Local oscillator phase drift prediction compensation method based on machine learning
Through the local oscillator phase drift prediction and compensation method based on machine learning, the linear regression model is used to predict the phase difference in real time and compensate, which solves the signal amplitude distortion caused by the factor local oscillator phase drift in the broadband interleaved sampling architecture, and realizes that users do not need to wait real-time calibration and high accuracy prediction.
Patent Information
- Application Number
- CN202411175861.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-26
- Publication Date
- 2024-12-03
AI Technical Summary
In the prior art, in the broadband interleaved sampling architecture, the subband decomposition filter causes adjacent subband overlapping frequency bands due to the non-ideal characteristics of the hardware, and introduces phase deviation. The traditional phase drift monitoring compensation requires the user to pause the use of equipment for calibration, and the factor cannot be solved in real time. The problem of signal amplitude distortion caused by local oscillator phase drift.
Using the local oscillator phase drift prediction and compensation method based on machine learning, the temperature data and phase difference calibration values are obtained in real time, and the phase difference is predicted using a linear regression model. If the predicted value exceeds the threshold, the overlap band phase initialization calibration will be initiated and the model will be updated in real time. Parameters to achieve real-time compensation without the need for users to pause the use of the device.
Improve the reduction of user waiting times, realize accurate prediction of phase difference, avoid instrument measurement deviation caused by phase drift, have high prediction accuracy of multiple regression methods and universality of the model, and are suitable for different times and equipment. Initial state.
Smart Images

Figure CN119066626A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of communications, and in particular to a local oscillator phase drift prediction and compensation method based on machine learning. Background Art
[0002] High-speed digital storage oscilloscopes based on broadband interleaved sampling architecture can break through the limitations of existing commercial ADC chip bandwidth and sampling rate, and use multiple low-speed ADC chips to achieve waveform capture functions under higher analysis bandwidth and sampling rate by interleaving and splicing in time and frequency domains; on the premise of meeting the same core indicators of high-speed direct sampling ADC chip architecture, broadband interleaved sampling architecture can reduce the R&D pressure of the whole machine itself to a certain extent. Figure 1 As shown in the figure, the broadband interleaved sampling technology uses sub-band decomposition filtering and analog mixing filtering to first down-convert the input signal to the corresponding low-frequency band, so as to meet the low-performance ADC chip's distortion-free sampling of the input signal. Then, the target sampling effect of the original input signal is restored through difference filtering, up-conversion filtering and sub-band splicing.
[0003] Sub-band decomposition filters will cause overlapping frequency bands in adjacent sub-bands due to their non-ideal hardware characteristics. The same-frequency signals in the overlapping frequency bands will enter the two adjacent sub-bands at the same time, and phase deviations will be introduced due to the differences in hardware architecture and characteristics of different sub-bands. In order to ensure that the amplitude of the signal after reconstruction and splicing at the overlapping band of the two sub-bands is distortion-free, the traditional overlapping band phase compensation is generally based on the elimination of the phase difference between sub-bands that does not change with time based on the cascaded integer delay / drop point, fractional delay FIR filter and nonlinear phase compensation IIR filter. However, the irrational characteristics of the device itself will cause the phase difference between sub-bands to change with time and produce phase drift. The main reason is that the local oscillator signal is not homologous and the internal voltage of the phase-locked loop used to lock the initial phase of the analog local oscillator is seriously affected by temperature. Considering the traditional phase drift monitoring and compensation, it is necessary to estimate the phase difference between the two sub-bands in real time. However, the estimation of the phase difference requires the user to suspend the use of the device, collect the calibration signal and calculate the phase difference. Therefore, how to solve the problem of amplitude distortion of the signal after reconstruction and splicing at the overlapping band caused by the phase drift of the sub-band local oscillator is currently needed. Summary of the invention
[0004] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a local oscillator phase drift prediction and compensation method based on machine learning to solve the deficiencies of the prior art.
