A precoding model training method and training system

Abstract

The present application provides a precoding model training method and training system, wherein the precoding model training method includes: processing the received signals obtained by several receiving antennas to obtain the covariance matrix corresponding to the received signals; based on the covariance matrix, calculating the bit error rate corresponding to the current channel; mapping the bit error rate with the pixels in the color spectrum to obtain the channel image corresponding to the current channel; inputting the channel image into the precoding model, extracting and reducing the channel features in the channel image; creating a fully connected layer adapted to the current channel features, and outputting a precoding matrix adapted to the current channel features; the present application visualizes the channel features by constructing a color spectrum, and adaptively trains the precoding model based on a deep learning neural network, effectively improving the generalization performance of the precoding model.

Description

Technical Field

[0001] The present application belongs to the field of model training technology, and specifically relates to a precoding model training method and training system. Background Art

[0002] In the field of wireless communications, precoding technology is used to preprocess the transmitted signal to adapt to channel characteristics, which is an important means to improve system channel capacity and data transmission reliability. Traditional precoding methods rely on CSI (Channel State Information) to construct optimization targets for calculation and generate precoding matrices through mathematical operations such as matrix inversion and eigendecomposition.

[0003] However, most traditional precoding methods directly use the vectorized channel matrix as the neural network input, ignoring the two-dimensional spatial structure and multi-dimensional statistical characteristics of the channel parameters, resulting in the model's insufficient ability to represent the global trends and local details of the channel. At the same time, due to the complexity of the communication environment, the fixed network architecture lacks a dynamic adjustment mechanism for different communication environments. When there is a difference between the actual deployment environment and the training data distribution, the model's generalization performance will be significantly reduced, making it difficult to meet the reliable communication needs in complex wireless environments, and there are certain drawbacks. Summary of the Invention

[0004] This application provides a precoding model training method and training system, which visualizes channel features by constructing a color map, and adaptively trains the precoding model based on a deep learning neural network, effectively improving the generalization performance of the precoding model.

[0005] In order to solve the above technical problems, the present application provides a precoding model training method, comprising the following steps:

[0006] Processing received signals obtained by multiple receiving antennas to obtain a covariance matrix corresponding to the received signals;

[0007] Calculating a bit error rate corresponding to the current channel based on the covariance matrix;

[0008] Mapping the bit error rate to pixels in the color map to obtain a channel image corresponding to the current channel;

[0009] Inputting the channel image into a precoding model, extracting and reducing the channel features in the channel image;

[0010] A fully connected layer adapted to the current channel characteristics is created, and a precoding matrix adapted to the current channel characteristics is output.

[0011] As a further improvement of the present application, obtaining the covariance matrix corresponding to the received signal includes:

[0012] Converting received signals obtained by a plurality of receiving antennas into received signal vectors;

[0013] A covariance matrix corresponding to the received signal is calculated based on the received signal vector.

[0014] As a further improvement of the present application, the bit error rate corresponding to the current channel is calculated based on the covariance matrix, including:

[0015] Obtaining a covariance matrix of a transmitted signal according to a covariance matrix of an interference signal, a covariance matrix of noise, and a covariance matrix of a received signal;

[0016] Calculating the effective signal power passing through the current channel, the interference signal power received by the receiving end, and the noise signal power received by the receiving end based on the covariance matrix of the transmitted signal;

[0017] Combining the effective signal power, the interference signal power and the noise signal power to obtain a signal-to-noise ratio corresponding to the current signal;

[0018] The bit error rate corresponding to the current channel is obtained by calculating the signal-to-noise ratio.

[0019] As a further improvement of the present application, mapping the bit error rate to pixels in the color map to obtain a channel image corresponding to the current channel includes:

[0020] The bit error rate is mapped one-to-one with pixels in the color spectrum through a color mapping function to obtain a channel image corresponding to the current channel, so that each pixel in the channel image represents the bit error rate of the corresponding position of the current channel.

[0021] As a further improvement of the present application, the color spectrum includes at least any one of the RGB color spectrum, the CMYK color spectrum, the HSV color spectrum and the Lab color spectrum.

[0022] As a further improvement of the present application, the step of inputting the channel image into a precoding model and extracting and reducing the channel features in the channel image includes:

[0023] Inputting the channel image into the precoding model;

[0024] Extracting local channel features from the channel image;

[0025] The extracted channel features are downsampled to retain local areas where the channel features have changed.

