convolution times to be calculated, thereby reducing the complexity of the deep learning algorithm in channel estimation; the deep neural network has higher precision and keeps good performance when the signal-to-noise ratio is low; the activation function used in the invention improves the accuracy of the deep neural network for estimating the channel.

Description

A channel estimation method for massive MIMO systems based on deep learning

technical field

The invention relates to the field of channel estimation in communication systems, in particular to a channel estimation method for massive MIMO systems based on deep learning.

Background technique

MIMO technology refers to the use of multiple transmit and receive antennas at the transmitter and receiver, respectively, so that signals are transmitted and received through multiple antennas at the transmitter and receiver, thereby improving communication quality. Massive MIMO technology is considered to be one of the key technologies for 5G and future cellular communication systems. Massive multiple-input multiple-output systems utilize massive antennas that are much larger than traditional MIMO systems by deploying massive antennas at base stations in a distributed or centralized manner. degrees of spatial freedom. Without increasing spectral resources and transmit power, such large-scale MIMO systems can exponentially increase channel capacity and significantly reduce inter-user interference by making full use of space resources.

In systems using massive MIMO, accurate uplink and downlink channel state information is crucial for signal detection, beamforming, resource allocation, signal preprocessing, etc. In fact, the channel is unknown to the transmitter and must first be estimated by pilots at the receiver. In time division duplex (TDD) mode, downlink channel state information can be obtained from uplink channel state information by exploiting channel reciprocity between uplink and downlink. However, in the time division duplex mode, channel reciprocity does not always exist, and the corresponding calibration process is difficult and complicated, so it may be inaccurate to obtain downlink channel state information from the uplink channel state.

Traditional sparsity and compressed sensing methods for channel estimation lead to solving a complex optimization problem at each coherence interval, the complexity of which increases with the number of antennas, which will gradually become impossible with the use of massive MIMO. Row. Therefore, the channel estimation method based on compressed sensing cannot adapt to this mechanism in terms of complexity, power consumption or pilot overhead, requires a lot of computing time and resources, and cannot meet the real-time processing requirements in actual deployment.

The application of deep neural networks to various communication and signal processing problems using deep learning algorithms has shown great potential to revolutionize communication systems, with applications in modulation identification, signal detection, CSI feedback and channel estimation, network routing, and service control. When current models based on deep learning algorithms perform the channel estimation function of communication systems, the channel matrix is regarded as a two-dimensional image, and high memory requirements and computational complexity constitute the main obstacles to the practical deployment of neural networks in communication systems.

SUMMARY OF THE INVENTION

Purpose of the invention: In view of the above problems, the purpose of the present invention is to provide a channel estimation method for massive MIMO systems based on deep learning.

Technical solution: A method for channel estimation of massive MIMO system based on deep learning of the present invention includes the following steps: transmitting a pilot signal through a MIMO communication system, receiving the communication signal at the receiving end, and using a least squares algorithm to perform preliminary channel estimation. Estimate, input the preliminary estimated channel state information into the pre-trained deep neural network to determine the channel estimation;

The deep neural network includes a grouped convolutional neural network, a point convolutional neural network, and a depthwise separable convolutional neural network. The first, fourth, seventh, and ninth layers use point convolutional neural networks, and the second layer uses grouped convolutional neural networks. The third, fifth, sixth, and eighth layers of the network use a depthwise separable convolutional neural network.

Further, an activation function tanh is added to the outputs of the first eight layers of the deep neural network.

Further, the training process of the deep neural network includes:

Step 1, input the sample data into the deep neural network, and use the mean square error function between the output estimated channel state information and the real channel matrix as the cost function of training the network;

Step 2, using the ADAM optimization algorithm to update the network parameters of the deep neural network, setting the initial learning rate, and the subsequent optimization algorithm will use the training process to adaptively update the learning rate;

Step 3: Use the sample data to perform offline training on the deep neural network, and deploy the trained network at the receiving end.

Further, the sample data includes the original channel matrix data directly generated by using the channel model, the original channel matrix data is preprocessed to generate relatively low-precision channel matrix data of the same dimension, and the original channel matrix data and the preprocessed channel matrix data are obtained. The data are labeled together as sample data for training deep neural networks.

Further, the generation process of the relatively low-precision channel matrix data is as follows: perform discrete Fourier transform on the original channel matrix data to the sparse domain, sample and interpolate the transformed channel matrix data to restore the original dimension, and then add the channel matrix data to the sparse domain. noise to obtain low-precision channel matrix data.

