Fsvr-dqn-based fi-dac system amplitude-frequency error correction method

Abstract

The application provides an FSVR-DQN-based FI-DAC system amplitude-frequency error correction method, belongs to the technical field of time domain testing, and solves the problems of insufficient fitting precision, complex weight design, overfitting and insufficient adaptability to complex signal processing in amplitude-frequency error correction, and comprises the following steps: single-tone signals are generated at a fixed frequency interval within a set frequency range, a target waveform table is generated for each frequency point through digital domain preprocessing, and an IFFT transformation is performed on the input frequency domain signals to obtain time domain signals; the time sequence signals are input into an FI-DAC system, an FFT transformation is performed on the time sequence signals by using a DQN-SVR algorithm to obtain frequency domain analog signals; the parameters of a spectrum analyzer are set, the frequency domain analog signals are collected by using the spectrum analyzer, and the amplitude-frequency characteristics of a low-frequency sub-path and a high-frequency sub-path are tested respectively; the unit of the collected signals is converted from dBm into a normalized amplitude value, the collected signals are compared according to the amplitude-frequency characteristics of ideal frequency division signals, amplitude-frequency errors are extracted, and the amplitude-frequency error correction of the FI-DAC system is completed.

Description

TECHNICAL FIELD

[0001] The application relates to an FSVR-DQN-based FI-DAC system amplitude-frequency error correction method and belongs to the technical field of time domain testing. BACKGROUND

[0002] The digital and analog modules in the FI-DAC system have non-ideal characteristics, including the zero-order hold characteristic of the DAC, the insertion loss and impedance mismatch in the hardware link, etc. These factors will cause amplitude-frequency response deviation, especially in the frequency overlap band and high-frequency sub-band. Traditional amplitude-frequency error correction methods mostly use window functions or support vector regression (SVR) technology. Although these methods improve the amplitude-frequency characteristics to some extent, they have problems such as insufficient fitting accuracy, overfitting, and insufficient adaptability to complex signals. Fuzzy support vector regression (FSVR) improves the fitting performance of the traditional SVR method by introducing fuzzy weights, but still has problems such as complex fuzzy weight design and insufficient dynamic adaptability.

[0003] FI-DAC (Frequency Interleaving Digital-to-Analog Converter) is designed to expand the bandwidth of digital-to-analog converters (DACs) through analog processing systems. FI-DAC uses analog processing components such as mixers, filters, and combiners to enhance the bandwidth performance of DACs. By interleaving DACs in the frequency domain, it combines multiple lower-rate DACs to output a wider bandwidth analog signal.

[0004] FSVR (Fuzzy Support Vector Regression) is a machine learning method that combines fuzzy logic and support vector regression (SVR). It improves the fitting performance of the traditional SVR method by introducing fuzzy weights. In FSVR, each data point is assigned a fuzzy weight that reflects its importance or uncertainty in the overall data set. FSVR uses these weights to adjust the error term, thereby improving the model's ability to fit non-linear and complex data sets. FSVR is particularly suitable for data sets that are difficult for traditional SVR to handle, with high uncertainty and non-linearity.

[0005] DQN (Deep Q-Network) is an algorithm that combines deep learning and Q-learning to solve reinforcement learning problems. DQN uses a deep neural network to approximate the Q function, which is the expected utility of an action in a given state. This network can learn a policy directly from raw pixels or other high-dimensional inputs without the need for hand-designed features. One of the key innovations of DQN is experience replay, which stores and reuses past experiences to break the correlation between samples, improving the stability and efficiency of learning. DQN has shown excellent performance in various games and simulation environments, handling complex decision-making problems.