[0005] The object of the present invention is achieved by the following technical solution: a local oscillator phase drift prediction and compensation method based on machine learning, the prediction and compensation method comprising:
[0006] After the system is powered on, an overlap band phase initialization calibration step is performed, and the overlap band phase initialization calibration step includes inter-subband phase difference estimation and overlap band phase calibration;
[0007] The temperature data acquired in real time and the phase difference calibration value acquired at irregular intervals are input into the model to obtain the phase difference prediction value; if the phase difference prediction value is greater than the preset threshold, the overlapping band phase initialization calibration step is started to perform phase compensation between sub-bands, and the b parameter of the model is updated and corrected in real time to achieve secondary calibration of the model;
[0008] If the predicted phase difference value is not greater than the preset threshold, the predicted phase difference value continues to be predicted.
[0009] The step of estimating the phase difference between subbands comprises:
[0010] The frequency interval of the overlapping band is determined by the amplitude-frequency response characteristics of the transition band between two adjacent sub-bands, and the frequency point where the maximum amplitude-frequency difference of the same frequency point of the two sub-bands is PdB is set as the boundary frequency point of the overlapping band;
[0011] The system is input with a sinusoidal signal of a frequency point within the overlapping frequency range. After the i-th sub-band is up-mixed and filtered, the output signal of the j-th frequency point is: Where i = 1, ..., S, S is the number of subbands in the system framework, j = 1, ..., P, P is the number of frequency points in the overlapping frequency interval, A i,j is the output signal amplitude, ω j is the angular frequency of the input signal, k is the number of sampling points, is the initial phase of the simulated local oscillator for the i-th subband, is the initial phase of the digital local oscillator of the i-th sub-band, is the initial phase of the input signal, λ i,j is the delay introduced by the j-th frequency signal passing through the i-th sub-band path device;
[0012] The first sub-band does not require mixing operation, and the initial phase of the relevant local oscillator is zero, that is, and Then the actual phase difference introduced by the j-th frequency signal of the two sub-bands is in, is the phase difference caused by the initial phase asynchrony between the digital and analog local oscillators, λ i+1,j -λ i,j is the frequency-dependent nonlinear and linear phase difference.
[0013] The overlapping band phase calibration step comprises:
[0014] The integer loss value m is calculated by linearly fitting the phase difference curve of the frequency points in the overlapping band interval int , the phase difference of the frequency points in the overlapping band frequency domain interval after integer loss is obtained:
[0015] The formula Substitute into the formula Get Among them, λ i+1,j -λ i,j -m int ω j To remove the nonlinear phase difference of the overlapping frequency band after the linear part is removed;
[0016] The phase difference introduced by setting the inter-subband factor and the initial phase of the local oscillator to be asynchronous is if The system updates the initial phase of the digital local oscillator of the i-th subband to if The system updates the initial phase of the (i+1)th sub-band digital local oscillator to
[0017] The simultaneous real-time updating and correction of the b parameter of the model includes:
[0018] Set the parameter update strategy of model b to Among them, y standard is the actual value of the phase difference between the corresponding frequency sub-bands after the digital local oscillator is updated, that is, the phase difference calibration value. Represents the phase difference calibration to y standard The phase difference prediction value at time, Δb is the change in parameter b, b new is the final value after update.
