Modeling Method, Device, Equipment and Medium for Wideband Nonlinear Characteristics of Radio Frequency Channels

Abstract

The present invention provides a method, device, equipment and medium for modeling broadband nonlinear characteristics of a radio frequency channel, wherein the method includes: obtaining an input signal and an output signal to be modeled, detecting the bandwidth type of the input signal and / or the output signal, obtaining multiple sets of model parameters based on the bandwidth type, and inputting the multiple sets of model parameters into a preset model respectively, thereby obtaining multiple pre-trained models, selecting a preset number of pre-trained models as target pre-trained models, training each target pre-trained model, obtaining a nonlinear characteristic measurement model after training, and selecting the nonlinear characteristic measurement model with the smallest second error value as the target nonlinear characteristic measurement model. The beneficial effects of the present invention are: greatly improving the speed and accuracy of parameter identification by model iteration. Compared with re-modeling using traditional methods, the present application greatly improves the speed of model parameter identification while ensuring model accuracy.

Description

Technical Field

[0001] The present invention relates to the field of electronic information technology, and in particular to a method, device, equipment and medium for modeling broadband nonlinear characteristics of a radio frequency channel. Background Art

[0002] RF devices are core components of modern electronic information systems and play a critical role in the evolution of these systems. Currently, RF devices are widely used in various electronic information systems, including communications, radar, and navigation. In a typical communication system architecture, RF modules perform key functions such as electromagnetic wave signal generation, modulation and demodulation, power amplification, and filtering and noise reduction. Their performance parameters directly determine the transmission quality of the entire information link. It can be said that the performance of RF devices is directly related to the stability and reliability of electronic information systems.

[0003] However, with the rapid development and widespread adoption of electronic information systems, the signal power and bandwidth that RF devices must process continue to increase, making nonlinear effects unavoidable. These effects, stemming from the high-power, wide-bandwidth signals entering RF devices, can severely impact system performance. For example, in satellite navigation anti-interference systems, if RF devices exhibit nonlinear effects, existing adaptive filtering algorithms will be unable to properly process interfering signals, significantly degrading the system's anti-interference performance.

[0004] Establishing nonlinear models for RF devices is the primary approach to addressing nonlinear effects. Model structure and parameter extraction methods are two key factors influencing modeling effectiveness. Model structure determines the upper limit of nonlinear models. Numerous nonlinear models exist, ranging from the earliest Volterra series to the pruned MP model, GMP model, and DDR model. Machine learning nonlinear models offer unique advantages in modeling nonlinear systems and extracting model parameters. Nonlinear models based on technologies such as RNNs, LSTMs, and CNNs have achieved significant progress in both modeling accuracy and computational complexity.

[0005] The method of extracting model parameters determines the lower limit of nonlinear models. The essence of parameter extraction is to obtain parameters that minimize the difference between the estimated value and the actual value. There are many parameter extraction methods based on nonlinear models, which can be roughly divided into analytical methods and approximation methods. Analytical methods, represented by the least squares method, can obtain accurate model parameters, but they have problems with high complexity and slow response when the model parameters are large. Approximation methods, represented by the minimum mean square error method, can extract model parameters in a relatively short time, but are prone to falling into local optimality, and the model accuracy is not as good as analytical methods. Summary of the Invention

[0006] The main purpose of the present invention is to provide a method, device, equipment and medium for modeling broadband nonlinear characteristics of radio frequency channels, aiming to solve the problem of poor speed of existing nonlinear characteristic modeling.

[0007] The present invention provides a method for modeling broadband nonlinear characteristics of a radio frequency channel, comprising:

[0008] Obtain the input signal and output signal to be modeled;

[0009] detecting a bandwidth type of the input signal and / or the output signal;

[0010] Acquire multiple sets of model parameters based on the bandwidth type, and input the multiple sets of model parameters into a preset model respectively, thereby obtaining multiple pre-trained models; wherein the preset model is a nonlinear model;

[0011] Inputting the input signal into each of the pre-trained models respectively to obtain a corresponding prediction signal, and calculating a first error value based on the output signal and the prediction signal;

[0012] According to the values of the first error values, a preset number of pre-training models are selected in order from small to large and recorded as target pre-training models;

[0013] Acquire a training data set of the same type based on the bandwidth type; wherein a set of training data in the training data set includes a first input signal and a first output signal;

[0014] Inputting the training data set into each of the target pre-training models for training, and obtaining corresponding nonlinear characteristic measurement models after training;

[0015] The second error value of each of the nonlinear characteristic measurement models is obtained, and the nonlinear characteristic measurement model with the smallest second error value is selected as the target nonlinear characteristic measurement model.

[0016] Furthermore, the step of obtaining the second error value of each of the nonlinear characteristic measurement models includes:

[0017] Acquire a test data set of the same type as the bandwidth type; wherein a set of test data in the test data set includes a test input signal and a test actual output signal;

[0018] Inputting a test input signal in the test data set into each of the nonlinear characteristic measurement models to obtain a test prediction output signal;

[0019] According to the formula Calculate the second error value; wherein NMSE is the second error value, N is the number of test data groups in the test data set, The actual output signal for the nth test, Predict the output signal for the nth test.

[0020] Furthermore, the step of obtaining the second error value of each of the nonlinear characteristic measurement models includes:

[0021] Acquire a test data set of the same type as the bandwidth type; wherein a set of test data in the test data set includes a test input signal and a test actual output signal;

[0022] Inputting a test input signal in the test data set into each of the nonlinear characteristic measurement models to obtain a test prediction output signal;

[0023] According to the formula Calculate the second error value; wherein, NMSE ’ is the second error value, N is the number of test data sets in the test data set, The actual output signal for the nth test, Predict the output signal for the nth test, .

[0024] Furthermore, the nonlinear model is a time domain model or a frequency domain model.

[0025] Furthermore, the nonlinear model is a generalized method of moments estimation model.

[0026] Furthermore, the step of inputting the training data set into each of the target pre-training models for training, and obtaining corresponding nonlinear characteristic measurement models after training, respectively, includes:

[0027] The training data set is input into each of the target pre-training models in sequence, and the model parameters in each of the target pre-training models are updated by the least mean square method during each training, so as to obtain corresponding nonlinear characteristic measurement models after the training is completed.

[0028] Furthermore, the step of obtaining multiple groups of model parameters based on the bandwidth type includes:

[0029] Inputting the input signal and / or the output signal into a preset time domain model to obtain a corresponding frequency spectrum;

[0030] Multiple sets of model parameters are matched based on the frequency spectrum.

[0031] The present invention also provides a device for modeling broadband nonlinear characteristics of a radio frequency channel, comprising:

[0032] A first acquisition module is used to obtain input signals and output signals to be modeled;

[0033] a detection module, configured to detect a bandwidth type of the input signal and / or the output signal;

[0034] a second acquisition module, configured to acquire multiple sets of model parameters based on the bandwidth type, and input the multiple sets of model parameters into a preset model, thereby obtaining multiple pre-trained models; wherein the preset model is a nonlinear model;

[0035] A first input module, configured to input the input signal into each of the pre-trained models, obtain a corresponding prediction signal, and calculate a first error value based on the output signal and the prediction signal;

[0036] A selection module, configured to select a preset number of pre-trained models as target pre-trained models in ascending order according to the values of the first error values;

[0037] A third acquisition module is configured to acquire a training data set of the same type based on the bandwidth type; wherein a set of training data in the training data set includes a first input signal and a first output signal;

[0038] A second input module is used to input the training data set into each of the target pre-training models for training, and respectively obtain a nonlinear characteristic measurement model after the training is completed;

[0039] The fourth acquisition module is configured to acquire the second error value of each of the nonlinear characteristic measurement models, and select the nonlinear characteristic measurement model with the smallest second error value as the target nonlinear characteristic measurement model.

