Method and system for compensating for power amplifier distortion

Abstract

A compensator compensates for distortion of a power amplifier circuit. A power amplifier neural network (PAN) is trained to model the power amplifier circuit using a predetermined input and output signal pair characterizing the power amplifier circuit. The compensator is then trained to pre-distort a signal received by the PAN. The compensator uses the trained neural network to optimize a loss between an input to the compensator and an output of the PAN, and the loss is computed according to a multi-objective loss function including one or more time-domain loss functions and one or more frequency-domain loss functions. The trained compensator performs signal compensation, outputting a pre-distorted signal to the power amplifier circuit.

Description

Technical Field

[0001] Embodiments of the present invention relate to using neural networks to optimize power amplifiers. Background Technology

[0002] In most radio frequency (RF) communication systems, the power amplifier (PA) is a major source of nonlinearity and memory effects that can lead to severe spectral regrowth. Spectral regrowth significantly degrades signal quality due to high out-of-band power. Memory effects cause asymmetry in the transmitted signal. These issues are critical to next-generation 5G technologies. Furthermore, in the edge devices of 5G communication systems, high transmit power and limited supply voltage both exacerbate the nonlinearity of the power amplifier.

[0003] To overcome the aforementioned problems of power amplifiers, and considering integration complexity and effectiveness, digital pre-distortion (DPD) has become a common and practical mechanism. DPD generates a pre-distorted signal to compensate for the damage caused by the power amplifier. Traditional techniques for DPD design typically involve deriving a mathematical power amplifier model and calculating the corresponding inverse function characterizing the DPD. Due to the complex nature of 5G communication systems, designing radio frequency (RF) components through mathematical modeling is a challenging task. Furthermore, these mathematical models require extensive manual adjustments to meet various specification requirements. Therefore, improvements to power amplifier design are needed. Summary of the Invention

[0004] In one embodiment, a method for compensating for power amplifier distortion is provided. The method includes the steps of: training a power amplifier neural network (PAN) to model the power amplifier circuit using predetermined input and output signal pairs characterizing the power amplifier circuit; and training a compensator to pre-distort the signal received by the PAN. The compensator uses the trained neural network to optimize the loss between the compensator input and the PAN output, and the loss is calculated based on a multi-objective loss function comprising one or more time-domain loss functions and / or one or more frequency-domain loss functions. The method further includes the step of performing signal compensation via the trained compensator to output the pre-distorted signal to the power amplifier circuit.

[0005] In another embodiment, a system for compensating for power amplifier distortion is provided. The system includes a memory storing a neural network model and processing hardware coupled to the memory. The processing hardware is used to train a neural network analog (PAN) to model the power amplifier circuit using predetermined input and output signal pairs characterizing the power amplifier circuit; and to train a compensator to pre-distort the signal received by the PAN. The compensator uses the trained neural network to optimize the loss between the compensator input and the PAN output, the loss being calculated based on a multi-objective loss function comprising one or more time-domain loss functions and / or one or more frequency-domain loss functions. The processing hardware is also used to perform signal compensation via the trained compensator, thereby outputting a pre-distorted signal to the power amplifier circuit.

[0006] The multi-objective loss function includes at least frequency domain normalization loss, which is the difference between the adjacent channel leakage power ratio (ACLR) input to the compensator and the ACLR output by the PAN, wherein the ACLR is the ratio of the filtered average power centered on the assigned channel frequency to the filtered average power centered on the adjacent channel frequency.

[0007] The multi-objective loss function used to train the PAN and the compensator includes at least the frequency domain normalization loss, the frequency domain mean absolute error (MAE), and the time domain mean square error (MSE). The MAE is calculated using the difference between the short-time Fourier transform (STFT) of the compensator input and the STFT of the PAN output, and the MSE is calculated using the difference between the compensator input and the PAN output.

[0008] The multi-objective loss function includes at least the frequency domain average absolute error (MAE) calculated using the difference between the STFT input to the compensator and the STFT output by the PAN.

