Buoy satellite communication system channel estimation method and system in marine environment

Abstract

The invention discloses a buoy satellite communication system channel estimation method and system in a marine environment, and the method comprises the steps: carrying out the preliminary channel estimation at a receiving end through a least square algorithm, and then achieving the fine optimization through an ARCENet model based on an attention mechanism and a residual network, the residual structure improves the training stability; and the interpolation layer adopts a generative adversarial network to realize high-precision response reconstruction. The method has the advantages that the calculation load is reduced through the lightweight network design, and the dynamic characteristics of the channel are captured through nonlinear mapping; the sparse pilot frequency configuration is kept to save spectrum resources, and the robustness under the condition of low signal-to-noise ratio is remarkably improved. And finally, high-precision recovery of channel response under a complex sea condition is realized, and the reliability of a buoy satellite communication link is effectively guaranteed.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence and wireless communication technology, in particular to a channel estimation method and system for a buoy satellite communication system in a marine environment. BACKGROUND

[0002] The buoy satellite communication system in a marine environment has important application value in the fields of marine monitoring and weather forecasting. This kind of system realizes remote return of buoy monitoring data through satellite link, and the communication quality directly affects the real-time and accuracy of marine data. However, the marine communication environment has significant particularity: the combined action of seawater reflection, atmospheric turbulence and floating body swing causes complex time-varying fading and multipath interference of the signal in the transmission process. This dynamic channel characteristic makes the received signal inevitably distorted, which seriously restricts the reliability of the communication system. In order to ensure effective data transmission, the receiving end must accurately estimate the channel characteristics to eliminate the influence of transmission distortion on the signal.

[0003] The current mainstream channel estimation technology generally relies on pilot symbols for estimation. These pilots are embedded in the time-frequency grid according to a certain pattern, forming block, comb or grid type structures. The receiver uses traditional algorithms such as least squares to interpolate and recover the channel response according to the known pilot position and value. However, this kind of method has obvious limitations in the marine dynamic environment: when the channel characteristics change rapidly or the noise interference increases significantly, the interpolation estimation based on linear model will have large deviation. Although the sparse configuration of pilots can save spectrum resources, it is difficult to capture the instantaneous change characteristics of the channel, resulting in insufficient estimation accuracy. At the same time, although increasing the pilot density can improve the estimation effect, it will sacrifice the valuable data transmission bandwidth, which does not meet the communication constraints of the buoy terminal with narrow bandwidth.

[0004] In recent years, deep learning technology has provided a new idea for channel estimation. By extracting the nonlinear characteristics of the channel through neural networks, it can theoretically break through the precision limit of traditional methods. However, in practical application, deep models face multiple challenges: as the number of network layers increases, the model structure tends to be complex, which not only leads to the expansion of parameter size, but also causes problems such as training efficiency decline and gradient information attenuation. More importantly, complex models put a huge pressure on the limited computing resources of buoy terminals, making it difficult to meet the real-time requirements of embedded devices. In addition, the existing deep learning scheme still has imperfect ability to capture the time-frequency joint characteristics of the channel, and the high-frequency detail recovery effect is not good.

[0005] The existing pilot structure design also has inherent defects. The standard sparse pilot layout is prone to position deviation when the buoy is violently shaken, which aggravates the estimation error. When recovering the complete channel response, the traditional interpolation method often appears curve oscillation phenomenon, and it is difficult to realize smooth transition. These limitations jointly restrict the reliability of buoy satellite communication in high sea state. Therefore, it is urgent to develop a new channel estimation scheme to realize high-precision and low-complexity channel estimation in dynamic marine environment on the premise of ensuring spectrum efficiency. SUMMARY

[0006] The present application provides a buoy satellite communication system channel estimation method and system in marine environment.

[0007] In order to achieve the above application purposes, the technical solutions adopted by the present application are as follows:

[0008] A buoy satellite communication system channel estimation method in marine environment, comprising the following steps:

[0009] At the receiving end, receiving an orthogonal frequency division multiplexing (OFDM) signal, the signal being transmitted through a non-terrestrial network-tap delay line (NTN-TDL) model channel and being affected by additive white Gaussian noise (AWGN);

[0010] Extracting pilot symbols from the received signal, the pilot symbols being configured as a sparse pilot structure in the time-frequency domain to reduce the occupation of spectrum resources;

[0011] Using a least squares (LS) algorithm to preliminarily estimate the channel of the pilot position, and obtaining a preliminary channel response matrix HLS, wherein HLS is a complex matrix with a dimension of N f ×N s , N f representing the number of subcarriers, and N s representing the number of symbols;

[0012] Pretreating HLS, including separating the real part and the imaginary part, forming a real number array with a dimension of 2×N f ×N s as input;

[0013] Inputting the pretreated data into an ARCENet model based on attention mechanism and residual network, performing fine channel estimation through the model, and obtaining an optimized channel response matrix wherein the ARCENet model is realized through a nonlinear mapping function f θ is a model training parameter;

[0014] Using the optimized channel response matrix to perform channel equalization and restore the original transmitted signal.

