Filter passband fluctuation automatic compensation method and system

Abstract

The invention provides a filter passband fluctuation automatic compensation method and system. The method comprises the following steps: acquiring multi-dimensional tensor data of filter data; extracting features of the multi-dimensional tensor data to obtain a time-frequency domain joint feature matrix; the time-frequency domain joint feature matrix is input into a preset double-branch LSTM network, the preset double-branch LSTM network comprises a main branch and an auxiliary branch, a first adjustment parameter set and a hidden state vector are generated through the main branch, reverse correction is performed on the first adjustment parameter set according to the hidden state vector through the auxiliary branch, and compensation parameters are generated; performing dynamic coupling mapping according to the compensation parameters and preset physical constraint conditions to obtain a constrained parameter set; and controlling corresponding hardware to carry out passband fluctuation compensation on the filter based on the constrained parameter set.

Description

Technical Field

[0001] This application relates to the field of electronic technology, and in particular, to a method and system for automatically compensating the passband ripple of a filter. Background Art

[0002] A filter is an essential key component in an electronic system, and its performance directly affects the quality of signal processing and the stability of the system. In practical applications, the passband ripple of a filter may deviate due to factors such as temperature changes, component aging, manufacturing tolerances, and external interference, thereby affecting the flatness of its frequency response and the reliability of signal transmission. To solve this problem, existing technologies usually introduce active components (such as operational amplifiers, digital potentiometers, or voltage-controlled components) to dynamically adjust the parameters of the filter to achieve compensation for the passband ripple. The existing compensation circuit designs often have the following problems: 1. The cooperation between multiple circuit modules is relatively complex, and it is difficult to achieve automated collaborative work; 2. The parameter settings of each module usually rely on manual debugging, which is time-consuming and easily affected by human errors; 3. Traditional methods lack the ability to adapt to environmental changes in real time and cannot meet the application requirements of high precision and high stability. Summary of the Invention

[0003] This application provides a method and system for automatically compensating the passband ripple of a filter, which is used to automatically coordinate multiple circuit modules for filter passband compensation, reduce manual intervention, and perform adaptive passband ripple compensation to improve the overall performance and reliability of the filter.

[0004] In a first aspect, an embodiment of this application provides a method for automatically compensating the passband ripple of a filter. The method includes: Obtain multi-dimensional tensor data of the data processed by the filter; Extract the features of the multi-dimensional tensor data to obtain a time-frequency domain joint feature matrix; Input the time-frequency domain joint feature matrix into a preset dual-branch LSTM network. The preset dual-branch LSTM network includes: a main branch and a sub-branch. Generate a first set of adjustment parameters and a hidden state vector through the main branch, and through the sub-branch, reverse-correct the first set of adjustment parameters according to the hidden state vector to generate compensation parameters; Perform dynamic coupling mapping according to the compensation parameters and preset physical constraint conditions to obtain a set of constrained parameters; Control the corresponding hardware based on the set of constrained parameters to perform passband ripple compensation on the filter.

[0005] In a second aspect, an embodiment of this application provides a device for automatically compensating the passband ripple of a filter. The device for automatically compensating the passband ripple of a filter is used to execute the method for automatically compensating the passband ripple of a filter according to any one of the embodiments of this application. The device includes: A data acquisition module, configured to acquire multi-dimensional tensor data of a filter for data; A feature extraction module, configured to extract features of the multi-dimensional tensor data to obtain a joint time-frequency domain feature matrix; A compensation calculation module, configured to input the joint time-frequency domain feature matrix into a preset dual-branch LSTM network. The preset dual-branch LSTM network includes: a main branch and a secondary branch. The first adjustment parameter set and the hidden state vector are generated through the main branch. Through the secondary branch, the first adjustment parameter set is reversely corrected according to the hidden state vector to generate compensation parameters; A constraint mapping module, configured to perform dynamic coupling mapping according to the compensation parameters and preset physical constraint conditions to obtain a constrained parameter set; A hardware control module, configured to control corresponding hardware based on the constrained parameter set to perform passband ripple compensation on the filter.

