Digital pre-delay with hybrid basis function-based actuators and neural networks

Abstract

Systems, devices, and methods relating to hybrid basis function, neural network-based digital predistortion (DPD) are provided. An exemplary apparatus for a radio frequency (RF) transceiver includes a digital predistortion (DPD) actuator for receiving an input signal associated with a nonlinear component of the RF transceiver and outputting a predistorted signal. The DPD actuator includes a basis function-based actuator for performing a first DPD operation using a set of basis functions associated with a first nonlinear characteristic of the nonlinear component. The DPD actuator further includes a neural network-based actuator for performing a second DPD operation using a first neural network associated with a second nonlinear characteristic of the nonlinear component. The predistorted signal is based on a first output signal of the basis function-based actuator and a second output signal of the neural network-based actuator.

Description

Technical Field

[0001] Cross - reference to Related Applications This application claims the priority and benefit of U.S. Provisional Patent Application No. 63 / 187,536, filed on May 12, 2021, entitled "DIGITAL PREDISTORTION FOR POWER AMPLIFIER LINEARIZATION USING NEURAL NETWORKS", and U.S. Non - Provisional Patent Application No. 17 / 732,764, filed on April 29, 2022, entitled "DIGITAL PREDISTORTION WITH HYBRID BASIS - FUNCTION - BASED ACTUATOR AND NEURAL NETWORK", and as fully described below and for all applicable purposes, they are hereby incorporated by reference in their entirety into this specification.

[0002] This disclosure generally relates to electronic devices, and more specifically, to digital predistortion (DPD) with a hybrid basis - function - based actuator and a neural network.

Background Art

[0003] A radio frequency (RF) system is a system that transmits and receives signals in the form of electromagnetic waves in the RF range of approximately 3 kilohertz (kHz) to 300 gigahertz (GHz). RF systems are commonly used for wireless communication, with cellular / wireless mobile technology being a prominent example, but may also be used for cable communication such as cable television. In both of these types of systems, the linearity of various components within them plays an important role.

[0004] The linearity of an RF component or system such as an RF transceiver is theoretically easy to understand. That is, linearity generally refers to the ability of a component or system to provide an output signal that is directly proportional to the input signal. In other words, if a component or system is perfectly linear, the relationship between the ratio of the output signal to the input signal is a straight line. Achieving this behavior in actual components and systems is far more complex, and many of the challenges to linearity must often be overcome at the expense of several other performance parameters, such as efficiency and / or output power.

[0005] Power amplifiers (PAs), made from semiconductor materials that are inherently nonlinear and must operate at relatively high power levels, are typically the first components analyzed when considering the design of an RF system in terms of linearity. PA outputs with nonlinear distortion can result in reduced modulation accuracy (e.g., reduced error vector magnitude, EVM) and / or out-of-band emissions. Therefore, both wireless RF systems (e.g., Long-Term Evolution (LTE) and millimeter-wave or 5th generation (5G) systems) and cable RF systems have stringent specifications regarding PA linearity.

[0006] The linearity of a PA can be improved by applying a DPD (Digital Pre-Distortion). Typically, DPD involves applying pre-distortion to the signal provided as input to the PA in the digital domain to reduce and / or cancel the distortion expected to be caused by the PA. Pre-distortion can be characterized by a PA model, which can be updated based on feedback from the PA (i.e., based on the PA's output). The more accurate the PA model is in predicting the distortion introduced by the PA, the more effective the pre-distortion of the input to the PA is in reducing the impact of distortion caused by the amplifier. [Overview of the project] [Means for solving the problem]

[0007] Implementing DPD in RF systems is not an easy task, as various factors can affect its cost, quality, and robustness. Physical constraints such as space / surface area and regulations can further restrict the requirements or specifications of the DPD. DPD has become particularly challenging as sampling rates used in state-of-the-art RF systems are constantly increasing, thus requiring trade-offs and ingenuity in DPD design.

[0008] To provide a more complete understanding of this disclosure and its features and benefits, please refer to the following description in conjunction with the attached figures, where similar reference numbers represent similar parts. [Brief explanation of the drawing]

[0009] [Figure 1] This disclosure provides schematic block diagrams of exemplary radio frequency (RF) transceivers in which hybrid basis functions and neural network-based digital pre-distortion (DPD) may be implemented according to several embodiments of this disclosure. [Figure 2] This disclosure provides schematic diagrams illustrating exemplary arrangements of hybrid basis functions and neural network-based DPD blocks according to several embodiments of this disclosure. [Figure 3] This disclosure provides schematic diagrams illustrating exemplary arrangements of hybrid basis functions and neural network-based DPD blocks according to several embodiments of this disclosure. [Figure 4] This disclosure provides schematic diagrams illustrating exemplary arrangements of hybrid basis functions for neural network-based DPDs according to several embodiments of this disclosure. [Figure 5] This disclosure provides schematic diagrams illustrating exemplary arrangements of hybrid basis functions for neural network-based DPDs according to several embodiments of this disclosure. [Figure 6]This disclosure provides illustrative diagrams of schemes for offline training and online operation of hybrid basis functions, neural network-based DPDs, according to several embodiments of this disclosure. [Figure 7] This disclosure provides schematic diagrams illustrating exemplary arrangements of hybrid basis functions for neural network-based DPDs according to several embodiments of this disclosure. [Figure 8] This disclosure provides schematic diagrams illustrating exemplary arrangements of hybrid basis functions and combiners in neural network-based DPDs according to several embodiments of this disclosure. [Figure 9] This disclosure provides schematic diagrams illustrating exemplary arrangements of hybrid basis functions and combiners in neural network-based DPDs according to several embodiments of this disclosure. [Figure 10] This disclosure provides schematic diagrams illustrating exemplary arrangements of hybrid basis functions and combiners in neural network-based DPDs according to several embodiments of this disclosure. [Figure 11] This disclosure provides schematic diagrams illustrating exemplary PA state estimation and prediction implementations in hybrid basis functions and neural network-based DPDs according to several embodiments of this disclosure. [Figure 12] This disclosure provides schematic diagrams illustrating exemplary hardware architectures for hybrid basis functions and neural network-based actuators in neural network-based DPDs, according to several embodiments of this disclosure. [Figure 13] This disclosure provides flowcharts illustrating methods for performing hybrid basis functions and neural network-based DPD according to several embodiments of this disclosure. [Modes for carrying out the invention]

[0010] <Overview> Each of the systems, methods, and devices described herein has several innovative embodiments, and no single one is solely responsible for all of the desired attributes disclosed herein. Details of one or more implementations of the subject matter described herein are given in the following description and accompanying drawings.

[0011] To illustrate the neural network-based DPD proposed herein, it is helpful to first understand the phenomena that may occur in RF systems. The following basic information can be seen as the basis for the proper explanation of this disclosure. Such information is provided for illustrative purposes only and should not be construed in any way as to limit the broad scope of this disclosure or its potential applications.

[0012] As described above, PA is typically the first component to analyze when considering the design of an RF system from the perspective of linearity. Having a linear and efficient PA is essential for wireless and cable RF systems. While linearity is also important for small-signal amplifiers, such as low-noise amplifiers, the challenges of linearity are particularly pronounced for PAs, as such amplifiers typically need to generate relatively high levels of output power and are therefore particularly susceptible to certain operating conditions where nonlinear behavior can no longer be ignored. On the one hand, the nonlinear behavior of the semiconductor materials used to form the amplifier tends to worsen when the amplifier operates at high-power signal levels (operating conditions commonly referred to as "operating at saturation"), increasing the amount of nonlinear distortion in the output signal, which is highly undesirable. On the other hand, amplifiers operating at relatively high power levels (i.e., operating at saturation) also typically operate at their highest efficiency, which is highly desirable. As a result, linearity and efficiency (or power level) are two performance parameters for which an acceptable trade-off must be found, in that improvements in one of these parameters often come at the expense of other parameters being suboptimal. For this purpose, the term "backoff" is used in the art to describe the degree to which the input power (i.e., the power of the signal supplied to the amplifier being amplified) should be reduced in order to achieve the desired output linearity (for example, backoff can be measured as the ratio of the input power that yields maximum power to the input power that yields the desired linearity). Thus, reducing the input power can improve linearity but will result in a decrease in the efficiency of the amplifier.

[0013] Furthermore, as described above, a DPD can pre-distort the input to the PA in order to reduce and / or cancel the distortion caused by the amplifier. At a high level, to achieve this function, a DPD involves forming a model of how the PA may affect the input signal, and this model defines the coefficients of a filter applied to the input signal (such coefficients called "DPD coefficients") in an attempt to reduce and / or cancel the distortion of the input signal caused by the amplifier. In this way, a DPD attempts to compensate an amplifier for applying an undesirable nonlinear modification to the transmitted signal by applying a modification corresponding to the input signal provided to the amplifier.

[0014] The models used in DPD algorithms are typically adaptive models, meaning they are formed through an iterative process by gradually adjusting coefficients based on a comparison between data entering an amplifier's input and data exiting the amplifier's output. Estimating DPD coefficients is based on acquiring a finite sequence of input and output data (i.e., inputs to and outputs from the PA), commonly referred to as "capture," and forming a feedback loop in which the model is adapted based on the analysis of the captures. More specifically, conventional DPD algorithms are based on generalized memory polynomial (GMP) models, which involve forming a set of polynomial equations commonly called "update equations" and updating the model of the PA by searching for appropriate solutions to the equations in a wide solution space. To this end, the DPD algorithm solves an inverse problem, which is the process of calculating the random factors that generated a series of observations from a set of observations.

[0015] Solving inverse problems in the presence of nonlinear effects can be difficult and potentially inappropriate. In particular, the inventors of this disclosure have recognized that GMP-based PA models may have limitations due to the signal dynamics required to store polynomial data and the limited memory depth, especially when sampling rates used in state-of-the-art RF systems are constantly increasing.

[0016] Solid-state devices that can be used at high frequencies are very important in modern semiconductor technology. In part, due to their large bandgaps and high mobilities, III-N-based transistors such as GaN-based transistors (i.e., transistors using a compound semiconductor material having a first sublattice of at least one element from Group III of the periodic table (e.g., Al, Ga, In) and a second sublattice of nitrogen (N) as the channel material) can be particularly advantageous for high-frequency applications. In particular, a PA can be constructed using GaN transistors.

[0017] GaN transistors have desirable characteristics with respect to cutoff frequency and efficiency, but their behavior is complicated by an effect known as charge trapping, where defect sites within the transistor channel trap charge carriers. The density of trapped charge depends strongly on the gate voltage, which typically scales with the signal amplitude. To further complicate matters, an opposing effect can compete with the effect of charge trapping simultaneously. That is, when some charge carriers are trapped by defect sites, for example, due to thermal activation, other charge carriers are released from the traps. These two effects have very different time constants. Each time the gate voltage rises, the defect sites may be quickly filled with trapped charge, while the release of trapped charge occurs more slowly. The release time constant can be from tens of microseconds to a maximum of milliseconds, and its effect is typically very evident on the time scale of the symbol period of 4G or 5G data, particularly data containing bursts.

[0018] Various embodiments of the present disclosure aim to provide a system and method that improve one or more of the above disadvantages to provide a linear and efficient amplifier (such as, but not limited to, a PA) for an RF system (such as, but not limited to, a wireless RF system of millimeter wave / 5G technology). In particular, aspects of the present disclosure provide a DPD arrangement that utilizes a combination of a basis function-based actuator and a neural network (NN)-based actuator.

[0019] As used herein, a basis function-based actuator may refer to a DPD actuator that performs DPD operation as part of pre-distorting an input signal to a non-linear component (e.g., a PA) using a set of basis functions. A basis function may refer to an element of a particular basis of a function space. Just as all vectors in a vector space can be represented as a linear combination of basis vectors, all functions in a function space can be represented as a linear combination of basis functions. Some examples of basis functions may be based on Volterra series, GMP models (a subset of Volterra series), and / or linear piecewise functions.

[0020] As used herein, a neural network-based actuator may refer to a DPD actuator that uses one or more neural networks to perform DPD operation as part of pre-distorting input signals to a nonlinear component (e.g., PA). A neural network is a deep learning model, a subset of machine learning. As an example, a neural network may include multiple layers, e.g., an input layer, followed by one or more hidden layers (e.g., a fully connected layer or a convolutional layer), and an output layer. Each layer may include a set of weights and / or biases that can transform the input received from the previous layer, and the resulting output can be passed to the next layer. The weights and / or biases of each layer can be trained and adapted to process, for example, input signals and / or feedback signals representing observed signals or the output of a nonlinear component, to post-process the output signals of a basis function-based actuator, to update the parameters of a basis function-based actuator when pre-distorting input signals, and / or to combine signals generated by a basis function-based actuator and a neural network-based actuator. In general, one or more neural networks used by a neural network-based actuator can have any suitable architecture (e.g., convolutional neural network, recurrent neural network, etc.).

[0021] According to one aspect of the present disclosure, an apparatus for an RF transceiver may include a DPD actuator for receiving an input signal associated with a nonlinear component (e.g., PA) of the RF transceiver and outputting a pre-distorted signal. The DPD actuator may include a basis function-based actuator and a neural network-based actuator. The basis function-based actuator may perform a first DPD operation using a set of basis functions associated with a first nonlinear characteristic of the nonlinear component. In some embodiments, the set of basis functions may be based on a Volterra series, a GMP model, and / or a dynamic deviation reduction (DDR) model. The neural network-based actuator may perform a second DPD operation using one or more neural networks associated with a second nonlinear characteristic of the nonlinear component. The first and second nonlinear characteristics may be the same or different. Generally, the first and second nonlinear characteristics may include any suitable order (e.g., third, fifth, tenth, eleventh, etc.) and / or any number of nonlinear characteristics of the nonlinear characteristics of the nonlinear component. A DPD actuator can output a pre-distorted signal based on a first output signal from a basis function-based actuator and a second output signal from a neural network-based actuator. The basis function-based actuator and the neural network-based actuator can be arranged in various configurations, such as parallel or cascaded configurations, and can interact with each other in various ways to generate the pre-distorted signal.

