Non-ground communication channel estimation method based on convolutional neural network

Abstract

The invention relates to the technical field of communication, in particular to a non-ground communication channel estimation method based on a convolutional neural network, which comprises the following steps that: a terminal receives an OFDM (Orthogonal Frequency Division Multiplexing) time domain signal transmitted by a non-ground base station and converts the OFDM time domain signal into a frequency domain signal; the terminal physical layer calculates the mapping position of the DMRS in the time-frequency domain according to the DMRS related signaling indicated by the high layer, and extracts a DMRS demodulation reference signal; the method comprises the following steps: performing LS estimation on a channel state at a position where a DMRS demodulation reference signal is located by using a frequency domain LS estimation algorithm to obtain a channel state at a DMRS, and performing rough estimation on a channel response at a non-pilot frequency position by using a linear interpolation algorithm to obtain a preliminary channel response matrix in a complex form; preprocessing the preliminary channel response matrix to convert the preliminary channel response matrix into a real number matrix; and inputting the pre-processed preliminary channel response matrix into the trained channel estimation model based on the residual convolutional neural network for processing to obtain a high-precision target channel response matrix, so that the channel estimation accuracy is remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of communication technologies, and particularly to a non-terrestrial communication channel estimation method based on a convolutional neural network. Background Art

[0002] In today's wireless communication field, non-terrestrial communication systems (such as low-earth orbit satellite communication systems, etc.) are developing rapidly, and they play a key role in achieving global seamless coverage, ensuring emergency communication, etc. However, non-terrestrial communication faces a series of severe challenges, bringing great difficulties to channel estimation work.

[0003] On the one hand, non-terrestrial communication network devices such as low-earth orbit satellites are in a high-speed moving state, which causes extremely high Doppler frequency shifts. This frequency shift causes a deviation between the received signal frequency and the transmitted signal frequency, resulting in serious signal distortion, greatly affecting the demodulation performance. Moreover, the signal transmission distance is far, and it will suffer a large attenuation when passing through the atmosphere, making it difficult for traditional channel estimation methods to accurately track the rapid changes of the channel, and thus having an adverse impact on subsequent signal detection and demodulation.

[0004] On the other hand, in the 5G system, in order to ensure that the receiving end can normally demodulate data, pilot signals (DMRS, Demodulation Reference Signal) are usually inserted into the transmitted data (such as the Physical Downlink Shared Channel PDSCH). In theory, increasing the density of DMRS can improve the accuracy of channel estimation and enhance the demodulation ability of the system, but this will occupy a large amount of time-frequency domain resources and reduce the efficiency of the communication system.

[0005] Traditional channel estimation methods, such as Least Squares (LS) estimation and Linear Minimum Mean Square Error (LMMSE) estimation, perform poorly in such a complex environment. LS estimation is sensitive to noise and has low estimation accuracy; although LMMSE estimation has better performance in theory, it has high computational complexity, poor real-time performance, and relies on accurate channel statistical information, which is difficult to obtain in non-terrestrial communication.

[0006] In recent years, deep learning methods have been studied and applied in the field of channel estimation. Due to its powerful feature extraction ability, the convolutional neural network (CNN) has been used for channel estimation tasks. By analogy with the OFDM (orthogonal frequency division multiplexing) channel as a two-dimensional image, the application of CNN in the OFDM communication field has also been explored. However, the existing deep learning methods based on CNN have obvious defects: First, in order to pursue higher model performance, the method of increasing the convolution kernel size or model depth is usually adopted. Although this can improve the accuracy to a certain extent, it will lead to a significant increase in the system's computational complexity, which is unacceptable for communication systems that are extremely sensitive to computational latency. Second, as the number of network layers continues to deepen, the gradient will gradually decay. During the backpropagation process, after the error value is multiplied by the weight, the weight continuously decreases, and the network is prone to over-saturation phenomena and is difficult to be effectively trained. Continuing to increase the number of layers will even lead to a decrease in accuracy. Third, the signals transmitted in communication are complex signals and contain negative numbers, while ordinary convolutional neural networks are usually designed to process real-valued data and cannot directly extract features from complex signals. If the convolution is modified to achieve direct feature extraction of complex signals, a large amount of additional computational complexity will be introduced. At the same time, the activation function of ordinary convolutional networks generally performs a zeroing operation on negative values, resulting in the physical meaning carried by negative numbers in communication signals not being able to propagate in the network.

[0007] In summary, there are many deficiencies in the existing technologies for non-terrestrial communication channel estimation, and a new method is urgently needed to solve these problems. Summary of the Invention

[0008] The present invention aims to overcome the defects of the existing channel estimation technologies in the non-terrestrial communication scenario and provide a high-performance channel estimation algorithm. Specifically, the present invention is committed to solving the following key problems: First, effectively overcome the problem of low accuracy of channel estimation results caused by the Doppler effect; Second, balance the algorithm complexity to ensure that the proposed algorithm can be efficiently deployed in the communication system; Third, make full use of the learning and fitting capabilities of the convolutional neural network, deeply explore the potential of the estimation algorithm, and significantly improve the accuracy of channel estimation.