[0009] The multi-objective loss function includes at least the temporal error vector magnitude EVM calculated using the difference (discrepancy) between the PAN output symbol and the ideal quadrature amplitude modulation (QAM) symbol.

[0010] The multi-objective loss function includes at least the time-domain mean square error (MSE) calculated using the difference between the compensator input and the PAN output.

[0011] Other aspects and features will become apparent to those skilled in the art when they read the following detailed description of specific embodiments in conjunction with the accompanying drawings. Attached Figure Description

[0012] The present invention is illustrated in the accompanying drawings by way of example rather than limitation, wherein the same reference numerals indicate similar elements. It should be noted that different designations for "a" or "an" embodiment in the present invention do not necessarily refer to the same embodiment; such designations imply at least one. Furthermore, when a particular feature, structure, or characteristic is described in connection with an embodiment, it means that such a feature, structure, or characteristic can be implemented in conjunction with other embodiments, whether or not explicitly described, to the knowledge of those skilled in the art.

[0013] Figure 1A A system for compensating for distortion caused by a power amplifier according to a first embodiment is shown.

[0014] Figure 1B A block diagram of the training configuration according to the first embodiment is shown.

[0015] Figure 2A A system for compensating for distortion caused by a power amplifier according to a second embodiment is shown.

[0016] Figure 2B This is a block diagram illustrating the training configuration according to the second embodiment.

[0017] Figure 3A A system for compensating for distortion caused by a power amplifier according to a third embodiment is shown.

[0018] Figure 3B This is a block diagram illustrating the training configuration according to the third embodiment.

[0019] Figure 4A A system for compensating for distortion caused by a power amplifier according to a fourth embodiment is shown.

[0020] Figure 4B This is a block diagram illustrating the training configuration according to the fourth embodiment.

[0021] Figure 5 A flowchart of a method for compensating for distortion caused by a power amplifier according to one embodiment is shown.

[0022] Figure 6 A block diagram of a device including a compensator according to one embodiment is shown. Detailed Implementation

[0023] Numerous specific details are set forth in the following description. However, it should be understood that embodiments of the invention can be practiced without these specific details. In other instances, well-known circuits, structures, and techniques have not been shown in detail so as not to obscure the understanding of the invention. However, those skilled in the art will understand that the invention can be practiced without these specific details. Those of ordinary skill in the art will be able to achieve appropriate functionality without excessive experimentation based on the included description.

[0024] Embodiments of the present invention provide an artificial-intelligent (AI) method and system for compensating for damage caused by RF circuitry, such as distortion caused by power amplifiers. Initially, a first neural network is trained to model the characteristics of the power amplifier. A compensator is then connected to the input of the trained first neural network, wherein the compensator includes a second neural network. In one embodiment, the second neural network is a coefficient generator neural network (CGN), trained to generate filter coefficients for a digital pre-distorter (DPD). The DPD generates a pre-distorted signal to cancel the distortion of the power amplifier. In another embodiment, the second neural network is a DPD neural network trained to generate a pre-distorted signal to cancel the distortion of the power amplifier.

[0025] According to embodiments of the present invention, the training of the first neural network and the second neural network can be based on a combination of one or more time-domain losses and one or more frequency-domain losses. For example, the time-domain loss may include mean square error (MSE) and error vector magnitude (EVM). The frequency-domain loss may include mean absolute error (MAE) and specification loss. The combination of the two frequency-domain losses can simultaneously reduce spectral regeneration and optimize circuit design according to communication specifications.