[0022] In a parallel configuration, each of the basis function-based actuator and the neural network-based actuator may process at least one of an input signal or a feedback signal indicating the output of a nonlinear component (where the output is looped back from the transmitter path to the receiver path for observation). In this regard, the basis function-based actuator may perform a first DPD operation by using a set of basis functions to process at least one of the input signal or feedback signal to generate a first output signal. The neural network-based actuator may perform a second DPD operation by using one or more neural networks to process at least one of the input signal or feedback signal to generate a second output signal. Furthermore, in one embodiment, the DPD actuator may further include a combiner for combining the input signal, the first output signal, and / or the second output signal to provide a pre-distorted signal. The combiner may perform various data transformations (e.g., signal alignment, upsampling, filtering, processing using another neural network, etc.) on the input signal, the first output, and / or the second output signal before summing these signals together. In another embodiment, as part of performing a second DPD operation, instead of combining the first and second output signals, the second output signal of a neural network-based actuator can be used to update the parameters of a basis function-based actuator, which can then use the updated parameters to generate the first output signal.

[0023] In a cascade configuration, basis function-based actuators and neural network-based actuators may be arranged sequentially in series. In one embodiment, a neural network-based actuator may be coupled to the output of a basis function-based actuator. In this case, the basis function-based actuator may perform a first DPD operation by using a set of basis functions to process at least one of an input signal or a feedback signal indicating the output of a nonlinear component to generate a first output signal. Subsequently, the neural network-based actuator may perform a second DPD operation by using one or more neural networks to process the first output signal of the basis function-based actuator to generate a second output signal. In some cases, the neural network-based actuator may further generate a second output signal by using one or more neural networks to further process the input signal. In another embodiment, a basis function-based actuator may be coupled to the output of a neural network-based actuator. In this case, the neural network-based actuator may perform a second DPD operation by using one or more neural networks to process at least one of an input signal or a feedback signal to generate a second output signal. Subsequently, the basis function-based actuator can perform a first DPD operation by using a set of basis functions to process the second output signal of the neural network-based actuator and generate a first output signal.

[0024] In some embodiments, for parallel and / or cascaded configurations, basis function-based actuators and / or neural network-based actuators may preprocess (e.g., transform) their respective inputs as part of a first DPD operation and / or a second DPD operation, respectively. Additionally or alternatively, for parallel configurations, the neural network-based actuator may further downsample at least one of the input signal or feedback signal as part of a second DPD operation.

[0025] In some embodiments, one or more neural networks used by a neural network-based actuator for a second DPD operation may include an estimated neural network model and a predictive neural network model. The neural network-based actuator may select between the estimated neural network model and the predictive neural network model based on the availability of a feedback signal. In this regard, if a feedback signal is available or effective, the neural network-based actuator may use the estimated neural network model to process the input signal and the feedback signal to generate a second output signal. On the other hand, if a feedback signal is not available, the neural network-based actuator may use the predictive neural network model to process the input signal to generate a second output signal. Furthermore, in some embodiments, one or more neural networks may have recursive internal states. In this regard, the neural network-based actuator may use the selected one of the estimated neural network model and the predictive neural network model to process at least one of the input signal or the feedback signal and further process previous state information associated with at least one of the estimated neural network model and the predictive neural network model to generate a second output signal. A neural network-based actuator may further update state information associated with at least one of the estimated neural network model or the predictive neural network model based on a second output signal.

[0026] Furthermore, in some embodiments, the neural network-based actuator may include a neural network processor (e.g., a hardware accelerator) for performing neural network-specific operations (e.g., convolution, rectified linear unit (ReLU) operations, etc.). The neural network-based actuator may further include memory for storing parameters (e.g., trained weights) associated with one or more neural networks (e.g., an estimated neural network model and a predictive neural network model). The neural network-based actuator may utilize the neural network processor to perform a second DPD operation based on the stored parameters.

[0027] The systems, schemes, and mechanisms described herein advantageously utilize neural networks to assist DPD operation. For example, there may be limitations to using basis functions in DPD. Neural networks can generally offer higher degrees of freedom than basis functions and can model any complex function that may not be readily expressible by analytical or mathematical functions. For example, neural network-based actuators can model and modify nonlinear characteristics and / or long-term memory effects that basis function-based actuators are not designed to model and / or modify (e.g., due to hardware resource limitations and / or desired utilization) and / or cannot model due to changes (e.g., analog gain settings, temperature, signal power, etc.). Furthermore, in contrast to conventional DPDs where observed data (e.g., feedback signals looped back from transmitter to receiver) is used only for adaptation and may have long delays on the order of seconds from observation time to operation time, the neural network-based actuators disclosed herein can use observed data for operation or compensation, which may have shorter delays on the order of microseconds from observation time to operation time. Therefore, this disclosure can improve DPD performance when linearizing nonlinear components. Although aspects of this disclosure are discussed in the context of linearizing PAs in RF systems, the disclosed hybrid basis functions, neural network-based DPD can be applied to linearize any suitable nonlinear component. Furthermore, the disclosed hybrid basis functions, neural network-based DPD technique is suitable for use in radio base stations and / or radio mobile handsets (e.g., user equipment).

[0028] Exemplary RF transceiver with hybrid basis functions and neural network-based DPD configuration Figure 1 provides a schematic block diagram of an exemplary RF transceiver 100 in which hybrid basis functions, neural network-based DPDs may be implemented according to several embodiments of the present disclosure. As shown in Figure 1, the RF transceiver 100 may include a DPD circuit 110, a transmitter circuit 120, a PA 130, an antenna 140, and a receiver circuit 150.

[0029] The DPD circuit 110 is configured to receive an input signal 102, represented by x, which may be a sequence of digital samples or a vector. Generally, as used herein, each of the lowercase, bold, italicized single-letter labels used in the figures (e.g., labels x, z, y, and y' shown in Figure 1) refers to a vector. In some embodiments, the input signal 102x may include one or more active channels in the frequency domain, but for simplicity, an input signal having only one channel (i.e., a single frequency range of in-band frequencies) is described. In some embodiments, the input signal x may be a baseband digital signal. The DPD circuit 110 is configured to generate an output signal 104, which may be represented by z, based on the input signal 102x. The DPD output signal 104z may be further provided to the transmitter circuit 120.

[0030] According to aspects of this disclosure, the DPD circuit 110 may include a basis function-based actuator 112 and a neural network-based actuator 114. The basis function-based actuator 112 may perform a first DPD operation using a set of basis functions associated with a first nonlinear characteristic of a nonlinear component. In some embodiments, the set of basis functions may be a Volterra series or a subset of a Volterra series (e.g., GMP and / or DDR). The neural network-based actuator 114 may perform a second DPD operation using one or more neural networks associated with a second nonlinear characteristic of a nonlinear component. The DPD circuit 110 may output a pre-distorted signal 104z based on a first output signal of the basis function-based actuator 112 and a second output signal of the neural network-based actuator 114. Generally, the basis function-based actuator 112 and the neural network-based actuator 114 may be implemented using any preferred combination of hardware and / or software. In certain embodiments, the basis function-based actuator 112 may utilize a lookup table (LUT) to store an associated set of DPD coefficients that can be represented by a set of basis functions and c, and the neural network-based actuator 114 may include a hardware accelerator (e.g., a neural network processor) for performing neural network operations. The basis function-based actuator 112 and the neural network-based actuator 114 can be arranged in various configurations, e.g., parallel or cascaded configurations, and may interact with each other in various ways to generate pre-distorted signals. In some embodiments, the DPD circuit 110 may optionally include a combiner 116, as shown by a dashed box, to combine the outputs of the basis function-based actuator 112 and the neural network-based actuator 114, for example, when the basis function-based actuator 112 and the neural network-based actuator 114 are arranged in a parallel configuration.The mechanism for performing DPD using basis function-based actuator 112 and neural network-based actuator 114 is described in more detail below. In some cases, the DPD circuit 110 may be called a hybrid basis function, neural network-based DPD.

[0031] The transmitter circuit 120 may be configured to upconvert the signal 104z from a baseband signal to a higher frequency signal such as an RF signal. The RF signal generated by the transmitter 120 may be supplied to the PA 130, which may be implemented as a PA array containing N individual PAs. The PA 130 may be configured to amplify the RF signal generated by the transmitter 120 (thus the PA 130 may be driven by a drive signal based on the output of the DPD circuit 110) and output an amplified RF signal 131 which can be represented by y (e.g., a vector).

[0032] In some embodiments, the RF transceiver 100 may be a wireless RF transceiver, in which case it also includes an antenna 140. In the context of wireless RF systems, an antenna is a device that acts as an interface between radio waves propagating wirelessly through space and electric currents moving within a metal conductor used in a transmitter, receiver, or transceiver. During transmission, the transmitter circuit of the RF transceiver may supply an electrical signal, which is amplified by a PA, and the amplified version of the signal is provided to the antenna's terminals. The antenna can then radiate the energy from the signal output by the PA as radio waves. Antennas are essential components of all wireless equipment and are used in devices such as radio broadcasting, broadcast television, two-way radio, communication receivers, radar, mobile phones, and satellite communications.

[0033] An antenna with a single antenna element typically broadcasts a radiation pattern that radiates equally in all directions within the spherical wavefront. A phased antenna array generally refers to a collection of antenna elements used to concentrate electromagnetic energy in a specific direction, thereby creating a main beam, a process commonly called "beamforming." Phased antenna arrays offer many advantages over single antenna systems, including high gain, the ability to perform directional steering, and simultaneous communication. Therefore, phased antenna arrays are more frequently used in countless different applications, such as mobile / cellular radio technology, military applications, aircraft radar, automotive radar, industrial radar, and Wi-Fi technology.

[0034] In an embodiment where the RF transceiver 100 is a wireless RF transceiver, the amplified RF signal 131y may be supplied to an antenna 140, which may be implemented as an antenna array containing multiple antenna elements, for example, N antenna elements. The antenna 140 is configured to wirelessly transmit the amplified RF signal 131y.

[0035] In embodiments where the RF transceiver 100 is a wireless RF transceiver for a phased antenna array system, the RF transceiver 100 may further include a beamformer configuration configured to steer the beam generated by the antenna array 140 by changing the input signals provided to the individual PAs of the PA array 130. Such beamformer configurations can be implemented in different ways, for example, as an analog beamformer (i.e., the input signals amplified by the PA array 130 are modified in the analog domain, i.e., after these signals have been converted from the digital domain to the analog domain), as a digital beamformer (i.e., the input signals amplified by the PA array 130 are modified in the digital domain, i.e., before these signals have been converted from the digital domain to the analog domain), or as a hybrid beamformer (i.e., the input signals amplified by the PA array 130 are modified partly in the digital domain and partly in the analog domain), and are not specifically shown in Figure 1.

[0036] Ideally, the amplified RF signal 131y from PA130 should be simply an upconverted and amplified version of the output of transmitter circuit 120, e.g., an upconverted, amplified, and beamformed version of input signal 102x. However, as discussed above, the amplified RF signal 131y may have distortion outside the main signal component. Such distortion may arise from nonlinearity in the response of PA130. As discussed above, it may be desirable to reduce such nonlinearity. Therefore, RF transceiver 100 may further include a feedback path (or observation path) that allows the RF transceiver to analyze the amplified RF signal 131y from PA130 (in the transmission path). In some embodiments, the feedback path may be implemented as shown in Figure 1A, where a feedback signal 151y' may be provided from PA130 to receiver circuit 150. However, in other embodiments, the feedback signal may be a signal from a probe antenna element configured to sense a radio RF signal transmitted by antenna 140 (not specifically shown in Figure 1A).

[0037] Therefore, in various embodiments, at least a portion of the output of PA130 or the output of antenna 140 may be provided to receiver circuit 150 as a feedback signal 151. The output of receiver circuit 150 is coupled to DPD circuit 110, in particular to basis function-based actuator 112 and / or neural network-based actuator 114. In this way, the output signal 151(y') of receiver circuit 150, which is the output signal 131(y) from PA130, may be provided to basis function-based actuator 112 and / or neural network-based actuator 114 via receiver circuit 150. As discussed above, the basis function-based actuator 112 and the neural network-based actuator 114 can be arranged in parallel or cascaded configurations. In the parallel configuration, each of the basis function-based actuator 112 and the neural network-based actuator 114 may process at least one of the input signal 102x or the feedback signal 151y'. In a cascade configuration, the basis function-based actuator 112 and the neural network-based actuator 114 may be arranged sequentially in series. One of the basis function-based actuator 112 and the neural network-based actuator 114 may process at least one of the input signal 102x or the feedback signal 151y' to generate a first output signal, while the other of the basis function-based actuator 112 and the neural network-based actuator 114 may use one or more neural networks to process the first output signal of the basis function-based actuator to generate a second output signal. Details of the parallel and cascade configurations are discussed more fully below with reference to Figures 2 to 5. In some cases, the DPD circuit 110 may optionally include a DPD adaptive circuit for processing the received signal (e.g., input signal 102x and / or feedback signal 151y') to update the DPD coefficient c applied to the input signal 102x by the DPD actuator circuit 112 to generate the actuator output 104z.The signal based on the actuator output z is provided as an input to PA130, meaning that the DPD actuator output z can be used to control the operation of PA130.

