Channel estimation method for 5G-OFDM system based on deep learning

Abstract

The invention discloses a channel estimation method for a 5G-OFDM system based on deep learning, and the method comprises the steps: building an OFDM orthogonal frequency division multiplexing system, and forming a low-resolution channel matrix; establishing a deep learning neural network channel estimation model; the low-resolution channel matrix is input into a deep learning neural network channel estimation model to be processed, the deep learning neural network channel estimation model adopts an interpolation-free end-to-end learning framework, and feature extraction is directly conducted on channel response data which are not interpolated; and outputting a complete channel frequency response matrix through the deep learning neural network channel estimation model to realize channel estimation. Under a 3GPP TDL-A multipath channel, the model provided by the invention has excellent performance in a signal-to-noise ratio range of 1-30dB: compared with an existing deep learning model, the estimation precision has the advantage that the parameter quantity is reduced by 93% under similar precision; the reasoning time is only 12.44 ms, the speed is increased by 25.7% compared with that of an SRCNN, the model calculation amount is only 38.24 MMac, and the method is easy to deploy in a mobile terminal with limited resources.

Description

Technical Field

[0001] This invention relates to the fields of deep learning and communication technology, and in particular to a channel estimation method based on deep learning for 5G-OFDM systems. Background Technology

[0002] In recent years, 5G communication technology, with its significant advantages of high speed, low latency, and large capacity, has become a key force driving social development and technological progress. Orthogonal Frequency Division Multiplexing (OFDM), as one of the core technologies of 5G communication, effectively improves spectral efficiency and multipath fading resistance by dividing a high-speed data stream into multiple low-speed sub-data streams and transmitting them simultaneously on multiple orthogonal subcarriers. In the complex communication environment of 5G OFDM, accurate channel estimation is crucial, as it is the foundation for reliable data transmission and efficient communication. Traditional channel estimation methods, such as those based on Minimum Mean Square Error (MMSE) and Least Square (LS), can achieve certain results under relatively simple channel conditions. However, with the increasing complexity of communication environments, the impact of factors such as multipath fading and noise interference on the accuracy of channel estimation has become increasingly significant. The limitations of these traditional methods have gradually become apparent, making it difficult to meet the stringent requirements of 5G communication for high-precision channel estimation.

[0003] In recent years, with the rapid development of deep learning technology, its application in the physical layer of wireless communication has become a research hotspot, including channel decoding, signal detection, channel equalization, and channel estimation. Significant progress has been made, particularly in the field of channel estimation. Existing research can be mainly divided into the following three categories: Method 1: End-to-end deep learning methods: A novel end-to-end model based on a 5-layer fully connected network was proposed, automatically learning the joint probability distribution of the wireless channel and the modulated signal through a joint training strategy. However, the fully connected structure leads to high computational complexity, limiting its application in real-time systems. Method 2: Combination of compressed sensing and deep learning methods: HE et al. combined compressed sensing theory with deep learning, proposing a beamspace sparsity method that effectively solves the channel estimation problem of millimeter-wave massive MIMO (Multiple-Input Multiple-Output) systems, providing a new approach for high-dimensional channel processing. Method 3: Deep learning methods based on image super-resolution: These methods treat the channel matrix as a two-dimensional image for processing: The proposed ChannelNet uses a cascaded structure of SRCNN and DNCNN, modeling channel estimation as an image super-resolution reconstruction problem. This literature review has explored several approaches to channel estimation. First, it describes the construction of an image restoration network based on convolutional neural networks, which improves network performance by introducing residual connections. Second, it presents a complex generative adversarial network model that fully utilizes the complex characteristics of signals to achieve more accurate channel estimation. Third, it describes the innovative introduction of a dynamic attention module into the proposed PDAAB network, which performs well in both SISO (Single-Input Single-Output) and MIMO systems. Fourth, it describes the use of a convolutional blind denoising network, which achieves good results over a wide signal-to-noise ratio range through interactive processing of noise level estimation and non-blind denoising. Analysis of existing literature reveals that deep learning methods based on image super-resolution exhibit unique advantages in channel estimation. These methods transform the channel matrix estimation problem into an image restoration problem, extracting channel features through deep learning networks, effectively balancing computational complexity and estimation accuracy. In particular, schemes incorporating advanced network structures such as attention mechanisms and residual connections significantly improve the accuracy of channel estimation while maintaining reasonable computational complexity.

[0004] However, existing channel estimation neural network models generally suffer from long inference times and a large number of parameters. This not only makes it impossible to quickly and accurately complete channel estimation in real-time communication scenarios, affecting the real-time performance and smoothness of communication; but also makes it difficult for resource-constrained mobile terminals to support such a large model due to limited hardware resources, seriously hindering the widespread application and popularization of 5G communication technology on mobile terminal devices.