[0009] To achieve the purpose of the present invention, the following technical solutions are adopted:

[0010] A non-terrestrial communication channel estimation method based on a convolutional neural network, comprising the following steps:

[0011] The terminal receives the OFDM time-domain signal transmitted from the non-terrestrial base station and performs a discrete Fourier transform to obtain the frequency-domain signal;

[0012] The terminal physical layer calculates the mapping position of the DMRS in the time-frequency domain according to the DMRS-related signaling indicated by the upper layer, and extracts the DMRS demodulation reference signal from the received frequency-domain signal;

[0013] Use the frequency-domain LS estimation algorithm to perform LS estimation on the channel state at the position of the DMRS demodulation reference signal, obtain the channel state at the DMRS, and use the linear interpolation algorithm to roughly estimate the channel response at non-pilot positions to obtain a preliminary channel response matrix in complex form;

[0014] Preprocess the preliminary channel response matrix to convert it into a real matrix;

[0015] Input the preprocessed preliminary channel response matrix into the trained channel estimation model based on the residual convolutional neural network for processing to obtain a target channel response matrix with high accuracy.

[0016] A further improvement is that all terminals extract the DMRS demodulation reference signal from the received frequency-domain signal according to the relevant provisions of the 5G standard protocol TS 38.211.

[0017] A further improvement is that the method of using the frequency-domain LS estimation algorithm to perform LS estimation on the channel state at the position of the DMRS demodulation reference signal to obtain the channel state at the DMRS includes:

[0018] Assume that the received frequency-domain signal obtained by the terminal through OFDM demodulation is Y. According to the DMRS position index, extract the pilot signal Y in the received frequency-domain signal pilot , the DMRS symbol at the transmitting end is denoted as X pilot , use LS estimation to simply calculate the channel response H at the pilot ls，P , and the expression is:

[0019]

[0020] Map the estimation result to a channel response matrix according to the DMRS position index. The size of the channel response matrix is [RB*12,14], where RB is the minimum scheduling unit of 5G NR. It is one OFDM symbol in the time domain and 12 subcarrier widths in the frequency domain, and 14 represents the number of OFDM symbols fixed in a time slot specified by 5G NR.

[0021] A further improvement is that the mathematical expression of the received frequency-domain signal Y is:

[0022] Y = X * H + N

[0023] where X is the transmitted frequency-domain signal, H represents the channel, and N represents the noise.

[0024] A further improvement lies in that the specific method for preprocessing the preliminary channel response matrix to convert it into a real number matrix includes: separating each preliminary channel response matrix into a real part matrix and an imaginary part matrix, and reorganizing them in the third dimension into a real number matrix.

[0025] A further improvement lies in that the size of the real number matrix is 12*RB×14×2.

[0026] A further improvement lies in that the specific method for inputting the preprocessed preliminary channel response matrix into the trained channel estimation model based on the residual convolutional neural network for processing to obtain a highly accurate target channel response matrix includes:

[0027] Inputting the preprocessed preliminary channel response matrix into the trained channel estimation model based on the residual convolutional neural network for processing, predicting the noise components of the real part and the imaginary part respectively and outputting them;

[0028] Subtracting the noise components output by the channel estimation model from the input preliminary channel response matrix to obtain a three-dimensional estimated channel response matrix;

[0029] Reorganizing along the third dimension of the estimated channel response matrix according to the real part and the imaginary part, and finally obtaining a two-dimensional target channel response matrix.

[0030] A further improvement lies in that the channel estimation model based on the residual convolutional neural network successively includes:

[0031] Feature extraction layer: used to extract features from the sampled preprocessed preliminary channel response matrix input, and obtain the output of high-dimensional channels;

[0032] Activation function layer: using the parametric rectified linear unit PReLU as the activation function of the model, which is used to ensure the effective backpropagation of negative numbers in the network, and at the same time introduce a learnable parameter to adaptively adjust the slope of the activation function in the negative value region and improve the fitting ability of the model;

[0033] Intermediate hidden layer: alternately stacked by a variable number of 3*3 convolutional layers and preluLayer, where the convolutional layer is responsible for extracting the features of the input data, and convolutional kernels of different sizes can capture feature information of different scales; the PReLU activation function introduces non-linearity into the model, enabling the model to learn more complex patterns and rules;

[0034] Channel attention module: used to enable the model to automatically learn the importance of each channel, calculate the channel weights of each channel, multiply the obtained channel weights with the feature map output by the intermediate hidden layer by channel, enhance the feature response of important channels, and suppress the feature response of unimportant channels;

[0035] Output convolutional layer: The number of output channels of the output convolutional layer is 2, which is consistent with the third dimension of the preprocessed preliminary channel response matrix of the input, and is used to predict the real and imaginary noise components respectively.