[0038] As further shown in Figure 1, in some embodiments, the transmitter circuit 120 may include a digital filter 122, a digital-to-analog converter (DAC) 124, an analog filter 126, and a mixer 128. In such a transmitter, the pre-distorted signal 104z may be filtered in the digital domain by the digital filter 122 to produce a filtered pre-distorted input, a digital signal. The output of the digital filter 122 may then be converted to an analog signal by the DAC 124. The analog signal generated by the DAC 124 may then be filtered by the analog filter 126. The output of the analog filter 126 may then be upconverted to RF by the mixer 128, which can receive a signal from a local oscillator (LO) 162 and convert the filtered analog signal from the analog filter 126 from baseband to RF. Other methods of implementing the transmitter circuit 120 are also possible and within the scope of this disclosure. For example, in another implementation (not illustrated in these drawings), the output of the digital filter 122 can be directly converted into an RF signal by the DAC 124 (e.g., in a direct RF architecture). In such an implementation, the RF signal provided by the DAC 124 can then be filtered by the analog filter 126. In this implementation, since the DAC 124 directly synthesizes the RF signal, the mixer 128 and local oscillator 162 illustrated in Figure 1A can be omitted from the transmitter circuit 120 in such an embodiment.

[0039] As further shown in Figure 1, in some embodiments, the receiver circuit 150 may include a digital filter 152, an analog-to-digital converter (ADC) 154, an analog filter 156, and a mixer 158. In such a receiver, the feedback signal 151 may be down-converted to baseband by a mixer 158 that receives a signal from a local oscillator (LO) 160 (which may be the same as or different from the local oscillator 160) and converts the feedback signal 151 from RF to baseband. The output of the mixer 158 may then be filtered by the analog filter 156. The output of the analog filter 156 may then be converted to a digital signal by the ADC 154. The digital signal generated by the ADC 154 may then be filtered in the digital domain by the digital filter 152 to produce a filtered down-converted feedback signal 151y', which may be a sequence of digital values ​​representing the output y of PA 130, and which may also be modeled as a vector. The feedback signal 151y' may be provided to the DPD circuit 110. Other methods of implementing the receiver circuit 150 are also possible and within the scope of this disclosure. For example, in another implementation (not illustrated in these drawings), the RF feedback signal 151y' may be directly converted to a baseband signal by the ADC 154 (e.g., in a direct RF architecture). In such an implementation, the down-converted signal provided by the ADC 154 can then be filtered by the digital filter 152. In this implementation, since the ADC 154 directly synthesizes the baseband signal, the mixer 158 and local oscillator 160 illustrated in Figure 1A can be omitted from the receiver circuit 150 in such an embodiment.

[0040] Further modifications are possible to the RF transceiver 100 described above. For example, while up-conversion and down-conversion have been described with respect to the baseband frequency, in other embodiments of the RF transceiver 100, an intermediate frequency (IF) may be used instead. The IF may be used in a superheterodyne radio receiver in which the received RF signal is shifted to the IF before final detection of information in the received signal takes place. Conversion to IF may be useful for several reasons. For example, when several stages of filters are used, they can all be set to a fixed frequency, which makes them easier to construct and tune. In some embodiments, the mixer of the RF transmitter 120 or receiver 150 may include several such stages of IF conversion. In another embodiment, a single-path mixer is shown for each of the transmit (TX) path (i.e., the signal path for signals processed by the transmitter 120) and receive (RX) path (i.e., the signal path for signals processed by the receiver 150) of the RF transceiver 100. However, in some embodiments, the TX path mixer 128 and the RX path mixer 158 may be implemented as quadrature upconverters and downconverters, respectively, in which case each of them would include a first mixer and a second mixer. For example, with respect to the RX path mixer 158, the first RX path mixer may be configured to perform downconversion to produce an in-phase (I) downconverted RX signal by mixing the feedback signal 151 with the in-phase component of the local oscillator signal provided by the local oscillator 160. A second RX path mixer may be configured to perform downconversion to produce a quadrature (Q) downconverted RX signal by mixing the feedback signal 151 with the quadrature component of the local oscillator signal provided by the local oscillator 160 (the quadrature component is the component offset by 90 degrees from the common-mode component of the local oscillator signal). The output of the first RX path mixer may be supplied to the I signal path, and the output of the second RX path mixer may be supplied to the Q signal path, which may be substantially 90 degrees out of phase with the I signal path.Generally, the transmitter circuit 120 and the receiver circuit 150 may utilize a zero-IF architecture, a direct conversion RF architecture, a complex IF architecture, a high (real) IF architecture, or any other suitable RF transmitter and / or receiver architecture.

[0041] Generally, the RF transceiver 100 can be any device / apparatus or system configured to support the transmission and reception of signals in the form of electromagnetic waves in the RF range of approximately 3 kHz to 300 GHz. In some embodiments, the RF transceiver 100 may be used for wireless communication in base station (BS) or user equipment (UE) devices of any suitable cellular wireless communication technology such as the Global System for Mobile Communication (GSM), Code Division Multiple Access (CDMA), or LTE. In further embodiments, the RF transceiver 100 may be used as, or within, a BS or UE device of millimeter-wave wireless technology such as 5G wireless (i.e., having frequencies in the range of approximately 20 to 60 GHz, corresponding to wavelengths in the high-frequency / short-wavelength spectrum, e.g., in the range of approximately 5 to 15 millimeters). In yet another embodiment, the RF transceiver 100 may be used for wireless communication in Wi-Fi-enabled devices such as desktops, laptops, video game consoles, smartphones, tablets, smart TVs, digital audio players, cars, and printers, using Wi-Fi technology (e.g., a 2.4 GHz frequency band corresponding to a wavelength of about 12 cm, or a 5.8 GHz frequency band corresponding to a wavelength of about 5 cm, spectrum). In some implementations, the Wi-Fi-enabled device may be, for example, a node in a smart system configured to communicate data with other nodes, such as smart sensors. In yet another embodiment, the RF transceiver 100 may be used for wireless communication using Bluetooth technology (e.g., a frequency band of about 2.4 to about 2.485 GHz corresponding to a wavelength of about 12 cm). In other embodiments, the RF transceiver 100 may be used to transmit and / or receive wireless RF signals for purposes other than communication, for example, in automotive radar systems or in medical applications such as magneto-resonance imaging (MRI).In yet another embodiment, the RF transceiver 100 may be used for cable communication, for example, in a cable television network.

[0042] Exemplary parallel and cascaded configurations of basis function-based actuators and neural network-based actuators. As described above, the basis function-based actuator 112 and the neural network-based actuator 114 can be arranged in parallel or cascaded configurations and can interact with each other in various ways to generate a pre-distorted signal 104z from the input signal 102x. At a high level, the input to the basis function-based actuator 112 can be any subset or transformed version of the input signal 102x, the feedback signal 151y', or the output of the neural network-based actuator 114. Similarly, the input to the neural network-based actuator 114 can be any subset or transformed version of the input signal 102x, the feedback signal 151y', or the output of the basis function-based actuator 112. The pre-distorted signal 104z can be obtained from the output of the basis function-based actuator 112, the output of the neural network-based actuator 114, or a combined waveform of the input signal x, the output of the basis function-based actuator 112, and / or the output of the neural network-based actuator 114. Figures 2 and 3 are discussed in relation to Figure 1, where the basis function-based actuator 112 and the neural network-based actuator 114 are arranged in parallel. Figures 4 and 5 are discussed in relation to Figure 1, where the basis function-based actuator 112 and the neural network-based actuator 114 are arranged in a cascaded configuration. For simplification, Figures 2 through 5 may use the same reference numerals as Figure 1 to refer to the same elements or signals.

[0043] Figure 2 provides a schematic illustrative diagram of exemplary configuration 200 for hybrid basis function, neural network-based DPD according to some embodiments of the present disclosure. For example, the basis function-based actuator 112 and the neural network-based actuator 114 of the DPD circuit 110 in Figure 1 can be arranged in a parallel configuration as shown in configuration 200.

[0044] As shown in Figure 2, each of the basis function-based actuator 112 and the neural network-based actuator 114 may take at least one of the following as input: input signal 102x or feedback signal 151y' (representing the output signal 131y output by PA 130). For example, the basis function-based actuator 112 may perform a first DPD operation by using a set of basis functions (e.g., Volterra series, GMP, DDR function, piecewise linear function, etc.) to process at least one of the input signal 102x or feedback signal 151y' to generate an output signal 202. The neural network-based actuator 114 may perform a second DPD operation by using one or more neural networks (e.g., convolutional neural network, recurrent neural network, etc.) to process at least one of the input signal 102x or feedback signal 151y' to generate an output signal 204.

[0045] More specifically, in one embodiment, as part of a first operation, a basis function-based actuator 112 may process an input signal 102x using a set of basis functions to generate an output signal 202. In another embodiment, as part of a first operation, a basis function-based actuator 112 may process an input signal 102x and a feedback signal 151y' using a set of basis functions to generate an output signal 202. Similarly, in one embodiment, as part of a second DPD operation, a neural network-based actuator 114 may process an input signal 102x using one or more neural networks to generate an output signal 204. In another embodiment, as part of a second DPD operation, a neural network-based actuator 114 may process an input signal 102x and a feedback signal 151y' with one or more neural networks to generate an output signal 204. That is, one or more neural networks may be trained to update and adapt their parameters to generate at least a portion or one component of the pre-distorted signal 104z.

[0046] Furthermore, in some embodiments, the basis function-based actuator 112 may preprocess its input (e.g., input signal 102x and / or feedback signal 151y') before applying a set of basis functions to the input. Similarly, the neural network-based actuator 114 may preprocess its input (e.g., input signal 102x and / or feedback signal 151y') before applying one or more neural networks to the input. Some embodiments of preprocessing may include envelope magnitude calculation, downsampling, etc. In certain embodiments, the basis function-based actuator 112 may be a LUT-based actuator in which a set of basis functions and associated linear coupling coefficients (e.g., c) are stored as a LUT. The basis function-based actuator 112 may calculate the magnitude of the input signal 102x (e.g., a complex baseband in-phase / quadrature-phase (I / Q) signal) and use the calculated magnitude to generate an output signal 202 based on a table lookup from the LUT. The neural network-based actuator 114 can pass the input signal 102x and / or feedback signal 151y' through the network layer of at least one first neural network among one or more neural networks.

[0047] As further shown in Figure 2, the combiner 116 may combine the input signal 102x, the output signal 202 of the basis function-based actuator 112, and / or the output signal 204 of the neural network-based actuator 114 to generate a pre-distorted signal 104z. The pre-distorted signal 104z can then be sent to the DAC 124 for transmission. The combiner 116 may have various structures, as will be discussed more fully below with reference to Figures 8 to 10. In some cases, the input signal 102x, the pre-distorted signal 104z, the feedback signal 151y', the output signal 202 of the basis function-based actuator 112, and the output signal 204 of the neural network-based actuator 114 are digital baseband I / Q signals (including complex I / Q samples).

[0048] Arrangement 200 is advantageous in that it allows the basis function-based actuator 112 and the neural network-based actuator 114 to operate independently, for example, to be optimized and / or adapted separately. In this regard, a set of basis functions used by the basis function-based actuator 112 may be configured to linearize a first nonlinear characteristic of PA 130, and one or more neural networks used by the neural network-based actuator 114 may be configured to linearize a second nonlinear characteristic. The first and second nonlinear characteristics may be different (e.g., different orders of nonlinearity, or different combinations of orders of nonlinearity). Alternatively, the first and second nonlinear characteristics may be the same, and one or more neural networks in the neural network-based actuator 114 may be trained to adapt to dynamic conditions that can change the nonlinearity of PA 130.

[0049] Figure 3 provides a schematic illustrative diagram of exemplary configuration 300 for a hybrid basis function, neural network-based DPD according to several embodiments of the present disclosure. For example, the basis function-based actuator 112 and the neural network-based actuator 114 of the DPD circuit 110 in Figure 1 can be arranged in a parallel configuration as shown in configuration 300. Configuration 300 in Figure 3 may be substantially similar to configuration 200 in Figure 2. For example, the basis function-based actuator 112 in configuration 300 may operate in substantially the same manner as the basis function-based actuator 112 in configuration 200 discussed above. However, the pre-distorted signal 104z corresponds to the output of the basis function-based actuator 112, rather than the combined output of the basis function-based actuator 112 and the neural network-based actuator 114, as shown in Figure 2.

[0050] As shown in Figure 3, the output signal 304 of the neural network-based actuator 114 is provided to the basis function-based actuator 112. More specifically, the neural network-based actuator 114 may perform DPD operation by processing at least one of the input signal 102x or the feedback signal 151y' using one or more neural networks (e.g., convolutional neural network, recurrent neural network, etc.) to generate the output signal 304. The output signal 304 may be used to update the parameters of the basis function-based actuator 112. Thus, one or more neural networks may be trained to generate DPD features (e.g., nonlinear characteristics) of PA130 and update the parameters of the basis function-based actuator 112 to pre-distort the input signal. In certain embodiments, the output signal 304 may be used to update (or program a LUT) the DPD coefficients used by the basis function-based actuator 112 to generate the pre-distorted signal 104z.

[0051] Furthermore, in some embodiments, a basis function-based actuator 112 may preprocess its input (e.g., input signal 102x and / or feedback signal 151y') before applying a set of basis functions, and / or a neural network-based actuator 114 may preprocess its input (e.g., input signal 102x and / or feedback signal 151y') before applying one or more neural networks, as discussed above with reference to Figure 2.

[0052] Arrangement 300 advantageously uses a neural network-based actuator 114 to support a basis function-based actuator 112, allowing it to adapt to dynamic conditions that the basis function-based actuator 112 might otherwise not be able to achieve. In some embodiments, the neural network-based actuator 114 can adapt to dynamic changes in PA 130 and can be used in place of a DPD adaptive circuit to update the basis function-based actuator 112 (and its DPD coefficients and / or LUT).

