Method and apparatus for machine learning based radio frequency (RF) front end calibration

Abstract

Certain aspects of the present disclosure provide techniques and apparatuses for calibrating radio frequency (RF) circuitry using machine learning. One example method generally includes calibrating a first subset of RF circuitry calibration parameters. Values of a second subset of RF circuitry calibration parameters are predicted based on a machine learning model and the first subset of RF circuitry calibration parameters. The second subset of RF circuitry calibration parameters can be different from the first subset of RF circuitry calibration parameters. At least the first subset of RF circuitry calibration parameters is verified, and after the verification, at least the first subset of RF circuitry calibration parameters is written to a memory associated with the RF circuitry.

Description

[0001] Cross Reference to Related Applications

[0002] This patent application claims priority to U.S. Patent Application No. 17 / 650,334, filed February 8, 2022, which is expressly incorporated by reference herein in its entirety. BACKGROUND

[0003] Aspects of the present disclosure relate to radio frequency (RF) circuit calibration.

[0004] Radio frequency circuits generally allow for conversion of signaling to and from radio frequency bandwidths for transmission to or reception from other devices. These RF circuits are generally manufactured as a set of electronic circuits that form a complete unit. However, due to variations in manufacturing processes, the characteristics of each RF circuit can vary. Thus, to allow these RF circuits to perform at similar levels, each RF circuit can be individually calibrated, and parameters used to calibrate each RF circuit can be written to a memory associated with that RF circuit.

[0005] The complexity involved in calibrating RF circuits generally increases as these RF circuits are designed to support new functionality. For example, RF circuits can be calibrated for operation in various frequency bands (e.g., FR1 bands between 4.1 GHz and 7.125 GHz, FR2 bands between 24.25 GHz and 52.6 GHz, etc.), operation in multiple-input multiple-output (MIMO) systems using a single antenna or using multiple antennas, etc. For RF systems that implement MIMO techniques or allow for communication using millimeter wave frequencies, the number of uplinks and downlinks within the system can further increase the number of parameters to be calibrated. Thus, as RF circuits become more complex and as the number of configuration parameters for these RF circuits increases, the process of calibrating the RF circuits can become increasingly resource intensive.

[0006] Accordingly, there is a need for techniques for efficiently calibrating parameters of RF circuits. SUMMARY

[0007] Certain aspects provide a method for calibrating a radio frequency (RF) circuit. The method generally includes calibrating a first subset of RF circuit calibration parameters. Values of a second subset of RF circuit calibration parameters are predicted based on a machine learning model and the first subset of RF circuit calibration parameters. The second subset of RF circuit calibration parameters can be different from the first subset of RF circuit calibration parameters. At least the first subset of RF circuit calibration parameters is verified, and after the verification, at least the first subset of RF circuit calibration parameters is written to a memory associated with the RF circuit.

[0008] Other aspects provide a processing system configured to perform the aforementioned methods and those described herein, a non-transitory computer-readable medium comprising instructions that, when executed by one or more processors of a processing system, cause the processing system to perform the aforementioned methods and those described herein, a computer program product embodied on a computer-readable storage medium, the computer program product comprising code for performing the aforementioned methods and those further described herein, and an apparatus comprising means for performing the aforementioned methods and those further described herein.

[0009] The following description and associated drawings set forth certain illustrative features of the one or more implementations. BRIEF DESCRIPTION OF DRAWINGS

[0010] The accompanying drawings illustrate certain aspects of the one or more implementations, and thus are not to be considered limiting the scope of the disclosure.

[0011] FIG. 1 An example pipeline for calibrating radio frequency (RF) circuitry is depicted.

[0012] FIG. 2A-2B An example pipeline for calibrating RF circuitry using a machine learning model is illustrated in accordance with aspects of the present disclosure.

[0013] FIG. 3A-3B An example pipeline for calibrating RF circuitry using a machine learning model is illustrated in accordance with aspects of the present disclosure.

[0014] FIG. 4 An example pipeline for calibrating RF circuitry using a machine learning model is illustrated in accordance with aspects of the present disclosure.

[0015] FIG. 5 Example operations that can be performed by an RF circuitry calibration system to calibrate RF circuitry using a machine learning model are illustrated in accordance with aspects of the present disclosure.

[0016] FIG. 6 Example implementations of a processing system in which RF circuitry can be calibrated using a machine learning model are illustrated in accordance with aspects of the present disclosure.

[0017] To facilitate understanding, like reference numerals have been used, where possible, to designate like elements across the figures. It is contemplated that elements and features of one implementation can be beneficially incorporated into other implementations without further recitation. DETAILED DESCRIPTION

[0018] Aspects of the disclosure provide apparatuses, methods, processing systems, and computer readable media for calibrating radio frequency (RF) circuits using machine learning techniques.

[0019] RF circuits are subject to variability in manufacturing and operational parameters, which can affect operation of such integrated circuits. For example, variations in manufacturing, such as variations in etch depth, metal or oxide layer thickness, impurity concentration, etc., can affect the resistance, threshold voltage, etc. of individual circuits. Moreover, as RF circuits become more complex, additional parameters can be configured for each RF circuit. For example, the introduction of additional functionality to support MIMO techniques, higher frequency bands (e.g., FR1 or FR2), etc. can introduce new parameters for calibration to account for the increase in uplink and downlink within the RF circuit, and more generally to account for additional components within the RF circuit to enable support for the additional functionality.

[0020] Aspects of the disclosure provide techniques for calibrating RF circuits using at least some parameters predicted using machine learning techniques. Generally, the predicted parameters can be a subset of the full domain of parameters that an RF circuit can be configured with, and can be predicted based on a set of calibration parameters for the RF circuit. By calibrating a first set of calibration parameters for an RF circuit and using the first set to predict (e.g., using machine learning) a second set of calibration parameters for the RF circuit, computational resources for calibrating the RF circuit can be reduced relative to calibrating the full domain of RF calibration parameters for any given RF circuit.