[0036] A further improvement lies in that the number of channels in the intermediate hidden layer is set in a gradually decreasing pattern, gradually decreasing to be close to the dimension of the input data.

[0037] A further improvement lies in that the training method of the channel estimation model based on the residual convolutional neural network is: using the Adam optimizer to train the channel estimation model, using the mean square error function as the cost function of the channel estimation model, and performing offline training with a large number of sample data, continuously updating the model parameters to obtain a trained channel estimation model based on the residual convolutional neural network.

[0038] The beneficial effects of the present invention are as follows:

[0039] The channel estimation model ResNet_CE based on the residual convolutional neural network proposed by the present invention performs excellently in terms of computational complexity control. Its computational complexity is of the same order of magnitude as that of LMMSE, effectively avoiding the problem of a sharp increase in computational complexity caused by increasing the model complexity in traditional deep learning methods. At the same time, a significant improvement has been achieved in the estimation accuracy. Compared with the frequency-domain LS interpolation and the 5G Practical estimation algorithm provided by MATLAB official, the estimation accuracy of the model has a gain of 1 - 4 dB. This means that in the same communication environment, the algorithm of the present invention can more accurately estimate the channel state, providing a strong guarantee for subsequent signal processing and communication quality improvement.

[0040] By adopting a relatively small number of convolutional layers (not exceeding 10 layers), the present invention can effectively learn the noise distribution law of the communication channel. In the processing of model input, it innovatively retains the feature correlation between the real and imaginary parts of the input matrix, and at the same time solves the problem that negative elements in the matrix cannot be effectively transmitted in the ordinary convolutional neural network, enabling meaningful elements to fully participate in the iterative process of the algorithm. These optimized designs not only improve the performance of the model, but also enhance the stability and reliability of the model, making it have significant technical advantages in the field of non-terrestrial communication channel estimation.

[0041] The channel estimation method of the present invention has wide applicability and can be applied to various pilot structures, including the comb-shaped and block pilots specified by the 5G NR standard. At the same time, for pilot scheme systems with different densities, this method can also operate stably, showing good adaptability and scalability. This enables the algorithm of the present invention to play its advantages in different communication scenarios and system configurations, and has high practical value. Description of the Drawings

[0042] Figure 1Flowchart of a non-terrestrial communication channel estimation method based on convolutional neural network according to the present invention;

[0043] Figure 2 Algorithm model framework diagram of the present invention;

[0044] Figure 3 Schematic diagram of frequency-domain LS estimation and linear interpolation process;

[0045] Figure 4 Pseudo-code schematic diagram of generation and preprocessing of model training data;

[0046] Figure 5 Performance comparison diagram of channel estimation algorithms under different SNR conditions;

[0047] Figure 6 Comparison diagram with channel estimation algorithms proposed by other researchers;

[0048] Figure 7 Performance comparison diagram of channel estimation algorithms under different Doppler frequency shift conditions. Detailed implementation manners

[0049] In order to enable those skilled in the art to better understand the solution of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0050] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and do not necessarily need to be used to describe a specific order or sequence. It should be understood that such data can be interchanged under appropriate circumstances so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "comprising" and "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device comprising a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0051] Please refer to the attached Figure 1 - attached Figure 7 In the embodiments of the present invention, a non-terrestrial communication channel estimation method based on convolutional neural network is proposed. As Figure 1 shown, it includes the following steps:

[0052] Step S1: The terminal receives the OFDM time-domain signal transmitted from a non-terrestrial base station (such as a satellite or a high-altitude platform) and performs a discrete Fourier transform (DFT) to obtain a frequency-domain signal, providing the basic data format for subsequent channel estimation operations. This step is the starting point of the entire channel estimation process.

[0053] It should be understood that in a 5G communication system, a time slot contains 14 OFDM symbols. The resources in the frequency domain consist of several RBs (Resource Blocks). The number of RBs depends on the subcarrier spacing. At this time, it can be regarded as a two-dimensional time-frequency resource block. Since data modulation and demodulation in a wireless communication system are usually performed in the frequency domain, converting the signal to the frequency domain helps to more accurately analyze and process the received information.

[0054] Step S2: The physical layer of the terminal calculates the mapping position of the DMRS in the time-frequency domain according to the DMRS-related signaling indicated by the higher layer, and extracts the DMRS demodulation reference signal from the received frequency-domain signal.

[0055] It should be understood that the DMRS, as a known reference signal, is used to evaluate the channel state. Accurately extracting the DMRS is crucial for subsequent channel state estimation because it provides key information about the channel characteristics. The position of the DMRS is specified by higher-layer signaling and protocols and is related to the configuration parameters of the PDSCH. These parameters include carrier configuration, time slot format, DMRS port configuration, number of DMRSs, etc. By analyzing these configuration information, the specific position of the DMRS in the time-frequency resources can be determined.

[0056] Step S3: Use the frequency-domain LS (Least Squares) estimation algorithm to perform LS estimation on the channel state at the position where the DMRS demodulation reference signal is located, obtaining the channel state at the DMRS. The linear interpolation algorithm is used to roughly estimate the channel response at non-pilot positions on the entire time-frequency resource block, obtaining a complete but less accurate complex-valued preliminary channel response matrix.