[0006] In a third aspect, an embodiment of the present application provides an electronic device, where the electronic device includes a memory and a processor; The memory is used to store a computer program; The processor is configured to execute the computer program and, when executing the computer program, implement any one of the filter passband ripple automatic compensation methods in the embodiments of the present application.

[0007] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, where the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the processor is caused to implement any one of the filter passband ripple automatic compensation methods in the embodiments of the present application.

[0008] An embodiment of the present application provides a method for automatically compensating the passband ripple of a filter. The method includes: obtaining multi-dimensional tensor data of the data of the filter; extracting the features of the multi-dimensional tensor data to obtain a joint time-frequency domain feature matrix; inputting the joint time-frequency domain feature matrix into a preset dual-branch LSTM network. The preset dual-branch LSTM network includes: a main branch and a secondary branch. Through the main branch, a first set of adjustment parameters and a hidden state vector are generated. Through the secondary branch, according to the hidden state vector, the first set of adjustment parameters is reversely corrected to generate compensation parameters; dynamic coupling mapping is performed according to the compensation parameters and preset physical constraint conditions to obtain a set of constrained parameters; based on the set of constrained parameters, the corresponding hardware is controlled to perform passband ripple compensation on the filter. Through the above method, the data representation ability is improved through multi-dimensional tensor data fusion, and the passband ripple pattern is accurately captured; through the main branch of the dual-branch LSTM network, a set of adjustment parameters is generated, and the secondary branch reversely corrects the compensation parameters based on the hidden state vector, effectively suppressing the negative impact of parameter adjustment on group delay and insertion loss; through physical constraint conditions, the compensation parameters are dynamically mapped to avoid over-adjustment or device damage, and based on the set of constrained parameters, the hardware is directly driven to achieve real-time compensation ability with environmental parameter adaptability. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0010] Figure 1 It is a schematic flowchart of a method for automatically compensating the passband ripple of a filter provided by an embodiment of the present application; Figure 2 It is a schematic block diagram of a device for automatically compensating the passband ripple of a filter provided by an embodiment of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0011] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art without creative efforts based on the embodiments of the present invention belong to the scope of protection of the present invention.

[0012] The flowchart shown in the drawings is only an example illustration and does not necessarily include all the contents and operations / steps, nor does it necessarily need to be executed in the described order. For example, some operations / steps can also be decomposed, combined, or partially merged. Therefore, the actual execution order may be changed according to the actual situation.

[0013] It should also be understood that the terms used in the specification of this application are only for the purpose of describing specific embodiments and are not intended to limit this application. As used in the specification of this application and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include the plural forms.

[0014] It should be further understood that the term "and / or" used in the specification of this application and the appended claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0015] Currently, there are various methods for compensating bandpass filters: Adaptive filtering: An adaptive filter can automatically adjust its parameters according to the error between the input signal and the desired output. In the case of passband fluctuations, a control system can be designed to monitor the output of the filter and adjust the filter coefficients based on the difference between the actual output and the ideal output.

[0016] Feedback control: By introducing a feedback mechanism, the performance of the filter can be monitored in real time and a response can be made to any situation that deviates from the design specifications. For example, in a feedback loop, the actual transfer function of the filter can be measured and compared with the theoretical value, and then the filter elements (such as resistors, capacitors or the gain of operational amplifiers) can be adjusted to compensate for this deviation.

[0017] Temperature compensation: For components that are greatly affected by temperature, temperature compensation techniques can be adopted. This may include using components with opposite temperature coefficients, or integrating temperature sensors and correction circuits to automatically adjust the filter characteristics.

[0018] Digital compensation: In the digital domain, automatic compensation can be achieved through software algorithms. Digital filters can apply the latest correction values each time data is processed, and these correction values may be based on pre-established models or real-time measurement results.

[0019] Utilizing adjustable components: Using tunable electronic components, such as digital potentiometers or voltage-controlled oscillators (VCOs), the center frequency and bandwidth of the filter can be changed as needed, thereby achieving automatic compensation.