[0040] The present invention also provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of any one of the above methods when executing the computer program.

[0041] The present invention also provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps of any of the above-mentioned methods are implemented.

[0042] The beneficial effects of the present invention are as follows: by obtaining the input signal and output signal to be modeled, detecting the bandwidth type of the input signal and / or the output signal, obtaining multiple sets of model parameters based on the bandwidth type, and inputting the multiple sets of model parameters into the preset model respectively, correspondingly obtaining multiple pre-trained models, selecting a preset number of pre-trained models as target pre-trained models, training each target pre-trained model, obtaining a nonlinear characteristic measurement model after training, selecting the nonlinear characteristic measurement model with the smallest second error value as the target nonlinear characteristic measurement model, thereby greatly improving the speed and accuracy of parameter identification by model iteration. Compared with re-modeling using traditional methods, the present application greatly improves the speed of model parameter identification while ensuring model accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 Schematic diagram of the modeling strategy of the time domain nonlinear model and the frequency domain nonlinear model according to an embodiment of the present invention; wherein, Figure 1 (a) is a schematic diagram of the time domain nonlinear model modeling strategy. Figure 1 (b) is a schematic diagram of the frequency domain nonlinear model modeling strategy;

[0044] Figure 2 This is a schematic diagram of the deviation between the local optimum and the global optimum of a system according to an embodiment of the present invention;

[0045] Figure 3 This is a general schematic diagram of a parameter reuse multi-starting point search algorithm according to an embodiment of the present invention;

[0046] Figure 4 Schematic diagram of the convergence process of a multi-starting point search algorithm according to an embodiment of the present invention;

[0047] Figure 5 This is a schematic diagram of a modeling process of a parameter reuse method according to an embodiment of the present invention;

[0048] Figure 6 : is a schematic diagram of three-tone signal modeling according to an embodiment of the present invention; wherein, Figure 6 (a) is a time domain diagram of three-tone signal modeling using different methods. Figure 6 (b) is the frequency domain diagram of the three-tone signal modeled by different methods;

[0049] Figure 7 100-tone signal modeling frequency domain diagram according to an embodiment of the present invention; wherein, Figure 7 (a) is a time domain diagram of 100-tone signal modeling using different methods. Figure 7 (b) is the frequency domain diagram of 100-tone signal modeling using different methods;

[0050] Figure 8 1. This is a graph showing a change in the normalized mean square error of a multi-tone signal excitation parameter multiplexing model according to an embodiment of the present invention;

[0051] Figure 9 1. It is a schematic diagram showing the changes of various order terms of the multi-tone signal excitation parameter multiplexing according to an embodiment of the present invention;

[0052] Figure 10 1 is a schematic diagram of the LMS iteration results for some multi-tone signal model parameters according to an embodiment of the present invention; wherein, Figure 10 (a) is a schematic diagram of the iteration results from the 1st to the 40th iteration starting with the multi-tone parameter. Figure 10 (b) is a schematic diagram of the results of iterations from 1 to 16,000 with the initial parameter being 0;

[0053] Figure 11 Schematic diagram of the relationship between bandwidth and normalized mean square error of a parameter multiplexing method according to an embodiment of the present invention;

[0054] Figure 12 This is a schematic diagram of the changes in various order terms of broadband Gaussian white noise signal excitation parameter multiplexing according to an embodiment of the present invention:

[0055] Figure 13 : is a schematic diagram of the LMS iteration results for some broadband Gaussian white noise signal model parameters according to one embodiment of the present invention; wherein, Figure 13 (a) is a schematic diagram of the 1st to 40th iterations of the LMS model parameters for some broadband Gaussian white noise signals. Figure 13 (b) is a schematic diagram of the 1st to 16000th iterations of LMS for some broadband white Gaussian noise signal model parameters;

[0056] Figure 14 Schematic diagram of the normalized power of each order of BSPK parameter multiplexing excited by a broadband white Gaussian noise signal according to one embodiment of the present invention;

[0057] Figure 15 FIG. 1 is a schematic diagram of the LMS iteration results for some BPSK signal model parameters according to an embodiment of the present invention; wherein, Figure 15 (a) is a schematic diagram of the 1st to 40th iterations of LMS for some BPSK signal model parameters. Figure 15 (b) is a schematic diagram of the 1st to 16000th iterations of the LMS BPSK signal model parameters;

[0058] Figure 16 is a schematic diagram of an experimental platform according to an embodiment of the present invention;

[0059] Figure 17 1 is a schematic diagram of modeling results of a broadband noise signal using various multi-tone signals according to an embodiment of the present invention;

[0060] Figure 18Schematic diagram of the LMS iteration results for some measured multi-tone signal model parameters according to one embodiment of the present invention; wherein, Figure 18 (a) is a schematic diagram of the 1st to 40th iterations of LMS for some measured multi-tone signal model parameters. Figure 18 (b) is a schematic diagram of the 1st to 16000th iterations of LMS for some measured multi-tone signal model parameters;

[0061] Figure 19 1 is a schematic diagram of iterative multiplexing of 0 parameters and 18 tone parameters according to an embodiment of the present invention;

[0062] Figure 20 1 is a schematic diagram of modeling results of various bandwidth signals for a broadband noise signal according to an embodiment of the present invention;

[0063] Figure 21 : is a graph showing the relationship between the normalized mean square error and the number of iterations of local optimal searches for Gaussian white noise signals of different bandwidths at each starting point in an embodiment of the present invention; wherein, Figure 21 (a) is a schematic diagram of the 1st to 40th iterations of the normalized mean square error of the local optimal search at each starting point of Gaussian white noise signals with different bandwidths. Figure 21 (b) is a schematic diagram of the normalized mean square error of the local optimal search for each starting point of the 1 MHz signal from the 1st to the 15000th iteration;

[0064] Figure 22 1 is a schematic diagram of modeling results of BPSK signals using signals of various bandwidths according to an embodiment of the present invention;

[0065] Figure 23 is a schematic diagram of a relationship between the normalized mean square error of the local optimal search for each starting point of a BPSK signal and the number of iterations according to an embodiment of the present invention; wherein, Figure 23 (a) is a schematic diagram of the normalized mean square error of the local optimal search for each starting point of the BPSK signal from the 1st to the 40th iteration. Figure 23 (b) is a schematic diagram of the normalized mean square error of the local optimal search for each starting point of the BPSK signal from the 1st to the 15000th iteration;

[0066] Figure 24 1 is a flow chart of a method for modeling broadband nonlinear characteristics of a radio frequency channel according to an embodiment of the present invention;

[0067] Figure 25 This is a schematic block diagram of a method for modeling broadband nonlinear characteristics of a radio frequency channel according to an embodiment of the present invention;

[0068] Figure 26 This is a schematic block diagram of the structure of a computer device according to an embodiment of the present application.

