A Power Amplifier Linearization Method Based on Conditional Generative Adversarial Network

Abstract

The present application provides a power amplifier linearization method based on conditional generative adversarial network, which solves the technical problem that the modeling performance of digital pre-distortion model is difficult to improve in the existing power amplifier linearization technology. The method comprises the following steps: obtaining randomly generated input data X(n) and actual output data Y(n) of the power amplifier; constructing a CGAN-DPD model, which includes a nonlinear distortion degree extraction network C net , Generative Adversarial Network; Generative Adversarial Network includes Generator Network G net , Discriminator Network D net ; Train the nonlinear distortion extraction network C net , generator network G net and the discriminator network D net ; Randomly generate a transmission signal and input it into the trained CGAN-DPD model. The input data can obtain a linearized output after passing through a power amplifier. The application is applied to the field of communication technology.

Description

Technical Field

[0001] The present application relates to a power amplifier linearization modeling method, and more specifically, to a power amplifier linearization method based on a conditional generative adversarial network. Background Art

[0002] The power amplifier (PA) is responsible for amplifying the RF signal of the base station to a sufficient power level for efficient transmission to the mobile device. However, the nonlinear characteristics of the PA may cause a variety of adverse effects in 5G communications, so 5G communication system designers usually use advanced PA linearization techniques, such as digital pre-distortion (DPD), to ensure that the required linearity is provided while maintaining high efficiency, thereby supporting the high rate and low latency requirements of 5G communications.

[0003] Traditional DPD models, including the original Volterra model and several variations of the Volterra model, have been widely used in the linearization correction of broadband PAs. However, the high correlation between the polynomial bases in these models makes it difficult to improve their modeling performance. Summary of the invention

[0004] In order to solve the problem that the modeling performance of the digital predistortion model is difficult to improve in the existing power amplifier linearization technology, the technical solution adopted in this application is: to provide a power amplifier linearization method based on a conditional generative adversarial network, comprising the following steps:

[0005] Obtain randomly generated input data X(n) and actual output data Y(n) of the power amplifier;

[0006] Construct a conditional generative adversarial network pre-distortion model CGAN-DPD, which includes a nonlinear distortion degree extraction network C net , Generative Adversarial Network; Generative Adversarial Network includes Generator Network G net , Discriminator Network D net ;

[0007] Training the nonlinear distortion extraction network C net , generator network G net and the discriminator network D net ;

[0008] The randomly generated transmission signal is input into the trained conditional generative adversarial network pre-distortion model CGAN-DPD. The input data can obtain a linearized output after passing through the power amplifier.

[0009] Preferably, the nonlinear distortion degree extraction network C net The specific training method is as follows:

[0010] (1) Based on the training data pair, obtain the actual distortion degree data D(n) of the power amplifier;

[0011] (2) Convert the two-dimensional data of X(n) Input nonlinear distortion extraction network C net , calculate the loss function ;

[0012] (3) Determine the loss function Is it greater than the threshold? If so, execute step (4); otherwise, end the loop;

[0013] (4) Calculate the loss function Extract the partial derivatives of network parameters with respect to the degree of nonlinear distortion ,in C net The parameters in the k-1th cycle, C net Find the loss function for the parameters of the kth cycle The local minimum of the nonlinear distortion extraction network is updated .

[0014] Preferably, the generator network G net and the discriminator network D net The specific training method is as follows:

[0015] (1) Convert the two-dimensional data of Y(n) Input nonlinear distortion extraction network C net , and obtain the distortion level prediction result C;

[0016] (2) Combined with the distortion level prediction result C, we get , input generator network G net , output false data result X gen ;

[0017] (3) Convert the two-dimensional data of X(n) , generator network G net The output result X gen Input to the discriminator network D net , we get Y real , Y gen ;

[0018] Among them, Y real is the discriminator network D net For two-dimensional input data The judgment result, Y gen is the discriminator network D net For false data result Xgen The judgment result of

[0019] (4) Calculate the generator network G net The loss function G loss and G NMSE , Discriminator Network D net The loss function D loss ;

[0020] According to G loss Using the forward propagation of the gradient descent algorithm, the generator network G net The parameters of G are updated. NMSE Used to determine whether the model training reaches the threshold; if G NMSE If it is greater than the threshold, then execute steps (5) and (6) to update the generator network G. net , Discriminator Network D net Otherwise, end the loop;

[0021] (5) Calculate the discriminator network loss function D loss For the discriminator network parameters The gradient of ,in D net The parameters in the i-1th loop, D net Parameters in the i-th loop;

[0022] (6) Calculate the generator network loss function G loss For the generator network parameters The gradient of ,in G net The parameters in the i-1th loop, G net Parameters for the i-th loop.

