Channel data generation method, electronic device, base station and readable storage medium

Abstract

The present application belongs to the field of wireless communications. Disclosed are a channel data generation method, an electronic device, a base station and a readable storage medium. The method comprises: acquiring a target noise signal; inputting the target noise signal into a target machine learning model to perform channel prediction processing, so as to obtain predicted channel data; and outputting the predicted channel data, wherein the target machine learning model is a trained machine learning model for channel prediction, the channel prediction processing comprises multi-dimensional channel feature extraction, and multi-dimensional channel features comprise spatial-domain channel features.

Description

Method for generating channel data, electronic device, base station and readable storage medium

[0001] Cross-references

[0002] This application claims priority to a Chinese patent application filed with the Patent Office of China on June 21, 2023, with application number 202310749420.9 and titled “Method, electronic device, base station and readable storage medium for generating channel data”, the entire contents of which are incorporated by reference into this application. Technical Field

[0003] The present application belongs to the field of wireless communications, and specifically relates to a method for generating channel data, an electronic device, a base station, and a readable storage medium. Background Art

[0004] Currently, in many situations, it is necessary to apply real channel data from wireless communication environments. However, due to limited conditions, the amount of real channel data obtained is often small. Therefore, it is extremely important to obtain data that is equivalent to real channel data.

[0005] Related technologies can use wireless channel prediction to obtain predicted channel data that is comparable to real channel data. However, these methods often have difficulty predicting complex channels and cannot better meet application requirements.

[0006] Summary of the Invention

[0007] The embodiments of the present application provide a method for generating channel data, an electronic device, a base station, and a readable storage medium, which can solve the problem that the related technologies for wireless channel prediction often have difficulty in predicting complex channels and cannot better meet application requirements.

[0008] In a first aspect, an embodiment of the present application provides a method for generating channel data, comprising: obtaining a target noise signal; inputting the target noise signal into a target machine learning model for channel prediction processing to obtain predicted channel data; and outputting the predicted channel data; wherein the target machine learning model is a machine learning model that has been trained for channel prediction; the channel prediction processing includes multidimensional channel feature extraction, and the multidimensional channel features include spatial channel features.

[0009] In a second aspect, an embodiment of the present application provides an electronic device, comprising: a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed, the steps of the method described in the first aspect are implemented.

[0010] In a third aspect, an embodiment of the present application provides a base station, which includes the electronic device described in the second aspect.

[0011] In a fourth aspect, an embodiment of the present application provides a readable storage medium, on which a program or instruction is stored, and when the program or instruction is executed, the steps of the method described in the first aspect are implemented. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1-1 is a schematic diagram of the structure of a generative adversarial network provided in an embodiment of the present application;

[0013] Figure 1-2 is a schematic diagram of a structural model of an example generator provided in an embodiment of the present application;

[0014] Figures 1-3 are schematic diagrams of a structural model of an example discriminator provided in an embodiment of the present application;

[0015] FIG2 is a flow chart of a method for generating channel data provided in an embodiment of the present application;

[0016] FIG3 is a flow chart of another method for generating channel data provided in an embodiment of the present application;

[0017] FIG4 is a flow chart of another method for generating channel data provided in an embodiment of the present application;

[0018] FIG5 is a flow chart of another method for generating channel data provided in an embodiment of the present application;

[0019] FIG6 is a flow chart of the training process of the target machine learning model in an embodiment of the present application;

[0020] FIG7 is a structural block diagram of a device for generating channel data provided by an embodiment of the present application;

[0021] FIG8 is a structural block diagram of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0022] The following will be combined with the accompanying drawings in the embodiments of the present application to clearly describe the technical solutions in the embodiments of the present application. Obviously, the embodiments described are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field are within the scope of protection of this application.

[0023] The terms "first," "second," and the like in the specification and claims of this application are used to distinguish similar objects, and are not used to describe a specific order or precedence. It should be understood that the terms used in this manner are interchangeable where appropriate, so that the embodiments of this application can be implemented in an order other than that illustrated or described herein, and that the objects distinguished by "first," "second," and the like are generally of the same type, and do not limit the number of objects; for example, the first object can be one or more. In addition, the term "and / or" in the specification and claims refers to at least one of the connected objects, and the character " / " generally indicates that the objects connected are in an "or" relationship.

[0024] As described in the background section, many current scenarios require the use of real channel data from wireless communication environments. However, due to limited conditions, the amount of real channel data available is often small. Related technologies typically use wireless channel prediction to generate predicted channel data that is comparable to real channel data. However, since these technologies often use mathematical modeling, they struggle to predict complex channels and thus fail to meet these requirements.

[0025] The embodiment of the present application provides a method for generating channel data, obtaining a target noise signal; inputting the target noise sample into a target machine learning model for channel prediction processing to obtain predicted channel data; wherein the channel prediction processing includes multi-dimensional channel feature extraction; outputting the predicted channel data; wherein the target machine learning model is a trained machine learning model; and the multi-dimensional channel features include spatial channel features. In this way, by extracting features from the multi-dimensional channel, the multi-dimensional channel characteristics and distribution characteristics can be reflected simultaneously. At the same time, since the multi-dimensional channel feature extraction includes spatial channel features, it can adapt to most communication environments, making the generated channel data more accurate, which can to a certain extent solve the problem that the existing technology cannot fully reflect the joint distribution characteristics of the multi-dimensional channel.

[0026] In an embodiment of the present application, feature extraction is performed on spatial domain channel features, time domain channel features, and / or frequency domain channel features through a two-dimensional convolutional layer, so as to achieve the purpose of simultaneously extracting spatial domain channel features, time domain channel features, and / or frequency domain channel features. When the discriminator is working, the influence of spatial domain channel features can be considered while considering the time domain channel features and / or frequency domain channel features, thereby ensuring that the predicted channel data obtained is more three-dimensional and does not lose spatial correlation, thereby better meeting application requirements.

[0027] In this embodiment of the present application, the parameters of the generator are frozen, and the discriminator is trained using the second training sample to obtain a trained target discriminator. After obtaining the trained target discriminator, the parameters of the target discriminator are frozen. By freezing the generator parameters and training the discriminator, the discriminator and generator are trained synchronously, allowing iteration of the entire generative adversarial network to improve the accuracy of the generator's prediction of channel data.