[0069] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0070] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0071] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative position relationship, movement status, etc. between the various components under a certain specific posture (as shown in the accompanying drawings). If the specific posture changes, the directional indication will also change accordingly. The connection can be a direct connection or an indirect connection.

[0072] The term "and / or" in this article is only a description of the association relationship between associated objects, indicating that there can be three relationships. For example, A and B can mean: A exists alone, A and B exist at the same time, and B exists alone.

[0073] In addition, in the present invention, descriptions such as "first" and "second" are for descriptive purposes only and should not be understood as indicating or implying their relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of such features. In addition, the technical solutions between the various embodiments can be combined with each other, but this must be based on the fact that they can be implemented by ordinary technicians in this field. When the combination of technical solutions is contradictory or cannot be implemented, it should be deemed that such combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

[0074] Both time-domain nonlinear models and frequency-domain nonlinear models are common nonlinear modeling methods. Compared with frequency-domain models, time-domain models do not require additional Fourier transforms on the signal, have lower computational complexity, and are more widely used. Most time-domain models are derived from the Volterra series (a mathematical functional that describes the relationship between the input and output of a nonlinear system). The Volterra series is equivalent to a Taylor series with memory. It can be viewed as a combination of linear convolution and nonlinear power series, and has the ability to simultaneously handle static nonlinear problems and memory effects. The discrete Volterra series can be expressed as

[0075] (1)

[0076] in, for k Order Volterra kernel function; is the nth order Volterra series, is the discrete Volterra series; M is the memory depth of the model; K is the highest order of the Volterra series, represents the memory depth of the first-order Volterra kernel function, represents the memory depth of the k-th order Volterra kernel function, Indicates the original signal sampling points. The Volterra series has advantages in physical meaning, discretization, and parameter extraction. However, it also has problems such as a large number of parameters, difficult and time-consuming calculations, and limited contribution of some terms to the nonlinearity of the model. Currently, most of the Volterra series used for modeling are simplified or trimmed Volterra series. The generalized memory polynomial (GMP) model used in this paper is a simplified Volterra model. The GMP model has low complexity and high modeling accuracy. Its discrete form is expressed as follows:

[0077] (2)

[0078] in, is the excitation signal, is the memory depth of the basis function part of the MP model, is the nonlinear order; and are the memory depth and nonlinear order of the lag term respectively, and the lag delay is ; and are the memory depth and nonlinear order of the lead term respectively, and the lead time is , represents the sum of GMP model outputs, Indicates the first sampling points, Indicates the first sampling points, Indicates the first sampling points.

[0079] Frequency domain nonlinear models originated from time domain nonlinear models. Since the frequency domain characteristics of some nonlinear effects are more obvious, frequency domain nonlinear models have more advantages in analyzing nonlinear effects. Generalized frequency response function (GFRF) is a widely used frequency domain nonlinear model. Its core is the multidimensional Fourier transform of the Volterra kernel function. The discrete frequency domain nonlinear model can be expressed as

[0080] (3)

[0081] in, and represent the discrete input spectrum and discrete output spectrum of the model respectively, is the kth frequency domain kernel, discrete frequency , is the number of sampling points, Represents the maximum order of the frequency domain kernel. The relationship between the time domain kernel and the frequency domain kernel is

[0082] (4)

[0083] Frequency domain nonlinear models have advantages in describing nonlinear effects such as memory effects. However, when modeling spread spectrum signals, the frequency domain kernel requires a large number of parameters to accurately describe the nonlinear characteristics. In formula (4), The time domain kernel has parameters, the frequency domain kernel has Parameters. Generally speaking, in order to ensure the accuracy of parameter extraction, the number of sampling points much higher than Frequency-domain nonlinear models require far more computation than time-domain nonlinear models. Therefore, frequency-domain models are primarily used for offline analysis of nonlinear effects in RF devices. Time-domain models remain the primary method for real-time nonlinear modeling of RF devices.

[0084] The parameter extraction method determines the lower limit of the nonlinear model. The mainstream time-domain nonlinear parameter extraction methods can be divided into analytical methods represented by the least squares method (LS) and iterative methods represented by the least mean square algorithm (LMS). LS is a data fitting method based on statistical regression, which determines the optimal parameters by searching for the minimum sum of the squares of the difference between the target data and the observed data. It has the advantages of fast convergence speed and high accuracy, and the obtained parameters are generally considered to be optimal. Taking the GMP model as an example, the basis function part of Equation (2) can be rewritten as

[0085] (5)

[0086] in, , ,

[0087]

[0088] Represents the number of signal points used for modeling. The model parameters can be estimated by LS ;

[0089] (6)

[0090] However, in practical applications, LS suffers from pathological problems. Specifically, in strongly nonlinear scenarios, the number of model parameters required for LS calculation increases significantly. This ultimately leads to excessive correlation between the model basis functions, resulting in observation errors. To minimize the impact of observation errors, the number of training samples must significantly exceed the number of model parameters. Furthermore, the more accurate the description of nonlinear effects, the greater the number of model parameters. Excessive parameters significantly slows down LS calculations, resulting in insufficient real-time performance.

[0091] The least mean square method (LMS) is an adaptive filtering algorithm based on the gradient descent method. It optimizes the model parameters by minimizing the mean square error so that the error between the algorithm output and the expected output is minimized. Taking the basis function of the GMP model as an example, LMS is composed of the input signal , expected signal , error signal It consists of three parts, and the objective function is

[0092] (7)

[0093] In LMS, the model parameters Update via gradient descent

[0094] (8)

[0095] in, is the step size factor, which determines the convergence speed and stability of the algorithm. Its value usually satisfies , [ ] indicates averaging. The LMS algorithm can identify the dynamic characteristics of an unknown model and update it in real time, with low computational complexity. However, the LMS algorithm is essentially a linear adaptive filtering algorithm and is prone to falling into local optima. When applied to nonlinear models with many parameters, the algorithm takes a long time to converge, and the modeling accuracy is inferior to that of the LS algorithm.

[0096] The parameters of the nonlinear model reflect the nonlinear characteristics of the RF device. If the nonlinear characteristics of two signals are similar, then the model parameters of the two signals will also be similar. In other words, for signals that produce similar nonlinear effects, their nonlinear model parameters can be reused. If the time domain model is represented by the frequency domain method, combining equations (3) and (4), the discrete frequency domain nonlinear model can be further written as

[0097] (9)

[0098] in, is the model order, For memory depth, is the number of signal sampling points, is the sampling rate, Indicates that the functions in the brackets are convolved with each other Second-rate.

[0099] From formula (9), we can see that the time domain model is essentially a simplified frequency domain model. The time domain k-order Volterra model includes the input signal The order terms of , each delay item and their cross terms, each type of term can be regarded as a separate signal. The time domain model matches the spectrum of each item of the input signal, which greatly reduces the amount of calculation. From the meaning of the parameters, the frequency domain nonlinear model reserves a parameter for each frequency. Finally, the various frequencies are combined to obtain the estimated output. The time domain nonlinear model reserves parameters for each item. One parameter can describe the spectral power of this item as a whole, and the spectrum is summed to obtain the estimated output. The difference between the two modeling strategies is as follows Figure 1 shown.