[0006] FSVR-DQN combines the advantages of fuzzy support vector regression (FSVR) and deep Q-network (DQN) for the amplitude-frequency error correction of FI-DAC systems. This algorithm uses the reinforcement learning mechanism of DQN to adjust the fuzzy weights in FSVR in real time to adapt to the dynamic changes of frequency characteristics. By intelligently selecting weight adjustment actions, FSVR-DQN can accurately fit the passband region while maintaining the stability of the stopband weights, avoiding model overfitting. This dynamic optimization strategy not only improves the fitting accuracy but also enhances the robustness of the system in complex frequency environments, providing a nonlinear amplitude-frequency error correction scheme for FI-DAC systems. SUMMARY

[0007] To solve the problems of insufficient fitting accuracy, complex fuzzy weight design, overfitting, and insufficient adaptability to complex signals in amplitude-frequency error correction, the present application proposes a FI-DAC system amplitude-frequency error correction method based on FSVR-DQN, which specifically includes:

[0008] Step 1: Generate single-tone signals at fixed frequency intervals within a specified frequency range, generate a target waveform table for each frequency point through digital domain preprocessing, and perform IFFT transformation on the input frequency domain signal to obtain a time-domain signal;

[0009] Step 2: Input the time sequence signal into the FI-DAC system and use the DQN-SVR algorithm to perform FFT transformation to obtain a frequency domain analog signal;

[0010] Step 3: Set the parameters of the spectrum analyzer, use the spectrum analyzer to collect the frequency domain analog signal, and test the amplitude-frequency characteristics of the low-frequency sub-path and the high-frequency sub-path respectively;

[0011] Step 4: Convert the collected signal unit from dBm to normalized amplitude, compare the collected signal with the ideal frequency division signal amplitude-frequency characteristics, extract the amplitude-frequency error, and complete the amplitude-frequency error correction of the FI-DAC system.

[0012] Preferably, the digital domain preprocessing step in step 1 includes:

[0013]

[0013] Step 1.1: The generated single-tone signal is used as test data, and the original test data is locally polynomial fitted using the Savitzky-Golay filter to extract the global trend and smooth the signal fluctuations;

[0014] Step 1.2: Calculate the residual of the original test data and the global trend and clean the residual, identify outliers in the original test data after cleaning the residual based on the Isolation Forest algorithm, and remove the mutant data points;

[0015] Step 1.3: Re-stack the original test data after removing the mutant data points and the global trend to generate smooth test data;

[0016] Step 1.4: Repeat steps 1.1-1.3 until the preset number of times, take the average of the preset number of smooth test data until the ideal performance is met, complete the digital domain preprocessing and generate the target waveform table for each frequency point.

[0017] Preferably, the step of inputting the time sequence signal into the FI-DAC system using the FSVR-DQN algorithm in step 2 includes:

[0018] Step 2.1: The time domain signal output in step 1 is used as the training data set of the FSVR model, and the amplitude-frequency error pre-equalizer is set , wherein is a feature matrix, which includes the cosine value of each frequency point, the target value vector and the corresponding target amplitude-frequency response;

[0019] Step 2.2: Set the maximum number of iterations and the convergence threshold;

[0020] Step 2.3: Initialize the FSVR model parameters and the DQN model parameters;

[0021] Step 2.4: Train the SVR model to output FSVR model parameters w, b and obtain the frequency domain simulation signal.

[0022] Preferably, the initialization of the FSVR model parameters in step 2.3 includes: the parameter epsilon of the insensitive loss function, the regularization parameter C and the fuzzy weight vector s;

[0023] The initialization of the DQN model parameters includes: the experience replay pool D, the original Q value network, the target Q value network parameters and the learning parameters.

[0024] Preferably, step 2.4 specifically includes:

[0025] Step 2.4.1: The FSVR model is directly fitted by minimizing the loss function to obtain the initial coefficient ;

[0026] Step 2.4.2: using the initialized FSVR model parameters to make prediction, fitting the FSVR model to get the passband frequency loss and fuzzy weight variance;

[0027] Step 2.4.3: taking the passband frequency loss and fuzzy weight variance as the current state, the DQN model root epsilon-greedy strategy obtains the corresponding action;

[0028] Step 2.4.4: updating the FSVR weight parameters s according to the action, refitting the FSVR model, updating ;

[0029] Step 2.4.5: calculating the new prediction and error to update the state, calculating the immediate reward based on the new prediction and error;

[0030] Step 2.4.6: storing the calculated new prediction, error and immediate reward as transition information to the experience recycling pool D;