[0028] In an embodiment of the present application, the target machine learning model includes a target generator obtained by training based on a generative adversarial network, the target generator includes the multidimensional convolutional layer, and the predicted channel data is output by the target generator. The channel prediction method without using a generative adversarial network is limited to specific scenarios and has poor generalization, or is only applicable to small-batch channel environment prediction. Compared with the method without using a generative adversarial network, the method of generating channel data provided by the embodiment of the present application is suitable for most communication environments, realizes large-scale prediction of channel data, and can simultaneously learn the correlation between time domain, frequency domain and spatial domain, so that the generated channel data is more accurate and the time-frequency characteristics are more in line with the real channel, filling the existing research gap.

[0029] At the same time, in an embodiment of the present application, during the training process of the target machine learning model, a first training sample for the generator is obtained, where the first training sample is used to reflect the real channel data; the loss value of the first training sample is determined through the generator loss function; based on the loss value of the first training sample, the generator is iteratively trained to obtain a target generator; wherein the generator loss function is associated with the real channel data and the simulated channel data generated by the generator. In this way, through this specific explicitly constrained generator loss function, the convergence speed can be increased, making the model training time shorter.

[0030] The embodiments of the present application can be used to reconstruct a real wireless channel environment in the field. Based on this, combined with software and hardware simulation systems, the performance of new systems and methods in real field conditions can be studied. For example, the predicted channel data obtained can be used to verify the performance of solutions or algorithms in wireless network simulation platforms. The embodiments of the present application can also be deployed in base stations as channel predictors to sense the wireless environment in advance, improve product performance, or improve the performance of existing technologies in specific field conditions, identify problems, and locate them.

[0031] It should be understood that the discussion of the same noun or situation in the embodiments of the present application can be referenced before and after. That is, the discussion of a certain noun or situation in one embodiment can also be applied to the description of the same noun or situation in other embodiments, as long as there is no logical contradiction.

[0032] The method provided in the embodiment of the present application is described in detail below through specific embodiments and their application scenarios in conjunction with the accompanying drawings.

[0033] The method for generating channel data provided in an embodiment of the present application is applicable to scenarios in the field of wireless communications where future channel data is predicted and generated based on actual channel data. The method for generating channel data provided in an embodiment of the present application can be executed by a target device, where the target device can be an electronic device. The electronic device can be, for example, a terminal device such as a laptop or tablet, or a server.

[0034] The method for generating channel data provided in the embodiments of the present application may involve a target machine learning model, which is a pre-trained machine learning model for channel prediction. The target machine learning model can be any deep learning model suitable for channel prediction. For example, the target machine learning model can be a generative adversarial network. In the field of wireless communications, a generative adversarial network includes a generator and a discriminator.

[0035] Figure 1-1 is a structural diagram of a generative adversarial network 101 provided in an embodiment of the present application. As shown in Figure 1-1, during the training process, the generator learns the characteristics of the training set data and, under the guidance of the discriminator, fits the random noise distribution as much as possible to the real distribution of the training data, thereby generating similar data with the characteristics of the training set. The discriminator is responsible for distinguishing whether the input data is real data or false data generated by the generator and feeding it back to the generator. The two are trained alternately, and their capabilities are improved synchronously until the data generated by the generator can pass the discriminator and reach a certain balance with the capabilities of the discriminator. After the training is completed, the trained generative adversarial network 101 can be applied to the reasoning process. During the reasoning process, the input of the generator can be a noise signal, such as Gaussian white noise, and the output of the generator can be predicted channel data. In this way, the predicted channel data can be obtained using the generator 101 in the trained generative adversarial network.

[0036] Figure 1-2 is a schematic diagram of the structural model of an example generator 102 provided in an embodiment of the present application. As shown in Figure 1-2, the input of the generator 102 is noise, and its input is three layers of convolution-regularization-activation function layers. The kernel size of the convolution layer is 1 and the convolution layer can be a multi-dimensional convolution layer, such as a two-dimensional convolution layer. After each regularization layer, a rectified linear unit is used as the activation function. The generator 102 uses an encoder-decoder mode. The encoder consists of several convolution-regularization-activation function layers, with a maximum pooling layer at the end of the even layer and a loss layer added at the end; the decoder consists of several convolution-regularization-activation function layers, with a deconvolution layer before the odd layer and a loss layer added at the end. Finally, there is a residual connection between the encoder and the decoder, and the final channel data is generated through several convolution-regularization-activation function layers. It should be understood that the structure of the generator 102 shown in Figure 1-2 is only an example, and the embodiment of the present application does not specifically limit the number of layers of each part of the generator.

[0037] In an embodiment of the present application, Figures 1-3 are schematic diagrams of the structural model of an example discriminator 103 provided in an embodiment of the present application. As shown in Figures 1-3, the discriminator 103 in Figures 1-3 is composed of several convolution-regularization-activation function-maximum pooling-loss layers, several residual blocks and several fully connected layers. Among them, the convolutional network can be designed as a multi-dimensional convolution layer to simultaneously extract features in the time domain, frequency domain and spatial domain; the residual block can expand the learning field of the discriminator 103; finally, the two fully connected layers extract features and determine whether the input signal is a real signal. It should be understood that the embodiment of the present application does not specifically limit the number of layers of each part of the discriminator 103, and is only used as an example.

[0038] FIG2 is a flow chart of a method for generating channel data provided in an embodiment of the present application. As shown in FIG2 , the method for generating channel data provided in an embodiment of the present application may include the following steps:

[0039] Step 210: Acquire a target noise signal.

[0040] In the embodiment of the present application, the target noise signal can be any noise obtained, including artificial noise or noise caused by natural weather. The target noise signal can be Gaussian white noise.

[0041] In the embodiment of the present application, acquiring the target noise signal may specifically be that the electronic device receives a noise signal transmitted from an external device, or obtains an external noise signal through detection.