[0100] From the frequency domain, when the output overlap of each frequency point of the signal that produces nonlinear effects is consistent or the frequency point power is extremely small, the obtained model parameters can be reused. In other words, the more similar the two signals are, the more accurate the parameter reuse result will be. Performing an inverse Fourier transform on equation (4), we have

[0101] (10)

[0102] in, is the unit impulse function. When the time domain model parameters are reused, The time domain kernel variation is

[0103] (11)

[0104] in, express The time-domain kernel variation, , Indicates the set negation operation. Indicates the intersection of sets. Indicates the union of sets. hour, and The frequencies not included are all set to 0. It is composed of the intersection term and the inverse term of the intersection of frequency domain multiplexing parameters. The value of the output signal Affected by the combination of components at each frequency point. For spread spectrum signals, the time domain model parameters are mainly affected by the power within the frequency domain model band, but when the signal characteristics are not much different, the distribution of the signal within the band is similar, and the influence of the intersection term of the frequency domain multiplexing parameters is mostly offset. Since the out-of-band signal has low power, the intersection inverse term of the frequency domain multiplexing parameters has little effect on the overall time domain model parameters. Therefore, the residual cancellation within the band and the unaccounted error outside the band cause the time domain model parameters to deviate during multiplexing. The smaller the difference in signal characteristics, the smaller the deviation. This type of deviation is ultimately reflected in the fact that when the signal type changes, the optimal model parameters will be offset relative to the original signal, or the local optimum and global optimum of the model will be offset. The diagram of the offset between the local optimum and the global optimum is shown in the figure below. Figure 2 shown.

[0105] when When the nonlinear characteristics are not fully excited, the frequency domain kernel There are more items than , the incompletely excited nonlinear characteristics become the main source of error, and the modeling method is similar to using the model parameters excited by narrowband signals for broadband signal modeling.

[0106] when When the nonlinear characteristics of each frequency point are fully excited, The power of the frequency points outside the bandwidth is small enough to be ignored for the overall modeling results. However, too many frequency points lead to the nonlinear characteristics overlapping at the same frequency points. For example, for a broadband signal, if there is

[0107] (12)

[0108] but

[0109] (13)

[0110] exist Under the encouragement of Contains the nonlinear characteristics of the two components of the above formula, but if Contains only This amount, The output results of the frequency points will have errors, which is why the frequency domain model is sensitive to noise. This type of modeling method is similar to using model parameters excited by a broadband signal to model narrowband signals.

[0111] Reference Figure 24 The present invention proposes a method for modeling broadband nonlinear characteristics of a radio frequency channel, comprising:

[0112] S1: Obtain the input signal and output signal to be modeled;

[0113] S2: Detecting a bandwidth type of the input signal and / or the output signal;

[0114] S3: obtaining multiple sets of model parameters based on the bandwidth type, and inputting the multiple sets of model parameters into a preset model respectively, thereby obtaining multiple pre-trained models; wherein the preset model is a nonlinear model;

[0115] S4: inputting the input signal into each of the pre-trained models respectively to obtain a corresponding prediction signal, and calculating a first error value based on the output signal and the prediction signal;

[0116] S5: Selecting a preset number of pre-training models in ascending order according to the values of the first error values as target pre-training models;

[0117] S6: Acquire a training data set of the same type based on the bandwidth type; wherein a set of training data in the training data set includes a first input signal and a first output signal;

[0118] S7: Inputting the training data set into each of the target pre-training models for training, and obtaining corresponding nonlinear characteristic measurement models after training;

[0119] S8: Obtain the second error value of each of the nonlinear characteristic measurement models, and select the nonlinear characteristic measurement model with the smallest second error value as the target nonlinear characteristic measurement model.

[0120] As described in the above step S1, the input signal and output signal to be modeled are obtained, wherein the input signal and output signal are the signals to be modeled, which can be obtained by manual uploading or by crawling.

[0121] As described in step S2 above, the bandwidth type of the input signal and / or the output signal is detected, referring to Figure 3 , use FFT (Fast Fourier Transform) to calculate the spectrum of the received signal. Fourier transform converts the signal from the time domain to the frequency domain to help analyze the frequency components in the signal.

[0122] As described in step S3 above, multiple sets of model parameters are obtained based on the bandwidth type, and the multiple sets of model parameters are respectively input into the preset model to obtain multiple pre-trained models. Figure 4Essentially, model parameters are the optimal solution that minimizes the difference between the estimated and actual results. However, due to the large number of nonlinear model parameters, existing parameter extraction methods can often only obtain locally optimal model parameters. In simulation analysis, because variables such as noise and nonlinear characteristics are stable and controllable, the extracted parameters converge at similar local optimal points. Therefore, the model parameters obtained by simulation analysis often show a certain regularity as parameters such as signal type and bandwidth vary. However, in actual measurements, noise and nonlinear characteristics are uncontrollable, and the local optimal points in the solution space vary dramatically due to disturbances such as noise. This results in large differences in the obtained model parameters even when the signals are acquired very close in time.

[0123] Parameter reuse methods require a local optimum search. They expect the local optimum in the solution space of the other signals to be identical to that of the excitation signal. However, due to noise and nonlinear variations, the local optima of the two solution spaces often deviate. Therefore, it is necessary to use the local optimum in the solution space of the excitation signal as a starting point to search for local optima near the other signals, thereby achieving higher accuracy. Compared to starting convergence from scratch, searching from the local optimum in the solution space of the excitation signal can reduce computational effort and improve convergence stability.

[0124] As described in step S4 above, the input signal is input into each of the pre-trained models respectively to obtain a corresponding prediction signal, and a first error value is calculated based on the output signal and the prediction signal, wherein the first error value is calculated by calculating its mean square error. Since the model parameters have been input into the pre-trained model, a corresponding prediction signal will be obtained after the input signal, and then the mean square error is calculated based on the prediction signal and the output signal.

[0125] As described in step S5 above, a preset number of pre-trained models are selected in order from small to large according to the values of each of the first error values and recorded as target pre-trained models. According to theoretical analysis, the closer the source signal of the reused parameters and the signal to be modeled are in the frequency domain, the higher the accuracy of the parameter reuse. Therefore, parameters with bandwidths close to the source signal and the signal to be modeled should be preferentially reused. At the same time, in actual situations, the results of parameter reuse vary greatly due to the influence of noise and convergence path, and even a false local optimum may occur, that is, the normalized mean square error of reuse is extremely low, but the normalized mean square error increases sharply in the early stage of iteration. Based on this phenomenon, a strategy of similar bandwidth and appropriate generalization is adopted to construct a LUT (Look-Up Table). That is, multiple parameters obtained from signals with bandwidths close to the signal to be modeled are preferentially selected, and another parameter with the best reuse effect can also be selected to jointly construct the LUT to ensure a balance between computational complexity and accuracy.

[0126] As described in steps S6-S7 above, a training data set of the same type is obtained based on the bandwidth type, and the training data set is input into each of the target pre-training models for training, thereby obtaining corresponding nonlinear characteristic measurement models after training.

[0127] Since it is necessary to build a model based on the input signal bandwidth, a training data set of the same type is obtained based on the bandwidth type to ensure that the final model obtained by training is related to the input signal. The specific training method is based on the least mean square method. The least mean square method (LMS) is an adaptive filtering algorithm based on the gradient descent method. The model parameters are optimized by minimizing the mean square error to minimize the error between the algorithm output and the expected output. Taking the basis function of the GMP model as an example, the LMS is composed of the input signal , expected signal , error signal It consists of three parts, and the objective function is

[0128] (7)

[0129] In LMS, the model parameters Update via gradient descent

[0130] (8)

[0131] in, is the step size factor, which determines the convergence speed and stability of the algorithm. Its value usually satisfies The LMS algorithm can identify the dynamic characteristics of unknown models and update them in real time with low computational complexity.

