Optimization pilot implementation method based on image super-resolution in non-stationary channel scene

Abstract

The application discloses an optimization pilot implementation method based on image super-resolution in a non-stationary channel scene, wherein in an offline stage, RB data of the non-stationary channel is generated through a DMRS densest pilot pattern and a geometry-based random model, a channel frequency domain response of pilot dimension is generated as a training set through a random two-dimensional pilot muting mechanism and a random pilot RE / data RE power ratio setting, a pilot dimension super-resolution convolutional neural network is trained, a high-resolution channel frequency domain response is obtained according to the trained Pilot-SRCNN, and after interpolation processing, the high-resolution channel frequency domain response is further used as a training set for training a resource block super-resolution convolutional neural network; in an online stage, according to to-be-measured RB data, all selectable two-dimensional pilot muting mechanisms and pilot RE / data RE power ratios are selected, and the to-be-measured RB data is obtained through the trained resource block super-resolution convolutional neural network to obtain a mean square error of the to-be-measured RB data, and when the mean square error satisfies a maximum energy efficiency standard, the pilot is set according to the RB data. The application can avoid resource waste and significantly improve system energy efficiency by optimizing system energy efficiency (EE) of different non-stationary channel scenes.

Description

Technical Field

[0001] This invention relates to a technology in the field of wireless communication, specifically an optimized pilot implementation method based on image super-resolution in non-stationary channel scenarios. Background Technology

[0002] One important function of pilots is to serve as known signals at both the transmitting and receiving ends for channel estimation to assess channel fading. In fourth-generation (4G) Long Term Evolution (LTE) systems, the diamond pilot pattern proved to be optimal, but it is no longer applicable in fifth-generation (5G) New Radio (NR) systems. Existing adaptive pilot design methods can be broadly categorized into two types. One type pre-defines a basic pilot configuration and adjusts the pilot spacing in the time domain axis based on real-time Doppler spread of the channel, and adjusts the pilot spacing in the frequency domain axis based on real-time delay spread of the channel. The Doppler (delay spread) spread is negatively correlated with the pilot spacing in both the time and frequency domain axes. The other type uses a pre-defined densest pilot configuration and optimizes one-dimensional pilot shutdown selection to achieve adaptive pilot design requirements. In 5G NR systems, one type of pilot demodulation reference signal (DMRS) has four different pattern formats for different channels: slowly time-varying channels, frequency-selective channels, fast time-varying channels, and dual-selective channels. However, with the introduction of the concept of non-stationary channels in Beyond 5G (B5G) systems and the increasing attention to green communication, the demand for autonomy in real-time pilot design of channels has further increased.

[0003] A search of existing technologies revealed that Chinese patent document CN114944895A, published on August 26, 2022, discloses an optimized pilot implementation method for non-stationary channel scenarios of unmanned aerial vehicles (UAVs). The method estimates the channel time correlation function of the current resource block (RB) based on the wireless protocol configuration and UAV communication operation data. Then, it selects the shutdown of pilot sequences based on different symbol positions and performs pilot and data power matching based on the principle of maximizing RB energy efficiency. Finally, it designs pilots with maximized energy efficiency based on the channel time correlation function and power matching. However, the existing pilot sequence only shuts down in the time domain, without exploring shutdown in the frequency domain. In fact, joint time-frequency domain silence selection is more suitable for non-stationary channels. On the other hand, the energy efficiency improvement has not been maximized. Firstly, the shutdown strategy of the pilot sequence in the time domain of the existing technology has the side effect of shutting down the data resource elements (RE) in the shut-down pilot sequence symbols, which reduces the system throughput. Secondly, because the pilot is only shut down in the time domain, while joint time-frequency domain silence can reduce the system's energy consumption to a minimum, there is room for further improvement in the system's energy efficiency. Summary of the Invention

[0004] To address the aforementioned shortcomings of existing technologies, this invention proposes an optimized pilot implementation method based on image super-resolution for non-stationary channel scenarios. In such scenarios, based on the densest pilot pattern configuration and a two-dimensional pilot mute mechanism, a channel estimation process nested based on image super-resolution compensates for the increased channel estimation error caused by the mute mechanism. This invention optimizes the system energy efficiency (EE) of resource blocks (RBs) for different non-stationary channel scenarios, avoiding resource waste and significantly improving system energy efficiency.