[0023] Preferably, in step (1), the training data pair is actual output data Y(n) and real distortion degree data D(n);

[0024] ;

[0025] Among them, Y(n) is the output data when input X(n), and K is the gain of the power amplifier.

[0026] Preferably, step (2) is specifically implemented as follows:

[0027] As the degree of nonlinear distortion extraction network C net Input, output eigenvalue results C1(n), C2(n);

[0028] Calculate the distance between the actual distortion data D(n) of the power amplifier and the output eigenvalue results C1(n) and C2(n) to obtain the loss function ;

[0029] ;

[0030] in, The loss function of the network is extracted for the degree of nonlinear distortion, is the amplitude component of D(n), is the phase component of D(n), C1(n) is the nonlinear distortion extraction network C net The real deviation of the output eigenvalue result, C2(n) is the nonlinear distortion degree extraction network C net Output the imaginary deviation of the eigenvalue result.

[0031] Preferably, in step (2), splicing means, Combined with C, we get ;

[0032] ;

[0033] Among them, C3(n) and C4(n) are C net The two eigenvalues ​​of the output, C refers to C net The output result is:

[0034] Preferably, in step (4), the loss function is calculated as follows:

[0035] ;

[0036] ;

[0037] .

[0038] Preferably, the nonlinear distortion degree extraction network C net It includes input layer, hidden layer 1, hidden layer 2 and output layer; hidden layer 1 is composed of a fully connected layer and an activation layer of multiple neurons, and hidden layer 2 is composed of a fully connected layer and an activation layer of 2 neurons; the activation layer selects the ReLU function.

[0039] Preferably, the generator network G net It consists of two fully connected layers, two activation layers and an output layer. The activation layer uses the ReLU function. The output layer consists of 2 neurons. The output I DPD and Q DPD Two signal paths.

[0040] Preferably, the discriminator network Dnet is composed of four convolutional layers, four activation layers, and one fully connected layer.

[0041] The beneficial effects of the present invention are that the present invention applies the conditional generative adversarial network for the first time to perform pre-distortion processing on the power amplifier. The conditional generative adversarial network is a special deep learning model. Compared with traditional deep neural networks and shallow neural networks, CGAN has stronger generation capabilities and better feature learning effects, and can more accurately describe the characteristics of nonlinear systems. The present invention first builds a power amplifier nonlinear distortion degree extraction network to extract the power amplifier distortion level prediction results, and applies them to the generator network. By introducing the distortion level prediction results, CGAN can control the specific properties of the generated samples. This control power enables the model to more accurately reflect the input conditional information during generation, providing strong support for various application scenarios. The conditional generative adversarial network pre-distortion model CGAN-DPD constructed by the present invention includes a nonlinear distortion degree extraction network C net , Generative Adversarial Network; Generative Adversarial Network includes Generator Network G net , discriminator network Dnet. After simulation experiments, the trained conditional generative adversarial network pre-distortion model CGAN-DPD is superior to GAN-DPD in ACPR, NMSE, and EVM of Doherty PA for 100MHz broadband signal, and is superior to NN-DPD by -2.86dB, -3.379dB, and -1.217dB, respectively, and superior to MP-DPD by -5.333dB, -6.725dB, and -2.9323dB, respectively. It is superior to the existing pre-distortion models in all three evaluation indicators. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. 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 paying creative labor.

