A power amplifier predistortion method and system based on heterogeneous neural networks

Abstract

This invention provides a power amplifier predistortion method and system based on heterogeneous neural networks, belonging to the field of power amplifier predistortion technology. The method includes: constructing a predistortion neural network, including a backbone network and an auxiliary network. The backbone network is used to perform nonlinear fitting on the power signal of the power amplifier, and the auxiliary network is used to generate dynamic modulation parameters based on changes in the circuit state of the power amplifier. The power signal of the target power amplifier is used as the input signal of the predistortion neural network. The backbone network performs nonlinear fitting on the input signal to generate fitting features. Based on the dynamic modulation parameters output by the auxiliary network, the fitting features output by the backbone network are modulated to output a predistortion signal. Beneficial effects: By constructing a predistortion neural network including a backbone network and an auxiliary network, and dynamically modulating the fitting features, the system can dynamically adapt to changes in the operating state of the power amplifier, improving the accuracy and real-time performance of predistortion compensation, and reducing computational resource requirements.

Description

Technical Field

[0001] This invention relates to the field of power amplifier predistortion technology, and in particular to a power amplifier predistortion method and system based on heterogeneous neural networks. Background Technology

[0002] A power amplifier (PA) is an amplifier that can produce maximum power output to drive a load under a given distortion rate. When amplifying signals, especially when operating near the saturation point, a power amplifier introduces amplitude-to-amplitude distortion (AM-AM) and phase-to-phase distortion (AM-PM), resulting in a decrease in the linearity of the output signal.

[0003] The predistortion method preprocesses the signal before it enters the power amplifier, introducing distortion that is opposite to the nonlinear characteristics of the power amplifier, thereby enabling the final output signal after amplification by the power amplifier to maintain a high degree of linearity.

[0004] However, the operating state of a power amplifier is time-varying and strongly coupled. Its nonlinear characteristics and memory effects dynamically drift with continuous changes in various circuit state parameters such as signal bandwidth, center frequency, peak-to-average power ratio, bias voltage, input power, and envelope tracking voltage, leading to a rapid decrease in the adaptability of the predistortion model. Traditional digital predistortion (DPD) typically employs a multi-look-up table (LUT) switching mechanism. This involves pre-collecting data for different operating states and training multiple sets of model parameters. During system runtime, the corresponding lookup table is called based on the current circuit state to implement predistortion, adapting to different circuit operating state switching. This increases the software storage burden. Moreover, the parameters of the lookup table are fixed and limited, failing to cover all possible operating scenarios and creating compensation blind spots in parameter transition regions.

[0005] Furthermore, while traditional neural network-based digital predistortion schemes possess some nonlinear fitting capabilities, most employ a single network structure or homogeneous architecture, focusing solely on modeling the static nonlinearity and memory effect of the power amplifier. They lack dedicated compensation mechanisms for dynamic circuit state drift, thus failing to achieve adaptive adjustment of model parameters. Some schemes directly input circuit state information as additional features into the backbone network, which, while improving adaptability to some extent, significantly increases network computational complexity and inference latency, and easily introduces redundant feature interference, leading to a decrease in the stability of the predistortion system. Summary of the Invention

[0006] To address the above technical problems, this invention provides a power amplifier predistortion method based on heterogeneous neural networks; on the other hand, it also provides a power amplifier predistortion system based on heterogeneous neural networks.

[0007] The technical problem solved by this invention can be achieved by the following technical solutions: A power amplifier predistortion method based on heterogeneous neural networks includes: Step S1: Construct a predistortion neural network, which includes a backbone network and an auxiliary network. The backbone network is used to perform nonlinear fitting on the power signal of the power amplifier, and the auxiliary network is used to generate dynamic modulation parameters based on the changes in the circuit state of the power amplifier. Step S2: Obtain the power signal of the target power amplifier, use the power signal of the target power amplifier as the input signal of the predistortion neural network, perform nonlinear fitting on the input signal through the backbone network to generate fitting features, and modulate the fitting features output by the backbone network according to the dynamic modulation parameters output by the auxiliary network to output the predistortion signal.