[0019] The prediction compensation method also includes collecting data to build a model and training it, which includes:
[0020] Given a data set D = {(x 11 ,x 12 ,…,x 1M ,y 1 ) 1 ,…(x N1 ,x N2 ,…,x NM ,y N ) N}, where M represents the number of eigenvalues, m = 1, 2, 3, ..., M, N represents the number of samples, n = 1, 2, 3, ..., N, the single sample temperature data in the data set (x n1 ,x n2 ,…,x nM ) is the RF board temperature collected in real time by M temperature sensors located at different positions on the RF board, y n The phase difference data calculated for the corresponding temperature samples;
[0021] Learn a linear model that accurately reflects the corresponding relationship between input and output from a given data set, that is, nm The function of the linear combination of is f n (x n1 ,x n2 ,…,x nM )=w 1 x n1 +w 2 x n2 +…+w M x nM +b, called f n (x n1 ,x n2 ,…,x nM ) is y n Yes (x n1 ,x n2 ,…,x nM ), the linear regression function is represented by the vector in, Indicates the weight of the corresponding attribute in the prediction result, b is the model b parameter, is the feature input, [·] T is the transpose of the vector; f n (x n ) is the prediction function;
[0022] By formula Solve the model parameters to minimize the loss function J(w,b), where:
[0023] The prediction compensation method includes the step of evaluating the model, which includes:
[0024] Coefficient of determination And the root mean square error As an evaluation index, the coefficient of determination R 2 The maximum value is 1. The closer it is to 1, the better the model's goodness of fit. The smaller the RMSE, the better the model's predictive ability. is the regression fitting value, Yes n The average value of .
[0025] The present invention has the following advantages: a local oscillator phase drift prediction and compensation method based on machine learning, which reduces the number of waiting times for users and does not require users to suspend use when the phase difference does not exceed a threshold; the accuracy is improved, and real-time prediction can be performed to avoid the situation where the phase drift is too large and the instrument measurement deviation occurs; a multivariate regression method is adopted, and the accuracy of the prediction applied by the model has a higher prediction accuracy than that of the linear fitting using a single temperature; a secondary calibration of the model is performed, which has universal applicability and can be used at different times and in different initial states of the device after offline training. BRIEF DESCRIPTION OF THE DRAWINGS
[0026] Figure 1 This is a schematic diagram of the broadband cross-sampling system architecture;
[0027] Figure 2 The figure is a schematic diagram of the overall process;
[0028] Figure 3 It is a single temperature and phase difference relationship diagram;
[0029] Figure 4 This is a schematic diagram of the effect of irregular calibration (correction);
[0030] Figure 5 This is a schematic diagram of the timing correction effect;
[0031] Figure 6 This is a schematic diagram of the effect when the calibration time interval is greater than 15 minutes. DETAILED DESCRIPTION
[0032] In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. The components of the embodiments of the present application described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the detailed description of the embodiments of the present application provided below in conjunction with the drawings is not intended to limit the scope of protection of the application claimed for protection, but merely represents the selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without making creative work belong to the scope of protection of the present application. The present invention is further described below in conjunction with the drawings.
[0033] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a local oscillator phase drift prediction and compensation method based on machine learning. The model trained by machine learning predicts the change of the phase difference between two sub-bands over time in real time. Once the predicted phase difference is found to exceed the preset threshold value, the system will activate the digital local oscillator coefficient update module and calibrate the phase drift deviation by changing the initial phase value of the digital local oscillator.
[0034] like Figure 2 As shown, first the system is turned on to perform overlap band initialization calibration. As time changes, the phase difference between the two sub-bands drifts, which eventually causes data distortion after the sub-bands are spliced. Since the phase drift is strongly correlated with the temperature change of the RF device, the present invention installs a temperature sensor in the system to collect the temperature data of the RF device in real time, and calculates the overlap band phase difference at the same time. Using the collected and calculated data, the relationship model between temperature and phase difference is trained offline. The phase drift is predicted in real time using the trained model and the real-time collected RF device temperature. During this period, if the predicted phase difference value is greater than the threshold set in advance, the system will perform phase compensation between sub-bands through overlap band phase calibration. However, the update of the digital local oscillator, including the update of the digital local oscillator during automatic calibration at startup, may cause some parameters in the offline trained model to be incompatible with the current machine state, so it is necessary to use the real phase difference data calculated when the system digital local oscillator is updated and the temperature at the corresponding time to perform secondary calibration on the model parameters.