[0053] In general, the parallel DPD configurations 200 and / or 300 discussed above can operate at different sampling rates. For example, the basis function-based actuator 112 can operate at the full signal sampling rate, while the neural network-based actuator 114 can operate at a lower rate (downsampled rate), as discussed below with reference to Figures 6 and 7. In some cases, the neural network-based actuator 114 may also include multiple neural networks operating at different sampling rates, as discussed more thoroughly below with reference to Figure 7. In this way, the neural network-based actuator 114 can account for dynamic changes and / or long-term memory effects that cannot be tracked or modeled by a set of basis functions used by the basis function-based actuator 112.

[0054] Figure 4 provides a schematic illustrative diagram of exemplary configuration 400 for hybrid basis function, neural network-based DPD according to some embodiments of the present disclosure. For example, the basis function-based actuator 112 and the neural network-based actuator 114 of the DPD circuit 110 in Figure 1 can be arranged in a cascaded configuration as shown in configuration 400.

[0055] As shown in Figure 4, the basis function-based actuator 112 is coupled to the output of the neural network-based actuator 114. For example, the neural network-based actuator 114 may perform a DPD operation by processing at least one of the input signal 102x or the feedback signal 151y' using one or more neural networks (e.g., convolutional neural networks, recurrent neural networks, etc.) to generate an output signal 402. The basis function-based actuator 112 may perform another DPD operation by processing the output signal 402 of the neural network-based actuator 114 using a set of basis functions (e.g., Volterra series, GMP, DDR function, piecewise linear function, etc.) to generate a pre-distorted signal 104z. That is, one or more neural networks may be trained and adapted to pre-process at least one of the input signal 102x or the feedback signal 151y' (e.g., to generate features) before processing by the basis function-based actuator 112.

[0056] Furthermore, in some embodiments, the neural network-based actuator 114 may preprocess its input (e.g., input signal 102x and / or feedback signal 151y') before applying one or more neural networks. Additionally or alternatively, the basis function-based actuator 112 may preprocess its input (e.g., output signal 402 of the neural network-based actuator 114) before applying a set of basis functions.

[0057] Arrangement 400 can advantageously assist the basis function-based actuator 112 in generating features from the input signal 102x and / or feedback signal 151y' using the neural network-based actuator 114. In some embodiments, the output signal 402 of the neural network-based actuator 114 may contain information associated with the features of PA 130 or transformations on the input signal 102x that cannot be readily captured or represented by the basis function-based actuator 112.

[0058] Figure 5 provides a schematic illustrative diagram of exemplary configuration 500 for a hybrid basis function, neural network-based DPD according to several embodiments of the present disclosure. For example, the basis function-based actuator 112 and the neural network-based actuator 114 of the DPD circuit 110 in Figure 1 can be arranged in a cascaded configuration as shown in configuration 500. Configuration 500 in Figure 5 may be substantially similar to configuration 400 in Figure 4, except that the order of the basis function-based actuator 112 and the neural network-based actuator 114 in the signal path is swapped. As shown in Figure 5, the neural network-based actuator 114 is coupled to the output of the basis function-based actuator 112.

[0059] For example, a basis function-based actuator 112 may perform a DPD operation by processing at least one of the input signal 102x or the feedback signal 151y' using a set of basis functions (e.g., Volterra series, GMP, DDR function, piecewise linear function, etc.) to generate an output signal 502. A neural network-based actuator 114 may then perform another DPD operation by processing the output signal 502 of the basis function-based actuator 112 to generate a pre-distorted signal 104z. Furthermore, in some embodiments, the neural network-based actuator 114 may also process the input signal 102x or the feedback signal 151y' using one or more neural networks. Thus, one or more neural networks (used by the neural network-based actuator 114) may be trained and adapted to post-process the output signal of the basis function-based actuator 112 and / or process the input signal 102x or the feedback signal 151y'.

[0060] Furthermore, in some embodiments, the basis function-based actuator 112 may preprocess its input (e.g., input signal 102x and / or feedback signal 151y') before applying one or more neural networks to the input. Additionally or alternatively, the neural network-based actuator 114 may preprocess its input (e.g., output signal 502 of the basis function-based actuator 112) before applying a set of basis functions to the input.

[0061] Configuration 500 may favorably utilize a neural network-based actuator 114 to account for features not modeled by the basis function-based actuator 112 (e.g., dynamic changes and / or nonlinear effects). For example, the basis function-based actuator 112 may have limited memory and therefore may not be able to modify or linearize certain long-term memory effects of PA 130. Thus, the neural network-based actuator 114 may be trained to modify the output of the basis function-based actuator 112 to compensate for long-term fluctuations. Additionally or alternatively, the basis function-based actuator 112 may perform less complex DPD operations, while the neural network-based actuator 114 may perform more complex DPD operations. For example, basis functions may be based on analytical equations and therefore may have limitations (e.g., they may be able to model certain nonlinearities but not other more complex or higher-order nonlinearities). Therefore, the neural network-based actuator 114 can be used to model any nonlinear function (for example, one with higher degrees of freedom), which may or may not be expressible by analytical mathematical equations.

[0062] As can be seen from the above description, the various DPD configurations 200, 300, 400, and 500 have different advantages. Therefore, the DPD circuit 110 may be configured using DPD configurations 200, 300, 400, or 500 depending on the availability of hardware resources (e.g., memory), the nonlinear characteristics of PA 130 under linearization (e.g., trapped charge, memory effect, etc.), and / or the target linearization performance metric. Furthermore, in some embodiments, the neural network-based actuator 114 within the DPD configurations 200, 300, 400, and / or 500 may include a neural network processor or accelerator (e.g., accelerator 670 in Figure 6 and / or accelerator 760 in Figure 7). The neural network processor or accelerator may have an architecture and / or hardware computing that enables efficient computation of neural network-specific processing (e.g., layer processing with multiplication and addition operations to combine weighted outputs from previous layers in a convolutional layer, ReLU operations, bias operations, etc.).

[0063] Exemplary DPD configuration with neural network model One aspect of the present disclosure provides a DPD configuration having an amplifier state estimation branch, which is based on a neural network model and is configured to estimate, predict, and compensate for slow effects such as signal dynamics or GaN charge traps.

[0064] Neural networks have been shown to have the ability to approximate arbitrary nonlinear functions with good accuracy across a wide range of applications. A neural network model can be configured to take both the transmission (e.g., input signal 102x) and loopback observations (e.g., feedback signal 151y') of the PA output (e.g., the output of PA130) and generate a useful feature vector (PA state estimate). Such a neural network can also be trained on a wide variety of waveforms having different dynamic behaviors, e.g., behaviors that cause charge trapping or junction temperature changes in GaN PAs. In addition, in some embodiments, the state dynamics model can be configured to take the previous state estimate and the next input signal and predict the next state (PA state prediction), assuming that PA state changes are primarily caused by input excitations. Thus, predicting future states may no longer require an observation channel. To address much longer time spans of signal / system dynamics, in some embodiments, this model can be trained on downsampled waveforms that may primarily (e.g., only) include the long-term evolution of the system. Together with a pre-distortion model operating at the sampling rate (which may be either a neural network model or a GMP model in various embodiments), the two combinations can perform pre-distortion using a dynamic system via a combiner model, i.e., a model similar to combiner 116 in Figure 2, which takes the output and output I / Q samples of both models into a DAC (e.g., DAC124). Since the PA state estimation network operates at a low sample rate, accelerators (e.g., accelerator 670 in Figure 6 and / or accelerator 760 in Figure 7) can be constructed to handle the operation of the model.

[0065] Modeling long-term effects in PAs over the microsecond to millisecond range based purely on the transmitted signal typically requires long capture buffers (e.g., tens of thousands of samples) and very deep memory models. The DPD configuration with amplifier state estimation branch proposed herein is configured to initially estimate the current PA state by comparing short captures of both the transmitted and observed signals in operation (in contrast to adaptations in conventional implementations). Such a DPD configuration may be further configured to track and predict future PA states using the current PA state and future transmitted signals without observation. In some embodiments, such a DPD configuration may be configured to utilize the estimated PA state to generate a sample-by-sample correction signal to compensate for the main DPD operation.

[0066] Figures 6 to 12 illustrate various embodiments of DPD configurations using neural network models. Figure 6 provides an illustrative diagram of scheme 600 for offline training and online operation for hybrid basis functions, neural network-based DPDs, according to several embodiments of the present disclosure. Scheme 600 includes offline training shown on the left side of Figure 6 and online operation on the right side of Figure 6. To avoid confusing the drawings, Figure 6 shows only the elements prominent in DPD operation, but online operation may include a transmitter circuit between the output of combiner 116 and the input of PA 130 (e.g., similar to transmitter circuit 120 including at least DAC 124) and a receiver circuit between the output of PA 130 and capture buffer 660 (e.g., similar to receiver circuit 150 including ADC 154). In some embodiments, the DPD circuit 110 of Figure 1 and / or configuration 200 of Figure 2 may be trained and deployed using scheme 600. For simplification, Figure 6 may use the same reference numerals as Figures 1 and 2 to refer to the same elements or signals.

[0067] As shown by the online operation on the right side of Figure 6, the basis function-based actuator 112 and the neural network-based actuator 114 of the DPD circuit 110 in Figure 1 are arranged in substantially the same manner as in the arrangement 200 in Figure 2. For example, the online operation may include a basis function-based actuator 112 that processes an input signal 102x using a set of basis functions and a set of associated DPD coefficients to generate an output signal 202. The online operation may further include a neural network-based actuator 114 for processing the captured signal 662 to generate an output signal 204. A combiner 116 may combine the output signal 202 of the basis function-based actuator 112 and the output signal 204 of the neural network-based actuator 114 to generate a pre-distorted signal 104z. Generally, the combiner 116 can combine the output signal 202 from the basis function-based actuator 112, the output signal 204 from the neural network-based actuator 114, and / or the input signal 102x to generate a pre-distorted signal 104z.

[0068] The combiner 116 can have various structures. Figure 6 illustrates one exemplary structure for the combiner 116. In the illustrated example, the combiner 116 may include a post-processing circuit 640 for post-processing the output signal 204 and the input signal 102x to output a post-processed signal 642. The combiner 116 may further include a signal summing circuit 630 for adding the post-processed signal 642 to the output signal 202 (from the basis function-based actuator 112) to generate a pre-distorted signal 104z. Furthermore, delays can be added to any of the output signals 202 from the basis function-based actuator 112, the output signal 204 from the neural network-based actuator 114, and / or the input signal 102x for time alignment before combining, as will be more fully described below with reference to Figure 7. Various exemplary structures for the combiner 116 are more fully described below with reference to Figures 8 to 10. Referring to Figure 1, as discussed above, the pre-distorted signal 104z can be sent to PA130 for transmission. Furthermore, online operation may include a capture memory or buffer 660 (for example, implemented as part of the DPD circuit 110). The capture buffer 660 may perform multiple captures of the pre-distorted signal 104z (or transmission signal) and the feedback signal 151y' (or observation signal) representing the output of PA130. For example, each capture may include a certain number of digital I / Q samples. In some embodiments, the captures may be performed according to a specific duty cycle, for example, depending on the available memory for storing the captures and / or the memory effect of PA130. For example, short segments of the pre-distorted signal 104z (e.g., N samples) and / or short segments of the feedback signal 151y' may be captured at a specific time interval.

[0069] In the illustrated embodiment of Figure 6, the neural network-based actuator 114 may include a neural network accelerator 670. The neural network accelerator 670 may perform processing for a neural network model 610 that has been specifically trained to perform DPD operations (e.g., PA state prediction, PA state estimation, and / or nonlinear pre-compensation) together with a basis function-based actuator 112, for example. In some embodiments, the neural network accelerator 670 may have an architecture and / or hardware computing that enables efficient computation of neural network-specific processing (e.g., layer processing units with multiplication and addition operations to combine weighted outputs from previous layers in a convolutional layer, ReLU operations, bias operations, etc.). As further shown in Figure 6, the neural network-based actuator 114 may include a downsampling circuit 682 for downsampling the captured signal 662 to provide a downsampled signal 683. The neural network accelerator 670 may process the downsampled signal 683 according to the trained neural network model 610 to provide an output signal 676. In some embodiments, the neural network accelerator 670 may optionally track the internal state of the neural network model 610 (e.g., shown as 674), for example, when the neural network model 610 is a recurrent neural network model, it may process the downsampled signal 683 and the internal state 674 to generate an output signal 676. The neural network-based actuator 114 may further include an upsampling circuit 680 for upsampling the output signal 676 of the neural network model 610 to generate a signal 204. That is, the operations within 604 (including the basis function-based actuator 112 and combiner 116) may be performed at the full sampling rate (e.g., the sampling rate of the input signal 102x), while the operations within 608 (including the neural network-based actuator 114) may be performed at a lower sampling rate.For example, the downsampling circuit 682 may downsample its input by a coefficient of K, and the upsampling circuit 680 may upsample its input signal by the same coefficient of K, where K can be 2, 3, 4, or any appropriate value. Figure 6 illustrates a neural network-based actuator 114 operating at a downsampled rate, but in other embodiments, the neural network-based actuator 114 can operate at the same full signal sampling rate as the basis function-based actuator 112.

[0070] In one embodiment, the PA 130 compensated by the combiner 116 may exhibit short-term memory effects using the output (e.g., signal 204) of PA state estimation and tracking implemented by the neural network accelerator 670. These short-term memory effects can be processed (corrected) by the basis function-based actuator 112. In some embodiments, the basis function-based actuator 112, the neural network-based actuator 114, and the combiner 116 may be included within a DPD actuator circuit (e.g., DPD circuit 110). In some cases, the basis function-based actuator 112 may be referred to as the main DPD or sample rate actuator, and the neural network model 610 may be referred to as the PA state estimation model.

