Channel noise estimation method

Abstract

The invention discloses a channel noise estimation method, which comprises the following steps of: performing channel denoising by adopting a denoising network model to obtain denoised channel information, and estimating to obtain channel noise according to the complete channel information and the denoised channel information. Complete channel information containing noise is obtained through the deconvolution network, and the channel noise is learned through the expansion convolution network, so that the noise is suppressed and removed to obtain more accurate channel estimation, and the estimation of the channel noise is realized.

Description

Technical Field

[0001] The invention relates to the technical field of satellite navigation, and in particular to a channel noise estimation method. Background Art

[0002] With the rapid development and widespread application of wireless communication technology, channel noise estimation is particularly important in communication systems. The accuracy of channel noise estimation directly affects the performance and reliability of the system. In modern communication systems, channel noise mainly includes thermal noise, interference noise, and multipath effects. These noises not only reduce the quality of the signal, but also increase the bit error rate at the receiving end. Therefore, the research and development of efficient channel noise estimation methods and devices has important practical significance and broad application prospects. Traditional channel noise estimation methods mainly rely on empirical models and statistical methods, such as maximum likelihood estimation and minimum mean square error estimation. These methods can effectively estimate noise to a certain extent, but it is often difficult to ensure high accuracy and high robustness when channel conditions are complex and changeable. With the rise of machine learning and deep learning technologies, data-driven methods have gradually become a hot topic in channel noise estimation research. By introducing deep learning models, such as convolutional neural networks (CNNs) and recurrent neural networks (RNNs), complex channel features can be automatically extracted from a large amount of communication data, thereby improving the accuracy and adaptability of noise estimation.

[0003] In practical applications, noise characteristics often change over time, so it is necessary to track the noise dynamically. Kalman filter and autoregressive moving average model (ARMA) play an important role in noise time series analysis. Kalman filter can track the changes of noise in real time and dynamically adjust the noise model, while ARMA model uses past noise data for prediction and provides robust noise estimation. In addition, sliding window technology and variational Bayes method are also applied to the real-time update process of spectrum to ensure the real-time and reliability of noise estimation. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a channel noise estimation method, which obtains complete channel information containing noise through a deconvolution network, and learns the channel noise through an expanded convolution network, thereby suppressing the noise to obtain a more accurate channel estimation, thereby realizing the estimation of channel noise.

[0005] The technical solution of the present invention is:

[0006] A channel noise estimation method, which uses a denoising network model to perform channel denoising to obtain denoised channel information, and then estimates the channel noise based on the complete channel information and the denoised channel information;

[0007] The denoising network model includes a deconvolution network and a dilated convolution network. The processing process of the deconvolution network is shown in the following formula (1):

[0008] size out =s(size in -1)+size kernel (1);

[0009] In formula (1), size in For input information, size out For output information, size kernel is the size information of the convolutional layer, s is the step size; the input information is the preliminary pilot channel information H LS , the output information is the complete channel information H containing noise obtained after being processed by the deconvolution network;

[0010] The dilated convolutional network includes twelve network layers. The first network layer is a convolutional layer. 64 convolution kernels of size 1×1 are used to process the channel estimation data obtained in the previous stage, and ReLU is used as the activation function. Then ten network layers are cascaded. Each network layer is a basic module. Each basic module includes a dilated convolutional layer, a BN batch normalization module and a ReLU activation function. The last network layer is a convolutional layer. 64 convolution kernels of size 3×3 are used to process the channel estimation data obtained in the previous stage to obtain the complete channel information after denoising.

[0011] The channel noise is obtained by the following formula (2):

[0012] H=H ′ +Z (2);

[0013] In formula (2), H ′ is the complete channel information, Z is the channel noise, and H is the complete channel information including noise.

[0014] The deconvolution network includes five network layers, the first four network layers are deconvolution layers, and the processing process is shown in the following formula (4):

[0015] H i＝ f(W i *H (i-1) +b i ), i=1,2,3,4 (4);

[0016] In formula (4), when i=1, H (i-1) That is H0, H0 is H LS ; W i is the weight of the i-th deconvolution layer, * is the deconvolution operation, b i is the bias value, f() is the activation function;

[0017] The fifth network layer is the convolution layer. The output data of the fourth deconvolution layer is used as the input data of the convolution layer. Convolution processing is performed through two convolution kernels of size 1×1 to reduce the dimension of feature data so that the output image size is consistent with the image size of the preliminary pilot channel information.