[0005] This invention is achieved through the following technical solution:

[0006] This invention relates to a pilot implementation method based on image super-resolution two-dimensional pilot mute mechanism. In the offline stage, non-stationary channel RB data is generated using the DMRS densest pilot pattern and a geometry-based random model (GBSM). The channel frequency domain response of the pilot dimension is generated under the random two-dimensional pilot mute mechanism and the power ratio setting of random pilot RE / data RE as a training set to train the pilot dimension super-resolution convolutional neural network (Pilot-SRCNN). The high-resolution channel frequency domain response (CFR) is obtained from the trained Pilot-SRCNN, and after interpolation, it is used as a training set to further train the resource block super-resolution convolutional neural network (RB-SRCNN). In the online stage, all possible two-dimensional pilot mute mechanisms and the power ratio of pilot RE / data RE are selected based on the RB data to be tested. The mean square error of the RB data to be tested is obtained through the trained resource block super-resolution convolutional neural network. When it meets the maximum energy efficiency standard, the pilot is set according to the RB data.

[0007] When training the super-resolution convolutional neural network for the pilot dimension, the loss function of Pilot-SRCNN is set to the mean square error of the minimum pilot dimension.

[0008] The training resource block super-resolution convolutional neural network described above sets the loss function of RB-SRCNN to minimize the mean square error of the data in the RB dimension.

[0009] Both the Pilot-SRCNN network and the RB-SRCNN network include: an input convolutional layer, an intermediate convolutional layer, and an output convolutional layer.

[0010] This invention relates to a system for implementing the above-mentioned method, comprising: a data generation unit, a pilot design unit, a channel estimation unit, and an energy efficiency optimization unit, wherein: the data generation unit performs non-stationary channel modeling based on the GBSM model and obtains channel data of the non-stationary channel under the GBSM model; the pilot design unit generates two-dimensional pilot silence selection and power ratio of pilot RE / data RE based on a two-dimensional pilot silence mechanism model, and obtains channel data after the two-dimensional pilot silence mechanism; the channel estimation unit performs non-stationary channel recovery processing on the channel data obtained by the pilot design unit based on an image super-resolution-based channel estimation framework, and obtains the mean square error (MSE) of the data in the channel; and the energy efficiency optimization unit implements corresponding optimal pilot silence decisions for channel energy efficiency under different conditions based on the two-dimensional silence mechanism.

[0011] Technical effect

[0012] This invention optimizes pilot signals for non-stationary channels. Through a two-dimensional pilot mute mechanism and a channel estimation framework based on image super-resolution, it significantly improves the accuracy of channel estimation. Compared with existing fixed pilot configurations, this invention also offers a certain performance improvement in RB energy efficiency gain. Attached Figure Description

[0013] Figure 1 This is a flowchart of the present invention;

[0014] Figure 2 This is a schematic diagram of pilot implementation for two-dimensional pilot-based mute.

[0015] Figure 3 This is a schematic diagram illustrating the changes in the RB dimension of the example.

[0016] Figure 4 This is a schematic diagram illustrating the convergence of the two SRCNN networks in the embodiment;

[0017] Figure 5 This is a schematic diagram comparing the system energy efficiency under different communication environments in the embodiments;

[0018] Figure 6 This is a schematic diagram illustrating the system energy efficiency gain under different communication environments in the embodiments. Detailed Implementation

[0019] like Figure 1As shown, this embodiment relates to a pilot implementation method for a two-dimensional pilot mute mechanism based on image super-resolution. In the offline stage, non-stationary channel RB data is generated using the DMRS densest pilot pattern and GBSM. The channel frequency domain response of the pilot dimension is generated under the random two-dimensional pilot mute mechanism and the power ratio setting of random pilot RE / data RE, and used as a training set to train Pilot-SRCNN. Based on the high-resolution channel frequency domain response obtained from the trained Pilot-SRCNN, it is interpolated and used as a further training set to train RB-SRCNN. In the online stage, all available two-dimensional pilot mute mechanisms and the power ratio of pilot RE / data RE are selected based on the RB data to be tested. The average mean square error of the RB data to be tested is obtained through the trained resource block super-resolution convolutional neural network. When it meets the maximum energy efficiency standard, the pilot is set according to the RB data, specifically including:

[0020] Step 1) Generate non-stationary channel RB data using the DMRS densest pilot pattern and GBSM. Then, generate the channel frequency domain response in the pilot dimension as a training set using a random two-dimensional pilot mute mechanism and a random pilot RE / data RE power ratio setting.