[0057] It should be understood that in this step, the channel response information is initially obtained through frequency-domain LS estimation and linear interpolation, providing the basic data for subsequent more accurate estimation. Specifically, LS estimation provides a simple and effective method to obtain the channel state at the DMRS position. Although linear interpolation can quickly give a rough channel estimate across the entire frequency band, laying the foundation for further high-precision estimation.

[0058] Step S4: Preprocess the preliminary channel response matrix to convert it into a real matrix.

[0059] It should be understood that since the traditional convolutional neural network design is mainly aimed at real - number data, with its convolutional kernel parameters being real numbers and unable to directly and accurately process complex - number data information, the complex - number data is converted into a real - number format. This conversion not only preserves all the information of the original data but also enables the model to effectively process this data.

[0060] Step S5: Input the pre - processed preliminary channel response matrix into the trained channel estimation model based on the residual convolutional neural network for processing to obtain a target channel response matrix with high accuracy.

[0061] It can be understood that using a deep - learning model, especially a residual convolutional neural network, can capture complex patterns and features in the channel response, thus providing a channel estimation result with higher accuracy than traditional methods. This method is particularly suitable for non - terrestrial communication systems with complex propagation environments.

[0062] Specifically, in step S2, 3GPP proposed in technical report TR 38.811 that in view of the current non - terrestrial network deployment scenario, there is no need to adjust the content related to the demodulation reference signal (DMRS) in the 5G new radio (NR) specification. Therefore, the terminal extracts the DMRS demodulation reference signal from the received frequency - domain signal according to the relevant provisions of the 5G standard protocol TS 38.211.

[0063] In a preferred solution of this embodiment, as Figure 3 shown, in step S3, the method of using the frequency - domain LS estimation algorithm to perform LS estimation on the channel state at the position of the DMRS demodulation reference signal to obtain the channel state at the DMRS includes:

[0064] Assume that the received frequency - domain signal obtained by the terminal through OFDM demodulation is Y, and the mathematical expression of the received frequency - domain signal Y is:

[0065] Y = X * H+N

[0066] where X is the transmitted frequency - domain signal, H represents the channel response (i.e., the influence of the channel in the frequency domain), and N represents noise.

[0067] According to the DMRS position index, extract the pilot signal Y in the received frequency - domain signal pilot , the DMRS symbol at the transmitting end is denoted as X pilot , use LS estimation to simply calculate the channel response H at the pilot ls，P , and the expression is:

[0068]

[0069] Map the estimation result to a channel response matrix according to the DMRS position index. The size of the channel response matrix is [RB * 12, 14]. Here, the dimension [RB * 12, 14] corresponds to the layout of a resource block (RB) of the 5G NR system in the time-frequency resource grid. RB is the smallest scheduling unit of 5G NR. It is one OFDM symbol in the time domain and 12 subcarrier widths in the frequency domain. 14 represents the number of OFDM symbols contained in one time slot specified by 5G NR.

[0070] It should be understood that this mapping process allows us to organize the initially estimated channel state information into a structured form, facilitating subsequent operations such as interpolation, and thus completing the channel response estimation for the entire time-frequency resource block. This process not only helps improve the accuracy of channel estimation but also provides a suitable data format for the deep learning model.

[0071] After subsequent rough estimation of the channel response at non-pilot positions on the entire time-frequency resource block using the linear interpolation algorithm, a preliminary channel response matrix H in complex form that is complete but has low accuracy is obtained. ls 。

[0072] In a preferred solution of this embodiment, in step S4, the specific method for preprocessing the preliminary channel response matrix H ls to convert it into a real number matrix includes: separating each preliminary channel response matrix into two matrices, the real part and the imaginary part, and reorganizing them in the third dimension into a real number matrix. The size of the real number matrix is 12 * RB × 14 × 2.

[0073] It can be understood that since the preliminary channel response matrix is a complex value matrix representing the amplitude and phase changes of the signal after passing through the channel, it is necessary to decompose it into two independent real number matrices to represent the real part and the imaginary part of these complex values respectively. Specifically, if the size of the original channel response matrix is [RB * 12, 14], then after separation, we will obtain two real number matrices of the same size (one representing the real part and the other representing the imaginary part). Then, these two matrices will be reorganized in the third dimension to form a new matrix with a size of [12 * RB, 14, 2]. Here: 12 * RB represents the number of subcarriers in the frequency domain (each resource block RB contains 12 subcarriers); 14 represents the number of OFDM symbols in the time domain; 2 represents the third dimension of the new matrix, used to distinguish the real part and the imaginary part of the original complex matrix.