[0042] In the embodiment of the present application, once the target noise signal is obtained, channel data may be generated.

[0043] Step 220: Input the target noise sample into the target machine learning model for channel prediction processing to obtain predicted channel data; wherein, the channel prediction processing includes multi-dimensional channel feature extraction.

[0044] In an embodiment of the present application, the target machine learning model may be a model obtained after the training of the generative adversarial network model shown in Figure 1-1. The multidimensional channel feature extraction includes but is not limited to extracting any one of the three channel features, namely, time domain features, frequency domain features, and spatial domain features.

[0045] Step 230: output the predicted channel data; wherein, the target machine learning model is a trained machine learning model; and the multidimensional channel features include spatial channel features.

[0046] In an embodiment of the present application, a target noise signal is obtained; the target noise sample is input into a target machine learning model for channel prediction processing to obtain predicted channel data; wherein the channel prediction processing includes multi-dimensional channel feature extraction; the predicted channel data is output; wherein the target machine learning model is a trained machine learning model; the multi-dimensional channel features include spatial channel features. In this way, by extracting features from the multi-dimensional channel, the effect of dimensional expansion is achieved, and the multi-dimensional channel features and distribution characteristics can be reflected simultaneously. At the same time, since the multi-dimensional channel feature extraction includes spatial channel features, the influence of spatial channel features can be considered, ensuring that the obtained predicted channel data is more three-dimensional and does not lose spatial correlation, thereby better meeting application requirements.

[0047] In an embodiment of the present application, the target machine learning model may include a multidimensional convolution layer, which is used to extract spatial channel features and target channel features, and the target channel features are time domain channel features and / or frequency domain channel features.

[0048] In an embodiment of the present application, the target machine learning model includes a two-dimensional convolution layer, which is used to extract spatial domain channel features and time domain channel features, or the two-dimensional convolution layer is used to extract spatial domain channel features and frequency domain channel features.

[0049] In an embodiment of the present application, feature extraction is performed on spatial domain channel features, time domain channel features, and / or frequency domain channel features through a two-dimensional convolutional layer, so as to achieve the purpose of simultaneously extracting spatial domain channel features, time domain channel features, and / or frequency domain channel features. When the discriminator is working, the influence of spatial domain channel features can be considered while considering the time domain channel features and / or frequency domain channel features, thereby ensuring that the predicted channel data obtained is more three-dimensional and does not lose spatial correlation, thereby better meeting application requirements.

[0050] In an embodiment of the present application, the predicted channel data is used to verify the performance of a solution or algorithm in a simulation platform of a wireless network, or the predicted channel data is used to predict the channel condition of a base station.

[0051] In an embodiment of the present application, the target machine learning model can be a target generator obtained by training based on a generative adversarial network, the target generator includes the multidimensional convolutional layer, and the predicted channel data is output by the target generator. The channel prediction method without using a generative adversarial network is limited to specific scenarios and has poor generalization, or is only applicable to small-batch channel environment prediction. Compared with the method without using a generative adversarial network, the method of generating channel data provided in the embodiment of the present application outputs predicted channel data through a target generator obtained based on a generative adversarial network, which can adapt to most communication environments and realize large-scale prediction of channel data. It can also simultaneously learn the correlation between the time domain, frequency domain and spatial domain, so that the generated channel data is more accurate and the time-frequency characteristics are more in line with the real channel.

[0052] FIG3 is a flow chart of a method for generating channel data provided in an embodiment of the present application. As shown in FIG3 , the method for generating channel data provided in an embodiment of the present application may include the following steps:

[0053] Step 310: Obtain a first training sample for a generator, where the first training sample is used to reflect real channel data.

[0054] In an embodiment of the present application, before using a trained target machine learning model, the target machine learning model needs to be trained. First, a first training sample for training a generator is obtained, and the first training sample includes but is not limited to using a real noise signal.

[0055] The generator may use the generator model shown in Figure 1-1 or Figure 1-2. For detailed discussion, please refer to the discussion above regarding the generator model shown in Figure 1-1 or Figure 1-2. Once the first training sample for training the generator is obtained, the training of the generator may begin.

[0056] Step 320: Determine the loss value of the first training sample through the generator loss function.

[0057] In an embodiment of the present application, the loss value of the first training sample is determined by a generator loss function. An exemplary method for determining the loss value of the first training sample may be as follows: as shown in FIG1-1 or FIG1-2, the first training sample is input into the generator, and the loss value is obtained within the generator using the generator loss function.

[0058] In one embodiment of the present application, the loss function of the generator can be

[0059] Wherein, L is the generator loss function, A is the old part of the generator loss function, λ is the coefficient, H is the real channel data, To simulate channel data, k is an integer greater than 1, and N is a constant.

[0060] Step 330: Iteratively train the generator based on the loss value of the first training sample to obtain a target generator with target parameters; wherein the generator loss function is associated with the real channel data and the simulated channel data generated by the generator.

[0061] In an embodiment of the present application, after obtaining the loss value of the first training sample, the parameters of the model can be adjusted based on the loss value.

[0062] When the training is completed and a target generator with target parameters is obtained, the target generator can be used for the next step.

[0063] Step 340: Acquire a target noise signal.

[0064] In the embodiment of the present application, the target noise signal can be any noise obtained, including artificial noise or noise caused by natural weather. The target noise signal can be Gaussian white noise.

[0065] In the embodiment of the present application, acquiring the target noise signal may specifically be that the electronic device receives a noise signal transmitted from an external device, or obtains an external noise signal through detection.

[0066] In the embodiment of the present application, once the target noise signal is obtained, channel data may be generated.

[0067] Step 350: Input the target noise sample into a target generator for channel prediction processing to obtain predicted channel data; wherein the channel prediction processing includes multi-dimensional channel feature extraction.

[0068] In an embodiment of the present application, the target machine learning model may be a model obtained after the training of the generative adversarial network model shown in Figure 1-1. The multidimensional channel feature extraction includes but is not limited to extracting any one of the three channel features, namely, time domain features, frequency domain features, and spatial domain features.

[0069] Step 360: output the predicted channel data; wherein the multi-dimensional channel characteristics include spatial channel characteristics.

