Millimeter wave mMIMO channel estimation method and system based on neural network

By directly inferring CSI from time domain signals through neural network-based methods, the problems of high channel estimation complexity and insufficient noise resistance in millimeter wave large-scale MIMO systems are solved, and efficient channel estimation in low signal-to-noise ratio environments are realized , suitable for large bandwidth systems.

CN119071109APending Publication Date: 2024-12-03HUNAN UNIV OF TECH
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202410994671.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-24
Publication Date
2024-12-03

AI Technical Summary

Technical Problem

In millimeter wave large-scale MIMO systems, the estimation complexity of channel state information (CSI) is high, especially in low signal-to-noise ratio environments, which affects system performance.

Method used

Using a neural network-based channel estimation method, CSI is directly inferred from the received time domain signals through a pre-trained neural network model, reducing the computational complexity and performing channel estimation without OFDM demodulation.

Benefits of technology

It reduces the computational complexity, improves the performance and robustness of channel estimation, and is especially suitable for systems with large bandwidth and large subcarriers, improving the overall throughput and signal transmission quality of the system.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119071109A_ABST
    Figure CN119071109A_ABST
Patent Text Reader

Abstract

The invention relates to a millimeter wave mMIMO channel estimation method and system based on a neural network, and the method comprises the steps that user equipment receives time domain signals of transmitting signals of a plurality of transmitting antennas of a base station based on an MIMO-OFDM system architecture; and inputting the time domain signal of the sending signal of each transmitting antenna into a pre-trained neural network model to obtain a channel estimation result predicted by the neural network model. According to the method, the CSI can be directly deduced from the received time domain signal under the condition that OFDM demodulation is not carried out. According to the method, the number of signals needing to be processed is reduced, so that the calculation complexity is reduced, the method is particularly suitable for a system with a large bandwidth and a large number of subcarriers, and estimation challenges in a complex channel environment can be effectively coped with.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of wireless communication technology, and in particular to a millimeter wave mMIMO channel estimation method and system based on neural network. Background Art

[0002] Future wireless communication systems require extremely high transmission rates to support emerging applications such as high-quality 3D video, virtual reality, and augmented reality. Massive multiple-input multiple-output (MIMO) technology can significantly improve spectral efficiency and system capacity by deploying a large number of antennas on the transceiver and utilizing additional degrees of freedom in the spatial domain. On the other hand, millimeter wave (mmWave) communication is considered to be another key supporting technology, and its huge available bandwidth enables higher throughput. Due to the natural complementary characteristics of the two technologies, the mmWave band is very suitable for the deployment of massive MIMO. Specifically, the short wavelength of mmWave makes it possible to integrate large-scale antenna units in a limited space, while its severe path loss and signal blocking problems can be effectively compensated by the high beamforming gain provided by massive MIMO.

[0003] However, in millimeter-wave hybrid massive MIMO (mMIMO) systems, obtaining channel state information (CSI) becomes a thorny challenge. Channel estimation, as a means of obtaining CSI, is the first step in the entire processing chain of decoding data packets. Its goal is to identify the complex signal transformation imposed by the channel on the transmitted wireless signal, which is inferred by special information bits in the packet preamble. In spatial multiplexing systems, this complex transformation is captured by CSI. Knowing CSI enables the transmitter to perform additional precoding functions to maximize the transmission effect of signal energy in the target direction. Therefore, delays or errors in CSI calculation will seriously affect system performance, and the accuracy of CSI estimation is critical to the performance of other modules in the system. For example, signal detection, CSI feedback, CSI prediction, and beamforming.

[0004] Since the number of antennas in the millimeter wave mMIMO system has increased significantly compared to traditional MIMO, the number of channel parameters that need to be estimated has increased dramatically, resulting in extremely high computational complexity. In addition, the complexity and variability of the wireless communication channel itself further exacerbates this problem. During signal transmission, it will be affected by various factors such as path loss, shadow fading, multipath effect, and Doppler effect. Multipath effect can cause frequency selective fading, and in severe cases, it can cause inter-symbol interference of the received signal. In addition, in non-static scenarios, the relative movement between the transceiver devices can cause signal frequency deviation, which in turn produces Doppler shift. When the Doppler shift is severe, the received signal will experience time selective fading. Therefore, it is necessary to design an algorithm with low complexity and high estimation performance to accurately estimate the frequency and time selective fading channels to obtain accurate CSI. This is crucial for coherent detection, demodulation, and equalization.

