Machine learning based digital predistortion for power amplifiers

Abstract

Example embodiments relate to machine learning based power amplifier digital predistortion. A device can utilize a power amplifier to amplify and transmit a signal. The signal can be received by an internal feedback receiver of the device. The device can further include a first machine learning model configured to emulate an external feedback receiver and generate an emulated feedback signal based on the internal feedback signal. The device can further include a second machine learning model configured to determine a digital predistortion parameter for the power amplifier based on the emulated feedback signal. Apparatuses, methods, and computer programs are disclosed.

Description

TECHNICAL FIELD

[0001] The present application relates generally to power amplifiers. In particular, some example embodiments of the present invention relate to machine learning based digital predistortion for power amplifiers. BACKGROUND

[0002] Power amplifiers affect the performance and throughput of a communication system. For example, the nonlinearity of a power amplifier can generate re-growth of the spectrum, which can result in adjacent channel interference or violate regulatory body prescribed out-of-band emission standards. In addition, the nonlinearity can cause in-band distortion that degrades the magnitude of the error vector and ultimately the bit error rate (BER) and data throughput. Machine learning (ML) or other automated processes can be used for different applications in different types of devices, such as, for example, mobile phones. Generally, machine learning enables a computational model, e.g., a neural network, to be trained to perform a specific task on input data. SUMMARY

[0003] This summary is provided to introduce a selection of concepts in a simplified form that are further described below in the detailed description. This summary is not intended to identify key features or essential features of the claimed subject matter, nor is it intended to be used to limit the scope of the claimed subject matter.

[0004] Example embodiments improve the determination of digital predistortion parameters for power amplifiers. This advantageous effect can be achieved by the features of the independent claims. Further, implementation forms are provided in the independent claims, the description and the drawings.

[0005] According to an aspect, there is provided an apparatus, which can comprise a transmitter configured to transmit a signal, wherein the transmitter comprises a power amplifier; an internal feedback receiver configured to receive the transmitted signal to obtain an internal feedback signal; a first machine learning model configured to emulate an external feedback receiver and generate an emulated feedback signal based on the internal feedback signal; and a second machine learning model configured to determine a digital predistortion parameter for the power amplifier based on the emulated feedback signal.

[0006] According to an aspect, there is provided a method, which can comprise transmitting a signal, wherein the signal is amplified with a power amplifier; receiving the transmitted signal with an internal feedback receiver to obtain an internal feedback signal; emulating an external feedback receiver and generating an emulated feedback signal based on the internal feedback signal with a first machine learning model; and determining a digital predistortion parameter for the power amplifier based on the emulated feedback signal with a second machine learning model.

[0007] According to an aspect, there is provided a system, which can include a reference device configured to transmit a set of test signals, the reference device comprising an internal feedback receiver configured to receive the set of test signals to obtain an internal feedback signal; an external feedback receiver configured to receive the set of test signals to obtain an external feedback signal; means for training a first baseline machine learning model to emulate the external feedback receiver and to generate an emulated feedback signal based on the internal feedback signal, wherein the training of the first baseline model is based on the internal feedback signal and the external feedback signal received from the external feedback receiver; means for training a second baseline machine learning model to determine digital pre-distortion parameters for a power amplifier of the reference device based on the emulated feedback signal, wherein the training of the second baseline machine learning model is based on the emulated feedback signal and the set of test signals.

[0008] According to an aspect, there is provided a method, which can include transmitting, by a reference device, a set of test signals, the reference device comprising an internal feedback receiver configured to receive the set of test signals to obtain a set of internal feedback signals; training a first baseline machine learning model to emulate an external feedback receiver and to generate an emulated feedback signal based on the internal feedback signal, wherein the training of the first baseline model is based on the internal feedback signal and an external feedback signal received from the external feedback receiver, the external feedback signal corresponding to the set of test signals; training a second baseline machine learning model to determine digital pre-distortion parameters for a power amplifier of the reference device based on the emulated feedback signal, wherein the training of the second baseline machine learning model is based on the emulated feedback signal and the set of test signals.

[0009] According to an aspect, there is provided a computer program, which can include instructions for causing an apparatus to perform at least the following: transmitting a signal, wherein the signal is amplified with a power amplifier; receiving, by an internal feedback receiver, the transmitted signal reception to obtain an internal feedback signal; emulating an external feedback receiver and generating, with a first machine learning model, an emulated feedback signal based on the internal feedback signal; determining, with a second machine learning model, digital pre-distortion parameters for the power amplifier based on the emulated feedback signal.

[0010] According to an aspect, there is provided a computer program, which can comprise instructions for causing an apparatus to perform at least the following: training a plurality of devices with a subset of test signals, wherein the training of each device of the plurality of devices comprises: initializing an instance of a first machine learning model based on a first baseline machine learning model; initializing an instance of a second machine learning model based on a subset of a second baseline machine learning model; training the instance of the first machine learning model to emulate an external feedback receiver and to generate an emulated feedback signal based on a device-specific internal feedback signal received from an internal feedback receiver of the device, wherein the training of the instance of the first machine learning model is based on the device-specific internal feedback signal and an external feedback signal received from the external feedback receiver; and training the instance of the second machine learning model to determine a digital predistortion parameter for a power amplifier of the device based on the emulated feedback signal, wherein the training of the second baseline machine learning model instance is based on the emulated feedback signal and the subset of test signals.

[0011] According to an aspect, there is provided an apparatus, which can comprise at least one processor, and at least one memory including a computer program code; the at least one memory and the computer program code configured to, with the at least one processor, cause the apparatus at least to: transmit, by a transmitter, a signal, wherein the transmitter comprises a power amplifier; receive, by an internal feedback receiver, the transmitted signal to obtain an internal feedback signal; emulate, by a first machine learning model, an external feedback receiver and generate, by the first machine learning model, an emulated feedback signal based on the internal feedback signal; and determine, by a second machine learning model, a digital predistortion parameter for the power amplifier based on the emulated feedback signal

[0012] A number of additional features will be more readily understood as the same becomes better understood by reference to the following detailed description when considered in connection with the accompanying drawings. BRIEF DESCRIPTION OF DRAWINGS

[0013] In order to provide a further understanding of example embodiments of the application and to constitute a part of the detailed description, the accompanying drawings are included in which:

[0014] Figure 1 FIGURE 1 illustrates an example of a digital predistortion system according to an example embodiment

[0015] Figure 2 FIGURE 2 illustrates an example of a digital predistortion system for sub-6 GHz limited bandwidth according to an example embodiment

[0016] Figure 3 FIGURE 3 illustrates an example of a digital predistortion system for high bandwidth and millimeter wave support according to an example embodiment FIGURE 4 illustrates an example of a digital predistortion system for high bandwidth and millimeter wave support according to an example embodiment

[0017] Figure 4 The illustration shows an example of a device configured to practice one or more example embodiments.

[0018] Figure 5 An example of a neural network according to an exemplary embodiment is illustrated.

[0019] Figure 6 An example of a basic computing unit according to an exemplary embodiment is illustrated.

[0020] Figure 7 An example of a convolutional neural network according to an exemplary embodiment is illustrated.

[0021] Figure 8 An example of a three-stage approach to digital predistortion calibration according to an example embodiment is illustrated.

[0022] Figure 9 The illustration shows an example used to train a machine learning-based digital predistortion model.

[0023] Figure 10 The illustration shows an example of applying a general adversarial neural network (GAN) to train a digital predistortion system according to an example embodiment.

[0024] Figure 11 The illustration shows an example of training a machine learning-based digital predistortion model according to an example embodiment.

[0025] Figure 12 The illustration depicts an example of instance-passive learning for a machine learning-based digital predistortion system according to an example embodiment.

[0026] Figure 13 The illustration shows an example flowchart of general training and device-specific training of a machine learning-based digital predistortion system according to an example embodiment.

[0027] Figure 14 The illustration shows an example of the deployment and retraining of a machine learning-based digital predistortion system according to an example embodiment.

[0028] Figure 15 The illustration shows another example of the deployment and retraining of a machine learning-based digital predistortion system according to the example embodiment.

[0029] Figure 16 The illustration shows an example of a method for applying a machine learning-based digital predistortion system according to an example embodiment.