[0026] Figure 1AA system 100 for compensating distortion caused by a power amplifier 110 according to an embodiment is shown. System 100 can be deployed in a wireless device. More specifically, in the transmitter section of a wireless device. System 100 includes a DPD 120, a CGN 130, and a digital clipping unit 150. DPD 120 and CGN 130 together are referred to as compensator 180. System 100 is coupled to the power amplifier 110. In an embodiment where the power amplifier 110 is an analog power amplifier, system 100 may also include a digital-to-analog converter (DAC) 151 and an analog-to-digital converter (ADC) 152. In another embodiment where the power amplifier 110 is a digital power amplifier, digital signals can be transmitted between system 100 and power amplifier 110 without conversion by DAC 151 and ADC 152. Therefore, DAC 151 and ADC 152 are shown with dashed boxes to indicate that they may not be included in system 100 in alternative embodiments. The following description of signal compensation uses an analog power amplifier as an example. It is understandable that the disclosed signal compensation also applies to digital power amplifiers.

[0027] DPD 120 receives the input signal x and generates a pre-distorted signal z to compensate for the distortion caused by power amplifier 110. In one embodiment, DPD 120 may be configured with a set of compensator coefficients. The pre-distorted signal z output from DPD 120 is clipped by digital clipping unit 150 to limit the amplitude of the signal received by power amplifier 110. Clipping can accelerate convergence during the training phase of CGN 130. The clipped signal is converted into an analog signal by DAC 151, which is amplified by power amplifier 110 to form the output signal x'. For each time t, the output signal x'(t) is generated from the time series x(tp:t) of the input signal, where p is a configurable value.

[0028] The DPD coefficients are calculated by the CGN 130. In one embodiment, the CGN 130 calculates the filter coefficients c(i) at time t using the clipped signal and the power amplifier output x'(t-1-q:t-1) after ADC conversion, where q is a configurable value and i is the running index for the total Nth-order filter coefficients. The CGN 130 can update the filter coefficients c(i) at runtime.

[0029] In one embodiment, CGN 130 is a neural network trained to generate coefficients for DPD 120. Alternatively, power amplifier 110 can be modeled by a neural network, and a power amplifier neural network (PAN) can be trained to use the trained PAN during CGN training.

[0030] Figure 1BThis illustrates a system 100 according to one embodiment. Figure 1A A block diagram of the training configuration for the power amplifier neural network (PAN). The system 100 can be trained in two training phases. In the first phase, the power amplifier neural network (PAN) 111 is trained to characterize non-differentiable power amplifier circuits (e.g., Figure 1A The nonlinear behavior and memory effect of the power amplifier 110 in the PAN 111 are investigated. After training of PAN 111 is completed, the weights of PAN 111 are fixed and used in the second training phase. CGN 130 is trained in the second training phase. In one embodiment, one or both of PAN 111 and CGN 130 are trained to optimize a multi-objective loss function. In another embodiment, PAN 111 and CGN 130 can be trained to optimize the same multi-objective loss function or two different multi-objective loss functions.

[0031] Training data for PAN 111 can be obtained from multiple actual power amplifiers (i.e., power amplifier circuits) and processed as time-series data with two channels, including an in-phase (I) channel and a quadrature (Q) channel (e.g., x = {I...}). t Q t}, t=0,1,...T). When training PAN 111, the output of the actual power amplifier is the ground truth, and the loss is measured in terms of the difference between the ground truth and the PAN output. After training PAN111 and fixing the filter weights of PAN111, the trained PAN 111 is used in the training of CGN 130. When training CGN130, the input of DPD 120 is the ground truth, and the loss is measured in terms of the difference between that ground truth and the PAN output. In the following description of the multi-objective loss function, the ground truth is denoted as x, and the PAN output is denoted as... It's important to note that the reference truth x in PAN 111 training is the output of the actual power amplifier, while the reference truth x in CGN 130 training is the input signal. Before calculating the loss, compare the reference truth x and the PAN output. Apply amplitude normalization.

[0032] In one embodiment, the multi-objective loss function comprises a combination of temporal and frequency domain losses. These losses can be weighted and summed to produce a total loss value. At each epoch of the training phase, the loss calculator 160 evaluates the loss function (e.g., calculates the total loss value) and the weight update module 170 calculates the gradient with respect to the neural network weights. As an example, the weight update module 170 may implement a gradient-based optimization algorithm, such as the Adam algorithm (Kingma et al., ADAM: A method for stochastic optimization, arXiv: 1412.6980).