[0008] The power amplifier predistortion method based on heterogeneous neural networks described in this invention includes a backbone network comprising: The input layer is used to receive input signals, which include at least one or more combinations of the instantaneous power of the target power amplifier at the past m moments, the current moment, and the future m moments, in-phase components, and quadrature components, where m is a positive integer; A preprocessing layer, connected to the input layer, is used to preprocess the input signal; The memory layer, connected to the preprocessing layer, is used to extract time-dependent and memory effect features from the preprocessed input signal to obtain multi-scale memory features. A fully connected layer connects the output of the memory layer and the auxiliary network, respectively, and is used to perform nonlinear transformation and dynamic parameter modulation on multi-scale memory features; The output layer, connected to the fully connected layer, is used to output the predistorted signal.

[0009] The power amplifier predistortion method based on heterogeneous neural networks described in this invention includes a preprocessing layer comprising a regularization module, which is used to perform uniform-scale regularization processing on the input signal.

[0010] The power amplifier predistortion method based on heterogeneous neural networks described in this invention includes a memory layer comprising: A multi-scale temporal coding module, comprising a dual-branch recurrent neural network structure, performs forward propagation and enhanced feature extraction along the time dimension to simultaneously extract short-term memory features and long-term memory features; Multiple convolutional neural network modules, each of which is connected to the output of all branches of the recurrent neural network, are used to fuse and aggregate multi-scale memory features to obtain aggregated features; Multiple learnable activation modules, each connected to a convolutional neural network module, are used to perform adaptive nonlinear transformation on aggregated features through piecewise spline curves and learnable activation functions to obtain fitted features.

[0011] The power amplifier predistortion method based on heterogeneous neural networks described in this invention uses gated recurrent units for each neuron in the dual-branch recurrent neural network structure.

[0012] The power amplifier predistortion method based on heterogeneous neural networks described in this invention uses a non-uniform piecewise spline construction for the piecewise spline curve-learnable activation function.

[0013] The power amplifier predistortion method based on heterogeneous neural networks described in this invention includes a fully connected layer comprising: The parameter modulation module, which is connected to the memory layer and the auxiliary network respectively, is used to linearly modulate the fitting features output by the learnable activation module of the backbone network according to the dynamic modulation coefficients output by the auxiliary network.

[0014] The power amplifier predistortion method based on heterogeneous neural networks described in this invention includes an auxiliary network comprising: The dynamic prediction layer is used to predict the state of the power amplifier based on the circuit state parameters, which include one or more combinations of signal bandwidth, cutoff frequency, peak-to-average power ratio, bias voltage, input power, and envelope tracking voltage. A dynamic prediction output layer, connected to the dynamic prediction layer, is used to output dynamic modulation coefficients, which include scaling coefficients and translation coefficients.

[0015] The power amplifier predistortion method based on heterogeneous neural networks described in this invention includes a dynamic prediction layer comprising multiple convolutional neural networks configured in parallel, with the outputs of all the convolutional neural networks connected to the dynamic prediction output layer.

[0016] On the other hand, a power amplifier predistortion system based on heterogeneous neural networks is also provided for implementing the power amplifier predistortion method based on heterogeneous neural networks as described above, including: A network construction unit is used to construct a predistortion neural network, which includes a backbone network and an auxiliary network. The backbone network is used to perform nonlinear fitting on the power signal of the power amplifier, and the auxiliary network is used to generate dynamic modulation parameters based on the changes in the circuit state of the power amplifier. The predistortion unit, connected to the network construction unit, is used to acquire the power signal of the target power amplifier, use the power signal of the target power amplifier as the input signal of the predistortion neural network, perform nonlinear fitting on the input signal through the backbone network to generate fitting features, and modulate the fitting features output by the backbone network according to the dynamic modulation parameters output by the auxiliary network to output the predistortion signal.

[0017] The advantages or beneficial effects of the technical solution of this invention are as follows: This invention constructs a predistortion neural network including a backbone network and an auxiliary network, dynamically modulates the fitting features, and can dynamically adapt to changes in the operating state of the power amplifier, thereby improving the accuracy and real-time performance of predistortion compensation and reducing the computational resource requirements. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the power amplifier predistortion method based on heterogeneous neural networks in a preferred embodiment of the present invention. Figure 2 A block diagram of the predistortion neural network is shown in a preferred embodiment of the present invention. Figure 3 This is a schematic diagram of the predistortion neural network in a preferred embodiment of the present invention; Figure 4 A waveform diagram of adaptive non-uniform piecewise linear fitting is shown in a preferred embodiment of the present invention. Figure 4 The horizontal axis represents the input amplitude, and the vertical axis represents the output amplitude. Figure 5 This is a schematic diagram of an offline training environment in a preferred embodiment of the present invention; Figure 6 This is a schematic diagram of the offline training process in a preferred embodiment of the present invention; Figure 7 This is a schematic diagram of the structure of a power amplifier predistortion system based on a heterogeneous neural network, which is a preferred embodiment of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0021] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but this is not intended to limit the scope of the invention.