[0132] As described in step S8 above, the second error value of each nonlinear characteristic measurement model is obtained, and the nonlinear characteristic measurement model with the smallest second error value is selected as the target nonlinear characteristic measurement model. The second error value can be a normalized mean square error or a corrected normalized mean square error. Since multiple nonlinear characteristic measurement models have been selected, the nonlinear characteristic measurement model with the smallest error can be selected as the target nonlinear characteristic measurement model. The speed and accuracy of parameter identification by model iteration are greatly improved. Compared with remodeling using traditional methods, this application greatly improves the speed of model parameter identification while ensuring model accuracy.

[0133] In one embodiment, the step S8 of obtaining the second error value of each nonlinear characteristic measurement model includes:

[0134] S801: Acquire a test data set of the same type as the bandwidth type; wherein a set of test data in the test data set includes a test input signal and a test actual output signal;

[0135] S802: Inputting a test input signal in the test data set into each of the nonlinear characteristic measurement models to obtain a test prediction output signal;

[0136] S803: According to the formula Calculate the second error value; wherein NMSE is the second error value, N is the number of test data groups in the test data set, The actual output signal for the nth test, Predict the output signal for the nth test.

[0137] As described in steps S801-S803 above, in order to compare the difference between the actual test output signal and the predicted test output signal, a second error value is generally used to evaluate the nonlinear model fitting accuracy. The second error value is defined as:

[0138] (14)

[0139] Among them, NMSE is the second error value, N is the number of test data groups in the test data set, The actual output signal for the nth test, Predict the output signal for the nth test.

[0140] In one embodiment, the step S8 of obtaining the second error value of each nonlinear characteristic measurement model includes:

[0141] S811: Acquire a test data set of the same type as the bandwidth type; wherein a set of test data in the test data set includes a test input signal and a test actual output signal;

[0142] S812: Inputting the test input signal in the test data set into each of the nonlinear characteristic measurement models to obtain a test prediction output signal;

[0143] S813: According to the formula Calculate the second error value; wherein, NMSE ’ is the second error value, N is the number of test data sets in the test data set, The actual output signal for the nth test, Predict the output signal for the nth test, .

[0144] As described in steps S811-S813 above, in actual situations, the parameters extracted from the test prediction output signal and the parameters extracted from the test actual output signal may have phase reversal or ratio inconsistency problems. These problems are caused by the inconsistency between the power of the test prediction output signal and the test actual output signal. Therefore, when calculating the fitting accuracy of the test actual output signal for parameter reuse, it should be corrected using the least squares method. The corrected second error value is defined as

[0145] (15)

[0146] in

[0147] (16)

[0148] The corrected second error value is the error function of the parameter reuse method.

[0149] The process of the multi-starting point search algorithm is as follows Figure 5 As shown: If the input signal of the nonlinear model to be modeled is , the output signal is .

[0150] Step 1: Use existing model parameters Respectively with the input signal Calculated by nonlinear model , and calculate and The corrected second error value.

[0151] Step 2: In the existing model parameters The selection is made in the criterion that the parameters with the best multiplexing effect among different types of signals are selected, and the selection criterion is that the corrected second error value is minimized.

[0152] Step 3: Select the model parameters As the starting point of LMS, check the convergence of LMS with different starting points. The convergence criterion is .

[0153] Step 4: Choose the best as new parameters for the nonlinear model.

[0154] In one embodiment, the nonlinear model is a time-domain model or a frequency-domain model. Both time-domain nonlinear models and frequency-domain nonlinear models are common nonlinear modeling methods. Compared to frequency-domain models, time-domain models do not require additional Fourier transforms of the signal, resulting in lower computational complexity and wider application.

[0155] In one embodiment, the nonlinear model is a generalized method of moments model. A generalized memory polynomial (GMP) model is a simplified Volterra model. The GMP model has low complexity and high modeling accuracy, which can meet the needs of this application.

[0156] In one embodiment, the step S7 of inputting the training data set into each of the target pre-training models for training, and obtaining the corresponding nonlinear characteristic measurement models after training, respectively, includes:

[0157] S701: Input the training data set into each of the target pre-training models in sequence, and update the model parameters in each of the target pre-training models by the least mean square method during each training, so as to obtain the corresponding nonlinear characteristic measurement models after the training is completed.

[0158] The least mean square method (LMS) is an adaptive filtering algorithm based on the gradient descent method. It optimizes the model parameters by minimizing the mean square error so that the error between the algorithm output and the expected output is minimized. Taking the basis function of the GMP model as an example, LMS is composed of the input signal , expected signal , error signal It consists of three parts, and the objective function is

[0159] (7)

[0160] In LMS, the model parameters Update via gradient descent

[0161] (8)

[0162] in, is the step size factor, which determines the convergence speed and stability of the algorithm. Its value usually satisfies The LMS algorithm can identify the dynamic characteristics of unknown models and update them in real time with low computational complexity.

[0163] In one embodiment, the step S3 of acquiring multiple sets of model parameters based on the bandwidth type includes:

[0164] S301: Inputting the input signal and / or output signal into a preset time domain model to obtain a corresponding frequency spectrum;

[0165] S302: Match multiple groups of model parameters based on the frequency spectrum.

[0166] As described in steps S301-S302 above, the parameters of the nonlinear model reflect the nonlinear characteristics of the RF device. If the nonlinear characteristics generated by two signals are similar, the model parameters of the two signals are likely to be similar. From a frequency domain perspective, when the output overlap of the various frequencies of the signals generating the nonlinear effect is consistent or the frequency power is extremely low, the resulting model parameters can be reused. In other words, the more similar the two signals are, the more accurate the parameter reuse results will be, so multiple sets of parameters can be matched based on the spectrum.

[0167] In a specific embodiment, the simulation process is as follows: With signal Input the nonlinear model separately to get the distorted output signal 、 . Use the GMP model to analyze the signal With signal Model them separately and get the model parameters of the two 、 Finally, the model parameters Used for Get the output , model parameters Used for Get the output .

[0168] Since the existing simulation methods cannot fully simulate the time-varying characteristics of nonlinear models, in order to make the simulation process as consistent as possible with the actual situation, the parameters of the nonlinear model are derived from the mean of the high-order GMP model parameters obtained in the experiment, and the GMP model used in the simulation modeling is a low-order model.

[0169] According to the theoretical analysis of the proposed method, it can be inferred that although the parameters of multi-tone signals cannot be reused due to the large difference between their frequency domain characteristics and those of wideband signals, if the number of multi-tone signal tones is continuously increased within the wideband range, the modeling error of the obtained model parameters for wideband signals will gradually decrease.

[0170] In order to verify the correctness of the proposed theory, the three-tone signal and the multi-tone signal with 100 tones commonly used in the frequency domain model are used as excitations, and the obtained model parameters are used for broadband signals. Among them, the multi-tone signal is evenly distributed within a bandwidth of 5MHz, and the broadband signal bandwidth is 5MHz. The modeling results are shown in Figure 2. Figure 6 and Figure 7 shown.