[0071] As shown on the left side of Figure 6, pre-training may include the following processes. First, data may be captured from the transmit and observe paths (e.g., of target hardware as shown on the right side of Figure 6) having various input stimuli to PA130 to form a training dataset (shown as capture 602). Capture 602 may include captures of the pre-distorted signal 104z and / or feedback signal 151y'. Next, an optimization algorithm may be used to train a neural network model 610 to estimate the PA state based on the captured data 602. Finally, a combiner model (modeling the combiner 116) may be trained together with the PA state estimation model to generate a correction signal (e.g., the pre-distorted signal 104z). In some embodiments, the PA state estimation model may be trained together with both the main DPD model (modeling the basis function-based actuator 112) and the combiner model to estimate the PA state based on the captured data 602. In other embodiments, the main DPD model may be pre-trained based on the initially captured data 602, and then the PA state estimation model may be trained together with the pre-trained main DPD model and combiner to estimate the PA state based on the captured data 602. In other words, the neural network model 610 may be trained using capture 602, for example by passing captures of the pre-distorted signal 104z and / or feedback signal 151y' through layers of the neural network model 610, and using backpropagation to update the weights (or parameters) of the neural network model 610 until the output of the neural network model 610 is optimized (i.e., when the error between the output predicted by the neural network model 610 and the desired signal is minimized or meets a certain criterion). In some embodiments, after the PA state estimation model or the neural network model 610 has been trained, further pruning and quantization 612 and / or any other post-processing may be performed to generate a neural network model representation 614.In some embodiments, the neural network model representation 614 may include trained weight parameters. The neural network model representation 614 may be used by a neural network-based actuator 114 for online operation. For example, the neural network model representation 614 (trained parameters) may be stored in memory, and the neural network accelerator 670 may use the stored parameters to perform neural network processing.

[0072] As further illustrated by the right side of Figure 6, the operation after deployment may include the following processes: Firstly, a short segment of the transmitted signal (e.g., pre-distorted signal 104z) from the output of the main DPD (e.g., basis function-based actuator 112) may be captured. Optionally, a short segment of the observed signal (e.g., feedback signal 151y') from the output of PA 130 may be captured. Next, a pre-trained PA state estimation model (e.g., model 610) may estimate the PA state using the captured signal 662. In some embodiments, the pre-trained PA state estimation model may directly estimate the PA state by comparing the captured transmitted signal (e.g., a capture of pre-distorted signal 104) with the captured observed signal (e.g., a capture of feedback signal 151), if observations are available. In contrast to prior art DPD implementations where observations are used only for adaptation (i.e., an optimization algorithm that uses observations to find the next set of coefficients in the actuator), with a delay of several seconds from the time of capture to the effect in operation, the PA state estimation and compensation DPD configuration shown on the right side of Figure 6 (similar to configuration 200 in Figure 2) has observations sent directly to the actuator (e.g., basis function-based actuator 112) for compensation, with a delay of microseconds from the time of capture to operation. In some embodiments, a pre-trained PA state estimation model may estimate the PA state using previous PA states and captured transmitted signals (if observations are unavailable). This is in contrast to prior art DPD implementations that use feedforward models that lack the ability to self-correct model drift from observations as needed and do not have an explicit observation receiver duty cycling mechanism for optimal performance. In the next process, the combiner 116 may combine the output of the main DPD (e.g., signal 202) and the estimated PA state vector (e.g., signal 204) to generate a final pre-distortion signal, which is then supplied to PA 130. Subsequently, the main DPD can perform its own adaptation without knowledge of the PA state estimation model and the combiner model.Finally, the PA state estimation model can undergo adaptation, and parameter updates can be performed using any suitable known optimization algorithm (e.g., stochastic gradient descent).

[0073] With respect to sample decimation, in some embodiments, the PA state estimation model included in the neural network accelerator 670 can operate at a decimated sample rate to capture memory effects over long horizons, for example, on the order of microseconds. Any deterministic transformation can be applied to the input before the first downsampling. In some embodiments, the PA state estimation model may include multiple downsampling stages, as shown in Figures 7 and 11.

[0074] Figure 7 provides a schematic illustrative diagram of exemplary configuration 700 for a hybrid basis function, neural network-based DPD according to several embodiments of the present disclosure. Configuration 700 may be substantially similar to the DPD hardware used for online operation in configuration 200 of Figure 2 and / or scheme 600 of Figure 6, and may provide a more detailed diagram of the internal elements of the neural network-based actuator 114 and the interactions (or joint actions) between the basis function-based actuator 112, combiner 116, and the neural network-based actuator 114. For simplification, Figure 7 may use the same reference numerals as Figures 1, 2, and 6 to point to the same elements or the same signals.

[0075] As shown in Figure 7, the configuration 700 may include a basis function-based actuator 112, a neural network-based actuator 114, a combiner 116, an alignment buffer 720, delay circuits 710, 712, 714, and an upsampling circuit 762. The basis function-based actuator 112 may use a set of basis functions and associated DPD coefficients to process an input signal 102x (e.g., a digital baseband I / Q signal encoded with data for transmission) to generate an output signal 202. The alignment buffer 720 may be similar to the capture buffer 660. The alignment buffer 720 may include capturing a feedback signal 151y' (an observed signal indicating the output of PA 130) (e.g., N I / Q samples), capturing the output signal 202 of the basis function-based actuator 112 (e.g., N I / Q samples), and / or capturing a pre-distorted signal 104z (the output of the combiner 116) (e.g., N I / Q samples). The feedback signal 151y', the pre-distorted signal 104z, and the output signal 202 are at the total signal sampling rate (e.g., Fs), are time-aligned, and can be stored in the alignment buffer 720. For example, delays due to processing and / or signal propagation delays in the transmission and feedback paths can be added to each of the output signal 202, the pre-distorted signal 104z, and the feedback signal 151y', such that a one-to-one correspondence at the sample level can exist between these signals stored in the buffer 720.

[0076] As further shown in Figure 7, the neural network-based actuator 114 may include a transformation and feature generation block 730 (e.g., a processor that executes digital circuits and instruction codes), one or more downsampling circuits 740 similar to the downsampling circuit 682, one or more serial-to-parallel (S / P) circuits 750, a neural network accelerator 760 similar to the neural network accelerator 670, and an upsampling circuit 762 similar to the upsampling circuit 680. The transformation and feature generation block 730 may take a time-aligned feedback signal 151y', a pre-distorted signal 104z, an output signal 202 and / or additional features 732 as inputs and generate a signal 734 (e.g., features related to the nonlinear characteristics of PA130) based on the inputs. Some embodiments of the additional feature 732 may include, but are not limited to, analog gain settings (in the transmit and / or receive paths), temperature (e.g., current operating temperature), symbol power of the feedback signal 151y', and / or any operating parameters and / or measurements that may cause the DPD to behave differently. The additional feature 732 may also be provided by other circuits (e.g., detection and / or measurement circuits) and / or read from CPU-accessible registers. As an example, the analog gain setting may be obtained directly from a circuit (e.g., detection and / or measurement circuits). Alternatively, the analog gain setting may be digitally controlled and read back via a register. In another embodiment, a DPD device or RF transceiver including basis function-based actuator 112 and neural network-based actuator 114 may include a temperature sensing circuit that can provide an indication of the current operating temperature via a register read.

[0077] Signal 734 can be downsampled by the downsampling circuit 740. The S / P circuit 750 can perform a serial-to-parallel conversion on the downsampled signal and provide the converted signal 752 to the neural network accelerator 760. For example, the neural network accelerator 760 can process signal 752 using parallel processing. As an example, a neural network model (e.g., Model 610) may operate on a block of downsampled samples at a time, and the S / P circuit 750 can provide the target sample block to the neural network accelerator 760. The neural network-based actuator 114 can optionally include multiple parallel downsampling paths, as shown by the dashed lines for the downsampling circuit 740 and associated S / P circuit 750. In some embodiments, different downsampling paths are executed at different downsampling rates. That is, the neural network accelerator 760 can operate multiple neural network models at different rates for, for example, estimating or predicting different PA state information. After processing signal 752 through the corresponding neural network, the neural network accelerator 760 may generate output signal 204. The upsampling circuit 762 may upsample output signal 204 to signal 763 before providing the output of the neural network-based actuator 114 to the combiner 116. In some embodiments, if the neural network-based actuator 114 includes multiple downsampling paths with different downsampling coefficients, the upsampling circuit 762 may upsample the output from the neural network accelerator 760 according to the corresponding downsampling coefficient so that signal 763 can be returned at the full signal sampling rate. Generally, upsampling can be performed as part of the neural network-based actuator 114, as part of processing in the neural network accelerator 760, or outside of the neural network-based actuator 114.

[0078] As further shown in Figure 7, the input signal 102x, the signal 202 generated by the basis function-based actuator 112, and the signal 763 generated by the neural network-based actuator 114 can be delayed by delay circuits 710, 712, and 714, respectively, before being provided to the combiner 116. Similarly, as discussed above, there may be processing and / or signal propagation delays from different paths. Therefore, the delay circuits 710, 712, and 714 can time-align the input signal 102x, signal 202, and signal 763 before the combiner 116. For example, delay circuit 710 can delay the input signal 102x by K samples, delay circuit 712 can delay signal 202 by L samples, and delay circuit 714 can delay signal 763 by P samples, where K, L, and P can have different values. The combiner 116 can generate a pre-distorted signal 104z by combining a delayed input signal 102x, a delayed output signal 202 from a basis function-based actuator 112, and a delayed output signal 763 from a neural network-based actuator 114. The combiner 116 may have a structure as shown in Figure 6, or other structures as discussed below with reference to Figures 8 to 10.

[0079] As discussed above, the combiner 116 used to combine the output of the basis function-based actuator 112 with the output of the neural network-based actuator 114 shown in Figures 2 and 6-7 can have various structures. Figure 6 illustrates one exemplary structure for the combiner 116. Figures 8-10 illustrate other variations of the combiner structure.

[0080] Figure 8 provides a schematic illustrative diagram of an exemplary arrangement 800 for a combiner in a hybrid basis function, neural network-based DPD, according to several embodiments of the present disclosure. For example, a combiner 116 within a DPD circuit 110 may be arranged using arrangement 800. As shown in Figure 8, the combiner 116 may include an adder circuit 810, a multiplier circuit 820, a low-pass filter (LPF) circuit 830, and an upsampling circuit 840.

[0081] In configuration 800, the upsampling circuit 840 may upsample the signal 204 output by the neural network-based actuator 114 according to the downsampling performed by the neural network-based actuator 114, as discussed above. For example, if the neural network-based actuator 114 operates at half the total signal sampling rate (of the input signal 102x), the upsampling circuit 840 may upsample the output signal 204 by 2, and as a result, the upsampled signal 842 may have the same total signal sampling rate. The LPF circuit 830 may pass the upsampled signal 842 through the LPF. The multiplier circuit 820 may multiply the input signal 102x by the filtered signal 832 to provide the signal 822. The adder circuit 810 may add the signal 822 with the signal 202 output by the basis function-based actuator 112 to generate the pre-distorted signal 104z.

[0082] Although not shown in Figure 8, arrangement 800 may include delay circuits similar to delay circuits 710, 712, and / or 714 to time-align signals 102 and 832 before multiplication in multiplication circuit 820, and / or to time-align signals 202 and 822 before addition in adder circuit 810. Furthermore, in some embodiments, upsampling operation can be optionally performed, for example, when the neural network-based actuator 114 operates at the same rate as the basis function-based actuator 112.

[0083] Figure 9 provides a schematic illustrative diagram of exemplary arrangement 900 for a combiner in a hybrid basis function, neural network-based DPD, according to several embodiments of the present disclosure. For example, a combiner 116 in a DPD circuit 110 may be arranged using arrangement 900. As shown in Figure 9, combiner 116 may include an adder 910, a multiplier 920, an LPF circuit 930, and an upsampling circuit 940 substantially similar to arrangement 800. However, in arrangement 900, combiner 116 does not utilize the input signal 102x to generate a pre-distorted signal 104z.

[0084] For example, the upsampling circuit 940 may upsample the signal 204 output by the neural network-based actuator 114 according to the downsampling performed by the neural network-based actuator 114, as discussed above. The LPF circuit 930 may pass the upsampled signal 942 through the LPF. The multiplier circuit 920 may multiply the signal 202 by the filtered signal 932 to provide the signal 922. The adder circuit 910 may add the signal 922 with the signal 202 output by the basis function-based actuator 112 to generate the pre-distorted signal 104z.

[0085] Although not shown in Figure 9, arrangement 900 may include delay circuits similar to delay circuits 710, 712, and / or 714 to time-align signals 932 and 202 before multiplication in multiplication circuit 920, and / or before addition in adder circuit 910. Furthermore, in some embodiments, upsampling operation can be optionally performed, for example, when the neural network-based actuator 114 operates at the same rate as the basis function-based actuator 112.

[0086] Figure 10 provides a schematic illustrative diagram of an exemplary arrangement 1000 for a combiner in a hybrid basis function, neural network-based DPD, according to some embodiments of the present disclosure. For example, a combiner 116 within a DPD circuit 110 may be arranged using arrangement 1000. As shown in Figure 10, the combiner 116 may include an adder circuit 1010, a neural network 1020, an LPF circuit 1030, and an upsampling circuit 1040.

[0087] In configuration 1000, the upsampling circuit 1040 may upsample the signal 204 output by the neural network-based actuator 114 according to the downsampling performed by the neural network-based actuator 114, for example, as discussed above. The LPF circuit 1030 may pass the upsampled signal 1042 through the LPF. The neural network 1020 (executed by a neural network processor or accelerator such as accelerator 670 in Figure 6, accelerator 760 in Figure 7, and / or neural network processor core 1240 in Figure 12) may process the upsampled signal 1032 and the input signal 102x to provide the signal 1022. The neural network 1020 may include weights trained to combine the output of the neural network-based actuator 114 with the input signal 102x. The summing circuit 1010 can add signal 1022 with signal 202 output by basis function-based actuator 112 to generate a pre-distorted signal 104z.