[0022] In a preferred embodiment of the present invention, based on the above-mentioned problems existing in the prior art, a power amplifier predistortion method based on heterogeneous neural networks is provided, such as... Figure 1 As shown, it includes: Step S1, construct a predistortion neural network 100; as follows Figure 2 As shown, the predistortion neural network 100 includes a backbone network 10 and an auxiliary network 20; wherein, the backbone network 10 is used to perform nonlinear fitting on the power signal of the power amplifier; and the auxiliary network 20 is used to generate dynamic modulation parameters according to the changes in the circuit state of the power amplifier. Step S2: Obtain the power signal of the target power amplifier, use the power signal of the target power amplifier as the input signal of the predistortion neural network 100, perform nonlinear fitting on the input signal through the backbone network 10 to generate fitting features, and modulate the fitting features output by the backbone network 10 according to the dynamic modulation parameters output by the auxiliary network 20 to output the predistortion signal.

[0023] Specifically, addressing the shortcomings of traditional predistortion neural networks, which are mostly single or homogeneous network architectures, this embodiment provides a heterogeneous neural network for power amplifier predistortion. It uses a backbone network 10 for nonlinear fitting and an auxiliary network 20 to dynamically modulate the characteristics output by the backbone network 10 according to changes in circuit state, thereby improving the accuracy and real-time performance of predistortion compensation and reducing computational resource requirements.

[0024] The power amplifier predistortion method based on heterogeneous neural networks of the present invention, such as... Figure 2 As shown, the backbone network 10 includes an input layer 101, a preprocessing layer 102, a memory layer 103, a fully connected layer 104, and an output layer 105; wherein, the input layer 101 is used to receive input signals, and the input signals include at least the instantaneous power of the target power amplifier at the past m time points, the current time point, and the future m time points |X(t±m)|. 2 In-phase component X i (t±m) and orthogonal component Xq One or more combinations of (t±m), where m is a positive integer; the preprocessing layer 102 is connected to the input layer 101 and is used to preprocess the input signal, including but not limited to regularization; the memory layer 103 is connected to the preprocessing layer 102 and is used to extract time-dependent and memory effect features from the preprocessed input signal to obtain multi-scale memory features; the fully connected layer 104 is connected to the memory layer 103 and the output of the auxiliary network 20 respectively, and is used to perform nonlinear transformation and dynamic parameter modulation on the multi-scale memory features; the output layer 105 is connected to the fully connected layer 104, such as... Figure 3 As shown, it includes an output module 1051 for outputting a predistortion signal.

[0025] Figure 3 In this context, the instantaneous power of the signal |X(t±m)|² is used to characterize the square of the signal amplitude at past and future times, in order to model the amplitude nonlinearity and memory effect of the power amplifier; the in-phase component Xi(t±m) and the quadrature component Xq(t±m) together provide phase information and I / Q cross-coupling characteristics, which are used to compensate for AM-PM distortion and quadrature imbalance, where t−m represents the past m sampling points and t+m represents the future m sampling points.

[0026] The power amplifier predistortion method based on heterogeneous neural networks of the present invention, such as... Figure 3 As shown, the preprocessing layer 102 includes a regularization module 1021, which is used to perform regularization processing on the input signal at a uniform scale to eliminate the dimensional differences between signals of different dimensions and improve the stability of network training.

[0027] The power amplifier predistortion method based on heterogeneous neural networks of the present invention, such as... Figure 3 As shown, memory layer 103 includes: The multi-scale temporal coding module includes a bi-branch recurrent neural network (RNN) structure, which performs forward and backward propagation along the time dimension to enhance feature extraction, thereby simultaneously extracting short-term memory features and long-term memory features. Multiple convolutional neural network modules 1032, each of which is connected to the output of all branches of the recurrent neural network, are used to fuse and aggregate multi-scale memory features to obtain aggregated features; Multiple learnable activation modules 1033 are provided, each of which is connected to a convolutional neural network module. These modules are used to perform adaptive nonlinear transformation on the aggregated features through a piecewise spline curve-learnable activation function to obtain the fitted features.

