Method and system for enabling real-time adaptive radio frequency transmitter optimization - Patents.com

The integration of SDR and GANs for load-pull extrapolation addresses the real-time optimization challenge of RF transmitters, enhancing their adaptability and efficiency in dynamic spectral environments.

JP2024544357AActive Publication Date: 2024-11-29BAYLOR UNIVERSITY
View PDF 1 Cites 0 Cited by

Patent Information

Application Number
JP2024522228
Authority / Receiving Office
JP · JP
Patent Type
Applications
Current Assignee / Owner
Priority Date
2021-10-15
Filing Date
2022-10-17
Publication Date
2024-11-29
Estimated Expiration
2042-10-17

AI Technical Summary

Technical Problem

Traditional RF transmitters struggle to dynamically optimize their transmission configurations in real-time due to computational intensity, making them unsuitable for rapidly changing environments and leading to inefficient spectrum utilization.

Method used

Utilizing a software-defined radio (SDR) with load-pull extrapolation and generative adversarial networks (GANs) for deep learning image completion to quickly identify optimal load impedances, reducing the number of measurements required for real-time optimization of RF transmitters.

Benefits of technology

Enables rapid and efficient real-time optimization of RF transmitters, improving spectrum utilization by significantly reducing the time needed for impedance tuning and enhancing the adaptability of RF systems to dynamic spectral environments.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 2024544357000001_ABST
    Figure 2024544357000001_ABST
Patent Text Reader

Abstract

A disclosed radio frequency (RF) system, such as a cognitive radar, includes a software defined radio (SDR), an adaptive transmit amplifier, and a host computer. The system performs an optimization operation including selecting an initial impedance as a load impedance for the RF device and iteratively performing an image completion operation until a convergence criterion is met. The image completion operation may include measuring the performance of the RF device to obtain a measured performance corresponding to the load impedance, storing the measured performance as a point on a measured load pull contour image, performing a load pull extrapolation to extrapolate a predicted optimal impedance from the load impedance, and saving the predicted impedance as the load impedance for the next iteration of the image completion operation. The convergence criterion is met if the predicted impedance matches one of the previously measured impedances.
Need to check novelty before this filing date? Find Prior Art

Description

[Technical field]

[0001] The present disclosure relates generally to adaptive radio frequency (RF) transmitters, and more particularly to real-time optimization of RF transmitters.

[0002] (CROSS REFERENCE TO RELATED APPLICATIONS) Under 35 USC § 119(e), this application claims the benefit of and priority to U.S. Application No. 63 / 256,001, filed October 15, 2021, which is referred to herein as the "Provisional Application," the disclosure of which is incorporated herein by reference in its entirety.

[0003] (STATEMENT REGARDING FEDERALLY SPONSORED RESEARCH OR DEVELOPMENT) Not applicable.

[0004] (See attached) Not applicable. [Background technology]

[0005] As the use of high-bandwidth communications applications increases, the available radio spectrum becomes congested, making it impossible for many other users of the radio spectrum to continue operating in a static, inflexible manner. Among these users are radar systems, for whose use large amounts of bandwidth have historically been reserved. To adapt and coexist with new technologies in a dynamically managed environment, next-generation radars will need to be able to reconfigure and optimize their spectrum configuration in real time.

[0006] In 2009, then-Federal Communications Commission (FCC) Chairman Julius Genachowski said, "I believe the greatest threat to America's mobile future is a looming spectrum crisis." The crisis he foresaw was the inability to provide sufficient spectrum for future wireless communication applications and systems, such as fifth-generation wireless (5G) technology. At the time, mobile data usage was expected to grow from 6PB / month in 2008 to 400PB / month in 2013. More recently, Ericsson projected that global mobile data usage will reach 226EB / month in 2026. To address this lack of spectrum availability, the 2010 National Broadband Plan directed the FCC to auction some areas of spectrum previously allocated to other uses, and new reallocation proposals are continually under review. These auctions have helped to alleviate spectrum pressure, but because the auction and reallocation process takes years to complete, the usefulness of this solution is limited to situations in which the need to shuffle spectrum can be recognized in advance.