[0033] In one embodiment, the multi-objective loss function may include temporal loss, such as the baseline truth x and the PAN output. The mean square error (MSE) between them. MSE can be expressed as follows:

[0034]

[0035] In the spectral domain of the Fourier transform, each complex number within an interval represents a specific frequency range. The absolute value describes the power amplitude at a specific frequency point. Since the power amplitude of the transmitted signal is typically greater than that of the out-of-band signal, the mean absolute error (MAE) can be used to replace or supplement the time-domain loss (MSE) for a fairer assessment of the loss. Therefore, multi-objective loss functions may include frequency-domain losses, such as the STFT and PAN outputs of the reference truth x. MAE between STFTs, where STFT represents the application to x and The Short-Time Fourier Transform (MAE) of both. MAE can be represented as follows:

[0036]

[0037] Instead of or in addition to the losses described above, multi-objective loss functions can include frequency domain losses, such as the canonical loss calculated using the adjacent channel leakage power ratio (ACLR). According to the 3GPP specification, ACLR is the ratio of the filtered average power centered at the assigned channel frequency (i.e., in-band frequency) to the filtered average power centered at the adjacent channel frequency (i.e., out-of-band frequency). The canonical loss is defined as minimizing the baseline truth x and the PAN output. The ACLR differences between them. The ACLR formula and specification loss are as follows:

[0038]

[0039]

[0040] Instead of or in addition to the losses described above, multi-objective loss functions can include time-domain losses, such as error vector magnitude (EVM). EVM measures the distance of a constellation point of the signal from its ideal position, for example, the difference (i.e., the error vector) between a PAN output symbol and an ideal quadrature amplitude modulation (QAM) symbol. EVM can be calculated as the root mean square (RMS) average magnitude of the error vector, whose magnitude is normalized to the ideal signal reference magnitude. Methods for measuring the EVM of a transmitter are known in the art. EVM can be used to quantify performance losses in the PAN output, and EVM can be optimized during the training of the CGN 130.

[0041] In one embodiment, the operation of the loss calculator 160 and the weight update module 170 can be performed by a general-purpose processor on the device housing the power amplifier 110. In another embodiment, system 100 may include dedicated hardware or an accelerator for training the neural network. The multi-objective loss function used in training the PAN 111 and CGN 130 may include the same or different combinations of the losses described above.

[0042] Figure 2A This is a block diagram of a system 200 for compensating distortion caused by a power amplifier 110 according to another embodiment. System 200 can be deployed in a wireless device, more specifically, in the transmitter section of the wireless device. System 200 includes a DPD 120 and a CGN 230, which together are referred to as compensator 280. If the power amplifier 110 is an analog power amplifier, system 200 may also include a DAC 151 and an ADC 152. CGN 230 receives an input including the output of the power amplifier 110 and the input signal x. Based on the input, CGN 230 generates delta coefficients Δc(i), where i is the running index for a total of Nth-order filter coefficients. The delta coefficient is the increment or decrement of the coefficients between two consecutive update times. Compensator 280 includes an accumulator 250 to accumulate the delta coefficients and sends the accumulated output (i.e., the filter coefficient c(i)) to the DPD 120. The coefficients can be initialized to zero.

[0043] Figure 2B This is a block diagram illustrating components used for training the CGN 230 according to one embodiment. Training follows the procedures outlined above. Figure 1BThe same two-stage training process is described. When training PAN 111 and CGN 230, the loss calculator 260 calculates a multi-objective loss function that measures the baseline truth x and the PAN output. The differences between them are as follows: In PAN 111 training, the baseline truth x is the output of the actual power amplifier, while in DPD neural network 420 training, the baseline truth x is the input signal. The weight update module 270 updates the neural network weights based on the gradient with respect to the neural network weights. The multi-objective loss function can be a combination of one or more time-domain losses (e.g., MSE, EVM) and one or more frequency-domain losses (e.g., MAE, ACLR-based canonical loss). The multi-objective loss function used in the training of PAN 111 and CGN 230 can include the same or different combinations of the above losses.