[0030] Figure 17 An example of a method for training a machine learning-based digital predistortion system according to an exemplary embodiment is illustrated.

[0031] Similar reference numerals are used to designate similar parts in the figures. Detailed Implementation

[0032] References will be made in detail in the exemplary embodiments, examples of which are illustrated in the accompanying drawings. The following detailed description of the drawings is intended as a description of the presented examples and is not intended to represent the only form in which this example can be constructed or utilized. The description clarifies the function of the example and the feasible order of steps for constructing and operating the example. However, the same or equivalent function and order may be implemented by other examples.

[0033] Nonlinear power amplifiers can produce various side effects. To reduce nonlinearity, power amplifiers can be operated at low power, such as in back-off mode within the linear portion of their operating curve. However, some transmission systems, such as Wideband Code Division Multiple Access (WCDMA) systems and Orthogonal Frequency Division Multiplexing (OFDM) systems, such as Wireless Local Area Networks (WLANs), 3GPP Long Term Evolution (LTE), and 3GPP 5G New Radio (NR), can cause transmitted signals to have a peak-to-average power ratio (PAPR), meaning the signal waveform has large variations. Therefore, the power amplifier may need to back off to significantly below its maximum saturation output power to avoid distortion of the peak values. This can result in low efficiency, for example, below 10%. With more than 90% of the power lost and converted into heat, the amplifier's performance, reliability, and operating costs can be severely degraded. Therefore, example embodiments of this disclosure provide a machine learning-based digital predistortion method for power amplifier linearization.

[0034] According to an example embodiment, the device can utilize a power amplifier to amplify and transmit a signal. This signal can be received by the device's internal feedback receiver. The device may also include a first machine learning model configured to simulate an external high-quality feedback receiver and generate a simulated feedback signal based on the internal feedback signal. The external high-quality feedback receiver can be used during the production phase to train the first machine learning model to mimic the external feedback signal based on the internal feedback signal. The device may also include a second machine learning model configured to determine digital predistortion parameters for the signal amplifier based on the simulated feedback signal. Therefore, even after the device is deployed in the field, sufficient quality of the feedback signal can be maintained regardless of the limitations of the internal feedback receiver, thereby improving the linearity of the power amplifier.

[0035] Figure 1An example of a digital predistortion system 100 according to an exemplary embodiment is illustrated. The digital predistortion system 100 can be embodied in a device, such as a transmitter. An input signal x(t) can be provided to a digital predistorter (DPD) 102, which can be configured to nonlinearly amplify the compressed nonlinearity of a supplemental power amplifier (PA) 104 to the predistorted signal z(t). As a result, the output y(t) of the cascaded DPD 102 and PA 104 is linearized with respect to the input x(t), as shown below. Figure 1 The diagram illustrates this. Ideally, y(t) is strictly linear with respect to x(t). However, it should be understood that in practical applications, linearizing a power amplifier can include making the output of the power amplifier generally or sufficiently linear, for example, for a specific application. DPP 102 and PA 104 can be utilized up to their saturation point while maintaining sufficient linearity. Therefore, the transmitter's output power capacity can be increased for a given linearity target, or improved linearity can be used to increase the transmitter's efficiency at a given backoff output power, for example, by rebiasing PA 104 for a lower saturation point. DPD 102 can be approximated as the inverse of PA 104.

[0036] The DPD 102 can be implemented based on a memoryless model suitable for power amplifiers with memoryless nonlinearity. In this case, the output of PA 104 can depend on the current input. This instantaneous nonlinearity can be characterized by the AM / AM (amplitude modulation / amplitude modulation) and AM / PM (amplitude modulation / phase modulation) responses of PA 104, where the amplitude and phase deviation of the output signal of the power amplifier can be provided as a function of its current input amplitude. For example, a memoryless model can be implemented using algorithm-based memoryless polynomials and lookup tables (LUTs).

[0037] However, a simple memoryless model may be insufficient for systems with wide bandwidths where the power amplifier may exhibit memory effects, for example, due to the thermal constant of active devices or bias components with frequency-dependent behavior. The output of PA 104 may additionally depend on past input values, and therefore the memoryless DPD 102 may not be able to adequately linearize PA 104.

[0038] The method used to linearize such a power amplifier is to apply a DPD that includes a memory structure. Examples of such algorithms include those based on Voltra series and their derivatives, such as Wiener, Hammerstein, Wiener-Hammerstein, or parallel Wiener structures and memory polynomial models.

[0039] One approach to constructing a memory-based DPD 102 is to directly find the inverse of PA 104 (DLA, Direct Learning Architecture). However, this may not directly yield the inverse of a nonlinear system with memory. Another approach is the Indirect Learning Architecture (IDLA), which allows for the avoidance of model assumptions and parameter estimations related to PA 104.

[0040] Figure 2 An example of a digital predistortion (DPD) system 200 for a limited bandwidth below 6 GHz, according to an exemplary embodiment, is illustrated. This example DPD system may be suitable for frequencies below 6 GHz and examples with limited signal bandwidth (e.g., <100 MHz). An example device, such as a user equipment (UE) of a cellular communication system, may include a baseband circuitry system 210 and an RF (radio frequency) circuitry system 220. The RF circuitry system 220 may include a digital front-end (DFE), a digital-to-analog converter (DAC) and an analog-to-digital converter (ADC), a mixer 222, and a power amplifier (PA) 224. One or more RF oscillators may be coupled to the mixer 222. The RF circuitry system 220 may be configured to transmit RF signals via an antenna 230. The DPD system 200 may be based on the transmission of a known reference signal at the factory. A built-in feedback receiver (FRx) may be configured to capture the transmitted signal and transmit the signal, for example, via the DFE 226, back to the baseband circuitry system 210 for calculating DPD coefficients. This workaround can be applied when performance changes significantly, such as due to unforeseen aging or the effects of the external environment, or if sufficient power back-off is incorporated to absorb environmental impacts and maintain compliance with specifications.

[0041] Figure 3 An example of a digital predistortion system for high bandwidth and millimeter wave support, according to an exemplary embodiment, is illustrated. A device, such as a UE, may include baseband circuitry 310, intermediate frequency (IF) circuitry system 320, and RF circuitry system 330. IF circuitry system 320 may include a DFE, DAC, mixer, IF oscillator, and / or power amplifier to provide and receive IF signals (IF / V, IF / H) to / from RF circuitry system 330. Furthermore, clock signals (CLK) and control data (CTRL) may be exchanged between IF circuitry system 320 and RF circuitry system 330. The RF circuitry may include any suitable components, such as a PA, mixer, RF oscillator, phase shifter (PS), and power detector (PD), to transmit and receive signals via an antenna module coupled to RF circuitry system 330.

[0042] The DPD system 300 may also include an external test device 340 (test box). This external test device 340 can be configured to capture transmitted signals with a bandwidth higher than the transmitted signal bandwidth, for example, up to three times the bandwidth. The external test device 340 can transmit the signal back to the baseband circuitry 310 to calculate the DPD coefficients. However, while this approach can calculate the DPD coefficients more accurately, it does not account for dynamic changes, such as the constantly changing load effects during effective operation after the equipment is deployed. In this case, a static set of DPD coefficients may not provide sufficient accuracy. Furthermore, using a similar... Figure 2 Such a system might be impossible because, during field operation, it might not be able to capture a sufficiently large bandwidth with the required dynamic range online using a built-in feedback receiver. Therefore, for example, using a system similar to... Figure 2 The feedback receiver (FRx) may not be able to properly capture the antenna load effect.

[0043] Typically, methods for determining DPD parameters, such as the mmWave DPD (mDPD) coefficients, can be adversely affected by the following factors:

[0044] 1) High Bandwidth. At high bandwidths, PAs can exhibit a storage effect, and therefore may require detailed representative data of real-time operating conditions to characterize them. Furthermore, the number of DPD coefficients can be significantly higher compared to DPDs without a storage effect (e.g., due to lower bandwidth). For example, millimeter wave in 5G NR Release 15 is defined as 400MHz with carrier aggregation (CA) options, allowing even early devices to support bandwidths of 400-1200MHz. In DPD applications, the measurement bandwidth used to characterize DPDs can even be more than three times the bandwidth of the transmitted signal. Therefore, a high dynamic range measurement receiver on silicon with a bandwidth of 2.4-3.6GHz may be impractical.