[0005] A deep learning-based OFDM channel estimation and signal detection method is disclosed in Chinese patent document N111404849B. This method first generates a sample database of OFDM multipath channel model matrices under complex environments based on existing Non-WSSUS channel models. Then, it constructs a neural network including a channel estimation subnetwork and a signal detection subnetwork, and trains the neural network using the data samples from the multipath channel model matrix. Finally, the trained neural network is applied offline to an OFDM data link system for unmanned aerial vehicles (UAVs) in complex environments to estimate the channel and detect the signal. This invention generates a large set of channel samples that reflects the characteristics of OFDM channels in complex environments, enabling the entire network to effectively reflect the nonlinear characteristics of wireless channels and transmitted signals. However, this deep learning-based OFDM channel estimation and signal detection method is not suitable for online deployment.

[0006] A deep learning-assisted OFDM channel estimation and signal detection method is disclosed in Chinese patent document CN115987722B. This method first generates a dataset using an OFDM system, then constructs a channel estimation model using a self-normalized neural network (SNN), and a signal detection model using a convolutional neural network (CNN) and a bidirectional gated recurrent unit (BiGRU). Next, the dataset is input into the neural network to train the model, and the network is optimized using an Adam optimizer and a mean squared error loss function. Finally, the method is compared with LS, BiLSTM, GRU, and DNN+BiLSTM models. This invention effectively solves the problems of low channel estimation and signal detection accuracy, severe network loss, and poor bit error rate performance in OFDM systems with a limited number of cyclic prefixes and pilots. Simulation results show that, with fewer cyclic prefixes and pilots, this method outperforms LS, BiLSTM, GRU, and DNN+BiLSTM models. However, this deep learning-assisted OFDM channel estimation and signal detection method contains a multi-layer cascaded structure, resulting in a large number of parameters and high computational complexity.

[0007] A method and system for channel estimation in MIMO-OFDM visible light communication using genetic algorithms and deep learning are disclosed in Chinese patent document CN117176517A. This method and system employ a genetic algorithm with stochastic global search optimization capabilities to simulate biological evolution and optimize the weight values ​​of a 6-layer conventional deep neural network (DNN) model. The deep learning method uses a feedforward DNN model, with initial training data from a 4×4 MIMO-OFDM VLC system. The optimized deep network model then learns channel characteristics and recovers the channel information of the online deployment system. The received bit error rate (BER) is calculated using the online deployment system. The BER performance was tested against that of conventional DNNs and existing channel estimation algorithms. Experiments on an indoor 4×4 MIMO-OFDM VLC system verified the advantages of the deep learning method in complex environments with severe distortion and interference in visible light channels, and demonstrated the feasibility of this invention. This invention has strong transmission and reception performance and is an effective channel estimation scheme in the field of indoor visible light communication. However, the method and system for channel estimation of MIMO-OFDM visible light communication using genetic algorithms and deep learning rely on global search optimization using genetic algorithms, involving crossover, mutation and other operations, resulting in large computational overhead and poor real-time performance.

[0008] To address the shortcomings of the existing technologies, it is worthwhile to propose a channel estimation method based on deep learning for 5G-OFDM systems. Summary of the Invention

[0009] The purpose of this invention is to overcome the shortcomings of long inference time and large number of parameters, and to provide a channel estimation method based on deep learning for 5G-OFDM systems, which achieves the technical effect of reducing the number of parameters and computational overhead.

[0010] The objective of this invention is achieved through the following technical solution:

[0011] A channel estimation method based on deep learning for 5G-OFDM systems includes the following steps:

[0012] Step 1: Following the 5G New Radio standard protocol established by the 3rd Generation Partnership Project, build an OFDM orthogonal frequency division multiplexing system, obtain channel response data at the pilot in the 5G-OFDM orthogonal frequency division multiplexing system, form a low-resolution channel matrix, and recover the full subcarrier channel matrix through a neural network;

[0013] Step 2: Establish a deep learning neural network channel estimation model;

[0014] Step 3: Input the low-resolution channel matrix into the deep learning neural network channel estimation model for processing. The deep learning neural network channel estimation model adopts an interpolation-free end-to-end learning framework to directly extract features from the uninterpolated channel response data.