[0035] Specifically include the following:
[0036] 1. Overlap band phase initialization calibration;
[0037] (1) Inter-subband phase difference estimation;
[0038] The overlapping frequency range is determined by the amplitude-frequency response characteristics of the transition band between two adjacent sub-bands. According to engineering experience, the frequency point where the maximum amplitude frequency difference between the two sub-bands is 25dB can be set as the overlapping frequency point. The system is input with a sinusoidal signal of a frequency point within the overlapping frequency range. After the i-th sub-band is up-mixed and filtered, the output signal of the j-th frequency point is:
[0039]
[0040] Where i=1,…,S, where S is the number of subbands in the system framework, j=1,…,P, where P is the number of frequency points in the overlapping frequency interval, A i,j is the output signal amplitude, ω j is the angular frequency of the input signal, k is the number of sampling points, is the initial phase of the simulated local oscillator for the i-th subband, is the initial phase of the digital local oscillator of the i-th sub-band, is the initial phase of the input signal, λ i,j is the delay introduced by the j-th frequency signal passing through the i-th sub-band path device. Figure 1 As shown, the first sub-band does not require mixing operation, and the initial phase of the relevant local oscillator can be set to zero, such as and According to formula (1), the actual phase difference introduced by the j-th frequency signal of the two sub-bands is:
[0041]
[0042] in, is the phase difference caused by the initial phase asynchronism of the digital and analog local oscillators, which has nothing to do with the frequency. i+1,j -λ i,j is the frequency-dependent nonlinear and linear phase difference.
[0043] (2) Overlap band phase calibration;
[0044] The actual phase difference after mixing in the overlapping band is Δ i,i+1 The compensation can generally be completed based on the design of cascaded integer delay / loss point, fractional delay FIR filter and nonlinear phase compensation IIR filter. This method design requires the addition of integer loss point modules, FIR and IIR filters in each sub-band to complete the phase calibration between adjacent sub-bands of the system. However, in actual applications, the system design will add an all-pass filter to perform nonlinear phase calibration on the full bandwidth of the system after the sub-band splicing full-pass amplitude-frequency calibration. Considering that the purpose of system overlapping band phase compensation is mainly to make the phase difference of each frequency point between adjacent sub-bands away from the -pi / pi area, and at the same time, in order to reduce the complexity of system design, the overlapping band phase calibration can be simplified to phase compensation with only an integer loss point module. However, since the initial phase of the first frequency point in the overlapping band frequency domain interval is different each time the power is turned on, the integer loss point method designed based on this method will introduce phase difference fluctuations of each frequency point after the overlapping band phase calibration. Therefore, the present invention first calculates the integer loss point value m by linearly fitting the frequency point phase difference curve in the overlapping band interval. int The phase difference of the frequency points in the overlapping frequency domain interval after integer loss is:
[0045]
[0046] Substituting formula (2) into formula (3) yields:
[0047]
[0048] Among them, λ i+1,j -λ i,j -m int ω j In order to remove the nonlinear phase difference of the overlapping band frequency point after the linear part, this nonlinear phase difference can be compensated by the all-pass filter after the sub-band splicing; The phase difference caused by the asynchronism of the digital and analog local oscillators can be eliminated by changing the initial phase value of the sub-band digital local oscillator. The phase difference introduced between sub-bands due to the asynchronism of the initial phase of the digital and analog local oscillators is defined as:
[0049]
[0050] if The system needs to update the initial phase of the digital local oscillator of the i-th sub-band to:
[0051]
[0052] if The system needs to update the initial phase of the (i+1)th sub-band digital local oscillator to:
[0053]
[0054] Due to the non-ideal characteristics of the RF device itself, the initial phase of the analog local oscillator will drift over time. It is a variable that changes with time, which will cause the phase difference between the calibrated sub-bands to change with time, thereby affecting the waveform distortion phenomenon in the frequency domain of the overlapping band after the sub-bands are spliced. To solve this problem, the present invention proposes a method for predicting and compensating the phase drift of an analog local oscillator based on machine learning. By predicting the phase drift, the system performs automatic digital local oscillator updates at irregular intervals to achieve the purpose of phase drift compensation.