[0028] Specifically, this embodiment employs a dual-loop gated memory network. By adding a reverse loop network, the prediction accuracy of the anti-causal memory effect of the RF power amplifier can be improved. The combination of forward and reverse branches can better characterize the thermoelectric memory effect of the power amplifier.

[0029] This invention combines a bidirectional gated recurrent unit (BiGRU) with the anti-causal memory effect of power amplifiers and subsequent innovative modules to produce a synergistic effect, specifically adapted to the anti-causal memory and thermoelectric memory characteristics of RF power amplifiers. The memory effect of a power amplifier depends not only on historical input signals but also on future input signals, such as the correlation between thermal accumulation and future signals; therefore, a reverse recurrent branch is required. The output of the recurrent neural network is not directly used as the final pre-distortion signal but is sequentially fed into a learnable activation module, a fully connected layer 104, and modulated by an auxiliary network before being output.

[0030] The power amplifier predistortion method based on heterogeneous neural networks of the present invention uses gated recurrent units for each recurrent neural network neuron 1031 in the dual-branch recurrent neural network structure.

[0031] Specifically, each of the 1031 recurrent neural network neurons in the bi-branch recurrent neural network employs a gated recurrent unit (GRU). Using a gated recurrent unit avoids the vanishing and exploding gradient problems, while simultaneously reducing network computational complexity and improving real-time performance.

[0032] The power amplifier predistortion method based on heterogeneous neural networks of the present invention uses non-uniform piecewise spline construction for the piecewise spline curve-learnable activation function.

[0033] Specifically, addressing the issue of traditional activation functions' poor adaptability to digital predistortion (DPD), particularly their lack of ability to fit higher-order nonlinearities, this embodiment employs a learnable activation function that satisfies the gradient descent update rule. This function can be trained synchronously with the network neuron parameters, significantly reducing the number of hidden layer neurons required. Simultaneously, the use of non-uniform spline segmentation allows for more accurate fitting of the power amplifier's non-ideal characteristics for signals of different amplitudes, i.e., the nonlinear distortion characteristics of the AM-AM curve.

[0034] The spline nodes are non-uniformly distributed. Let the node set T = {t0, t1, ..., t} K}, t 0= h min ,t K= h max , where h min h represents the minimum value of a spline node. max This represents the maximum value of a spline node.

[0035] spline spacing Δi=t i+1 -t iThe values ​​i = 0, 1, ..., K-1 represent non-equidistant spacing settings, with the spacing adaptively allocated based on the local nonlinearity intensity of the AM-AM characteristics of the power amplifier. Specifically, in regions where the AM-AM curve has a large curvature and exhibits drastic nonlinear changes, such as the power amplifier compression region or the transition zone from the weak nonlinear region to the strong nonlinear region, the node arrangement is denser, and the spline spacing Δi is smaller; in the approximately linear region, the node arrangement is sparser, and the spline spacing Δi is larger.

[0036] The preferred node placement strategy is as follows:

[0037] Where i = 1, 2, ..., K-1; This represents the weight of the k-th node; Among them, weight It is proportional to the absolute value of the second derivative of the AM-AM curve:

[0038] in, express; Indicate; indicate; The learnable activation function is the pointwise activation following the output of the GRU, and its mathematical expression is:

[0039] Among them, B j,p (h) represents a p-order B-spline basis function defined on a non-uniform node set T, c j ∈R represents the learnable parameters (control point coefficients). These learnable parameters, along with other parameters in the neural network, such as GRU weights and fully connected layer weights, are updated together using gradient descent.

[0040] like Figure 4 The waveform diagram for adaptive non-uniform piecewise linear fitting is shown when m=6.

[0041] The power amplifier predistortion method based on heterogeneous neural networks of the present invention includes a fully connected layer comprising: The parameter modulation module 1041 is connected to the memory layer and the auxiliary network respectively, and is used to linearly modulate the fitting features output by the learnable activation module of the backbone network according to the dynamic modulation coefficients output by the auxiliary network.