[0015] Step 4: Output the complete channel frequency response matrix through the deep learning neural network channel estimation model to achieve channel estimation. The network design proposed in this invention breaks through the computational paradigm of traditional deep learning channel estimation methods that rely on interpolation preprocessing. Compared with existing solutions (such as CNN or Transformer models based on interpolation reconstruction), this method directly utilizes the sparsity characteristics of the pilot signal. Through the channel-space dynamic mapping mechanism of the PixelShuffleHW module, the low-dimensional features of the pilot signal are directly upsampled to the target channel response dimension, eliminating the explicit interpolation step. End-to-end joint optimization suppresses the frequency domain phase distortion caused by interpolation while significantly reducing computational complexity and saving model inference time.

[0016] Optionally, the OFDM orthogonal frequency division multiplexing system in step one adopts a single-input single-output architecture, focusing the research on the 5G physical downlink shared channel. In the 5G New Radio system, the 5G physical downlink shared channel serves as a key channel for carrying downlink user data and control information. Through flexible subcarrier spacing, resource block configuration, and cyclic prefix configuration parameters, it efficiently supports large-scale spectrum utilization and parallel transmission of multiple users.

[0017] Data is transmitted through resource units, which are the smallest units of time-frequency resources. A specific resource grid implementation is as follows: Figure 2 As shown, it corresponds to one OFDM (Orthogonal Frequency Division Multiplexing) symbol in the time domain and one subcarrier in the frequency domain. In 5G NR, a time slot is a unit of time-domain signal resources, and each time slot contains... One OFDM symbol, typically in NR systems, The value is set to 14 or 12, depending on the subcarrier spacing used. A resource block is the smallest unit of frequency domain resource allocation, and each resource block contains... Each subcarrier, as specified in the 5G NR standard. In a time slot RB, of which There is no specific requirement for the quantity; it can be set flexibly according to actual needs.

[0018] The set parameter value is , That is, one time slot contains 12 symbols and 512 subcarriers. The resource grid is shown in the figure. The gray RE represents the demodulation reference signal, i.e. the pilot signal. The blank square represents the PDSCH RE, which is used to transmit the transmission signal information. In the NR system, the location and number of DMRS can be flexibly configured. In the OFDM (Orthogonal Frequency Division Multiplexing) system model, the time-domain signal after OFDM modulation at the transmitting end is:

[0019]

[0020] in: For the first Modulation symbols on each subcarrier The total number of subcarriers in the system (including guard band) and the cyclic prefix length. Dynamically adjusted according to 3GPP TDL-A channel delay spread;

[0021] After passing through a multipath channel, the received signal It can be represented as:

[0022]

[0023] in: For the first The complex channel gain of each path, It is additive white Gaussian noise. It is a multipath number;

[0024] After removing the CP from the received signal and performing an FFT, the frequency domain received signal is:

[0025]

[0026] in: For the first Channel frequency response of each subcarrier This is frequency domain noise;

[0027] Finally, the LS estimate at the pilot frequency is... Mapped to a low-resolution channel matrix The full subcarrier channel matrix is ​​recovered using a neural network. .

[0028] Optionally, the deep learning neural network channel estimation model in step two includes: a feature extraction layer, a composite residual module, and an upsampling module;

[0029] The feature extraction layer is used to perform shallow feature extraction on the input data. It uses a 3×3 convolutional layer combined with layer normalization.

[0030] The composite residual module is used to deepen feature extraction and adopts a residual connection structure.

[0031] The upsampling module is used to map low-resolution features to a high-resolution space.

[0032] Optionally, the channel response value at the pilot signal of the feature extraction layer is used for shallow feature extraction via the initial convolutional layer, and a 3×3 convolutional kernel is used to achieve spatial context awareness, which is formally represented as follows:

[0033]

[0034] in Layer normalization (LayerNorm) is calculated as follows: the mean of each sample is calculated along the normalization dimension. and variance :

[0035]

[0036]

[0037] Where C is the number of channels, and the mean and variance are calculated independently of the batch and spatial dimensions (H, W), that is, each sample is calculated separately;

[0038] Normalize the input using the mean and variance:

[0039]

[0040] in, It is a local minimum value to prevent the denominator from being zero. The normalized mean is 0 and the variance is 1.

[0041] Introducing learnable scaling parameters and bias parameters To restore the expressive power of the model:

[0042]

[0043] in, and The shape is the same as C;

[0044] The channel response processed by the feature extraction layer is used as a preliminary feature input to the composite residual module, allowing the model to decide whether to retain or adjust the normalized distribution. The data input to the network in this invention is uninterpolated, which eliminates the data preprocessing process and allows feature extraction to be performed directly on the channel response data at the uninterpolated original pilot, further realizing end-to-end channel estimation.