[0055] 2. Phase drift prediction;
[0056] (1) Data collection relationship analysis;
[0057] like Figure 3 As shown in the figure, a scatter plot of actual data is drawn with a single temperature as the horizontal axis and phase difference as the vertical axis. From the figure, we can see that the two have a strong linear relationship. Therefore, the linear regression model in machine learning can be used to establish the relationship between temperature and phase difference.
[0058] (2) Model construction and training;
[0059] Given a data set D = {(x 11 ,x 12 ,…,x 1M ,y 1 ) 1 ,…(x N1 ,x N2 ,…,x NM ,y N ) N}, where M represents the number of eigenvalues, m = 1, 2, 3, ..., M, N represents the number of samples, n = 1, 2, 3, ..., N, the single sample temperature data in the data set (x n1 ,x n2 ,…,x nM ) is the RF board temperature collected in real time by M temperature sensors located at different positions on the RF board, y nis the phase difference data (i.e., true value) calculated for the corresponding temperature sample. From this data set, learn a linear model that reflects the corresponding relationship between input and output as accurately as possible. The linear model here is the attribute x nm The linear combination function can be expressed by formula (8):
[0060] f n (x n1 ,x n2 ,…,x nM )=w 1 x n1 +w 2 x n2 +…+w M x nM +b, (8)
[0061] Can be called f n (x n1 ,x n2 ,…,x nM ) is y n Yes (x n1 ,x n2 ,…,x nM ), whose vector representation is:
[0062]
[0063] in and b are model parameters, w represents the weight of the corresponding attribute in the prediction result, and b is the model b parameter; is the feature input, [·] T is the transpose of the vector; f n (x n ) is the prediction function. For the data set D, we need to make f n (x n ) is close to y n , that is, the model parameters are calculated through formula (10) to minimize the loss function J(w,b).
[0064]
[0065] in The solution method of formula (10) can adopt the standard gradient descent method. According to the above model description, the relationship between temperature and phase difference of the present invention can collect temperature phase difference data of multiple days for offline training, and finally obtain the optimal model. After experimental judgment, the best prediction effect is achieved when the number of temperatures M is 8. The multiple linear regression model (MLR) constructed using the above data input is used to train the model to obtain the best prediction result, and the best model is saved for subsequent real-time prediction.
[0066] (3) Model evaluation;
[0067] The prediction fitting performance of the present invention is based on the coefficient of determination (R 2 ) and root mean square error (RMSE) were used as evaluation indicators.
[0068]
[0069] In formulas (11) and (12), is the regression fitted value, y n is the true value, Yes n The average value of (n represents the nth sample point, n∈(0,N)). Generally speaking, in regression analysis, the coefficient of determination R 2 The maximum value is 1. The closer it is to 1, the better the model fit is; the smaller the RMSE, the better the predictive ability of the model.
[0070] 3. Second calibration of prediction model;
[0071] Each time the device is calibrated, the model parameters also need to be adjusted. This is because each calibration is equivalent to shifting the phase trend with temperature drift to the standard initial phase difference, which will cause the intercept parameter (i.e., b parameter) of the model to change. The device needs to perform model calibration in the following two scenarios: a) after the initial DBI calibration performed at each startup; b) after the digital local oscillator update calibration caused by the predicted phase difference drift exceeding the specified threshold.
[0072] In order to maintain the accuracy of model prediction, the present invention obtains the b parameter of the real-time update model. The b parameter update strategy is as follows:
[0073]
[0074] b new =b+Δb, (14)
[0075] y standardis the actual value of the phase difference between the corresponding frequency sub-bands after the digital local oscillator is updated. Represents the phase difference calibration to y standard The phase difference prediction value at time, Δb is the change in parameter b, b new is the final value after update.