[0007] It is important to note that spectrum auctions alone are unlikely to be a long-term solution to spectrum congestion, as recognized by the President's Council of Advisors on Science and Technology (PCAST) in 2012. PCAST ​​argued that auctioning federal spectrum is unsustainable due to cost, time, and disruption to government missions, but that is not the only problem. Indeed, as long as there are intermittent or infrequent spectrum applications (including various military and aerospace systems), long-term static allocations are fundamentally incompatible with the need for high spectrum utilization (the ratio of occupied spectrum to total spectrum). With traditional approaches to spectrum allocation, spectrum utilization studies have revealed that densely populated areas such as New York City, Chicago, and Dublin may utilize as little as 13-17% of available spectrum at peak times, and utilization within a particular band often reflects the nature of the applications assigned to that band. See, for example, Mark A. McHenry et al., Spectrum Occupancy Measurements Location 4 of 6, Republican National Convention, New York City, NY, August 30, 2004 - September 3, 2004, Rev. 2 (Shared Spectrum, Inc., August 15, 2005) (reported utilization of television bands is up to 77%, cellular bands is up to 46%, and radar and aerospace application bands are 0.2-5%). Clearly, not all wireless applications are equal in terms of spectrum utilization. However, traditional application-based spectrum allocation (where specific portions of the spectrum are allocated to specific uses) ignores this utilization imbalance, resulting in some bands being congested while other bands are very underutilized in comparison.

[0008] In the field of high power RF transmitters, including radar transmitters, techniques have been described that use tunable amplifier matching networks to optimize transmission configurations, including but not limited to carrier frequency, bandwidth, waveform type, etc. However, such techniques are generally too computationally intensive and therefore not suitable for use cases requiring real-time optimization. Summary of the Invention

[0009] The disclosed subject matter addresses the inability of conventional RF transmitters to dynamically optimize the transmit configuration in real time by presenting a faster method for optimizing the transmit configuration of an RF amplifier using a software-defined radio (SDR) configured to perform a load-pull extrapolation-based search to quickly identify the optimal load impedance.

[0010] More efficient approaches to spectrum management include dynamically allocating spectrum resources "on the fly" depending on the users in a particular environment. Various spectrum management approaches described in the prior art address the benefits and challenges faced in alternative arrangements. Some of these approaches are beginning to see real-world application under the Citizens Broadband Radio Service (CBRS), and additional spectrum sharing options are currently being explored as part of the US Mid Band Initiative. The introduction of phased array antenna technology adds an additional dimension of spectrum control with further control over transmission direction, further enhancing the potential for efficient use of spectrum resources.

[0011] To take full advantage of the flexibility afforded by a dynamic allocation system, devices must be able to be reconfigured in time to operate efficiently in the available bands. For high power applications such as radar or cell phone base stations, this reconfiguration requires optimizing the system's transmit chain for optimal performance at the selected frequency. Work has been done to develop optimization methods for radar transmit amplifiers using adjustable load impedance tuners, but the measurement techniques used are insufficient for real-time operation, with each individual measurement set taking several seconds to complete. Active load pull systems have demonstrated significant improvements in tuning time, but are impractical for use in deployed systems. Software-defined radios (SDRs), on the other hand, have been used to demonstrate measurement techniques. The subject matter disclosed herein is the use of SDRs to perform load pull extrapolation to achieve significant improvements in the search for optimal configurations. The predicted optimal impedance is used to guide the search, alternating between selecting a region of interest by extrapolation and taking new measurements in the region of interest to identify a new predicted optimal impedance. The efficiency of this alternating process is determined primarily by the reconfiguration period of the components and the computational requirements of the adaptation algorithm.

[0012] Typically, high-power, mechanically actuated impedance tuners cannot adapt fast enough to provide and maintain optimal performance for rapidly adapting RF systems. A cognitive radar that can create its own radar pulses designed to fit the existing spectrum environment is one example of such a system. Typically, the pulse repetition interval (PRI) of a radar ranges from 10 μs to 100 μs, while a fast mechanically actuated impedance tuner may take 40 ms to 100 ms per tuning operation. As a result, such tuners can only be used to optimize the average performance of such a radar over a large number of pulses. Unfortunately, optimization typically requires considering multiple transmit configurations, each of which must be evaluated at multiple impedances during each search iteration. While the measurement period can be adjusted to accommodate this reality, many transmit configurations have similar effects on the transmit amplifier (since they can be treated as identical during the search process), so most measurements may provide redundant information. It is therefore reasonable to consider a grouping strategy that treats the transmit configurations within each group as essentially equivalent, greatly reducing the number of measurements required during optimization.