[0074] Although a complex matrix is decomposed into two parts, namely the real part and the imaginary part, all the information of the original complex data, including both amplitude and phase information, is still retained through the way of reorganizing in the third dimension. This method simplifies the subsequent data processing process and allows the use of standard deep learning frameworks and technologies to train and apply models without special adjustments for complex operations. Generally speaking, this processing method not only cleverly retains the correlation between the real part and the imaginary part in the complex number, but also effectively avoids the problem of excessive computational complexity caused by directly inputting complex numbers, enabling the processed data to better meet the input requirements of the subsequent channel estimation model based on the residual convolutional neural network.

[0075] In a preferred solution of this embodiment, in step S5, the specific method of inputting the preprocessed preliminary channel response matrix into the trained channel estimation model based on the residual convolutional neural network to obtain a highly accurate target channel response matrix includes:

[0076] Step S51: Input the preprocessed preliminary channel response matrix into the trained channel estimation model ResNet_CE based on the residual convolutional neural network for processing, and respectively predict the noise components of the real part and the imaginary part and output them.

[0077] It can be understood that the design purpose of the channel estimation model is to identify and predict the noise components in the input signal. Since the input signal is organized in a form containing real part and imaginary part information, the channel estimation model can process these two parts respectively and output the corresponding noise component prediction results.

[0078] Step S52: Subtract the noise components output by the channel estimation model from the input preliminary channel response matrix H ls to obtain a three-dimensional estimated channel response matrix

[0079] It can be understood that similar to the concept of the residual connection mode of the conventional convolutional neural network but with the opposite operation, the present invention uses the noise components output by the model to subtract these noise components from the original input preliminary channel response matrix H ls to remove the interference factors therein. Through the above operations, a new three-dimensional matrix is obtained, and this matrix is regarded as the estimated channel response matrix after noise reduction processing which retains the structural information of the original signal but reduces the influence of noise.

[0080] Step S53: Reorganize along the third dimension of the estimated channel response matrix by real part and imaginary part, and finally obtain a two-dimensional target channel response matrix

[0081] It can be understood that for the three-dimensional estimated channel response matrix obtained in step S52 is reorganized according to its third dimension (i.e., real and imaginary parts information), that is, the originally separated real and imaginary parts information is recombined to form a complete complex representation form. In this way, a two-dimensional target channel response matrix with the size adjusted back to the original size is finally obtained Target channel response matrix Integrates the denoised channel state information and provides a more accurate and reliable channel response description than the preliminary channel response matrix H ls

[0082] The present invention regards H ls as a three-dimensional picture containing noise, and estimates and removes noise through a deep learning model, which can significantly improve the accuracy of channel estimation, especially in complex environments or non-terrestrial communication scenarios. By converting the complex channel response matrix into a real form for processing and then reorganizing it back, it not only simplifies the design and training of the deep learning model, but also ensures the integrity and usability of the final output. It is particularly suitable for dynamically changing channel conditions, can quickly adapt to different communication environments, and provides stable and efficient channel estimation performance.

[0083] Specifically, in this embodiment, as Figure 2 shown, the channel estimation model based on the residual convolutional neural network sequentially includes:

[0084] Feature extraction layer: a 9*9 convolutional layer with a large receptive field, used to extract features from the sampled preliminary channel response matrix after preprocessing the input, and obtain the output of high-dimensional channels, so as to be able to capture the global feature information in the input data.

[0085] Activation function layer: Adopts the parametric rectified linear unit PReLU (Parametric Rectified Linear Unit) as the activation function of the model, which is used to ensure the effective backpropagation of negative numbers in the network, and at the same time introduces a learnable parameter to adaptively adjust the slope of the activation function in the negative value region, improving the fitting ability of the model and reducing the risk of overfitting.

[0086] It can be understood that by using the parametric rectified linear unit (PReLU) as the activation function, different from the traditional ReLU, PReLU allows non-zero slopes in the negative number region, which helps to solve the "dead" ReLU problem.

[0087] ​Intermediate hidden layer: It is composed of an adjustable number of alternating stacks of 3*3 convolutional layers and preluLayer. This design enables the model to effectively extract multi-scale features from the input data. Among them, the convolutional layer is responsible for extracting the features of the input data. Convolution kernels of different sizes can capture feature information at different scales, which is particularly important for understanding complex channel states. The PReLU activation function introduces non-linearity into the model, enabling the model to learn more complex patterns and rules. The number of channels in the intermediate hidden layer is set in a gradually decreasing pattern, gradually decreasing to be close to the dimension of the input data. This design helps to retain key feature information while reducing the model complexity.

[0088] Channel attention module: It is used to enable the model to automatically learn the importance of each channel. By calculating the channel weights of each channel, the obtained channel weights are multiplied with the feature map output by the intermediate hidden layer channel by channel, enhancing the feature response of important channels and suppressing the feature response of unimportant channels. Further improving the model's ability to capture and utilize key information, it helps the model focus on the most important features and improve the accuracy and efficiency of prediction.

[0089] Output convolutional layer: The number of output channels of the output convolutional layer is 2, which is consistent with the third dimension (real part and imaginary part) of the preprocessed initial channel response matrix of the input, and is used to predict the noise components of the real part and the imaginary part respectively.