[0088] Although not shown in Figure 10, arrangement 1000 may include delay circuits similar to delay circuits 710, 712, and / or 714 to time-align signals 1032 and 102 before processing by neural network 1020, and / or to time-align signals 1022 and 202 before addition in adder circuit 1010. Furthermore, in some embodiments, upsampling operation can be optionally performed, for example, when the neural network-based actuator 114 operates at the same rate as the basis function-based actuator 112.

[0089] Generally, the combiner 116 can combine the input signal x, the signal 202 output by the basis function-based actuator 112, and / or the signal 204 output by the neural network-based actuator 114 to generate a pre-distorted signal 104z. In some embodiments, the combiner 116 can upsample and / or filter the output signal 204 output by the neural network-based actuator 114 before combining. Additionally or alternatively, the combiner 116 can multiply the signal 204 output by the neural network-based actuator 114 with the input signal 102x before combining. Additionally or alternatively, the combiner 116 can multiply the signal 204 output by the neural network-based actuator 114 with the signal 202 output by the basis function-based actuator 112 before combining. Additionally or alternatively, the combiner 116 can pass 204 and / or the input signal 102x through the neural network before combining.

[0090] Figure 11 provides a schematic illustrative diagram of exemplary PA state estimation and prediction implementation 1100 for a hybrid basis function, neural network-based DPD according to several embodiments of the present disclosure. For example, a neural network-based actuator 114 of the DPD circuit 110 may be implemented as shown in implementation 1100. As shown in Figure 11, the neural network-based actuator 114 may include a PA state estimation phase 1102 and a PA state prediction phase 1104. The neural network-based actuator 114 may include an estimation neural network model 1116 and a prediction neural network 1126. In some embodiments, the estimation neural network model 1116 and the prediction neural network model 1126 may be executed by a neural network processor or accelerator (e.g., accelerator 670 in Figure 6, accelerator 760 in Figure 7, and / or neural network processor core 1240 in Figure 12).

[0091] During the PA state estimation phase 1102, the neural network-based actuator 114 may utilize an estimated neural network model 1116 for PA state estimation. For example, the neural network-based actuator 114 may estimate the PA state 1118 based on an input signal 102x (e.g., a digital baseband I / Q signal carrying data for transmission by PA 130) and a feedback signal 151y' indicating the output of PA 130. The neural network-based actuator 114 may perform one or more stages of transformation 1110 and / or downsampling 1112 on the input signal 102x and the feedback signal 151y', and provide the transformed and / or downsampled signal 1114 to the estimated neural network model 1116 for processing. The estimated neural network model 1116 may include multiple neural network layers (e.g., an input layer, one or more hidden layers, and an output layer), each having a set of weights, and the transformed and / or downsampled signal 1114 may pass through each of the layers for processing to provide an estimated PA state 1118.

[0092] During the PA state prediction phase 1104, the neural network-based actuator 114 may utilize a predictive neural network model 1126 for PA state prediction. For example, the neural network-based actuator 114 may predict the PA state 1128 based on an input signal 102x (e.g., a digital baseband I / Q signal carrying data for transmission by PA 130). The neural network-based actuator 114 may perform one or more stages of transformation 1120 and / or downsampling 1122 on the input signal 102x and provide the transformed and / or downsampled signal 1124 to the predictive neural network model 1126 for processing. The predictive neural network model 1126 may include multiple neural network layers (e.g., an input layer, one or more hidden layers, and an output layer), each having a set of weights, and the transformed and / or downsampled signal 1124 may pass through each of the layers for processing to provide the predicted PA state 1128.

[0093] In some embodiments, the neural network-based actuator 114 can switch between a PA state estimation phase 1102 and a PA state prediction phase 1104 based on the availability of a feedback signal 151y'. In this regard, the neural network-based actuator 114 can select between an estimation neural network model 1116 or a prediction neural network model 1126 based on the availability of a feedback signal 151y'. For example, if the feedback signal 151y' (observation) is available, the neural network-based actuator 114 may select the estimation neural network model 1116. However, if the feedback signal 151y' is unavailable, the neural network-based actuator 114 may select the prediction neural network model 1126.

[0094] Furthermore, in some embodiments, the neural network-based actuator 114 may include shared storage 1130 (e.g., memory) for storing state information associated with the estimated neural network model 1116 and / or the predictive neural network model 1126. The state information can be fed back to the estimated neural network model 1116 and / or the predictive neural network model 1126 for processing to generate the respective PA states. The neural network-based actuator 114 may further include a delay circuit 1132 that delays the PA state information before processing by the estimated neural network model 1116 and / or the predictive neural network model 1126. As an example, the estimated neural network model 1116 may process the transformed and / or downsampled signal 1114 and previous state information 1134 (e.g., previous estimated PA state 1118 and / or previous predictive PA state 1128) to output a new estimated PA state 1118. The state information in the shared storage 1130 can then be updated with the new estimated PA state 1118. Similarly, the predictive neural network model 1126 may process the transformed and / or downsampled signal 1124 and previous state information 1134 (e.g., previous estimated PA state 1118 and / or previous predicted PA state 1128) to output a new predicted PA state 1128. The state information in the shared storage 1130 can then be updated with the new predicted PA state 1128.

[0095] In other words, for example, if the feedback signal 151y' (observation) is unavailable due to duty cycling in the observation receiver (e.g., receiver circuit 150), the neural network-based actuator 114 may use the predictive neural network model 1126 to predict further PA states according to the input signal 102x and previous state information 1134. On the other hand, if the feedback signal 151y' (observation) is available, the neural network-based actuator 114 may use the estimation neural network model 1116 to estimate future PA states according to the input signal 102x and feedback signal 151y'. In some embodiments, the estimation neural network model 1116 may estimate future PA states based further on the previous state information 1134. Generally, the estimation neural network model 1116 and the predictive neural network model 1126 may operate alternately depending on the availability of the feedback signal 151y'.

[0096] In some embodiments, the neural network-based actuator 114 may include a further processing step after the estimated neural network model 1116 to process the PA state 1118 and generate an estimated PA state 1119. Similarly, the neural network-based actuator 114 may include a further processing step after the predictive neural network model 1126 to process the PA state 1128 and generate a predicted PA state 1129.

[0097] Figure 11 illustrates two distinct paths for performing an action during the PA state estimation phase 1102 and the PA state prediction phase 1104, but the neural network-based actuator 114 can be implemented in any suitable way, for example, by sharing at least some hardware blocks and circuits.

[0098] Figure 12 provides a schematic illustrative diagram of an exemplary hardware architecture 1200 for a neural network-based actuator in a hybrid basis function, neural network-based DPD, according to several embodiments of the present disclosure. Architecture 1200 may be used in conjunction with implementation 1100 of Figure 11. For example, the neural network-based actuator 114 of the DPD circuit 110 may be implemented using implementation 1100 and architecture 1200. As shown in Figure 12, architecture 1200 may include a weight memory 1210, a multiplexer 1220, an inference controller 1230, a neural network processor core 1240, an activation memory 1250, and a direct memory access (DMA) controller 1260.

[0099] The weight memory 1210 may be any suitable volatile or non-volatile memory. Some embodiments of the memory may include double data rate random access memory (DDR RAM), synchronous RAM (SRAM), dynamic RAM (DRAM), flash memory, read-only memory (ROM), etc. The activation memory 1250 may be configured to store estimated neural network model information 1212 and predicted neural network model information 1214. For example, the estimated neural network model information 1212 may include weight parameters and / or any other information associated with the neural network layers of the estimated neural network model 1116. Similarly, the predicted neural network model information 1214 may include weight parameters and / or any other information associated with the neural network layers of the predicted neural network model 1126.

[0100] The multiplexer 1220 may select between estimated neural network model information 1212 or predicted neural network model information 1214 stored in memory 1210, based on the feedback signal (or observation signal) validity indicator signal 1222. For example, if the feedback signal validity indicator signal 1222 indicates that the feedback signal 151y' is valid (or available), the multiplexer 1220 may select estimated neural network model information 1212 and output it to the neural network processor core 1240. However, if the feedback signal validity indicator signal 1222 indicates that the feedback signal 151y' is invalid (or unavailable), the multiplexer 1220 may select predicted neural network model information 1214 and output it to the neural network processor core 1240.

[0101] The activation memory 1250 may be any suitable volatile memory, e.g., DDR RAM, SRAM, DRAM, etc. The activation memory 1250 may be configured to store input data 1252, recursive state information 1254, and output data 1256. For example, the input data 1252 may include a capture of the input signal 102x (pre-distorted by a hybrid basis function, a neural network-based DPD, as disclosed herein, before transmission by the PA 130) and a capture of the feedback signal 151y' (indicating the output of the PA 130). The recursive state information 1254 may include previous PA states estimated by the neural network defined by the estimated neural network model information 1212 (e.g., PA states 1118, 1119) and / or previous PA states predicted by the neural network defined by the predicted neural network model information 1214 (e.g., PA states 1128, 1129). The output data 1256 may include an output signal 204 generated by a neural network-based actuator 114. The output signal 204 can be used to generate a pre-distorted signal 104z for PA 130 as disclosed herein.

[0102] The DMA controller 1260 may be a hardware device configured to provide memory access (e.g., for streaming data) between the neural network processor core 1240 and hybrid basis functions, and other elements in the neural network-based DPD circuit 110. For example, the signal path may be coupled to basis function-based actuators 112 and combiners 116 in the DPD circuit 110, as shown in Figures 6 to 10. In some embodiments, the DMA controller 1260 may transfer input data 1252 from a DPD capture buffer (e.g., capture buffer 660) to activation memory 1250. In some embodiments, the signal path may include other circuits such as downsampling circuits (e.g., downsampling circuits 682, 740, 1112), upsampling circuits (e.g., upsampling circuits 680, 840, 940, 1040), filtering circuits (e.g., LPF circuits 830, 930, 1030), multiplier circuits (e.g., multiplier circuits 820, 920), neural network processors (e.g., accelerator 670 in Figure 6 and / or accelerator 760 in Figure 7, and / or neural network processor 1240), adders (e.g., adders 630, 810, 910, 1010), and so on.

[0103] The neural network processor core 1240 may be configured to perform neural network-specific operations (e.g., convolution, ReLU operation, bias operation, etc.). Depending on the selection in the multiplexer 1220, the neural network processor core 1240 may process the input signal 102x (block of digital I / Q samples) and / or the feedback signal 151y' (block of digital I / Q samples) captured in the input data 1252 using either estimated neural network model information 1212 or predicted neural network model information 1214. For example, when the feedback signal 151y' is available, the neural network processor core 1240 may use the estimated neural network model information 1212 to process the input signal 102x and the feedback signal 151y' to generate an estimated PA state (future PA state). When the input data 1252 does not have the availability of the feedback signal 151y', the neural network processor core 1240 may use the predictive neural network model information 1214 to process the input signal 102x and generate a predicted PA state (future PA state).

[0104] The inference controller 1230 may be configured to trigger inference of the neural network processor core 1240 at a given cadence. For example, the inference controller 1230 can synchronize the inputs, outputs, and / or operations of the neural network processor core 1240 so that data can be streamed between the neural network-based actuator 114 and the combiner 116 and / or basis function-based actuator 112. Furthermore, the inference controller 1230 can inform the neural network processor core 1240 of the neural network model to execute (e.g., an estimated neural network model or a predictive neural network model). For example, in some cases, the inference controller 1230 can provide the neural network processor core 1240 with other information associated with PA state estimation and / or prediction.

[0105] In general, the weight memory 1210, multiplexer 1220, inference controller 1230, neural network processor core 1240, activation memory 1250, and DMA controller 1260 can be arranged in any preferred manner. In some embodiments, the weight memory 1210, multiplexer 1220, inference controller 1230, neural network processor core 1240, activation memory 1250, and / or DMA controller 1260 can be implemented as part of a neural network hardware accelerator (e.g., accelerator 670 in Figure 6 and / or accelerator 760 in Figure 7).

[0106] Figure 12 illustrates the context of a neural network-based actuator 114 utilizing an estimated neural network model and a predictive neural network model, but the embodiments are not limited thereto. For example, architecture 1200 can be used by a neural network-based actuator 114 utilizing any appropriate number of neural networks, where parameters associated with each neural network may be stored in a weighted memory 1210, and a neural network processor 1240 may process the input using these parameters. For example, architecture 1200 can be used in the neural network-based actuator 114 in the configurations 200, 300, 400, and / or 500 discussed above.

[0107] Figure 13 provides a flowchart illustrating method 1300 for performing a hybrid basis function, neural network-based DPD according to several embodiments of the present disclosure. Method 1300 can be implemented by a hybrid basis function, neural network-based DPD that pre-distorts an input signal (e.g., input signal 102x) into a nonlinear component (e.g., PA130), as discussed above with reference to Figures 1 to 12. The operations are illustrated in Figure 13, each once and in a specific order, but the operations can be performed in parallel, rearranged, and / or repeated as desired.

[0108] In 1302, a first signal (e.g., signal 202) can be generated using a first actuation circuit (e.g., basis function-based actuator 112) based on at least one of a set of basis functions, DPD coefficients, and a feedback signal (e.g., feedback signal 151y') indicating the input signal or output of a nonlinear component. The set of basis functions and DPD coefficients can be associated with a first nonlinear characteristic of the nonlinear component.

[0109] In 1304, a second signal (e.g., signal 204) can be generated using a second actuator circuit (e.g., a neural network-based actuator 114) based on a neural network (e.g., neural network model 610, neural network 1020, estimated neural network model 1116, predictive neural network model 1126, estimated neural network model information 1212, and / or predictive neural network model information 1214) and at least one of an input signal or a feedback signal. The neural network is associated with a second nonlinear characteristic of a nonlinear component.