[0070] In an embodiment of the present application, during the training process of a target machine learning model, a first training sample for a generator is obtained, where the first training sample is used to reflect real channel data; a loss value of the first training sample is determined through a generator loss function; and based on the loss value of the first training sample, the generator is iteratively trained to obtain a target generator; wherein the generator loss function is associated with the real channel data and the simulated channel data generated by the generator. In this way, through this specific explicitly constrained generator loss function, the convergence speed can be increased, resulting in a shorter model training time.

[0071] FIG4 is a flow chart of a method for generating channel data provided in an embodiment of the present application. As shown in FIG4 , the method for generating channel data provided in an embodiment of the present application may include the following steps:

[0072] Step 410: Sample a target channel in a wireless channel environment to obtain real channel information, where the target channel includes a spatial domain channel and at least one of a time domain channel and a frequency domain channel.

[0073] Step 420: convert the real channel information into a target matrix, wherein the target matrix is ​​used to reflect the characteristics of the real channel information in the spatial channel, and at least one of the time channel and the frequency channel, and use the target matrix as the first training sample.

[0074] In the embodiment of the present application, the target matrix may be a three-dimensional matrix, which can simultaneously reflect the characteristics of the real channel information in the time domain, frequency domain, and spatial domain. Of course, the target matrix may also be a two-dimensional matrix.

[0075] Step 430: Determine the loss value of the first training sample through the generator loss function.

[0076] In an embodiment of the present application, the loss value of the first training sample is determined by a generator loss function. An exemplary method for determining the loss value of the first training sample may be as follows: as shown in FIG1-1 or FIG1-2, the first training sample is input into the generator, and the loss value is obtained within the generator using the generator loss function.

[0077] In one embodiment of the present application, the loss function of the generator can be

[0078] Wherein, L is the generator loss function, A is the old part of the generator loss function, λ is the coefficient, H is the real channel data, To simulate channel data, k is an integer greater than 1, and N is a constant.

[0079] Step 440: Iteratively train the generator based on the loss value of the first training sample to obtain a target generator with target parameters; wherein the generator loss function is associated with the real channel data and the simulated channel data generated by the generator.

[0080] In an embodiment of the present application, after obtaining the loss value of the first training sample, the parameters of the model can be adjusted based on the loss value.

[0081] When the training is completed and a target generator with target parameters is obtained, the target generator can be used for the next step.

[0082] Step 450: Acquire a target noise signal.

[0083] In the embodiment of the present application, the target noise signal can be any noise obtained, including artificial noise or noise caused by natural weather. The target noise signal can be Gaussian white noise.

[0084] In the embodiment of the present application, acquiring the target noise signal may specifically be that the electronic device receives a noise signal transmitted from an external device, or obtains an external noise signal through detection.

[0085] In the embodiment of the present application, once the target noise signal is obtained, channel data may be generated.

[0086] Step 460: Input the target noise sample into a target generator for channel prediction processing to obtain predicted channel data; wherein the channel prediction processing includes multi-dimensional channel feature extraction.

[0087] In an embodiment of the present application, the target machine learning model may be a model obtained after the training of the generative adversarial network model shown in Figure 1-1. The multidimensional channel feature extraction includes but is not limited to extracting any one of the three channel features, namely, time domain features, frequency domain features, and spatial domain features.

[0088] Step 470: output the predicted channel data; wherein the multi-dimensional channel characteristics include spatial channel characteristics.

[0089] In an embodiment of the present application, a target channel in a wireless channel environment is sampled to obtain real channel information, where the target channel includes a spatial channel and at least one of a time domain channel and a frequency domain channel; the real channel information is converted into a three-dimensional matrix, where the three-dimensional matrix is ​​used to reflect the characteristics of the real channel information in the time domain, frequency domain, and spatial domain, and the three-dimensional matrix is ​​used as a first training sample, where the first training sample is used to reflect real channel data. By sampling the target channel of the wireless channel and converting the sampled information into a three-dimensional matrix, information that simultaneously reflects the spatial channel, the time domain channel, and the frequency domain channel is obtained. This allows for simultaneous learning of the correlations in the time domain, frequency domain, and spatial domain, resulting in higher accuracy in the generated channel data and a time-frequency characteristic that is more consistent with the real channel.

[0090] FIG5 is a flow chart of a method for generating channel data provided in an embodiment of the present application. As shown in FIG5 , the method for generating channel data provided in an embodiment of the present application may include the following steps:

[0091] Step 510: Sample a target channel in a wireless channel environment to obtain real channel information, where the target channel includes a spatial domain channel and at least one of a time domain channel and a frequency domain channel.

[0092] Step 520: convert the real channel information into a multidimensional matrix, wherein the multidimensional matrix is ​​used to reflect the characteristics of the real channel information in the time domain, frequency domain and spatial domain, and use the multidimensional matrix as a first training sample, wherein the first training sample is used to reflect the real channel data.

[0093] In the embodiment of the present application, the target matrix may be a three-dimensional matrix, which can simultaneously reflect the characteristics of the real channel information in the time domain, frequency domain, and spatial domain. Of course, the target matrix may also be a two-dimensional matrix.

[0094] Step 530: Obtain a second training sample for training the discriminator.

[0095] In an embodiment of the present application, the second training sample includes the first training sample and the output result of the generator.

[0096] Step 540: Freeze the parameters of the generator, and train the discriminator using the second training sample to obtain a trained target discriminator.

[0097] In the embodiment of the present application, the parameters of the generator include but are not limited to the target parameters of the generator described above in Figures 3 and 4. This embodiment of the present application does not make specific requirements for this. After freezing the parameters of the generator, the discriminator is trained. The process of training the discriminator can refer to the training process of the generator as shown in Figures 3 and 4 above. The second training sample is input to the discriminator, and the loss value of the second training sample is obtained using the discriminator loss function. Based on the loss value of the second training sample, the parameters of the discriminator are obtained, and repeated iterations are performed to further obtain the trained target discriminator.

[0098] Step 550: After obtaining the trained target discriminator, freeze the parameters of the target discriminator.