[0013] However, during the optimization process, it is difficult to obtain enough information to confirm the equivalence of different transmit configurations. Because of the generally regular nature of amplifier performance contours on the Smith chart, it is possible to use partial knowledge of the performance of the circuits associated with a particular transmit configuration to extrapolate a more complete understanding that can be used to find transmit configurations with substantially identical performance characteristics. The disclosed subject matter applies deep learning, at least in some respects, to generative adversarial networks (GANs) to enable highly efficient image completion using extrapolated load-pull information. Such techniques provide a viable optimization solution for adaptive RF transmitters, and may also be suitable for quickly starting the initialization of other algorithms or accelerating the load-pull process by significantly reducing the number of evaluation impedances required to generate accurate load-pull contours, often requiring as few as six measurements.

[0014] In one aspect, the disclosed radio frequency (RF) device system includes a software-defined radio (SDR), an adaptive transmit amplifier, and a host computer communicatively coupled to the SDR and the adaptive transmit amplifier. The RF device itself may be a cognitive radar. The host computer includes a central processing unit and a computer-readable memory including processor-executable instructions that, when executed by the processor, cause the system to perform an optimization operation, the optimization operation including selecting an initial impedance as the load impedance of the RF device, and iteratively performing an image complement operation until a convergence criterion is met. The image complement operation may include obtaining a measured performance corresponding to the load impedance, storing the measured performance as a point on a measured load pull contour image, performing a load pull extrapolation to extrapolate a predicted optimal impedance from the load impedance, and saving the predicted impedance as the load impedance for the next iteration of the image complement operation. The convergence criterion may be met if the difference between the predicted impedance and the closest previously measured load impedance is less than a predetermined threshold. In an embodiment in which the impedance is represented as a pixel in the image, the convergence criterion may be met if the predicted impedance matches one of the previously measured impedances.

[0015] The load-pull extrapolation may be implemented as a gradient-based image completion process applied to a generative adversarial network (GAN) trained on a known load-pull contour. The load-pull contour image may correspond to an array of image pixels, with each pixel representing a corresponding load impedance. The image completion operation may include measuring the performance of the RF device with one or more additional impedances and adding a point corresponding to each of the one or more additional impedances to the load-pull contour image. The predicted impedance may correspond to a first image pixel, and the one or more additional impedances may correspond to one or more additional pixels, with the additional pixels selected at least in part based on their proximity to the first image pixel. For example, the one or more additional pixels may include one or more pixels selected from a group of pixels adjacent to and surrounding the first pixel.

[0016] The GAN may include a generative network and a discriminative network, and performing the load pull extrapolation may include generating a predicted load pull contour image using the generative network, determining a degree of agreement between points of the measured load pull contour image and corresponding points of the predicted load pull contour image, and searching for generative network inputs that generate a predicted load pull contour image that minimizes a loss metric. The GAN may be implemented as a Wasserstein GAN, and the discriminative network may include an evaluation network. In such an embodiment, the loss metric may include a contextual loss component indicative of a degree of similarity between points of the generated load pull contour image and the measured load pull contour image, and a perceptual loss component indicative of a degree of dissimilarity between the generated load pull contour image and the training dataset. [Brief description of the drawings]