[0045] 2) Antenna Load Effect: For millimeter-wave operation, the antenna may include an active antenna built into the final industrial design. This means that users of the operating equipment may touch the antenna (or the material covering it), thus affecting the antenna load. Once this antenna load changes, the PA characteristics will be affected, and the DPD coefficients may need to be updated. Therefore, static solutions characterized in the lab or on the production line may be insufficient. If real-time operation is not supported for DPD training, the PA design may need to establish a wide buffer power margin for load mismatch scenarios, which can lead to suboptimal performance compared to normal conditions.

[0046] 3) Device Reference Signal. Due to spurious emission requirements, UEs may be required not to transmit above approximately -47 dBm / 1 MHz, unless permitted by examples in the 3GPP specification. A suitable non-3GPP standard UE TX reference / test signal can be used to characterize the PA's behavior. The power of this test signal may need to represent real-time UL transmissions to place the PA under proper large-signal conditions during training. This may only be possible in a laboratory or UE generation setting, as the final device may not be able to use a large reference / test signal in the field.

[0047] 4) Online Measurement: During PA and mDPD characterization, the device can operate in SISO (Single-Input Single-Output) mode. This allows the UE to transmit on one MIMO (Multiple-Input Multiple-Output) branch and use other MIMO branches to receive feedback signals to complete PA characterization. However, for uplink MIMO, this workaround may not work because both available TX branches may be occupied.

[0048] Therefore, the example embodiments disclosed herein provide a machine learning (ML)-based DPD architecture that enables more accurate determination of DPD parameters after the UE is deployed in the field. This example embodiment can be applied to an example system where the bandwidth of the built-in feedback receiver is insufficient for accurate DPD parameter determination. This example embodiment is able to simulate high-fidelity, such as high-resolution (high-bandwidth) feedback signals in the absence of external test equipment, taking into account antenna load effects and aging effects, using a staged training method where the architecture of the reference device can be transmitted to different UEs during production and in the field, and online training is triggered once the UE is in operation in the field.

[0049] Figure 4 An example of device 400 is illustrated, configured to practice one or more example embodiments. Device 400 may include at least one processor 402. The at least one processor may include, for example, one or more various processing devices, such as a coprocessor, microprocessor, controller, digital signal processor (DSP), processing circuitry with or without an accompanying DSP, or various other processing devices including integrated circuits, such as application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), microcontroller units (MCUs), hardware accelerators, special purpose computer chips, or similar devices.

[0050] The device may also include at least one memory 404. The memory may be configured to store, for example, computer program code or similar items, such as operating system software and application software. The memory may include one or more volatile memory devices, one or more non-volatile memory devices, and / or combinations thereof. For example, the memory may be embodied as a magnetic storage device (such as a hard disk drive, floppy disk, magnetic tape, etc.), an optical storage device, or a semiconductor memory, such as a mask ROM, PROM (programmable ROM), EPROM (erasable PROM), flash memory, RAM (random access memory), etc.

[0051] Device 400 may also include a communication interface 408 configured to enable device 400 to transmit and / or receive information, such as signals according to one or more wireless communication standards. The communication interface 408 may be configured to provide at least one radio connection, such as a 3GPP mobile broadband connection (e.g., 3G, 4G, 5G); a wireless local area network (WLAN) connection, such as a connection standardized by the IEEE 802.11 series or the Wi-Fi Alliance; a short-range wireless network connection, such as Bluetooth, NFC (Near Field Communication), or RFID; a local wired connection, such as a local area network (LAN) connection or a Universal Serial Bus (USB) connection; or a wired internet connection.

[0052] The device 400 may also include a user interface 410, which includes input and / or output devices. The input devices can take various forms, such as a keyboard, touchscreen, or one or more embedded control buttons. The output devices may include, for example, a display, speaker, vibration motor, or similar items.

[0053] When the device is configured to perform certain functions, some components and / or components of device 400, such as at least one processor 402 and / or memory 404, can be configured to perform those functions. Furthermore, when at least one processor 402 is configured to perform certain functions, those functions can be implemented using program code 406, including, for example, in memory 404.

[0054] The functions described herein can, at least in part, be implemented by one or more computer program product components, such as software components. According to an example embodiment, the device includes a processor or processor circuitry, such as a microprocessor, configured by program code at execution time to perform the operations and functions. Additionally, or furthermore, the functions described herein can, at least in part, be implemented by one or more hardware logic components. For example, and not limited to, illustrative types of hardware logic components that can be used include programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), system-on-a-chip (SoCs), complex programmable logic devices (CPLDs), and graphics processing units (GPUs).

[0055] The apparatus 400 may include components for implementing at least one method described herein. In one example, the components include instructions for a processor 402, at least one memory 404, and program code 406, the at least one memory 404 and the program code 406 being configured with the at least one processor 402 to cause the apparatus 400 to implement the method.

[0056] Device 400 may include computing devices such as mobile phones, tablets, laptops, Internet of Things (IoT) devices, or such items. Examples of IoT devices include, but are not limited to, customer electronics, wearable devices, and smart home devices. In one example, device 400 may include vehicles, such as automobiles. Although device 400 is illustrated as a single device, it is understood that the functionality of device 400 may be distributed across multiple devices, as long as applicable herein, for example, by implementing the example embodiment as a cloud computing service.

[0057] Figure 5 An example of a neural network according to an exemplary embodiment is illustrated. The neural network may include a computation graph having several computational layers. For example, neural network 500 may include an input layer, one or more hidden layers, and an output layer. Nodes i1 to i2 of the input layer are shown. n It can be connected to one or more (n) nodes in the m nodes of the first hidden layer. 11 up to n 1m A node in the first hidden layer can be connected to one or more (n) nodes in the k nodes of the second hidden layer. 21 up to n 2k Connected to each other. It is understandable that, despite... Figure 5 The neural network example illustration shows two hidden layers, but neural networks can apply any number and type of hidden layers. A neural network 500 can also include an output layer. The nodes of the final hidden layer, such as... Figure 5 Example, node n in the second hidden layer 21 up to n 2k It can be connected to one or more nodes (o1 to o) in the output layer.j The nodes are connected. It's worth noting that the number of nodes in each layer of the network can vary. A node can also be called a neuron, computational unit, or basic computational unit. The terms neural network, neural mesh, network, and model are used interchangeably. Machine learning models or algorithms can include neural networks, but machine learning can also be implemented as any other suitable learning model. The weights of a neural network can be called learnable parameters or simply parameters. Figure 5 In the example, one or more of these layers can be comprehensive connection layers, where each node in the layer is connected to every node in the previous layer.

[0058] Feedforward neural networks are an example of neural network architectures. Any feedback loop can exist in a feedforward neural network. Each layer can take input from one or more preceding layers and provide its output to one or more subsequent layers. Furthermore, individual nodes in some layers may take input from nodes in one or more preceding layers and provide their output to nodes in one or more subsequent layers. In linear bypass network architectures, one or more nodes in the input layer can be connected to one or more nodes in the output layer. In digital predistortion applications of power amplifiers, this allows hidden layers to focus on compensating for the nonlinearity of the power amplifier.

[0059] Figure 6 An example of a neural network node 601 according to an exemplary embodiment is illustrated. This node 601 can be configured to receive one or more inputs (a1 to a1) from one or more nodes in one or more preceding layers. n The node calculates the output based on the received input values. Generally, a node can also receive feedback from one or more nodes in one or more subsequent layers. Inputs can be associated with parameters to adjust the effect of a particular input on the output. For example, with inputs a1 to a... n The associated weights w1 to w n It can be used to multiply input values ​​a1 to a n Node 601 can also be configured to combine inputs and outputs, or to activate. For example, node 601 can be configured to aggregate modified input values. A bias or offset b can also be applied to add a constant to the modified input combination. Weights and biases can be learnable parameters. For example, when the neural network is trained for a specific task or situation, the values ​​of the weights and / or biases associated with different inputs and different nodes can be updated to reduce the error associated with performing the task to an acceptable level.