[0015] Furthermore, the pilot symbols adopt QPSK modulation and are configured as narrow bandwidth, short burst OFDM frame structures.

[0016] In the pilot structure, each frame contains N ps There are 1 pilot symbol, and N is placed on each pilot symbol. pf One pilot subcarrier, receiving pilot signal Y p With the transmission of pilot signal X p The relationship satisfies Y p =H p ·X p +Z p H p Z represents the channel gain at the pilot position. p This is the noise term.

[0017] Furthermore, the system architecture of the ARCENet model includes a transmitter and a receiver:

[0018] At the transmitting end, the baseband signal is QPSK modulated and pilot information is inserted. It is allocated to the OFDM frame subcarrier position through resource mapping, and then formed into a time domain signal by inverse fast Fourier transform (IFFT) and adding a cyclic prefix.

[0019] At the receiving end, the received signal is converted into frequency domain data through cyclic prefix removal, serial-to-parallel conversion, and Fast Fourier Transform (FFT), and the pilot signal Y is extracted. pilot , used for initial LS estimation.

[0020] Furthermore, the ARCENet model includes an input layer, a convolutional layer, multiple residual modules, and an interpolation layer connected in sequence:

[0021] The input layer receives data in a 2×N dimension. f ×N s An array of real numbers;

[0022] The convolutional layer consists of a first two-dimensional convolutional layer using two 3×3 kernels for initial feature extraction and noise suppression, followed by a Gaussian error linear unit (GeLU) activation function, the expression of which is:

[0023] Furthermore, the residual module is a composite residual structure, comprising four cascaded residual blocks, wherein the second and fourth residual blocks introduce an SE (Squeeze-and-Excitation) attention mechanism to form an attention residual block;

[0024] The attention residual block includes two convolutional layers and an SE module. The SE module compresses the feature map using global average pooling and calculates... Then, a channel attention map H is generated through a fully connected layer and GeLU activation.ex =σ(W2δ(W1H sq )),where σ is a Sigmoid function, δ is a GeLU activation function, and W1 and W2 are weight matrices.

[0025] Residual connections are established between the initial convolutional layer and the first attention residual module, and between adjacent attention residual modules.

[0026] Further, the interpolation layer adopts an ESRGAN (Enhanced Super-Resolution Generative Adversarial Network) model to realize interpolation processing of channel responses, and restore low-resolution feature maps to high-resolution outputs, so as to avoid the oscillation problem of traditional interpolation methods.

[0027] The ESRGAN generation network includes 16 basic residual blocks, each of which contains a double convolutional layer, a batch normalization layer and a ReLU activation function, and uses a sub-pixel convolutional layer with a scale of 2 to increase the size of the feature map.

[0028] Further, the training of the ARCENet model includes the following steps:

[0029] Under different signal-to-noise ratio conditions and NTN-TDL channel models, a data set is generated through simulation, including the LS preliminary estimation result HLS and the real channel response H;

[0030] The ARCENet model is trained using the data set, and the parameters θ are adjusted to minimize the estimation error until the model converges.

[0031] The trained model parameters are applied to the test system to optimize the channel estimation result.

[0032] Further, the method is applied to a single-input single-output (SISO) scenario of a satellite communication uplink of an ocean monitoring buoy, and each frame of the OFDM system contains N symbols, the channel impulse response changes slowly within a single frame, and the maximum path delay is less than the cyclic prefix duration.

[0033] The application further discloses a buoy satellite communication system channel estimation system in an ocean environment, which can be used to implement the above-mentioned buoy satellite communication system channel estimation method in an ocean environment, and specifically comprises:

[0034] The transmitting end module is responsible for QPSK modulation of the baseband signal, insertion of pilot symbols at predetermined subcarrier positions of the OFDM frame, conversion of the signal into a time-domain waveform through IFFT transformation, and addition of a cyclic prefix to form a final transmission signal.