[0042] The power amplifier predistortion method based on heterogeneous neural networks of the present invention, such as... Figure 2As shown, the auxiliary network 20 includes a dynamic prediction layer 201 and a dynamic prediction output layer 202. The dynamic prediction layer 201 is used to predict the state of the power amplifier based on its circuit state parameters. These parameters include one or more combinations of signal bandwidth Bw, cutoff frequency Fc, peak-to-average power ratio (PAPR), bias voltage Bias, input power Pin, and envelope tracking voltage ET. The signal bandwidth Bw is the frequency width of the current input signal; a larger bandwidth leads to a more significant memory effect in the power amplifier, resulting in more complex nonlinear distortion. The cutoff frequency Fc is the frequency of the RF carrier; frequency changes affect the impedance characteristics of the matching network, thereby altering the nonlinear response of the power amplifier. The peak-to-average power ratio (PAPR) is... The ratio of peak power to average power (PAPR) is a key performance indicator. A higher PAPR indicates greater instantaneous power fluctuations, increasing the probability of the power amplifier entering saturation and exacerbating nonlinear distortion. Bias is the static operating point set for the power amplifier transistors, directly affecting the amplifier's gain, efficiency, and linearity. Changing Bias adjusts its operating state. Input power (Pin) represents the average power level of the input signal. The amplifier's nonlinearity (such as AM-AM and AM-PM distortion) changes with the input power, and Pin is a core driving factor. Envelope tracking (ET) is a technique that dynamically adjusts the power amplifier's supply voltage based on the signal envelope. When ET is enabled, the amplifier's characteristics change rapidly with the supply voltage, making it a highly dynamic scenario. The dynamic prediction output layer 202 connects to the dynamic prediction layer 201 and is used to output dynamic modulation coefficients, which include scaling and translation coefficients.

[0043] The power amplifier predistortion method based on heterogeneous neural networks of the present invention includes a dynamic prediction layer 201 comprising multiple parallel convolutional neural networks 2011 (CNNs), the outputs of all convolutional neural networks 2011 being connected to the dynamic prediction output layer 202.

[0044] Specifically, addressing the issues of high storage overhead and inability to cover all continuous operating scenarios caused by switching between multiple lookup tables (LUTs) in traditional digital predistortion (DPD), this embodiment adds a dynamic prediction layer to dynamically predict the modulation coefficients based on circuit state parameters. The dynamic prediction layer learns and trains circuit state parameters, including but not limited to bandwidth Bw, cutoff frequency Fc, peak-to-average power ratio PAPR, bias Bias, input power Pin, and envelope tracking ET, to predict actual circuit state changes.

[0045] Compared to the coarse architecture that directly uses circuit state information as the input of the backbone network, this embodiment uses a parameter modulation module 1041 to scale the coefficients of the fully connected layer 104 to achieve dynamic compensation, which is lightweight and more adaptable.

[0046] The dynamic prediction process includes the following steps: First, the current circuit state vector ct=[Bw,Fc,PAPR,Bias,Pin,ET] is input into the auxiliary network 20 to extract the influence mode of the circuit state on the nonlinear characteristics of the power amplifier; this process does not require the acquisition of the power amplifier output signal and can be run independently online.

[0047] Then, based on the learned nonlinear mapping relationship, the auxiliary network 20 dynamically predicts and generates multiple pairs of modulation coefficients for the current circuit state, each pair of modulation coefficients including a scaling factor. Translation coefficient .

[0048] Finally, the scaling factor is used. Translation coefficient Fitting features of the backbone network 10 output Feature linear modulation is performed. A heterogeneous neural network employs a master-slave collaborative architecture, where the auxiliary network modulates the output features of the master network. The master and auxiliary networks use different structures. Mathematically, the modulation operation is expressed as scaling and translation transformations:

[0049] in, This represents the signal after adaptive modulation based on circuit state prediction. Indicates the scaling factor; Indicates the output of the backbone network; This represents the translation coefficient.