[0043] Figure 1 A schematic flow chart of a power amplifier linearization method based on a conditional generative adversarial network provided in one embodiment of the present application;

[0044] Figure 2 A schematic diagram of a network structure for extracting the degree of nonlinear distortion provided in an embodiment of the present application;

[0045] Figure 3 A schematic diagram of a generator network structure provided in one embodiment of the present application;

[0046] Figure 4 A schematic diagram of a discriminator network structure provided in an embodiment of the present application;

[0047] Figure 5 A schematic diagram of simulation results provided in one embodiment of the present application. DETAILED DESCRIPTION

[0048] In order to make the technical problems, technical solutions and beneficial effects to be solved by the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0049] See also Figure 1 , is a flow chart of a power amplifier linearization method based on a conditional generative adversarial network provided in an embodiment of the present application. For the sake of convenience, only the part related to the present embodiment is shown, which is described in detail as follows:

[0050] In one embodiment, a power amplifier linearization method based on a conditional generative adversarial network comprises the steps of:

[0051] S101, obtaining input data X(n) and actual output data Y(n) of a power amplifier;

[0052] Specifically, the input data X(n) is 300 frames of randomly generated data, and the output data Y(n) is the actual output data corresponding to the measured power amplifier. Preferably, X(n) is an OFDM time domain signal with a bandwidth of 100Mhz that complies with the 5G communication standard.

[0053] S102: Construct a conditional generative adversarial network pre-distortion model (CGAN-DPD), which includes a nonlinear distortion degree extraction network C net , Generative Adversarial Network; Generative Adversarial Network includes Generator Network G net , Discriminator Network D net ; This step and step S101 are interchangeable in order.

[0054] Specifically, the nonlinear distortion degree extraction network C net , the generator network and the discriminator network structures are as follows:

[0055] See also Figure 2 , is the nonlinear distortion degree extraction network C net Structural diagram. Nonlinear distortion degree extraction network C net It includes input layer, hidden layer 1, hidden layer 2 and output layer. The input layer is used to input two-dimensional input data, hidden layer 1 is composed of a fully connected layer and an activation layer of multiple neurons, and hidden layer 2 is composed of a fully connected layer and an activation layer of 2 neurons; the activation layer uses the ReLU function.

[0056] See also Figure 3 , is a schematic diagram of the generator network structure. Generator network G net It consists of two fully connected layers, two activation layers, an input layer and an output layer. The activation layer uses the ReLU function. The output layer consists of 2 neurons. The output I DPD and Q DPD Two signal paths.

[0057] Finally, see Figure 4 , is a schematic diagram of the discriminator network structure. Discriminator network D net It consists of a fully connected layer, four activation layers (activation layer 1, activation layer 2, activation layer 3, and activation layer 4), an input layer, an output layer, and convolutional layers (convolutional layer 1, convolutional layer 2, convolutional layer 3, and convolutional layer 4).

[0058] S103, training nonlinear distortion degree extraction network C net , generator network G net and the discriminator network D net ;

[0059] In one embodiment, the nonlinear distortion level extraction network C net The specific training method is as follows:

[0060] (1) Based on the training data pair, obtain the actual distortion degree data D(n) of the power amplifier;

[0061] Specifically, the nonlinear distortion degree extraction network C net The training data pair is the actual output data Y(n) of the power amplifier and the real distortion degree data D(n) of the power amplifier.

[0062] ;

[0063] Wherein, Y(n) is the output data when the power amplifier inputs X(n), and K is the gain of the power amplifier.

[0064] (2) Convert the two-dimensional data of X(n) Input nonlinear distortion extraction network C net , calculate the loss function ;

[0065] Specifically, As the degree of nonlinear distortion extraction network C net Input, output eigenvalue results C1(n), C2(n);

[0066] Calculate the distance between the actual distortion data D(n) of the power amplifier and the output eigenvalue results C1(n) and C2(n) to obtain the loss function ;

[0067] ;

[0068] in, The loss function of the network is extracted for the degree of nonlinear distortion, is the amplitude component of D(n), is the phase component of D(n), C1(n) is the nonlinear distortion extraction network C net The real deviation of the output eigenvalue result, C2(n) is the nonlinear distortion degree extraction network C net Output the imaginary deviation of the eigenvalue result.

[0069] Furthermore, the two-dimensional conversion is explained. Two-dimensional conversion refers to converting one-dimensional data into a two-dimensional graph. This is because the discriminator network is composed of a convolutional neural network, so the column vector of the input data (only related to time) needs to be converted into a two-dimensional input to adapt to the input characteristics of the convolution layer. For example, if the data that needs to be converted from one dimension to two dimensions is p(n), then the conversion to two dimensions should include the I / Q components of p(n) and the envelope related items of the current and past signals. Input matrix It is expressed as follows:

[0070]

[0071] Among them, M is the memory depth and N is the nonlinearity.