[0017] [Figure 1] FIG. 1 is a graphical depiction of the determination of a complete load pull contour from partial contour information. [Diagram 2]FIG. 2 is a graphical representation of the generative and discriminative network functions of a GAN configured for use in combination with the disclosed image completion functions. [Diagram 3] FIG. 3 is a graphical representation of the generative and discriminative networks of FIG. 3 configured to perform image completion in accordance with the disclosed teachings. [Figure 4] FIG. 4 illustrates an example network topology for a load-pull identification network. [Diagram 5] FIG. 5 illustrates an example network topology of a load pull generation network. [Figure 6] FIG. 6 shows a flow diagram of the optimization method using load pull extrapolation. [Figure 7] FIG. 7 illustrates an exemplary radar system suitable for use in combination with the disclosed subject matter for real-time optimization. [Figure 8] FIG. 8 illustrates an exemplary host computer suitable for use in conjunction with the subject matter disclosed in FIGS. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0018] The drawings above and the following specific structural and functional description are not presented to limit the scope of applicant's invention or the scope of the appended claims. Rather, the drawings and description are provided to teach one of ordinary skill in the art how to make and use the invention for which patent protection is sought. Those skilled in the art will appreciate that, for the sake of clarity and understanding, not all features of the commercial embodiments of the present invention have been described or shown. Those skilled in the art will also appreciate that the development of an actual commercial embodiment incorporating aspects of the present disclosure will require numerous implementation-specific decisions to achieve the developer's end goal for the commercial embodiment. Such implementation-specific decisions may include, but are not limited to, compliance with system-related, business-related, government-related, and other constraints. These constraints may vary depending on the particular implementation, location, or time. While the developer's efforts may be complex and time-consuming in absolute terms, such efforts will nevertheless be routine for those skilled in the art having the benefit of the present disclosure.

[0019] It should be understood that the invention disclosed and taught herein is susceptible to numerous and various modifications and alternative forms. The use of singular terms, such as "a," is not intended to limit the number of items. Furthermore, various methods and embodiments of the system can be included in combination with each other to create variations of the disclosed methods and embodiments. Discussion of a singular element may include a plural element and vice versa. Reference to at least one item may include one or more items. Additionally, various aspects of the embodiments can be used in combination with each other to achieve the understood objectives of the disclosure.

[0020] Unless the context dictates otherwise, the term "comprise" or variations such as "comprises" or "comprising" should be understood to mean the inclusion of at least the recited elements or steps, groups of elements or steps, or equivalents thereof, and not the exclusion of a greater number or other elements or steps, groups of elements or steps, or equivalents thereof. Terms such as "coupled," "coupling," "coupler," and the like are used broadly herein and may include any method or device for fastening, joining, adhering, fastening, attaching, joining, inserting therein, forming on or within, communicating with, or otherwise associating one or more members, for example, mechanically, magnetically, electrically, chemically, operatively, directly or indirectly with an intermediate element, and may further include, but is not limited to, forming one functional member integrally with another functional member. Coupling may occur in any orientation, including rotational orientation.

[0021] The order of the disclosed steps can be performed in various orders unless specifically limited. The disclosed steps can be combined with other steps, combined with the steps described, or divided into multiple steps. Similarly, the disclosed elements are described functionally and can be embodied as separate components or combined into components having multiple functions. Some elements may be designated by device names for brevity, and will be understood to include systems of related components known to those skilled in the art, which may or may not be specifically described.

[0022] Structural elements may be disclosed through examples provided in the specification and drawings that perform various functions. These examples are not limited in shape, size, and description, but serve as exemplary structures that can be modified as known to those skilled in the art in light of the teachings contained herein. Thus, the use of the term "exemplary" is an adjective form of the noun "example," and similarly refers to exemplary structures, not necessarily preferred embodiments. Element numbers with letter suffixes such as "A," "B," etc., designate different elements within a group of similar elements with similar structure or function, while corresponding element numbers without such letters generally refer to one or more similar elements. Any element numbers in the claims that correspond to elements disclosed in this application are exemplary and not exclusive, as multiple embodiments may be disclosed that use various element numbers for similar elements.

[0023] SUMMARY: The provisional application contains subject matter disclosing various systems, techniques, and applications related to real-time optimization of cognitive radar transmitters. This subject matter spans topics such as determining spatial-spectral transmission constraints, real-time performance evaluation of RF devices using software-defined radios (SDRs), real-time optimization of cognitive radar transmit amplifiers, the impact of impedance tuning on range-Doppler processing, and partial load-pull extrapolation with deep image completion. While there is overlap in these topics, the present application is primarily directed to real-time optimization of adaptive RF transmit amplifiers and image completion using load-pull extrapolation.