[0005] Due to the high dimensionality in mMIMO, especially when OFDM technology is involved, most existing solutions use complex and deep architectures to estimate large channel matrices. These solutions treat the multi-dimensional input signal as a whole and usually require additional pre- or post-estimation steps. While the use of very deep architectures is a growing trend, their complexity usually limits their use on edge devices with limited power and processing capabilities.

[0006] In low signal-to-noise ratio environments, the noise resistance performance of existing channel estimation algorithms needs to be improved. These methods fail to fully consider the characteristics of low signal-to-noise ratio scenarios during training, so they perform poorly when faced with noise and interference in actual applications. Summary of the invention

[0007] Based on this, it is necessary to provide a millimeter wave mMIMO channel estimation method and system based on neural network to address the above technical problems, which can reduce computational complexity and improve performance.

[0008] In a first aspect, the present application provides a millimeter wave mMIMO channel estimation method based on a neural network. The method comprises:

[0009] The user equipment receives time domain signals of signals transmitted by multiple transmitting antennas of the base station;

[0010] The time domain signal of the transmission signal of each transmitting antenna is input into a pre-trained neural network model to obtain a channel estimation result predicted by the neural network model.

[0011] In a second aspect, the present application also provides a millimeter wave mMIMO channel estimation system based on a neural network. The system comprises:

[0012] A receiving module, used for receiving time domain signals of transmission signals of multiple transmitting antennas of a base station;

[0013] The prediction module is used to input the time domain signal of the transmission signal of each transmitting antenna into a pre-trained neural network model to obtain a channel estimation result predicted by the neural network model.

[0014] The above-mentioned neural network-based millimeter wave mMIMO channel estimation method and system is based on the MIMO-OFDM system. The user equipment receives the time domain signals of the transmission signals of multiple transmitting antennas of the base station, and inputs the time domain signals of the transmission signals of each transmitting antenna into a pre-trained neural network model to obtain the channel estimation results predicted by the neural network model. Therefore, this method can directly infer CSI from the received time domain signal without OFDM demodulation. This method reduces the number of signals that need to be processed, thereby reducing the computational complexity, and is particularly suitable for systems with large bandwidth and a large number of subcarriers. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 1 is a flow chart of a method for estimating a millimeter wave mMIMO channel based on a neural network in one embodiment;

[0016] Figure 2 is a schematic diagram of a MIMO-OFDM system architecture in one embodiment;

[0017] Figure 3 is a schematic diagram of a conventional millimeter wave mMIMO channel detection process in one embodiment;

[0018] Figure 4 A schematic diagram of a millimeter wave mMIMO channel detection process of the present application in one embodiment;

[0019] Figure 5 A schematic diagram of a data generation process for sending a signal in one embodiment;

[0020] Figure 6 A schematic diagram of the structure of a neural network model in one embodiment;

[0021] Figure 7 is a schematic diagram of the structure of an attention guidance module in one embodiment;

[0022] Figure 8 is a schematic diagram of the structure of a residual shrinkage module in one embodiment;

[0023] Fig. 9 4 is a structural block diagram of a millimeter wave mMIMO channel estimation system based on a neural network in one embodiment. DETAILED DESCRIPTION

[0024] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0025] Traditional channel estimation algorithms mainly include the least square (LS) algorithm and the linear minimum mean square error (LMMSE) algorithm. The LS algorithm is widely used due to its low computational complexity, but its disadvantage is that the noise power has a significant impact on the estimation performance, and it does not consider the noise in the received signal and the interference between subcarriers, and performs poorly in a low signal-to-noise ratio (SNR) environment. As one of the classic channel estimation algorithms, the LMMSE algorithm has good noise resistance, but in millimeter-wave massive MIMO systems, the computational complexity is extremely high due to the large dimension of matrix operations. In addition, other traditional linear estimation methods rely on statistical models, which are highly complex and lack robustness in highly dynamic environments. In contrast, deep learning networks can extract the inherent characteristics of the channel from a large amount of training data and estimate the channel more accurately through efficient parallel computing methods. Deep learning methods can not only effectively deal with noise and interference problems, but also maintain good estimation performance in complex and changing channel environments.