[0072] (3) Determine the loss function Is it greater than the threshold? If so, execute step (4); otherwise, end the loop. In this embodiment, the nonlinear distortion degree extraction network C net Set the learning rate to 0.001, the threshold to -30db, and the loop k = 1, 2, ..., 40.

[0073] (4) Use the Adam gradient descent algorithm to calculate the loss function Extract the partial derivatives of network parameters with respect to the degree of nonlinear distortion ,in C net The parameters in the k-1th cycle, C net Find the loss function for the parameters of the kth cycle The local minimum of the nonlinear distortion extraction network is updated , thus achieving reverse propagation.

[0074] In one embodiment, the generator network G net and the discriminator network D net The specific training method is as follows:

[0075] (1) Convert the two-dimensional data of Y(n) Input nonlinear distortion extraction network C net , and obtain the distortion level prediction result C;

[0076] (2) Combined with the distortion level prediction result C, we get , input generator network G net , output false data result X gen ;

[0077] Among them, splicing specifically refers to Combined with C, we get ;

[0078] ;

[0079] Among them, C3(n) and C4(n) are C net The two eigenvalues ​​of the output, C refers to C net The output result, in this example, is the vector .

[0080] The distortion level prediction result C of the actual output data Y(n) of the power amplifier is integrated into the generative adversarial network (GAN) as conditional information.

[0081] Specifically, the framework concatenates the distortion level prediction results C with the actual output data Y(n) of the power amplifier, so that the generative adversarial network GAN can effectively combine these distortion level prediction results C. This not only extracts the information of the actual output data Y(n) of the power amplifier, but also extracts the real distortion level data. The more features extracted, the more beneficial it is for the nonlinear distortion level extraction network C. net effectiveness.

[0082] Moreover, during the linearization process, the generator network G net It can not only extract the current features (I(n), Q(n)), historical features (I(n-1), Q(n-1)) and envelope-related properties (|X(n)|, |X(n)|) of the input signal N-1 , |X(n-1)|, |X(n-1)| N-1 ), and the distortion level prediction result C can also be used to obtain the false data result X gen , this integration makes the generator network G net Capable of capturing more comprehensive and sensitive features.

[0083] (3) Two-dimensional input data , generator network G netThe output result X gen Input to the discriminator network D net , we get Y real , Y gen ; Y real is the discriminator network D net For two-dimensional input data The discrimination result of the real input data is between [0,1]; gen is the discriminator network D net For false data result X gen The discriminant network D net The output result of represents the probability that the input data is true data. The closer the result is to 1, the greater the probability that it is true data.

[0084] (4) Calculate the generator network G net The loss function G loss and G NMSE , Discriminator Network D net The loss function D loss According to G loss Using the forward propagation of the gradient descent algorithm, the generator network G net The parameters of G are updated. NMSE Used to determine whether the model training reaches the threshold; if G NMSE If it is greater than the threshold, then execute steps (5) and (6) to update the generator network G. net , Discriminator Network D net Otherwise, end the loop;

[0085] In this embodiment, the learning rates of the generator network and the discriminator network are both 0.001, and G NMSE The threshold is -28dB.

[0086] Specifically, the calculation formula of the above loss function is as follows:

[0087] ;

[0088] ;

[0089] ;

[0090] (5) Calculate the discriminator network loss function D loss For the discriminator network parameters The gradient (i.e. partial derivative) of ,in D net The parameters in the i-1th loop, D netParameters in the i-th loop;

[0091] (6) Calculate the generator network loss function G loss For the generator network parameters The gradient (i.e. partial derivative) of ,in G net The parameters in the i-1th loop, G net Parameters for the i-th loop.

[0092] S104, randomly generate a transmission signal and input it into the trained conditional generative adversarial network pre-distortion model CGAN-DPD. The input data can obtain a linearized output after passing through a power amplifier.

[0093] Specifically, an OFDM signal that meets the 5G communication standard is randomly generated as a transmission signal, which is input into the trained conditional generative adversarial network pre-distortion model CGAN-DPD, and the output of the CGAN-DPD model is input into the power amplifier. The simulation results show that the transmission signal can be linearized by passing through the CGAN-DPD model and then input into the power amplifier.