[0110] In 1306, a pre-distorted signal (e.g., a pre-distorted signal 104z) can be generated based on the first signal and the second signal.

[0111] In some embodiments, the first and second actuator circuits may be arranged in parallel. Thus, generating the first signal in 1302 may include applying a set of basis functions and DPD coefficients to at least one of the input signal or feedback signal. Furthermore, generating the second signal in 1304 may include applying a neural network to at least one of the input signal or feedback signal. In some embodiments, generating the pre-distorted signal in 1306 may include combining the first signal generated by the first actuator circuit with the second signal generated by the second actuator circuit, as discussed above with reference to, for example, Figures 2, 6, 7-10. In some embodiments, generating the pre-distorted signal in 1306 may further include combining the input signal with the first signal generated by the first actuator circuit and the second signal generated by the second actuator circuit, as discussed above with reference to, for example, Figures 8-10. In some embodiments, generating a pre-distorted signal in 1306 may include updating the parameters of the first actuator circuit based on a second signal from the second actuator circuit, as discussed above with reference to Figure 3.

[0112] In some embodiments, the first and second actuator circuits may be arranged in a cascade configuration. Thus, in one embodiment, generating a first signal in 1302 may include applying a set of basis functions and DPD coefficients to at least one of the input signal or feedback signal, and generating a second signal in 1304 may include applying a neural network to the first signal generated by the first actuator circuit, for example, as discussed above with reference to Figure 5. In another embodiment, generating a second signal in 1304 may include applying a neural network to at least one of the input signal or feedback signal, and generating a first signal in 1302 may include applying a set of basis functions and DPD coefficients to the first signal generated by the first actuator circuit, for example, as discussed above with reference to Figure 4.

[0113] In some embodiments, the neural network may include an estimated neural network model (e.g., model 1116) and a predictive neural network model (e.g., model 1126), as discussed above with reference to Figures 11-12. Thus, generating a second signal in 1304 may include selecting between the estimated neural network model and the predictive neural network model based on whether the feedback signal is active. Generating a second signal in 1304 may further include applying the estimated neural network model to the input signal and the feedback signal in response to the feedback signal being active. Generating a second signal in 1304 may further include applying the predictive neural network model to the input signal in response to the feedback signal being inactive.

[0114] Embodiment 1 provides an apparatus for a radio frequency (RF) transceiver. The apparatus includes a digital pre-distortion (DPD) actuator for receiving an input signal associated with a nonlinear component of the RF transceiver and outputting a pre-distorted signal, wherein the DPD actuator includes a basis function-based actuator for performing a first DPD operation using a set of basis functions associated with a first nonlinear characteristic of the nonlinear component, and a neural network-based actuator for performing a second DPD operation using a first neural network associated with a second nonlinear characteristic of the nonlinear component, wherein the pre-distorted signal is based on a first output signal of the basis function-based actuator and a second output signal of the neural network-based actuator.

[0115] Embodiment 2 provides an apparatus according to one or more of the preceding and / or subsequent embodiments, wherein a basis function-based actuator performs a first DPD operation by processing at least one of an input signal or a set of basis functions, which is a feedback signal indicating the output of a nonlinear component, to generate a first output signal.

[0116] Embodiment 3 provides an apparatus according to one or more of the preceding and / or following embodiments, wherein a basis function-based actuator further performs a first DPD operation by applying a conversion operation to at least one of the input signal or feedback signal.

[0117] Embodiment 4 provides an apparatus according to one or more of the prior and / or subsequent embodiments, wherein a neural network-based actuator performs a second DPD operation by updating the parameters of a basis function-based actuator based on a second output signal of the neural network-based actuator, and the basis function-based actuator further performs a first DPD operation by generating a first output signal using the updated parameters.

[0118] Example 5 provides an apparatus according to one or more of the prior and / or following embodiments, wherein a basis function-based actuator performs a first DPD operation by processing a second output signal of a neural network-based actuator using a set of basis functions to generate a first output signal.

[0119] Embodiment 6 provides an apparatus according to one or more of the prior and / or following embodiments, wherein a neural network-based actuator performs a second DPD operation by processing at least one of an input signal or a feedback signal indicating the output of a nonlinear component using a first neural network to generate a second output signal.

[0120] Embodiment 7 provides an apparatus according to one or more of the prior and / or subsequent embodiments, wherein a neural network-based actuator further performs a second DPD operation by applying a conversion operation to at least one of the input signal or feedback signal. In one embodiment, the conversion includes generating a feature based on at least one of an analog gain setting, temperature, or symbol power associated with the feedback signal.

[0121] Example 8 provides an apparatus according to one or more of the prior and / or following embodiments, wherein a neural network-based actuator further performs a second DPD operation by downsampling at least one of the input signal or feedback signal.

[0122] Example 9 provides an apparatus according to one or more of the prior and / or following embodiments, wherein a neural network-based actuator performs a second DPD operation by processing a first output signal of a basis function-based actuator using a first neural network to generate a second output signal.

[0123] Example 10 provides an apparatus according to one or more of the prior and / or following embodiments, wherein a neural network-based actuator further performs a second DPD operation by processing an input signal using a first neural network.

[0124] Example 11 provides an apparatus according to one or more of the prior and / or following embodiments, wherein the DPD actuator further includes a combiner for generating a pre-distorted signal by combining a first output signal of a basis function-based actuator with a second output signal of a neural network-based actuator.

[0125] Example 12 provides an apparatus according to one or more of the preceding and / or following embodiments, wherein the combiner further combines the input signal with a first output signal of a basis function-based actuator and a second output signal of a neural network-based actuator to generate a pre-distorted signal.

[0126] Example 13 provides an apparatus according to one or more of the prior and / or subsequent embodiments, wherein the combiner further multiplies the second output signal of the neural network-based actuator with the input signal or first output signal of the basis function-based actuator before combining the first output signal of the basis function-based actuator with the second output signal of the neural network-based actuator.

[0127] Example 14 provides an apparatus according to one or more of the prior and / or subsequent embodiments, wherein the combiner further applies a conversion operation to the second output signal of a neural network-based actuator before combining the first output signal of a basis function-based actuator with the second output signal of a neural network-based actuator.

[0128] Example 15 provides an apparatus described in one or more of the prior and / or following embodiments, wherein a transformation operation applied to a second output signal of a neural network-based actuator is associated with at least one of the following: upsampling, filtering, signal alignment, or a second neural network different from the first neural network.

[0129] Example 16 provides an apparatus according to one or more of the prior and / or following embodiments, wherein the DPD actuator further includes a combiner for processing at least one of an input signal or a second output signal of a neural network-based actuator using a second neural network different from the first neural network to generate a third output signal, and for combining the third output signal with the first output signal of a basis function-based actuator to generate a pre-distorted signal.

[0130] Example 17 provides an apparatus according to one or more of the prior and / or subsequent embodiments, wherein the combiner further applies a conversion operation to a second output signal of a neural network-based actuator.

[0131] Example 18 provides an apparatus according to one or more of the prior and / or following embodiments, wherein a transformation operation applied to a second output signal of a neural network-based actuator is associated with at least one of an upsampling operation or a filtering operation.

[0132] Example 19 provides an apparatus according to one or more of the prior and / or subsequent embodiments, wherein a neural network-based actuator further uses a second neural network to perform a second DPD operation, and the first and second neural networks operate at different sampling rates. For example, a basis function-based actuator performs a first DPD operation at a first sampling rate, and a neural network-based actuator performs a second DPD operation at a second sampling rate different from the first sampling rate.

[0133] Example 20 provides an apparatus for a radio frequency (RF) transceiver. The apparatus includes a DPD actuator for performing digital pre-distortion (DPD) on an input signal associated with a nonlinear component of the RF transceiver, the DPD actuator comprising: a first actuator for processing the input signal based on a set of basis functions and DPD coefficients to generate a first output signal; a second actuator for processing at least one of the input signal or a feedback signal indicating the output of the nonlinear component using one or more neural networks to generate a second output signal; and a combiner for generating a pre-distorted signal based on the first and second output signals, wherein the set of basis functions, DPD coefficients, and one or more neural networks are each associated with one or more nonlinear characteristics of the nonlinear component.

[0134] Example 21 provides an apparatus according to one or more of the prior and / or following embodiments, wherein a second actuator selects between a first neural network and a second neural network from among one or more neural networks based on the availability of a feedback signal.

[0135] Example 22 provides an apparatus described in one or more of the prior and / or subsequent embodiments, wherein a second actuator processes an input signal and a feedback signal by using a first neural network to generate a second output signal, and the use of the first neural network is based on the availability of the feedback signal.

[0136] Example 23 provides an apparatus described in one or more of the prior and / or subsequent embodiments, wherein a second actuator processes an input signal by using a second neural network to generate a second output signal, and the use of the second neural network is based on the lack of availability of a feedback signal.

[0137] Example 24 provides an apparatus according to one or more of the prior and / or subsequent embodiments, wherein a second actuator generates a second output signal by using one of the first or second neural networks, selected from the first and second neural networks, and processes at least one of the input signal or feedback signal and previous state information associated with at least one of the first or second neural networks, and updates the state information associated with at least one of the first or second neural networks based on the second output signal.

[0138] Example 25 provides an apparatus according to one or more of the prior and / or subsequent embodiments, further comprising a memory for storing parameters associated with one or more neural networks, and a neural network processor for performing neural network-specific operations, wherein a second actuator processes at least one of an input signal or a feedback signal by using the neural network processor and the stored parameters.

[0139] Example 26 provides a method for performing digital pre-distortion (DPD) to pre-distort an input signal to a nonlinear component. The method includes: generating a first signal using a first actuator circuit based on a set of basis functions, DPD coefficients, and at least one of an input signal or a feedback signal indicating the output of a nonlinear component, wherein the set of basis functions and DPD coefficients are associated with a first nonlinear characteristic of the nonlinear component; generating a second signal using a second actuator circuit based on a neural network and at least one of an input signal or a feedback signal, wherein the neural network is associated with a second nonlinear characteristic of the nonlinear component; and generating a pre-distorted signal based on the first and second signals.

[0140] Example 27 provides a method according to one or more of the prior and / or subsequent examples, wherein generating a first signal includes applying a set of basis functions and DPD coefficients to at least one of the input signal or feedback signal, and generating a second signal includes applying a neural network to at least one of the input signal or feedback signal.

[0141] Example 28 provides a method according to one or more of the prior and / or subsequent examples, wherein generating a pre-distorted signal involves combining a first signal generated by a first actuator circuit with a second signal generated by a second actuator circuit.

[0142] Example 29 provides a method according to one or more of the prior and / or subsequent examples, further comprising generating a pre-distorted signal by combining an input signal with a first signal generated by a first actuator circuit and a second signal generated by a second actuator circuit.

[0143] Example 30 provides a method according to one or more of the prior and / or subsequent examples, wherein generating a pre-distorted signal includes updating the parameters of a first actuator circuit based on a second signal from a second actuator circuit.

[0144] Example 31 provides a method according to one or more of the prior and / or subsequent examples, wherein generating a first signal includes applying a set of basis functions and DPD coefficients to at least one of an input signal or a feedback signal, and generating a second signal includes applying a neural network to the first signal generated by a first actuator circuit.

[0145] Example 32 provides a method according to one or more of the prior and / or subsequent examples, wherein generating a second signal includes applying a neural network to at least one of an input signal or a feedback signal, and generating a first signal includes applying a set of basis functions and DPD coefficients to a first signal generated by a first actuator circuit.

[0146] Example 33 provides a method according to one or more of the prior and / or following examples, wherein the neural network includes an estimated neural network model and a predictive neural network model, and generates a second signal, which includes selecting between the estimated neural network model and the predictive neural network model based on whether a feedback signal is valid, applying the estimated neural network model to the input signal and the feedback signal in response to the feedback signal being valid, and applying the predictive neural network model to the input signal in response to the feedback signal being invalid.

[0147] Modification and implementation forms Various embodiments of DPDs using a combination of basis function-based and neural network-based actsuations are described herein with reference to the “input signal for PA,” i.e., a signal generated based on an input signal x, which is a drive signal for PAs. However, in other embodiments of DPDs using a combination of basis function-based and neural network-based actsuations, the “input signal for PAs” may be a bias signal used to bias N PAs. Thus, embodiments of the present disclosure also include DPD arrangements involving a combination of basis function-based and neural network-based actuators, similar to those described herein and illustrated in the figures, except that instead of modifying the drive signal for PAs, the DPD arrangement may be configured to modify the bias signal for PAs, which may be done based on a control signal generated by a DPD adaptive circuit (e.g., the DPD adaptive circuit described herein), where the output of the PAs is based on a bias signal used to bias the PAs. In other embodiments of the present disclosure, both the drive signal and the bias signal for PAs may be tuned as described herein to implement DPDs using a neural network.

[0148] While some of the descriptions provided herein refer to PAs, in general, the various embodiments of DPDs presented herein, including combinations of basis function-based actuators and neural network-based actuators, are applicable to amplifiers other than PAs, such as low-noise amplifiers and variable-gain amplifiers, as well as to nonlinear electronic components of RF transceivers other than amplifiers (i.e., components that may exhibit nonlinear behavior). Furthermore, while some of the descriptions provided herein refer to millimeter-wave / 5G technologies, in general, the various embodiments of DPDs using neural networks presented herein are applicable to any technology or standard wireless communication system other than millimeter-wave / 5G, any wireless RF system other than a wireless communication system, and / or RF system other than a wireless RF system.

[0149] While embodiments of this disclosure have been described above with respect to exemplary implementation configurations as shown in Figures 1 to 13, those skilled in the art will understand that the various teachings described above are applicable to a wide variety of other implementation configurations.

[0150] In certain circumstances, the features discussed herein may be applicable to automotive systems, safety-critical industrial applications, medical systems, scientific instruments, wireless and wired communications, radio, radar, industrial process control, audio and video equipment, current sensing, instruments (which may be extremely precise), and other digital processing-based systems.