[0021] Example radio frequency (RF) circuit calibration pipeline

[0022] FIG. 1 An example pipeline 100 for calibrating an RF circuit is depicted. As shown, the pipeline 100 includes an RF parameter calibrator 110, temporary storage 120, an automated test engine 125, a code verifier 130, and a field programmable read only memory (FPROM) 140.

[0023] An RF circuit being calibrated generally includes a number of circuits that allow upconversion of baseband or intermediate frequency signals to radio frequency signals for transmission and allow downconversion of received radio frequency signals to intermediate frequency signals or baseband frequency signals for processing. These circuits can include, for example, power amplifiers, low noise amplifiers, mixers, filters, frequency synthesizers, phase locked loops, etc. Generally, the number of components within an RF circuit can increase as the complexity of the RF circuit increases, for example to support additional frequency bands (e.g., for carrier aggregation), a larger number of antennas (e.g., for diversity reception), etc. As the number of components in the RF circuit increases, the number of circuit parameters to be calibrated can also increase.

[0024] For N number of parameters to be calibrated, the RF parameter calibrator 110 can test and calibrate each parameter and generate N calibration codes for validation. These calibration codes can include, for example, codes that adjust power usage parameters, control current bias within the RF circuit, measure and control voltage at the RF circuit, etc. Generally, as the total number of circuit parameters increases, the total number of calibrations performed by the RF parameter calibrator 110 can increase.

[0025] The N calibration codes can be written to the temporary memory 120, and the automated test engine 125 can access the N calibration codes stored in the temporary memory to evaluate the RF circuit being calibrated (e.g., by measuring a first pass of the system). Generally, the automated test engine 125 can perform various measurements of the RF circuit being calibrated (such as current consumption, voltage, etc.), and can determine whether additional calibration is to be performed. Further, the automated test engine 125 can record the calibration codes generated by the RF parameter calibrator 110 for future use.

[0026] The N calibration codes can also be provided to the code validator 130 for validation. In some aspects, the N calibration codes can be provided to the code validator 130 for validation based on instructions generated by the automated test engine 125 to complete calibration of the RF circuit. Generally, in validating these calibration codes, the code validator 130 can determine whether the RF circuit configured based on the N calibration codes meets a target set of performance parameters (e.g., by measuring a second pass of the system). If the target set of performance parameters is met, the code validator 130 can write the N calibration codes to the FPROM 140 associated with the RF circuit (also referred to as burning the calibration codes to the FPROM). The calibration codes written to the FPROM 140 can subsequently be used by the RF circuit during operation.

[0027] As discussed, generating and validating calibration codes for an RF circuit can be a resource intensive process. First, the number of calibration codes to be generated and validated can increase as the complexity of the RF circuit increases. Additionally, for each parameter, calibration can be a multi-step process of generating a calibration code, validating the calibration code, and refining the generated calibration code based on the output of the validation process. However, it can also be observed that some RF circuit parameters can change systematically (e.g., specific to a die and manufacturing parameters of that die), while other circuit parameters can change randomly. Because the parameters that change systematically can be predictable, aspects of the present disclosure can leverage this predictability to reduce the number of RF circuit calibration parameters generated and validated, while maintaining the performance of the RF circuit and minimizing or at least reducing yield loss from circuits that cannot be calibrated to comply with a defined set of performance parameters.

[0028] Example machine learning model based radio frequency (RF) circuit calibration pipeline

[0029] To take advantage of the predictability of some RF circuit calibration parameters, aspects of the present disclosure use a machine learning model to identify relationships between different circuit calibration parameters and generate a set of parameters whose calibration can be predicted with minimal or at least reduced yield loss. By doing so, a first set of RF circuit calibration parameters can be generated by a parameter calibrator, and the first set of RF circuit calibration parameters can be used to predict a second set of RF circuit calibration parameters. Because the first set of RF circuit calibration parameters can be a subset of the full domain of RF circuit calibration parameters, aspects of the present disclosure can reduce the computational resources used to calibrate RF circuits by reducing the number of calibration codes that are generated and verified using the calibration pipeline discussed above.

[0030] FIG. 2A and FIG. 2B Example pipelines 200A and 200B for off-chip calibration of RF circuits using a machine learning model are illustrated in accordance with aspects of the present disclosure. In these pipelines 200A and 200B, a code prediction machine learning model can predict values of a set of RF circuit calibration codes based on outputs of an off-chip automated test engine, and can output the predicted values of the set of RF circuit calibration codes for verification or directly write the predicted values of the set of RF circuit calibration codes to an FPROM or other memory associated with the RF circuit.

[0031] FIG. 2A An example pipeline 200A for predicting RF circuit calibration codes through a machine learning model is illustrated. As shown, for a set of N RF circuit parameters, an RF parameter calibrator 110 can be configured to generate calibration codes for N-M parameters. The M parameters can be parameters that can be predicted by a code prediction machine learning model 210, as discussed in further detail below, while the N-M parameters can be parameters that will be calibrated by the RF parameter calibrator 110 and used by the code prediction machine learning model 210 to predict calibration codes for the M parameters. With the RF circuit calibrated, the N-M calibration codes can be written to temporary memory 120 on the chip.

[0032] Off-chip, an automated test engine 125 can access the N-M calibration codes for evaluation. Further, the automated test engine 125 can provide the N-M calibration codes to a code prediction machine learning model 210, which can use at least a portion of the N-M calibration codes to predict values of the M calibration codes. As shown, the code prediction machine learning model 210 can write the M calibration codes to temporary memory 120.

[0033] After the code prediction machine learning model 210 generates M calibration codes and writes them to temporary memory 120, the code verifier 130 retrieves N calibration codes and verifies that these codes cause the RF circuit to perform according to a defined set of performance parameters. The N calibration codes retrieved from temporary memory 120 typically include a first subset of RF circuit calibration codes corresponding to the NM parameters calibrated by RF parameter calibrator 110, and a second subset of RF circuit calibration codes corresponding to the M parameters predicted by the code prediction machine learning model 210. After the code verifier 130 verifies that the N calibration codes result in performance that satisfies the defined set of performance parameters, the code verifier 130 may write the N calibration codes to FPROM 140.