[0094] Finally, a simulation experiment is conducted using the conditional generative adversarial network pre-distortion model CGAN-DPD trained by the present invention:

[0095] The power amplifier model selected for this simulation is NXP™ Airfast LDMOS Doherty PA, with a gain K of 113.0241. The transmission data is an OFDM signal similar to 5G, with a bandwidth of 100MHz. Each frame of data contains 6 OFDM symbols, each subcarrier carries a 16QAM symbol, the oversampling rate is 5, the number of iterations is 300, and the nonlinear distortion degree extraction network C is used. net and the discriminator network D net The learning rate is set to 0.001, G NMSE The threshold is set to -28dB. Define the nonlinear distortion level extraction network C net The number of neurons in the hidden layer 1 is 10, and the generator network G is defined net The number of neurons in the middle fully connected layers FC1 and FC2 is 40;

[0096] Define the generator network G net The activation layers Act1 and Act2 are ReLU functions; define C net The activation function of the first two layers is ReLU, and the activation function of the third layer is Sigmod function; define the discriminator network D netThe convolution kernel size of the first convolution layer is 2*2, and the convolution kernel size of the next three layers is 3*3, with step sizes of 2, 2, 1, and 2 respectively;

[0097] Define M=5, N=5, batch=4096.

[0098] The power spectrum of the output signals of various pre-distortion models and the power spectrum of the original signal without pre-distortion are simulated as follows: Figure 5 As shown in Table 1, the evaluation index values ​​of each pre-distortion model are shown in Table 1. The adjacent channel power ratio ACPR, normalized mean square error NMSE, and error magnitude EVM of the generative adversarial network pre-distortion model GAN-DPD are better than those of the neural network pre-distortion model NN-DPD and the memory polynomial pre-distortion model Memory Polynomial DPD, which are -0.934dB, -0.585dB, and -0.0897dB better than NN-DPD respectively. Under the same training settings, the ACPR, NMSE, and EVM of the conditional generative adversarial network pre-distortion model CGAN-DPD of the present invention are better than those of GAN-DPD, and are -2.86dB, -3.379dB, and -1.217dB better than those of NN-DPD respectively, and -5.333dB, -6.725dB, and -2.9323dB better than those of MP-DPD respectively.

[0099] Table 1 Performance comparison of different digital pre-distortion models

[0100]

[0101] The present invention provides a power amplifier linearization method based on a conditional generative adversarial network. The conditional generative adversarial network is used for the first time to perform pre-distortion processing on the power amplifier. The conditional generative adversarial network is a special deep learning model. Compared with traditional deep neural networks and shallow neural networks, CGAN has stronger generation capabilities and better feature learning effects, and can describe the characteristics of nonlinear systems more accurately. The present invention first builds a power amplifier nonlinear distortion degree extraction network to extract the power amplifier distortion level prediction results, and applies them to the generator network. By introducing the distortion level prediction results, CGAN can control the specific properties of the generated samples. This control power enables the model to more accurately reflect the input conditional information when generating, providing strong support for various application scenarios. The training model of CGAN-DPD is then proposed. After simulation experiments, for 100MHz broadband signals, CGAN-DPD is better than GAN-DPD in terms of ACPR, NMSE, and EVM for Doherty PA, and is better than NN-DPD by -2.86dB, -3.379dB, and -1.217dB, respectively, and better than MP-DPD by -5.333dB, -6.725dB, and -2.9323dB, respectively. It is better than the existing pre-distortion models in all three evaluation indicators.

[0102] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed in the present invention can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0103] The embodiments described above are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, a person skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application, and should all be included in the protection scope of the present application.