[0044] Figure 3A This is a block diagram of a system 300 for compensating distortion caused by a power amplifier 110 according to another embodiment. System 300 can be deployed in a wireless device, more specifically, in the transmitter section of the wireless device. System 300 includes a DPD 120 and a CGN 330, which together are referred to as compensator 380. If the power amplifier 110 is an analog power amplifier, system 300 may also include a DAC 151 and an ADC 152. CGN 330 receives input including the output of power amplifier 110, the input signal x, and previously generated coefficients. Based on the input, CGN 330 generates filter coefficients c(i)', where i is the running index for a total of Nth-order filter coefficients. The output of CGN 330 is sent to DPD 120.

[0045] Figure 3B This is a block diagram illustrating components for training a CGN 330 according to one embodiment. Training follows the steps outlined above. Figure 1B The description is the same for the two-stage training process. When training PAN 111 and CGN 330, the loss calculator 360 calculates a multi-objective loss function that measures the baseline truth x and the PAN output. The differences between them are as follows: In PAN 111 training, the baseline truth x is the output of the actual power amplifier, while in DPD neural network 420 training, the baseline truth x is the input signal. The weight update module 370 updates the neural network weights based on the gradient with respect to the neural network weights. The multi-objective loss function can be a combination of one or more time-domain losses (e.g., MSE, EVM) and one or more frequency-domain losses (e.g., MAE, ACLR-based canonical loss). The multi-objective loss function used in the training of PAN 111 and CGN 330 can include the same or different combinations of the above losses.

[0046] Figure 4AThis is a block diagram of a system 400 for compensating for distortion caused by a power amplifier 110 according to another embodiment. System 400 includes a DPD neural network 420, which is a neural network trained to generate a pre-distorted signal z. The DPD neural network 420 is a compensator for the power amplifier 110. In an embodiment where the power amplifier is analog, the pre-distorted signal z is used by the DAC 151 of the power amplifier 110 to convert it into an analog signal.

[0047] Figure 4B This is a block diagram illustrating components for training a DPD neural network 420 according to one embodiment. Training follows the steps outlined above. Figure 1B The same two-stage training process is described. When training the PAN 111 and DPD neural networks 420, the loss calculator 460 calculates a multi-objective loss function that measures the baseline truth x and the PAN output. The differences between them are as follows: In PAN111 training, the baseline truth x is the output of the actual power amplifier; in DPD neural network 420 training, the baseline truth x is the input signal. The weight update module 470 updates the neural network weights based on the gradient with respect to the neural network weights. The multi-objective loss function can be a combination of one or more time-domain losses (e.g., MSE, EVM) and one or more frequency-domain losses (e.g., MAE, ACLR-based canonical loss). The multi-objective loss functions used in the training of PAN 111 and DPD neural network 420 can include the same or different combinations of the above losses.

[0048] Figure 5 This is a flowchart of a method 500 for compensating power amplifier distortion according to one embodiment. Method 500 can be performed by an electronic device, such as... Figure 6 The device 600 in the middle. In some embodiments, the method 500 may be by Figure 1A , 2A The method can be executed by any of systems 100, 200, 300, and 400 in 3A and 4A. Alternatively, method 500 can be executed by a server computer system including memory and processing hardware. The processing hardware performs neural network training and provides the trained neural network to the device where the power amplifier resides. In step 510, the system trains the power amplifier neural network (PAN) to model the power amplifier circuit using predetermined input and output signal pairs characterizing the power amplifier circuit. In step 520, the system trains a compensator to predistort the signal received by the PAN. The compensator can be... Figure 1A The compensator 180 in the system includes both DPD 120 and CGN 130. Alternatively, the compensator can be... Figure 2AThe compensator 280 includes a DPD 120, an accumulator 250, and a CGN 230. In another embodiment, the compensator may be... Figure 3A The compensator 380 includes DPD 120 and CGN 330. In yet another embodiment, the compensator may be... Figure 4A The compensator 420 in the diagram is also known as the DPD neural network 420.

