FI-DAC system amplitude-frequency error correction method based on FSVR-DQN

Abstract

The invention provides an FI-DAC system amplitude-frequency error correction method based on FSVR-DQN, belongs to the technical field of time domain testing, solves the problems of insufficient fitting precision, complex fuzzy weight design, over-fitting and insufficient adaptability to complex signal processing in amplitude-frequency error correction, and comprises the following steps: within a set frequency range, generating single-tone signals at fixed frequency intervals; generating a target waveform table for each frequency point through digital domain preprocessing, and carrying out IFFT (Inverse Fast Fourier Transform) on an input frequency domain signal to obtain a time domain signal; inputting the time sequence signal into an FI-DAC system, and carrying out FFT conversion by using a DQN-SVR algorithm to obtain a frequency domain analog signal; parameters of a spectrum analyzer are set, the spectrum analyzer is used for collecting the frequency domain analog signals, and the amplitude-frequency characteristics of the low-frequency sub-circuit and the high-frequency sub-circuit are tested respectively; converting the collected signal unit from dBm into a normalized amplitude value, comparing the collected signals according to the amplitude-frequency characteristic of the ideal frequency division signal, extracting an amplitude-frequency error, and completing the amplitude-frequency error correction of the FI-DAC system.

Description

Technical Field

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

[0002] The digital and analog modules in the FI-DAC system have non-ideal characteristics, including the zero-order hold characteristics of the DAC, the insertion loss and impedance mismatch in the hardware link, etc. These factors can cause amplitude-frequency response deviations, 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 have improved the amplitude-frequency characteristics to a certain extent, they still have problems such as insufficient fitting accuracy, overfitting, and insufficient adaptability to complex signal processing. Fuzzy support vector regression (FSVR) improves the fitting performance of the traditional SVR method by introducing fuzzy weights, but there are still 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, multiple lower-rate DACs are combined to output analog signals with wider bandwidths.

[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 traditional SVR methods by introducing fuzzy weights. In FSVR, each data point is assigned a fuzzy weight that reflects the importance or uncertainty of the point in the overall data set. FSVR uses these weights to adjust the error terms, thereby improving the model's ability to fit nonlinear and complex data sets. FSVR is particularly suitable for data sets with high uncertainty and nonlinearity that are difficult for traditional SVR to handle.

[0005] DQN (Deep Q-Network) is an algorithm that combines deep learning and Q-learning to solve reinforcement learning problems. DQN approximates the Q function, which is the expected utility of an action in a given state, by using a deep neural network. This network is able to learn strategies 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 experience to break the correlation between samples, thereby improving the stability and efficiency of learning. DQN performs well in a variety of games and simulation environments and is able to handle complex decision-making problems.

[0006] FSVR-DQN combines the advantages of fuzzy support vector regression (FSVR) and deep Q network (DQN) for amplitude-frequency error correction of FI-DAC system. The 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 and avoiding excessive complexity of the model. This dynamic optimization strategy not only improves the fitting accuracy, but also enhances the robustness of the system in complex frequency band environments, providing a nonlinear amplitude-frequency error correction solution for FI-DAC system. Summary of the invention

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

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

[0009] Step 2: Input the timing 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 spectrum analyzer parameters, 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: The collected signal unit is converted from dBm to normalized amplitude, and the collected signal is compared with the amplitude-frequency characteristics of the ideal frequency-divided signal to 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] Step 1.1: Use the generated single-tone signal as test data, and use the Savitzky-Golay filter to perform local polynomial fitting on the original test data to extract the global trend and smooth the signal fluctuation;

[0014] Step 1.2: Calculate the residuals between the original test data and the global trend and clean the residuals. Based on the isolation forest algorithm, identify the outliers in the original test data after cleaning the residuals and remove the mutation data points.

[0015] Step 1.3: Re-superimpose the original test data after removing the mutation data points with the global trend to generate smooth test data;

[0016] Step 1.4: Repeat steps 1.1 to 1.3 until the preset number of times is reached, and average 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.