[0076] Furthermore, if Figure 2 As shown in FIG. 1 , if the predicted phase difference value is greater than the pre-set threshold, the system will automatically activate the digital local oscillator update module to perform inter-subband phase compensation. The range of the phase drift threshold here is determined according to the following two criteria:
[0077] (1) According to engineering experience, the amplitude flatness of the acquisition system after splicing fluctuates within ±1dB, which can be said to have good amplitude flatness. According to formula (5), the phase difference caused by the initial phase drift of the digital-analog local oscillator is Assume that the phase difference of each frequency point in the overlapping band after system calibration is Where i is the subband number and j is the frequency point number. Therefore, the amplitude-frequency response value after subband splicing caused by phase drift can be calculated as:
[0078]
[0079] Among them, A i (ω j ),A i+1 (ω j ) are the low-frequency sub-band and the high-frequency sub-band in the overlapping band at the frequency ω j The amplitude-frequency response value at . for The amplitude-frequency response value of the corresponding frequency point after the system overlap band calibration. j The amplitude fluctuation at the frequency point can be expressed by the following formula:
[0080]
[0081] Consider all frequencies in the overlap band where ω L ,ω H are the upper and lower bounds of the overlapping band frequency points. By changing the formula (15) The size of the value needs to be found The maximum value of , so that:
[0082]
[0083] The optimal solution at this time is the maximum boundary value of the phase drift threshold.
[0084] (2) To ensure that the system does not amplify noise after sub-band splicing, the following phase difference must be met: in is the local oscillator drift variable, and its actual allowable drift range is:
[0085]
[0086] Define the phase drift threshold as Θ, and the above two criteria must be met at the same time, namely:
[0087]
[0088] This can ensure good amplitude flatness and prevent the noise from being amplified after splicing.
[0089] The test results of the present invention are as follows:
[0090] (1) Evaluation of prediction effect;
[0091] The equipment used for the test is three high-speed digital storage oscilloscopes with a sampling rate of 80GSa / s and a bandwidth of 18GHz, which collected temperature and phase difference data for multiple days. The model prediction effect was verified on three machines using R 2 The two indicators of R and RMSE are used for evaluation, and the results are shown in Table 1. The training data and the validation data set are from different days. 2 The average is 0.9988, indicating that the model has a good goodness of fit. The RMSE on the training set is low, indicating that the prediction accuracy is high. The RMSE on the validation set is also excellent, indicating that the model has good generalization ability.
[0092] Table 1. Model evaluation table
[0093]
[0094] (2) Comparison with the single temperature effect;
[0095] The experiment used multiple temperatures (1 to 8 temperatures) as input and used four days of temperature phase difference data for verification. Table 2 shows the RMSE of the model trained with different temperatures. It can be seen that if a single temperature is used for fitting, there will be a large prediction error. When the number of temperatures is 8, the prediction error is the smallest.
[0096] Table 2. Prediction RMSE table at different temperatures
[0097]
[0098] (3) Compare with the timing calibration scheme;
[0099] Take the data of device 2 as an example. According to the above two criteria, it can be calculated that the phase drift threshold of device 2 is set to Θ = [-1.5, 1.5] (rad / s). The timing calibration scheme is: set the calibration time, and the machine detects and calibrates the phase difference according to the fixed calibration time. The following compares this scheme with the timing calibration scheme from the two aspects of taking the same number of calibrations and limiting the number of calibrations for irregular calibrations. And the error time is used as an indicator to measure the advantages and disadvantages of the two schemes. The error time is defined as the time when the phase difference drifts out of the allowed range Θ during the user's use time period.