[0049] The compensator uses a trained neural network to optimize the loss between the compensator input and the PAN output. The loss is calculated based on a multi-objective loss function that includes one or more time-domain losses and one or more frequency-domain losses. After the training phase in steps 510 and 520 is completed, the system begins the inference phase in step 530, where the compensator performs signal compensation to output a pre-distorted signal to the power amplifier circuit.

[0050] In one embodiment, the multi-objective loss function may include at least a frequency-domain normalization loss, which is the difference between the ACLR of the compensator input and the ACLR of the PAN output. In one embodiment, the multi-objective loss function may include at least a frequency-domain MAE calculated using the difference between the STFT of the compensator input and the STFT of the PAN output. In one embodiment, the multi-objective loss function may include at least a time-domain EVM calculated using the difference between the PAN output symbol and the ideal quadrature amplitude modulation (QAM) symbol. In one embodiment, the multi-objective loss function may include at least a time-domain MSE calculated using the difference between the compensator input and the PAN output. In one embodiment, the multi-objective loss function used to train the PAN and the compensator may include any combination of the above-described frequency-domain and time-domain losses; non-limiting examples include combinations of time-domain MSE, frequency-domain MAE, and frequency-domain normalization loss.

[0051] In one embodiment, training the compensator includes training the CGN to generate filter coefficients for a DPD that pre-distorts the signal received by the PAN. The CGN is trained to optimize the loss between the DPD's input and the PAN's output. The CGN's input may include the PAN's output and the digitally clipped output of the DPD.

[0052] In another embodiment, training the compensator includes training a CGN to generate Δ coefficients and accumulating these Δ coefficients over time to generate filter coefficients for a DPD that pre-distorts the signal received by the PAN. The CGN is trained to optimize the loss between the DPD's input and the PAN's output. In yet another embodiment, training the compensator includes training a DPD neural network that pre-distorts the signal received by the PAN. The DPD neural network is trained to optimize the loss between the DPD neural network's input and the PAN's output.

[0053] Figure 6 This is a schematic diagram illustrating a device 600 according to one embodiment, which includes a compensator to compensate for distortion caused by a power amplifier. Device 600 may be a wireless device. Device 600 includes processing hardware 630, which may include any general-purpose and / or special-purpose computing circuitry, such as a central processing circuit (CPU), graphics processing unit (GPU), digital signal processor (DSP), media processor, neural processing circuit (NPU), AI accelerator, application-specific integrated circuit (ASIC), etc. In one embodiment, processing hardware 630 may evaluate the aforementioned multi-objective loss function and neural network weight updates during training of a power amplifier neural network (PAN), a coefficient generator neural network (CGN), and / or a digital pre-distorter (DPD) neural network.

[0054] Device 600 further includes memory 620. Memory 620 may include on-chip and off-chip memory devices, such as dynamic random access memory (DRAM), static RAM (SRAM), flash memory, and other volatile or non-volatile memory devices. Memory 620 may include instructions that, when executed by processing hardware 610, cause processing hardware 610 to perform neural network training for PAN, CGN, and / or DPD neural networks. Memory 620 may also store neural network models for use as the aforementioned PAN, CGN, and / or DPD neural networks.

[0055] The neural networks or neural network modules described herein may include one or more of the following: fully-connected networks (FC), convolutional neural networks (CNN), recurrent neural networks (RNN), graph neural networks (GNN), and self-attestation-based networks (e.g., transformers). The aforementioned PAN, CGN, and / or DPD neural networks may be the same or different neural networks.