[0060] Furthermore, the activation function f() can be applied to control the timing and manner in which node 601 provides output. The activation function can be, for example, a nonlinear function that is substantially linear in the region of 0, but limits the node's output as the input increases or decreases. Examples of activation functions include, but are not limited to, the step function, the sigmoid function, the hyperbolic tangent function, the ReLU (Rectified Linear Unit) function, and the softmax function. The output can be provided to nodes in one or more subsequent layers of the network, and / or to one or more nodes in one or more preceding layers of the network.

[0061] Forward propagation or forward pass may include feeding input data sets through layers of neural network 500 and generating outputs. In this process, the weights and biases of neural network 500 affect the excitation of individual nodes, thereby affecting the output provided by the output layer.

[0062] One property of neural networks and other machine learning tools is their ability to learn properties from input data, whether in a supervised or unsupervised manner. Learning can be based on the network taught by the training algorithm or on training signals provided by a meta-neural network.

[0063] Typically, training algorithms can involve modifying some properties of a neural network so that its output is as close as possible to the desired output. For example, in the case of DPD parameter generation, the neural network can be trained to provide an output that produces the desired (linear) output compared to a reference power amplifier. During training, the produced output can be compared to the desired output to calculate an error value, or loss value. This error can be calculated based on a loss function. The neural network can then be updated based on calculating the derivatives with respect to the network's learnable parameters. This can be done using an example of backpropagation, which determines the gradient for each layer of the network, starting from the last layer, until the gradients of the learnable parameters are obtained. The parameters of each layer can then be updated accordingly to iteratively reduce the loss. An example of this loss is the mean squared error (MSE) between the system output and the desired output data. In deep learning, training can include an iterative process where, in each iteration, the algorithm modifies the parameters of the neural network to gradually improve the network's output, i.e., gradually reduce the loss.

[0064] Figure 7An example of a convolutional neural network 700 is illustrated. The convolutional neural network 700 includes at least one convolutional layer. The convolutional layer can perform convolution operations to extract information from input data 702 to form multiple feature maps 706. Feature maps can be generated by applying filters or kernels to a subset of the input data, such as block 704 of input data 702, and by sliding the filters through the input data to obtain the value of each element of the feature map. The filter can include a matrix or tensor, for example, it can be multiplied by the input data to extract features corresponding to that filter. Multiple feature maps can be generated based on the application of multiple filters. Further convolutional layers can take the feature maps from the previous layer as input and apply the same filtering principles to feature maps 706 to generate another set of feature maps 708. Filter weights can be scientific parameters, and they can be updated during the training phase, similar to the parameters of neural network 500. Similar to node 600, activation functions can be applied to the outputs of the filters. The convolutional neural network 700 may also include one or more other types of layers, such as fully connected layers 710 before, after, and / or between convolutional layers. The input is provided by output layer 712. For the DPD parameter generation task, the input can be a signal sample from the built-in feedback receiver, which will be discussed further below, and the output can include DPD coefficients suitable for the power amplifier characteristics.

[0065] Figure 8 The illustration depicts an example of a three-stage method for digital predistortion calibration according to an exemplary embodiment. Stage 1 may include a laboratory characterization phase. Stage 1 may be performed on a reference device or a group of reference devices, for example, in a laboratory environment prior to initializing mass production. In other words, during this phase, the device under test (DUT) may serve as the reference device. The reference device may be referred to as a reference cell, gold cell, or reference UE. The objective of Stage 1 may be to train a first machine learning algorithm to mimic a high-fidelity (hi-fi) signal based on a reduced-fidelity feedback signal. Another objective is to train a second ML model to learn the behavior of the reference device PA. The first ML model and / or the second ML model may be used as multiple baseline ML models for subsequent fine-tuning of multiple corresponding ML modules in different devices.

[0066] Phase 1 may include training the ML model on various parameter sets using reference devices. For example, exhaustive training may be performed on a range configured for the parameter sets. Therefore, in Phase 1, the stimulus may include full test coverage. The feedback (FB) path may include a feedback path via an external test box (refer to external test device 340) and a built-in feedback receiver from an onboard feedback path. The external test box may provide a high-fidelity signal, and the built-in feedback receiver may provide a reduced-fidelity signal. The output of Phase 1 may include a baseline DPD coefficient set.

[0067] Phase 2 may include UE generation testing. In Phase 2, the training methods of Phase 1 can be applied to multiple devices, such as all UEs in production. However, the stimuli may include reduced test coverage, such as a subset or range of parameters used in Phase 1. Feedback paths may include paths from external test boxes and onboard feedback paths, similar to Phase 1. Therefore, both high-fidelity and low-fidelity signals can be used. In Phase 2, the baseline ML model can be tailored to the characteristics of each device. Therefore, the output of Phase 2 may include a device-specific set of DPD coefficients.

[0068] Phase 3 may include a field-on-field adjustment phase. Phase 3 may be performed on one or more of a plurality of devices. In this phase, the device may have already been deployed to the field and used by a customer or user. Therefore, the stimulus includes a real-time transmitted signal and a feedback path including a loaded feedback path. The feedback path via an external test box may be unavailable. Therefore, the device may need to perform field-on-field adjustment by retraining the first ML model and / or the second ML model based on a defidelity signal. Reliable retraining based on the defidelity feedback signal is achieved by the first ML model, which was trained by Phases 1 and 2 to mimic the output of the external test box, i.e., the high-fidelity signal. The output domain of Phase 3 includes device-specific DPD coefficients adapted to the current conditions.

[0069] Figure 9 An example of a system for training a machine learning-based digital predistortion model is illustrated according to an exemplary embodiment. The system 900 may include a baseband signal generator configured to generate a signal (e.g., a test signal s). n The signal may include a baseband signal. The signal may be described by one or more signals relating to environmental parameters, such as at least one of the following: transmission power, signal bandwidth, beam configuration, modulation and / or error correction coding scheme, carrier frequency, temperature, battery level, or antenna load. Therefore, the signal may be described by s n The parameters are represented as (a, b, c, d, e, f, g, h), where the parameter ah represents a signal parameter. In this example embodiment, the number of parameters is 8; however, the number of parameters is variable. The baseband signal generator 901 can generate a set of test signals with varying parameters and / or parameter values.

[0070] The system's RF chain may include a power amplifier 902, which may cause nonlinearity in the signal. Therefore, the system 900 may include an internal feedback receiver 904 (iRX) configured to receive the transmitted signal to obtain an internal feedback signal y. n(a, b, c, d, e, f, g, h). The amplified signal can be radiated using antenna 903. However, the iRX 904 can receive the amplified signal internally without a radio interface. The iRX 904 can also be referred to as a built-in (feedback) receiver or an inherent (feedback) receiver.

[0071] System 900 may also include an external feedback receiver 914 (eRX), which can be configured to receive signals via a radio interface, for example via antenna 913, to obtain an external feedback signal r. n (a,b,c,d,e,f,g,h). The eRX 914 can also be referred to as an external (feedback) receiver. The eRX 914 can be configured to receive a set of test signals to obtain multiple external feedback signals, for example, during Phase 1 and Phase 2.

[0072] System 900 may also include a first machine learning model, ML-hi-fi model 905 (ML high-fidelity model), which can be configured to simulate eRX 914 and is based on the internal feedback signal y. n (a,b,c,d,e,f,g,h) generates the simulation feedback signal g. n (a,b,c,d,e,f,g,h). The feedback signal g in this simulation. n (a,b,c,d,e,f,g,h) may include a regenerated version of the signal received by the eRX914 via the antenna 913.

[0073] As discussed above, the capabilities of the iRX 904 may be more limited than those of the eRX 914. For example, the eRX 914 can be located at an external test device, which is not limited in terms of cost or size, for example. In contrast, the iRX 914 can be located at the device under test (DUT), such as a mobile phone. For example, the bandwidth of the iRX 904 may be lower than that of the eRX 904. Furthermore, the dynamic range of the iRX 904 may be lower than that of the eRX 904. The iRX 904 may also be inferior with respect to one or more parameters. Therefore, the ML-hi-fi model 905 can be trained in stages 1 and 2 to mimic the external feedback signal provided by the eRX 914.