[0026] Based on this, the present application provides a millimeter wave mMIMO channel estimation method based on neural network, such as Figure 1 As shown, the method includes:

[0027] Step 102: The user equipment receives time domain signals of signals transmitted by multiple transmitting antennas of the base station.

[0028] The MIMO-OFDM system architecture of an embodiment is as follows: Figure 2 As shown, the MIMO-OFDM system includes a transmitter and a receiver. The system architecture adopts the time division duplex (TDD) mode and assumes that the channel changes slowly (the coherence time is 10 to 100 milliseconds). In this mode, downlink transmission includes two key stages: first, the channel detection stage, the user equipment (UserEquipment, UE) estimates the CSI of the complete MIMO channel and sends the estimation result to the base station (Base Station, BS); second, the data transmission stage, the base station uses the received CSI estimation value to calculate the precoding weight of the directional beam to optimize the data transmission performance.

[0029] In order to complete the two phases of channel detection and data transmission within the channel coherence time, CSI (Channel State Information) estimation must be completed quickly. Strict time requirements ensure that the base station can use the channel reciprocity of the wireless front end for downlink transmission. By reducing the overhead associated with CSI estimation, the time of the channel detection phase can be shortened, so that more data can be transmitted within a given channel coherence time, ultimately improving the overall throughput of the system.

[0030] In the explicit channel detection phase, it is necessary to calculate not only the mMIMO channel for each signal pair, but also the mMIMO channel for each subcarrier. If the CSI matrix is ​​not calculated accurately, the beamforming effect between the mMIMO base station and the user equipment will be weakened, resulting in an increase in the bit error rate (BER) during data transmission. In addition, if the CSI matrix is ​​not calculated in time within the channel coherence time, it will have an adverse effect on subsequent data transmission because the channel coefficients used for beamforming are outdated and cannot reflect the current channel status.

[0031] In the millimeter wave hybrid massive MIMO-OFDM system, this application uses Figure 2 The hybrid architecture shown in the figure has N t transmit antennas and The receiving end is equipped with N r receiving antennas and N receive RF chains. s Data flows through N t and N r For data transmission, the number of antennas is much larger than the data stream, that is,

[0032] N between the receiving end and the transmitting end in the delay domain t ×N r The channel matrix is:

[0033]

[0034] In the formula, L represents the number of corresponding paths, represents the gain of the lth path, is the average power gain, τ l is the delay of the lth path, θ l and φ l ∈[0,2π] are the arrival and departure angles (AoA / AoD) of the receiver and transmitter respectively. r (θ l )and are the array steering vectors of the uniform linear array (ULA), which can be expressed as

[0035]

[0036] Where d is the distance between adjacent antennas and λ is the signal wavelength.

[0037] according to Figure 2 The channel model in OFDM is:

[0038]

[0039] Where fs is the sampling rate and K is the number of OFDM subcarriers.

[0040] The signal is modulated by a set of precoding weights, propagated through the channel, and recovered by a set of combining weights. Assume that the initial signal X transmitted by the transmitter is an N s The columns of the matrix represent the data stream. After hybrid precoding at the transmitting end and hybrid merging at the receiving end, the k-th subcarrier received signal at the baseband of the receiving end can be expressed as

[0041] Y k =W k (H k F k X k +N k ) (5)

[0042] In the formula, F k =F rf F bb is an N s ×N t Matrix, representing the precoding weight; W k =W rf W bb is an N r ×N s Matrix, representing the merge weight; N k is additive white Gaussian noise and obeys complex Gaussian distribution.

[0043] Step 104: input the time domain signal of the transmission signal of each transmitting antenna into a pre-trained neural network model to obtain a channel estimation result predicted by the neural network model.