[0056] A non-limiting example of the aforementioned PAN, CGN, and / or DPD neural networks could be a CNN consisting of six convolutional layers, where the first convolutional layer has an input sequence of 128 sample points (input length). In each sample, two input channels are used for I and Q, respectively. Each layer is a one-dimensional (1-D) convolution with a kernel size of 3 and a stride of 1, and the output channel number of each layer is 16-32-64-64-128-2. Except for the last output layer, each convolutional layer is followed by a batch normalization layer and a PReLU activation function.

[0057] Device 600 also includes RF circuitry 640, which further includes at least power amplifier circuitry 610. Power amplifier circuitry 610 can be either analog or digital. Distortion caused by power amplifier circuitry 610 can be compensated by compensator 680, which can be compensator 180 (…). Figure 1A ), 280 Figure 2A ), 380 Figure 3A ) or 420 ( Figure 4A Any one of the following. The compensator 680 includes a neural network for generating filter coefficients, Δ coefficients, or the pre-distorted signal. The compensator 680 may include general-purpose or special-purpose hardware to perform the neural network operation. It is understood that... Figure 6 The embodiments described are simplified for illustrative purposes. Additional hardware components may be included.

[0058] refer to Figure 1B , 2BOne or more of the following components—3B and 4B, PAN 111, compensators (180, 280, 380, 420), CGN (130, 230, 330), and DPD (120, 220, 320)—may be implemented in hardware circuitry, software executed by the hardware circuitry, or a combination of hardware and software. The hardware circuitry may be dedicated or general-purpose hardware. The software may be stored on any non-transitory computer-readable medium for use by device 600 or by methods executed by device 600.

[0059] This invention proposes a learning-based architecture for power amplifier compensation. This architecture improves the performance of power amplifiers used in 5G communication networks. The framework uses a deep neural network (DNN) to learn the behavior and characteristics of the power amplifier. A trained neural network representing a non-differentiable power amplifier circuit is used, and the corresponding pre-distortion compensation is then learned in an end-to-end training paradigm. Furthermore, two frequency domain losses (i.e., MAE and ACLR-based canonical loss) can be simultaneously used to minimize spectral regeneration and optimize circuit design according to communication specifications.

[0060] This document has described various functional components, blocks, or modules. As those skilled in the art will understand, functional blocks or modules can be implemented by circuitry (dedicated or general-purpose circuitry that operates under the control of one or more processors and coded instructions), which typically includes multiple transistors configured to control the operation of the circuitry according to the functions and operations described herein.

[0061] While the invention has been described with reference to several embodiments, those skilled in the art will recognize that the invention is not limited to the described embodiments and can be practiced with modifications and variations within the spirit and scope of the appended claims. This description is therefore to be considered illustrative rather than restrictive.

Claims

1. A method for compensating for distortion in a power amplifier, characterized in that, include: The power amplifier neural network PAN is trained to model the power amplifier circuit using predetermined input and output signal pairs that characterize the power amplifier circuit. A compensator is trained to predistort the signal received by the PAN, wherein the compensator uses a trained neural network to optimize the loss between the compensator input and the PAN output, and the loss is calculated based on a multi-objective loss function comprising: one or more time-domain loss functions and / or one or more frequency-domain loss functions; and Signal compensation is performed using a trained compensator to output a pre-distorted signal to the power amplifier circuit; Training the compensator includes: A coefficient generator neural network (CGN) is trained to generate filter coefficients for a digital predistorter (DPD) that predistorts the signal received by the PAN. The CGN is trained to optimize the loss between the input of the DPD and the output of the PAN. The input of the CGN includes the PAN output and the digitally clipped output of the DPD. or, The coefficient generator neural network (CGN) is trained to generate Δ coefficients; and the Δ coefficients are accumulated over time to generate filter coefficients for a digital predistorter (DPD) that predistorts the signal received by the PAN, wherein the CGN is trained to optimize the loss between the input of the DPD and the output of the PAN.