[0018] The preliminary pilot channel information H LS It is obtained based on the least squares estimation, that is, the preliminary pilot channel information is obtained by multiplying the receiver's received signal matrix with the inverse matrix of the satellite's sent matrix pilot.

[0019] In the deconvolution network, the convolution kernel sizes of the deconvolution operations from the first deconvolution layer to the fourth deconvolution layer are all 5×5, the step sizes are all 1, and the numbers of convolution kernels are 128, 32, 8, and 4, respectively.

[0020] The denoising network model uses actual channel test data and utilizes genetic algorithm or particle swarm algorithm to iteratively optimize parameters.

[0021] Advantages of the present invention:

[0022] (1) The deconvolution network of the present invention utilizes the advantage of its interpolation and gradual amplification to achieve a better channel reconstruction process, that is, the deconvolution network does not need to undergo traditional interpolation operations to obtain data of the same size as the label channel, and avoids the problems of too many parameters and complex calculations brought about by the multi-layer convolution network process. The process of gradually amplifying the channel data is achieved only by relying on a lightweight network structure; at the same time, in order to change the output dimension of the feature map and make the output data size consistent with the label data size dimension, a 1×1 convolution kernel is used in the last layer of the deconvolution network. The 1×1 convolution kernel can achieve cross-channel interaction and information integration, and perform dimensionality reduction operations on the number of channels, so that the width and height of the convolution layer output data will not change, but only the number of channels of the output data will change. The convolution operation ensures that the output data size is consistent with the label data size, and at the same time, the nonlinearity of the network can be increased, so that the network can express more complex features.

[0023] (2) The dilated convolutional network of the present invention can learn channel noise, thereby subtracting noise from noisy channel information to obtain more accurate channel estimation and prevent the gradient vanishing problem of the front layer network during training; the dilated convolutional layer of the dilated convolutional network injects holes into the basic convolution kernel to increase the receptive field, which can not only learn the local correlation characteristics of the channel, but also further utilize the correlation characteristics of the wider frequency dimension to obtain more accurate estimation; in the basic module, the BN module as batch normalization can accelerate the network convergence speed and improve the training stability, and the ReLU as an activation function can better mine the signal features and fit the training data. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0025] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0026] See Figure 1 A channel noise estimation method, specifically, using a denoising network model to perform channel denoising, the denoising network model first uses a shallow deconvolution network to extract features and interpolate channels from a low-resolution preliminary pilot channel information image, thereby gradually achieving effective channel estimation while enlarging the channel data size: at the same time, in order to reduce the influence of noise and improve the estimation performance, an expanded convolution network is used to further suppress the channel noise, thereby obtaining denoised channel information, and then estimating the channel noise based on the complete channel information and the denoised channel information;

[0027] The denoising network model includes a deconvolution network and a dilated convolution network;

[0028] S1, the processing process of the deconvolution network is shown in the following formula (1):

[0029] size out =s(size in -1)+size kernel (1);

[0030] In formula (1), size in For input information, size out For output information, size kernel is the size information of the convolutional layer, s is the step size; the input information is the preliminary pilot channel information H LS , preliminary pilot channel information H LS It is obtained based on the least squares method (LS), that is, the preliminary pilot channel information is obtained by multiplying the receiver's received signal matrix with the inverse matrix of the satellite's sent matrix pilot; the output information is the complete channel information H containing noise obtained after being processed by the deconvolution network;

[0031] The deconvolution network consists of five network layers. The first four network layers are deconvolution layers. The processing process is shown in the following formula (4):

[0032] H i＝ f(W i *H (i-1) +bi ), i=1,2,3,4 (4);

[0033] In formula (4), when i=1, H (i-1) That is H0, H0 is H LS ; W i is the weight of the i-th deconvolution layer, * is the deconvolution operation, b i is the bias value, f() is the activation function; the convolution kernel size of the deconvolution operation from the first to the fourth deconvolution layer is 5×5, the step size is 1, and the number of convolution kernels is 128, 32, 8, and 4 respectively;

[0034] The fifth network layer is a convolutional layer. The output data of the fourth deconvolution layer is used as the input data of the convolutional layer. Convolution processing is performed through two convolution kernels of size 1×1 to reduce the dimension of feature data so that the output image size is consistent with the image size of the preliminary pilot channel information.