[0171] Among them, the normalized mean square error of the three-tone signal parameter reuse is 0.96dB, and the normalized mean square error of the 100-tone signal parameter reuse is -2.85dB. From the modeling results, it can be found that when the nonlinear model parameters obtained by using the three-tone signal as the excitation are reused on a white noise signal with a bandwidth of 5MHz, the model accuracy is not as good as the model parameters identified by the excitation signal with a higher number of tones. When the number of tones in the excitation signal gradually increases, the normalized mean square error of the parameter reuse shows a clear downward trend, as shown in Figure 2. Figure 8 shown.

[0172] When the number of tones is small, the normalized mean square error (NMEE) of parameter reuse varies dramatically, with the highest-accuracy NME reaching -8.94dB. However, as the number of tones increases, the NMEE shows a clear downward trend. This indicates that as the input signal approaches a wideband signal in the frequency domain, the resulting nonlinear model closely matches the nonlinear characteristics of the wideband signal.

[0173] However, the problem of large variations in the accuracy of the parameter reuse model under the condition of a low number of tones seems to violate this rule, because this cannot explain why the multi-tone excitation with a tonal number of 9 ultimately obtains an accuracy far higher than that of other excitations. This is because when the number of tones is small, the frequency domain excitation points of the multi-tone signal will change greatly with the change in the number of tones, and the model parameters will also be quite different. The parameters of the multi-tone signal with a tonal number of 9 happen to fall on a local optimum point of the broadband signal, which is why the parameter reuse model under the condition of a low number of tones exhibits large variations in accuracy. The model parameters of each tone are analyzed. The various order terms of the multi-tone signal excitation parameter reuse are as follows: Figure 9 shown.

[0174] Before the number of polyphonic signals reaches 20, the distribution of model parameters is relatively chaotic. After the number of polyphonic signals exceeds 20, the distribution of model parameters gradually converges. This also indirectly shows that signals with fewer polyphonic signals have a small coverage range in the frequency domain and are difficult to stimulate the broadband nonlinear characteristics of the nonlinear model, which ultimately shows low modeling accuracy and poor stability. In order to verify whether the model parameters obtained from these polyphonic signals are suitable for broadband signals, LMS is used to iterate the model parameters of some polyphonic signals. The iteration results are shown in the figure below. Figure 10 shown.

[0175] Clearly, all LMS methods using the multitone signal parameters as a starting point failed to converge downward to the correct optimization path. However, the LMS method with an initial parameter of 0 found the correct convergence direction, ultimately achieving a modeling result with much higher accuracy than the multitone signal. Therefore, the model parameters derived from multitone signal excitation are not suitable for broadband signal modeling.

[0176] In order to study the modeling ability of the parameter multiplexing method for broadband signals, Gaussian white noise signals with different bandwidths are used as excitations, and the model parameters are applied to Gaussian white noise signals and BPSK signals. The simulation results are shown in Figure 2. Figure 12 As shown in , for Gaussian white noise signals, the bandwidth of the signal to be modeled is set to 6MHz, 7MHz, 8MHz, and 9MHz respectively, and the bandwidth of the excitation signal increases from 5MHz to 10MHz with a step of 0.1MHz. The model parameters obtained from each excitation signal are used to calculate the normalized mean square error of the multiplexed signal. Figure 11 As shown in Table 1, the closer the bandwidths of the excitation signal and the reused signal are, the higher the accuracy of the model parameter reuse. It is worth noting that as the signal bandwidth increases, the modeling accuracy of the nonlinear model gradually decreases, which also leads to a decrease in the accuracy of the parameter reuse method. As can be seen from Table 1, the accuracy of the parameter reuse method decreases faster than that of the re-modeling method.

[0177] Table 1 Comparison of normalized mean square error between remodeling and parameter reuse

[0178]

[0179] Taking a broadband Gaussian white noise broadband signal with a multiplexing signal bandwidth of 7 MHz as an example, the normalized power of each order of the broadband signal excitation parameter multiplexing is as follows: Figure 12 As shown in Figure 3, for broadband signals, the model parameters gradually converge as the bandwidth increases. Figure 12 Items 1 to 4 correspond to first-order model terms, items 4 to 24 correspond to second-order model terms, and items 25-44 correspond to third-order model terms. The 7MHz signal model parameters with the highest multiplexing accuracy have larger coefficients in the third-order terms, and third-order parameters contribute more to model nonlinearity than second-order parameters.

[0180] The model parameters of Gaussian white noise signals with bandwidths of 5MHz, 7MHz, and 9MHz are used as the starting points of LMS iterations. Compared with the starting point of 0 parameters, the iteration results are as follows: Figure 13 As shown in the figure. In short-term iterations of 40 sampling points, the parameter reuse method achieves lower initial normalized mean square error and faster response than the traditional LMS method. In long-term iterations of 16,000 sampling points, the optimal normalized mean square error of the 7MHz and 9MHz parameter reuse methods remains lower than that of the traditional LMS method, indicating that some of the iteration starting points of the proposed method are closer to a more optimal local minimum.

[0181] Similarly, for BPSK signals, Gaussian white noise of different bandwidths is used as the model parameters for excitation to perform modeling, and the results are as follows: Figure 14As shown in the figure, it can be seen that the broadband white noise signal can effectively fit the BPSK signal with similar bandwidth. When the white noise bandwidth is 4.7MHz, the parameter reuse model has the highest accuracy, which can reach -22.92dB. The Gaussian white noise signal model parameters with bandwidths of 3MHz, 4.7MHz, and 6MHz are reused on the BPSK signal and iterated. The iterative results are compared with the results of the zero-parameter starting point as shown in the figure. Figure 15 As shown in the figure, under both short-term and long-term iteration conditions, the 4.7MHz bandwidth parameter reuse performance is better than the 0-parameter starting point. It can be said that approximately broadband white noise can be used for nonlinear modeling and reuse of BPSK signals.

[0182] Figure 16 The experimental platform consists of a transmitter 300, a low-noise amplifier 200, and a receiver 100. The transmit signal is written to the transmitter 300 (Agilent E4438C) via a personal computer (PC 400). The transmit signal then travels through the RF line to the low-noise amplifier 200 and receiver 100. The RF front-end, except for the low-noise amplifier 200, is integrated into the receiver 100.

[0183] The experimental transmit signal was generated in MATLAB. After entering transmitter 300, it was modulated onto a carrier frequency of 1563.2 MHz. The transmit signal consisted of in-phase and quadrature (I / Q) signals with a sampling rate of 25.039 MHz. The clock was aligned with the receiver 100 clock. The transmitted multi-tone and wideband signals were generated periodically using periodic baseband data, and the transmit power was manually adjusted by transmitter 300.

[0184] The received signal was acquired using the Chipscope platform, with the sampling rate also set to 25.039 MHz. The received RF signal was down-converted to an intermediate frequency (IF) before sampling via the front-end. It's important to note that clock synchronization between receiver 100 and transmitter 300 is essential; if not, the received and transmitted signals will exhibit significant phase deviation.

[0185] The multi-tone signals collected in the experiment are used to build a model. The obtained model parameters are applied to a broadband noise signal with a bandwidth of 1 MHz. The results of the normalized mean square error varying with the number of tones are shown in the figure below. Figure 17 As shown in the figure, the normalized mean square error (NMEE) of the reused model decreases as the number of tones increases. Specifically, as the input signal approaches the wideband signal in the frequency domain, the resulting nonlinear model closely matches the nonlinear characteristics of the wideband signal. The lowest NMEE achieved with parameter reuse is -21.14 dB, while the NMEE achieved using the LS method is -24.13 dB, representing a 2.99 dB difference in accuracy.