Claims

1. A power amplifier linearization method based on conditional generative adversarial network, characterized in that: The following steps are involved: Obtain randomly generated input data X(n) and actual output data Y(n) of the power amplifier; Construct a conditional generative adversarial network pre-distortion model CGAN-DPD, which includes a nonlinear distortion degree extraction network C net , Generative Adversarial Network; Generative Adversarial Network includes Generator Network G net , Discriminator Network D net ; Training the nonlinear distortion extraction network C net , generator network G net and the discriminator network D net ; The randomly generated transmission signal is input into the trained conditional generative adversarial network pre-distortion model CGAN-DPD, and the output data can be linearized after passing through the power amplifier; The nonlinear distortion degree extraction network C net The specific training method is as follows: (1) Based on the training data pair, obtain the actual distortion degree data D(n) of the power amplifier; The training data pair is the actual output data Y(n) and the real distortion degree data D(n); ； Where, Y(n) is the output data when input X(n), and K is the gain of the power amplifier; (2) Convert the two-dimensional data of X(n) Input nonlinear distortion extraction network C net , calculate the loss function ; The specific implementation of step (2) is as follows: As the degree of nonlinear distortion extraction network C net Input, output eigenvalue results C1(n), C2(n); Calculate the distance between the actual distortion data D(n) of the power amplifier and the output eigenvalue results C1(n) and C2(n) to obtain the loss function ; ； in, The loss function of the network is extracted for the degree of nonlinear distortion, is the amplitude component of D(n), is the phase component of D(n), C1(n) is the nonlinear distortion extraction network C net The real deviation of the output eigenvalue result, C2(n) is the nonlinear distortion degree extraction network C net Output the imaginary deviation of the eigenvalue result; (3) Determine the loss function Is it greater than the threshold? If so, execute step (4); otherwise, end the loop; (4) Calculate the loss function Extract the partial derivatives of network parameters with respect to the degree of nonlinear distortion ,in C net The parameters in the k-1th cycle, C net Find the loss function for the parameters of the kth cycle The local minimum of the nonlinear distortion extraction network is updated ; The generator network G net and the discriminator network D net The specific training method is as follows: (1) Convert the two-dimensional data of Y(n) Input nonlinear distortion extraction network C net , and obtain the distortion level prediction result C; (2) Combined with the distortion level prediction result C, we get , input generator network G net , output false data result X gen ; (3) Convert the two-dimensional data of X(n) , generator network G net The output result X gen Input to the discriminator network D net , we get Y real , Y gen ; Among them, Y real is the discriminator network D net For two-dimensional input data The judgment result, Y gen is the discriminator network D net For false data result X gen The judgment result of (4) Calculate the generator network G net The loss function G loss and G NMSE , Discriminator Network D net The loss function D loss ; According to G loss Using the forward propagation of the gradient descent algorithm, the generator network G net The parameters of G are updated. NMSE Used to determine whether the model training reaches the threshold; if G NMSE If it is greater than the threshold, then execute steps (5) and (6) to update the generator network G. net , Discriminator Network D net Otherwise, end the loop; (5) Calculate the discriminator network loss function D loss For the discriminator network parameters The gradient of ,in D net The parameters in the i-1th loop, D net Parameters in the i-th loop; (6) Calculate the generator network loss function G loss The parameters of the generator network The gradient of ,in G net The parameters in the i-1th loop, G net Parameters in the i-th loop; The two-dimensional conversion refers to converting one-dimensional data into a two-dimensional graph, and the conversion into two dimensions should include I / Q components of the data and envelope correlation items of current and past signals.

2. The power amplifier linearization method based on conditional generative adversarial network according to claim 1, characterized in that: In step (2), the splicing refers to splicing C into The last two columns get ; ； Among them, C3(n) and C4(n) are C net The two eigenvalues ​​of the output, C refers to C net The output result is:

3. The power amplifier linearization method based on conditional generative adversarial network according to claim 1, characterized in that: In step (4), the calculation formula of the loss function is as follows: ； ； 。 4. The power amplifier linearization method based on conditional generative adversarial network according to claim 1, characterized in that: The nonlinear distortion degree extraction network C net It includes input layer, hidden layer 1, hidden layer 2 and output layer; hidden layer 1 is composed of a fully connected layer and an activation layer of multiple neurons, and hidden layer 2 is composed of a fully connected layer and an activation layer of 2 neurons; the activation layer selects the ReLU function.

5. The power amplifier linearization method based on conditional generative adversarial network according to claim 1, characterized in that: The generator network G net It consists of two fully connected layers, two activation layers and an output layer. The activation layer uses the ReLU function. The output layer consists of 2 neurons. The output I DPD and Q DPD Two signal paths.

6. The power amplifier linearization method based on conditional generative adversarial network according to claim 1, characterized in that: The discriminator network D net It consists of four convolutional layers, four activation layers, and one fully connected layer.

Images