[0044] In this application, a compact convolutional neural network model is designed to keep the computation time low, and it is trained to learn a joint approximation of OFDM demodulation and LS / LMMSE channel estimation methods.

[0045] Figure 3It shows how the DNN model can replace the processing blocks related to demodulation and channel estimation in a typical mMIMO channel detection process. Unlike the previously proposed state of the art, the method of this application, such as Figure 4 As shown in Figure 1, the received time domain waveform corresponding to the LTF obtained after synchronization is used as the input of the model. This can completely avoid the OFDM demodulation required for CSI estimation, further reducing the computational burden associated with this step in systems with large bandwidths and requiring a large number of subcarriers. The model can directly infer the spectral components of the received time domain signal without explicit demodulation.

[0046] The present application is a millimeter wave mMIMO channel estimation method based on a neural network. Based on the MIMO-OFDM system, the user equipment receives the time domain signals of the transmission signals of multiple transmitting antennas of the base station, and inputs the time domain signals of the transmission signals of each transmitting antenna into a pre-trained neural network model to obtain the channel estimation results predicted by the neural network model. Thus, the method can directly infer the CSI from the received time domain signal without OFDM demodulation. This method reduces the number of signals that need to be processed, thereby reducing the computational complexity, and is particularly suitable for systems with large bandwidths and a large number of subcarriers.

[0047] In another embodiment, the neural network-based millimeter wave mMIMO channel estimation method also includes: the transmitting end generates a baseband signal using orthogonal mapping, and modulates and inverse fast Fourier transforms the baseband signal to generate a time domain signal suitable for multiple transmitting antennas of the base station; in combination with the millimeter wave mMIMO system, the time domain signal is processed through hybrid beamforming technology so that the time domain signal is transmitted in parallel to the channel through multiple transmitting antennas.

[0048] Specifically, the data generation process of sending the signal, such as Figure 5 As shown in Figure 1, considering a MIMO-OFDM system operating in the millimeter wave frequency band, a standard long training frame (Long Training Field, LTF) is generated, which contains a pilot sequence spanning the pilot and data available subcarriers. This LTF signal is then copied to different orthogonal positions of all transmit antennas. Using millimeter wave mMIMO beamforming, the base station transmits all N t Channel sounding frames are transmitted in parallel on the transmit antennas.

[0049] In the LTF sequence of the preamble, each channel detection frame spans L OFDM symbols. In order to avoid interference during the initial omnidirectional simultaneous transmission after the orthogonal signals are generated, a specific mapping sequence needs to be applied to each signal. These mapping sequences adjust the phase and amplitude of the signal so that the signals remain orthogonal when all transmitting antennas transmit simultaneously, thereby effectively avoiding interference. Finally, the inverse fast Fourier transform (IFFT) is used to generate the final time domain signal to be sent at each antenna unit of the base station. In general, the use of millimeter wave mMIMO beamforming combined with orthogonal mapping sequences not only effectively improves the accuracy and efficiency of channel detection, but also ensures the transmission quality of the signal in the multi-antenna system, improving the overall performance of the wireless communication system.

[0050] In another embodiment, the time domain signal includes a long training field (LTF) signal and an orthogonal coding sequence associated with a transmit antenna.

[0051] Specifically, in the experiment, the model input consists of arriving at a specific receiving antenna N r The LTF signal p nR and the given orthogonal coding sequence O associated with the transmitting antenna nT Cascaded, the size of the input tensor is [(T P +T O )×1], where T P is the number of symbols belonging to the LTF sequence in the time domain, T O is the length of the orthogonal coding sequence. The reason for concatenating the LTF sequence and the orthogonal coding sequence into the input is as follows: for the LTF signal received on a given receiving antenna, all channel states relative to each transmitting antenna must be restored. This means that it is necessary to determine the channel state of the given receiving antenna N. r To achieve this, an orthogonal coding sequence is applied to each signal to ensure that the signals from different transmitting antennas are distinguishable at the receiving end, thereby accurately estimating each channel state. Without an orthogonal coding sequence, the input signal p nR In all cases, the signal received by the receiver cannot distinguish which transmit antenna it comes from, and the model will not be able to generate the correct channel state estimate. On the other hand, the model output is a [K×1] tensor, where k is the number of subcarriers in the target system. This output tensor corresponds to the model's prediction of the input LTF signal during the propagation process, the transmit antenna N t and receiving antenna N r The channel frequency response experienced between.

