A channel estimation simulation method based on multi-feature input GRU

Abstract

The application relates to the field of communication technology research, in particular to a channel estimation simulation method based on a multi-feature input GRU, which comprises the following steps: collecting channel estimation at a pilot under a real scene (such as a high-speed scene) and arranging the pilot signal in an OFDM time-frequency domain resource grid, taking the pilot time, frequency, antenna index and channel estimation value as input data for model training; inputting the training data into a multi-feature GRU model for training to obtain a pre-trained GRU channel estimation model; applying the trained channel estimation model to an online simulation platform, when the MSE increases, the GRU model parameters are fine-tuned online through a Reptile meta-learning algorithm to obtain a real-time GRU channel estimation model; through the combination of the Reptile algorithm and the GRU model, the whole channel estimation system has channel estimation and real-time adaptation capabilities; the application has a simpler and clearer structure and fewer parameters, which makes the model training more efficient and easier to realize on a resource-limited simulation platform.

Description

Technical Field

[0001] This invention relates to the field of communication technology research, specifically to a channel estimation simulation method based on a multi-feature input GRU. Background Technology

[0002] In recent years, with the continuous development of my country's economy and technology, the emergence of 4G and 5G networks has improved people's communication capabilities and continues to evolve. More and more services are relying on high-quality, high-efficiency wireless mobile communication systems. However, in certain special scenarios, such as high-speed mobile scenarios and densely populated urban scenarios, frequency shifts and phase changes caused by Doppler effects and multipath effects still pose significant challenges to the performance of communication systems.

[0003] In traditional channel estimation processes, model-based methods, such as least squares and maximum likelihood estimation, are typically used to predict and compensate for signal distortion. These methods rely on accurate modeling of channel characteristics, but when faced with complex scenarios such as high-speed movement, multipath effects, and Doppler shift, traditional models may fail to adequately capture the dynamic changes in the channel, leading to a decrease in estimation accuracy.

[0004] Neural networks, by simulating the connections and information transmission of neurons and training network models with input data, can learn and recognize complex patterns. This method can better adapt to the nonlinear and time-varying characteristics of channels, especially in complex environments where traditional methods are difficult to handle. Through deep learning, neural networks can be trained to recognize patterns and structures in the channel, thereby achieving more efficient and accurate channel estimation in various environments. Patent No. CN113285899A discloses a time-varying channel estimation method and system based on deep learning. By constructing a channel model, utilizing historical channel information and received pilot signals, and combining least squares estimation (LS) and first-order autoregressive (AR) models with a backpropagation neural network (BP neural network), channel estimation is performed. In the offline training phase, this method optimizes network parameters using quantized conjugate gradient descent to achieve highly efficient online channel estimation; deep learning techniques, especially BP neural networks, capture channel variation characteristics to improve estimation accuracy. However, this method uses received complex signals as input for training data. These complex signals typically contain time, frequency, and user data or pilot information. If the same complex signal is received in different time-frequency domain resource grids, the same channel estimate will be obtained. Therefore, incorporating user data into the training process leads to inaccuracies in the channel estimation model. Furthermore, this method is only applicable to high-speed mobile scenarios, exhibiting significant limitations. Patent CN110266620A discloses a channel estimation method for 3D MIMO-OFDM systems based on convolutional neural networks. This method trains a channel estimation model using a CECNN convolutional neural network, ultimately inputting data and predicting the complete channel response value of the 3D MIMO-OFDM system. The CECNN convolutional neural network used in this method has three convolutional layers, which increases training overhead and time. Additionally, this method is also only applicable to 3D MIMO scenarios, similarly exhibiting limitations. Therefore, based on these specific problems, it is especially necessary to simplify the network structure to accommodate the massive amount of pilot information during communication and reduce training overhead; then, it is also necessary to optimize the model input, remove the influence of user data on channel estimation, and improve the accuracy of simulation; finally, it is especially necessary to improve the generalization ability of the neural network to adapt to different scenarios. Summary of the Invention

[0005] The main objective of this invention is to improve the accuracy and adaptability of LTE physical channel estimation algorithms, accelerate the training speed of channel estimation models, reduce training overhead, improve the real-time performance and versatility of models, and provide a channel estimation simulation method based on multi-feature input GRU to solve the problems mentioned in the background art.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] A channel estimation simulation method based on a multi-feature input GRU, the method comprising:

[0008] S100. Design a method for organizing training data at the pilot, shielding the influence of received pilot data on the channel estimation value, and reducing the factors affecting the channel estimation result to time, frequency and the antenna where the pilot is located;

[0009] S200. Design a GRU model based on multi-feature input to simultaneously process three key features that affect the channel estimation value, and more comprehensively capture the dynamic characteristics of channel changes.