[0099] In an embodiment of the present application, after obtaining the target discriminator, the parameters of the target discriminator are frozen so as to start training the generator to obtain the target generator.

[0100] Step 560: Determine the loss value of the first training sample through the generator loss function.

[0101] In an embodiment of the present application, the loss value of the first training sample is determined by a generator loss function. An exemplary method for determining the loss value of the first training sample may be as follows: as shown in FIG1-1 or FIG1-2, the first training sample is input into the generator, and the loss value is obtained within the generator using the generator loss function.

[0102] In one embodiment of the present application, the loss function of the generator can be

[0103] Wherein, L is the generator loss function, A is the old part of the generator loss function, λ is the coefficient, H is the real channel data, To simulate channel data, k is an integer greater than 1, and N is a constant.

[0104] Step 570: Iteratively train the generator based on the loss value of the first training sample to obtain a target generator with target parameters; wherein the generator loss function is associated with the real channel data and the simulated channel data generated by the generator.

[0105] In an embodiment of the present application, after obtaining the loss value of the first training sample, the parameters of the model can be adjusted based on the loss value.

[0106] When the training is completed and a target generator with target parameters is obtained, the target generator can be used for the next step.

[0107] Step 580: Acquire the target noise signal.

[0108] In the embodiment of the present application, the target noise signal can be any noise obtained, including artificial noise or noise caused by natural weather. The target noise signal can be Gaussian white noise.

[0109] In the embodiment of the present application, acquiring the target noise signal may specifically be that the electronic device receives a noise signal transmitted from an external device, or obtains an external noise signal through detection.

[0110] In the embodiment of the present application, once the target noise signal is obtained, channel data may be generated.

[0111] Step 590: Input the target noise sample into the target machine learning model for channel prediction processing to obtain predicted channel data; wherein, the channel prediction processing includes multi-dimensional channel feature extraction.

[0112] In an embodiment of the present application, the target machine learning model may be a model obtained after the training of the generative adversarial network model shown in Figure 1-1. The multidimensional channel feature extraction includes but is not limited to extracting any one of the three channel features, namely, time domain features, frequency domain features, and spatial domain features.

[0113] Step 595: output the predicted channel data; wherein the multi-dimensional channel characteristics include spatial channel characteristics.

[0114] In this embodiment of the present application, the parameters of the generator are frozen, and the discriminator is trained using the second training sample to obtain a trained target discriminator. After obtaining the trained target discriminator, the parameters of the target discriminator are frozen. By freezing the generator parameters and training the discriminator, the discriminator and generator are trained synchronously, allowing iteration of the entire generative adversarial network to improve the accuracy of the generator's prediction of channel data.

[0115] FIG6 is a flow chart of the training process of the target machine learning model in the embodiment of the present application. As shown in FIG6, the training process of the target machine learning model in the embodiment of the present application may include:

[0116] Step S1: Establish the generator model and discriminator model in the generative adversarial network and define the loss function and

[0117] In an embodiment of the present application, the process of constructing the generator model in step S1 can be as follows: construct a generator model, the input of the generator is noise, and its input is three layers of convolution-regularization-activation function layers, and the kernel size of the convolution layer is 1. After each regularization layer, a rectified linear unit is used as the activation function. The generator uses an encoder-decoder mode, the encoder consists of several convolution-regularization-activation function layers, with a maximum pooling layer at the end of the even layer, and a loss layer added at the end; the decoder consists of several convolution-regularization-activation function layers, with a deconvolution layer before the odd layer, and a loss layer added at the end. Finally, there is a residual connection between the encoder and the decoder, and the final channel data is generated through several convolution-regularization-activation function layers.

[0118] In this embodiment of the present application, the process of constructing the discriminator model in step S1 can be as follows: Constructing a discriminator model, which consists of several convolution-regularization-activation function-max pooling-dropout layers, several residual blocks, and several fully connected layers. The convolutional network needs to be designed as a two-dimensional convolutional layer to simultaneously extract features in the time domain, frequency domain, and spatial domain; the residual block expands the discriminator's learning field of view; finally, two fully connected layers extract features and determine whether the input signal is a real signal.

[0119] In this embodiment of the present application, the noise signal z can be defined as obeying the probability distribution p z , the generator and discriminator models are G and D respectively, and the generator parameter is θ G , using H and Represent the original channel data and the generated channel data respectively, use λ as the proportional coefficient, and define the generator loss function

[0120] in, Indicates finding a generator parameter set θ G , so that the following calculation formula is minimized; Indicates that the input noise obeys the noise distribution; G(z,H) indicates that when the current training sample is H, the input noise z, the channel data generated by the generator D(x,H) represents the probability that the input discriminator data x is judged to be true channel data when the current training sample is H. The value is 0-1, and 1-D(x,H) represents the probability that it is judged to be false channel data. log(1-D(G(z,H),H)) represents the probability that the generator generates The probability of being judged as fake by the discriminator; Represents the channel data generated by the generator and the difference norm of the true noise data H.

[0121] Then, the discriminator input noise signal z is defined, and the real channel data H obeys the probability distribution pdata , the discriminator parameter is θ D , the data input from the generator to the discriminator is x, and the discriminator loss function is defined

[0122] in, Indicates finding a discriminator parameter set θ D ; Indicates that the input data x of the discriminator follows the real channel data distribution, that is, x is randomly extracted from the real channel data set, and x and H both come from the same real channel data set; [logD(x,H)] represents the probability of judging the real data as true; E z [log(1-D(G(z,H),H))] represents the channel data generated by the generator The probability of being false.

[0123] Step S2: collect channel information H′, organize it into three dimensions: time domain, frequency domain, and spatial domain, normalize the data, and obtain the first training sample H.

[0124] In the process of collecting channel information, the channel in the wireless channel environment can be sampled at fixed time slots to collect the real channel information H′, which is organized into three dimensions: time domain, frequency domain, and spatial domain. The matrix is ​​obtained by the maximum-minimum linear normalization method. That is, first calculate the maximum amplitude H of all elements of the complex matrix H max With the minimum amplitude H min The difference is that all element amplitudes are subtracted from H min Then divide by this difference;

[0125] Step S3, initialize the generator parameters θ G , discriminator parameters θ D , learning rate η and learning batch B, so that θ G and θ D Follows Gaussian distribution Where μ and σ represent the mean and variance of the initialization parameters respectively.