[0031] Step 2.4.7: collecting small batch data N from the experience recycling pool D;

[0032] Step 2.4.8: updating the original Q value network parameters and the target Q value network parameters;

[0033] Step 2.4.9: repeating steps 2.4.1-2.4.8 until the FSVR model convergence condition or the maximum iteration number is obtained, outputting the FSVR model parameters w, b;

[0034] (1);

[0035] In formula (1), is the loss function of fuzzy support vector regression, is the weighted penalty term of the loss function, is the weight vector of the regression model, and the square norm of the weight vector is used to control the complexity of the model to avoid overfitting, is the bias term of the regression model, is the fuzzy weight, is the epsilon-insensitive region;

[0036] The calculation formula of the FSVR model parameter prediction is:

[0037] (2);

[0038] In formula (2), and are the initial coefficients of the first training process;

[0039] The calculation formula of the current state of the FSVR model is:

[0040] (3);

[0041] In formula (3), is the loss function of the model in the passband frequency range in the first state, is the loss function of the model in the passband frequency range in the second state, is the variance of the sample blur weight in the first state, is the variance of the sample blur weight in the second state, which is used to quantify the uncertainty or dispersion degree of the data distribution;

[0042] The expression of the epsilon-greedy strategy is:

[0043] (4);

[0044] In formula (4), is the action output by the DQN model, is the current state;

[0045] The expression for updating the FSVR weight parameter s is:

[0046] (5);

[0047] In formula (5), is the FSVR weight parameter before updating, is the action update;

[0048] The calculation formula of the new prediction and error is:

[0049] (6);

[0050] The calculation formula of the immediate reward is:

[0051] (7);

[0052] In formula (7), is a weight coefficient for adjusting the contribution of loss improvement to the reward, is a weight coefficient for adjusting the contribution of the change in the variance of the blur weight to the reward;

[0053] The expression for storing the transition information into the experience recycling pool D is:

[0054] (8);

[0055] The expression for collecting small batches of data N is:

[0056] (9);

[0057] In formula (9), state, current action, immediate reward and next state, respectively;

[0058] The expression for updating the original Q value network parameter is:

[0059] (10);

[0060] In formula (10), is a loss function;

[0061] The expression for updating the target Q value network parameter is:

[0062] (11);

[0063] In formula (11), is an update proportion coefficient.

[0064] The beneficial effects of the present application are:

[0065] 1. The constellation diagram of the system after correction is more clear, the signal quality is significantly improved, especially in the high frequency band, the correction effect is more prominent.

[0066] 2. The present application can effectively reduce EVM, improve SNR, and significantly improve the signal quality and stability of the system through the correction of low frequency and high frequency signals. This correction method is particularly important in high frequency signals, and DQN-FSVR correction shows superior performance in the case of large signal distortion.

[0067] 3. The present application designs an amplitude-frequency error pre-equalizer to compensate for the non-ideal characteristics of each module in the system, especially the amplitude-frequency error of the frequency overlap band and the high frequency sub-band of the sub-path signal, thereby improving the amplitude-frequency response accuracy of the output signal. BRIEF DESCRIPTION OF DRAWINGS

[0068] Figure 1 The flowchart of the FSVR-DQN-based FI-DAC system amplitude-frequency error correction method provided by the present application;

[0069] Figure 2 The flowchart of the test data preprocessing provided by the present application;

[0070] Figure 3 The flowchart of the FSVR-DQN algorithm provided by the present application;

[0071] Figure 4 The error test schematic diagram provided by the present application;

[0072] Figure 5 This is a schematic diagram of the QAM signal test of the FI-DAC system before correction provided by the present invention;

[0073] Figure 6 This is a schematic diagram of the QAM signal test of the corrected FI-DAC system provided by the present invention. Detailed Implementation

[0074] Specific implementation method one: Combining Figures 1-4 This implementation method is described as follows: Figure 1 As shown, the steps of the amplitude-frequency error correction method for the FI-DAC system based on FSVR-DQN described in this embodiment include:

[0075] S1: Within a set frequency range, generate single-tone signals at fixed frequency intervals, generate target waveform tables for each frequency point through digital domain preprocessing, and perform IFFT transformation on the input frequency domain signal to obtain the time domain signal;

[0076] In this embodiment, the fixed frequency interval is 10 MHz, and the digital domain preprocessing steps include:

[0077] S101: Trend Learning: Using the generated monotone signal as test data, the Savitzky-Golay filter is used to perform local polynomial fitting on the original test data to extract the global trend and smooth signal fluctuations.