[0010] S300. When performing online channel estimation, the Reptile algorithm is used to learn the best estimate of the current model parameters online, and to obtain the GRU network model with the best parameters and the channel estimates at other locations except for the pilots.

[0011] Preferably, in the implementation process of training data organization at pilot S100:

[0012] First, the channel estimation scenario is determined, and pilot data is obtained using the LS method via MATLAB simulation software, or DMRS pilot data within a specific time period is collected for that scenario. The collected pilot data is then screened and cleaned, with missing and outlier values ​​imputed to ensure the validity of the training data.

[0013] Furthermore, the time, frequency, and antenna information at the pilot points are organized, and the DMRS pilot signal on the i-th antenna at time slot i and frequency f is denoted as P. i t,f Then the training task formed at that pilot frequency is:

[0014]

[0015] Among them, t P Indicates the time slot in which the pilot signal is located, f P Indicates the frequency domain location of the pilot signal, i P H indicates the antenna subscript where the pilot signal is located. P This represents the channel estimate measured at the receiver for the pilot signal, and is in complex form.

[0016] In the simulation environment, since the pilot signal received by the receiver is represented in complex form, it is necessary to interpret the pilot P... i t,f The channel estimate is decomposed to obtain the real and imaginary parts of the channel estimate at the pilot, denoted as:

[0017]

[0018] The training data is divided into two groups, where the training data for each step size is represented as follows:

[0019]

[0020] Since the DMRS signals in different time slots are arranged at equal intervals, and the interval between two consecutive time slots is 7 OFDM symbols (0.5 ms), then using 0.5 ms as the step size for the input time series of the GRU model, the real and imaginary input matrices are obtained as follows:

[0021]

[0022] Preferably, in the GRU model based on multi-feature input designed in S200:

[0023] First, the algorithm structure is determined. The GRU algorithm sets up two gating units: an update gate and a reset gate. These two gating units control the degree of influence of historical pilot information on current pilot information.

[0024] Next, an update gate is designed to control how much of the previous pilot state information is retained and how much new pilot state information is introduced during the current hidden transition. The goal is to output a value between 0 and 1 using the sigmoid function, which determines the weight of retaining the hidden state from the previous time step. If the update gate value is close to 1, it means more historical information is retained; if it is close to 0, it means more historical pilot information is forgotten, while new pilot information is introduced. By using the sigmoid function, the transition between weights is smooth. This smooth transition helps the model find a balance between retaining historical pilot information and introducing new pilot information, avoiding drastic changes. Due to the use of nonlinear decision-making, this helps the model capture more complex channel change characteristics and patterns, improving the model's adaptability to different environments. By introducing the update gate, the GRU model can more flexibly handle the relationship between historical and new pilot information when dealing with channel estimation problems. The update gate calculation formula is:

[0025] z t =σ(W z ·[h t-1 ,x t,f,i ])

[0026] Where, x t,f,i This represents the pilot signal input at the current time step (i.e., the t-th time slot), where f and i represent the frequency and antenna index of the current pilot signal, respectively. σ represents the sigmoid activation function, and W... z It is the weight matrix of the updated gate, h t-1 This is the pilot hidden state from the previous time step. t,f,iThe input dimension is three, which means that under this model, the channel estimation result is only affected by three factors: time, frequency and the antenna. This eliminates the interference caused by traditional convolutional neural networks using received pilot signals carrying pilot data as training data, and further improves the accuracy of the network model.

[0027] Furthermore, the reset gate for the GRU model is designed. Similar to the update gate, the calculation formula for the reset gate is as follows:

[0028] r t =σ(W r ·[h t-1 ,x t,f,i ])

[0029] Among them, W r This is the weight matrix of the reset gate. The output of the reset gate is related to the previous hidden state h. t-1 Element-wise multiplication is performed to weight the historical pilot information, controlling its influence on the current output. In traditional RNNs, errors accumulate over time, leading to long-term dependency issues. The reset gate reduces this error propagation and improves the model's generalization ability by allowing the model to reset historical pilot information when necessary.

[0030] Furthermore, the candidate hidden states are designed using the following formula:

[0031]

[0032] Where W is the weighted matrix of the candidate hidden states, and * denotes element-wise multiplication. Candidate hidden states A new hidden state is proposed for the current time step. If the reset gate has forgotten most of the pilot history information, then the candidate hidden state will be mainly determined by the current input.