[0074] Training the ML-hi-fi model can include receiving external feedback signals r from the eRX 914. n (a,b,c,d,e,f,g,h). The training of the ML-hi-fi model 905 can be based on the internal feedback signal y. n (a,b,c,d,e,f,g,h) and external feedback signal rn (a,b,c,d,e,f,g,h). For example, training the ML-hi-fi model 905 to mimic the eRX 914 can be based on generative adversarial networks (GANs), for example, involving... Figure 10 Detailed description. However, other ML methods for manipulating and / or extrapolating samples can also be used. The output g of the ML-hi-fi model 905. n (a,b,c,d,e,f,g,h) can include those close to r n The regenerated high-fidelity signal of (a,b,c,d,e,f,g,h).

[0075] System 900 may also include a second machine learning model, ML-mDPD model 906 (ML millimeter-wave digital predistortion model). This ML-mDPD model 906 can be configured to use a simulation-based feedback signal g. n (a, b, c, d, e, f, g, h) determines at least one digital predistortion parameter for the power amplifier 902. The ML-mDPD model 906 can be implemented in any suitable manner, for example as a cascade of deep neural networks, which can be trained to learn the composite PA response and its inverse, as shown in the diagram. Figure 5 Further discussion is needed. The power amplifier 902 can be linearized based on digital predistortion parameters determined by the ML-mDPD model 906. For example, the digital predistortion parameters may include a subset or portion of the ML-mDPD 906, as discussed regarding... Figure 11 Further discussion is needed.

[0076] In Phase 1, the DUT may include a reference device. Therefore, the reference device may include a BB signal generator 901, a PA 902, an antenna 903, an iRX 904, an ML-hi-fi model 905, and an ML-mDPD model 906. During Phase 1, the reference device may be configured to transmit a set of test signals. The iRX 904 of the reference device may be configured to receive the set of test signals to obtain an internal feedback signal. A first baseline machine learning model, such as a baseline ML-hi-fi model, may then be trained to simulate the eRX 914. The baseline ML-hi-fi model may also be trained to generate the simulated feedback signal based on the internal feedback signal. The training of the baseline ML-hi-fi model may be based on the internal feedback signal and the external feedback signal received from the eRX 914 of the reference device.

[0077] Furthermore, a second baseline ML model, such as baseline ML-mDPD, can be trained to determine at least one digital predistortion parameter for the PA902 reference device. The training of the baseline ML-mDPD model can be based on the simulated feedback signal g. n(a,b,c,d,e,f,g,h), and the test signal set s n (a,b,c,d,e,f,g,h). During the inference phase, such as when using the model after deployment, the baseline ML-mDPD 906 can obtain the simulation feedback signal g. n As input, it provides digital predistortion parameters for linearizing PA 902 as output. This output may include a subset or a portion of the trained ML-mDPD model 906.

[0078] Figure 10 The illustration depicts an example of applying a generalized adversarial neural network (GAN) to train a digital predistortion system, according to an example embodiment. The GAN may include a generator neural network 1002 configured to take as input multiple internal feedback signals from an iRX 904. The GAN also includes a discriminator neural network 1004, which may be configured to take as input either generated samples from the generator neural network 1002 or real samples from an eRX 914, representing real samples or underlying real data. The discriminator neural network 1004 may be configured to alternately receive inputs from the eRX 914 and the generator neural network 1002. The task of the discriminator neural network 1004 is to evaluate whether its input is real or fake, i.e., whether the input comes from the eRX 914 or the generator neural network 1002.

[0079] In step 1006, generator 1002 and discriminator 1004 can be trained based on whether discriminator 1004 correctly estimates the input sample source. The correct source can be determined based on the position of switch 1008, which is known. Discriminator 1004 can be rewarded for correctly evaluating the input sample source, and generator 1002 can be rewarded for deceiving discriminator 1004, for example, when discriminator 1004 makes an incorrect evaluation. The reward is determined based on a function that can depend on the state and the agent's actions, such as the actions taken by discriminator 1004. The reward function, or an approximation thereof, can take any suitable form. A quirk could mean that generator 1002 is able to create a false signal that discriminator 1004 considers a true signal. The error in determining whether a false or true signal occurs relates to the quantification of the deception performance. This competitive approach allows generator 1002 to be trained to mimic the operation of eRX914 based on internal feedback signals from iRX904. After deployment, the generator neural network 1002 can be used with the ML-hi-fi model 905 to improve the quality of the internal feedback signal without needing to access the eRX914.

[0080] Figure 11The illustration shows an example of training a machine learning-based digital predistortion system according to an example embodiment. The ML-mDPD model 906 may include two ML models, ML-mDPD(PA) 1102 and ML-mDPD(PA). -1 A cascade of these structures can be trained to model the composite PA response and its inverse. This model can be implemented, for example, as a 500-layer feedforward neural network or a 700-layer k-layer convolutional neural network (each layer having N layers). k An example of a layer (e.g., 1 neuron). The activation function of the layer can be linear or nonlinear, such as ReLU, softmax, or similar. The ML-mDPD model 906 can be trained with samples collected at the output of PA 902, with the goal of minimizing the difference, e.g., the output of ML-mDPD 906 versus signal s. n The mean squared error between undistorted samples (a, b, c, d, e, f, g, h). A subset or part of the ML-mDPD model 906 can be used as a digital predistortion model to linearize PA 902. This subset may include neural networks that inversely model the PA response, such as ML-mDPD(PA 902). -1 ) or its parameters. Therefore, the digital predistortion parameters determined by the ML-mDPD model 906 may include parameters characterizing the inverse of the response portion of the ML-mDPD model 906, ML-mDPD(PA) -1 Ideally, the output from PA 902 includes the same signal s. n (a, b, c, d, e, f, g, h). However, in practice, the output signal may include the input signal s. n Approximate values ​​of (a,b,c,d,e,f,g,h).

[0081] Figure 12 The illustration depicts an example of transfer learning for a machine learning-based digital predistortion system according to an exemplary embodiment. In phase 2, such as the UE production phase, multiple devices' ML-based digital predistortion systems can be trained such that a baseline ML model trained for a reference device is adapted to device-specific characteristics.

[0082] Reference device 1202 may include a first baseline ML model, baseline ML-hi-fi model 905-1. The first ML model, ML-hi-fi model 905-2, of the current DUT 1204 may be initialized based on the baseline ML-hi-fi model 905-1 of reference device 1202. For example, prior to fine-tuning, the ML-hi-fi model 905-2 of the current DUT may include a copy of the ML-hi-fi model 905-1 of reference device 1202.

[0083] Reference device 1202 may also include a second baseline ML model, a baseline ML-mDPD model 906-1. The second ML model, ML-mDPD model 906-2 of the current DUT 1204, may be initialized based on the baseline ML-mDPD model 906-1 of reference device 1206, for example, as a copy of ML-mDPD model 906-1.

[0084] However, according to the example embodiment, the similarity between the current DUT 1204 and the reference device 1202 can be considered when initializing the ML-mDPD model 906-2. For example, the ML-mDPD model 906-2 can be initialized using a subset of the baseline ML-mDPD model 906-1. This subset may include a subset of the layers of the baseline ML-mDPD model 906-1. This subset can be determined based on a similarity metric between the current DUT 1204 and the reference device 1202. This similarity metric can be calculated, for example, using health signals from both devices, their power amplifier batch numbers, or the like. The health signal may include indications of the performance of any transmitter block. The similarity metric can be calculated by a similarity function 1208, which may be external to the current DUT 1204 and the reference device 1202. In this case, the training system in phase 2 can determine a subset of the baseline ML-mDPD 906-1 for initializing the ML-mDPD 906-2. Alternatively, similarity function 1208 can be included in the current DUT 1204. In this case, DUT 1204 can determine a subset. Figure 12 The settings enable the ML-hi-fi 905-2 and ML-mDPD 906-2 of the DUT 1204 to be adapted to the specific characteristics of the device. Furthermore, the initialization of the ML-mDPD906-2 based on a similarity function accelerates the adaptation process while providing ample freedom for fine-tuning the ML-mDPD 906-2.

[0085] Figure 13 The illustration shows an example flowchart of general training and device-specific training for a machine learning-based predistortion system, according to an example embodiment. The operations of method 1300 can be performed in phase 1 and / or phase 2, and, for example, in a similar manner... Figure 12 It is used in the system.

[0086] In 1301, the method may include selecting one or more reference devices.