[0078] S102: Outlier Detection: Calculate the residual between the original test data and the global trend and clean up the residual. Based on the isolated forest algorithm, identify outliers in the original test data after residual cleanup, remove abrupt data points, and improve data consistency.

[0079] S103: Signal Reconstruction: The original test data after removing abrupt data points is overlaid with the global trend to generate smooth test data;

[0080] S104: Repeat S101-S103 until the preset number of times is reached. Take the average value of the smoothed test data of the preset number of times to further remove the interference of random fluctuations on the data trend until the ideal performance is met. Complete the digital domain preprocessing and generate a target waveform table for each frequency point.

[0081] S2: Input the timing signal into the FI-DAC system and perform FFT transformation using the DQN-SVR algorithm to obtain the frequency domain analog signal;

[0082] The joint algorithm based on fuzzy support vector regression (FSVR) and deep Q network (DQN) provided by the application realizes high-precision fitting of a passband by dynamically optimizing fuzzy weights, and effectively suppresses stopband errors; support vector regression (SVR) introduces an insensitive bandwidth, and only calculates a loss when a deviation between a predicted value and a target value exceeds a set threshold, thereby optimizing fitting precision while controlling model complexity; based on this, the application further introduces fuzzy weights by fuzzy support vector regression (FSVR), and different importance is allocated to different frequency regions; in order to ensure the fitting precision of the passband, FSVR allocates higher weights to the passband region and lower weights to the stopband region, so as to reduce the interference of non-target frequency signals on model fitting; however, the static allocation of fuzzy weights has limitations, and it is difficult to cope with the characteristic changes of complex regions such as frequency overlap bands and high-frequency subbands, which may lead to local underfitting or overfitting; therefore, the application combines the dynamic optimization strategy of deep Q network (DQN), and uses a reinforcement learning mechanism to adjust fuzzy weights in real time; DQN intelligently selects weight adjustment actions by evaluating the current fitting error and the state of fuzzy weight distribution, so that the fitting of the passband region is more accurate, while the stability of the stopband weight is maintained, and the model is prevented from being excessively complicated; the joint algorithm can adapt to dynamic changes of frequency characteristics, improve fitting precision and robustness in a complex frequency band environment, and provides an efficient and flexible optimization scheme for amplitude-frequency error correction of an FI-DAC system.

[0083] The pseudo code flow of the DQN-SVR algorithm is shown in Table 1.

[0084] Table 1

[0085]

[0086] S3: The frequency domain analog signal is collected by using a spectrum analyzer, and the amplitude-frequency characteristics of the low-frequency subpath and the high-frequency subpath are tested respectively; the resolution of the test data is determined by the parameter setting of the spectrum analyzer, so as to ensure sufficient precision;

[0087] S4: The unit of the collected signal is converted from dBm to a normalized amplitude, and the collected signal is compared according to the amplitude-frequency characteristics of the ideal frequency division signal, the amplitude-frequency error is extracted, and the amplitude-frequency error correction of the FI-DAC system is completed.

[0088] Specific implementation method two: combination Figure 5 and Figure 6 The present embodiment is described, in order to verify the amplitude-frequency error correction effect of the application, the present embodiment tests 16-QAM, 32-QAM, 64-QAM, 128-QAM, 256-QAM, 512-QAM and 1024-QAM signals, the center frequencies are 1GHz and 2.5GHz respectively, and the bandwidth is 200MHz, effectively verifying the characteristics of the low-frequency and high-frequency subpaths, such asFigure 6 As shown, the corrected system constellation is more clear, and the signal quality is significantly improved, especially in the high frequency band, the correction effect is more prominent.