[0126] Step S4, sampling a training batch m from the training data set H; sampling a training batch m from the data set H, m should be smaller than the amount of data in the entire data set, but the maximum value that can be obtained when the computing power supports it.

[0127] Step S5, freeze the generator parameters θ G ; Input the training batch m into the discriminator to start training the discriminator alone.

[0128] Step S6: Update the discriminator parameters through the Adaptive Moment Estimation (Adam) optimizer Through the loss function and Adam optimizer to update the discriminator parameters

[0129] Step S7, freeze the discriminator parameters θ D ; Input the noise signal z into the generator and start training the generator alone.

[0130] Step S8, update the generator parameters through the Adam optimizer Through the loss function And Adam optimizer updates the generator parameters

[0131] Step S9: Determine whether the generator has converged. If not, return to S5. If converged, end the process, and finally obtain a trained generative adversarial network and a predictively generated wireless channel.

[0132] The training process of FIG6 is specifically described below through two embodiments. It should be understood that the examples are for better illustrating the embodiments of the present application, rather than for limitation.

[0133] In one embodiment of the present application, the training process of the target machine learning model provided in the embodiment of the present application may include:

[0134] First, establish the generative adversarial network generator and discriminator models and define the loss function and

[0135] The generator model shown in Figure 1-1 is established. The generator input is noise, and it feeds into three layers of convolutional-regularization-activation function layers, with a kernel size of 1. After each regularization layer, a rectified linear unit is used as the activation function. The generator uses an encoder-decoder model. The encoder consists of six convolutional-regularization-activation function layers, with max pooling layers at the end of the second, fourth, and sixth layers, and a dropout layer at the end. The decoder consists of six convolutional-regularization-activation function layers, with deconvolution layers before the first, third, and fifth layers, and a dropout layer at the end. Finally, a residual connection is established between the encoder and decoder, and the final channel data is generated through three convolutional-regularization-activation function layers.

[0136] The discriminator model is constructed as shown in Figure 1-2. It consists of three convolution-regularization-activation function-max pooling-dropout layers, three residual blocks, and two fully connected layers. The convolutional network needs to be designed as a two-dimensional convolutional layer to simultaneously extract features in the time, frequency, and spatial domains. The residual block expands the discriminator's learning field of view. Finally, the two fully connected layers extract features and determine whether the input signal is real.

[0137] For the generator, the noise signal z is defined to obey the probability distribution p z , the generator and discriminator models are G and D respectively, and the generator parameter is θ G , using H and Represent the original channel data and the generated channel data respectively, use λ as the proportional coefficient, and define the generator loss function The value of λ is 0.5.

[0138] in, Indicates finding a generator parameter set θ G , so that the following calculation formula is minimized; Indicates that the input noise obeys the noise distribution; G(z,H) indicates that when the current training sample is H, the input noise z, the channel data generated by the generator D(x,H) represents the probability that the input discriminator data x is judged to be true channel data when the current training sample is H. The value is 0-1, and 1-D(x,H) represents the probability that it is judged to be false channel data. log(1-D(G(z,H),H)) represents the probability that the generator generates The probability of being judged as fake by the discriminator; Represents the channel data generated by the generator and the difference norm of the true noise data H.

[0139] For the discriminator, the discriminator input noise signal z is defined, and the real channel data H obeys the probability distribution p data , the discriminator parameter is θ D , the data input from the generator to the discriminator is x, and the discriminator loss function is defined

[0140] in, Indicates finding a discriminator parameter set θ D ; Indicates that the input data x of the discriminator follows the real channel data distribution, that is, x is randomly extracted from the real channel data set, and x and H both come from the same real channel data set; [logD(x,H)] represents the probability of judging the real data as true; E z [log(1-D(G(z,H),H))] represents the channel data generated by the generator The probability of being false.

[0141] All frequency domain channels in the wireless channel environment are sampled every fixed time slot to collect the real channel information H′, which is organized into three dimensions: time domain, frequency domain, and spatial domain. The matrix is ​​obtained by the maximum-minimum linear normalization method. That is, first calculate the maximum amplitude H of all elements of the complex matrix H max With the minimum amplitude H min The difference is that all element amplitudes are subtracted from H min Then divide it by the difference to get the training sample H.

[0142] Initialize the generator parameters θ G , discriminator parameters θ D , learning rate η and learning batch B, so that θ G and θ D Follows Gaussian distribution Where μ and σ represent the mean and variance of the initialization parameters respectively.

[0143] Sample training batch m from the dataset H. m should be smaller than the amount of data in the entire dataset, but the maximum value that can be obtained if the computing power supports it.

[0144] Freeze generator parameters θ G , input the training batch m into the discriminator to start training the discriminator alone.

[0145] Through the loss function and Adam optimizer to update the discriminator parameters

[0146] Freeze the discriminator parameters θ D , input the noise signal z into the generator and start training the generator separately.

[0147] Through the loss function And Adam optimizer updates the generator parameters

[0148] Determine the generator parameter θ G Whether convergence is achieved, if not, return to the discussion shown in step S5 of Figure 6, if convergence is achieved, then end. Finally, a trained generative adversarial network and a wireless channel generated predictively by the target generator are obtained.

[0149] In another embodiment of the present application, in one embodiment of the present application, the training process of the target machine learning model provided by the embodiment of the present application may include:

[0150] Build a generative adversarial network generator and discriminator model and define the loss function and

[0151] Among them, a generator model is established. The input of the generator is noise, and its input is five layers of convolution-regularization-activation function layers. The kernel size of the convolution layer is 1. After each regularization layer, a rectified linear unit is used as the activation function. The generator uses an encoder-decoder mode. The encoder consists of eight convolution-regularization-activation function layers, with maximum pooling layers at the end of the 2nd, 4th, 6th and 8th layers, and a dropout layer is added at the end; the decoder consists of eight convolution-regularization-activation function layers, with deconvolution layers before the 1st, 3rd, 5th and 7th layers, and a dropout layer is added at the end. Finally, there is a residual connection between the encoder and decoder, and the final channel data is generated through three convolution-regularization-activation function layers.