[0087] At 1302, the method may include training a baseline ML model. Operation 1302 may include transmitting a set of test signals by reference device 1202. As discussed above, the reference device may include an internal feedback receiver configured to receive the set of test signals to obtain an internal feedback signal. The method may also include training a first baseline ML model, ML-hi-fi 905-1, to simulate an external feedback receiver and generate a simulated feedback signal based on the internal feedback signal. Training of the first baseline model ML-hi-fi 905-1 may be based on the internal feedback signal and an external feedback signal received from the external feedback receiver. The external feedback signal may correspond to the set of test signals received by the external feedback receiver, for example, through a radio interface between antennas 903 and 913.

[0088] In 1303, the method may include selecting a device as the current DUT 1202.

[0089] At 1304, the method may include training a device-specific ML model. For example, the ML-hi-fi model 905-2 and ML-mDPD model 906-2 of the current DUT 1204 may be trained, as described above, to adapt the baseline ML model to the device-specific characteristics of the current DUT 1204. For example, operation 1304 may include initializing an instance of ML-hi-fi model 905-2 based on the baseline ML-hi-fi model 905-1. Operation 1304 may also include initializing an instance of ML-mDPD model 906-2 based on a subset of ML-mDPD 906-1. Operation 1304 may also determine a subset of ML-mDPD model 906-1 based on a similarity metric between the device and a reference device. The similarity metric may be determined based on at least one of the following: the batch number of the power amplifier of the reference device 1202 and the power amplifier of the current DUT 1204, and the transmission performance metrics of the reference device 1202 and the current DUT 1204. Lot numbers may include serial numbers that describe components obtained from the same primary material or similar materials within the same manufacturing cycle. Therefore, lot numbers can be used as a similarity measure between two power amplifiers. Transmission performance metrics may include indications of one or more transmission blocks or components. Transmission performance can be determined based on health signals 1206-1 and 1206-2. An example health signal is triggered when the modem records a high level of adjacent channel leakage.

[0090] The similarity metric can be determined based on the distance between the current DUT 1204 and the reference device 1202. The similarity metric can be expressed as, for example, taking three discrete values: -1 = dissimilar products, 0 = uncertain, and 1 = similar products. These three values ​​can be calculated by comparing the characteristics of the current DUT with those of the reference device. For example, similarity = 1 if {(DUT batch == reference device batch) & (DUT ACLR - reference device ACLR < threshold)}. Based on the similarity metric, all layers, a subset of layers, or no layers can be transferred when initializing the ML-mDPD model 906-2 for the current DUT. For example, if the similarity metric equals 1, no layers can be transferred; if the similarity metric equals 0, the input and output layers can be transferred; if the similarity metric equals 1, all layers can be transferred.

[0091] Operation 1304 may also include training instances of the ML-hi-fi model 905-2 to simulate an external feedback receiver and generating simulated feedback signals based on device-specific internal feedback signals received from the internal feedback receiver of the current DUT 1204. Instances of training the ML-hi-fi model 905-2 may be based on both device-specific internal feedback signals and external feedback signals received from the external feedback receiver. For example, using... Figure 10 The GAN method allows the ML-hi-fi model 905-2 to be trained to mimic an external feedback receiver. This enables a more accurate simulation of the external feedback receiver once the current DUT 1204 is deployed in the field and the external feedback signal is no longer available. For example, when the ML-hi-fi model 905-2 is used to improve the quality of the feedback signal from the internal feedback receiver, device-specific characteristics of the current DUT 1204's power amplifier can be taken into account.

[0092] Operation 1304 can also train instance ML-hi-fi model 905-2 to determine at least one digital predistortion parameter for the power amplifier used in the current DUT 1204 based on the simulation feedback signal. Training ML-mDPD 906-2 can be similar to training baseline ML-mDPD 906-1.

[0093] However, according to example embodiments, test coverage can be reduced by using a subset of the test signals during the training of ML-hi-fi 905-2 and ML-mDPD 906-2. The subset of the test signals may correspond to a subset of parameters, and / or a subset of parameter values. For example, to accelerate the adaptation process, the test signals s n(a, b, c, d, e, f, g, h) can be generated for a subset of parameter ah, for a finite range of parameters, and / or for specific values ​​of parameters. Training of an instance of ML-mDPD906-2 can be based on a subset of simulated feedback and test signals. This parameter set can include at least one of the following: transmission power, signal bandwidth, beam configuration, modulation and / or error correction coding scheme, carrier frequency, temperature, battery level, or antenna load.

[0094] At 1305, the method may include determining whether any devices remain. If all devices in the current device set have been trained, the method may terminate. If there are still devices to be trained, the process may move to operation 1306 to determine whether the baseline ML model should be updated or further reference devices should be selected. Alternatively, the process may move directly to operation 1303 to select the next device as the current DUT 1204.

[0095] At 1306, the method may include determining whether a time period and / or performance offset has been triggered. If not, the process may move to operation 1303 to select the next device as the current DUT 1204 for use with an adapted ML model that can be executed at 1304. Therefore, the method may include training multiple devices using a subset of the test signal. Training for each device may include the operations described above.

[0096] In response to the detection of a time period or performance shift triggered at 1306, the process can move to 1302 to update the baseline ML model. Devices can be selected for more extensive testing, for example, at regular periods or other predetermined intervals, or whenever a performance shift is observed. This comprehensive laboratory description can be revisited periodically to update the baseline ML model during device production (Phase 2). This can absorb any shifts due to component batch variations or other effects. Therefore, the method can include retraining the baseline ML-hi-fi model 905-1 and / or the baseline ML-mDPD model 906-1 with a set of test signals, after training a predetermined number of devices, and / or in response to detecting shifts in the power amplifier linearization performance across multiple devices. For example, if the performance of the DPD coefficients generated by a device differs from the performance of previously trained devices or the average performance of previously trained devices by a specific threshold, the system can initiate retraining of the baseline ML model to keep it updated and reduce the number of training iterations in fine-tuning the ML model for further devices.

[0097] Alternatively, from 1306, the process can move back to 1301 to select (multiple) different reference devices, such as random selection. Then, the baseline ML model can be updated based on the new reference devices at 1302, and further devices can be initialized based on the updated baseline ML model at 1304.

[0098] Method 1300 enables efficient training of ML-based digital predistorters and adapts ML models to device-specific characteristics. Furthermore, this method maintains good training efficiency for dynamic adaptation of the baseline ML model, even with variable characteristics within production devices.

[0099] Figure 14 The illustration shows an example of the deployment and retraining of a machine learning-based digital predistortion system according to an example embodiment. The illustrated procedure can be applied to Phase 3, for example, when the device is operating in the field and cannot obtain received external feedback. For example, the device could transmit a signal conforming to 3GPP specifications, rather than a test signal. n (a,b,c,d,e,f,g,h). The device may include a baseband signal generator 901, an antenna 903, an internal feedback receiver 904, and an ML-hi-fi model 905. The device may also include a digital predistortion ML model, such as ML mDPD(PA). -1 The 1104 is configured to linearize the output of the power amplifier 902. The device may also include an ML-mDPD model (not shown).

[0100] Since the device may not be connected to an external feedback receiver, the ML-hi-fi model 905 may not be trained at this stage. However, if triggered, the ML-hi-fi model 905 can receive internal feedback signals from the iPX904 and generate the simulated feedback signal g. n To mimic an external feedback receiver for retraining the ML-mDPD model. When the learning trigger is off, the device can operate according to the "deployment" branch of the procedure. Upon detecting the learning trigger or transitioning to the on state, the device can operate according to the "retraining" branch of the procedure. This learning trigger state can be obtained by the network, for example, it can be determined based on SINR (signal-interference-signal-noise ratio) reports from nearby devices. Alternatively, the learning trigger state can be determined locally, for example, periodically, which will be discussed further below. When ML-mDPD training is triggered, for example, when a flag is raised, the signal... n It can be stored in buffer 1402. The ML-hi-fi model 905 can also be activated. The output of the ML-hi-fi model 905, the simulation feedback signal gn, can also be buffered. In response to the deterministic signal s n and gn It has been adequately buffered, and ML-mDPD can switch from deployment to learning.