[0026] As a further improvement of the present application, the step of creating a fully connected layer adapted to the current channel characteristics includes:

[0027] Freeze the convolutional layer weights during training;

[0028] Create a fully connected layer, dynamically adjust the parameters of the fully connected layer according to the current channel environment, and optimize and train the parameter matrix of the newly created fully connected layer.

[0029] As a further improvement of the present application, the outputting of a precoding matrix adapted to the current channel characteristics includes:

[0030] Performing a linear transformation on the channel features after dimensionality reduction processing by an activation function, mapping the channel features into a precoding parameter vector;

[0031] The precoding parameter vector is reorganized into a precoding matrix, and a precoding matrix adapted to the current channel characteristics is output.

[0032] As a further improvement of the present application, the precoding model is a CNN model, and the CNN model includes a convolutional layer, a pooling layer and a fully connected layer;

[0033] Wherein, the convolutional layer is used to extract local channel features in the channel image;

[0034] The pooling layer is used to perform dimensionality reduction processing on the extracted channel features;

[0035] The fully connected layer is used to map the channel features after dimensionality reduction processing into a precoding parameter vector, and reorganize the parameter vector into a precoding matrix.

[0036] Based on the above-mentioned precoding model training method, the present application also provides a precoding model training system, including:

[0037] a processing unit, configured to process received signals acquired by a plurality of receiving antennas to obtain a covariance matrix corresponding to the received signals;

[0038] a calculation unit, configured to calculate a bit error rate corresponding to a current channel based on the covariance matrix;

[0039] a mapping unit, configured to map the bit error rate to pixels in a color map to obtain a channel image corresponding to a current channel;

[0040] an extraction unit, configured to input the channel image into a precoding model, and extract and perform dimensionality reduction processing on channel features in the channel image;

[0041] The output unit is used to create a fully connected layer adapted to the current channel characteristics and output a precoding matrix adapted to the current channel characteristics.

[0042] The precoding model training method and training system provided in the present application first process the received signal of the receiving antenna to obtain the covariance matrix corresponding to the received signal, calculate the bit error rate of the current channel based on the covariance matrix, map the bit error rate to the pixels in the color map to obtain the channel image, thereby converting the abstract bit error rate data into an intuitive image form, characterizing the channel characteristics through image features, and using the image structure to retain the distribution characteristics of the channel space. Combined with the convolution layer to extract local detail features and the pooling layer to retain the characteristics of the feature change area, it effectively captures the local details and global trends of the channel space, avoiding the problem of structural information loss caused by directly using the vectorized channel matrix as the input of the neural network; through the design of freezing the convolution layer weights and dynamically adjusting the fully connected layer parameters, the model can adaptively optimize the fully connected layer parameters according to the current channel environment, and output a precoding matrix adapted to the current channel characteristics, thereby improving the generalization performance and adaptation accuracy of the precoding model, and improving the performance of the communication system in various channel environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0044] Figure 1 A flowchart of the precoding model training method provided in an embodiment of the present application.

[0045] Figure 2 This is a flowchart for calculating the covariance matrix in the precoding model training method provided in an embodiment of the present application.

[0046] Figure 3 This is a flowchart for calculating the bit error rate in the precoding model training method provided in an embodiment of the present application.

[0047] Figure 4 This is a flowchart of channel feature extraction in the precoding model training method provided in an embodiment of the present application.

[0048] Figure 5 This is a flowchart of outputting a precoding matrix in the precoding model training method provided in an embodiment of the present application.

[0049] Figure 6 A schematic diagram of the structure of the precoding model training system provided in an embodiment of the present application. DETAILED DESCRIPTION

[0050] In order to make the purpose, technical solutions and advantages of this application more clearly understood, this application is further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0051] In the description of this application, the meaning of "plurality" is at least two, for example, two, three, etc., unless otherwise specifically defined. All directional indications in the embodiments of this application (such as up, down, left, right, front, back, etc.) are only used to explain the relative position relationship and movement of the components under a specific posture (as shown in the accompanying drawings). If the specific posture changes, the directional indication will also change accordingly. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions.