[0024] As disclosed in the provisional application, a technique for optimizing the average performance of an RF transmit circuit involves observing each transmit configuration multiple times for each impedance value. This approach can require an undesirably large number of observations, especially when the spectral conditions are highly variable or complex. However, many transmit configurations produce performance contours with respect to impedance that are similar enough that they can be used interchangeably during a search with little impact on the search's ability to converge to an optimal solution. Grouping substantially equivalent transmit configurations reduces the number of transmit configurations and improves the efficiency of the optimization process. Unfortunately, the amount of configuration measurement data that must be obtained to confirm the valid equivalence of any two transmit configurations with the necessary certainty can also be burdensome. However, if comparison of transmit configurations across different impedances were possible, the benefits of grouping configurations would be greatly enhanced, as the amount of data required to reach a conclusion of equivalence during system operation would be greatly reduced.

[0025] Performance data for different configurations and impedances can be compared by extrapolating the performance of each configuration to a common impedance. Using this load pull extrapolation technique, a complete set of load pull contours can be derived from an incomplete data set. Figure 1 shows a complete set of load pull contours 102 plotted on a Smith chart 100 derived from an incomplete data set 101.

[0026] The objective depicted graphically in FIG. 1, i.e., deriving a perfect load-pull contour from an incomplete one, suggests an image completion problem, in which each load-pull contour is represented as a group of pixels in an image of a Smith chart that plots a particular value of a performance parameter, such as output power, gain, etc., as a function of normalized real and imaginary components of a set of load impedances that generate an applicable value of the performance parameter. In such an image, each pixel of the image corresponds to a particular impedance. In at least some embodiments, a deep learning image completion process is employed to achieve the depiction objective. In an exemplary implementation, a gradient-based deep learning image completion process is applied to a generative adversarial network (GAN) trained with load-pull contours of known amplifiers. The training data can be obtained from measurements or simulations of linear or nonlinear amplifiers. As shown in FIG. 2, the GAN 200 can include a discriminative network 201 and a generative network 202 trained with opposing objectives. The discriminative network 201 may be trained to recognize whether a given load pull contour image 203 represents actual load pull contours, i.e., those contained in the training data, or generated load pull contours, i.e., those generated by the generative network 202, while being careful not to overfit the training data. Meanwhile, the generative network 202 may be trained to generate generated load pull contour images 204 that the discriminative network 201 would classify as a set of valid load pull contours. In at least some embodiments, the discriminative network 201 and the generative network 202 are trained in an alternating manner, allowing both networks to adapt to weaknesses in the operation of the other.

[0027] Image completion may be performed in the manner illustrated in FIG. 3 by iteratively searching for inputs 301 to a generative network 202 that generate predicted contours 302 that closely match known contours 303 of the image to be completed using a properly trained GAN 200. Because the GAN 200 is pre-trained prior to image completion, this approach provides an excellent basis for quickly evaluating unknown devices. In at least some embodiments, the GAN 200 may be implemented as a Wasserstein GAN (WGAN), in which case a predetermined distance, called the Wasserstein distance, is used as a loss metric when training the generative network 202 and the discriminative network 201 (which may be referred to as the critic network in this context). Using the Wasserstein distance avoids metric saturation that can occur if one of the networks performs too well, e.g., the discriminative network 201 never makes a mistake. Avoiding this saturation greatly improves the stability and robustness of the network training process by preventing the gradient calculations used to update the generative network 202 from disappearing if the discriminative network 201 performs too well, and helps prevent the generator network 202 from learning only a few specific images that can fool the discriminative network 201.