[0101] At 1404, ML-mDPD can be based on the simulated feedback signal g. n and signal s n The signal is retrained and is governed by transmission. Retraining ML-mDPD can be a response to the detection that a retraining period has expired. For example, the device can be configured to receive signal information about retraining periods, such as daily, weekly, or similar. Once a retraining period expires, the device can initiate retraining. Once learning is complete, the buffer can be refreshed at 1406. The retraining period can be pre-configured or determined based on the device type. Device types can include factory robots, handheld devices, sensors, or similar devices. Different retraining periods can be pre-configured or signaled to the device by network nodes. For example, retraining could be configured to occur every few days for the former and every few months for the latter (as seen with some devices).

[0102] Alternatively or additionally, ML-mDPD retraining can be initiated in response to the detection of changes in power amplifier performance. For example, if transmission performance metrics, such as adjacent channel leakage, no longer meet configuration or signal requirements, retraining can be launched. This allows the device to adapt to, for example, varying antenna loads and hardware component degradation due to aging.

[0103] Alternatively, or in addition, retraining the ML-mDPD can respond to the detection of updates at at least one signal transmission parameter. For example, changes in transmission parameters, such as bandwidth or modulation or coding scheme, can affect the operation of the power amplifier 902. Therefore, retraining the ML-mDPD to respond to such detected changes enables the maintenance of good DPD performance even when transmission parameters change dynamically.

[0104] Alternatively or additionally, retraining ML-mDPD can be performed in response to instructions received from network nodes, such as 5G base stations or gNBs. This enables network nodes to operate within their coverage area and control device performance, such as UEs.

[0105] Figure 15 The illustration depicts an application according to an example embodiment and another example of retraining a machine learning-based digital predistortion system. The device may include something similar to... Figure 14The device may also include an aging tracking function 1510, configured to control the learning trigger. The aging tracking function 1510 can run continuously and periodically in the background. At 1512, the aging tracking function can determine whether a period corresponding to an expiration time (ET) has expired. If not, the learning trigger can be set or maintained in the off state at 1514. If it has expired, the learning trigger can be set to the on state at 1516. The device may also include a reset function 1518 associated with ML-mDPD model training. In response to the completion of ML-mDPD retraining, similar to 1406, the device can refresh the buffer 1402. Furthermore, the timer can be reset and the learning trigger can be set back to the off state. Figure 14 and Figure 15 The example implementations enable ML-based digital predistortion systems to be retrained to mitigate the effects of aging and different types of environmental factors.

[0106] The exemplary embodiments of this disclosure can improve the accuracy of digital predistortion applied to linearized power amplifiers. This exemplary embodiment is helpful for millimeter-wave systems, but it is understood that it can also be applied to other frequency bands.

[0107] Figure 16 An example of a method 1600 for applying a machine learning-based digital predistortion system, according to an example embodiment, is illustrated.

[0108] In 1601, the method may include transmitting a signal, wherein the signal is amplified by a power amplifier.

[0109] In 1602, the method may include receiving the transmitted signal by an internal feedback receiver to obtain an internal feedback signal.

[0110] In 1603, the method may include simulating an external feedback timer and generating a simulated feedback signal based on an internal feedback signal using a first machine learning model.

[0111] In 1604, the method may include using a second machine learning model to determine digital predistortion parameters for the power amplifier based on the feedback signal from the simulation.

[0112] According to the example embodiment, the bandwidth of the internal feedback receiver may be lower than that of the external feedback receiver, and / or the dynamic range of the internal feedback receiver may be lower than that of the external feedback receiver.

[0113] According to an example embodiment, the method further includes linearizing the power amplifier based on digital predistortion parameters, wherein the digital predistortion parameters include a subset of a second machine learning model.

[0114] According to an example embodiment, the method may further include retraining the second machine learning model based on simulation-based feedback signals and signals responding to one of the following: detecting an expiration during retraining; detecting a change in the performance of the power amplifier; detecting an update of at least one transmission parameter of the signal; or receiving instructions from a network node to perform retraining.

[0115] According to an example embodiment, the method may also include determining the retraining period based on the type of device.

[0116] According to an example embodiment, the method may further include receiving an external feedback signal from an external feedback receiver; training a first machine learning model based on the internal feedback signal and the external feedback signal; and determining a second machine learning model based on the simulated feedback signal and the signal.

[0117] According to an example embodiment, the method further includes initializing a first machine learning model using a first baseline machine learning model trained using at least one reference device, and initializing a second machine learning model using a subset of a second baseline machine learning model trained using at least one reference device.

[0118] According to an example embodiment, the subset of the second baseline machine learning model may include a subset of the layers of the second baseline machine learning model.

[0119] According to an example embodiment, the first machine learning model includes a network generator for a generative adversarial network (GAN).

[0120] Figure 17 The training method for a machine learning-based digital predistortion system is illustrated in the example embodiment.

[0121] In 1701, the method may include transmitting a set of test signals via a reference device, the reference device including an internal feedback receiver configured to receive the set of test signals to obtain an internal feedback signal set.

[0122] In 1702, the method may include training a first baseline machine learning model to simulate an external feedback receiver and generating a simulated feedback signal based on an internal feedback signal, wherein training of the first baseline model is based on the internal feedback signal and an external feedback signal received from the external feedback receiver, the external feedback signal corresponding to a set of test signals.

[0123] In 1703, the method includes training a second baseline machine learning model to determine digital predistortion parameters for a power amplifier used in a reference device based on simulated feedback signals, wherein the training of the second baseline machine learning model is based on simulated feedback signals and a set of test signals.

[0124] According to an example embodiment, the method may further include training a plurality of devices using a subset of test signals, wherein training each of the plurality of devices includes: initializing an instance of a first machine learning model based on a first baseline machine learning model; initializing an instance of a second machine learning model based on a subset of a second baseline machine learning model; training the instance of the first machine learning model to simulate an external feedback receiver and generating a simulated feedback signal based on device-specific internal feedback signals received from an internal feedback receiver of the device, wherein training the instance of the first machine learning model is based on the device-specific internal feedback signals and the external feedback signals received from the external feedback receiver; and training an instance of the second machine learning model to determine digital predistortion parameters for a power amplifier of the device based on the simulated feedback signals, wherein training the instance of the second baseline machine learning model is based on the simulated feedback signals and a subset of test signals.

[0125] According to an example embodiment, the method further includes retraining a first baseline machine learning model and a second baseline machine learning model using a test signal set after a predetermined number of devices have been trained.

[0126] According to an example embodiment, the method may further include retraining a first baseline machine learning model and a second baseline machine learning model using a set of test signals in response to detecting a shift in the linearization performance of the power amplifier among multiple devices.

[0127] According to an example embodiment, the method may further include determining a subset of the second baseline model based on a similarity metric between the device and a reference device.

[0128] According to an example embodiment, the method may further include determining a similarity metric based on at least one of the following: the batch number of the power amplifier of the reference device and the power amplifier of the device; and the transmission performance index of the reference device and the device.

[0129] According to the example embodiment, the subset of the second baseline machine learning model includes a subset of the second baseline machine learning model layers.

[0130] According to an example embodiment, the test signal set can be characterized by at least one of the following: transmission power, signal bandwidth, beam configuration, modulation and / or error correction coding scheme, carrier frequency, temperature, battery level, or antenna load.

[0131] According to the example embodiment, the subset of the test signal may correspond to a subset of the parameters, and / or the subset of the test signal may correspond to a subset of the parameter values.

[0132] Further features of this method derive directly from the functionality and parameters of the apparatus and / or DPD training system or architecture as described in the appended claims and throughout the specification, and therefore will not be repeated here. It should be noted that one or more steps of the method may be performed in a different order.

[0133] An apparatus may be configured to perform or cause any aspect of the methods described herein. Further, a computer program may include instructions that, when executed, cause the apparatus to perform any aspect of the methods described herein. Further, a computer program is configured to, when executed, cause the apparatus to perform at least any aspect of the methods described herein. Further, a computer program product or computer-readable medium may include program instructions to cause the apparatus to perform any aspect of the methods described herein. Further, the apparatus may include components to perform any aspect of the methods described herein. According to an example embodiment, the component includes at least one processor and at least one memory including computer program code; the at least one memory and the computer program code are configured, along with the at least one processor, to cause the apparatus to perform at least any aspect of the method.

[0134] Any ranges or device values ​​provided herein may be expanded or changed without losing the desired effect. Furthermore, unless expressly prohibited, any embodiment may be combined with any other embodiment.