[0035] S2. The dilated convolutional network includes twelve network layers. The first network layer is a convolutional layer. 64 convolution kernels of size 1×1 are used to process the channel estimation data obtained in the previous stage, and ReLU is used as the activation function. Then ten network layers are cascaded. Each network layer is a basic module. Each basic module includes a dilated convolutional layer, a BN batch normalization module and a ReLU activation function. The last network layer is a convolutional layer. 64 convolution kernels of size 3×3 are used to process the channel estimation data obtained in the previous stage to obtain the complete channel information after denoising.

[0036] The channel noise is obtained by the following formula (2):

[0037] H=H ′ +Z (2);

[0038] In formula (2), H ′ is the complete channel information, Z is the channel noise, and H is the complete channel information with noise;

[0039] S3. Collect actual channel test data, divide the actual channel test data into a test set and a training set, iteratively optimize the parameters of the training set using a genetic algorithm or a particle swarm algorithm to obtain a trained denoising network model, and test the trained denoising network model on the test set to obtain a denoising network model with optimized parameters.

[0040] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A channel noise estimation method, characterized in that: A denoising network model is used to perform channel denoising to obtain denoised channel information, and then the channel noise is estimated based on the complete channel information and the denoised channel information; The denoising network model includes a deconvolution network and a dilated convolution network. The processing process of the deconvolution network is shown in the following formula (1): size out =s(size in -1)+size kernel (1); In formula (1), size in For input information, size out For output information, size kernel is the size information of the convolutional layer, s is the step size; the input information is the preliminary pilot channel information H LS , the output information is the complete channel information H containing noise obtained after being processed by the deconvolution network; The dilated convolutional network includes twelve network layers. The first network layer is a convolutional layer. 64 convolution kernels of size 1×1 are used to process the channel estimation data obtained in the previous stage, and ReLU is used as the activation function. Then ten network layers are cascaded. Each network layer is a basic module. Each basic module includes a dilated convolutional layer, a BN batch normalization module and a ReLU activation function. The last network layer is a convolutional layer. 64 convolution kernels of size 3×3 are used to process the channel estimation data obtained in the previous stage to obtain the complete channel information after denoising. The channel noise is obtained by the following formula (2): H=H ′ +Z (2); In formula (2), H ′ is the complete channel information, Z is the channel noise, and H is the complete channel information including noise.

2. A channel noise estimation method according to claim 1, characterized in that: The deconvolution network includes five network layers, the first four network layers are deconvolution layers, and the processing process is shown in the following formula (4): H i＝ f(W i *H (i-1) +b i ),i=1,2,3,4 (4); In formula (4), when i=1, H (i-1) That is H0, H0 is H LS ; W i is the weight of the i-th deconvolution layer, * is the deconvolution operation, b i is the bias value, f() is the activation function; The fifth network layer is the convolution layer. The output data of the fourth deconvolution layer is used as the input data of the convolution layer. Convolution processing is performed through two convolution kernels of size 1×1 to reduce the dimension of feature data so that the output image size is consistent with the image size of the preliminary pilot channel information.

3. A channel noise estimation method according to claim 1, characterized in that: The preliminary pilot channel information H LS It is obtained based on the least squares estimation, that is, the preliminary pilot channel information is obtained by multiplying the receiver's received signal matrix with the inverse matrix of the satellite's sent matrix pilot.

4. A channel noise estimation method according to claim 2, characterized in that: In the deconvolution network, the convolution kernel sizes of the deconvolution operations from the first deconvolution layer to the fourth deconvolution layer are all 5×5, the step sizes are all 1, and the numbers of convolution kernels are 128, 32, 8, and 4, respectively.

5. A channel noise estimation method according to claim 1, characterized in that: The denoising network model uses actual channel test data and utilizes genetic algorithm or particle swarm algorithm to iteratively optimize parameters.

Images