Further, using the least squares algorithm to perform preliminary channel estimation, and inputting the preliminary estimated channel state information into the pre-trained deep neural network to determine the channel estimation includes: the receiving end extracts the received signal at the pilot position, and calculates the initial channel according to the least squares algorithm. Channel estimation matrix: The real part and imaginary part of the initial channel estimation matrix form a two-dimensional real number channel matrix, the two-dimensional real number channel matrix is input into the deep neural network, and the two-dimensional real number matrix formed by the real part and the imaginary part of the channel is output.

Beneficial effect: Compared with the prior art, the present invention has the following significant advantages:

1. The present invention utilizes a point convolution layer, a grouped convolution layer and a depthwise separable convolution layer to construct a deep neural network, which reduces the number of parameters that need to be stored and the number of convolutions that need to be calculated, thereby reducing the depth of the learning algorithm. complexity in channel estimation;

2. The deep neural network of the present invention has higher precision and maintains good performance at low signal-to-noise ratio; 3. The activation function used in the present invention improves the accuracy of the deep neural network to estimate the channel.

Description of drawings

Fig. 1 is the flow chart of the channel estimation method of the present invention;

Figure 2 shows the normalized mean square error-signal-to-noise ratio curves of the estimated channel and the real channel when different network frameworks and different activation functions are used respectively.

Detailed ways

In order to make the objectives, technical solutions and advantages of the present application more clear, the present application will be further described in detail below with reference to the accompanying drawings and embodiments.

As shown in FIG. 1 , a deep learning-based channel estimation method for a massive MIMO system described in this embodiment includes the following steps: transmitting a pilot signal through a MIMO communication system, receiving the communication signal at the receiving end, and using the minimum The square algorithm performs preliminary channel estimation, and the preliminary estimated channel state information is input into the pre-trained deep neural network to determine the channel estimation.

The deep neural networks include grouped convolutional neural networks, point convolutional neural networks, and depthwise separable convolutional neural networks. The first, fourth, seventh, and ninth layers use point convolutional neural networks, and the second layer uses grouped convolutional neural networks. , the third, fifth, sixth, and eighth layers use a depthwise separable convolutional neural network. In this embodiment, the output dimension of the point convolutional neural network of the first layer is 96, and the output dimensions of the fourth, seventh, and ninth layers of the point convolutional neural network are 16, 8, and 2, respectively. The grouped convolutional neural network of the second layer divides the output features of the upper layer into groups of two, and the size of the convolution kernel is 5×5. The depthwise separable convolutional neural networks of the third, fifth, sixth and eighth layers have filter bank numbers of 48, 16, 16, and 8, respectively, and the convolution kernel size is 3×3. The specific parameters of each layer of the network are shown in Table 1.

The activation function tanh is added to the output of the first eight layers of the deep neural network, and the output layer of the deep neural network does not use the activation function but outputs directly. In the case of large signal-to-noise ratio, other linear activation functions can also be selected to speed up the convergence of the network.

Table 1 Network parameters of each layer of deep neural network

The training process of the deep neural network in this embodiment includes:

Step 1, input the sample data into the deep neural network, and use the mean square error function between the output estimated channel state information and the real channel matrix as the cost function of training the network, and the expression is:

where H _LS is the initial channel state information estimated by the least squares algorithm, H is the training sample data, the function f is the mapping relationship from the input to the output of the deep neural network, T represents the number of training samples, and θ represents the weight of the neural network. value parameter.

Step 2, use ADAM optimization algorithm to update the network parameters of the deep neural network, set the initial learning rate, and the subsequent optimization algorithm will use the training process to adaptively update the learning rate. In this embodiment, the initial learning rate is set to 0.01.

Step 3: Use the sample data to perform offline training on the deep neural network, and deploy the trained network at the receiving end.

The sample data includes the original channel matrix data directly generated by the channel model. The original channel matrix data is preprocessed to generate relatively low-precision channel matrix data of the same dimension, and the original channel matrix data and the preprocessed channel matrix data are marked together. as sample data for training deep neural networks.

The generation process of relatively low-precision channel matrix data is as follows: perform discrete Fourier transform on the original channel matrix data to the sparse domain, sample and interpolate the transformed channel matrix data to restore the original dimension, and then add noise to the channel matrix data, Obtain low-precision channel matrix data.