[0035] Channel transmission module: simulates the signal transmission characteristics of the marine environment, uses the NTN-TDL channel model, and introduces additive white Gaussian noise interference to realistically restore the signal distortion and attenuation process of the buoy-to-satellite link.

[0036] Receiving end preprocessing module: performs a cyclic prefix removal operation on the received time domain signal, extracts frequency domain data through serial / parallel conversion and FFT transformation, separates the pilot signal therefrom, and calculates the preliminary channel response matrix using the least squares algorithm.

[0037] Core processing module: based on the attention mechanism and residual network ARCENet architecture, receives the preprocessed channel data. First, feature extraction is performed through a convolutional layer, then four serial residual blocks (two of which integrate channel attention mechanisms) are used to capture key features, and finally, high-precision channel response reconstruction is achieved through an ESRGAN interpolation layer. This module significantly improves the channel estimation accuracy in complex sea conditions.

[0038] Channel equalization module: uses the optimized channel response matrix output by the core processing module to perform distortion compensation on the received signal, restores the original transmitted data, and completes information recovery of the communication link.

[0039] Compared with the prior art, the advantages of the present application are:

[0040] 1. Significantly improves the channel estimation accuracy in complex environments

[0041] By combining deep learning methods with traditional estimation algorithms, the advantages of neural networks in capturing nonlinear features are fully utilized. The attention mechanism can dynamically enhance the learning ability of key channel response features, more accurately restoring signal distortion details in dynamic ocean channel conditions. This hybrid architecture effectively compensates for the lack of adaptability of traditional linear interpolation methods in fast-changing channels, making the data restored by the receiving end closer to the true information.

[0042] 2. Enhances the anti-interference performance of feature extraction

[0043] The introduction of a composite structure combining residual networks and attention mechanisms significantly optimizes the learning process of channel features. The attention module can autonomously select high-frequency components that are crucial for channel restoration while suppressing noise-related features. This intelligent feature selection mechanism allows the system to maintain stable performance in low signal-to-noise ratio environments, overcoming the problem of rapid estimation error growth in traditional methods under strong interference.

[0044] 3. Achieves efficient utilization of spectral resources

[0045] The interpolation compensation mechanism based on deep learning greatly reduces the dependence on the number of pilots. Through the high-precision interpolation realized by the ESRGAN generative adversarial network, the channel response can be completely reconstructed under sparse pilot configuration. This design breaks through the contradiction between pilot density and estimation accuracy in traditional methods, maximizes the data transmission bandwidth while ensuring communication reliability.

[0046] 4. Optimizing the deployment of computing resources

[0047] The introduction of the residual connection structure effectively alleviates the gradient attenuation phenomenon of deep networks and significantly improves the model training efficiency. The carefully designed lightweight network architecture avoids parameter redundancy, making the model more suitable for the limited computing power of the buoy terminal. This hardware-friendly design ensures that the system can run in real time in embedded devices, overcoming the deployment difficulties of complex deep learning models.

[0048] 5. Improving the robustness of high sea state communication

[0049] The system has adaptive learning ability for channel time-varying characteristics, especially for time-frequency offset problems caused by buoy movement. The attention mechanism can dynamically track the channel change trajectory and compensate for the estimation deviation caused by pilot position offset. This strong robustness design ensures that the system still maintains a reliable communication link in harsh marine environments. BRIEF DESCRIPTION OF DRAWINGS

[0050] Figure 1 is the pilot structure diagram of the embodiment of the present application;

[0051] Figure 2 is the system architecture diagram of the ARCENet channel estimator of the embodiment of the present application;

[0052] Figure 3 is the network structure diagram based on the attention mechanism and residual network of the embodiment of the present application;

[0053] Figure 4 is the GeLu activation function diagram of the embodiment of the present application;

[0054] Figure 5 is the ESRGAN model generation network structure diagram of the embodiment of the present application, (a) ARCENet network structure, (b) attention residual block, (c) ESRGAN model process structure. DETAILED DESCRIPTION

[0055] To make the purpose, technical scheme and advantages of the present application clearer, the following embodiments are described based on the drawings.