[0034] In the example illustrated in pipeline 200A, the number of calibrations performed on the RF circuit parameters can be reduced by M. However, code verifier 130 can verify all N calibration codes (or at least a portion of the N calibration codes, including some of the M calibration codes) before writing the N calibration codes to FPROM 140.

[0035] In some respects, additional performance gains can be achieved by reducing the amount of calibration code that is verified by the code verifier (and thus refined via additional calibration). For example... FIG. 2B As shown, pipeline 200B allows M calibration codes predicted by code prediction machine learning model 210 to be directly written to FPROM 140 without verification by code verifier 130. In pipeline 200B, automated test engine 125 can receive NM calibration codes generated by RF parameter calibrator 110 and instruct code verifier 130 to verify these NM calibration codes. Simultaneously, code prediction machine learning model 210 can use the NM calibration codes to predict the values ​​of calibration codes for M circuit parameters not calibrated by RF parameter calibrator 110.

[0036] Therefore, in pipeline 200B, there can be two write paths for writing calibration codes to FPROM 140. The first write path from code verifier 130 can be used to write calibration codes to FPROM 140 for parameters that cannot (or are specified not to) be predicted using code prediction machine learning model 210, and therefore will be calibrated using RF parameter calibrator 110 and verified using code verifier 130. The second write path from code prediction machine learning model 210 can be used to write predicted calibration codes to FPROM 140 for those parameters that can be predicted (and are specified to be predicted) based on the NM calibration codes generated by RF parameter calibrator 110. Because M calibration codes can be directly written to FPROM 140 in pipeline 200A without being processed by code verifier 130, both the number of calibrations and verifications can be reduced by M, thus further reducing the computational resources used for calibrating the RF circuitry.

[0037] In some respects, RF circuit parameter calibration can be performed on-chip without using an off-chip automated test engine. FIG. 3A and FIG. 3B An example pipeline for on-chip calibration of RF circuits using a machine learning model is illustrated according to various aspects of this disclosure.

[0038] like FIG. 3A As shown in pipeline 300A, RF parameter calibrator 110 can receive N RF circuit parameters and generate NM calibration codes for a subset of calibration parameters that cannot (or are specified not to) be predicted using code prediction machine learning model 310. The NM calibration codes can be written to temporary memory 120 and provided to code prediction machine learning model 310. Code prediction machine learning model 310 can use the NM calibration codes as input to predict the values ​​of M parameters for which RF parameter calibrator 110 has not generated calibration codes. These M calibration codes can also be written to temporary memory 120, resulting in N calibration codes being stored in temporary memory 120. Similar to... FIG. 2A As illustrated in pipeline 200A, the code verifier 130 in pipeline 300A can thus verify N calibration codes, and once verified, write the N calibration codes to FPROM 140. Therefore, similar to pipeline 200A, the number of calibration operations performed in pipeline 300A can be reduced by M, while the number of verification operations can remain at N.

[0039] Similar to assembly line 200B, FIG. 3B The 300B pipeline further reduces computational resource usage in the calibration and verification of RF circuits. For example... FIG. 3BAs shown, the code prediction machine learning model 310 can use NM calibration codes to predict calibration codes for M parameters that have not been calibrated by the RF parameter calibrator 110. The M calibration codes can be directly written to the FPROM 140 (e.g., in the absence of verification by the code verifier 130). Simultaneously, the NM calibration codes can be verified by the code verifier 130, and once verified, they can be written to the FPROM 140. In pipeline 300B, directly writing the M calibration codes to the FPROM 140 reduces the number of calibration and verification operations by M.

[0040] In some respects, machine learning models can be used to predict calibration codes for a subset of RF circuit calibration parameters for direct use by the RF circuit. For example... FIG. 4 As shown in pipeline 400, NM calibration codes can be generated by RF parameter calibrator 110 for a subset of parameters that cannot (or are specified not to) be predicted by code prediction machine learning model 410. The NM calibration codes can be written to temporary memory 120 and verified by code verifier 130 (and may be refined based on the output of code verifier 130). The verified NM calibration codes can be written to FPROM 140.

[0041] For M parameters that can be predicted (and specified to be predicted) based on NM calibration codes, the code prediction machine learning model 410 can predict these codes while the RF circuit is operating. Therefore, M calibration codes can be output by combining the NM calibration codes written to FPROM 140 for calibrating the RF circuit, and the RF circuit can be configured with a total of N codes. Because M calibration codes can be generated while the RF circuit is operating, pipeline 400 can reduce the number of calibration operations and verification operations by M. Furthermore, because NM calibration codes are written to FPROM 140, the size of FPROM 140 can also be reduced by a factor of M.

[0042] Example machine learning model for predicting RF circuit calibration parameters

[0043] The machine learning model described in this paper can typically be trained to identify a subset of calibration parameters that can be predicted by yield loss below a threshold amount. That is, the calibration parameters predicted by the machine learning model described in this paper can produce calibration parameters that allow the RF circuit to operate according to a defined set of performance parameters without losing more than a threshold amount of circuitry due to the inability to operate according to the defined set of performance parameters. In general, the machine learning model can be used with the parameters shown in Figure 2 to... FIG. 4 The exemplary set of parameters corresponding to the parameters calibrated by the RF parameter calibrator 110 illustrated herein is used as input to generate a non-exemplary set of parameters.

[0044] In one example, a machine learning model can be trained based on pairwise correlations between parameters in a dataset of calibration parameters on a set of dies on which RF circuits are calibrated. These pairwise correlations can be clustered into multiple clusters. Within each cluster, one parameter from the exemplary parameter set can be associated with multiple parameters from non-exemplary parameter sets. Clustering can be optimized to generate a minimum number of clusters. While this example allows for the prediction of some parameters, values ​​of other parameters can be used to calibrate a minimum number of parameters while keeping yield loss below a threshold amount.