[0052] To provide a more detailed and complete description of the present disclosure, the following provides illustrative descriptions of the embodiments and examples of the present application; however, these descriptions are not intended to be the only ways to implement or use the embodiments of the present application. The embodiments cover features of various embodiments, as well as the method steps and sequences for constructing and operating these embodiments. However, other embodiments may also be used to achieve the same or equivalent functionality and step sequences.

[0053] In the field of wireless communications, precoding technology is used to preprocess the transmitted signal to adapt to channel characteristics. This is an important means to improve system channel capacity and data transmission reliability. Traditional precoding methods rely on channel state information (CSI) to construct optimization targets for calculation and generate precoding matrices through mathematical operations such as matrix inversion and eigendecomposition.

[0054] However, most traditional precoding methods directly use the vectorized channel matrix as neural network input, ignoring the two-dimensional spatial structure and multi-dimensional statistical characteristics of the channel parameters. This results in the model's inability to represent both global channel trends and local details. Furthermore, due to the complexity of communication environments, fixed network architectures lack dynamic adjustment mechanisms for different communication environments. When the actual deployment environment differs from the training data distribution, the model's generalization performance significantly degrades, making it difficult to meet the requirements for reliable communication in complex wireless environments. Therefore, there is an urgent need to provide a precoding model training method that can be adaptively adjusted to improve the accuracy of channel modeling and the adaptability of precoding optimization.

[0055] In view of this, please refer to Figures 1-6The embodiment of the present application provides a precoding model training method and training system, which visualizes channel features by constructing a color map, and adaptively trains the precoding model based on a deep learning neural network, effectively improving the generalization performance of the precoding model.

[0056] Please refer to Figure 1 , is a structural diagram of a precoding model training method provided in an embodiment of the present application, the precoding model training method comprising the following steps:

[0057] Step S1: Processing received signals acquired by a plurality of receiving antennas to obtain a covariance matrix corresponding to the received signals;

[0058] As an optional implementation, please refer to Figure 2 , is a flowchart of calculating the covariance matrix in the precoding model training method provided in an embodiment of the present application, wherein obtaining the covariance matrix corresponding to the received signal includes:

[0059] Step S10: converting the received signals obtained by the multiple receiving antennas into a received signal vector;

[0060] Step S11: Calculate the covariance matrix corresponding to the received signal according to the received signal vector.

[0061] In wireless communication systems, multiple receiving antennas are usually equipped. All receiving antennas will receive corresponding receiving signals at the same time. The required receiving signal vector is the combination of the receiving signals received by all receiving antennas at the same time.

[0062] After the received signal vector is obtained, the formula R=E[yy H ] to calculate the covariance matrix corresponding to the received signal; where y is the received signal vector, is the expectation operation, y H is the conjugate transposed vector of y, and R is the covariance matrix corresponding to the received signal.

[0063] Since actual received signals are often affected by multipath effects, interference, and noise, this application calculates the covariance matrix corresponding to the received signal to provide a preliminary description of the power distribution and spatial correlation of multiple antenna channels in a multipath channel. Furthermore, since this application processes the received signal in real time, it is applicable to stable modeling in time-varying channel environments.

[0064] Step S2: Calculating the bit error rate corresponding to the current channel based on the covariance matrix;

[0065] As an optional implementation, please refer to Figure 3 , is a flowchart of calculating the bit error rate in the precoding model training method provided in an embodiment of the present application, wherein the bit error rate corresponding to the current channel is calculated based on the covariance matrix, including:

[0066] Step S20: Obtaining a covariance matrix of the transmitted signal according to the covariance matrix of the interference signal, the covariance matrix of the noise, and the covariance matrix of the received signal;

[0067] Step S21: Calculating the effective signal power passing through the current channel, the interference signal power received by the receiving end, and the noise signal power received by the receiving end based on the covariance matrix of the transmitted signal;

[0068] Step S22: combining the effective signal power, the interference signal power, and the noise signal power to obtain a signal-to-noise ratio corresponding to the current signal;

[0069] Step S23: Calculate the bit error rate corresponding to the current channel through the signal-to-noise ratio.