[0056] The present application provides a channel estimation method for a buoy satellite communication system in a marine environment, which specifically includes the following contents:

[0057] 1. Uplink communication model for satellite-to-ground communication

[0058] First, for the characteristics of the single-input single-output of the ocean monitoring buoy, in the single-input single-output SISO scenario, an orthogonal frequency division multiplexing (OFDM) communication system suitable for uplink and downlink is constructed, and a channel environment based on the Non-Terrestrial Networks-Tapped Delay Line (NTN-TDL) model is introduced. The channel estimation algorithm proposed is simulated and verified to evaluate its performance in this channel condition. The system represents the transmitted signal as a time-domain signal x(t), and the impulse response of the Rayleigh fading channel as g(t). In this system model, it is assumed that the channel noise is additive white Gaussian noise (AWGN), and the noise term is represented by z(t). Then the received signal y(t) in the time domain can be represented as shown in equation 1:

[0059]

[0060] Suppose each frame contains N OFDM symbols, and assume that the channel impulse response changes slowly within a single frame. If the maximum path delay of the channel is less than the duration of the cyclic prefix (CP), the relationship between the OFDM symbols between the transmitting end and the receiving end can be represented by equation 2:

[0061] Y = H · X + Z (2)

[0062] wherein, represents the channel gain, represents the transmitted OFDM data, represents the received OFDM data, N f represents the number of subcarriers contained in an OFDM symbol, represents an additive white Gaussian noise matrix with mean and variance of zero.

[0063] 2. Pilot structure

[0064] To ensure high reliability and concealment of buoy satellite communication in high sea conditions, in view of the narrow bandwidth and large delay constraints of buoy satellite communication, the design of short data frame structure for buoy satellite communication is studied to realize reliable communication of buoy terminal.

[0065] Around the characteristics of high reliable communication, the narrow bandwidth and short burst OFDM air interface waveform system is designed, and the specific pilot frame structure is as shown in Figure 1 .

[0066] By analyzing the current research status of channel estimation, sparse pilot configuration scheme is widely used in current communication systems. This design is mainly based on the following two aspects: first, although dense pilot deployment can improve estimation accuracy, it will significantly sacrifice spectrum resources; second, the time-varying channel environment requires that the pilot design must consider the implementation complexity and estimation reliability. The specific pilot structure conforms to the frame structure defined in 3GPP TS 38.211 standard. These pilot symbols all use QPSK modulation mode.

[0067] In the frequency domain, N pf represents the number of pilot subcarriers placed on each pilot symbol, N ps represents the number of pilot symbols placed in each frame, the received pilot symbol and the transmitted symbol The relationship between them can be expressed as equation 3:

[0068] Y p = H p · X p + Z p (3)

[0069] Where, the channel gain coefficient at the pilot position is represented as The impact of AWGN on the pilot signal is represented as The non-blind channel estimation method depends on the known pilot signal and the received pilot signal value to obtain the predicted channel gain

[0070] 3. Channel estimation model based on attention mechanism and residual network

[0071] Compared with the above traditional pilot-based channel estimation algorithms, deep learning algorithms have significant advantages in extracting channel nonlinear features, especially suitable for dynamic communication scenarios in ocean satellite Internet of Things. Considering the low computational complexity of LS channel estimation algorithm, deep learning algorithm and LS channel estimation algorithm are used for channel estimation. The deep learning method can further correct the estimation error of the LS channel estimation algorithm, and use the nonlinear fitting ability of deep learning to capture the channel characteristics that the LS estimation fails to compensate, thereby reducing the error.

[0072] This model is designed based on a combination of an attention mechanism and a residual network structure (ARCENet) with the LS channel estimation method. When facing complex communication channels, it is necessary to extract key information for utilization while suppressing noise interference. The attention mechanism can selectively highlight key features while suppressing irrelevant or interfering information, thereby improving the model's focus on important channel features. Introducing a residual neural network structure allows for a direct mapping between input information and target output through residual connections, effectively improving the model's training efficiency. Simultaneously, the residual structure can alleviate the gradient vanishing problem that may occur during the training of deep neural networks, thus enhancing the network's stability and learning ability.

[0073] The channel estimation model based on the attention mechanism and residual network structure of this invention consists of convolutional layers, composite residual blocks and interpolation layers. The interpolation layer introduces a super-resolution generative adversarial network (ESRGAN) to perform interpolation processing, and continuously judges and improves the estimation effect through a discriminant network.

[0074] The following section will systematically introduce the system architecture design and network model based on the proposed channel estimation model of this invention.

[0075] 3.1 Channel Estimation Model System Architecture

[0076] The proposed deep learning-based channel estimation algorithm system architecture is as follows: Figure 2 As shown, the data information at the data sending end is represented by X. data This indicates that the inserted initial pilot information is represented by X. pilot This indicates that the data received by the receiver is represented by Y. data Representation. Symbol This represents the result of the preliminary LS channel estimation. The symbol x(t) represents the time-domain signal obtained after OFDM modulation, and the symbol y(t) represents the time-domain signal received by the receiver.