[0050] This invention provides a power amplifier predistortion system based on heterogeneous neural networks, used to implement the power amplifier predistortion method based on heterogeneous neural networks described above. Figure 7 As shown, it includes: Network building unit 1 is used to build a predistortion neural network 100. The predistortion neural network 100 includes a backbone network 10 and an auxiliary network 20. The backbone network 10 is used to perform nonlinear fitting on the power signal of the power amplifier, and the auxiliary network 20 is used to generate dynamic modulation parameters according to the changes in the circuit state of the power amplifier. The predistortion unit 2 is connected to the network construction unit 1. It is used to acquire the power signal of the target power amplifier and use the power signal of the target power amplifier as the input signal of the predistortion neural network 100. The input signal is nonlinearly fitted by the backbone network 10 to generate fitting features. According to the dynamic modulation parameters output by the auxiliary network 20, the fitting features output by the backbone network 10 are modulated to output the predistortion signal.

[0051] Furthermore, this invention employs an offline method to train the model and an online method to optimize the model in real time. Figure 5The offline training environment of the predistortion neural network 100 is shown. The PC sends circuit parameter configuration commands to the power amplifier (PA) to set its working state. At the same time, the training data is imported into the signal source to generate an RF excitation signal and input to the PA. The output signal of the PA after amplification is collected by the spectrum analyzer and sent back to the PC. The PC uses the complete input-output data to complete the training of the predistortion model and realize the fitting of the nonlinear inverse characteristics of the power amplifier.

[0052] Figure 6 The offline training process of the predistortion neural network 100 is shown, which is divided into three stages: training preparation, iterative training, and scenario verification, including the following steps: Training preparation phase A1, Initiate DPD training; A2, Change power amplifier frequency band / gain / voltage: Actively adjust the power amplifier's operating parameters to simulate different operating scenarios (frequency band, gain, power supply voltage) to ensure that the training data covers a variety of nonlinear states; A3, Set Instrument Frequency / Input Power: Configure the operating frequency and input power of test instruments such as signal sources and spectrum analyzers to match the current operating status of the power amplifier; A4, Selecting Training Data Sources: Determine the basic data to be used for model training, providing a data foundation for subsequent iterative training; Iterative training phase A5, Determine if the epoch (training round) has reached the preset value: Check if the current training round has reached the preset value; If the target is not met, proceed with the single-round training process from A6 to A10. If the preset number of rounds is reached, exit the training loop and proceed to A11; A6, Import Circuit Parameters: Load the circuit parameters (such as nonlinear model coefficients and operating point parameters) corresponding to the current operating condition of the power amplifier. A7, Predistortion data import signal source: Send the predistortion data generated by the model to the signal source to generate the predistortion excitation signal; A8, Spectrum analyzer acquires power amplifier output: acquires the output signal of the power amplifier after pre-distortion correction; A9, Alignment of Transmitted and Acquired Data: Synchronizes and aligns the raw data transmitted by the signal source with the acquired data output by the power amplifier in time, providing sample pairs for model training; A10, Training Model: Based on aligned emission / acquisition data, update the parameters of the DPD model to complete one round of training.

[0053] Scenario verification phase A11, Determine if scene coverage is sufficient: Check if enough power amplifier operating scenes (frequency band, gain, voltage combination) have been covered during the training process. If insufficient, return to A2, readjust the amplifier parameters, supplement the scene data, and continue training; If sufficient, proceed to A12; A12, Input validation data into the model to generate predistorted data: Input validation data that was not used in training into the trained DPD model to generate the corresponding predistorted data; A13, Predistortion data import signal source: Send the predistortion data for verification to the signal source to generate the excitation signal; A14, Spectrum analyzer acquires and demodulates the power amplifier output: acquires and demodulates the power amplifier's output signal, and extracts key performance indicators; A15, Record EVM / ACLR: Record indicators such as error vector magnitude (EVM) and adjacent channel leakage ratio (ACLR) to evaluate the DPD correction effect; At this point, the training and validation of the DPD model are complete.

[0054] The above are merely preferred embodiments of the present invention and are not intended to limit the implementation methods and protection scope of the present invention. Those skilled in the art should recognize that any equivalent substitutions and obvious changes made using the content of this specification and illustrations should be included within the protection scope of the present invention.