[0052] In another embodiment, the time domain signal of the transmission signal of each transmitting antenna is represented in complex form, and the time domain signal of the transmission signal of each transmitting antenna is input into a pre-trained neural network model to obtain a channel estimation result predicted by the neural network model, including: converting the real component and the imaginary component of each time domain signal into real value form respectively to obtain the real value of the real component and the real value of the imaginary component of each time domain signal; inputting the real value and the real value of the real component of each time domain signal into a pre-trained neural network model respectively to obtain the real part prediction value of the channel state information and the imaginary part prediction value of the channel state information in real value form predicted by the neural network model; combining the real part prediction value of the channel state information and the imaginary part prediction value of the channel state information in real value form into a complex number representation to obtain the channel estimation result.

[0053] Specifically, in order to reduce the size of the input signal and effectively capture the impact of the channel on the amplitude and phase components of the received signal, the real and imaginary parts of the input signal are selected to be processed independently. To this end, the real and imaginary data can be used to train the neural network model separately, and the structure of the neural network model can be the same. That is, two identical and independent specialized models are created to process the real and imaginary components respectively, both of which are input in real-valued format. The real-valued output of each model is then converted back to a complex representation to generate the final channel estimation result. Since the two models are independent, they can run in parallel, thereby improving processing efficiency. Such an approach can reduce the complexity of signal processing while ensuring accurate analysis of channel characteristics, thereby optimizing system performance and improving the accuracy and efficiency of signal decoding.

[0054] In this embodiment, the influence of the channel on the amplitude and phase components of the received signal can be effectively processed, and by independently processing the real part and the imaginary part of the input signal, the model can run in parallel, improving the processing efficiency and ensuring accurate analysis of the channel characteristics.

[0055] In another embodiment, the neural network model includes a convolutional layer, an attention guidance module, a first residual shrinkage module, a second residual shrinkage module, a first fully connected layer and a second fully connected layer from the input end to the output end; the attention guidance module dynamically adjusts the processing focus to give priority to processing the signal part with higher information density; the first residual shrinkage module and the second residual shrinkage module are used to eliminate noise and redundant information in the data.

[0056] Specifically, when building a neural network model, the choice of model architecture is crucial. Figure 6As shown, the network configuration selected in this application includes a convolution layer, an attention guidance module, a first residual shrinkage module, a second residual shrinkage module, a first fully connected layer, and a second fully connected layer. The architecture starts with a one-dimensional convolution layer with 64 filters, a kernel size of 7×7, and a step size of 1. The kernel size and step size are designed to effectively capture signal features. Next, an attention module is introduced, which dynamically adjusts the processing focus to prioritize the signal part with higher information density, thereby significantly improving the model's information processing ability and reducing the problem of information overload. In addition, two residual shrinkage modules are used to further remove noise and redundant information in the data and optimize the quality of the signal. The deep part of the network consists of two fully connected layers, of which the first fully connected layer is equipped with 1024 neurons and ReLU activation function to enhance nonlinear processing capabilities; the second fully connected layer uses a linear activation function, which is mainly used for the output of regression tasks. In order to improve the generalization ability of the model, batch normalization is performed before each fully connected layer, and a Dropout layer with a dropout rate of 15% is added between the second residual module and the first fully connected layer, which effectively reduces the risk of overfitting. Therefore, in order to obtain a complete millimeter wave mMIMO channel matrix, a matrix of size N can be constructed. t ×N r to process all required inputs in parallel and produce a K×N t ×N r The channel matrix is ​​taken as output.

[0057] In another embodiment, the attention guiding module includes a one-dimensional convolution layer, a batch normalization layer, a global average pooling layer and a fully connected layer from the input end to the output end; the one-dimensional convolution layer is used to extract and process the input features; the batch normalization layer is used to standardize the processed data; the global average pooling (GAP) layer is used to compress the features of each channel and highlight the global important information of each channel; the fully connected layer is used to generate weights for each channel.