[0070] Since the covariance matrix R corresponding to the received signal has been calculated in step S1, the covariance matrix also satisfies Ry=E[yy H ]=HRxH H +Ri+Rn; where Ri is the covariance matrix of the interference signal, Rn is the covariance matrix of the noise, and H is the channel matrix of the current channel. Therefore, the covariance matrix Rx of the transmitted signal can be inferred by combining the covariance matrix Ri of the interference signal, the covariance matrix Rn of the noise, and the channel matrix H of the current channel.

[0071] It should be noted that the channel matrix H is usually used to describe the channel characteristics between the transmitter and the receiver, and can be calculated by channel matrix estimation. For example, the channel matrix H is calculated by least square estimation or minimum mean square error estimation in related technologies. The covariance matrix Ri of the interference signal can be obtained by interference detection, and the covariance matrix of the noise can be obtained by the formula Rn= To calculate, the formula The noise power of each receiving antenna is directly measured, and I is the unit matrix; therefore, this application does not elaborate on the specific calculation and acquisition steps of the covariance matrix Ri of the above-mentioned interference signal, the covariance matrix Rn of the noise, and the channel matrix H of the current channel, and those skilled in the art should know this.

[0072] As an optional implementation, the covariance matrix Rx of the transmitted signal derived in step S20 can be used to further calculate the effective signal power Ps of the current channel, the interference signal power Pi received by the receiving end, and the noise signal power Pn received by the receiving end.

[0073] It should be noted that the calculation formula for the effective signal power Ps of the current channel is: Ps=Tr(HRxH H ); the calculation formula of the interference signal power Pi received by the receiving end is: Pi=Tr(Ri); the calculation formula of the noise signal power Pn received by the receiving end is: Pn=Tr(Rn); where, is the trace operation of the matrix.

[0074] Furthermore, the signal-to-interference plus noise ratio (SINR) corresponding to the current channel can be calculated by combining the effective signal power Ps, the interference signal power Pi, and the noise signal power Pn calculated above. Since the signal-to-interference plus noise ratio can reflect the degree to which the signal is affected by interference and noise during transmission, a higher signal-to-interference plus noise ratio means that the useful signal is relatively strong, the interference and noise are relatively weak, and the signal transmission quality is better. Therefore, the present application combines the effective signal power Ps of the current channel, the interference signal power Pi received by the receiving end, and the noise signal power Pn received by the receiving end to calculate the signal-to-noise ratio (SINR). The calculation formula of the signal-to-noise ratio (SINR) is as follows:

[0075]

[0076] At the same time, the bit error rate (BER) is a core indicator for measuring the reliability of a communication system. It directly reflects the error probability of signal transmission in the communication system and is closely related to the signal-to-interference-and-noise ratio (SINR). Therefore, after calculating the signal-to-interference-and-noise ratio (SINR), it is necessary to further calculate the bit error rate (BER) corresponding to the current channel in combination with the Q function.

[0077] In an optional embodiment, the Q function is defined as follows:

[0078]

[0079] Here, t is the specific value of the variable x, and the variable x is the signal-to-interference-plus-noise ratio (SINR) value calculated in step S22, ranging from 0 to ∞.

[0080] Furthermore, through the formula Calculate the bit error rate BER corresponding to the current channel.

[0081] In an embodiment of the present application, a relationship curve between the signal-to-noise ratio (SINR) and the bit error rate (BER) can be further drawn to represent the changing trend of the bit error rate under different channel conditions, thereby providing a verification benchmark for the subsequent verification of the precoding model. Of course, other methods that can intuitively display the relationship between the signal-to-noise ratio (SINR) and the bit error rate (BER) are also feasible, and the present application does not impose too many restrictions on this.

[0082] Step S3: Mapping the bit error rate with pixels in the color map to obtain a channel image corresponding to the current channel;

[0083] As an optional implementation manner, the above mapping of the bit error rate to the pixels in the color spectrum to obtain the channel image corresponding to the current channel includes:

[0084] The bit error rate is mapped one-to-one with pixels in the color spectrum through a color mapping function to obtain a channel image corresponding to the current channel, so that each pixel in the channel image represents the bit error rate of the corresponding position of the current channel.

[0085] As an optional implementation, the color spectrum includes at least any one of the RGB color spectrum, the CMYK color spectrum, the HSV color spectrum and the Lab color spectrum.