[0090] This channel estimation model based on the residual convolutional neural network aims to improve the model's expressiveness and robustness in each link from feature extraction, non-linear transformation, multi-scale feature learning to adaptive channel weighting, so as to provide more accurate channel estimation results in complex environments. In addition, by using the PReLU activation function and the channel attention mechanism, the learning ability and generalization performance of the model are further enhanced.

[0091] In this embodiment, the training method of the channel estimation model based on the residual convolutional neural network is as follows: The Adam optimizer is used to train the channel estimation model. The mean square error (MSE) function is used as the cost function of the channel estimation model. A large amount of sample data is used for offline training, and the model parameters (weights W, biases b) are continuously updated to obtain a trained channel estimation model based on the residual convolutional neural network.

[0092] It can be understood that the Adam optimizer combines the advantages of AdaGrad and RMSProp, can adaptively adjust the learning rate of each parameter, speeds up the convergence rate of the model during training, and improves the training efficiency. MSE calculates the average of the squares of the differences between the predicted values and the actual values, and is suitable for regression problems, especially when it is assumed that the errors follow a normal distribution. In this scenario, MSE effectively quantifies the channel estimation vector output by the model The difference from the true channel vector H can make the predicted value of the model as close as possible to the true value by minimizing the MSE:

[0093]

[0094] During the training process, the loss value under the current parameters is calculated through forward propagation, and the gradient is calculated based on the loss value using the backpropagation algorithm. Then, the Adam optimizer is used to adjust the model parameters (weights W and biases b) according to the calculated gradient to minimize the cost function. This process needs to be repeated multiple times until the model performance reaches a satisfactory level or no longer improves significantly. Finally, a channel estimation model based on a residual convolutional neural network that is fully trained and can accurately estimate the channel state is obtained.

[0095] The generation and preprocessing pseudocode of the model training data are as Figure 4 shown.

[0096] Experimental comparison:

[0097] Proposed ResNet is the basic algorithm for non-terrestrial communication channel estimation. Proposed ResNetWith SE is the final model algorithm after adding the channel attention mechanism to the basic algorithm; CNN is a common convolutional neural network with 8 convolutional layers; the frequency-domain LS interpolation algorithm is the frequency-domain LS interpolation algorithm provided by MATLAB; Practice is an algorithm for denoising using noise information based on the frequency-domain LS interpolation algorithm in MATLAB.

[0098] (1) Performance comparison of channel estimation algorithms under different SNR conditions, as Figure 5 shown.

[0099] Test conditions and data description: According to the proposed simulation scenario, link modeling is performed, and the demodulation reference signal DMRS is extracted at the receiving end as the input for different estimation algorithms. When simulating, the SNR value range is [0, 30] dB, and the step interval is 5 dB. The Doppler frequency shift value is fixed at 600 Hz. MSE is the mean square error between the output result of the estimation algorithm and the true channel matrix. The smaller the value, the closer the estimated matrix is to the true matrix, that is, the smaller the error. The vertical axis of the data graph uses a logarithmic scale, and the values of other parameters are as shown in Table 1 below:

[0100] Parameter Set Value Height of Non-Ground Base Station 600KM SNR [0,30] Doppler Shift 600Hz Subcarrier Spacing 30KHz Cyclic Prefix Type Normal Number of Resource Blocks (RB) 51 Physical Layer Cell Identity 1 Modulation Order 16QAM Code Rate 490 / 1024 DMRS Frequency Domain Mapping Type Type A DMRS Time Domain Mapping Type Type 1 DMRS Length 1 Transmission Mode SISO

[0101] Table 1

[0102] Result analysis: A total of 10 test samples are generated under each SNR condition, and finally the average of the estimated results (MSE) is taken.

[0103] 1. Low signal-to-noise ratio (SNR = 0 dB): When the signal and noise intensities are equal, the signal is almost submerged by the noise. Except for the LMMSE estimation algorithm that has reliable prior information of the channel matrix and noise statistical characteristics and can achieve good estimation results at SNR = 0 dB, other algorithms cannot make accurate estimations.

[0104] 2. Medium signal-to-noise ratio (5 - 15 dB): The signal is relatively clear but the noise is still obvious. The proposed algorithms ResNet and ResNet with SE are both superior to the ordinary CNN estimation algorithm and the practical channel estimation algorithm proposed by MATLAB, and are far better than the frequency-domain LS estimation interpolation method.

[0105] 3. High signal-to-noise ratio (15 - 30 dB): The signal intensity is significantly better than the noise, and the communication quality is good at this time. The recommended target SNR values for user terminals given by actual systems such as Starlink are also in this range. At this time, due to the small influence of noise, the denoising-based convolutional neural network estimation algorithms, namely the proposed ResNet and ResNet with SE, achieve the best results in the comparison of all algorithms.

[0106] (2) Performance comparison of channel estimation algorithms under different Doppler frequency shift conditions, as Figure 7 shown.