[0028] FIG. 4 illustrates an example discrimination network topology 400 of the discrimination network 201, and FIG. 5 illustrates an example generative network topology 500 of the generative network 201. The illustrated discrimination network topology 400 operates on a 32x32 pixel load-pull image 401 and generates a single value, distance 402, that indicates the magnitude of the distance between the input image and the trained set of valid images. The illustrated discrimination network topology 402 includes an input layer 410, four 2D convolutional layers 420-1, 420-2, 420-3, 420-4, and three leaky rectified linear unit activation functions 422-1, 422-2, 422-3. The outputs of the 2D convolutional layers 420-2 and 420-3 are layer-normalized via layer normalization functions 422-1 and 422-2, respectively. The generative network topology 500 shown in Figure 5 operates on a 128-element initialization vector 501 to generate a synthesized 32x32 pixel grayscale load-pull image 502. An input layer 510 receives the initialization vector 501, and the topology shown includes four 2D convolutional transposition layers 520-1, 520-2, 520-3, and 520-4 and three leaky ReLUs 522-1, 522-2, and 522-3. The outputs of the 2D convolutional transposition layers 520-1, 520-2, and 520-3 are batch normalized via batch normalization functions 521-1, 521-2, and 521-3, respectively. A hyperbolic tangent activation layer 523 provides the final scaling of the network output.

[0029] In an exemplary implementation, training data is generated via simulated output power contours of 100,000 randomly generated sets of amplifier scattering parameters (S-parameters), which represent the linear characteristics of the amplifier. The simulated contours thus represent linear device performance. In some embodiments, it may be desirable to provide additional training data for large signal behavior by running load-pull simulations using existing nonlinear device models across a range of settings (frequency, bias conditions, input power, etc.). In other embodiments, linear training data may be sufficient.

[0030] As mentioned before, the goal of image completion is to find an input z to a generative network that generates an image G(z) that is similar to a known sub-image x and matches the overall target dataset. To complete an image, a mask M is defined to encode the portion of the complete image provided by x. The mask is specified as follows:

number

[0031] In an exemplary embodiment, two loss metrics are used to determine the quality of G(z), where z is the scalar image: contextual loss and perceptual loss. The contextual loss indicates the similarity between the generated image and the provided subimage, and is defined as follows:

number

[0032] The perceptual loss represents how similar the generated image is to the members of the dataset that were trained according to the evaluation network C() and is defined as:

number

[0033] The two losses can be combined with a hyperparameter λ, which weights the relative importance of the two metrics. In this disclosure, the value of λ is chosen to be 1. Thus, the total loss L(z) is:

number

[0034] With respect to Equation 4, the optimization goal is to find z with z that minimizes L(z) with z. This is a stochastic gradient-based optimization problem that can be solved using any of a variety of suitable algorithms that will be familiar to those skilled in the art. An exemplary algorithm for identifying a solution to Equation 4 may include elements described in "Adam: A Method for Stochastic Optimization" by DPK Kingma at the 3rd International Conference on Learning Representations (submitted on December 22, 2014, published on May 5, 2015). Each extrapolated load-pull contour set may be the result of a fixed number of iterations (e.g., 1000) or a dynamic number of iterations of the applicable algorithm. In the dynamic case, the number of iterations may be determined by the behavior of the loss function at the completed iterations.

[0035] 6, a flow diagram illustrates an optimization method 600 using iterative load pull extrapolation. The illustrated method 600 embodies a "max-add" approach, in which the best predicted impedance from each iteration of the load pull extrapolation is selected for measurement and its performance is added to a load pull contour data set representing the load pull contour image being constructed.

[0036] As shown in FIG. 6, the illustrated method 600 begins with an initial operation 602 in which an initial impedance is selected as the output impedance of the RF transmitter. The illustrated method then enters an iterative loop 604 beginning with a convergence check (operation 606) in which a particular convergence criterion is evaluated against a predefined threshold. In at least one embodiment, the convergence criterion evaluates whether a recently identified best predicted impedance matches an already measured impedance. If convergence is not reached, the illustrated method measures the performance of the RF device (operation 610) and stores the measured performance as points on a load pull contour image, referred to herein as a measured load pull contour image. The illustrated method 600 then next performs a load pull extrapolation (operation 612) to identify the best predicted impedance.