[0058] Specifically, the Attention-Guided Module (AGM) is a technology designed to enhance the attention and efficiency of neural networks in processing features, improving the channel estimation performance in millimeter-wave massive MIMO-OFDM systems. This module effectively focuses on the key information in the input data by dynamically adjusting the weights of feature channels, thereby improving the performance of the model. This module consists of a one-dimensional convolutional layer, a batch normalization layer, a global average pooling layer (GAP), and the ECANet attention mechanism.

[0059] like Figure 7As shown in the figure, in the architecture of the attention guidance module (AGM), the input data first passes through the convolution layer for feature extraction, and the subsequent batch normalization layer helps to standardize the data, extract preliminary features, and improve the stability of the training process. Next, the global average pooling (GAP) layer is used to further compress the features of each channel and highlight the global important information of each channel. This information is passed to the fully connected layer, which is responsible for generating weights for each channel. The weights are adjusted by the sigmoid activation function and are used to weight the features of the original channel to enhance key information and suppress irrelevant information.

[0060] After channel global average pooling, the attention network uses one-dimensional convolution to achieve cross-channel information interaction, and the size of the convolution kernel k is adaptive through the function. Given the channel dimension C, the convolution kernel size k can be adaptively determined as:

[0061]

[0062] Among them, odd is an odd number, γ=2, and b=1.

[0063] The high dimensionality of millimeter-wave massive MIMO systems, especially when combined with OFDM technology, requires most existing solutions to require complex and deep architectures to estimate large channel matrices. Through its efficient ECA module, AGM uses one-dimensional convolution to achieve cross-channel information interaction and adaptively determines the size of the convolution kernel, thereby achieving efficient channel feature extraction and processing without increasing complexity. At the same time, the noise resistance of existing deep learning channel estimation algorithms in low signal-to-noise ratio environments needs to be improved, mainly because the characteristics of low signal-to-noise ratio scenarios are not fully considered during the training process. AGM dynamically adjusts feature weights, introduces residual connections and "soft thresholding" technology in combination with residual shrinkage modules, effectively removes noise and redundant information from the data, improves performance in low signal-to-noise ratio environments, and further realizes the efficiency and robustness of channel estimation in millimeter-wave hybrid massive MIMO-OFDM systems.

[0064] In another embodiment, Figure 8 As shown, the first residual shrinkage module and the second residual shrinkage module each include a convolution layer, a batch normalization layer, a ReLU activation function, and a soft threshold processing subnet from the input to the output end; the soft threshold processing subnet is used to adaptively adjust the threshold.

[0065] The residual shrinkage module aims to improve the overall efficiency and performance of the network by processing the input signal and significantly reducing the amount of redundant information passed to subsequent layers. The residual shrinkage module introduces "soft thresholding" as a "shrinkage layer" into the residual module. Its main function is to effectively remove noise from the input signal by adaptively setting the threshold. Noise is not just an interference signal in the traditional sense, but more broadly includes feature information that is irrelevant to the current task.

[0066] Specifically, the residual shrinkage module first processes the input signal through a series of convolutional layers, each followed by a batch normalization layer and a ReLU activation function to ensure the nonlinear expression and standardization of the signal. During the processing, the subnetwork of the residual shrinkage module further extracts and emphasizes important features through an absolute global average pooling layer, while compressing the feature space to reduce the complexity of data transmission, and its size is [V×1×1]. The subsequent fully connected layer further refines these features, and the ReLU and sigmoid activation functions ensure the effective transmission and conversion of features.

[0067] In the millimeter-wave massive MIMO-OFDM system, channel estimation needs to process a large amount of subcarrier and antenna data, which places higher requirements on the model's computing power. To address this challenge, the residual shrinkage module then introduces soft thresholding to compress irrelevant feature information and noise in the input data to near zero, further simplifying the amount of data processed by the model while also effectively suppressing noise. The formula for soft thresholding is as follows:

[0068]

[0069] Where x represents the input feature, y represents the output feature, and τ represents the threshold. The threshold τ needs to be set to a positive number, and its size should be appropriate and not too large to avoid the output feature y being all zero, which will cause the soft thresholding to lose its effect. At the same time, the derivative formula of the soft thresholding function is as follows:

[0070]

[0071] The derivative of this function shows that unless the absolute value of the input feature is within the threshold range, its derivative is 1, which is similar to the characteristics of the ReLU activation function. Therefore, the soft thresholding function is also helpful in preventing the problem of gradient disappearance and gradient explosion.