[0186] However, due to factors such as noise, the normalized mean square error (NME) of the reused model is unstable. Some reused parameter models are significantly inferior to the NME of the re-modeled broadband noise signal. This shows that when model parameters derived from multi-tone signals are applied to broadband signals, while modeling performance improves with the number of tones, they still exhibit poor stability.

[0187] The parameter with the best multiplexing effect among the 16-tone to 18-tone signals is selected as the starting point of the LMS iteration. The iteration results and the iteration process with the parameter 0 as the starting point are as follows: Figure 18 and Figure 19 As shown, for Figure 18 , the parameter iteration direction is from light to dark. The initial accuracy of the parameter reuse method is higher than that of the starting point parameter 0. However, as the number of iterations increases, the reused parameters gradually break away from the original local optimum and iterate in the same direction as the starting point parameter 0. It is worth noting that although the convergence trend of the parameters at each starting point is the same, the accuracy of the 18-tone signal parameters is consistently higher than that of the starting point parameter 0 during the long-term iteration process. This shows that the proposed method can overcome the convergence dilemma of the traditional LMS method and further improve the convergence accuracy of the LMS method.

[0188] The obtained model parameters are applied to a broadband white noise signal with a bandwidth of 1 MHz. The obtained normalized mean square error varies with the number of tones. Figure 20 As shown in Table 2, for the illustrated wideband excitation range of 0.4 MHz to 10 MHz, the 1 MHz parameter reuse method achieves the highest accuracy when the model parameters are applied to a 1 MHz wideband white noise signal, even surpassing the LS fitting accuracy. Comparing the remodeling of wideband signals acquired at different times reveals that the parameter reuse method with a 1 MHz bandwidth achieves the lowest average normalized mean square error. This indicates that, when the maximum accuracy remains similar, parameter reuse using signals of the same bandwidth provides more stable modeling results.

[0189] Table 2 Comparison of normalized mean square error between broadband white noise signal remodeling and parameter reuse

[0190]

[0191] The parameter with the highest normalized mean square error among the parameters in the range of 1MHz, 2MHz, and 3MHz and the parameter with the initial value of 0 are selected as the starting point of LMS. The 1MHz signal collected at different times is used as the optimization target. The normalized mean square error iteration results are as follows: Figure 21As shown in Table 3, when the number of iterations is small, the normalized mean squared error (NMSE) of parameter reuse exhibits two states. For algorithms starting at 1MHz and 2MHz, the model parameters converge quickly to the local optimum. For the algorithm starting at 3MHz, the model parameters converge slowly and gradually move toward other local optima, with the NMSE showing a slowly decreasing trend. It can be seen that the selected parameter starting points are very close to the local optimum, allowing for rapid convergence through local optimal search. However, when the initial parameter value is 0, the parameters have difficulty converging to a high accuracy. With a sufficient number of iterations, the local optimal search for parameter reuse can further iterate to the local optimum, while the parameters starting at 0 gradually converge toward the local optimum, but the convergence time and accuracy are still inferior to those of the proposed method. The proposed method significantly improves the parameter convergence speed. For the 3MHz parameter starting point, which has the worst convergence performance, the number of iterations is 38.9% of that for the 0-valued parameter. For the 1MHz parameter starting point, which has the best convergence performance, the number of iterations is only 2.5% of that for the 0-valued parameter.

[0192] Table 3 Changes in normalized mean square error of local optimal search at different starting points for 1MHz signal

[0193]

[0194] When the signal to be modeled is a BPSK signal with a bandwidth of 2.3 MHz, the parameter reuse result without iteration is as follows: Figure 22 As shown in the figure, unexpectedly, the multiplexing parameter accuracy of 5MHz bandwidth is higher than that of 2MHz and 3MHz, which seems to contradict the proposed theory. To further verify, the three parameters with the highest normalized mean square error of parameter multiplexing in the range of 2MHz, 3MHz, and 5MHz and the parameter with the initial value of 0 are selected as the starting point, and the BPSK signals collected at different times are used as the optimization target. The normalized mean square error iteration results are shown as follows: Figure 23 As shown in Table 4. When the number of iterations is small, 2MHz and 3MHz are closer to the local optimum, and the normalized mean square error does not change much, while the 5MHz signal parameter with the best initial multiplexing result shows the worst optimization result. Figure 23 From the above, we can see that this parameter first deviates from the original local optimum and then iterates again to find the optimal parameter. When the number of iterations is high, the 5MHz signal parameter performs worse than the best normalized mean square error (NMSE) obtained when the other parameters converge. The root cause of this result is a false local optimum, where the obtained parameters are located near an unstable local optimum far from the BPSK signal to be modeled. When iterations begin, the algorithm quickly escapes this unstable local optimum and converges towards a more stable optimum. This problem rarely occurs for bandwidths close to the signal to be modeled. Therefore, selecting parameters with similar bandwidths for reuse is a better strategy.

[0195] With the exception of the 5MHz parameter, the 2MHz and 3MHz parameters all demonstrated performance superior to the 0-parameter during the iteration process, especially for the 2MHz parameter reuse iteration starting point. Under the same number of iterations, the 2MHz parameter reuse consistently achieved higher accuracy than the 0-parameter. Using the optimal 0-parameter iteration result as the criterion, the number of iterations required to achieve the target was 69.7% of the 0-parameter iteration itself. Using a 1dB improvement over the iteration result as the criterion, the 2MHz parameter reuse iteration required 951 iterations, which is 27.5% of the 0-parameter iteration's 3463 iterations. This demonstrates that the proposed method outperforms existing 0-parameter iteration methods in both computational complexity and accuracy, with the same accuracy or number of iterations.

[0196] Table 4 Changes in normalized mean square error of local optimal search at different states for each starting point of BPSK signal

[0197]

[0198] This paper proposes a model parameter reuse multi-starting point iterative algorithm to address the slow iteration of nonlinear model parameters. It creates a LUT for existing nonlinear model parameters within the same model, enabling rapid response to sudden changes in nonlinear characteristics. To accurately describe the accuracy of the signal to be modeled with different model parameters, the LS method is applied to correct the normalized mean square error (NMSE), avoiding spurious NMSE increases caused by phase reversal and signal power variations. Furthermore, to further enrich the iterative starting points and avoid exhaustive calculation of all parameter starting points, a method is proposed to select optimal model parameters based on signal type. Specifically, the model first detects the bandwidth of the signal to be modeled and preferentially reuses model parameters obtained from signals with a bandwidth close to that of the modeled signal. Furthermore, this paper combines time-domain and frequency-domain models, analyzing the feasibility of time-domain model parameter reuse using the concept of the frequency-domain model and performing error analysis. Simulation and experimental results demonstrate that the proposed method significantly improves the speed and accuracy of parameter identification using LMS iteration. Compared with re-modeling using the LS method, the proposed method significantly improves model parameter identification speed at the expense of reduced accuracy.