[0086] In the embodiment of the present application, after the bit error rate BER corresponding to the current channel is calculated, it is necessary to establish a corresponding relationship between the bit error rate BER and the color spectrum pixels.

[0087] First, it is necessary to determine a suitable color spectrum, such as one of the three-channel RGB color spectrum, CMYK color spectrum, HSV color spectrum and Lab color spectrum. It can be understood that the color spectrums provided above are all feasible, and this application does not impose too many restrictions on the specific color spectrum form selected.

[0088] Next, we will use a three-channel RGB color map as an example. After determining that the color map is a three-channel RGB color map, it is necessary to establish a correspondence between the bit error rate (BER) and the color map pixels. That is, the calculated bit error rate (BER) value is converted into pixel values ​​in the color map. The channel image is constructed using these pixel values, thereby intuitively displaying the bit error rate distribution of the current channel.

[0089] Specifically, the three-channel color map is as follows:

[0090]

[0091] Where (x, y) represents the coordinate point in the channel space; C(x, y) represents the pixel value of the coordinate point (x, y) in the three-channel color spectrum; BER(x, y) represents the BER value of the coordinate point (x, y); colormap is the color mapping function, which is as follows:

[0092]

[0093]

[0094] Preferably, the bit error rate BER needs to be normalized first. In the formula, t is the normalized bit error rate BER value, BERmin is the minimum value of the bit error rate BER, and BERmax is the maximum value of the bit error rate BER.

[0095] In the embodiment of the present application, each coordinate point (x, y) in the channel space is determined, its bit error rate (BER) value is calculated, and the bit error rate (BER) value is mapped to the RGB color space through colormap transformation to form a color spectrum matrix C:

[0096]

[0097] Among them, C R , C G , C B They are the values ​​of the red, green, and blue channels in the RGB color space respectively. By filling these bit error rate RGB values ​​as pixel values ​​into the corresponding positions of the channel image, the bit error rate BER value of each coordinate point (x, y) in the channel space can be mapped to the RGB color space to construct the required channel image.

[0098] It can be understood that the RGB three-channel color spectrum is a common color representation method. It represents various colors through the numerical combination of the three color channels of red (R), green (G), and blue (B). Therefore, by determining the values ​​of the red, green, and blue channels in the RGB color space corresponding to the bit error rate BER, and processing all coordinate points in the channel space in sequence, a complete channel image can be obtained; at the same time, the application of the color mapping function here can also ensure the accuracy and consistency of the mapping between the bit error rate BER value and the RGB pixel value, so that each pixel in the channel image can accurately represent the bit error rate at the corresponding channel space position, which is convenient for the subsequent precoding model to learn the relationship between the bit error rate BER and the channel characteristics.

[0099] Step S4: inputting the channel image into a precoding model, extracting and reducing the channel features in the channel image;

[0100] As an optional implementation, please refer to Figure 4, which is a flow chart for extracting channel features in the precoding model training method provided in an embodiment of the present application, wherein the channel image is input into the precoding model, and the channel features in the channel image are extracted and dimensionality reduced, including:

[0101] Step S40: inputting the channel image into the precoding model;

[0102] Step S41: extracting local channel features from the channel image;

[0103] Step S42: down-sample the extracted channel features to retain the local areas where the channel features have changed.

[0104] In this embodiment of the present application, since the bit error rate BER value has been mapped one-to-one to the pixels in the color map in step S3, the channel image corresponding to the current signal has been obtained. Therefore, it is necessary to send the channel image as input into the precoding model to extract and reduce the channel features in the channel image.

[0105] Furthermore, the size of the channel image needs to be adjusted to the size expected by the precoding model. The pixel values ​​in the channel image can be normalized so that their values ​​range from 0 to 1, and then the processed image data can be input into the precoding model.

[0106] Specifically, assuming that the size of the input channel image is H×W×3H, where H is the height, W is the width, and 3 represents the RGB channel, the pixel value X of the input channel is normalized to [0, 1] using the following formula:

[0107]

[0108] Among them, X is the current pixel value, X norm is the normalized pixel value, X min is the minimum pixel value, X max is the maximum pixel value.

[0109] This application normalizes the pixel value X of the input channel image to [0, 1] to ensure that the data format of the channel image is compatible with the precoding model, providing suitable input for subsequent feature extraction and dimensionality reduction processing, while ensuring that the data of different channel images have a unified scale, which is conducive to the training and convergence of the subsequent precoding model.