[0021] like Figure 2 As shown, in this embodiment, resource block RB includes time-domain axis N. sym The 14 symbols and frequency domain axis N sub The resource element (RE) consists of 1 symbol and 1 subcarrier, with the first two symbols on the time axis used to transmit the control signal N. ctrl This refers to the Physical Downlink Control Channel (PDDCH); the densest DMRS configuration in the downlink, i.e., the time-domain interval Δ p t represents 3 symbols, and the frequency domain axis spacing Δ p f is two subcarriers, therefore the number of pilots on the time-domain axis is... The number of pilot axes is 4. The number of pilot REs D in the densest pilot configuration within the resource block is 6. p The number of data REs (Resource Extraction Objects) within a resource block is 24. base It is 120.

[0022] The GBSM mentioned is Where: q is the q-th receiving antenna, and P is the p-th transmitting antenna. The complex channel gain is the line-of-sight distance. Let K(t) be the non-line-of-sight complex channel gain, K(t) be the Rice factor, and τ be the non-line-of-sight complex channel gain. LOS (t) represents the time delay of the line-of-sight distance, N(t) represents the time-varying number of clusters, and M... n (t) represents the number of rays in the nth cluster, τ n(t) represents the delay of the nth cluster. The relative time delay is the m-th ray of the n-th cluster.

[0023] The channel frequency domain response in the pilot dimension is obtained by modeling the energy efficiency of a single RB based on a two-dimensional pilot silence mechanism, maximizing the energy efficiency of the RB, and setting constraints on the model. Specifically, it includes:

[0024] ① Energy efficiency modeling of a single RB η s =C s / E s Where: RB's channel throughput The energy consumption of the RB after the two-dimensional pilot mute mechanism is The number of pilots that are active within a resource block after the two-dimensional pilot mute mechanism. For the power of the pilot RE, The power of data RE, and the instantaneous spectral efficiency S(D) of RB. p ) = N base / N sub (N sym -N ctrl ), RB's post-confidence interference-to-noise ratio The average power of the noise. Let be the average power of the subcarrier interference in the system, which satisfies the following condition f d =vf c / c represents the maximum Doppler frequency shift, f c Let v be the subcarrier frequency, c be the system velocity, Δf be the subcarrier spacing, and δ be the system velocity. d The average MSE of the data RE within the resource block; the set of indices of the symbols in the RB that can be pilot muted on the time-domain axis. The set of indices of symbols for which pilot mute operations can be performed on the frequency domain axis of RB.

[0025] ② Maximize the energy efficiency of RB and set constraints for the model, i.e. And ρ < ρ max That is, while ensuring that the power of the pilot RE is greater than the power of the data RE, ρ max Determined by the peak-to-average power ratio (PAPR), where: P = {{p i p j}} is the identifier function for two-dimensional pilot silence. For the power ratio of pilot RE and data RE, The rule that the first column of pilot symbols defined in the 3GPP protocol should not be muted.

[0026] Step 2) Construct and train the pilot-dimensional super-resolution convolutional neural network (Pilot-SRCNN) using the training set obtained in Step 1. Based on the channel frequency domain response of the high-resolution pilot-dimensional training set obtained from the trained Pilot-SRCNN, interpolate it to generate a training set for further training of the resource block super-resolution convolutional neural network (RB-SRCNN), specifically including:

[0027] Step 2.1) Construct pilot-SRCNN, where: the input convolutional layer uses 64 3×3 convolutional kernels; the intermediate convolutional layers use 32 1×1 convolutional kernels, and both layers use the Rectified Linear Array (ReLU) as the activation function; the output convolutional layer uses 16 5×5 convolutional kernels, and the loss function is... in: The training dataset is for the pilot dimension. H represents the number of training datasets in the pilot dimension. p For the true CFR in the pilot dimension, ||*|| F It is the F-norm. The training set obtained through step 1 is the low-precision pilot dimension CFR obtained through the two-dimensional pilot mute mechanism. High-precision pilot-dimensional CFR for the output of pilot-SRCNN.

[0028] Step 2.2) Perform frequency domain DFT interpolation and time domain Gaussian interpolation on the high-precision pilot-dimensional channel frequency domain response obtained in Step 2.1 to obtain the training set. Specifically, this involves: high-precision pilot-dimensional channel time domain response. Perform DFT interpolation in the frequency domain: [The text abruptly ends here, likely due to an incomplete translation or a Pad with zeros until the length is N sub Right now Then perform the DFT operation. Then perform Gaussian interpolation. Obtain low precision of RB dimension Where: i p j is the sign index of the pilot on the time-domain axis. p This is the subcarrier index of the pilot on the frequency domain axis. These are the Gaussian interpolation coefficients.