[0107] Test conditions and data description: According to the proposed simulation scenario, link modeling is carried out, and the demodulation reference signal DMRS is extracted at the receiving end as the input of different estimation algorithms. During the simulation, SNR is fixed, with a value of SNR = 15 dB. The Doppler frequency shift value is fixed at 200 Hz to 800 Hz, with a step size of 100 Hz. MSE is the mean square error between the output result of the estimation algorithm and the true channel matrix. The smaller the value, the closer the estimated matrix is to the true matrix, that is, the smaller the error. The vertical axis of the data graph uses a logarithmic scale.

[0108] Result analysis:

[0109] A total of 10 test samples are generated under each Doppler frequency shift condition, and finally the average of the estimated results (MSE) is taken.

[0110] 1. When the Doppler frequency shift is low, that is, the Doppler frequency shift value is 200 - 300 Hz, the LMMSE estimation algorithm with reliable prior information of the channel matrix and noise statistical characteristics has good estimation results. Since the high Doppler frequency shift conditions (400 - 1000 Hz) are set during the training of the proposed ResNet and ResNet with SE, the channel estimation error is inferior to that of LMMSE and the practical estimation method of MATLAB at this time. To solve this problem, communication data with Doppler frequency shift can be added during the training of the algorithm model.

[0111] 2. When the Doppler frequency shift value is 300 - 500 Hz, the proposed ResNet with SE outperforms both the ordinary CNN estimation algorithm and the practical channel estimation algorithm proposed in MATLAB, is close to the LMMSE estimation method, and is far better than the frequency-domain LS estimation interpolation method.

[0112] 3. When the Doppler frequency shift value is 500 - 800 HZ, the proposed ResNet with SE shows better estimation performance than all other algorithms, has the smallest MSE error, and is closest to the true channel matrix.

[0113] From this, it can be concluded that the proposed ResNet with SE estimation algorithm achieves excellent estimation results under high-dynamic-range Doppler frequency shift conditions without relying on prior channel information, and is suitable for non-terrestrial communications, especially when a low-orbit satellite is used as the transmitting end (as proposed in the 3GPP technical report, the typical measured Doppler frequency shift in this scenario is 544 Hz). Moreover, since it does not require prior channel information, it is suitable for application in actual non-terrestrial communication scenarios. In actual non-terrestrial communication, due to the high-speed movement of low-orbit satellites and the movement of ground terminals, the channel state information provided by the base station may not be completely reliable, while the ResNet with SE algorithm that does not rely on prior information can better estimate the channel matrix.

[0114] Especially when the ground receiving terminal is a ground gateway station, since the gateway station itself does not move and the orbits of the accessed spaceborne base stations are relatively fixed, the range of Doppler frequency shift is also relatively fixed. Training ResNet with SE for a specific range of Doppler changes can achieve excellent channel estimation results.

[0115] (3) Comparison with the channel estimation algorithms proposed by other researchers, as Figure 6 shown.

[0116] For non-terrestrial communication channel estimation methods and data, they are not open source. Therefore, the open-source work of other researchers is selected for comparison. The extended urban channel (ETU) defined in 3GPP TS 36.101, which is consistent with the present invention, is used as the communication channel model.

[0117] The communication data provided by the present invention is used to train ResNet with SE, and the best model provided by the present invention is used for testing under completely the same conditions.

[0118] The test results show that the proposed ResNet with SE has performance similar to that of the Hybir architecture algorithms proposed by other researchers based on Transformer and hybrid architectures when the SNR is below 15 dB; when the SNR is greater than 15 dB, the performance of the proposed ResNet with SE is better than that of the algorithms proposed by the author. At the same time, the computational complexity of ResNet with SE improved based on CNN is much lower than that of the Hybir architecture based on Transformer and hybrid architectures proposed by other researchers, which requires computational operations in the fully connected layer and the multi-head attention layer, and the required complexity will increase sharply as the length of the input LS estimation sequence increases.

[0119] Therefore, it shows that the channel estimation performance of the proposed ResNet with SE in the present invention not only has high performance in non-terrestrial communications, but can also be applied in terrestrial communication networks, and at the same time has a small complexity and is more easily deployed in latency-sensitive communications.

[0120] The above-described embodiments are only used to illustrate the technical solutions of the present application, and are not intended to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the various embodiments of the present application, and should all be included in the protection scope of the present application.

Claims

1. A non-terrestrial communication channel estimation method based on convolutional neural network, characterized in that: The following steps are involved: The terminal receives the OFDM time domain signal transmitted from the non-ground base station and performs discrete Fourier transform to obtain the frequency domain signal; The terminal physical layer calculates the mapping position of DMRS in the time-frequency domain according to the DMRS-related signaling indicated by the higher layer, and extracts the DMRS demodulation reference signal from the received frequency-domain signal; The frequency domain LS estimation algorithm is used to perform LS estimation on the channel state at the location of the DMRS demodulation reference signal to obtain the channel state at the DMRS. The linear interpolation algorithm is used to roughly estimate the channel response at the non-pilot signal to obtain a preliminary channel response matrix in complex form. Preprocessing the preliminary channel response matrix to convert it into a real number matrix; The preprocessed preliminary channel response matrix is ​​input into the trained channel estimation model based on residual convolutional neural network for processing to obtain a high-precision target channel response matrix.