[0072] In the overall network structure, the subnetwork adaptively adjusts the threshold to ensure that the result of multiplying the absolute value average of the feature map by the output τ of the last fully connected layer after sigmoid activation is a positive number. The sigmoid function ensures that the value of τ is between 0 and 1, thereby avoiding setting all outputs to zero and maintaining the validity of the data. This design effectively solves the high dimensionality of channel estimation in millimeter-wave massive MIMO-OFDM systems and the noise resistance problem in low signal-to-noise ratio environments, making the residual shrinkage module perform well in channel estimation.

[0073] Based on the same inventive concept, the embodiment of the present application also provides a millimeter wave massive MIMO channel estimation system based on a neural network, such as Fig. 9 As shown, including:

[0074] The receiving module 902 is used to receive time domain signals of transmission signals of multiple transmitting antennas of the base station;

[0075] The prediction module 904 is used to input the time domain signal of the transmission signal of each transmitting antenna into a pre-trained neural network model to obtain a channel estimation result predicted by the neural network model.

[0076] In another embodiment, the system also includes a sending module, which is used for the transmitting end to generate a baseband signal using orthogonal mapping, and to modulate and inverse fast Fourier transform the baseband signal to generate a time domain signal suitable for multiple transmitting antennas of the base station; in combination with the millimeter wave mMIMO system, the time domain signal is processed through hybrid beamforming technology so that the time domain signal is transmitted in parallel to the channel through multiple transmitting antennas.

[0077] In another embodiment, the time domain signal includes an LTF signal and an orthogonal coding sequence associated with a transmit antenna.

[0078] In another embodiment, the time domain signal of the transmission signal of each transmitting antenna is represented in complex form, and the prediction module includes:

[0079] A signal processing module, used for converting the real component and the imaginary component of the time domain signal into real value forms respectively, to obtain the real value of the real component and the real value of the imaginary component of each of the time domain signals;

[0080] A processing module, used to input the real value of the real component and the real value of the imaginary component of each of the time domain signals into a pre-trained neural network model, respectively, to obtain a real-valued predicted value of the channel state information and an imaginary-valued predicted value of the channel state information predicted by the neural network model;

[0081] The conversion module is used to combine the real part prediction value of the channel state information and the imaginary part prediction value of the channel state information in the real value form into a complex number representation to obtain a channel estimation result.

[0082] In another embodiment, the neural network model includes, from the input end to the output end, a convolutional layer, an attention guidance module, a first residual shrinkage module, a second residual shrinkage module, a first fully connected layer, and a second fully connected layer.

[0083] In another embodiment, the attention guiding module includes, from the input end to the output end, a one-dimensional convolutional layer, a batch normalization layer, a global average pooling layer, and a fully connected layer.

[0084] In another embodiment, the first residual shrinkage module and the second residual shrinkage module each include a convolution layer, a batch normalization layer, a ReLU activation function, and a soft thresholding processing subnet in sequence from the input to the output; the soft thresholding processing subnet is used to adaptively adjust the threshold.

[0085] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. Non-relational databases may include distributed databases based on blockchains, etc., but are not limited to this. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but are not limited to this.

[0086] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0087] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be construed as limiting the scope of the present application. It should be noted that, for a person of ordinary skill in the art, several modifications and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A millimeter wave mMIMO channel estimation method based on neural network, It is characterized in that include: The user equipment receives time domain signals of signals transmitted by multiple transmitting antennas of the base station; The time domain signal of the transmission signal of each transmitting antenna is input into a pre-trained neural network model to obtain a channel estimation result predicted by the neural network model.