[0151] In the discussion of the embodiments described above, system components such as multiplexers, multipliers, adders, delay taps, filters, converters, mixers, and / or other components can be readily replaced, substituted, or otherwise modified to meet the needs of a particular circuit. Furthermore, it should be noted that the use of complementary electronic devices, hardware, software, etc., provides equally viable options for implementing the teachings of this disclosure relating to the application of model architecture lookup for hardware configurations in various communication systems.

[0152] Components of various systems for using hybrid basis functions, neural network-based DPD techniques, as proposed herein, may include electronic circuits for performing the functions described herein. In some cases, one or more components of the system may be provided by a processor specifically configured to perform the functions described herein. For example, the processor may include one or more application-specific components or programmable logic gates configured to perform the functions described herein. The circuits may operate in the analog domain, the digital domain, or the mixed-signal domain. In some cases, the processor may be configured to perform the functions described herein by executing one or more instructions stored in a non-temporary computer-readable storage medium.

[0153] In one exemplary embodiment, any number of electrical circuits shown in the figures may be mounted on a substrate of the associated electronic device. The substrate may be a general circuit board capable of holding various components of the internal electronic system of the electronic device and further providing connectors for other peripheral devices. More specifically, the substrate may provide electrical connections so that other components of the system can communicate electrically. Any suitable processor (including DSPs, microprocessors, support chipsets, etc.), computer-readable non-temporary memory elements, etc., can be appropriately coupled to the substrate based on specific configuration needs, processing requirements, computer design, etc. Other components such as external storage, additional sensors, audio / video display controllers, and peripheral devices may be attached to the substrate as plug-in cards via cables, or integrated into the substrate itself. In various embodiments, the functionality described herein may be implemented in emulation form as software or firmware operating within one or more configurable (e.g., programmable) elements located within a structure that supports these functionality. The software or firmware providing the emulation may be provided on a non-temporary computer-readable storage medium containing instructions that enable the processor to perform those functionality.

[0154] In another exemplary embodiment, the electrical circuits in this figure may be implemented as standalone modules (e.g., devices comprising related components and circuits configured to perform a specific application or function) or as plug-in modules into application-specific hardware for electronic devices. Note that certain embodiments of this disclosure can be readily contained, either partially or entirely, within a system-on-chip (SOC) package. An SOC represents an IC that integrates components of a computer or other electronic system onto a single chip. It may include digital, analog, mixed-signal, and often RF functions, all of which may be provided on a single chip substrate. Other embodiments may include a multi-chip module (MCM) comprising multiple separate ICs located within a single electronic package and configured to interact closely with one another through the electronic package.

[0155] Furthermore, it is essential to note that the specifications, dimensions, and relationships outlined herein (e.g., the apparatus, DPD arrangement, and / or the number of components of the RF transceiver shown in Figures 1 to 12) are provided solely for illustrative and teaching purposes. Such information may vary considerably without departing from the spirit of this disclosure or the scope of the appended claims. It should be understood that the system may be solidified in any preferred manner. In line with similar design alternatives, any of the illustrated circuits, components, modules, and elements can be combined in a variety of possible configurations, all clearly within the broad scope of this specification. In the foregoing description, exemplary embodiments have been described with reference to specific processor and / or component arrangements. Various modifications and changes can be made to such embodiments without departing from the scope of the appended claims. Therefore, the description and drawings should be taken in an illustrative rather than restrictive sense.

[0156] It should be noted that in many of the embodiments provided herein, interactions may be described with respect to two, three, four, or more electrical components. However, this is done for clarification and illustrative purposes only. It should be understood that the system can be solidified in any preferred manner. Along with alternatives to similar designs, any of the illustrated components, modules, and elements may be combined into a variety of possible configurations, all of which are clearly within the broad scope of this specification. In some cases, it may be easier to describe one or more of the functionality of a given set of flows by referring to only a limited number of electrical elements. It should be understood that the electrical circuits and their teachings in the figures are readily extensible and can accommodate a large number of components, as well as more complex / sophisticated arrangements and configurations. Therefore, the embodiments provided should not limit the scope or preclude the broad teachings of electrical circuits that are potentially applicable to countless other architectures.

[0157] In this specification, references to various features (e.g., elements, structures, modules, components, steps, operations, characteristics, etc.) included in “one embodiment,” “exemplary embodiment,” “embodiment,” “another embodiment,” “several embodiments,” “various embodiments,” “other embodiments,” “alternative embodiments,” etc., are intended to mean that any such feature is included in one or more embodiments of this disclosure, but may be combined in the same embodiment, or may not necessarily be combined. Also, when used herein, including in the claims, “or” in a list of items (e.g., a list of items preceded by phrases such as “at least one” or “one or more”) indicates a comprehensive list, for example, the list [at least one A, B, or C] means A or B or C or AB or AC or BC or ABC (i.e., A and B and C).

[0158] The various aspects of the illustrated embodiments are described using terminology commonly adopted by those skilled in the art to communicate their work to others skilled in the art. For example, the term “connected” means a direct electrical connection between things that are connected without any intermediate devices / components, while the term “coupled” means either a direct electrical connection between things that are connected, or an indirect connection via one or more passive or active intermediate devices / components. In another embodiment, the term “circuit” means one or more passive and / or active components arranged to cooperate with each other to provide a desired function. Also, as used herein, terms such as “substantially,” “approximately,” and “about” may be used to generally mean within + / - 20% of a target value, for example, within + / - 10% of a target value, based on the context of a particular value described herein or known in the art.

[0159] Numerous other changes, substitutions, modifications, alterations, and modifications may be apparent to those skilled in the art, and this disclosure is intended to encompass all such changes, substitutions, modifications, alterations, and modifications as they fall within the scope of the examples and the appended claims. It should be noted that all optional features of the apparatus described above may also be implemented in relation to the methods or processes described herein, and the details of the examples may be used anywhere in one or more embodiments. [Explanation of Symbols]

[0160] 100 RF Transceiver 102 Input Signal 104 Output signal 110 DPD circuit 112 DPD Actuator Circuit 114 Actuator 116 Combiner 120 Transmitter Circuit 122 Digital Filters 124 Digital-to-Analog Converter 126 Analog Filters 128 Mixer 130 PA array 131 Output signal 140 Antenna Array 150 Receiver Circuit 151 Feedback signal 152 Digital Filters 154 Analog-to-Digital Converter 156 Analog Filters 158 Mixer 160 Local Oscillator (LO) 200, 300, 400, 500, 700, 800, 900, 1000 DPD arrangement Output signals 202, 204, 304, 402, 502 600 Scheme 602 Data 610 Neural Network Models 612 Quantization 614 Neural Network Model Representations 630 Signal Adder Circuit 640 Post-processing circuit 642 Post-processing signal 660 capture buffers 662 signal 670 Neural Network Accelerators 674 Internal State 676 Output signal 680 Upsampling Circuit 682 Downsampling Circuit 683 Signal 710, 712, 714 Delay Circuits 720 Alignment Buffer 730 Feature generation block 734 signal 740 Downsampling Circuit 750 S / P circuit 752 signal 760 Neural Network Accelerators 762 Upsampling Circuit 763 Output signal 810 Adder Circuit 820 Multiplier Circuit 822 signal 830 LPF circuit 832 signal 840 Upsampling Circuit 842 signal 910 Adder Circuit 920 Multiplier Circuit 922 signal 930 LPF circuit 932 signal 940 Upsampling Circuit 942 signal 1010 Adder Circuit 1020 Neural Networks 1022 signal 1030 LPF circuit 1032 signal 1040 Upsampling Circuit 1042 signal 1100 Implementation 1102 PA State Estimation Phase 1104 PA State Prediction Phase 1110 conversion 1112 Downsampling Circuit 1114 Signal 1116 Estimated Neural Network Models 1118 Estimated PA status 1119 Estimated PA status 1120 conversion 1122 Downsampling 1124 Signal 1126 Predictive Neural Network Models 1128 Predicted PA status 1129 Predicted PA status 1130 Shared Storage 1132 Delay Circuit 1134 Status Information 1200 Hardware Architectures 1210 memory 1212 Estimated Neural Network Model Information 1214 Predictive Neural Network Model Information 1220 Multiplexer 1222 Feedback signal effectiveness indicator signal 1230 Inference Controller 1240 Neural Network Processors 1250 Activation Memory 1252 Input data 1254 Recursive state information 1256 Output data 1260 DMA Controller

Claims

1. A device for a radio frequency (RF) transceiver, wherein the device is The RF transceiver includes a digital pre-distortion (DPD) actuator for receiving an input signal associated with the nonlinear characteristics of the nonlinear component and outputting a pre-distorted signal, wherein the DPD actuator A basis function-based actuator for performing a first DPD operation using a set of basis functions associated with the first nonlinear characteristics of the nonlinear component, A neural network-based actuator for performing a second DPD operation using a first neural network associated with the second nonlinear characteristics of the aforementioned nonlinear component, A device comprising: a combiner for generating the pre-distorted signal by combining the input signal with a first output signal of the basis function-based actuator and a second output signal of the neural network-based actuator.

2. The apparatus according to claim 1, wherein the basis function-based actuator performs the first DPD operation by processing at least one of the input signals or the set of basis functions to process a feedback signal indicating the output of the nonlinear component to generate the first output signal.

3. The neural network-based actuator described above The second DPD operation is performed by updating the parameters of the basis function-based actuator based on the second output signal of the neural network-based actuator. The aforementioned basis function-based actuator further, The apparatus according to claim 2, further comprising using the updated parameters to generate the first output signal, thereby performing the first DPD operation.

4. The apparatus according to claim 1, wherein the basis function-based actuator performs the first DPD operation by processing the second output signal of the neural network-based actuator using the set of basis functions to generate the first output signal.

5. The apparatus according to claim 1, wherein the neural network-based actuator performs the second DPD operation by processing at least one of the input signal or the first neural network to generate the second output signal.

6. The apparatus according to claim 1, wherein the neural network-based actuator performs the second DPD operation by processing the first output signal of the basis function-based actuator using the first neural network to generate the second output signal.

7. The apparatus according to claim 1, wherein the combiner further applies at least one data transformation operation, among signal alignment, upsampling, filtering, or processing using another neural network, to the second output signal of the neural network-based actuator before combining the first output signal of the basis function-based actuator with the second output signal of the neural network-based actuator.

8. The DPD actuator, Using a second neural network different from the first neural network, at least one of the input signal or the second output signal of the neural network-based actuator is processed to generate a third output signal. The apparatus according to claim 1, further comprising a combiner for generating the pre-distorted signal by combining the third output signal with the first output signal of the basis function-based actuator.

9. A device for a radio frequency (RF) transceiver, wherein the device is The RF transceiver is equipped with a DPD actuator for performing digital pre-distortion (DPD) on an input signal associated with the nonlinear characteristics of the nonlinear component of the RF transceiver, wherein the DPD actuator, A first actuator for processing the input signal based on a set of basis functions and DPD coefficients to generate a first output signal, A second actuator for generating a second output signal by processing at least one of the input signal or a feedback signal indicating the output of the nonlinear component using one or more neural networks, wherein the second actuator selects from among the one or more neural networks, a first neural network and a second neural network, based on the availability of the feedback signal. A combiner for generating a pre-distorted signal based on the first output signal and the second output signal, An apparatus in which the set of basis functions, the DPD coefficients, and the one or more neural networks are each associated with one or more nonlinear characteristics of the nonlinear component.

10. The apparatus according to claim 9, wherein the second actuator generates the second output signal by using the first neural network to process the input signal and the feedback signal, and the use of the first neural network is based on the availability of the feedback signal.

11. The apparatus according to claim 9, wherein the second actuator processes the input signal by using the second neural network to generate the second output signal, and the use of the second neural network is based on the lack of availability of the feedback signal.

12. The second actuator described above, The second output signal is generated by using the selected one of the first neural network or the second neural network. The input signal or the feedback signal, at least one of the above, The process involves processing the previous feature vector of the nonlinear component associated with at least one of the first neural network or the second neural network, The apparatus according to claim 9, wherein the feature vector of the nonlinear component associated with at least one of the first neural network or the second neural network is updated based on the second output signal.

13. A memory for storing parameters associated with one or more neural networks, It further comprises a neural network processor for performing neural network-specific operations, The apparatus according to claim 9, wherein the second actuator processes at least one of the input signal or the feedback signal by using the neural network processor and the stored parameters.

14. A method for performing digital pre-distortion (DPD) to pre-distort an input signal to a nonlinear component, wherein the method is A first actuator circuit is used to generate a first signal based on a set of basis functions, DPD coefficients, and at least one of the input signal or a feedback signal indicating the output of the nonlinear component, wherein the set of basis functions and the DPD coefficients are associated with a first nonlinear characteristic of the nonlinear component. A second actuator circuit is used to generate a second signal based on a neural network and at least one of the input signal or the feedback signal, wherein the neural network is associated with the second nonlinear characteristic of the nonlinear component. This includes generating a pre-distorted signal based on the first signal and the second signal, The neural network includes an estimated neural network model and a predictive neural network model. The generation of the second signal is Based on whether the feedback signal is effective, select between the estimated neural network model and the predictive neural network model. In response to the validity of the feedback signal, the estimated neural network model is applied to the input signal and the feedback signal. A method comprising applying the predictive neural network model to the input signal in response to the feedback signal being invalid.

15. Generating the first signal includes applying the set of basis functions and the DPD coefficients to at least one of the input signal or the feedback signal, The method according to claim 14, wherein generating the second signal includes applying the neural network to at least one of the input signal or the feedback signal.

16. The method according to claim 15, wherein generating the pre-distorted signal includes combining the first signal generated by the first actuator circuit and the second signal generated by the second actuator circuit.

Images