[0033] Furthermore, the model will calculate the final hidden state, the current final hidden state h. t It is the hidden state h of the previous time step. t-1 Candidate hidden states at the current time step The weighted sum. The weights are determined by the output z of the updated gate. t The decision has been made. The details are as follows:

[0034]

[0035] This weighted sum allows the GRU to dynamically adjust its dependence on historical and new pilot information, thus better handling long-term dependencies in sequence data. In this way, the GRU can effectively capture dynamic changes in channel estimation at the pilots and improve model performance.

[0036] Furthermore, the real part is input into the Input matrix. re and the imaginary part of the input matrix. im The GRU model is trained by inputting the parameters into it, resulting in optimized model parameters and a trained GRU model. These parameters include the update gate, reset gate, weight matrix of candidate hidden states, and bias term. The trained GRU model can effectively capture temporal dependencies in the sequence based on the input real and imaginary parts of the data, and accurately estimate the channel state or other target variables. For online channel estimation, the time slot, frequency, and antenna index of the received pilot signal are first combined into three-dimensional data and input into the trained GRU model to obtain the channel estimate at the pilot. Then, the channel estimate of the PUSCH channel is obtained through interpolation.

[0037] Preferably, during online channel estimation in S300, the Reptile algorithm is used to efficiently update the parameters of the pre-trained GRU network model.

[0038] First, the structure of the Reptile algorithm is determined: the Reptile algorithm consists of a base learner and a meta-learner. The goal of the base learner is to quickly learn from the M pilot signals received online, adapting to the current pilot signals through a small number of gradient updates. After the base learner completes training in the previous time step, it feeds the learned parameters back to the meta-learner for meta-level updates. The meta-learner updates the parameters of the GRU neural network by collecting parameter update information from the base learners across multiple tasks. This update strategy enables the model to maintain good real-time performance and channel estimation performance even in complex channel environments. Furthermore, the parameter estimation of the base learner is designed, with the following formula:

[0039]

[0040] in, It is task T j The loss function of the base learner, where θ is the initial parameter value provided by the meta-learner to the base learner. The parameters in the base learner are updated through M iterations to obtain parameter estimates. Finally, the parameters in the meta-learner are updated:

[0041]

[0042] The Reptile algorithm allows for real-time fine-tuning of parameters in GRU networks. Furthermore, due to its simpler structure, Reptile avoids second-derivative calculations compared to the traditional MAML algorithm, reducing computational complexity. Especially for large-scale signal transceiver systems with 8 to 64 antennas, Reptile significantly reduces computation time, enabling the GRU channel estimation model to achieve better generalization capabilities.

[0043] Compared with the prior art, the beneficial effects achieved by the present invention are:

[0044] This invention first collects channel estimates and their time-frequency domain information at pilot points in real-world scenarios (such as high-speed scenarios), and classifies and labels the pilot signals in the OFDM time-frequency domain resource grid to obtain the time, frequency, antenna index, and channel estimate of the DMRS pilot signal as input data for model training. Then, the training data is input into the model for training, resulting in an optimized multi-feature GRU channel estimation model. Finally, the trained channel estimation model is applied to an online simulation platform, and the Reptile meta-learning algorithm is used for online fine-tuning, yielding an optimized multi-feature GRU channel estimation model. This model possesses real-time update and adaptive adjustment capabilities. Compared to existing channel estimation algorithms such as LS, the system of this invention significantly improves the accuracy and effectiveness of channel estimation in specific environments, enhancing the model's generalization ability. By utilizing a multi-feature GRU network and the Reptile meta-learning algorithm, this system can not only quickly adapt to new tasks with a small number of samples but also achieve efficient online fine-tuning in practical applications, further optimizing the performance of channel estimation. Attached Figure Description

[0045] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0046] Figure 1 This is a flowchart of the channel estimation simulation method based on a multi-feature input GRU according to the present invention;

[0047] Figure 2 This is a flowchart of a channel estimation simulation method based on a multi-feature input GRU according to the present invention. Detailed Implementation

[0048] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0049] Please see Figures 1-2 The present invention provides the following technical solution:

[0050] Example 1:

[0051] like Figure 1 As shown, the present invention includes the following steps:

[0052] S100. Design a method for organizing training data at the pilot. This method effectively shields the influence of received pilot data on the channel estimation value, reduces the factors affecting the channel estimation result to time, frequency and the antenna where the pilot is located, simplifies data training and improves the accuracy of the subsequent GRU model.