[0045] Optionally, the composite residual module employs residual connections. First, spatial features are encoded using a 5×5 depthwise convolution with the number of groups equal to the number of channels. Then, channel-dimensional layer normalization is used for stable training. Subsequently, the number of channels is expanded by a factor of four using a 1×1 convolution and a GELU activation function is applied, introducing Global Response Normalization (GRN) to enhance inter-channel interaction. Finally, the number of channels is compressed back to the original dimension using a 1×1 convolution. This "depthwise convolution-channel expansion-compression" design draws inspiration from the network structure design of Google's ConvNeXt, maintaining input-output spatial resolution. While improving efficiency, the cascaded structure of layer normalization and double 1×1 convolution balances local feature extraction capability with cross-channel information fusion efficiency. The gradient vanishing problem in deep networks is alleviated by identity mapping, which significantly improves the training stability and representation capability of the model. The residual structure allows gradients to be directly backpropagated through cross-layer short-circuit paths, avoiding gradient decay caused by continuous multiplication in traditional stacked layers. In order to reduce computational cost, the convolution operation in the deep residual module adopts depthwise separable convolution, which can reduce the number of parameters to 1 / C of the standard convolution (C is the number of channels).

[0046] Optionally, the upsampling module decouples the height and width scaling factors ( and ), decompose the channel dimension of the input tensor into The sub-blocks are used to achieve asymmetric upsampling through dynamic recombination of the channel-space dimension;

[0047] Specifically, the input shape is The feature map, after channel splitting, dimension permutation, and spatial merging operations, has its output resolution increased to [value missing]. The upsampling module is an extension and improvement of the classic PixelShuffle module, adapted to the upsampling requirements of asymmetric spatial resolution. The classic PixelShuffle is used in image super-resolution tasks, with the same scaling factor for height H and width W, without taking into account the inconsistent scaling factor of the number of pilots in the OFDM channel matrix in the frequency domain and time domain. This is crucial for flexible pilot configuration in 5G NR.

[0048] Optionally, the channel estimation method has a computational cost of no more than 38.24 MMac and an inference time of no more than 12.44 ms during the inference phase, and is applied to deployments on mobile terminals or IoT devices.

[0049] Positive and beneficial effects: 1. This channel estimation method based on deep learning for 5G-OFDM systems addresses the problem of high computational complexity of existing deep learning models in 5G-OFDM systems. Through an end-to-end learning framework without interpolation preprocessing, the model significantly reduces the number of parameters and computational overhead while ensuring estimation accuracy.

[0050] 2. The deep learning-based channel estimation method for 5G-OFDM systems demonstrates superior performance in the 1–30 dB signal-to-noise ratio range under 3GPP TDL-A multipath channels: compared with existing deep learning models (such as ChannelNet), the estimation accuracy is similar, but the number of parameters is reduced by 93%; the inference time is only 12.44 ms, which is 25.7% faster than SRCNN; the model computation is only 38.24 MMac, making it easy to deploy on resource-constrained mobile terminals.

[0051] 3. The lightweight design studied in this invention provides a feasible solution for the efficient deployment of deep learning in the 5G / 6G physical layer, which is especially suitable for resource-constrained mobile terminals and IoT devices. Attached Figure Description

[0052] Figure 1 This is a schematic diagram of a resource grid provided in an embodiment of the present invention;

[0053] Figure 2 This is a network structure model diagram provided in an embodiment of the present invention;

[0054] Figure 3 This is a flowchart of 5G NR simulation data provided in an embodiment of the present invention;

[0055] Figure 4 These are experimental graphs showing the mean squared error (MSE) performance of four networks;

[0056] Figure 5 This is an experimental diagram showing the impact of different pilot numbers on channel estimation accuracy;

[0057] Figure 6 This is a visualization experiment of the channel estimation results of a real channel algorithm;

[0058] Figure 7 This is a visualization experiment of the channel estimation results of the ESPCN-net algorithm;

[0059] Figure 8 This is a visualization experiment of the channel estimation results of different algorithms in ChannelNet;

[0060] Figure 9 This is a visualization experiment of the channel estimation results of the SRCNN algorithm;

[0061] Figure 10 This is a visualization experiment of the channel estimation results of the ReCNe algorithm;

[0062] Figure 11 This is a flowchart of the method of the present invention;

[0063] Figure 12 This is the simulation system parameter configuration table. Detailed Implementation

[0064] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0065] Example 1

[0066] like Figures 1 to 12 As shown, a channel estimation method based on deep learning for 5G-OFDM systems includes the following steps:

[0067] Step 1: Following the 5G New Radio standard protocol established by the 3rd Generation Partnership Project, build an OFDM orthogonal frequency division multiplexing system, obtain channel response data at the pilot in the 5G-OFDM orthogonal frequency division multiplexing system, form a low-resolution channel matrix, and recover the full subcarrier channel matrix through a neural network;

[0068] Step 2: Establish a deep learning neural network channel estimation model;

[0069] Step 3: Input the low-resolution channel matrix into the deep learning neural network channel estimation model for processing. The deep learning neural network channel estimation model adopts an interpolation-free end-to-end learning framework to directly extract features from the uninterpolated channel response data.