[0110] In an optional embodiment, the precoding model can be set in the form of a CNN convolutional neural network model. Typically, a CNN model includes a convolutional layer, a pooling layer, and a fully connected layer.

[0111] As an optional implementation, the present application extracts local channel features in the channel image through a convolutional layer, performs dimensionality reduction processing on the extracted channel features through a pooling layer, maps the channel features after dimensionality reduction processing into a precoding parameter vector through a fully connected layer, and reorganizes the parameter vector into a precoding matrix.

[0112] Specifically, the present application slides the convolution kernel on the channel image through the convolution layer, performs convolution operation on the local area of ​​the channel image, thereby extracting the local channel features in the channel image, and captures the local feature changes by extracting the local area of ​​the channel image.

[0113] For example, an area where the bit error rate (BER) value changes sharply can be extracted, and channel degradation may occur at this time; this application does not impose too many restrictions on the basis or specific circumstances for extracting local channel features, and those skilled in the art can adjust it according to actual needs.

[0114] In the CNN model, the convolution kernel is a small matrix. Each convolution kernel can be regarded as a feature detector used to detect specific features in the image. By sliding the convolution kernel row by row and column by column on the channel image, local features in the channel image can be extracted.

[0115] Specifically, assuming the convolution kernel size is k×k, the convolution operation can be defined as: ; Where W is the convolution kernel weight matrix, which is used to learn the local characteristic patterns of the channel, * represents the convolution operation, b is the bias term, and σ(.) is a nonlinear activation function (such as the ReLU activation function). The local characteristic change pattern of the bit error rate (BER) value distribution is extracted through the convolution layer, thereby learning the spatial distribution of the channel characteristics, reducing computational complexity and improving the generalization ability of the model.

[0116] Furthermore, after extracting the local channel features in the channel image through the convolution layer, it is necessary to use the pooling layer to downsample these channel features to retain the local areas where the channel features have changed.

[0117] Specifically, the dimensionality reduction operation can be defined as:

[0118]

[0119] Among them, Y poolIt represents the channel characteristics after downsampling, which retains the significant characteristics of the local area, such as the area where the bit error rate (BER) value changes sharply, the bit error rate (BER) value is extremely high, or the area where the bit error rate (BER) value is extremely low; max represents maximum pooling; i, j represent the pooling window size. This application retains the local area where the channel characteristics change through the pooling layer, reducing the size of the feature map and the complexity of subsequent calculations. At the same time, downsampling can reduce the impact of noise to a certain extent, thereby improving the robustness of the precoding model.

[0120] Step S5: Create a fully connected layer adapted to the current channel characteristics, and output a precoding matrix adapted to the current channel characteristics.

[0121] As an optional implementation, please refer to Figure 5 , which is a flowchart of outputting a precoding matrix in the precoding model training method provided in an embodiment of the present application. The above-mentioned creation of a fully connected layer adapted to the current channel characteristics includes:

[0122] Step S50: Freeze the convolutional layer weights during training;

[0123] Step S51: Create a fully connected layer, dynamically adjust the parameters of the fully connected layer according to the current channel environment, and optimize and train the parameter matrix of the newly created fully connected layer.

[0124] As an optional implementation manner, the outputting of the precoding matrix adapted to the current channel characteristics includes:

[0125] Step S52: performing a linear transformation on the channel features after the dimensionality reduction process by an activation function, and mapping the channel features into a precoding parameter vector;

[0126] Step S53: reorganize the precoding parameter vector into a precoding matrix, and output a precoding matrix adapted to the current channel characteristics.

[0127] Since the fully connected layer structure of the traditional CNN model is fixed and difficult to adapt to different channel environments, this application needs to redesign the fully connected layer so that it can be adaptively adjusted according to the current channel characteristics, and the local channel features obtained after the convolution layer and pooling layer operations are mapped to the global representation of the channel, and then the precoding matrix adapted to the current channel characteristics is output to meet the precoding prediction requirements under different channel environments.

[0128] In an embodiment of the present application, when recreating the fully connected layer, it is necessary to first freeze the weights of the convolutional layer during the training process to avoid changes in the weights of the convolutional layer during subsequent optimization training of the fully connected layer, thereby ensuring the stability of the extracted channel local features and focusing on the adaptive adjustment of the fully connected layer.