[0077] At the transmitter of this communication system, QPSK is used to modulate the baseband signal, followed by the insertion of pilot information. These signals are allocated to designated subcarrier positions in the OFDM frame according to a predetermined resource mapping scheme. The complete transmission frame undergoes frequency-time domain conversion via Inverse Fast Fourier Transform (IFFT), and a cyclic prefix is ​​added to finally form a time-domain waveform for transmission through the physical channel.

[0078] At the receiving end, the received signal y(t) is first subjected to cyclic prefix removal. Then, the received signal is converted into frequency domain data through serial-to-parallel conversion and Fast Fourier Transform (FFT). The pilot signal Y is extracted from the frequency domain data based on the known position information. pilot Using the received pilot information Y pilot and the known pilot information X from the transmitting endpilot After the LS channel estimator, a preliminary channel estimation is performed to obtain preliminary channel estimation information This is an N f ×N s complex matrix.

[0079] Then the preliminary channel estimation result obtained by the LS channel estimator needs to be preprocessed, the purpose of which is to separate the real part and the imaginary part of the complex matrix and convert it into a real number array so that it can be input to the neural network for subsequent processing. The dimension of the real number array is 2 × N f ×N s , where the first dimension represents the number of channels, and the first channel contains real part data and the second channel contains imaginary part data.

[0080] Under different signal-to-noise ratio conditions and various channel model conditions, data is randomly generated and collected and the real channel response H, a dataset containing training samples and corresponding labels is constructed. The core goal of ARCENet is to achieve the function of formula 4 through a neural network:

[0081]

[0082] where f represents the function mapping relationship represented by the neural network ARCENet, represents the refined result closer to the real channel response obtained through the neural network, and θ represents the parameters that need to be trained in ARCENet. First, the neural network is trained using the dataset to adjust the parameters of the neural network so that it converges to obtain the optimal parameters θ, and then the trained parameters are applied to the test communication system to optimize the results obtained by the LS channel estimation algorithm, thereby obtaining the channel response matrix Finally, after channel equalization, the original signal is restored.

[0083] 3.2 Structure design based on attention mechanism and residual network

[0084] Figure 3 The network model structure based on attention mechanism and residual network is shown. Among them Figure 3 (a) is the network structure of ARCENet model, the input of ARCENet model is 2 × N f ×N sThe model first passes through a two-dimensional convolutional layer (2 convolutional kernels, 3x3 kernel size) for preliminary feature extraction and noise suppression. Then four residual modules are connected in series, among which the second and fourth modules introduce an attention mechanism, forming an attention residual module. In order to enhance the efficiency of feature transmission, residual connections are established between the initial convolutional layer and the first attention residual module, as well as between adjacent attention residual modules, constructing a composite residual structure. Finally, an output is completed through a two-dimensional convolutional layer with the same parameters as the input layer. This design not only retains the advantages of residual networks, but also enhances the extraction ability of key features through attention mechanisms. The design of the composite residual structure not only enhances the feature expression ability, but also effectively alleviates the gradient vanishing problem caused by deepening the network, thereby improving the training efficiency and performance of the model.

[0085] The structure of the attention residual block is as shown in Figure 3 (b), after the second convolutional layer, an SE (Squeeze-and-Excitation Networks) structure is introduced to enhance the expression ability of channel features. The structure consists of a pooling layer, two fully connected layers, a GeLu activation function layer and a Sigmoid function layer. The reason for using GeLu activation function is that compared with traditional activation functions such as ReLU, GeLU has a smoother output curve, as shown in Figure 4 . This smoothness makes it behave like a nearly linear mapping in the vicinity of the zero value interval, which is conducive to the flow of gradients during backpropagation, thereby accelerating the convergence speed of the model. In addition, considering that there may be negative features in the channel response, the non-zero lower bound property of the GeLU function avoids the truncation of negative information, thereby more comprehensively preserving the original feature distribution and improving the fitting ability of the model. The introduction of the SE module helps the network to focus on important channel information while suppressing irrelevant or redundant features, making the overall feature extraction process more efficient and accurate.

[0086] Input layer design: The complex channel matrix obtained by LS channel estimation is used as the input of the network. Since neural networks require real inputs, the real and imaginary parts of the complex matrix are separated into two independent channels. The original dimension of the input data after data preprocessing in this experiment is 2x612x14, where "2" represents the real and imaginary parts of two channels, "612" represents 612 effective subcarriers contained in each symbol, and "14" represents the frame structure of the OFDM system using 14 symbols.