[0070] Step 4: Output the complete channel frequency response matrix through the deep learning neural network channel estimation model to achieve channel estimation. The network design proposed in this invention breaks through the computational paradigm of traditional deep learning channel estimation methods that rely on interpolation preprocessing. Compared with existing solutions (such as CNN or Transformer models based on interpolation reconstruction), this method directly utilizes the sparsity characteristics of the pilot signal. Through the channel-space dynamic mapping mechanism of the PixelShuffleHW module, the low-dimensional features of the pilot signal are directly upsampled to the target channel response dimension, eliminating the explicit interpolation step. End-to-end joint optimization suppresses the frequency domain phase distortion caused by interpolation while significantly reducing computational complexity and saving model inference time.

[0071] The OFDM (Orthogonal Frequency Division Multiplexing) system in Step 1 adopts a single-input single-output architecture. The research focuses on the 5G physical downlink shared channel. In the 5G New Radio system, the 5G physical downlink shared channel serves as a key channel for carrying downlink user data and control information. Through flexible subcarrier spacing, resource block configuration, and cyclic prefix configuration parameters, it efficiently supports large-scale spectrum utilization and parallel transmission of multiple users.

[0072] Data is transmitted through resource units, which are the smallest units of time-frequency resources. A specific resource grid implementation is as follows: Figure 2 As shown, it corresponds to one OFDM (Orthogonal Frequency Division Multiplexing) symbol in the time domain and one subcarrier in the frequency domain. In 5G NR, a time slot is a unit of time-domain signal resources, and each time slot contains... One OFDM symbol, typically in NR systems, The value is set to 14 or 12, depending on the subcarrier spacing used. A resource block is the smallest unit of frequency domain resource allocation, and each resource block contains... Each subcarrier, as specified in the 5G NR standard. In a time slot RB, of which There is no specific requirement for the quantity; it can be set flexibly according to actual needs.

[0073] The set parameter value is , That is, one time slot contains 12 symbols and 512 subcarriers. The resource grid is shown in the figure. The gray RE represents the demodulation reference signal, i.e. the pilot signal. The blank square represents the PDSCH RE, which is used to transmit the transmission signal information. In the NR system, the location and number of DMRS can be flexibly configured.

[0074] In the OFDM (Orthogonal Frequency Division Multiplexing) system model, the time-domain signal after OFDM modulation at the transmitting end is:

[0075]

[0076] in: For the first Modulation symbols on each subcarrier The total number of subcarriers in the system (including guard band) and the cyclic prefix length. Dynamically adjusted according to 3GPP TDL-A channel delay spread;

[0077] After passing through a multipath channel, the received signal It can be represented as:

[0078]

[0079] in: For the first The complex channel gain of each path, It is additive white Gaussian noise. It is a multipath number;

[0080] After removing the CP from the received signal and performing an FFT, the frequency domain received signal is:

[0081]

[0082] in: For the first Channel frequency response of each subcarrier This is frequency domain noise;

[0083] Finally, the LS estimate at the pilot frequency is... Mapped to a low-resolution channel matrix The full subcarrier channel matrix is ​​recovered using a neural network. .

[0084] The deep learning neural network channel estimation model in step two includes: a feature extraction layer, a composite residual module, and an upsampling module;

[0085] The feature extraction layer is used to perform shallow feature extraction on the input data. It uses a 3×3 convolutional layer combined with layer normalization.

[0086] The composite residual module is used to deepen feature extraction and adopts a residual connection structure;

[0087] The upsampling module is used to map low-resolution features to a high-resolution space.

[0088] The channel response value at the pilot frequency of the feature extraction layer is used for shallow feature extraction through the initial convolutional layer. A 3×3 convolutional kernel is used to achieve spatial context awareness, which is formally represented as follows:

[0089]

[0090] in Layer normalization (LayerNorm) is calculated as follows: the mean of each sample is calculated along the normalization dimension. and variance :

[0091]

[0092]

[0093] Where C is the number of channels, and the mean and variance are calculated independently of the batch and spatial dimensions (H, W), that is, each sample is calculated separately;

[0094] Normalize the input using the mean and variance:

[0095]

[0096] in, It is a local minimum value to prevent the denominator from being zero. The normalized mean is 0 and the variance is 1.