[0129] Furthermore, let the convolution output dimension be D and the recreated fully connected layer parameter matrix be W fc , the activation function uses ReLU, and the fully connected layer is defined as:

[0130]

[0131] Among them, W fc is the parameter matrix of the newly created fully connected layer, b fc is the bias of the fully connected layer; Y fc is the output of the fully connected layer, that is, the extracted precoding parameter vector, σ(.) is a nonlinear activation function (such as ReLU). Of course, it is also feasible to select other nonlinear activation functions, and this application does not impose too many restrictions on this.

[0132] This application can optimize and train the parameter matrix of the fully connected layer according to different channel environments, so that the fully connected layer can adjust its own parameters according to different channel environments and obtain fully connected layer parameters that are adapted to the current channel environment.

[0133] Furthermore, the channel features after dimensionality reduction need to be further processed before they can be converted into the parameters required by the precoding matrix. At this time, the introduction of the activation function can add nonlinear factors to the linear transformation, enabling the CNN model to learn more complex mapping relationships, thereby better mapping the channel features to the required precoding parameter vector.

[0134] Furthermore, the channel features after dimensionality reduction processing need to be input into the fully connected layer after training and optimization for linear transformation. The linear transformation here can be processed by activation functions such as ReLU and Sigmoid to obtain the required precoding parameter vector, but the precoding matrix usually has a specific structure and dimension. Therefore, the precoding parameter vector needs to be reorganized into a precoding matrix according to certain rules to meet the actual needs of the communication system; of course, the reorganization rules here can be set according to the specific situation, and this application does not impose further restrictions on this.

[0135] In an optional embodiment, the elements in the precoding parameter vector can be rearranged in a certain order to form the required precoding matrix according to the structure and dimensionality requirements of the precoding matrix, thereby outputting the precoding matrix adapted to the current channel characteristics. This precoding matrix serves as the precoding matrix of the transmitting end, which can optimize the structure of the transmitted signal, reduce the bit error rate, and improve the channel capacity and data transmission rate.

[0136] The precoding model training method provided in the present application first processes the received signal of the receiving antenna to obtain the covariance matrix corresponding to the received signal, calculates the bit error rate of the current channel based on the covariance matrix, maps the bit error rate to the pixels in the color spectrum to obtain the channel image, thereby converting the abstract bit error rate data into an intuitive image form, inputs the channel image into the precoding model to extract the channel features in the channel image and perform dimensionality reduction processing, thereby reducing data redundancy and computational complexity, creating a fully connected layer that can be adaptively adjusted according to the current channel characteristics, and outputting a precoding matrix adapted to the current channel characteristics, thereby improving the performance of the communication system in various channel environments.

[0137] Based on the above precoding model training method, please refer to Figure 6 , is a structural diagram of the precoding model training system provided in an embodiment of the present application, and the training system includes: a processing unit, a computing unit, a mapping unit, an extraction unit and an output unit.

[0138] The processing unit is used to process the received signals obtained by the plurality of receiving antennas to obtain a covariance matrix corresponding to the received signals;

[0139] The calculation unit is used to calculate the bit error rate corresponding to the current channel based on the covariance matrix;

[0140] The mapping unit is used to map the bit error rate with pixels in the color spectrum to obtain a channel image corresponding to the current channel;

[0141] The extraction unit is used to input the channel image into the precoding model, extract the channel features in the channel image and perform dimensionality reduction processing;

[0142] The output unit is used to create a fully connected layer adapted to the current channel characteristics and output a precoding matrix adapted to the current channel characteristics.

[0143] Compared with the existing technology, the present application can map the bit error rate to the pixels in the color map, characterize the channel features through image features, use the image structure to retain the distribution characteristics of the channel space, combine the convolution layer to extract local detail features, and the pooling layer to retain the characteristics of the feature change area, effectively capturing the local details and global trends of the channel space, avoiding the problem of structural information loss caused by directly using the vectorized channel matrix as the input of the neural network; by freezing the convolution layer weights and dynamically adjusting the fully connected layer parameters, the model can adaptively optimize the fully connected layer parameters according to the current channel environment, and generate a precoding matrix adapted to the current channel characteristics, thereby improving the generalization performance and adaptation accuracy of the precoding model.