Claims

1. A power amplifier predistortion method based on heterogeneous neural networks, characterized in that, include: Step S1: Construct a predistortion neural network, which includes a backbone network and an auxiliary network. The backbone network is used to perform nonlinear fitting on the power signal of the power amplifier, and the auxiliary network is used to generate dynamic modulation parameters based on the changes in the circuit state of the power amplifier. Step S2: Obtain the power signal of the target power amplifier, use the power signal of the target power amplifier as the input signal of the predistortion neural network, perform nonlinear fitting on the input signal through the backbone network to generate fitting features, and modulate the fitting features output by the backbone network according to the dynamic modulation parameters output by the auxiliary network to output the predistortion signal.

2. The power amplifier predistortion method based on heterogeneous neural networks according to claim 1, characterized in that, The backbone network includes: The input layer is used to receive input signals, which include at least one or more combinations of the instantaneous power of the target power amplifier at the past m moments, the current moment, and the future m moments, in-phase components, and quadrature components, where m is a positive integer; A preprocessing layer, connected to the input layer, is used to preprocess the input signal; The memory layer, connected to the preprocessing layer, is used to extract time-dependent and memory effect features from the preprocessed input signal to obtain multi-scale memory features. A fully connected layer connects the output of the memory layer and the auxiliary network, respectively, and is used to perform nonlinear transformation and dynamic parameter modulation on multi-scale memory features; The output layer, connected to the fully connected layer, is used to output the predistorted signal.

3. The power amplifier predistortion method based on heterogeneous neural networks according to claim 2, characterized in that, The preprocessing layer includes a regularization module, which is used to perform uniform-scale regularization processing on the input signal.

4. The power amplifier predistortion method based on heterogeneous neural networks according to claim 2, characterized in that, The memory layer includes: A multi-scale temporal coding module, comprising a dual-branch recurrent neural network structure, performs forward propagation and enhanced feature extraction along the time dimension to simultaneously extract short-term memory features and long-term memory features; Multiple convolutional neural network modules, each of which is connected to the output of all branches of the recurrent neural network, are used to fuse and aggregate multi-scale memory features to obtain aggregated features; Multiple learnable activation modules, each connected to a convolutional neural network module, are used to perform adaptive nonlinear transformation on aggregated features through piecewise spline curves and learnable activation functions to obtain fitted features.

5. The power amplifier predistortion method based on heterogeneous neural networks according to claim 4, characterized in that, In the described bi-branch recurrent neural network structure, each recurrent neural network neuron employs a gated recurrent unit.

6. The power amplifier predistortion method based on heterogeneous neural networks according to claim 4, characterized in that, The piecewise spline curve-learnable activation function is constructed using non-uniform piecewise splines.

7. The power amplifier predistortion method based on heterogeneous neural networks according to claim 2, characterized in that, The fully connected layer includes: The parameter modulation module, which is connected to the memory layer and the auxiliary network respectively, is used to linearly modulate the fitting features output by the learnable activation module of the backbone network according to the dynamic modulation coefficients output by the auxiliary network.

8. The power amplifier predistortion method based on heterogeneous neural networks according to claim 1, characterized in that, The auxiliary network includes: The dynamic prediction layer is used to predict the state of the power amplifier based on the circuit state parameters, which include one or more combinations of signal bandwidth, cutoff frequency, peak-to-average power ratio, bias voltage, input power, and envelope tracking voltage. A dynamic prediction output layer, connected to the dynamic prediction layer, is used to output dynamic modulation coefficients, which include scaling coefficients and translation coefficients.

9. The power amplifier predistortion method based on heterogeneous neural networks according to claim 8, characterized in that, The dynamic prediction layer comprises multiple convolutional neural networks configured in parallel, and the outputs of all the convolutional neural networks are connected to the dynamic prediction output layer.

10. A power amplifier predistortion system based on heterogeneous neural networks, characterized in that, A method for implementing the power amplifier predistortion method based on heterogeneous neural networks as described in any one of claims 1-9 includes: A network construction unit is used to construct a predistortion neural network, which includes a backbone network and an auxiliary network. The backbone network is used to perform nonlinear fitting on the power signal of the power amplifier, and the auxiliary network is used to generate dynamic modulation parameters based on the changes in the circuit state of the power amplifier. The predistortion unit, connected to the network construction unit, is used to acquire the power signal of the target power amplifier, use the power signal of the target power amplifier as the input signal of the predistortion neural network, perform nonlinear fitting on the input signal through the backbone network to generate fitting features, and modulate the fitting features output by the backbone network according to the dynamic modulation parameters output by the auxiliary network to output the predistortion signal.

Images