[0029] Step 2.3) Construct RB-SRCNN, where: the input convolutional layer uses 64 9×9 convolutional kernels; the intermediate convolutional layers use 32 5×5 convolutional kernels, and both of these layers use the Rectified Linear Array (ReLU) as the activation function; the output convolutional layer uses 3 5×5 convolutional kernels, and its loss function is... in: For the training dataset of RB dimension, Ω represents the number of training datasets in the RB dimension, H represents the true CFR in the RB dimension, and Ω represents the true CFR in the RB dimension. ref The index set of pilot time-frequency locations under the densest DMRS configuration. The training set obtained in step 2.2 High-precision pilot dimension CFR for RB-SRCNN output.

[0030] Step 3) In the online phase, iterate through each two-dimensional pilot mute mechanism selection and pilot RE / data RE power ratio of the RB channel data under test, and obtain the mean square error of data RE in the RB for each case using the resource block super-resolution convolutional neural network trained in Step 2. Based on the obtained δ d The RB efficiency can then be calculated for each case.

[0031] Based on specific practical experiments, this embodiment detects different Rice factors. Based on the speed θ, pilot design optimizes RB energy efficiency under a two-dimensional pilot mute mechanism. The simulation environment designs three different sets of channel characteristic Rice factors. The channel communication environment with speed θ is represented by the Rice factor of the channel characteristics. -InfdB and a velocity θ of 3m / s, the Rice factor of the channel characteristics Rice factor of channel characteristics at 15dB and speed θ of 3m / s The values ​​are -InfdB and the velocity θ is 250m / s, as shown in Table 1.

[0032] Detection environment subcarrier spacing Δ f =30KHz, carrier frequency f c =5.9GHz, Number of transmitting antennas N T =1. Number of receiving antennas N R =1. Power configuration of pilot RE and data RE ρ = [0:1:9] dB, power of data RE noise power The key parameters of SRCNN are as follows: Number of training set RBs N train =5000, the number of validation sets (RBs) N valid =250, learning rate l c=0.001 and batch size B size =8.

[0033] Table 1 Parameter settings under different communication environments

[0034]

[0035] like Figure 4 As shown, the normalized mean square error (NMSE) of the pilot-SRCNN network converges to around 0.1 as the signal-to-noise ratio (SNR) increases. This is because the two-dimensional pilot silence mechanism leads to an increase in the channel recovery error in the pilot dimension. However, since the system only focuses on the channel estimation error in the RB dimension, verifying the convergence of the pilot-SRCNN network is sufficient. On the other hand, the NMSE of the RB-SRCNN network converges to around 0.001 as the SNR increases, reaching the acceptable error range for channel estimation. Based on this simulation performance analysis, the usability of the image super-resolution-based channel estimation framework is demonstrated.

[0036] As shown in Table 1 and Figure 5 As shown, the Rice factor in the channel characteristics are respectively -InfdB and a velocity θ of 3m / s, the Rice factor of the channel characteristics Rice factor of channel characteristics at 15dB and speed θ of 3m / s The simulation analysis examines the optimal energy efficiency performance of the system under three conditions: -InfdB, a velocity θ of 250m / s, and the Rice factor of the channel characteristics. The system energy efficiency at -InfdB and a velocity θ of 3m / s is 280bps / Hz / J, and the channel characteristics are Rice factor. The system energy efficiency at 15 dB and speed θ of 3 m / s is 160 bps / Hz / J, and the channel characteristics are Rice factor. With a system efficiency of -InfdB and a speed θ of 250m / s, the system efficiency is 120bps / Hz / J. Therefore, it can be concluded that the present invention can achieve the optimal RB efficiency under different communication environments.

[0037] like Figure 6 As shown, the RB energy efficiency gain is compared under three communication environments: channel characteristics and Rice factor. The channel characteristics are -Inf dB and a velocity θ of 3 m / s, and the Rice factor. Rice factor of channel characteristics at 15dB and speed θ of 250m / s The system's energy efficiency gain is around 8dB under all three communication environments, with a value of -InfdB and a speed θ of 250m / s.

[0038] In summary, this invention proposes a two-dimensional pilot mute mechanism based on RB energy efficiency optimization for non-stationary channels. In order to ensure the accuracy of channel estimation under the two-dimensional pilot mute mechanism, a channel estimation framework based on image super-resolution is proposed. Both of these can bring energy efficiency gains to the communication system.