[0144] For specific details about how the precoding model training system implements the above technical solution, please refer to the relevant description of the above-mentioned precoding model training method. This application will not go into details about this.

[0145] It can be understood that the various technical features of the above embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the various 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.

[0146] The above embodiments are merely exemplary embodiments for illustrating the principles of the present application, but the present application is not limited thereto. Those skilled in the art may make various modifications and improvements without departing from the spirit and substance of the present application, and such modifications and improvements are also considered to be within the scope of protection of the present application.

Claims

1. A precoding model training method, characterized in that: The following steps are involved: Processing received signals obtained by multiple receiving antennas to obtain a covariance matrix corresponding to the received signals; Calculating a bit error rate corresponding to the current channel based on the covariance matrix; Mapping the bit error rate to pixels in the color map using a color mapping function to obtain a channel image corresponding to the current channel, so that each pixel in the channel image represents the bit error rate at a corresponding position of the current channel; Inputting the channel image into a precoding model, extracting local channel features from the channel image, downsampling the extracted channel features, and retaining local areas where the channel features have changed; A fully connected layer adapted to the current channel characteristics is created, and a precoding matrix adapted to the current channel characteristics is output.

2. The precoding model training method according to claim 1, wherein: Obtaining a covariance matrix corresponding to the received signal includes: Converting received signals obtained by a plurality of receiving antennas into received signal vectors; A covariance matrix corresponding to the received signal is calculated based on the received signal vector.

3. The precoding model training method according to claim 2, wherein: The calculating the bit error rate corresponding to the current channel based on the covariance matrix includes: Obtaining a covariance matrix of a transmitted signal according to a covariance matrix of an interference signal, a covariance matrix of noise, and a covariance matrix of a received signal; Calculating the effective signal power passing through the current channel, the interference signal power received by the receiving end, and the noise signal power received by the receiving end based on the covariance matrix of the transmitted signal; Combining the effective signal power, the interference signal power and the noise signal power to obtain a signal-to-noise ratio corresponding to the current signal; The bit error rate corresponding to the current channel is obtained by calculating the signal-to-noise ratio.

4. The precoding model training method according to claim 1, wherein: The color spectrum includes at least any one of the RGB color spectrum, the CMYK color spectrum, the HSV color spectrum and the Lab color spectrum.

5. The precoding model training method according to claim 1, wherein: The creating of a fully connected layer adapted to the current channel characteristics includes: Freeze the convolutional layer weights during training; Create a fully connected layer, dynamically adjust the parameters of the fully connected layer according to the current channel environment, and optimize and train the parameter matrix of the newly created fully connected layer.

6. The precoding model training method according to claim 5, wherein: The outputting a precoding matrix adapted to the current channel characteristics includes: Performing a linear transformation on the channel features after dimensionality reduction processing by an activation function, mapping the channel features into a precoding parameter vector; The precoding parameter vector is reorganized into a precoding matrix, and a precoding matrix adapted to the current channel characteristics is output.

7. The precoding model training method according to claim 6, wherein: The precoding model is a CNN model, which includes a convolutional layer, a pooling layer and a fully connected layer; Wherein, the convolutional layer is used to extract local channel features in the channel image; The pooling layer is used to perform dimensionality reduction processing on the extracted channel features; The fully connected layer is used to map the channel features after dimensionality reduction processing into a precoding parameter vector, and reorganize the parameter vector into a precoding matrix.

8. A precoding model training system, characterized in that: include: a processing unit, configured to process received signals acquired by a plurality of receiving antennas to obtain a covariance matrix corresponding to the received signals; a calculation unit, configured to calculate a bit error rate corresponding to a current channel based on the covariance matrix; a mapping unit, configured to map the bit error rate to pixels in a color map using a color mapping function to obtain a channel image corresponding to the current channel, so that each pixel in the channel image represents the bit error rate at a corresponding position of the current channel; an extraction unit, configured to input the channel image into a precoding model, extract local channel features from the channel image, downsample the extracted channel features, and retain local areas where the channel features have changed; The output unit is used to create a fully connected layer adapted to the current channel characteristics and output a precoding matrix adapted to the current channel characteristics.

Images