[0097] Introducing learnable scaling parameters and bias parameters To restore the expressive power of the model:

[0098]

[0099] in, and The shape is the same as C;

[0100] The channel response processed by the feature extraction layer is used as a preliminary feature input to the composite residual module, allowing the model to decide whether to retain or adjust the normalized distribution. The data input to the network in this invention is uninterpolated, which eliminates the data preprocessing process and allows feature extraction to be performed directly on the channel response data at the uninterpolated original pilot, further realizing end-to-end channel estimation.

[0101] The composite residual module employs residual connections. First, spatial features are encoded using 5×5 depthwise convolutions with the number of groups equal to the number of channels. Then, channel-dimensional normalization is used for stable training. Next, 1×1 convolutions expand the number of channels by a factor of four, and the GELU activation function is applied, introducing Global Response Normalization (GRN) to enhance inter-channel interactions. Finally, 1×1 convolutions compress the number of channels back to the original dimension. This "depthwise convolution-channel expansion-compression" design borrows from the network architecture of Google's ConvNeXt, maintaining input-output spatial resolution while... At the same time, by using the cascaded structure of layer normalization and double 1×1 convolution, the local feature extraction capability and cross-channel information fusion efficiency are balanced. The gradient vanishing problem of deep networks is alleviated by identity mapping, which significantly improves the training stability and representation capability of the model. The residual structure allows the gradient to be directly backpropagated through the cross-layer short-circuit path, avoiding the gradient decay caused by continuous multiplication in the traditional stacked layers. In order to reduce the computational cost, the convolution operation in the deep residual module adopts depthwise separable convolution, which can reduce the number of parameters to 1 / C of the standard convolution (C is the number of channels).

[0102] The upsampling module decouples the height and width scaling factors ( and ), decompose the channel dimension of the input tensor into The sub-blocks are used to achieve asymmetric upsampling through dynamic recombination of the channel-space dimension;

[0103] Specifically, the input shape is The feature map, after channel splitting, dimension permutation, and spatial merging operations, has its output resolution increased to [value missing]. The upsampling module is an extension and improvement of the classic PixelShuffle module, adapted to the upsampling requirements of asymmetric spatial resolution. The classic PixelShuffle is used in image super-resolution tasks, with the same scaling factor for height H and width W, without taking into account the inconsistent scaling factor of the number of pilots in the OFDM channel matrix in the frequency domain and time domain. This is crucial for flexible pilot configuration in 5G NR.

[0104] The channel estimation method has a computational cost of no more than 38.24 MMac and an inference time of no more than 12.44 ms during the inference phase, and can be applied to deployments on mobile terminals or IoT devices.

[0105] Example 2

[0106] This section presents a simulation and performance comparison of the proposed deep learning-based channel estimation scheme for OFDM communication systems. Waveforms and channels are generated based on the 3GPP 5G NR standard, and the simulation system parameter configuration is as follows: Figure 12 As shown;

[0107] like Figure 3 As shown, data generation is based on an end-to-end simulation process. This invention adopts a simulation data generation scheme based on the 3GPP 5G NR standard, constructing a channel estimation dataset that conforms to the specifications of actual communication systems using the MATLAB 5G Toolbox. The data generation process strictly follows the physical layer signal processing chain, ensuring a high degree of consistency between the simulation environment and the real 5G system. Each training sample contains an input-label pair, where the input is the channel estimate at the pilot after actual channel distortion and noise contamination, and the label is the ideal channel frequency response matrix. This method ensures that the model can learn an accurate mapping from impaired observations to the real channel. The key parameter configurations for data generation are shown in the table above. These parameters are uniformly managed through MATLAB's built-in parameter configuration functions, ensuring the consistency and repeatability of parameter settings. The dataset has a strict hierarchical structure and partitioning scheme to ensure the objectivity and reliability of model evaluation. The total sample size consists of 10,000 complete channel response matrices, with an SNR range of 1-30 dB. A scientific data partitioning strategy is employed to prevent overfitting and ensure generalization ability. The training set contains 9,000 samples (90% of the total samples), and the validation set contains 1,000 samples (10% of the total samples). The test set, within the 1-30 dB range, is generated in 1 dB increments, with 1,000 samples independently generated for each SNR level to test the model's performance under different signal-to-noise ratio conditions. To improve dataset diversity and model generalization ability, channel parameters, such as delay spread and maximum Doppler shift, are randomly selected in each iteration.