[0152] The discriminator model is constructed, consisting of three convolution-regularization-activation function-max pooling-dropout layers, three residual blocks, and two fully connected layers. The convolutional network requires two-dimensional convolutional layers to simultaneously extract features in the time, frequency, and spatial domains. The residual blocks expand the discriminator's learning field of view. Finally, the two fully connected layers extract features and determine whether the input signal is real.

[0153] For the generator, the noise signal z is defined to obey the probability distribution p z , the generator and discriminator models are G and D respectively, and the generator parameter is θ G , using H and Represent the original channel data and the generated channel data respectively, use λ as the proportional coefficient, and define the generator loss function The value of λ is 0.7.

[0154] in, Indicates finding a generator parameter set θ G , so that the following calculation formula is minimized; Indicates that the input noise obeys the noise distribution; G(z,H) indicates that when the current training sample is H, the input noise z, the channel data generated by the generator D(x,H) represents the probability that the input discriminator data x is judged to be true channel data when the current training sample is H. The value is 0-1, and 1-D(x,H) represents the probability that it is judged to be false channel data. log(1-D(G(z,H),H)) represents the probability that the generator generates The probability of being judged as fake by the discriminator; Represents the channel data generated by the generator and the difference norm of the true noise data H.

[0155] For the discriminator, the discriminator input noise signal z is defined, and the real channel data H obeys the probability distribution p data , the discriminator parameter is θ D, the data input from the generator to the discriminator is x, and the discriminator loss function is defined

[0156] in, Indicates finding a discriminator parameter set θ D ; Indicates that the input data x of the discriminator follows the real channel data distribution, that is, x is randomly extracted from the real channel data set, and x and H both come from the same real channel data set; [logD(x,H)] represents the probability of judging the real data as true; E z [log(1-D(G(z,H),H))] represents the channel data generated by the generator The probability of being false.

[0157] The frequency domain channel in the wireless channel environment is sampled every fixed time slot to collect the real channel information H′, which is organized into three dimensions: time domain, frequency domain, and spatial domain. The matrix is ​​obtained by the maximum-minimum linear normalization method. That is, first calculate the maximum amplitude H of all elements of the complex matrix H max With the minimum amplitude H min The difference is that all element amplitudes are subtracted from H min Then divide it by the difference to get the training sample H.

[0158] Initialize the generator parameters θ G , discriminator parameters θ D , learning rate η and learning batch B, so that θ G and θ D Follows Gaussian distribution Where μ and σ represent the mean and variance of the initialization parameters respectively.

[0159] Sample training batch m from the dataset H. m should be smaller than the amount of data in the entire dataset, but the maximum value that can be obtained if the computing power supports it.

[0160] Freeze generator parameters θ G , input the training batch m into the discriminator to start training the discriminator alone.

[0161] Through the loss function and Adam optimizer to update the discriminator parameters

[0162] Freeze the discriminator parameters θ D , input the noise signal z into the generator and start training the generator separately.

[0163] Through the loss function And Adam optimizer updates the generator parameters

[0164] Determine the generator parameter θ G Whether convergence is achieved, if not, return to the discussion shown in step S5 of Figure 6, if convergence is achieved, then end. Finally, the trained generative adversarial network and the target generator are obtained to perform predictive generation of a wireless channel.

[0165] FIG7 is a block diagram of a device for generating channel data according to an embodiment of the present application. Referring to FIG7 , the device 700 for generating channel data according to an embodiment of the present application includes:

[0166] The acquisition module 710 is configured to acquire a target noise signal.

[0167] The processing module 720 is used to input the target noise sample into the target machine learning model for channel prediction processing to obtain predicted channel data; wherein, the channel prediction processing includes multi-dimensional channel feature extraction.

[0168] The output module 730 is used to output the predicted channel data; wherein the target machine learning model is a trained machine learning model; and the multidimensional channel features include spatial channel features.

[0169] In an embodiment of the present application, a target noise signal is obtained; the target noise sample is input into a target machine learning model for channel prediction processing to obtain predicted channel data; wherein the channel prediction processing includes multi-dimensional channel feature extraction; the predicted channel data is output; wherein the target machine learning model is a trained machine learning model; the multi-dimensional channel features include spatial channel features. In this way, by extracting features from the multi-dimensional channel, the effect of dimensional expansion is achieved, and the multi-dimensional channel features and distribution characteristics can be reflected simultaneously. At the same time, since the multi-dimensional channel feature extraction includes spatial channel features, it can adapt to most communication environments, making the generated channel data more accurate, which can to a certain extent solve the problem that the existing technology cannot fully reflect the joint distribution characteristics of the multi-dimensional channel.

[0170] In one embodiment of the present application, the target machine learning model includes a multidimensional convolution layer, which is used to extract spatial channel features and target channel features, and the target channel features are time domain channel features and / or frequency domain channel features.

[0171] In one embodiment of the present application, the target machine learning model includes a two-dimensional convolution layer, which is used to extract spatial domain channel features and time domain channel features, or the two-dimensional convolution layer is used to extract spatial domain channel features and frequency domain channel features.

[0172] In one embodiment of the present application, the target machine learning model includes a target generator trained based on a generative adversarial network.

[0173] In one embodiment of the present application, in the target machine learning training model process, the processing module 720 is also used to obtain a first training sample for the generator, where the first training sample is used to reflect real channel data; the loss value of the first training sample is determined through a generator loss function; based on the loss value of the first training sample, the generator is iteratively trained to obtain a target generator with target parameters; wherein the generator loss function is associated with the real channel data and the simulated channel data generated by the generator.

[0174] In one embodiment of the present application, in the process of obtaining the first training sample for the generator, the processing module 720 is also used to sample the target channel in the wireless channel environment to obtain real channel information, where the target channel includes a spatial domain channel and at least one of a time domain channel and a frequency domain channel; convert the real channel information into a three-dimensional matrix, where the three-dimensional matrix is ​​used to reflect the characteristics of the real channel information in the time domain, frequency domain and spatial domain, and use the three-dimensional matrix as the first training sample.