[0037] In at least some embodiments, it may be beneficial or desirable to perform one or more additional measurements within the loop 604 such that with each iteration of the loop 604, multiple data points are added to the measured load pull contour image being constructed. The illustrated method 600 includes determining (operation 614) whether to measure (operation 616) one or more additional impedances during the current iteration of the loop 604. Measuring the additional data points is likely to be more beneficial if the additional data points are associated with impedances that are identifiable with little or no appreciable delay and that are likely to exhibit similar performance contours as the currently applied impedance. In the case of a pixel-based load pull contour image, selecting the additional impedances based on pixel proximity can satisfy both of these conditions. As an illustrative example, an implementation may perform a total of five measurements during each iteration of the loop 604, the five measurements corresponding to the best predicted impedance and impedances corresponding to pixels immediately north, south, east, and west of the active pixel (i.e., the pixel corresponding to the current best predicted impedance). In other implementations, more, fewer, and / or different additional pixels may be selected to be measured.

[0038] If the convergence check (operation 606) indicates that the applicable convergence criteria are met, the method 600 shown in FIG. 6 applies the best measured impedance to the load-pull contour image (operation 620) and the method ends.

[0039] 7, a cognitive radar system 700 suitable for implementing the disclosed real-time optimization functionality is shown. The illustrated radar system 700 includes a software radio 701 and an adaptive transmit amplifier 705, both of which are communicatively coupled to a host computer 702. The adaptive transmit amplifier 705 includes an amplifier 701 and a configurable or variable load impedance 710. The software radio 701 may be implemented using any of a variety of suitable commercially available SDRs, such as Ettus Research's USRP X310. In at least one embodiment, the host computer 702 includes software instructions that, when executed, perform the real-time optimization described herein.

[0040] 8, one or more of the elements depicted in FIGS. 1-2 may be implemented as or within an information handling system, as illustrated in information handling system 800 depicted in FIG. 8. The depicted information handling system includes one or more general purpose processors or central processing units (CPUs) 801 communicatively coupled to memory resources 810 and an input / output hub 820 to which various I / O resources and / or components are communicatively coupled. I / O resources explicitly depicted in FIG. 8 include a network interface 840, commonly referred to as a network interface card (NIC), storage resources 830, and additional I / O devices, components, or resources 850, including, by way of non-limiting examples, a keyboard, mouse, display, printer, speaker, microphone, and the like. The depicted information handling system 800 includes a baseboard management controller (BMC) 860 that provides out-of-band management resources that may be coupled to a management server (not shown), among other functions and services. In at least some embodiments, the BMC 860 may manage the information handling system 800 even when the information handling system 800 is powered off or in a standby state. BMC 860 may include a processor, memory, an out-of-band network interface that is separate and physically separate from an in-band network interface of information processing system 800, and / or other embedded information processing resources. In certain embodiments, BMC 860 may include or be an integral part of a remote access controller (e.g., a Dell Remote Access Controller or an Integrated Dell Remote Access Controller) or a chassis management controller.

[0041] Other and further embodiments may be devised that utilize one or more aspects of the invention described above without departing from the disclosed invention as defined in the claims. For example, some of the steps and teachings may be combined or arranged in different sequences and other variations limited only by the claims.

[0042] The present invention is described in the context of preferred and other embodiments, and not all embodiments of the present invention are described. Those skilled in the art can make obvious modifications and changes to the described embodiments. The disclosed and undisclosed embodiments are not intended to limit or restrict the scope or applicability of the invention conceived by the applicant, but rather, in accordance with the patent laws, the applicant intends to fully protect all such changes and improvements that fall within the scope of the following claims.

Claims

1. 1. A method for optimizing a transmit configuration of an adaptive radio frequency (RF) device, the method comprising: selecting an initial impedance as a load impedance for the RF device; performing an iterative image completion operation until a convergence criterion is met; The image completion operation includes: measuring a performance of the RF device to obtain a measured performance corresponding to the load impedance; storing the measured performance as points on a measured load pull contour image; performing a load pull extrapolation to extrapolate a predicted optimal impedance from the load impedance; storing the predicted impedance as the load impedance for a next iteration of the image completion operation; and the convergence criterion is met when a difference between the predicted impedance and a closest previously measured load impedance is less than a predetermined threshold. method.

2. 2. The method of claim 1, wherein performing the load-pull extrapolation comprises applying a gradient-based image completion process to a generative adversarial network (GAN) trained on known load-pull contours.

3. The method of claim 2 , wherein the load-pull contour image comprises an array of image pixels, each pixel representing a corresponding load impedance.