In this embodiment, the COST2100 model is used to generate the original channel matrix data, and an indoor scene of 5.3 GHz is used. In the indoor scene, the base station is centered on a square area of 20m×20m, and the user equipment moves randomly within the square. The base station adopts a uniform linear array, the number of transmit antennas is N _t =256, the user end has only one receive antenna, the system operates in the orthogonal frequency division multiplexing mode, and the number of subcarriers is K, where K=256.

The original channel matrix data is subjected to discrete Fourier transform, and the channel data obtained after the transformation is divided into 5 parts according to the ratio of 1:1:2:3:3, and is sampled and compressed, and the compression rate is 2 times, 4 times, 8 times, 16 times, 32 times. The compressed channel data is restored to the dimension of the original channel matrix data through interpolation, and the restored data is added with noise and marked with the original channel matrix data to form sample data, which is used for offline training of deep neural networks.

By performing discrete Fourier transform on the channel matrix, an approximate sparse channel matrix in the angle-delay domain is obtained, and the channel _Ha in the angle-delay domain is defined as follows:

H _a =F _a H _s F _b

where Fa and Fb are discrete Fourier transform matrices, and H _S is the MIMO channel in frequency division duplex mode.

Use the least squares algorithm to perform preliminary channel estimation, and input the preliminary estimated channel state information into the pre-trained deep neural network to determine the channel estimation. The real part and imaginary part of the initial channel estimation matrix form a two-dimensional real number channel matrix, the two-dimensional real number channel matrix is input into the deep neural network, and the two-dimensional real number matrix formed by the real part and the imaginary part of the channel is output.

In order to further verify the performance of the method in this embodiment, the performance of the channel estimation method in this embodiment is verified in combination with simulation. Figure 2 shows the accuracy of the channel estimation method proposed in this embodiment when estimating the channel, wherein by using different activation functions in the neural network, the channel estimation method proposed in this embodiment has relatively close performance in the case of a large signal-to-noise ratio , but there is a large gap at low signal-to-noise ratio, which proves the noise suppression ability among the three different methods. The deep neural network framework proposed in the present invention has a strong ability to suppress noise. Also has high accuracy.

Claims (6)

1. A large-scale MIMO system channel estimation method based on deep learning is characterized by comprising the following steps: transmitting a pilot signal through an MIMO communication system, receiving the communication signal by a receiving end, performing preliminary channel estimation by using a least square algorithm, and inputting preliminary estimation channel state information into a pre-trained deep neural network to determine channel estimation;

the deep neural network comprises a grouped convolutional neural network, a point convolutional neural network and a deep separable convolutional neural network, the point convolutional neural network is adopted in the first layer, the fourth layer, the seventh layer and the ninth layer, the grouped convolutional neural network is adopted in the second layer, and the deep separable convolutional neural network is adopted in the third layer, the fifth layer, the sixth layer and the eighth layer.

2. The channel estimation method according to claim 1, wherein the activation function tanh is added to the outputs of the first eight layers in the deep neural network.

3. The channel estimation method of claim 1, wherein the training process of the deep neural network comprises:

step 1, inputting sample data into a deep neural network, and taking a mean square error function between output estimated channel state information and a real channel matrix as a cost function of a training network;

step 2, updating network parameters of the deep neural network by using an ADAM optimization algorithm, setting an initial learning rate, and adaptively updating the learning rate by using a training process in a subsequent optimization algorithm;

and 3, performing off-line training on the deep neural network by using the sample data, and deploying the trained network at a receiving end.

4. The channel estimation method according to claim 3, wherein the sample data includes original channel matrix data directly generated by using a channel model, the original channel matrix data is preprocessed to generate channel matrix data with relatively low precision of the same dimension, and the original channel matrix data and the preprocessed channel matrix data are marked together to be used as sample data for training the deep neural network.

5. The channel estimation method of claim 4, wherein the relatively low-precision channel matrix data is generated by: and performing discrete Fourier transform on the original channel matrix data to a sparse domain, sampling and interpolating the transformed channel matrix data to restore the original dimension, and then adding noise to the channel matrix data to obtain low-precision channel matrix data.

6. The channel estimation method of claim 1, wherein the performing the preliminary channel estimation using a least squares algorithm, and inputting the preliminary channel state information into a pre-trained deep neural network to determine the channel estimation comprises: the receiving end extracts a receiving signal on a pilot frequency position, calculates an initial channel estimation matrix according to a least square algorithm, enables a real part and an imaginary part of the initial channel estimation matrix to form a two-dimensional real number channel matrix, inputs the two-dimensional real number channel matrix into a deep neural network, and outputs the two-dimensional real number matrix formed by the real part and the imaginary part of a channel.

Images