[0108] like Figure 4As shown, the mean squared error (MSE) performance of four networks—ESPCN-net, ChannelNet, ReCNet, and SRCNN—is presented within the signal-to-noise ratio (SNR) range. Experimental data indicates that the MSE of all networks monotonically decreases with increasing SNR, which is entirely consistent with theoretical expectations. Specifically, ESPCN-net and ChannelNet outperform ReCNet and SRCNN across the entire SNR range: ESPCN-net, with its innovative sub-pixel convolutional network structure, maintains a slight advantage in the low-to-medium SNR range (1-20dB); while ChannelNet, through its dual-network stacked structure, exhibits superior performance in the high SNR region (>25dB). Notably, when the SNR exceeds 25dB, ESPCN-net exhibits significant performance saturation. This is because, although its lightweight design significantly reduces computational complexity, its 47.25k parameter size may be insufficient for fine-grained channel estimation tasks under high SNR conditions, resulting in limited feature extraction capabilities and difficulty in fully capturing the subtle structures and long-range correlations in the channel matrix under high SNR conditions. Although the CRB module incorporates the ConvNeXt design concept, its 5×5 depthwise convolutional receptive field is still insufficient to model the complex spatial dependencies of high signal-to-noise ratio (SNR) channels. In contrast, ChannelNet, with its 682.34k parameters and cascaded structure, exhibits stronger performance in high SNR environments, demonstrating that its large parameter count supports more complex feature transformations. ESPCN-net strikes a balance between computational efficiency and estimation accuracy, but its lightweight architecture also limits its performance in high SNR regions.

[0109] like Figure 5 As shown, pilot design has a crucial impact on channel estimation performance. Theoretically, estimation accuracy improves with increasing pilot quantity, but too many pilots can crowd out spectrum resources and reduce data transmission efficiency. To investigate this trade-off, we conducted experimental evaluations under different signal-to-noise ratio (SNR) conditions. With each additional pilot, ESPCN-net's MSE performance improved by approximately 35% in the low-to-medium SNR region, but only by approximately 29% in the high SNR region. This indicates that pilot configuration should be dynamically adjusted according to channel conditions: in harsh low SNR environments, more pilots can be added to improve estimation accuracy, while in favorable high SNR environments, fewer pilots can be appropriately reduced to optimize spectrum utilization.

[0110] like Figures 6 to 10 As shown, in order to systematically evaluate the performance of different channel estimation algorithms, we conducted a comparative experiment in a test environment with a signal-to-noise ratio of 15dB, and performed a visualization analysis of the estimation results of each method. Figure 6 Five channel response images are shown, including real channel responses. Figure 6As a benchmark reference, and ESPCN-net Figure 7 ChannelNet Figure 8 ReCNet Figure 9 and SRCNN Figure 10 The estimation results of the four algorithms are presented. Visualization reveals that the channel matrix estimated by the SRCNN algorithm exhibits significant structural distortion and noise amplification. Obvious artifacts and local distortions are visible in its reconstructed image, indicating that this model struggles to effectively capture channel structural features in high-noise environments. The estimation results of ReCNet and ChannelNet are relatively close to the real channel in terms of overall contour, reflecting a certain reconstruction capability, but still suffer from visible detail loss and edge blurring, especially in regions with drastic changes in channel response, where the estimation results are overly smoothed and fail to fully preserve high-frequency components. In contrast, the channel matrix estimated by the ESPCN-net algorithm is visually closest to the real response. Its reconstructed image demonstrates excellent edge preservation and detail restoration, good structural continuity, and significant noise suppression.

[0111] The above is only used to illustrate the technical solution of the present invention and not to limit it. Any other modifications or equivalent substitutions made by those skilled in the art to the technical solution of the present invention, as long as they do not depart from the spirit and scope of the technical solution of the present invention, should be covered within the scope of the claims of the present invention.

Claims

1. A channel estimation method based on deep learning for 5G-OFDM systems, characterized in that, Includes the following steps: Step 1: Following the 5G New Radio standard protocol established by the 3rd Generation Partnership Project, build an OFDM orthogonal frequency division multiplexing system, obtain channel response data at the pilot in the 5G-OFDM orthogonal frequency division multiplexing system, form a low-resolution channel matrix, and recover the full subcarrier channel matrix through a neural network; Step 2: Establish a deep learning neural network channel estimation model; Step 3: Input the low-resolution channel matrix into the deep learning neural network channel estimation model for processing. The deep learning neural network channel estimation model adopts an interpolation-free end-to-end learning framework to directly extract features from the uninterpolated channel response data. Step 4: Output the complete channel frequency response matrix through the deep learning neural network channel estimation model to achieve channel estimation.