2. The neural network-based millimeter wave mMIMO channel estimation method according to claim 1, It is characterized in that The method further comprises: The transmitting end generates a baseband signal by using orthogonal mapping, and performs modulation and inverse fast Fourier transform on the baseband signal to generate a time domain signal suitable for multiple transmitting antennas of the base station; In combination with the millimeter wave mMIMO system, the time domain signal is processed by hybrid beamforming technology so that the time domain signal is transmitted to the channel in parallel through multiple transmitting antennas.

3. The neural network-based millimeter wave mMIMO channel estimation method according to claim 2, It is characterized in that The time domain signal includes a long training field signal and an orthogonal coding sequence associated with a transmit antenna.

4. The method for millimeter wave mMIMO channel estimation based on neural network according to claim 1, It is characterized in that The time domain signal of the transmission signal of each of the transmitting antennas is represented in a complex form, and the time domain signal of the transmission signal of each of the transmitting antennas is input into a pre-trained neural network model to obtain a channel estimation result predicted by the neural network model, including: Convert the real component and the imaginary component of each of the time domain signals into real value forms respectively, to obtain the real value of the real component and the real value of the imaginary component of each of the time domain signals; Inputting the real value of the real component and the real value of the imaginary component of each of the time domain signals into a pre-trained neural network model respectively, to obtain a real value prediction value of the channel state information and an imaginary value prediction value of the channel state information in real value form predicted by the neural network model; The real part prediction value of the channel state information and the imaginary part prediction value of the channel state information in the real value form are combined into a complex number representation to obtain a channel estimation result.

5. The neural network-based millimeter wave mMIMO channel estimation method according to claim 1 or 4, It is characterized in that The neural network model includes, from the input end to the output end, a convolutional layer, an attention guidance module, a first residual shrinkage module, a second residual shrinkage module, a first fully connected layer, and a second fully connected layer.

6. The neural network-based millimeter wave mMIMO channel estimation method according to claim 5, It is characterized in that The attention guidance module includes a convolution layer, a batch normalization layer, a global average pooling layer and a fully connected layer from the input end to the output end.

7. The neural network-based millimeter wave mMIMO channel estimation method according to claim 5, It is characterized in that The first residual shrinkage module and the second residual shrinkage module each include a convolution layer, a batch normalization layer, a ReLU activation function and a soft threshold processing subnet from the input to the output end; the soft threshold processing subnet is used to adaptively adjust the threshold.

8. A millimeter wave mMIMO channel estimation system based on neural network, It is characterized in that The system comprises: A receiving module, used for receiving time domain signals of transmission signals of multiple transmitting antennas of a base station; The prediction module is used to input the time domain signal of the transmission signal of each transmitting antenna into a pre-trained neural network model to obtain a channel estimation result predicted by the neural network model.

9. The neural network-based millimeter wave mMIMO channel estimation system according to claim 8, It is characterized in that The system further comprises: The transmitting module is used to generate a baseband signal by orthogonal mapping at the transmitting end, and modulate and inverse fast Fourier transform the baseband signal to generate a time domain signal suitable for multiple transmitting antennas of the base station; in combination with the millimeter wave mMIMO system, the time domain signal is processed through hybrid beamforming technology so that the time domain signal is transmitted to the channel in parallel through multiple transmitting antennas.

10. The neural network-based millimeter wave mMIMO channel estimation system according to claim 8, It is characterized in that The time domain signal of the transmission signal of each transmitting antenna is represented in complex form, and the prediction module includes: A signal processing module, used for converting the real component and the imaginary component of each of the time domain signals into real value forms, to obtain the real value of the real component and the imaginary component of each of the time domain signals; A processing module, used to input the real value of the real component and the real value of the imaginary component of each of the time domain signals into a pre-trained neural network model, respectively, to obtain a real part prediction value of the channel state information and an imaginary part prediction value of the channel state information in a real value form predicted by the neural network model; The conversion module is used to combine the real part prediction value of the channel state information and the imaginary part prediction value of the channel state information in the real value form into a complex number representation to obtain a channel estimation result.

Citation Information

Cited By

  • CNNLSTM-based energy storage equipment temperature control method and system

    CN119806240A