2. The method according to claim 1, characterized in that, The multi-objective loss function includes at least a frequency domain normalization loss, which is the difference between the adjacent channel leakage power ratio (ACLR) at the input of the compensator and the ACLR at the output of the PAN, wherein the ACLR is the ratio of the filtered average power centered on the assigned channel frequency to the filtered average power centered on the adjacent channel frequencies.

3. The method according to claim 2, characterized in that, The multi-objective loss function used to train the PAN and the compensator includes at least the frequency domain normalized loss, the frequency domain mean absolute error (MAE), and the time domain mean square error (MSE), wherein the MAE is calculated using the difference between the short-time Fourier transform (STFT) of the compensator input and the STFT of the PAN output, and the MSE is calculated using the difference between the compensator input and the PAN output.

4. The method according to claim 1, characterized in that, The multi-objective loss function includes at least the frequency domain average absolute error (MAE) calculated using the difference between the short-time Fourier transform (STFT) of the compensator input and the STFT of the PAN output.

5. The method according to claim 1, characterized in that, The multi-objective loss function includes at least the temporal error vector magnitude EVM calculated using the difference between the PAN output symbol and the ideal quadrature amplitude modulation (QAM) symbol.

6. The method as described in claim 1, characterized in that, The multi-objective loss function includes at least the time-domain mean square error (MSE) calculated using the difference between the compensator input and the PAN output.

7. A system for compensating for distortion in a power amplifier, characterized in that, include: Memory, storing neural network models; as well as Processing hardware, coupled to the memory, is used for: The power amplifier neural network PAN is trained to model the power amplifier circuit using predetermined input and output signal pairs that characterize the power amplifier circuit. A compensator is trained to predistort the signal received by the PAN, wherein the compensator uses a trained neural network to optimize the loss between the compensator input and the PAN output, the loss being calculated based on a multi-objective loss function comprising: one or more time-domain loss functions and / or one or more frequency-domain loss functions; and The signal is compensated by the trained compensator, thereby outputting a pre-distorted signal to the power amplifier circuit; The processing hardware is further used for: A coefficient generator neural network (CGN) is trained to generate filter coefficients for a digital predistorter (DPD) that predistorts the signal received by the PAN. The CGN is trained to optimize the loss between the input of the DPD and the output of the PAN. The input of the CGN includes the PAN output and the digitally clipped output of the DPD. or, The coefficient generator neural network (CGN) is trained to generate Δ coefficients; and the Δ coefficients are accumulated over time to generate filter coefficients for a digital predistorter (DPD) that predistorts the signal received by the PAN, wherein the CGN is trained to optimize the loss between the input of the DPD and the output of the PAN.

8. The system according to claim 7, characterized in that, The power amplifier circuit is a digital circuit.

9. The system according to claim 7, characterized in that, The power amplifier circuit is an analog circuit.

10. The system according to claim 7, characterized in that, The multi-objective loss function includes at least a frequency domain normalization loss, which is the difference between the adjacent channel leakage power ratio (ACLR) at the input of the compensator and the ACLR at the output of the PAN, wherein the ACLR is the ratio of the filtered average power centered on the assigned channel frequency to the filtered average power centered on the adjacent channel frequencies.

11. The system according to claim 7, characterized in that, The multi-objective loss function includes at least the frequency-domain average absolute error (MAE) calculated based on the difference between the short-time Fourier transform (STFT) of the compensator input and the STFT of the PAN output.

12. The system as described in claim 7, characterized in that, The multi-objective loss function includes at least the temporal error vector magnitude EVM calculated based on the difference between the PAN output symbol and the ideal quadrature amplitude modulation (QAM) symbol.

13. The system according to claim 7, characterized in that, The multi-objective loss function includes at least the time-domain mean square error (MSE) calculated based on the difference between the compensator input and the PAN output.

Images