[0017] Preferably, in step 2, the timing signal is input into the FI-DAC system and the FFT transformation is performed using the FSVR-DQN algorithm to obtain the frequency domain analog signal, which includes:

[0018] Step 2.1: Use 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 {x(ω), y(ω)}, where {x(ω), y(ω)} 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 convergence threshold;

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

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

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

[0023] Initialization of DQN model parameters includes: experience replay pool D, original Q value network, target Q value network parameters and 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 coefficients w1, b1;

[0026] Step 2.4.2: Use the initialized FSVR model parameters for prediction and fit the FSVR model to obtain 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 obtains the corresponding action based on the epsilon-greedy strategy;

[0028] Step 2.4.4: Update the FSVR weight parameter s according to the action, refit the FSVR model, and update (w i+1 ,b i+1 );

[0029] Step 2.4.5: Calculate new predictions and errors to update the state, and calculate immediate rewards based on the new predictions and errors;

[0030] Step 2.4.6: Store the calculated new prediction, error, and immediate reward as transfer information into the experience recycling pool D;

[0031] Step 2.4.7: Collect small batches of data N from the experience revisit pool D;

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

[0033] Step 2.4.9: Repeat steps 2.4.1 to 2.4.8 until the FSVR model convergence condition or the maximum number of iterations is obtained, and output the FSVR model parameters w, b;

[0034] The expression for minimizing the loss function is:

[0035]

[0036] In formula (1), L FSVR is, w is, b is, s is, ε is, w is;

[0037] The calculation formula for FSVR model parameter prediction is:

[0038]

[0039] In formula (2), w i and b i is the initial coefficient of the first training process;

[0040] The calculation formula for the current state of the FSVR model is:

[0041]

[0042] In formula (3), For Var i (s) is;

[0043] The expression of the epsilon-greedy strategy is:

[0044] action i =π(state i )(4);

[0045] In formula (4), action is the action output by the DQN model, and state is;

[0046] The expression for updating the FSVR weight parameter s is:

[0047]

[0048] In formula (5), is the FSVR weight parameter before update, action·adjustment is the action update; the calculation formula of the new prediction and error is:

[0049]

[0050] The calculation formula for immediate return is:

[0051]

[0052] In formula (7), α is, β is;

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

[0054] D←D∪(state i ,action i ,reward i ,state i+1 )(8);

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

[0056] (s,a,r,s′) N ∈D(9);

[0057] s,a,r,s′ are respectively;

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

[0059]

[0060] In formula (10), L(θ) is the loss function;

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

[0062] θ′←vθ+(1-v)θ′(11);

[0063] In formula (11), v is.

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

[0065] 1. After correction by the present invention, the system constellation diagram is analyzed more clearly, and the signal quality is significantly improved, especially in the high frequency band, where the correction effect is more prominent.

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

[0067] 3. The present invention 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 sub-channel signal frequency overlap band and the high-frequency sub-band, thereby improving the amplitude-frequency response accuracy of the output signal. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 A flow chart of the FI-DAC system amplitude-frequency error correction method based on FSVR-DQN provided by the present invention;

[0069] Figure 2 A flow chart of test data preprocessing provided by the present invention;

[0070] Figure 3 A schematic diagram of the flow of the FSVR-DQN algorithm provided by the present invention;

[0071] Figure 4 A schematic diagram of an error test provided by the present invention;

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

[0073] Figure 6 A schematic diagram of a QAM signal test of a FI-DAC system after correction provided by the present invention. DETAILED DESCRIPTION

[0074] Specific implementation method 1: Combination Figure 1-4 To illustrate this embodiment, Figure 1 As shown, the steps of the FI-DAC system amplitude-frequency error correction method based on FSVR-DQN described in this embodiment include:

[0075] S1: Generate a single tone signal at a fixed frequency interval within the set frequency range, generate a target waveform table for each frequency point through digital domain preprocessing, perform IFFT transformation on the input frequency domain signal, and obtain a time domain signal;

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

[0077] S101: Trend learning: Use the generated single tone signal as test data, use the Savitzky-Golay filter to perform local polynomial fitting on the original test data, extract the global trend and smooth the signal fluctuation;

[0078] S102: Outlier detection: Calculate the residuals between the original test data and the global trend and clean up the residuals. Based on the isolation forest algorithm, identify the outliers in the original test data after cleaning the residuals, remove the mutation data points, and improve the data consistency.

[0079] S103: Signal reconstruction: re-superimposing the original test data after removing the mutation data points 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 preset number of smooth test data, 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 use the DQN-SVR algorithm to perform FFT transformation to obtain a frequency domain analog signal;