[0135] Although the subject matter has been described in language specific to the resulting features and / or behaviors, it is understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or behaviors described above. Rather, the specific features and behaviors described above are disclosed as examples of implementing the claims and other equivalent features, and other equivalent features and behaviors are intended to be included within the scope of the claims.

[0136] It will be understood that the above benefits and advantages may apply to one or more embodiments. The embodiment is not limited to solving any or all of the described problems, or is limited to any or all of the described benefits and advantages. It will be further understood that the term "one" may refer to one or more of such items. Furthermore, the terms "at least one" or "one or more" may refer to one or more of such items.

[0137] The steps or operations of the methods described herein can be performed in any suitable order, or simultaneously where appropriate. Furthermore, individual blocks can be removed from any method without departing from the scope of the subject matter described herein. Aspects of any of the example embodiments described above can be combined with any aspect of the other example embodiments to form further example embodiments, and are not necessarily for the desired effect.

[0138] The term "comprising" as used herein includes a method, a block, or an identified element, but the block or element does not include an exclusive list, and a method or apparatus may include additional blocks or elements.

[0139] As used herein, the term "circuit" may refer to one or more of the following: (b) a purely hardware circuit implementation (such as an implementation in an emulated and / or digital circuit only) and (c) a combination of hardware circuitry and software, such as (if applicable): (i) a combination of emulated and / or digital hardware circuitry with software / firmware, and (ii) any portion of a device (such as a mobile phone or server) with software (including a digital signal processor), software, and memory that work together to enable the device to perform various functions; and (d) hardware circuitry and / or a processor, such as a microprocessor or a portion thereof, that requires software (such as firmware) to operate, but which may not be present when operation is not required. This definition of "circuit" applies to all uses of the term in this application, including all uses in any of the claims.

[0140] As a further example, as used herein, the term "circuit" also encompasses only hardware circuitry or a processor (or multiple processors) or portions thereof, and its (or their) accompanying software and / or firmware. For example, if applicable to a particular claim element, the term "circuit" also includes a baseband integrated circuit, or a processor integrated circuit, or a similar integrated circuit in a server, cellular network device, or other computing or networking device.

[0141] It will be understood that the above description is provided by way of example only, and various modifications can be made by those skilled in the art. The descriptions, examples, and data above provide a complete description of the results and uses of exemplary embodiments. Although the various embodiments described above have specific particularities, or reference one or more individual embodiments, those skilled in the art can make numerous modifications to the disclosed embodiments without departing from the scope of this specification.

Claims

1. A device for communication, comprising: A power amplifier, configured to amplify signals; A transmitter configured to transmit the amplified signal, wherein the transmitter includes a power amplifier; An internal feedback receiver is configured to receive the transmitted signal without using a radio interface to obtain an internal feedback signal; A first machine learning model is configured to simulate an external feedback receiver and generate a simulated feedback signal based on the internal feedback signal, wherein the simulated external feedback receiver is configured to receive the transmitted signal through the radio interface; A second machine learning model is configured to determine digital predistortion parameters for the power amplifier based on the feedback signal from the simulation.

2. The apparatus of claim 1, wherein the bandwidth of the internal feedback receiver is lower than the bandwidth of the external feedback receiver, and / or wherein the dynamic range of the internal feedback receiver is lower than the dynamic range of the external feedback receiver.

3. The apparatus according to claim 1, further comprising: A component for linearizing the power amplifier based on the digital predistortion parameters, wherein the digital predistortion parameters include a subset of the second machine learning model.

4. The apparatus according to claim 1, further comprising: The second machine learning model is retrained based on the feedback signal from the simulation and the signal in response to one of the following: The expiration of the retraining period has been detected; A change in the performance of the power amplifier was detected; An update to at least one transmission parameter of the transmitted signal is detected; or Receive instructions from network nodes to perform retraining.

5. The apparatus according to any one of the preceding claims further comprises: Components for receiving external feedback signals from the external feedback receiver; A component for training the first machine learning model based on the internal feedback signal and the external feedback signal; A component for training the second machine learning model based on the signal and the feedback signal from the simulation.

6. The apparatus according to claim 5, further comprising: Components for initializing a first machine learning model using a first baseline machine learning model, the first baseline machine learning model being trained using at least one reference device, and Components for initializing the second machine learning model using a subset of the second baseline machine learning model, the second baseline machine learning model being trained using the at least one reference device.

7. The apparatus of claim 5, wherein the first machine learning model comprises a generator network of a generative adversarial network (GAN).

8. A method of communication, comprising: Send a signal amplified by a power amplifier; The internal feedback signal is obtained by receiving the transmitted signal without using the radio interface from the internal feedback receiver; An external feedback receiver is simulated, and a simulated feedback signal is generated based on the internal feedback signal using a first machine learning model, wherein the simulated external feedback receiver is configured to receive the transmitted signal through the radio interface; Using a second machine learning model, the digital predistortion parameters for the power amplifier are determined based on the feedback signal from the simulation.

9. A system for communication, comprising: A reference device is configured to amplify a set of test signals using a power amplifier and to transmit the amplified set of test signals. The reference device includes an internal feedback receiver configured to receive the transmitted amplified set of test signals without a radio interface to obtain an internal feedback signal. An external feedback receiver is configured to receive the amplified set of test signals transmitted via the radio interface to obtain an external feedback signal; Components for training a first baseline machine learning model to simulate the external feedback receiver and for generating a simulated feedback signal based on the internal feedback signal, wherein the training of the first baseline machine learning model is based on the internal feedback signal and the external feedback signal received from the external feedback receiver. A component for training a second baseline machine learning model to determine digital predistortion parameters of the power amplifier for the reference device based on the feedback signal from the simulation, wherein the training of the second baseline machine learning model is based on the feedback signal from the simulation and the test signal set.

10. A method of communication, comprising: Amplify the test signal set using the power amplifier of the reference device; The reference device transmits an amplified set of test signals, the reference device including an internal feedback receiver configured to receive the transmitted amplified set of test signals without a radio interface to obtain an internal feedback signal set. A first baseline machine learning model is trained to simulate an external feedback receiver and to generate a simulated feedback signal based on the internal feedback signal, wherein the simulated external feedback receiver is configured to receive the transmitted signal via the radio interface, and wherein the training of the first baseline machine learning model is based on the internal feedback signal and an external feedback signal received from the external feedback receiver, the external feedback signal corresponding to the test signal set. A second baseline machine learning model is trained to determine the digital predistortion parameters of the power amplifier for the reference device based on the feedback signal from the simulation, wherein the training of the second baseline machine learning model is based on the feedback signal from the simulation and the test signal set.

11. The method of claim 10, further comprising: Multiple devices are trained using a subset of test signals, wherein training for each of the multiple devices includes: The instance of the first machine learning model is initialized based on the first baseline machine learning model; The second machine learning model instance is initialized based on a subset of the second baseline machine learning model; The instance of the first machine learning model is trained to simulate the external feedback receiver and to generate a simulated feedback signal based on a device-specific internal feedback signal received from the device's internal feedback receiver, wherein the training of the instance of the first machine learning model is based on the device-specific internal feedback signal and the external feedback signal received from the external feedback receiver. The instance of the second machine learning model is trained to determine the digital predistortion parameters of the power amplifier for the device based on the feedback signal of the simulation, wherein the training of the instance of the second baseline machine learning model is based on the feedback signal of the simulation and a subset of the test signal.

12. The method of claim 11, further comprising: After training a predetermined number of devices, the first baseline machine learning model and the second baseline machine learning model are retrained using the test signal set.

13. The method of claim 11, further comprising: In response to the detection of a deviation in the linearization performance of the power amplifier among the plurality of devices, the first baseline machine learning model and the second baseline machine learning model are retrained using the test signal set.

14. The method of claim 11, wherein the amplified test signal set is characterized by a set of parameters including at least one of the following: transmit power, signal bandwidth, beam configuration, modulation and / or error correction coding scheme, carrier frequency, temperature, battery level, or antenna load.

15. The method according to any one of claims 11 to 14, wherein the subset of the test signals corresponds to a subset of the parameters, and / or wherein the subset of the test signals corresponds to a subset of the parameter values.

Images