[0039] The above-described specific implementations can be partially adjusted by those skilled in the art in different ways without departing from the principles and purpose of the present invention. The scope of protection of the present invention is defined by the claims and is not limited to the above-described specific implementations. All implementation schemes within the scope of the claims are bound by the present invention.

Claims

1. A method for optimizing pilot signals based on image super-resolution in non-stationary channel scenarios, characterized in that, In the offline phase, non-stationary channel RB data is generated using the DMRS densest pilot pattern and a geometry-based stochastic model (GBSM). The channel frequency response (CFR) in the pilot dimension is generated using a random two-dimensional pilot mute mechanism and a random pilot RE / data RE power ratio setting. This CFR is then used as a training set to train a pilot-dimensional super-resolution convolutional neural network (Pilot-SRCNN). The trained Pilot-SRCNN yields a high-resolution channel frequency response (CFR), which is then interpolated and used as a further training set to train a resource block super-resolution convolutional neural network (RB-SRCNN). In the online phase, all available two-dimensional pilot mute mechanisms and pilot RE / data RE power ratios are selected based on the RB data to be tested. The mean square error of the RB data is obtained using the trained resource block super-resolution convolutional neural network. When the maximum energy efficiency standard is met, pilots are set based on this RB data. Both the Pilot-SRCNN network and the RB-SRCNN network include: an input convolutional layer, an intermediate convolutional layer, and an output convolutional layer; In the Pilot-SRCNN network: the input convolutional layer uses 64 3×3 convolutional kernels; the intermediate convolutional layers use 32 1×1 convolutional kernels, and both layers use the Rectified Linear Array (ReLU) activation function; the output convolutional layer uses 16 5×5 convolutional kernels, and the loss function is... ,in: The training dataset is for the pilot dimension. The number of training datasets in the pilot dimension. For the true CFR in the pilot dimension, It is the F-norm. The training set consists of low-precision pilot-dimensional CFRs obtained through a two-dimensional pilot mute mechanism. High-precision pilot-dimensional CFR for pilot-SRCNN output; In the RB-SRCNN network: the input convolutional layer uses 64 9×9 convolutional kernels; the intermediate convolutional layers use 32 5×5 convolutional kernels, and both of these layers use the Rectified Linear Array (ReLU) as the activation function; the output convolutional layer uses 3 5×5 convolutional kernels, and its loss function is... ,in: For the training dataset of RB dimension, The number of training datasets in the RB dimension. For the true CFR in the RB dimension, The index set of pilot time-frequency locations under the densest DMRS configuration. For the training set, High-precision pilot dimension CFR for RB-SRCNN output.

2. The optimized pilot implementation method based on image super-resolution in non-stationary channel scenarios according to claim 1, characterized in that, When training the super-resolution convolutional neural network for the pilot dimension, the loss function of Pilot-SRCNN is set to the mean square error of the minimum pilot dimension. The training resource block super-resolution convolutional neural network described above sets the loss function of RB-SRCNN to minimize the mean square error of the data in the RB dimension.