[0082] The joint algorithm based on fuzzy support vector regression (FSVR) and deep Q network (DQN) proposed in the present invention realizes high-precision fitting of the passband by dynamically optimizing fuzzy weights, while effectively suppressing stopband errors. Support vector regression (SVR) introduces an insensitive bandwidth and only calculates the loss when the deviation between the predicted value and the target value exceeds a set threshold, thereby optimizing the fitting accuracy while controlling the complexity of the model; based on this, the present invention further introduces fuzzy weights through fuzzy support vector regression (FSVR) to assign different importance to different frequency regions. To ensure the fitting accuracy of the passband, FSVR assigns a higher weight to the passband region and a lower weight 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 is difficult to cope with the characteristic changes in complex areas such as frequency overlapping bands and high-frequency sub-bands, which may lead to local insufficient fitting or overfitting. Therefore, the present invention combines the dynamic optimization strategy of the deep Q network (DQN) and uses the reinforcement learning mechanism to adjust the fuzzy weights in real time. The DQN intelligently selects the weight adjustment action by evaluating the current fitting error and the distribution state of the fuzzy weights, making the fitting of the passband area more accurate while maintaining the stability of the stopband weights and avoiding over-complexity of the model. The joint algorithm can adapt to the dynamic changes of frequency characteristics, improve the fitting accuracy and robustness in a complex frequency band environment, and provide an efficient and flexible optimization solution for the amplitude-frequency error correction of the FI-DAC system.

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

[0084] Table 1

[0085]

[0086]

[0087] S3: Use a spectrum analyzer to collect frequency domain analog signals and test the amplitude-frequency characteristics of the low-frequency sub-channel and the high-frequency sub-channel respectively. The resolution of the test data is set by the parameters of the spectrum analyzer to ensure sufficient accuracy;

[0088] S4: The collected signal unit is converted from dBm to a normalized amplitude, and the collected signal is compared with the amplitude-frequency characteristic of the ideal frequency-divided signal to extract the amplitude-frequency error, thereby completing the amplitude-frequency error correction of the FI-DAC system.

[0089] Specific implementation method 2: Combination Figure 5 and Figure 6This embodiment is described. To verify the amplitude-frequency error correction effect of the present invention, this embodiment tests 16-QAM, 32-QAM, 64-QAM, 128-QAM, 256-QAM, 512-QAM and 1024-QAM signals, with center frequencies of 1 GHz and 2.5 GHz, respectively, and a bandwidth of 200 MHz, effectively verifying the characteristics of the low-frequency and high-frequency sub-paths, such as Figure 6 As shown in the figure, after correction, the system constellation diagram has clearer analysis and the signal quality is significantly improved, especially in the high frequency band, where the correction effect is more prominent.

[0090] like Figure 5 and Figure 6 As shown in the figure, for high-frequency signals with a center frequency of 2.5GHz and a bandwidth of 200MHz, the EVM and SNR before correction show obvious distortion. For example, the EVM of 16-QAM modulation before correction is as high as 3.50%, and the SNR is only 26.52dB. However, after DQN-FSVR correction, the EVM is greatly reduced to 2.05%, the SNR is significantly improved to 31.21dB, and the gain reaches 4.69dB. Similarly, for 32-QAM and 64-QAM, the corrected EVM is reduced to 1.78% and 1.83%, respectively, while the SNR is improved to 30.99dB and 31.06dB, respectively. The distortion of high-frequency signals before correction is more serious, and the DQN-FSVR correction effectively improves this problem, reducing the EVM and significantly improving the SNR.

[0091] In summary, Figure 6 The data in the figure fully demonstrates the importance of DQN-FSVR correction in the FI-DAC system. By correcting low-frequency and high-frequency signals, DQN-FSVR can effectively reduce EVM, improve SNR, and significantly improve the signal quality and stability of the system. This correction method is particularly critical in high-frequency signals. In the case of large signal distortion, DQN-FSVR correction shows better performance.

[0092] The above is only a preferred embodiment of the present invention and does not limit the present invention in any form. Although the present invention has been disclosed as a preferred embodiment as above, it is not used to limit the present invention. Any technician familiar with this profession can make some changes or modify the technical contents disclosed above into equivalent embodiments without departing from the scope of the technical solution of the present invention. However, any simple modification, equivalent replacement and improvement made to the above embodiments without departing from the content of the technical solution of the present invention, based on the technical essence of the present invention, within the spirit and principles of the present invention, still fall within the protection scope of the technical solution of the present invention.

Claims

1. The FI-DAC system amplitude-frequency error correction method based on FSVR-DQN is characterized by: The steps of the FI-DAC system amplitude-frequency error correction method based on FSVR-DQN include: Step 1: Generate a single tone signal at a fixed frequency interval within the set frequency range, generate a target waveform table for each frequency point through digital domain preprocessing, perform IFFT transformation on the input frequency domain signal, and obtain a time domain signal; Step 2: Input the timing signal into the FI-DAC system and use the DQN-SVR algorithm to perform FFT transformation to obtain a frequency domain analog signal; Step 3: Set the spectrum analyzer parameters, 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: The collected signal unit is converted from dBm to normalized amplitude, and the collected signal is compared with the amplitude-frequency characteristics of the ideal frequency-divided signal to extract the amplitude-frequency error and complete the amplitude-frequency error correction of the FI-DAC system.