2. The channel estimation method for 5G-OFDM systems based on deep learning according to claim 1, characterized in that: The OFDM (Orthogonal Frequency Division Multiplexing) system in step one adopts a single-input single-output architecture and focuses on the 5G physical downlink shared channel. In the 5G New Radio system, the 5G physical downlink shared channel serves as a key channel for carrying downlink user data and control information. Through flexible subcarrier spacing, resource block configuration, and cyclic prefix configuration parameters, it efficiently supports large-scale spectrum utilization and parallel transmission of multiple users. Data is transmitted through resource units, which are the smallest units of time-frequency resources. A resource unit corresponds to an OFDM (Orthogonal Frequency Division Multiplexing) symbol in the time domain and a subcarrier in the frequency domain. In 5G NR, a time slot is the time-domain resource unit of a signal, and each time slot contains... One OFDM symbol, typically in NR systems, The value is set to 14 or 12, depending on the subcarrier spacing used. A resource block is the smallest unit of frequency domain resource allocation, and each resource block contains... Each subcarrier, as specified in the 5G NR standard. In a time slot RB, of which There is no specific requirement for the quantity; it can be set flexibly according to actual needs. The set parameter value is , That is, one time slot contains 12 symbols and 512 subcarriers; In the OFDM (Orthogonal Frequency Division Multiplexing) system model, the time-domain signal after OFDM modulation at the transmitting end is: in: For the first Modulation symbols on each subcarrier The total number of subcarriers in the system, and the cyclic prefix length. Dynamically adjusted according to 3GPP TDL-A channel delay spread; After passing through a multipath channel, the received signal Represented as: ； in: For the first The complex channel gain of each path, It is additive white Gaussian noise. It is a multipath number; After removing the CP from the received signal and performing an FFT, the frequency domain received signal is: ； in: For the first Channel frequency response of each subcarrier This is frequency domain noise; Finally, the LS estimate at the pilot frequency is... Mapped to a low-resolution channel matrix The full subcarrier channel matrix is ​​recovered using a neural network. .

3. The channel estimation method for 5G-OFDM systems based on deep learning according to claim 1, characterized in that: The deep learning neural network channel estimation model in step two includes: a feature extraction layer, a composite residual module, and an upsampling module; The feature extraction layer is used to perform shallow feature extraction on the input data. It uses a 3×3 convolutional layer combined with layer normalization. The composite residual module is used to deepen feature extraction and adopts a residual connection structure. The upsampling module is used to map low-resolution features to a high-resolution space.

4. The channel estimation method for 5G-OFDM systems based on deep learning according to claim 3, characterized in that: The channel response value at the pilot frequency of the feature extraction layer is used for shallow feature extraction via the initial convolutional layer. A 3×3 convolutional kernel is used to achieve spatial context awareness, which is formally represented as follows: ； in The representation layer normalization is calculated as follows: along the normalization dimension, the mean of each sample is calculated. and variance : ； ； Where C is the number of channels, and the mean and variance are calculated independently of the batch and spatial dimensions (H, W), that is, each sample is calculated separately; Normalize the input using the mean and variance: ； in, It is a local minimum value to prevent the denominator from being zero. The normalized mean is 0 and the variance is 1. Introducing scaling parameters and bias parameters To restore the expressive power of the model: ； in, and The shape is the same as C; The channel response processed by the feature extraction layer is used as the initial feature input to the composite residual module.

5. The channel estimation method for 5G-OFDM systems based on deep learning according to claim 3, characterized in that: The composite residual module employs residual connections. First, spatial features are encoded using a 5×5 depthwise convolution with the number of groups equal to the number of channels. Then, channel dimension layer normalization is used to stabilize the training. Subsequently, the number of channels is expanded by 4 times using a 1×1 convolution and the GELU activation function is applied, introducing Global Response Normalization (GRN) to enhance the interaction between channels. Finally, the number of channels is compressed back to the original dimension using a 1×1 convolution.

6. The channel estimation method for 5G-OFDM systems based on deep learning according to claim 3, characterized in that: The upsampling module decouples the height and width scaling factors ( and ), decompose the channel dimension of the input tensor into The sub-blocks are used to achieve asymmetric upsampling through dynamic recombination of the channel-space dimension; Specifically, the input shape is The feature map, after channel splitting, dimension permutation, and spatial merging operations, has its output resolution increased to [value missing]. .

7. The channel estimation method for 5G-OFDM systems based on deep learning according to claim 1, characterized in that: The channel estimation method has a computational cost of no more than 38.24 MMac and an inference time of no more than 12.44 ms during the inference phase, and can be applied to deployments on mobile terminals or IoT devices.

Images