[0089] As shown in FIG. 6, the EVM and SNR of the low frequency signal before correction show relatively obvious distortion, for example, the EVM of 16-QAM modulation before correction is as high as 1.50%, and the SNR is only 27.48 dB. However, after correction by DQN-FSVR, the EVM is reduced to 0.48%, and the SNR is significantly improved to 29.48 dB, with a gain of 2.00 dB. Similarly, for 32-QAM and 64-QAM, the EVM after correction is reduced to 0.48% and 0.49% respectively, and the SNR is improved to 29.48 dB and 29.49 dB respectively. The distortion of the low frequency signal before correction is relatively serious, and the DQN-FSVR correction effectively improves this problem, reduces the EVM and significantly improves the SNR. Figure 5 Figure 6 As shown in FIG. 7, for high frequency signals with a center frequency of 2.5 GHz and a bandwidth of 200 MHz, the EVM and SNR before correction show relatively obvious distortion, for example, the EVM of 16-QAM modulation before correction is as high as 3.50%, and the SNR is only 26.52 dB. However, after correction by DQN-FSVR, the EVM is reduced to 2.05%, and the SNR is significantly improved to 31.21 dB, with a gain of 4.69 dB. Similarly, for 32-QAM and 64-QAM, the EVM after correction is reduced to 1.78% and 1.83% respectively, and the SNR is improved to 30.99 dB and 31.06 dB respectively. The distortion of the high frequency signal before correction is relatively serious, and the DQN-FSVR correction effectively improves this problem, reduces the EVM and significantly improves the SNR.

[0090] In summary, Figure 6 The data in Table 1 fully prove the importance of DQN-FSVR correction in the FI-DAC system. Through the correction of low frequency and high frequency signals, DQN-FSVR can effectively reduce the EVM, improve the SNR, and significantly improve the signal quality and stability of the system. This correction method is particularly critical in high frequency signals, and DQN-FSVR correction exhibits superior performance in cases of greater signal distortion.

[0091] The above is only a preferred embodiment of the present application, and does not limit the present application in any form. Although the present application has been disclosed as above with a preferred embodiment, it is not intended to limit the present application. Any skilled person in the art can make some changes or modifications to the above disclosed technical content without departing from the scope of the technical solution of the present application, and any simple modification, equivalent replacement and improvement of the above embodiments within the scope of the technical solution of the present application and in the spirit and principle of the present application are also within the protection scope of the present application.​

Claims

1. A method for correcting the amplitude-frequency error of a FSRV-DQN-based FI-DAC system, characterized in that, The steps of the FSVR-DQN-based FI-DAC system amplitude-frequency error correction method include: Step 1: In a set frequency range, generate single-tone signals at a fixed frequency interval, generate a target waveform table for each frequency point through digital domain preprocessing, perform IFFT transformation on the input frequency domain signal to obtain a time domain signal; Step 2: Input the time sequence signal into the FI-DAC system and perform FFT transformation using the DQN-SVR algorithm to obtain a frequency domain analog signal; Step 3: Set the parameters of the spectrum analyzer, use the spectrum analyzer to collect the frequency domain analog signal, and test the amplitude-frequency characteristics of the low-frequency sub-path and the high-frequency sub-path respectively; Step 4: Convert the unit of the collected signal from dBm to normalized amplitude, compare the collected signal with the ideal frequency division signal according to the amplitude-frequency characteristics, extract the amplitude-frequency error, and complete the amplitude-frequency error correction of the FI-DAC system.

2. The FSVR-DQN-based FI-DAC system amplitude-frequency error correction method according to claim 1, characterized in that, The steps of digital domain preprocessing in step 1 include: Step 1.1: Use the Savitzky-Golay filter to perform local polynomial fitting on the original test data, extract the global trend and smooth the signal fluctuations; Step 1.2: Calculate the residual of the original test data and the global trend and clean the residual, identify outliers in the original test data after cleaning the residual based on the isolation forest algorithm, and remove the mutation data points; Step 1.3: Re-add the original test data after removing the mutation data points to the global trend to generate smooth test data; Step 1.4: Repeat steps 1.1-1.3 until the preset number of times, take the average of the preset number of smooth test data until the ideal performance is met, complete the digital domain preprocessing, and generate a target waveform table for each frequency point.