[0053] Preferably, firstly, DMRS is selected as the pilot signal, and the channel to be estimated is the PUSCH channel. Since DMRS is evenly distributed in each subframe with a step size of 0.5ms, it simplifies the selection of training data compared to selecting pilots at other locations. Next, the time-frequency domain, antenna index, and channel estimation value at the current pilot are labeled and sorted to obtain the original training input data. Subsequently, the channel estimation value is split to obtain the real and imaginary parts of the channel estimation at the pilot, forming two sets of training data respectively. This method of organizing training data effectively shields the model from interference from pilot data and avoids the situation where the receiver calculates the same channel estimation result when receiving the same complex pilot signal at different time-frequency domain resources, thereby improving the accuracy and effectiveness of channel estimation.

[0054] For a concrete example, in a MATLAB simulation environment, for an 8-antenna MIMO system, the receiver receives a DMRS pilot signal, denoted as 1+6i. The channel estimate at this pilot point is calculated as 9+2i using the LS method. However, since the channel estimate characterizes the channel's properties and is unrelated to the pilot signal data, and the pilot signal 1+6i carries pilot data, frequency information, and time information, directly using it as training data for the model would significantly reduce the accuracy of model training.

[0055] Taking this pilot as an example, the present invention obtains the current pilot position through function input parameters.

[0056] With subframe = 3, slot = 1, frequency = 15kHz, and antennaIndex = 1, the current pilot signal is located in the 7th time slot using the subframe and slot values. Therefore, the original training data at this pilot signal location can be expressed as: T j =[7,1,15,1,9+2i] T , j indicates that the data is the j-th group of data in the training task dataset. The complete original training dataset obtained by this method is T = [T1, T2, ... T]. j ,…T N ] T Compared to the previous method of directly using received pilot signals as training data, the training data organization method of this invention removes the interference of received data on channel estimation calculation. In addition, adding antenna dimension to the training data can also expand the versatility of subsequent models.

[0057] S200. Design a GRU model based on multiple feature inputs. Unlike the traditional time-based single-feature LSTM algorithm, the GRU model based on multiple feature inputs can simultaneously process three key features that affect the channel estimation, thereby capturing the dynamic characteristics of channel changes more comprehensively.

[0058] Preferably, a multi-feature-based GRU model is designed, where both the output and input of the model are real-valued signals. The specific implementation scheme is as follows: First, the collected complex values ​​of the channel estimation are split to obtain real-part data and imaginary-part data; then, the real-part data and imaginary-part data are respectively used as training data and input into the model for training to obtain the real-part GRU model and the imaginary-part GRU model; during online verification, the real-part channel estimate and the imaginary-part channel estimate are obtained by simultaneously inputting the time coordinate and frequency coordinate of the current RE into the real-part GRU model and the imaginary-part GRU model; finally, the obtained real-part and imaginary-part channel estimates are concatenated to obtain the complete channel response value.

[0059] For a specific example: the original pilot training data T = [T1, T2, ... T] j ,…T N ] T In MATLAB, the real and imaginary training data at the pilot frequency are obtained using the real and imaginary functions, respectively: T re =real(T), T im =imag(T), which sets T re and T imTwo identical multi-feature GRU models were trained as inputs, resulting in a real-part GRU model and an imaginary-part GRU model, respectively. During online validation, the time, frequency, and antenna index of the current RE were simultaneously input into both the real-part and imaginary-part GRU models to obtain real and imaginary channel estimates. Finally, the real and imaginary channel estimates were concatenated to obtain the complete channel response. This decomposition of the channel estimate effectively reduces the simulation complexity, simplifies the model structure, and improves computational efficiency. In the memory-constrained MATLAB simulation environment, decomposing the channel estimate at the original pilot reduces the output dimension of the model, decreases memory consumption during training, and improves the feasibility of the simulation.

[0060] Preferably, a GRU model based on multiple features is designed, and its specific implementation includes:

[0061] 1) Determine the input features and output of the model. Since the channel to be estimated is a PUSCH channel, the input of the model is the time slot, frequency, and antenna index of the DMRS pilot, and the output of the model is the channel estimate of the pilot.

[0062] 2) The influence of historical pilot information on current pilot information can be controlled by setting update gates and reset gates.

[0063] 3) Calculate candidate hidden states, which provide a new hidden state proposal for the current time step, thus improving the model's generalization ability.

[0064] 4) Perform weighted fusion of the candidate hidden state and the hidden state of the previous time step to calculate the final hidden state and obtain the output of the model.