[0199] Reference Figure 25 The present invention also provides a device for modeling broadband nonlinear characteristics of a radio frequency channel, comprising:

[0200] A first acquisition module 10 is used to acquire input signals and output signals to be modeled;

[0201] A detection module 20, configured to detect a bandwidth type of the input signal and / or the output signal;

[0202] A second acquisition module 30 is configured to acquire multiple sets of model parameters based on the bandwidth type, and input the multiple sets of model parameters into a preset model to obtain multiple pre-trained models; wherein the preset model is a nonlinear model;

[0203] A first input module 40 is configured to input the input signal into each of the pre-trained models to obtain a corresponding prediction signal, and calculate a first error value based on the output signal and the prediction signal;

[0204] A selection module 50 is configured to select a preset number of pre-trained models as target pre-trained models in ascending order according to the values of the first error values;

[0205] A third acquisition module 60 is configured to acquire a training data set of the same type based on the bandwidth type; wherein a set of training data in the training data set includes a first input signal and a first output signal;

[0206] A second input module 70 is used to input the training data set into each of the target pre-training models for training, and obtain a corresponding nonlinear characteristic measurement model after training;

[0207] The fourth acquisition module 80 is configured to acquire the second error value of each of the nonlinear characteristic measurement models, and select the nonlinear characteristic measurement model with the smallest second error value as the target nonlinear characteristic measurement model.

[0208] Other embodiments of the apparatus for modeling the wideband nonlinear characteristics of a radio frequency channel provided by the present invention are the same as the method for modeling the wideband nonlinear characteristics of a radio frequency channel, and are not described in detail here.

[0209] Reference Figure 26 In the embodiment of the present application, a computer device is also provided. The computer device may be a server, and its internal structure may be as follows: Figure 26 As shown. The computer device includes a processor, a memory, a network interface and a database connected via a system bus. The processor designed by the computer is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store various radio frequency channel broadband nonlinear characteristic modeling methods, etc. The network interface of the computer device is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, it can implement the radio frequency channel broadband nonlinear characteristic modeling method described in any of the above embodiments.

[0210] Those skilled in the art will understand that Figure 26The structure shown in is merely a block diagram of a portion of the structure related to the present application solution and does not constitute a limitation on the computer device to which the present application solution is applied.

[0211] An embodiment of the present application further provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the method for modeling the broadband nonlinear characteristics of a radio frequency channel described in any of the above embodiments can be implemented.

[0212] Those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the above-described method embodiments. Any reference to memory, storage, database, or other media provided herein and used in the embodiments may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct RAMbus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM).

[0213] It should be noted that, in this document, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, apparatus, article, or method comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, apparatus, article, or method. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, apparatus, article, or method comprising the element.

[0214] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of the claims.

Claims

1. A method for modeling broadband nonlinear characteristics of a radio frequency channel, characterized in that: include: Obtain the input signal and output signal to be modeled; detecting a bandwidth type of the input signal and / or the output signal; Acquire multiple sets of model parameters based on the bandwidth type, and input the multiple sets of model parameters into a preset model respectively, thereby obtaining multiple pre-trained models; wherein the preset model is a nonlinear model; Inputting the input signal into each of the pre-trained models respectively to obtain a corresponding prediction signal, and calculating a first error value based on the output signal and the prediction signal; According to the values of the first error values, a preset number of pre-training models are selected in order from small to large and recorded as target pre-training models; Acquire a training data set of the same type based on the bandwidth type; wherein a set of training data in the training data set includes a first input signal and a first output signal; Inputting the training data set into each of the target pre-training models for training, and obtaining corresponding nonlinear characteristic measurement models after training; obtaining second error values of the respective nonlinear characteristic measurement models, and selecting the nonlinear characteristic measurement model with the smallest second error value as a target nonlinear characteristic measurement model; The step of obtaining the second error value of each of the nonlinear characteristic measurement models includes: Acquire a test data set of the same type as the bandwidth type; wherein a set of test data in the test data set includes a test input signal and a test actual output signal; Inputting a test input signal in the test data set into each of the nonlinear characteristic measurement models to obtain a test prediction output signal; According to the formula Calculate the second error value; wherein, NMSE ’ is the second error value, N is the number of test data sets in the test data set, The actual output signal for the nth test, Predict the output signal for the nth test, .

2. The method for modeling broadband nonlinear characteristics of a radio frequency channel according to claim 1, wherein: The step of obtaining the second error value of each of the nonlinear characteristic measurement models includes: Acquire a test data set of the same type as the bandwidth type; wherein a set of test data in the test data set includes a test input signal and a test actual output signal; Inputting a test input signal in the test data set into each of the nonlinear characteristic measurement models to obtain a test prediction output signal; According to the formula Calculate the second error value; wherein NMSE is the second error value, N is the number of test data groups in the test data set, The actual output signal for the nth test, Predict the output signal for the nth test.

3. The method for modeling broadband nonlinear characteristics of a radio frequency channel according to claim 1, wherein: The nonlinear model is a time domain model or a frequency domain model.

4. The method for modeling the broadband nonlinear characteristics of a radio frequency channel according to claim 3, wherein: The nonlinear model is a generalized method of moments estimation model.

5. The method for modeling broadband nonlinear characteristics of a radio frequency channel according to claim 1, wherein: The step of inputting the training data set into each of the target pre-training models for training, and obtaining corresponding nonlinear characteristic measurement models after training, respectively, includes: The training data set is input into each of the target pre-training models in sequence, and the model parameters in each of the target pre-training models are updated by the least mean square method during each training, so as to obtain corresponding nonlinear characteristic measurement models after the training is completed.

6. The method for modeling broadband nonlinear characteristics of a radio frequency channel according to claim 1, wherein: The step of obtaining multiple groups of model parameters based on the bandwidth type includes: Inputting the input signal and / or the output signal into a preset time domain model to obtain a corresponding frequency spectrum; Multiple sets of model parameters are matched based on the frequency spectrum.

7. A device for modeling broadband nonlinear characteristics of a radio frequency channel, characterized in that: include: A first acquisition module is used to obtain input signals and output signals to be modeled; a detection module, configured to detect a bandwidth type of the input signal and / or the output signal; a second acquisition module, configured to acquire multiple sets of model parameters based on the bandwidth type, and input the multiple sets of model parameters into a preset model, thereby obtaining multiple pre-trained models; wherein the preset model is a nonlinear model; A first input module, configured to input the input signal into each of the pre-trained models, obtain a corresponding prediction signal, and calculate a first error value based on the output signal and the prediction signal; A selection module, configured to select a preset number of pre-trained models as target pre-trained models in ascending order according to the values of the first error values; A third acquisition module is configured to acquire a training data set of the same type based on the bandwidth type; wherein a set of training data in the training data set includes a first input signal and a first output signal; A second input module is used to input the training data set into each of the target pre-training models for training, and respectively obtain a nonlinear characteristic measurement model after the training is completed; a fourth acquisition module, configured to acquire second error values of the respective nonlinear characteristic measurement models, and select the nonlinear characteristic measurement model with the smallest second error value as a target nonlinear characteristic measurement model; The fourth acquisition module includes: Acquire a test data set of the same type as the bandwidth type; wherein a set of test data in the test data set includes a test input signal and a test actual output signal; Inputting a test input signal in the test data set into each of the nonlinear characteristic measurement models to obtain a test prediction output signal; According to the formula Calculate the second error value; wherein, NMSE ’ is the second error value, N is the number of test data sets in the test data set, The actual output signal for the nth test, Predict the output signal for the nth test, .

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images