[0045] In another example, a machine learning model can be trained based on yield rate similarity clustering. When generating the training dataset, the percentage yield loss can be calculated for each parameter pair in the historical dataset of RF circuit calibration parameters. Generally, the yield loss can be calculated based on the yield loss experienced by calibrating the second parameter in the parameter pair using the value of the first parameter. Generally, yield loss similarity can indicate whether a yield loss below a threshold can be maintained by replacing the value of the second parameter with the value of the first parameter. Yield loss similarity can be clustered such that the centroid of each cluster corresponds to the values ​​shown in Figure 2 to... FIG. 4 The RF parameter calibrator 110 illustrated in the diagram calibrates exemplary parameters, while other parameters in the cluster may correspond to non-exemplary parameters that can be predicted by a machine learning model. The number of clusters can be minimized such that the yield loss of each cluster is less than a threshold.

[0046] To allow for an additional reduction in the number of RF circuit calibration parameters that can be predicted using machine learning models, an iterative process can be used to evaluate the RF circuit calibration parameters and identify those with yield losses below a threshold. Therefore, an exemplary set of parameters can be selected using iterative or "greedy" methods, and predicted calibration codes can be generated based on one or more regression models. For example, the regression model could be a lasso regression model in the loss function, which optimizes by removing or minimizing features that are less important to the predictions generated by the model. In this case, circuit calibration may converge to a local optimum rather than a global optimum, and some predictable circuit calibration parameters may not actually be predicted using these techniques.

[0047] In yet another example, the machine learning model used to identify and predict the values ​​of RF circuit calibration parameters could be a dropout gradient descent network. Dropout gradient descent networks typically allow the machine learning model to use information from each of the exemplary parameters in a set of exemplary parameters to predict the values ​​of non-exemplary parameters, and can minimize this set of exemplary parameters based on a sparsity criterion. To construct this dropout gradient descent network (which could be a neural network), an exemplary search space can be initially generated on a per-parameter basis. For each parameter, the values ​​of other parameters in the entire RF circuit calibration parameter domain can be predicted, where identity connections for the parameters are masked in the neural network. For example, a candidate set of exemplary parameters can be selected based on a linear regression model, and the candidate set of non-exemplary parameters can be a set of parameters that can be predicted with a yield loss less than a threshold.

[0048] The candidate set of non-exemplary parameters can be further refined based on a dropout probability metric associated with each parameter in the candidate set of non-exemplary parameters. To this end, the dropout probability can be initialized based on weights extracted from the linear regression model discussed above. Each candidate non-exemplary parameter p... ne It can be associated with the target parameter T e weight Related, where p ne ≠T e For each parameter, according to the equation: max(p e For each p, e∈T, select the maximum weight across all objective parameters T. ne The dropout probability can be initialized as the output of the softmax function of all parameters p.

[0049] The final set of exemplary parameters can be selected to predict the values ​​of non-exemplary parameters by identifying the parameters in the candidate set of non-exemplary parameters that are actually exemplary parameters to be calibrated. To this end, a variational mask can be applied to the input layer to optimize the parameterized dropout rate. The loss function can penalize the count of features not dropped from the candidate set of non-exemplary values. Parameters with a retention probability exceeding a threshold can be retained in the set of non-exemplary parameters, and parameters with a retention probability below a threshold can be included in the set of exemplary parameters. Finally, based on the learned set of non-exemplary parameters, a linear network can be trained to identify the learned set of exemplary parameters. This may result in a model that does not drop out on non-exemplary parameters, which will be used to predict the values ​​of these non-exemplary parameters.

[0050] In another example, a dropout gradient descent network can be constructed as a linear network. In doing so, a candidate set of predictable parameters can be generated based on gradient descent optimization of a linear regression function. The linear regression function can be a regression function of the weights associated with each parameter in the candidate set.

[0051] Example RF circuit calibration using machine learning models

[0052] FIG. 5 Examples are shown (for example, from Figure 2 to...) FIG. 4 The RF circuit calibration pipeline 200A-B, 300A-B or 400 illustrated in the figure performs example operation 500 to calibrate the RF circuit using a machine learning model.

[0053] As shown in the figure, operation 500 can begin at box 510, where a first set of RF circuit calibration parameters is calibrated. The first subset of RF circuit calibration parameters can be an exemplary set of parameters that can be used by a machine learning model to predict the values ​​of one or more parameters in a second set of RF circuit calibration parameters (also known as a non-exemplary parameter set).

[0054] At box 520, the values ​​of a second subset of RF circuit calibration parameters are predicted based on a machine learning model and a first subset of RF circuit calibration parameters. For example, the second subset of RF circuit calibration parameters may include RF circuit calibration parameters that can be predicted to result in a yield loss of less than a threshold.

[0055] In some respects, various techniques can be used to identify a second subset of RF circuit calibration parameters. For example, to identify a second subset of RF circuit calibration parameters, pairwise correlations between parameters in a historical dataset of RF circuit calibration parameters can be extracted. These pairwise correlations can be clustered into multiple clusters. Each of these clusters can typically be associated with multiple parameters in both the second subset and the first subset of RF circuit calibration parameters. Clustering can be performed to generate a minimum number of clusters.

[0056] In some aspects, a second subset identifying the RF circuit calibration parameters can be performed on a per-parameter-pair basis. For each parameter pair, including the first and second parameters from a historical dataset of RF circuit calibration parameters, yield loss can be calculated by calibrating the second parameter using the value of the first parameter. Yield loss similarity values ​​for corresponding parameter pairs can be generated based on the yield loss calculated for each corresponding parameter pair, and pairwise correlations between each corresponding parameter pair can be clustered based on the yield loss similarity values ​​for each corresponding parameter pair. Clustering can be performed such that a minimum number of clusters are generated.

[0057] In some respects, a second subset of identifying RF circuit calibration parameters may include iteratively evaluating RF circuit calibration parameters to identify those with yield loss below a threshold.