[0065] For a specific example, for an 8-antenna MIMO system, the original training dataset at the pilot is: T = [T1, T2, ... T j ,…T N ] T The parameters of the GRU model are set as follows: input_size = 3, indicating that the dimension of the input features is 3; hidden_size = 10, indicating that the dimension of the hidden layer is 10; and the output dimension is set to 1 by default. Considering the complexity of the channel environment in real-world conditions and the fact that the model input has multiple features, setting a higher hidden_size can provide richer feature representations and improve the generalization ability of the multi-feature GRU model. Based on 2), the designed update gate is:

[0066] z t =σ(W z ·[h t-1 ,x t,f,i ])

[0067] The designed reset door is:

[0068] r t =σ(W r ·[h t-1 ,x t,f,i ])

[0069] Based on 3), calculate the hidden candidate states:

[0070]

[0071] Based on 4), calculate the final candidate state:

[0072]

[0073] The multi-feature GRU model designed using steps 1)-4) reduces one gate compared to the original LSTM model, resulting in a simpler structure, lower complexity, and significantly improved training speed when processing large-scale pilot data. Due to its fewer parameters, the GRU model also reduces memory consumption in the simulation environment, improving simulation efficiency and scalability.

[0074] S300: When performing online channel estimation, the Reptile algorithm is used to learn the best estimate of the current model parameters online, realize the online fine-tuning of the GRU model parameters, obtain the GRU network model with the best parameters, and obtain the channel estimates at other locations except for the pilots.

[0075] Preferably, an algorithm for online fine-tuning of GRU network model parameters is designed, and its specific implementation scheme is as follows:

[0076] First, a base learner is set up to quickly learn from M pilot signals received online, adapting to the current pilot signal through a small number of gradient updates. Next, a meta-learner is set up to update the parameters of the current model by collecting parameter update information from the base learners on M tasks. Through the online learning mechanism of the Reptile algorithm, real-time adjustment and optimization of the GRU network model parameters are achieved, thereby improving the model's ability to quickly adapt to new tasks. The Reptile algorithm, by rapidly training on M tasks based on the current channel and updating the offline-trained GRU model parameters in real time, enables the GRU model to quickly achieve good performance with a small number of gradient updates when the current channel state changes rapidly.

[0077] For a specific example, during online updates: First, when the slot of the current pilot is detected to be greater than a certain threshold, the pilot data of the previous M steps are collected as the support set, denoted as T. M =[T1,T2,…T m ] TThen the parameter estimates obtained by the base learner through M iterations can be calculated as follows:

[0078]

[0079] θ represents the parameters to be optimized in the current GRU model. Then, the initial values ​​of the parameters in the meta-learner are updated:

[0080]

[0081] After each training task is completed, the meta-learner will... i The base learner is re-inputted for optimization. After training on M sets of data in the support set, a pre-trained Reptile model is obtained. The network parameters output by this model include channel variation characteristics up to the (M+1)th pilot, exhibiting good real-time performance. Furthermore, for each new set of pilot data, the parameter estimates of the base learner and meta-learner are updated in real-time using the above two-step algorithm. When the mean square error (MSE) of the existing GRU network model exceeds a preset threshold, indicating a decline in model performance and inability to meet predetermined accuracy requirements, the system automatically triggers the Reptile algorithm. At this time, the Reptile algorithm uses its optimized network parameters to fine-tune the GRU network model online, aiming to maintain network stability while effectively reducing the bit error rate and enhancing the network's real-time performance. Compared to the original MAML meta-learner, Reptile simplifies the parameter update process through a first-order optimization method, reduces dependence on second-order derivatives, resulting in a simpler structure, more efficient training, reduced hardware requirements for the simulation environment, and easier implementation and debugging. Experiments show that Reptile and MAML achieve almost the same MSE, indicating that the two algorithms achieve nearly identical results.

[0082] Example 2:

[0083] The computer-readable storage medium of this embodiment stores a computer program that, when executed by a processor, implements the steps in a channel estimation simulation method based on a multi-feature input GRU according to Embodiment 1.

[0084] The computer-readable storage medium in this embodiment can be an internal storage unit of the terminal, such as the terminal's hard disk or memory; the computer-readable storage medium in this embodiment can also be an external storage device of the terminal, such as a plug-in hard disk, smart memory card, secure digital card, flash memory card, etc. equipped on the terminal; furthermore, the computer-readable storage medium can include both the terminal's internal storage unit and external storage devices.