3. The FSVR-DQN-based FI-DAC system amplitude-frequency error correction method according to claim 1, characterized in that, The steps of step 2 include: Step 2.1: Set the output time-domain signal in step 1 as the training data set of the FSVR model, and set the amplitude-frequency error pre-equalizer wherein, is a feature matrix, the feature matrix including a cosine value of each frequency point, a target value vector, and a corresponding target amplitude-frequency response; Step 2.2: Set the maximum number of iterations and the convergence threshold; Step 2.3: Initialize the FSVR model parameters and the DQN model parameters; Step 2.4: Train the SVR model to output the FSVR model parameters w, b, and obtain the frequency domain analog signal.

4. The FSVR-DQN-based FI-DAC system amplitude-frequency error correction method according to claim 3, characterized in that, The initialization of the FSVR model parameters in step 2.3 includes: the parameters epsilon of the insensitive loss function, the regularization parameter C, and the fuzzy weight vector s; The initialization of the DQN model parameters includes: the experience replay pool D, the original Q value network, the target Q value network parameters, and the learning parameters.

5. The FSVR-DQN-based FI-DAC system amplitude-frequency error correction method according to claim 3, characterized in that, Step 2.4 specifically includes: Step 2.4.1: FSVR model is fitted directly by minimizing the loss function, obtaining initial coefficients ; Step 2.4.2: Use the initialized FSVR model parameters for prediction, and perform fitting on the FSVR model to obtain the passband frequency loss and the fuzzy weight variance; Step 2.4.3: Use the epsilon-greedy strategy to obtain the corresponding action based on the passband frequency loss and the fuzzy weight variance as the current state; Step 2.4.4: Update FSVR weight parameters s, refit FSVR model, update ; Step 2.4.5: Calculate the new prediction and error to update the state, and calculate the immediate reward based on the new prediction and error; Step 2.4.6: Store the calculated new prediction, error, and immediate reward as transition information in the experience recycling pool D; Step 2.4.7: Collecting small batch data N from experience replay pool D; Step 2.4.8: Updating original Q value network parameters and target Q value network parameters; Step 2.4.9: Repeating step 2.4.1-step 2.4.8 until FSVR model convergence condition or maximum iteration number is obtained, and outputting FSVR model parameters w, b; The target function expression of FSVR fitting is: （1）； In formula (1), is a loss function of fuzzy support vector regression, is a weighted penalty term of the loss function, is a weight vector of the regression model, and the square norm of the weight vector is used to control the complexity of the model to avoid overfitting, is a bias term of the regression model, is a fuzzy weight, is an epsilon-insensitive region; The calculation formula of FSVR model parameter prediction is: （2）； In equation (2), and are initial coefficients for the first training process; The calculation formula of FSVR model current state is: （3）； In formula (3), For the first The loss function of the model in each state within the passband frequency range For the first Sample fuzzy weights under each state The variance is used to quantify the uncertainty or dispersion of the data distribution; The expression of epsilon-greedy strategy is: （4）； In equation (4), is the action output by the DQN model, is the current state; The expression of updating FSVR weight parameter s is: （5）； In equation (5), is the FSVR weight parameter before update, is the action update; The calculation formula of new prediction and error is: （6）； The calculation formula of immediate return is: （7）； In equation (7), is a weight coefficient for adjusting the contribution of the loss improvement to the reward, is a weight coefficient for adjusting the contribution of the change in the blur weight variance to the reward. The expression of storing transition information to experience recycling pool D is: （8）； The expression of collecting small batch data N is: （9）； In equation (9), state, action, reward, and next state, respectively. The expression of updating original Q value network parameters is: （10）； In equation (10), is a loss function; The expression of updating target Q value network parameters is: （11）； In equation (11), is an update ratio coefficient.

Images