2. According to claim 1, a non-terrestrial communication channel estimation method based on convolutional neural network is characterized in that: Therefore, the terminal extracts the DMRS demodulation reference signal from the received frequency domain signal in accordance with the relevant provisions of the 5G standard protocol TS 38.

211.

3. The non-terrestrial communication channel estimation method based on convolutional neural network according to claim 1, characterized in that: The method of performing LS estimation on the channel state at the location of the DMRS demodulation reference signal by using the frequency domain LS estimation algorithm to obtain the channel state at the DMRS includes: Assume that the received frequency domain signal obtained by the terminal through OFDM demodulation is Y, and extract the pilot signal Y from the received frequency domain signal according to the DMRS position index pilot , the DMRS symbol of the transmitter is represented by X pilot , use LS estimation to simply calculate the channel response H at the pilot ls , P, the expression is: The estimation result is mapped to a channel response matrix according to the DMRS position index. The channel response matrix size is [RB*12,14], where RB is the minimum scheduling unit of 5G NR. It is an OFDM symbol in the time domain and has a width of 12 subcarriers in the frequency domain. 14 represents the number of OFDM symbols fixed in a time slot specified by 5G NR.

4. The non-terrestrial communication channel estimation method based on convolutional neural network according to claim 3 is characterized in that: The mathematical expression of the received frequency domain signal Y is: Y=X*H+N Among them, X is the transmitted frequency domain signal, H represents the channel, and N represents the noise.

5. The non-terrestrial communication channel estimation method based on convolutional neural network according to claim 1, characterized in that: The specific method of preprocessing the preliminary channel response matrix to convert it into a real number matrix includes: separating each preliminary channel response matrix into two matrices of a real part and an imaginary part, and reorganizing them into a real number matrix in the third dimension.

6. The non-terrestrial communication channel estimation method based on convolutional neural network according to claim 5, characterized in that: The size of the real number matrix is ​​12*RB×14×2.

7. The non-terrestrial communication channel estimation method based on convolutional neural network according to claim 5, characterized in that: The specific method of inputting the preprocessed preliminary channel response matrix into the trained residual convolutional neural network-based channel estimation model for processing to obtain a high-precision target channel response matrix includes: The preprocessed preliminary channel response matrix is ​​input into the trained residual convolutional neural network-based channel estimation model for processing, and the real and imaginary noise components are predicted and output respectively; Subtract the noise component output by the channel estimation model from the input preliminary channel response matrix to obtain a three-dimensional estimated channel response matrix; The estimated channel response matrix is ​​reorganized according to the real and imaginary parts along the third dimension, and finally a two-dimensional target channel response matrix is ​​obtained.

8. The non-terrestrial communication channel estimation method based on convolutional neural network according to claim 5, characterized in that: The channel estimation model based on residual convolutional neural network includes: Feature extraction layer: used to extract features from the input preprocessed preliminary channel response matrix sampling to obtain the output of high-dimensional channels; Activation function layer: The parameterized rectified linear unit PReLU is used as the activation function of the model to ensure the effective back propagation of negative numbers in the network. At the same time, a learnable parameter is introduced to adaptively adjust the slope of the activation function in the negative value area to improve the fitting ability of the model; Middle hidden layer: It is composed of an adjustable number of 3*3 convolutional layers and preluLayer stacked alternately. The convolutional layer is responsible for extracting the features of the input data, and convolution kernels of different sizes can capture feature information of different scales. The PReLU activation function introduces nonlinearity to the model, enabling the model to learn more complex patterns and rules. Channel attention module: used to realize the model to automatically learn the importance of each channel. The channel weight of each channel is obtained by calculation, and the obtained channel weight is multiplied by the feature map output by the middle hidden layer by channel to enhance the feature response of important channels and suppress the feature response of unimportant channels. Output convolution layer: The number of output channels of the output convolution layer is 2, which is consistent with the third dimension of the input preprocessed preliminary channel response matrix, and is used to predict the real and imaginary noise components respectively.

9. The non-terrestrial communication channel estimation method based on convolutional neural network according to claim 8, characterized in that: The number of channels in the middle hidden layer is set to a gradually decreasing mode, gradually decreasing to a dimension close to the input data.

10. The non-terrestrial communication channel estimation method based on convolutional neural network according to claim 1, characterized in that: The training method of the channel estimation model based on the residual convolutional neural network is as follows: the channel estimation model is trained using the Adam optimizer, a mean square error function is used as the cost function of the channel estimation model, a large amount of sample data is used for offline training, and the model parameters are continuously updated to obtain a trained channel estimation model based on the residual convolutional neural network.

Images