4. The image complementing operation includes: Measuring the performance of the RF device at one or more additional impedances; adding a point corresponding to each of the one or more additional impedances to the load-pull contour image; The method of claim 3 , comprising:

5. 5. The method of claim 4, wherein the predicted impedance corresponds to a first image pixel and the one or more additional impedances correspond to one or more additional pixels, the pixels being selected based at least in part on their proximity to the first image pixel.

6. The method of claim 5 , wherein the one or more additional pixels include one or more pixels selected from a group of pixels adjacent to and surrounding the first pixel.

7. The GAN includes a generating network and an identifying network, and performing the load pull extrapolation includes: generating a predicted load-pull contour image with the generation network; determining a degree of correspondence between points of the measured load-pull contour image and corresponding points of the predicted load-pull contour image; searching for a generative network input that generates a predicted load-pull contour image that minimizes a loss metric; The method of claim 2 , comprising:

8. The GAN comprises a Wasserstein GAN, the discrimination network comprises an evaluation network, and the loss metric is a context loss component indicating the similarity between the generated load-pull contour image and the points of the measured load-pull contour image; a perceptual loss component indicative of the degree of dissimilarity between the generated load-pull contour image and a training dataset; The method of claim 7, comprising:

9. The method of claim 2 , wherein the GAN is trained only on contours corresponding to linear device operation.

10. The method of claim 2 , wherein the RF device comprises a cognitive radar device.

11. 1. A radio frequency (RF) device system, comprising: Software Defined Radio (SDR); an adaptive transmit amplifier; a host computer communicatively coupled to the SDR and the adaptive transmit amplifier; the host computer including a central processing unit and computer readable memory including processor executable instructions that, when executed, cause the system to perform an optimization operation, the optimization operation comprising: selecting an initial impedance as a load impedance for the RF device; performing an iterative image completion operation until a convergence criterion is met; The image completion operation includes: measuring a performance of the RF device to obtain a measured performance corresponding to the load impedance; storing the measured performance as points on a measured load pull contour image; performing a load pull extrapolation to extrapolate a predicted optimal impedance from the load impedance; storing the predicted impedance as the load impedance for a next iteration of the image completion operation; and the convergence criterion is met when a difference between the predicted impedance and a closest previously measured load impedance is less than a predetermined threshold. A radio frequency (RF) device system.

12. The RF device of claim 11 , wherein performing the load-pull extrapolation includes applying a gradient-based image completion process to a generative adversarial network (GAN) trained on known load-pull contours.

13. The RF device of claim 12 , wherein the load-pull contour image comprises an array of image pixels, each pixel representing a corresponding load impedance.

14. The image complementing operation includes: Measuring the performance of the RF device at one or more additional impedances; adding a point corresponding to each of the one or more additional impedances to the load-pull contour image; The RF device of claim 13 , comprising:

15. 15. The RF device of claim 14, wherein the predicted impedance corresponds to a first image pixel and the one or more additional impedances correspond to one or more additional pixels, the pixels being selected based at least in part on their proximity to the first image pixel.

16. 16. The RF device of claim 15, wherein the one or more additional pixels include one or more pixels selected from a group of pixels adjacent to and surrounding the first pixel.

17. The GAN includes a generating network and an identifying network, and performing the load pull extrapolation includes: generating a predicted load-pull contour image with the generation network; determining a degree of correspondence between points of the measured load-pull contour image and corresponding points of the predicted load-pull contour image; searching for a generative network input that generates a predicted load-pull contour image that minimizes a loss metric; The RF device of claim 12 , comprising:

18. The GAN comprises a Wasserstein GAN, the discrimination network comprises an evaluation network, and the loss metric is a context loss indicating the similarity between the generated load-pull contour image and the points of the measured load-pull contour image; a perceptual loss indicating the degree of dissimilarity between the generated load-pull contour image and a training dataset; 20. The RF device of claim 17, comprising:

19. The RF device of claim 12 , wherein the GAN is trained only with contours corresponding to linear device operation.

20. The RF device of claim 12 , wherein the RF device comprises a cognitive radar device.

Citation Information

Patent Citations

  • Method of characterizing nonlinear behavior of amplifiers using loadpull measurements

    WO2007081705A2