[0085] The computer-readable storage medium of this embodiment is used to store computer programs and other programs and data required by the terminal. The computer-readable storage medium can also be used to temporarily store data that has been output or will be output.

[0086] Example 3:

[0087] The computer device of this embodiment includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the channel estimation simulation method based on a multi-feature input GRU of Embodiment 1.

[0088] In this embodiment, the processor can be a central processing unit, or other general-purpose processors, digital signal processors, application-specific integrated circuits, off-the-shelf programmable gate arrays or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc. The memory can include read-only memory and random access memory, and provides instructions and data to the processor. A portion of the memory can also include non-volatile random access memory. For example, the memory can also store device type information.

[0089] Those skilled in the art will understand that the content disclosed in the embodiments can be provided as a method, system, or computer program product. Therefore, this solution can take the form of a hardware embodiment, a software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this solution 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 and optical storage) containing computer-usable program code.

[0090] This solution is described with reference to flowchart illustrations and / or schematic diagrams of methods and computer program products according to embodiments of this solution. It should be understood that each block of the flowchart illustrations and / or schematic diagrams, and combinations of blocks of the flowchart illustrations and / or schematic diagrams, can be implemented by computer program instructions; these computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing device, generate instructions for implementing the flowchart illustrations and / or block combinations. Figure 1 One or more processes and / or methods are illustrated. Figure 1 A device that provides the functions specified in one or more boxes.

[0091] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or methods are illustrated. Figure 1 The function specified in one or more boxes.

[0092] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or methods are illustrated. Figure 1 The steps of the function specified in one or more boxes.

[0093] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0094] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A channel estimation simulation method based on a multi-feature input GRU, characterized in that: The method includes: S100. Design a method for organizing training data at the pilot, shielding the influence of received pilot data on the channel estimation value, and reducing the factors affecting the channel estimation result to time, frequency and the antenna where the pilot is located; S200. Design a GRU model based on multi-feature input to simultaneously process three key features that affect the channel estimation value, and more comprehensively capture the dynamic characteristics of channel changes. S300. When performing online channel estimation, the Reptile algorithm is used to learn the best estimate of the current model parameters online, and to obtain the GRU network model with the best parameters and the channel estimates at other locations except for the pilots. One method for organizing training data at pilot frequencies in S100 includes: The PUSCH channel was selected as the simulation channel, and its pilot signal was DMRS. The signals at the pilot points are labeled to obtain the time, frequency, antenna index at the pilot point, and channel estimate of the pilot signals, which are then used as input data for model training. The channel estimate, expressed as a complex number, is decomposed to obtain the real and imaginary parts of the channel estimate at the pilot, which are then used to form two sets of training data. S200 includes: When applying the trained model online for channel estimation, the original pilot data is split to obtain real part training data and imaginary part training data. The real part training data and the imaginary part training data are respectively input into the multi-feature GRU model for training, to obtain the real part GRU model and the imaginary part GRU model. During online verification, the real part channel estimate and the imaginary part channel estimate are obtained by inputting the current pilot time, frequency and antenna index information into the real part GRU model and the imaginary part GRU model, respectively. The obtained channel estimates of the real and imaginary parts are concatenated to obtain the complete channel response value. The multi-feature input GRU model in S200 includes: The model's inputs are the time slot, frequency, antenna index, and channel estimate at the pilot location; the channel estimate is determined by the time slot, frequency, and antenna index of the pilot. The influence of historical pilot information on current pilot information can be controlled by setting update and reset gates; Calculate candidate hidden states, providing a proposal for a new hidden state for the current time step; The candidate hidden state and the hidden state of the previous time step are weighted and fused to calculate the final hidden state. By completing a specified number of training tasks, an offline pre-trained GRU model with better parameters is obtained. In S300, the optimal estimate of the current model parameters is learned online using the Reptile algorithm, including: Set up a base learner to quickly learn from M sets of pilot data up to the current time step and adapt to the current pilot signal through a small number of gradient updates; at the same time, set up a meta learner to update the global parameter initialization by collecting parameter update information from the base learners on multiple tasks. When the channel estimation capability of the existing GRU model parameters is poor, the parameters of the GRU model are updated online by activating the output of the Reptile algorithm.

2. A computer-readable storage medium having a computer program stored thereon, characterized in that: When executed by the processor, the program implements the steps in the channel estimation simulation method based on a multi-feature input GRU as described in claim 1.

3. A computer device, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the program, it implements the steps in the channel estimation simulation method based on a multi-feature input GRU as described in claim 1.

Images