[0058] In some aspects, identifying a second subset of RF circuit calibration parameters may include identifying the second subset of RF circuit calibration parameters based on a dropout gradient descent network. For example, the dropout gradient descent network may include a neural network. In this case, a second subset of RF circuit calibration parameters may be identified on a per-parameter basis. For each parameter in the global domain of RF circuit calibration parameters, other parameters in the global domain of RF circuit calibration parameters are predicted, wherein identity connections for the corresponding parameters are masked in the neural network. A candidate set of predictable parameters is generated based on the predicted parameters having a yield loss less than a threshold, and the candidate set of predictable parameters is refined based on a dropout probability metric associated with each parameter in the candidate set of predictable parameters.

[0059] In some respects, to generate a candidate set of predictable parameters, weights can be assigned to each corresponding parameter in the candidate set based on weights extracted from the neural network. Generally, the weights associated with a corresponding parameter correspond to the influence of that parameter on each target parameter. The maximum weight for each corresponding parameter in the candidate set is identified across the set of target parameters associated with that parameter. The dropout probability value for each corresponding parameter is initialized based on a softmax function computed on the maximum weight of each corresponding parameter in the candidate set.

[0060] In some respects, in order to generate a candidate set of predictable parameters, the candidate set is generated based on gradient descent optimization of a linear regression function with weights associated with each corresponding parameter in the candidate set.

[0061] At box 530, verify at least a first subset of the RF circuit calibration parameters.

[0062] At block 540, after verifying a first subset of at least the RF circuit calibration parameters, the first subset of at least the RF circuit calibration parameters is written to the memory associated with the RF circuit (e.g., FPROM 140).

[0063] In some respects, a second subset of the RF circuit calibration parameters can also be verified. After verifying the second subset of the RF circuit calibration parameters, the second subset of the RF circuit calibration parameters can be written to the memory associated with the RF circuit.

[0064] In some respects, the RF circuit can be operated based on a second subset of RF circuit calibration parameters written to a memory associated with the RF circuit and at least a first subset of the RF circuit calibration parameters. The second subset of the RF circuit parameters can be used to operate the RF circuit directly without a verification phase, or can be verified before being used to operate the RF circuit.

[0065] Example processing system for calibrating RF circuits using machine learning models

[0066] FIG. 6Describes for (such as this article, for example, about) FIG. 5 The example processing system 600 described uses a machine learning model to calibrate RF circuits.

[0067] The processing system 600 includes a central processing unit (CPU) 602, which in some examples may be a multi-core CPU. Instructions executed at the CPU 602 may be loaded, for example, from program memory associated with the CPU 602, or from memory 624.

[0068] The processing system 600 also includes additional processing components tailored for specific functions, such as a graphics processing unit (GPU) 604, a digital signal processor (DSP) 606, a neural processing unit (NPU) 608, a multimedia processing unit 610, and a wireless connectivity component 612.

[0069] NPUs (such as the NPU 608) are typically dedicated circuits configured to implement all the necessary control and arithmetic logic for executing machine learning algorithms, such as those for processing artificial neural networks (ANNs), deep neural networks (DNNs), random forests (RFs), etc. NPUs are sometimes alternatively referred to as neural signal processors (NSPs), tensor processing units (TPUs), neural network processors (NNPs), intelligent processing units (IPUs), vision processing units (VPUs), or graphics processing units.

[0070] NPUs (such as the NPU 608) are configured to accelerate the execution of common machine learning tasks, such as image classification, machine translation, object detection, and various other predictive models. In some examples, multiple NPUs may be instantiated on a single chip (such as a system-on-a-chip (SoC)), while in other examples, an NPU may be part of a dedicated neural network accelerator.

[0071] An NPU can be optimized for training or inference, or in some cases configured to balance performance between the two. For an NPU capable of performing both training and inference, these two tasks can typically still be performed independently.

[0072] NPUs designed to accelerate training are typically configured to accelerate the optimization of new models. This involves taking an existing dataset (usually labeled or sublabeled), iterating over the dataset, and subsequently tuning model parameters (such as weights and biases) to improve model performance—a highly computationally intensive operation. Generally, optimization based on incorrect predictions involves propagating back through the layers of the model and determining gradients to reduce prediction errors.

[0073] NPUs designed to accelerate inference are typically configured to operate on the full model. Thus, such NPUs can be configured to take new data segments as input and rapidly process those segments through an already trained model to generate model outputs (e.g., inference).

[0074] In one specific implementation, the NPU 608 is part of one or more of the CPU 602, GPU 604, and / or DSP 606.

[0075] The processing system 600 may also include one or more input and / or output devices 622, such as screens, touch-sensitive surfaces (including touch-sensitive displays), physical buttons, speakers, microphones, etc.

[0076] In some examples, one or more processors in the processing system 600 may be based on the ARM or RISC-V instruction set.

[0077] The processing system 600 also includes a memory 624, which represents one or more static and / or dynamic memories, such as dynamic random access memory, flash-based static memory, etc. In this example, the memory 624 includes a computer-executable component that can be executed by one or more of the aforementioned processors of the processing system 600.

[0078] Specifically, in this example, memory 624 includes an RF circuit calibration component 624A, a parameter value prediction component 624B, a parameter verification component 624C, and a parameter writing component 624D. The depicted components, as well as other undepicted components, may be configured to perform various aspects of the methods described herein.

[0079] Example clauses

[0080] Clause 1: A method for calibrating a radio frequency (RF) circuit, comprising: calibrating a first subset of RF circuit calibration parameters; predicting values ​​of a second subset of RF circuit calibration parameters based on a machine learning model and the first subset of RF circuit calibration parameters, the second subset of RF circuit calibration parameters being different from the first subset of RF circuit calibration parameters; verifying at least the first subset of RF circuit calibration parameters; and, after said verification, writing at least the first subset of RF circuit calibration parameters into a memory associated with the RF circuit.