[0087] Convolutional layer: According to the structure design of the ARCENet model, the first layer of convolutional module adopts two convolution kernels for feature extraction operation, and the size of each convolution kernel is 2x3x3, which is used to capture the local shallow features in the input signal. After this convolution operation, the model introduces Gaussian error linear unit (GeLu) as the activation function to enhance the non-linear expression ability, and its smoothness performs better in feature learning, which helps to improve the stability and convergence speed of training. The mathematical expression of GeLu is shown in equation 5:

[0088]

[0089] When the input data is processed by the second layer of convolution, the output CxHxZ feature map is obtained. Then, the feature map is sent to the SE module to further extract key features. In the SE module, first, the feature map is compressed in the channel dimension by global average pooling, and the compressed feature map Cx1x1 is obtained sq , the process of which is shown in equation 6:

[0090]

[0091] Subsequently, the compressed feature matrix is excited: the feature is first input into two fully connected layers, and the GeLu activation function is embedded in it, to further learn the feature representation in the channel dimension. Then, the feature vector output by the second fully connected layer is nonlinearly transformed by the Sigmoid function, thereby generating a channel attention map H ex consistent with the original feature map channel number:

[0092] H ex = σ (W2δ (W1H sq )) (7)

[0093] where δ represents the GeLu activation function, and are the weights of the two fully connected layers, and the parameter τ represents the scaling operation on the number of channels, since the number of channels is 2, τ = 1, i.e. no reduction of the number of channels. σ represents the Sigmoid function. The original input feature map is multiplied element by element with the obtained channel attention feature map H ex , to obtain the feature map with fused attention information

[0094] After the composite residual structure, interpolation processing is needed, and the channel response recovery process based on the interpolation method is greatly affected by the value at the pilot position, and when the number of interpolation nodes and the order of the interpolation polynomial increase, the interpolation curve is prone to oscillation and deviation problems. In order to solve this phenomenon, the application introduces a model based on a super-resolution generative adversarial network (ESRGAN) for interpolation processing. Specifically, the role of the channel interpolation layer can be understood as a super-resolution reconstruction process, that is, restoring a low-resolution single-channel image to a corresponding high-resolution image, thereby achieving more accurate and smooth interpolation effects.

[0095] Interpolation layer: the ESRGAN model used for interpolation processing, the generation network structure of the ESRGAN model is as shown in Figure 5 It contains 16 basic residual blocks, each residual block contains a double convolution layer, and the convolution layer is followed by a batch normalization layer and an activation function ReLu, and two sub-pixel convolution layers with a scale of 2 are used to increase the size of the feature map.

[0096] The model obtains a feature map with the same dimension as the complete channel time-frequency response through the interpolation layer, and finally outputs the complete channel estimation value through a two-dimensional convolution layer with two kernel sizes of 36x7.

[0097] In another embodiment of the application, a channel estimation system for a buoy satellite communication system in a marine environment is provided, which can be used to implement the channel estimation method for a buoy satellite communication system in a marine environment described above, and specifically comprises:

[0098] The transmitting end module is responsible for QPSK modulation of the baseband signal, insertion of pilot symbols at the predetermined subcarrier position of the OFDM frame, conversion of the signal to a time-domain waveform through IFFT transformation, and addition of a cyclic prefix to form a final transmission signal.

[0099] The channel transmission module simulates the signal transmission characteristics of the marine environment, adopts the NTN-TDL channel model, introduces additive white Gaussian noise interference, and truly restores the signal distortion and attenuation process of the buoy-to-satellite link.

[0100] The receiving end preprocessing module performs a cyclic prefix removal operation on the received time-domain signal, extracts frequency-domain data through serial / parallel conversion and FFT transformation, separates the pilot signal therefrom, and calculates a preliminary channel response matrix using a least squares algorithm.

[0101] Core processing module: based on the attention mechanism and residual network (ARCENet) architecture, receiving the preprocessed channel data. First, feature extraction is performed through the convolutional layer, then four residual blocks in series (two of which integrate channel attention mechanism) are used to capture key features, and finally high-precision channel response reconstruction is realized through the ESRGAN interpolation layer. This module significantly improves the channel estimation accuracy in complex sea conditions.