2. The FI-DAC system amplitude-frequency error correction method based on FSVR-DQN according to claim 1 is characterized in that: The steps of digital domain preprocessing in step 1 include: Step 1.1: Use the generated single-tone signal as test data, and use the Savitzky-Golay filter to perform local polynomial fitting on the original test data to extract the global trend and smooth the signal fluctuation; Step 1.2: Calculate the residuals between the original test data and the global trend and clean the residuals. Based on the isolation forest algorithm, identify the outliers in the original test data after cleaning the residuals and remove the mutation data points. Step 1.3: Re-superimpose the original test data after removing the mutation data points with the global trend to generate smooth test data; Step 1.4: Repeat steps 1.1 to 1.3 until the preset number of times is reached, and average 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 FI-DAC system amplitude-frequency error correction method based on FSVR-DQN according to claim 1 is characterized in that: In step 2, the timing signal is input into the FI-DAC system and the FSVR-DQN algorithm is used to perform FFT transformation to obtain the frequency domain analog signal. The steps include: Step 2.1: Use 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 {x(ω), y(ω)}, where {x(ω), y(ω)} is a feature matrix, which includes the cosine value of each frequency point, the target value vector and the corresponding target amplitude-frequency response; Step 2.2: Set the maximum number of iterations and convergence threshold; Step 2.3: Initialize FSVR model parameters and DQN model parameters; Step 2.4: Train the SVR model, output the FSVR model parameters w,b, and obtain the frequency domain simulation signal.

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

5. The FI-DAC system amplitude-frequency error correction method based on FSVR-DQN according to claim 3 is characterized in that: Step 2.4 specifically includes: Step 2.4.1: The FSVR model is directly fitted by minimizing the loss function to obtain the initial coefficients w1, b1; Step 2.4.2: Use the initialized FSVR model parameters for prediction and fit the FSVR model to obtain the passband frequency loss and fuzzy weight variance; Step 2.4.3: Taking the passband frequency loss and fuzzy weight variance as the current state, the DQN model obtains the corresponding action based on the epsilon-greedy strategy; Step 2.4.4: Update the FSVR weight parameter s according to the action, refit the FSVR model, and update (w i+1 ,b i+1 ); Step 2.4.5: Calculate new predictions and errors to update the state, and calculate immediate rewards based on the new predictions and errors; Step 2.4.6: Store the calculated new prediction, error, and immediate reward as transfer information into the experience recycling pool D; Step 2.4.7: Collect small batches of data N from the experience revisit pool D; Step 2.4.8: Update the original Q value network parameters and the target Q value network parameters; Step 2.4.9: Repeat steps 2.4.1 to 2.4.8 until the FSVR model convergence condition or the maximum number of iterations is obtained, and output the FSVR model parameters w, b; The objective function expression of FSVR fitting is: In formula (1), w is the loss function of fuzzy support vector regression (FSVR), C is the weighted penalty term of the loss function, w 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, b is the bias term of the regression model, s is the fuzzy weight, and ε is the epsilon-insensitive area; The calculation formula for FSVR model parameter prediction is: In formula (2), w i and b i is the initial coefficient of the first training process; The calculation formula for the current state of the FSVR model is: In formula (3), is the loss function of the model in the passband frequency range in the i-th state, Var i (s) is the variance of the sample fuzzy weight s in the i-th state, which is used to quantify the uncertainty or discreteness of the data distribution; The expression of the epsilon-greedy strategy is: action i =π(state i )(4); In formula (4), action is the action output by the DQN model, and state is the current state; The expression for updating the FSVR weight parameter s is: In formula (5), is the FSVR weight parameter before updating, action·adjustment is the action update; The new prediction and error are calculated as: The calculation formula for immediate return is: In formula (7), α is the weight coefficient used to adjust the contribution of loss improvement to reward, and β is the weight coefficient used to adjust the contribution of fuzzy weight variance change to reward; The expression for storing the transfer information into the experience recycling pool D is: D←D∪(state i ,action i ,reward i ,state i+1 )(8); The expression for collecting small batch data N is: (s,a,r,s′) N ∈D (9); In formula (9), s, a, r, s′ are the current state, current action, immediate reward, and next state respectively; The expression for updating the original Q value network parameters is: In formula (10), L(θ) is the loss function; The expression for updating the target Q value network parameters is: θ′←vθ+(1-v)θ′ (11); In formula (11), v is the update proportional coefficient.

Images