[0081] Clause 2: The method described in Clause 1 further includes identifying a second subset of the RF circuit calibration parameters.

[0082] Clause 3: The method according to Clause 2, wherein identifying the second subset of the RF circuit calibration parameters comprises: extracting pairwise correlations between parameters in a historical dataset of the RF circuit calibration parameters; and clustering the pairwise correlations into a plurality of clusters, each of the plurality of clusters being associated with a plurality of parameters in the second subset of the RF circuit calibration parameters and parameters in the first subset of the RF circuit calibration parameters.

[0083] Clause 4: The method according to Clause 3, wherein clustering the pairwise correlations includes generating a minimum number of clusters.

[0084] Clause 5: The method according to any one of Clauses 2 to 4, wherein identifying the second subset of the RF circuit calibration parameters comprises: for each corresponding parameter pair in a historical dataset including the first parameter and the second parameter: calculating the yield loss resulting from calibrating the second parameter by using the value of the first parameter, and generating a yield loss similarity value for the corresponding parameter pair based on the yield loss calculated for the corresponding parameter pair; and clustering the pairwise correlations between each corresponding parameter pair based on the yield loss similarity value for each corresponding parameter pair.

[0085] Clause 6: The method according to Clause 5, wherein clustering the pairwise correlations includes generating a minimum number of clusters such that the yield loss of each cluster is less than a threshold.

[0086] Clause 7: The method according to any one of Clauses 2 to 6, wherein identifying the second subset of the RF circuit calibration parameters includes iteratively evaluating the RF circuit calibration parameters to identify parameters having a yield loss of less than a threshold.

[0087] Clause 8: The method according to any one of Clauses 2 to 7, wherein identifying the second subset of the RF circuit calibration parameters includes identifying the second subset of the RF circuit calibration parameters based on a dropout gradient descent network.

[0088] Clause 9: The method according to Clause 8, wherein: the dropout gradient descent network comprises a neural network; and identifying a second subset of the RF circuit calibration parameters comprises, for each corresponding parameter in the global domain of the RF circuit calibration parameters: predicting other parameters in the global domain of the RF circuit calibration parameters, wherein identity connections for the corresponding parameters are masked in the neural network; generating a candidate set of predictable parameters based on predicted parameters having a yield loss less than a threshold; and refining the candidate set of predictable parameters based on a dropout probability metric associated with each parameter in the candidate set of predictable parameters.

[0089] Clause 10: The method according to Clause 9, wherein generating a candidate set of the predictable parameters comprises: assigning weights to each corresponding parameter in the candidate set based on weights extracted from the neural network, wherein the weights of the corresponding parameters correspond to the influence of the corresponding parameter on each target parameter; identifying the maximum weight of each corresponding parameter in the candidate set across a set of target parameters associated with the corresponding parameter; and initializing a dropout probability value for each corresponding parameter based on a softmax function computed on the maximum weight of each corresponding parameter in the candidate set.

[0090] Clause 11: The method according to Clause 9 or 10, wherein generating the candidate set of the predictable parameters comprises generating the candidate set based on gradient descent optimization of a linear regression function with weights associated with each corresponding parameter in the candidate set.

[0091] Clause 12: The method according to any one of Clauses 1 to 11, wherein the second subset of the RF circuit calibration parameters includes RF circuit calibration parameters that can be predicted with a yield loss of less than a threshold.

[0092] Clause 13: The method according to any one of Clauses 1 to 12 further includes verifying a second subset of the RF circuit calibration parameters.

[0093] Clause 14: The method according to Clause 13 further includes, after verifying the second subset of the RF circuit calibration parameters, writing the second subset of the RF circuit calibration parameters into the memory associated with the RF circuit.

[0094] Clause 15: The method according to any one of Clauses 1 to 14 further includes operating the RF circuit based on a second subset of the RF circuit calibration parameters written to the memory associated with the RF circuit and a first subset of the at least the RF circuit calibration parameters.

[0095] Clause 16: An apparatus comprising: a memory storing executable instructions; and a processor configured to execute the executable instructions to cause the apparatus to perform a method according to any one of Clauses 1 to 15.

[0096] Clause 16: An apparatus comprising: a component for performing the method according to any one of Clauses 1 to 15.

[0097] Clause 17: A non-transitory computer-readable medium having instructions stored thereon, which, when executed by a processor, cause the processor to perform the method according to any one of Clauses 1 to 15.

[0098] Clause 18: A computer program product embodied on a computer-readable storage medium, comprising code for performing a method pursuant to any one of Clauses 1 to 15.

[0099] Additional considerations

[0100] The foregoing description is provided to enable any person skilled in the art to practice the various embodiments described herein. The examples discussed herein are not intended to limit the scope, applicability, or embodiments set forth in the claims. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be applied to other embodiments. For example, the function and arrangement of the elements discussed may be changed without departing from the scope of this disclosure. Various procedures or components may be omitted, substituted, or added as appropriate in the examples. For example, the described methods may be performed in a different order than described, and various steps may be added, omitted, or combined. Furthermore, features described for some examples may be combined in others. For example, any number of aspects set forth herein may be used to implement an apparatus or practice. In addition, the scope of this disclosure is intended to cover such apparatuses or methods practiced using structures, functionalities, or structures and functionalities other than or different from the aspects of this disclosure set forth herein. It should be understood that any aspect of the disclosure herein may be embodied by one or more elements of the claims.

[0101] As used herein, the term “exemplary” means “serving as an example, instance, or illustration.” Any aspect described herein as “exemplary” is not necessarily to be construed as preferred or superior to other aspects.

[0102] As used in this article, the phrase “at least one of” in a list of items refers to any combination of these items (including a single member). As an example, “at least one of a, b, or c” is intended to cover a, b, c, ab, ac, bc, and abc, as well as any combination with multiple identical elements (e.g., aa, aaa, aab, aac, abb, acc, bb, bbb, bbb, bbc, cc, and ccc, or any other ordering of a, b, and c).