[0102] Channel equalization module: using the optimized channel response matrix output by the core processing module, the received signal is distorted and compensated, and the original transmitted data is restored, completing the information recovery of the communication link.

[0103] In another embodiment of the present application, a terminal device is provided, which includes a processor and a memory, the memory being used to store a computer program, the computer program including program instructions, and the processor being used to execute the program instructions stored in the computer storage medium. The processor can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc., which are the computing core and control core of the terminal, and are suitable for implementing one or more instructions, and are specifically suitable for loading and executing one or more instructions to implement corresponding method processes or corresponding functions; the processor of the embodiment of the present application can be used for the operation of the buoy satellite communication system channel estimation method.

[0104] In another embodiment of the present application, a storage medium is provided, specifically a computer readable storage medium (Memory), which is a memory device in a terminal device, used to store programs and data. It can be understood that the computer readable storage medium here can include the built-in storage medium in the terminal device, and of course can also include the expansion storage medium supported by the terminal device. The computer readable storage medium provides a storage space, which stores the operating system of the terminal. Moreover, one or more instructions suitable for being loaded and executed by the processor are also stored in the storage space, and these instructions can be one or more computer programs (including program codes). It should be noted that the computer readable storage medium here can be a high-speed RAM memory, or a non-volatile memory such as at least one disk memory.

[0105] The one or more instructions stored in the computer-readable storage medium can be loaded and executed by a processor to implement the corresponding steps of the channel estimation method for the buoy satellite communication system in the above-described embodiments.

[0106] Those skilled in the art will appreciate that embodiments of the present application can be supplied as a method, a system, or a computer program product. Therefore, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage devices, etc.) having computer-usable program code embodied therein.

[0107] The present application is described in reference to the flowcharts and / or block diagrams of the method, apparatus (system) and computer program product according to the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing apparatus to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing apparatus produce the functions specified in the flowcharts and / or block diagrams of the method, apparatus (system) and computer program product. Figure 1 one or more flows and / or blocks Figure 1 means for performing the function specified by the flow or flows and / or block or blocks.

[0108] These computer program instructions can also be stored in a computer-readable memory that can direct the computer or other programmable data processing apparatus to work in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including instruction means, which implements the functions specified in the flowcharts and / or block diagrams of the method, apparatus (system) and computer program product. Figure 1 one or more flows and / or blocks Figure 1 means for performing the function specified by the flow or flows and / or block or blocks.

[0109] These computer program instructions can also be loaded to the computer or other programmable data processing apparatus, so that a series of operation steps are performed on the computer or other programmable data processing apparatus to produce a computer-implemented process, so that the instructions executed on the computer or other programmable data processing apparatus provide the functions specified in the flowcharts and / or block diagrams of the method, apparatus (system) and computer program product. Figure 1 one or more flows and / or blocks Figure 1 means for performing the function specified by the flow or flows and / or block or blocks.

[0110] Those skilled in the art will appreciate that the embodiments described herein are presented for purposes of illustration and understanding of the methods of practicing the present application and are not intended to limit the scope of the present application in any way. Various modifications and alterations of this application can be made by those skilled in the art without departing from the scope and spirit of this application, which is measured by the appended claims and their equivalents.

Claims

1. A method for channel estimation of a buoy satellite communication system in a marine environment, characterized by, Specifically comprising the following steps: At the receiving end, an orthogonal frequency division multiplexing (OFDM) signal is received, which is transmitted through a non-terrestrial network-tap delay line (NTN-TDL) model channel and is affected by additive white Gaussian noise (AWGN); Pilot symbols are extracted from the received signal, which are configured as a sparse pilot structure in the time-frequency domain to reduce spectral resource occupation; A least square (LS) algorithm is used to make a preliminary channel estimation on the pilot positions to obtain a preliminary channel response matrix Hls, where Hls is a complex matrix with a dimension of N f ×N s , N f represents the number of subcarriers, and N s represents the number of symbols. HLS is pre-processed to include separation of real and imaginary parts, forming a real array of dimension 2 x N f as input s ​ The pretreated data is input into an ARCENet model based on an attention mechanism and a residual network, fine channel estimation is performed through the model, and an optimized channel response matrix is obtained The ARCENet model is implemented through a nonlinear mapping function f θ is a model training parameter Channel equalization is performed using the optimized channel response matrix to restore the original transmitted signal.