3. The optimized pilot implementation method based on image super-resolution in non-stationary channel scenarios according to claim 1, characterized in that, specifically... include: Step 1) Generate non-stationary channel RB data using the DMRS densest pilot pattern and GBSM. Generate channel frequency domain response in the pilot dimension as a training set using a random two-dimensional pilot mute mechanism and a random pilot RE / data RE power ratio setting. Step 2) Construct and train the pilot-dimensional super-resolution convolutional neural network using the training set obtained in Step 1. Based on the channel frequency domain response of the high-resolution pilot-dimensional training set obtained after training, interpolate it to generate a training set for further training of the resource block super-resolution convolutional neural network. Specifically, this includes: Step 2.1) Construct pilot-SRCNN, where: the input convolutional layer uses 64 3×3 convolutional kernels; the intermediate convolutional layers use 32 1×1 convolutional kernels, and both layers use the Rectified Linear Array (ReLU) as the activation function; the output convolutional layer uses 16 5×5 convolutional kernels, and the loss function is... ,in: The training dataset is for the pilot dimension. The number of training datasets in the pilot dimension. For the true CFR in the pilot dimension, It is the F-norm. The training set obtained through step 1 is the low-precision pilot dimension CFR obtained through the two-dimensional pilot mute mechanism. High-precision pilot-dimensional CFR for pilot-SRCNN output; Step 2.2) Perform frequency domain DFT interpolation and time domain Gaussian interpolation on the high-precision pilot-dimensional channel frequency domain response obtained in Step 2.1 to obtain the training set. Specifically, this involves: high-precision pilot-dimensional channel time domain response. Perform DFT interpolation in the frequency domain: [The text abruptly ends here, likely due to an incomplete sentence or a formatting error.] Pad with zeros until the length is Right now Then perform DFT operations. Then perform Gaussian interpolation. This yields a low-precision CFR in the RB dimension. ,in: This is the sign index of the pilot on the time-domain axis. This is the subcarrier index of the pilot on the frequency domain axis. These are the Gaussian interpolation coefficients; Step 2.3) Construct RB-SRCNN, where: the input convolutional layer uses 64 9×9 convolutional kernels; the intermediate convolutional layers use 32 5×5 convolutional kernels, and both of these layers use the Rectified Linear Array (ReLU) as the activation function; the output convolutional layer uses 3 5×5 convolutional kernels, and its loss function is... ,in: For the training dataset of RB dimension, The number of training datasets in the RB dimension. For the true CFR in the RB dimension, The index set of pilot time-frequency locations under the densest DMRS configuration. The training set obtained in step 2.2 For high-precision pilot dimension CFR output of RB-SRCNN; Step 3) In the online phase, iterate through each two-dimensional pilot mute mechanism selection and pilot RE / data RE power ratio of the RB channel data under test, and obtain the mean square error (MSE) of the data RE in the RB for each case through the resource block super-resolution convolutional neural network trained in Step 2. Based on the obtained The RB efficiency can then be calculated for each case.

4. The optimized pilot implementation method based on image super-resolution in non-stationary channel scenarios according to claim 3, characterized in that, The GBSM mentioned above is: ,in: For the first One receiving antenna, For the first One transmitting antenna, The complex channel gain is the line-of-sight distance. For non-line-of-sight complex channel gain, Rice factor, The time delay is the line-of-sight distance. The number of time-varying clusters, The number of rays in the nth cluster. The latency of the nth cluster, The relative time delay is the m-th ray of the n-th cluster.

5. The optimized pilot implementation method based on image super-resolution in non-stationary channel scenarios according to claim 3, characterized in that, The channel frequency domain response in the pilot dimension is obtained by modeling the energy efficiency of a single RB based on a two-dimensional pilot silence mechanism, maximizing the energy efficiency of the RB, and setting constraints on the model. Specifically, it includes: ① Energy efficiency modeling of a single RB Where: RB's channel throughput The energy consumption of the RB after the two-dimensional pilot mute mechanism is The number of pilots that are active within a resource block after the two-dimensional pilot mute mechanism. , For the power of the pilot RE, For the power of data RE, The post-confidence interference-to-noise ratio (SINR) of RB. , The average power of the noise. Let be the average power of the subcarrier interference in the system, which satisfies the following condition For the maximum Doppler frequency shift, For subcarrier frequency, Let c be the system velocity and c be the speed of light. For subcarrier spacing, The average MSE of the data RE within the resource block; the set of indices of the symbols in the RB that can be pilot muted on the time-domain axis. , The set of indices of symbols for which pilot mute operations can be performed on the frequency domain axis of RB. , ; ② Maximize the energy efficiency of RB and set constraints for the model, i.e. ,and That is, ensuring that the power of the pilot RE is greater than the power of the data RE, Determined by the peak-to-average power ratio (PAPR), where: This is the identifier function for two-dimensional pilot silence. For the power ratio of pilot RE and data RE, The rule that the first column of pilot symbols defined in the 3GPP protocol should not be muted.

6. A system for implementing the optimized pilot implementation method based on image super-resolution in any of the non-stationary channel scenarios described in claims 1-5, characterized in that, include: The system comprises a data generation unit, a pilot design unit, a channel estimation unit, and an energy efficiency optimization unit. Specifically: the data generation unit models a non-stationary channel based on the GBSM model and obtains channel data for the non-stationary channel under the GBSM model; the pilot design unit generates two-dimensional pilot silence selection and a power ratio of pilot RE / data RE based on a two-dimensional pilot silence mechanism model, obtaining channel data after the two-dimensional pilot silence mechanism; the channel estimation unit performs non-stationary channel recovery processing on the channel data obtained by the pilot design unit based on an image super-resolution-based channel estimation framework, obtaining the average mean square error of the data in the channel; and the energy efficiency optimization unit implements the corresponding optimal pilot silence decision for channel energy efficiency under different conditions based on the two-dimensional silence mechanism.

Images