[0103] As used herein, the term "determine" encompasses a wide variety of actions. For example, "determine" can include calculation, computation, processing, derivation, research, searching (e.g., looking in a table, database, or other data structure), ascertaining, and similar actions. Furthermore, "determine" can include receiving (e.g., receiving information), accessing (e.g., accessing data in memory), etc. Additionally, "determine" can include parsing, selecting, choosing, building, etc.

[0104] The methods disclosed herein include one or more steps or actions for implementing the methods. The steps and / or actions of the methods may be interchanged without departing from the scope of the claims. In other words, unless a specific order of steps or actions is specified, the order and / or use of a particular step and / or action may be modified without departing from the scope of the claims. Furthermore, the various operations of the methods described above can be performed by any suitable component capable of performing the corresponding function. This component may include various hardware and / or software components and / or modules, including but not limited to circuits, application-specific integrated circuits (ASICs), or processors. Generally, where operations illustrated in the figures are present, these operations may have corresponding components with similar numbering plus functional components.

[0105] Example components for calibration may include, for example, those shown in Figures 2 to 3. FIG. 4 The RF parameter calibrator 110 illustrated herein. Example components for prediction may include, for example, those shown in Figures 2 to 110. FIG. 4 The code illustrated herein predicts machine learning models 210, 310, or 410, etc. Example components for verification may include, for example, those shown in Figures 2 to 410. FIG. 4 The code validator 130, etc., is illustrated in the figure. Example components for writing may include, for example, those shown in Figures 2 to... FIG. 4 FIG. 4 The code validator 130, code prediction machine learning model 210, 310 or 410 and / or FPROM 140, etc., are illustrated in the example.

[0106] The following claims are not intended to be limited to the embodiments shown herein, but should be granted the full scope consistent with the language of the claims. Within the claims, unless otherwise specified, references to elements in the singular form are not intended to mean “one and only one”, but rather “one or more”. Unless otherwise specified, the term “some” means one or more. No element of any claim should be construed in accordance with 35 U.S.SC §112(f) unless the element is expressly referred to by the phrase “component for…” or, in the case of a method claim, by the phrase “step for…”. All structural and functional equivalents of elements of the aspects described herein that are known to or will be known hereafter by a person skilled in the art are expressly incorporated herein by reference and are intended to be covered by the claims. Furthermore, nothing disclosed herein is intended for public disclosure, whether or not such disclosure is expressly referred to in the claims.

Claims

1. A method for calibrating radio frequency (RF) circuits, comprising: The first subset of calibration parameters for RF circuits; The values ​​of a second subset of RF circuit calibration parameters are predicted based on a machine learning model and a first subset of the RF circuit calibration parameters, the second subset of the RF circuit calibration parameters being different from the first subset of the RF circuit calibration parameters, and the second subset of the RF circuit calibration parameters including RF circuit calibration parameters with predictable yield loss of less than a threshold. Verify at least a first subset of the RF circuit calibration parameters; as well as Following the verification, at least a first subset of the RF circuit calibration parameters are written into the memory associated with the RF circuit.

2. The method of claim 1, further comprising identifying a second subset of the RF circuit calibration parameters.

3. The method of claim 2, wherein the second subset identifying the RF circuit calibration parameters comprises: Extract pairwise correlations between parameters in the historical dataset of RF circuit calibration parameters; as well as The pairwise correlations are clustered into multiple clusters, each of which is associated with multiple parameters in a second subset of the RF circuit calibration parameters and parameters in a first subset of the RF circuit calibration parameters.

4. The method of claim 3, wherein clustering the pairwise correlations comprises generating a minimum number of clusters.

5. The method of claim 2, wherein the second subset identifying the RF circuit calibration parameters comprises: For each corresponding parameter pair of the first and second parameters in the historical dataset including RF circuit calibration parameters: Calculate the yield loss caused by calibrating the second parameter using the value of the first parameter; as well as A yield loss similarity value for the corresponding parameter pair is generated based on the yield loss calculated for the corresponding parameter pair. as well as Based on the yield loss similarity value of each corresponding parameter pair, the pairwise correlations between each corresponding parameter pair are clustered.

6. The method of claim 5, wherein clustering the pairwise correlations comprises generating a minimum number of clusters such that the yield loss of each cluster is less than a threshold.

7. The method of claim 2, wherein identifying the second subset of the RF circuit calibration parameters comprises iteratively evaluating the RF circuit calibration parameters to identify parameters having a yield loss of less than a threshold.

8. The method of claim 2, wherein identifying the second subset of the RF circuit calibration parameters comprises identifying the second subset of the RF circuit calibration parameters based on a dropout gradient descent network.

9. The method according to claim 8, wherein: The discard gradient descent network includes a neural network; and The second subset identifying the RF circuit calibration parameters includes each corresponding parameter across the entire RF circuit calibration parameter domain: Predict other parameters in the global domain of the RF circuit calibration parameters, wherein identity connections for the corresponding parameters are masked in the neural network; A candidate set of predictable parameters is generated based on the predicted parameters having yield loss less than a threshold; and The candidate set of predictable parameters is refined based on a dropout probability metric associated with each parameter in the candidate set of predictable parameters.

10. The method of claim 9, wherein generating the candidate set of the predictable parameters comprises: Weights are assigned to each corresponding parameter in the candidate set based on weights extracted from the neural network, wherein the weight of the corresponding parameter corresponds to the influence of the corresponding parameter on each target parameter; The maximum weight of each corresponding parameter in the candidate set is identified across the set of target parameters associated with the corresponding parameter. as well as The dropout probability value for each corresponding parameter is initialized based on the softmax function calculated on the maximum weight of each corresponding parameter in the candidate set.

11. The method of claim 9, wherein generating the candidate set of predictable parameters comprises generating the candidate set based on gradient descent optimization of a linear regression function with weights associated with each corresponding parameter in the candidate set.