2. The method of claim 1, wherein The pilot symbols use QPSK modulation and are configured as a narrow bandwidth, short burst OFDM frame structure; The pilot structure contains N ps pilot symbols in each frame, N pf pilot subcarriers in each pilot symbol, the relationship between the received pilot signal Y p and the transmitted pilot signal X p satisfies Y p = H p · X p + Z p , where H p is the pilot position channel gain and Z p is a noise term.

3. The method of claim 1, wherein The system architecture of the ARCENet model includes a transmitting end and a receiving end: At the transmitting end, the baseband signal is modulated using QPSK and pilot information is inserted, which is allocated to the OFDM frame subcarrier position through resource mapping, and then inverse fast Fourier transformed and cyclic prefix added to form a time domain signal; At the receiving end, the received signal is converted into frequency domain data by de-cyclic prefix, serial-to-parallel conversion, and fast Fourier transform, and the pilot signal Y pilot is extracted for LS preliminary estimation.

4. The method of claim 1, wherein The ARCENet model includes an input layer, a convolutional layer, a plurality of residual modules, and an interpolation layer connected in sequence: The input layer receives a real array of dimension 2 x N f x N s ; The convolutional layer includes a first two-dimensional convolutional layer using 2 convolution kernels with a kernel size of 3x3 for preliminary feature extraction and noise suppression, followed by a Gaussian error linear unit (GeLU) activation function, the expression of which is 5. The method of claim 4, wherein The residual module is a composite residual structure, including four residual blocks connected in series, and the second and fourth residual blocks introduce an SE (Squeeze-and-Excitation) attention mechanism to form an attention residual block; The attention residual block includes two convolutional layers and an SE module, the SE module compresses the feature map through global average pooling, and calculates and generates a channel attention map H through a full connection layer and a GeLU activation ex = σ(W2δ(W1H sq )), where σ is a Sigmoid function, δ is a GeLU activation function, and W1 and W2 are weight matrices. Residual connections are established between the initial convolutional layer and the first attention residual module, as well as between adjacent attention residual modules.

6. The method of claim 4, wherein The interpolation layer uses an ESRGAN (Enhanced Super-Resolution Generative Adversarial Network) based model to implement interpolation processing of the channel response, restoring low-resolution feature maps to high-resolution output to avoid the oscillation problem of traditional interpolation methods; The ESRGAN generation network includes 16 basic residual blocks, each containing double convolutional layers, batch normalization layers, and ReLU activation functions, and uses a sub-pixel convolutional layer with a scale of 2 to increase the feature map size.

7. The method of claim 1, wherein The training of the ARCENet model includes the following steps: Under different signal-to-noise ratio conditions and NTN-TDL channel models, a dataset is generated through simulation, including the LS preliminary estimation result HLS and the true channel response H; The ARCENet model is trained using the dataset, and the parameters θ are adjusted to minimize the estimation error until the model converges; The trained model parameters are applied to the test system to optimize the channel estimation result.

8. The method of claim 1, wherein The method is applied to a single-input single-output (SISO) scenario of an ocean monitoring buoy satellite communication uplink, and each frame of the OFDM system contains N symbols, the channel impulse response changes slowly within a single frame, and the maximum path delay is less than the cyclic prefix duration.

9. A channel estimation system for a buoy satellite communication system in a marine environment, characterized by The system can be used to implement the buoy satellite communication system channel estimation method of any one of claims 1 to 8, specifically comprising: The transmitting end module is responsible for QPSK modulation of the baseband signal, insertion of pilot symbols at the predetermined subcarrier position of the OFDM frame, conversion of the signal to a time domain waveform through IFFT transformation, and addition of a cyclic prefix to form the final transmitted signal. Channel transmission module: simulate the signal transmission characteristics of the ocean environment, use NTN-TDL channel model, introduce additive white Gaussian noise interference, and truly restore the signal distortion and attenuation process of the buoy-to-satellite link. Receive-end preprocessing module: remove the cyclic prefix of the received time-domain signal, extract the frequency-domain data through serial / parallel conversion and FFT transformation, separate the pilot signal from it, and calculate the preliminary channel response matrix using the least squares algorithm. Core processing module: based on the attention mechanism and residual network ARCENet architecture, receive the preprocessed channel data. First, extract features through convolutional layers, then capture key features using four serial residual blocks (two of which integrate channel attention mechanisms), and finally implement high-precision channel response reconstruction through the ESRGAN interpolation layer. This module significantly improves the channel estimation accuracy in complex sea conditions. Channel equalization module: use the optimized channel response matrix output by the core processing module to perform distortion compensation on the received signal, restore the original transmitted data, and complete the information recovery of the communication link.

Images