[0100] 1) The same calibration times are used for scheduled calibration and irregular calibration;
[0101] like Figure 4 The calibration results under the untimed calibration scheme. The original uncalibrated phase drift curve is the dotted line in the figure. The real phase curve and the predicted phase curve after adopting this scheme are the solid line and the line with the '*' in the figure, respectively. The two lines basically coincide. The calculated RMSE = 0.1128, and the prediction accuracy is high. The vertical dotted line in the figure represents the calibration at this moment. A total of 19 untimed calibrations were performed within 14 hours of machine operation, and there was no phase difference drift exceeding the threshold, that is, the error time was 0.
[0102] Figure 5 In order to use the calibration results of the scheduled calibration scheme in the same time, the calibration interval is set to 40 minutes under the condition of using the same number of calibrations as the unscheduled calibration. The error time at this time is 1 hour, 55 minutes and 53 seconds.
[0103] 2) When the calibration time difference of irregular calibration is limited;
[0104] from Figure 4 As can be seen in Figure 1, irregular calibrations are too frequent in certain time periods, which can cause a poor user experience. To avoid frequent calibrations in a short period of time, the time difference between two calibrations must not be less than 15 minutes. Figure 6 This is a calibration effect diagram of irregular calibration with a time difference of no less than 15 minutes between two calibrations. The error duration of this irregular calibration scheme with a limited calibration time interval is 36 minutes and 31 seconds, and the number of calibrations is 14. Compared with regular calibration, it has less error time and lower calibration frequency.
[0105] Table 3. Comparison of the irregular calibration scheme and the regular calibration scheme
[0106]
[0107] Combining the above two situations, the error time and correction times of irregular calibration are superior to those of regular calibration.
[0108] The above is only a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the form disclosed herein, and should not be regarded as excluding other embodiments, but can be used for various other combinations, modifications and improvements, and can be modified within the scope of the concept described herein through the above teachings or the technology or knowledge of the relevant field. The changes and modifications made by those skilled in the art do not deviate from the spirit and scope of the present invention, and should be within the scope of protection of the claims attached to the present invention.
Claims
1. A local oscillator phase drift prediction and compensation method based on machine learning, characterized in that: The prediction compensation method comprises: After the system is powered on, an overlap band phase initialization calibration step is performed, and the overlap band phase initialization calibration step includes inter-subband phase difference estimation and overlap band phase calibration; The temperature data acquired in real time and the phase difference calibration value acquired at irregular intervals are input into the model to obtain the phase difference prediction value; If the phase difference prediction value is greater than the preset threshold, the overlapping band phase initialization calibration step is started to perform inter-subband phase compensation, and the b parameter of the model is updated and corrected in real time to achieve secondary calibration of the model; If the predicted phase difference value is not greater than the preset threshold, the predicted phase difference value continues to be predicted.
2. The local oscillator phase drift prediction and compensation method based on machine learning according to claim 1, characterized in that: The step of estimating the phase difference between subbands comprises: The frequency interval of the overlapping band is determined by the amplitude-frequency response characteristics of the transition band between two adjacent sub-bands, and the frequency point where the maximum amplitude-frequency difference of the same frequency point of the two sub-bands is PdB is set as the boundary frequency point of the overlapping band; The system is input with a sinusoidal signal of a frequency point within the overlapping frequency range. After the i-th sub-band is up-mixed and filtered, the output signal of the j-th frequency point is: Where i = 1, ..., S, S is the number of subbands in the system framework, j = 1, ρ, P, P is the number of frequency points in the overlapping frequency interval, A i,j is the output signal amplitude, ω j is the angular frequency of the input signal, k is the number of sampling points, is the initial phase of the simulated local oscillator for the i-th subband, is the initial phase of the digital local oscillator of the i-th sub-band, is the initial phase of the input signal, λ i,j is the delay introduced by the j-th frequency signal passing through the i-th sub-band path device; The first sub-band does not require mixing operation, and the initial phase of the relevant local oscillator is zero, that is, and Then the actual phase difference introduced by the j-th frequency signal of the two sub-bands is in, is the phase difference caused by the initial phase asynchrony between the digital and analog local oscillators, λ i+1,j -λ i,j is the frequency-dependent nonlinear and linear phase difference.