12. The method of claim 1, further comprising verifying a second subset of the RF circuit calibration parameters.

13. The method of claim 12, further comprising, after verifying a second subset of the RF circuit calibration parameters, writing the second subset of the RF circuit calibration parameters into the memory associated with the RF circuit.

14. The method of claim 1, further comprising operating the RF circuit based on a second subset of the RF circuit calibration parameters written to the memory associated with the RF circuit and a first subset of the at least the RF circuit calibration parameters.

15. An apparatus for wireless communication, comprising: A radio frequency (RF) circuit having a first memory; A second memory, on which executable instructions are stored; and Processor, the processor being configured to execute the executable instructions in order to: The first subset of calibration parameters for RF circuits; The values ​​of a second subset of RF circuit calibration parameters are predicted based on a machine learning model and a first subset of the RF circuit calibration parameters, the second subset of the RF circuit calibration parameters being different from the first subset of the RF circuit calibration parameters, and the second subset of the RF circuit calibration parameters including RF circuit calibration parameters with predictable yield loss of less than a threshold. Verify at least a first subset of the RF circuit calibration parameters; as well as After the verification, at least a first subset of the RF circuit calibration parameters are written into the first memory.

16. The apparatus of claim 15, wherein the processor is further configured to identify a second subset of the RF circuit calibration parameters.

17. The apparatus of claim 16, wherein, in order to identify a second subset of the RF circuit calibration parameters, the processor is configured to: Extract pairwise correlations between parameters in the historical dataset of RF circuit calibration parameters; and The pairwise correlations are clustered into multiple clusters, each of which is associated with multiple parameters in a second subset of the RF circuit calibration parameters and parameters in a first subset of the RF circuit calibration parameters.

18. The apparatus of claim 16, wherein, in order to identify a second subset of the RF circuit calibration parameters, the processor is configured to: For each corresponding parameter pair of the first and second parameters in the historical dataset including RF circuit calibration parameters: Calculate the yield loss caused by calibrating the second parameter using the value of the first parameter; as well as A yield loss similarity value for the corresponding parameter pair is generated based on the yield loss calculated for the corresponding parameter pair. as well as Based on the yield loss similarity value of each corresponding parameter pair, the pairwise correlations between each corresponding parameter pair are clustered.

19. The apparatus of claim 18, wherein, in order to cluster the pairwise correlations, the processor is configured to generate a minimum number of clusters such that the yield loss of each cluster is less than a threshold.

20. The apparatus of claim 16, wherein, in order to identify a second subset of the RF circuit calibration parameters, the processor is configured to iteratively evaluate the RF circuit calibration parameters to identify parameters having a yield loss of less than a threshold.

21. The apparatus of claim 16, wherein, in order to identify a second subset of the RF circuit calibration parameters, the processor is configured to identify the second subset of the RF circuit calibration parameters based on a drop-down gradient descent network.

22. The apparatus according to claim 21, wherein: The discard gradient descent network includes a neural network; and To identify the second subset of the RF circuit calibration parameters, the processor is configured for each corresponding parameter across the entire RF circuit calibration parameter domain: Predict other parameters in the global domain of the RF circuit calibration parameters, wherein identity connections for the corresponding parameters are masked in the neural network; A candidate set of predictable parameters is generated based on the predicted parameters having yield loss less than a threshold; and The candidate set of predictable parameters is refined based on a dropout probability metric associated with each parameter in the candidate set of predictable parameters.

23. The apparatus of claim 22, wherein, in order to generate a candidate set of the predictable parameters, the processor is configured to: Weights are assigned to each corresponding parameter in the candidate set based on weights extracted from the neural network, wherein the weight of the corresponding parameter corresponds to the influence of the corresponding parameter on each target parameter; The maximum weight of each corresponding parameter in the candidate set is identified across the set of target parameters associated with the corresponding parameter. as well as The dropout probability value for each corresponding parameter is initialized based on the softmax function calculated on the maximum weight of each corresponding parameter in the candidate set.

24. The apparatus of claim 22, wherein, in order to generate the candidate set of the predictable parameters, the processor is configured to generate the candidate set based on gradient descent optimization of a linear regression function with weights associated with each corresponding parameter in the candidate set.

25. The apparatus of claim 15, wherein the processor is further configured to: Verify a second subset of the RF circuit calibration parameters; and After verifying the second subset of the RF circuit calibration parameters, the second subset of the RF circuit calibration parameters is written into the first memory.

26. The apparatus of claim 15, wherein the processor is further configured to operate the RF circuit based on a second subset of the RF circuit calibration parameters written to the first memory and a first subset of the at least the RF circuit calibration parameters.

27. An apparatus for calibrating radio frequency (RF) circuits, comprising: Components used to calibrate the first subset of calibration parameters for RF circuits; A component for predicting the value of a second subset of RF circuit calibration parameters based on a machine learning model and a first subset of the RF circuit calibration parameters, the second subset of the RF circuit calibration parameters being different from the first subset of the RF circuit calibration parameters, and the second subset of the RF circuit calibration parameters including RF circuit calibration parameters with predictable yield loss of less than a threshold. Components used to verify at least a first subset of the RF circuit calibration parameters; and A component for writing at least a first subset of the RF circuit calibration parameters into a memory associated with the RF circuit after the verification.

28. A non-transitory computer-readable medium having instructions stored thereon, the instructions, when executed by a processor, causing the processor to perform operations for calibrating radio frequency (RF) circuitry, the operations including: The first subset of calibration parameters for RF circuits; The values ​​of a second subset of RF circuit calibration parameters are predicted based on a machine learning model and a first subset of the RF circuit calibration parameters, the second subset of the RF circuit calibration parameters being different from the first subset of the RF circuit calibration parameters, and the second subset of the RF circuit calibration parameters including RF circuit calibration parameters with predictable yield loss of less than a threshold. Verify at least a first subset of the RF circuit calibration parameters; as well as Following the verification, at least a first subset of the RF circuit calibration parameters are written into the memory associated with the RF circuit.

Images