[0175] In one embodiment of the present application, during the target machine learning training model process, the processing module 720 is further used to obtain a second training sample for training the discriminator; freeze the parameters of the generator, train the discriminator through the second training sample, and obtain a trained target discriminator; after obtaining the trained target discriminator, freeze the parameters of the target discriminator.

[0176] In addition, as shown in FIG8 , an embodiment of the present application further provides an electronic device 800 , which may be various types of computers, etc. The electronic device 800 includes a processor 810 and a memory 820 , wherein the memory 820 stores programs or instructions. When the programs or instructions are executed by the processor 810 , the steps of any of the methods described above are implemented, such as the steps of the method for generating channel data shown in FIG2 to FIG5 .

[0177] An embodiment of the present application further provides a base station, which includes the electronic device provided by any embodiment of the present application, and the electronic device can execute the method for generating channel data provided by any embodiment of the present application.

[0178] An embodiment of the present application also provides a readable storage medium, on which a program or instruction is stored. When the program or instruction is executed by a processor, the various processes of the above-mentioned method embodiment are implemented and the same technical effect can be achieved. To avoid repetition, it will not be repeated here.

[0179] The processor is the processor in the electronic device described in the above embodiment. The readable storage medium includes a computer readable storage medium, such as a computer read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0180] An embodiment of the present application further provides a chip, which includes a processor and a communication interface, wherein the communication interface is coupled to the processor, and the processor is used to run programs or instructions to implement the various processes of the above-mentioned method embodiment and achieve the same technical effect. To avoid repetition, it will not be repeated here.

[0181] An embodiment of the present application provides a computer program product, which is stored in a storage medium. The program product is executed by at least one processor to implement the various processes of the above-mentioned method embodiment and can achieve the same technical effect. To avoid repetition, it will not be repeated here.

[0182] It should be noted that, in this article, the terms "comprise", "include" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the statement "comprises a ..." does not exclude the presence of other identical elements in the process, method, article or device comprising the element. In addition, it should be noted that the scope of the methods and devices in the embodiments of the present application is not limited to performing functions in the order shown or discussed, and may also include performing functions in a substantially simultaneous manner or in the opposite order according to the functions involved. For example, the described method may be performed in an order different from that described, and various steps may also be added, omitted, or combined. In addition, the features described with reference to certain examples may be combined in other examples.

[0183] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus the necessary general hardware platform, and of course can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art can be embodied in the form of a computer software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), including a number of instructions for enabling a terminal (which can be a mobile phone, computer, server, or network device, etc.) to execute the methods described in each embodiment of the present application.

[0184] The embodiments of the present application are described above in conjunction with the accompanying drawings, but the present application is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of this application, ordinary technicians in this field can also make many forms without departing from the purpose of this application and the scope of protection of the claims, all of which are within the protection of this application.

Claims

1. A method for generating channel data, comprising: Acquire target noise signal; Inputting the target noise signal into a target machine learning model for channel prediction processing to obtain predicted channel data; outputting the predicted channel data; Among them, the target machine learning model is a machine learning model that has been trained for channel prediction; the channel prediction processing includes multi-dimensional channel feature extraction, and the multi-dimensional channel features include spatial domain channel features.

2. The method according to claim 1, wherein: The target machine learning model includes a multidimensional convolution layer, which is used to extract spatial domain channel features and target channel features, and the target channel features include at least one of time domain channel features and frequency domain channel features.

3. The method according to claim 2, wherein: The multidimensional convolution layer includes a two-dimensional convolution layer, and the two-dimensional convolution layer is used to extract spatial domain channel features and time domain channel features, or the two-dimensional convolution layer is used to extract spatial domain channel features and frequency domain channel features.

4. The method according to claim 2, wherein: The target machine learning model includes a target generator obtained through training based on a generative adversarial network, the target generator includes the multi-dimensional convolutional layer, and the predicted channel data is output by the target generator.

5. The method according to claim 4, wherein: The training process of the target machine learning model includes: Acquire a first training sample for a generator, wherein the first training sample is used to reflect real channel data; Determining a loss value of the first training sample by using a generator loss function; Iteratively training the generator based on the loss value of the first training sample to obtain a target generator; Wherein, the generator loss function is associated with real channel data and simulated channel data generated by the generator.

6. The method according to claim 5, wherein: The generator loss function is expressed as follows: Among them, L is the generator loss function, A is the old part of the generator loss function, λ is the coefficient, H is the real channel data, H^ is the simulated channel data, k is an integer greater than 1, and N is a constant.

7. The method according to claim 5, wherein: The obtaining of a first training sample for a generator comprises: Sampling the target channel in the wireless channel environment to obtain real channel information, wherein the target channel includes a spatial domain channel, and at least one of a time domain channel and a frequency domain channel; The real channel information is converted into a target matrix, where the target matrix is ​​used to reflect the characteristics of the real channel information in a spatial domain channel, and at least one of a time domain channel and a frequency domain channel, and the target matrix is ​​used as a first training sample.

8. The method according to claim 5, wherein: Before determining the loss value of the first training sample by the generator loss function, the training process of the target machine learning model further includes: Obtaining a second training sample for training the discriminator; Freezing the parameters of the generator, and training the discriminator through the second training sample to obtain a trained target discriminator; After obtaining the trained target discriminator, the parameters of the target discriminator are frozen.

9. The method according to claim 1, wherein: The predicted channel data is used to verify the performance of a solution or algorithm in a simulation platform of a wireless network, or the predicted channel data is used to predict a channel condition of a base station.

10. An electronic device, comprising a processor and a memory, wherein the memory stores programs or instructions, and when the programs or instructions are executed by the processor, the steps of the method according to any one of claims 1 to 9 are implemented.

11. A base station, comprising the electronic device according to claim 10.

12. A readable storage medium storing a program or an instruction, wherein the program or the instruction, when executed, implements the steps of the method according to any one of claims 1 to 9.

Images