3. The local oscillator phase drift prediction and compensation method based on machine learning according to claim 2, characterized in that: The overlapping band phase calibration step comprises: The integer loss value m is calculated by linearly fitting the phase difference curve of the frequency points in the overlapping band interval int , the phase difference of the frequency points in the overlapping band frequency domain interval after integer loss is obtained: The formula Substitute into the formula Get Among them, λ i+1,j -λ i,j -m int ω j To remove the nonlinear phase difference of the overlapping frequency band after the linear part is removed; The phase difference introduced by setting the inter-subband factor and the initial phase of the local oscillator to be asynchronous is if The system updates the initial phase of the digital local oscillator of the i-th subband to if The system updates the initial phase of the (i+1)th sub-band digital local oscillator to 4. The local oscillator phase drift prediction and compensation method based on machine learning according to claim 1, characterized in that: The simultaneous real-time updating and correction of the b parameter of the model includes: Set the parameter update strategy of model b to b new =b+Δb, where y standard is the actual value of the phase difference between the corresponding frequency sub-bands after the digital local oscillator is updated, that is, the phase difference calibration value. Represents the phase difference calibration to y standard The phase difference prediction value at time, Δb is the change in parameter b, b new is the final value after update.
5. A local oscillator phase drift prediction and compensation method based on machine learning according to any one of claims 1 to 4, characterized in that: The prediction compensation method also includes collecting data to build a model and training it, which includes: Given a data set D = {(x 11 ,x 12 ,…,x 1M ,y1)1,…(x N1 ,x N2 ,…,x NM ,y N ) N }, where M represents the number of eigenvalues, m = 1, 2, 3, ..., M, N represents the number of samples, n = 1, 2, 3, ..., N, the single sample temperature data in the data set (x n1 ,x n2 ,…,x nM ) is the RF board temperature collected in real time by M temperature sensors located at different positions on the RF board, y n The phase difference data calculated for the corresponding temperature samples; Learn a linear model that accurately reflects the corresponding relationship between input and output from a given data set, that is, nm The function of the linear combination of is f n (x n1 ,x n2 ,…,x nM )=w1x n1 +w2x n2 +…+w M x nM +b, called f n (x n1 ,x n2 ,…,x nM ) is y n Yes (x n1 ,x n2 ,…,x nM ), the linear regression function is represented by the vector in, Indicates the weight of the corresponding attribute in the prediction result, b is the model b parameter, is the feature input, [·] T is the transpose of the vector; f n (x n ) is the prediction function; By formula Solve the model parameters to minimize the loss function J(w,b), where:
6. The local oscillator phase drift prediction and compensation method based on machine learning according to claim 5, characterized in that: The prediction compensation method includes the step of evaluating the model, which includes: Coefficient of determination And the root mean square error As an evaluation index, the coefficient of determination R 2 The maximum value is 1. The closer it is to 1, the better the model's goodness of fit. The smaller the RMSE, the better the model's predictive ability. is the regression fitting value, Yes n The average value of .
Citation Information
Patent Citations
Phase difference based kNN fingerprint localization algorithm
CN108710121A
DBF radar multi-channel digital signal amplitude phase error calibration circuit based on FPGA
CN112230188A
SAR internal calibration signal processing method
CN113391278A
Method and device for generating phase difference compensation value based on artificial intelligence
CN117783996A
All-pass amplitude-frequency rapid calibration method based on overlapping band amplitude-frequency data replacement
CN118011300A
Cited By
Electric energy equipment receiving end phase angle deviation correction method based on Z-shaped winding
CN120611170A
Method for correcting phase angle deviation of power receiving end of power equipment based on zigzag winding
CN120611170B
Pore water pressure gauge for monitoring safety of dam
CN120947894A
Method, device, system and medium for predicting phase drift of a quantum chip
CN122693880A