[0020] The above-mentioned bandpass filter compensations may exist simultaneously in a system with a very high degree of integration, but it is very difficult to coordinate them with each other. In extreme cases, there is also a situation of mutual interference. Therefore, a method is needed for comprehensive management to achieve automatic compensation adjustment.

[0021] Please refer to Figure 1 , Figure 1It is a schematic flowchart of a method for automatically compensating the passband ripple of a filter provided by an embodiment of the present application. As Figure 1 shown, the specific steps of the method for automatically compensating the passband ripple of the filter include: S101 - S105.

[0022] S101. Obtain the multi - dimensional tensor data of the data of the filter.

[0023] Exemplarily, through a multi - channel synchronous acquisition system, the frequency response, phase response, insertion loss, group delay, and environmental parameters of the filter are captured in real - time. The frequency response scans the amplitudes of each frequency point within the passband by a vector network analyzer and generates an amplitude - frequency characteristic curve matrix with a resolution of 0.1 dB. The phase response data extracts the phase offset through IQ demodulation technology, calculates the first - order derivative of the phase with respect to frequency based on the Hilbert transform, generates a group delay tensor, and integrates it into a phase - group delay joint matrix. The insertion loss synchronously measures the power values at the input and output ends by a high - precision power probe and calculates the logarithmic power ratio to generate an insertion loss matrix. The environmental parameters are collected by an embedded temperature - humidity sensor, and the component values are pre - corrected in combination with a linear temperature drift model to eliminate the basic environmental interference. The multi - source data realizes timestamp alignment through a hardware trigger signal. The high - frequency signal is down - sampled to match the sampling rate of the environmental parameters, and an outlier data segment is detected by using a sliding - window Mahalanobis distance model. The missing values are repaired by cubic spline interpolation, and a five - dimensional tensor structure is constructed, covering time, frequency, physical quantities (amplitude - frequency response, phase, group delay, insertion loss), and environmental variables (temperature, humidity), providing a complete data basis for subsequent analysis.

[0024] S102. Extract the features of the multi - dimensional tensor data to obtain a time - frequency domain joint feature matrix.

[0025] Exemplarily, in the frequency - domain analysis, calculate the range of the amplitude - frequency response within the passband to generate a ripple coefficient vector, which characterizes the passband ripple amplitude. Fit the slope of the amplitude - frequency curve by the least - squares method to generate a passband tilt index, quantifying the overall tilt trend. In the phase analysis, calculate the frequency - domain differential entropy of the phase offset to generate a phase non - linear index vector, reflecting the degree of phase distortion. Perform a sliding - window standard deviation statistic on the group delay tensor to generate a group delay stability parameter, detecting transient mutations. In the insertion - loss analysis, based on the abnormal interval marked by the group delay, calculate the root - mean - square value of the insertion - loss change rate to generate a transient - fluctuation energy vector. In the environmental coupling analysis, analyze the sensitivity coefficients of temperature and humidity to the insertion loss and phase, and construct an environmental coupling matrix. In the feature fusion stage, use the random forest algorithm to evaluate the importance of each feature, assign weights, and then align through t - SNE dimensionality reduction to generate a preliminary time - frequency domain joint feature matrix. Finally, dynamically activate relevant feature subsets according to the real - time environmental parameters (such as suppressing temperature - sensitive features at high temperatures), apply an exponentially decaying weight, and output an optimized time - frequency domain joint feature matrix.

[0026] S103. Input the time-frequency domain joint feature matrix into a pre-set dual-branch LSTM network. The pre-set dual-branch LSTM network includes: a main branch and a sub-branch. Generate a first set of adjustment parameters and a hidden state vector through the main branch. Through the sub-branch, reverse-correct the first set of adjustment parameters according to the hidden state vector to generate compensation parameters.

[0027] Exemplarily, the pre-set dual-branch LSTM network includes a main branch and a sub-branch. The main branch is composed of a bidirectional LSTM: the forward layer extracts historical time series dependencies, the backward layer captures future trends, and outputs a preliminary hidden state vector; the key frequency band features are dynamically weighted by combining environmental parameters through an attention mechanism to generate an enhanced set of hidden state vectors, and then the hidden state is output through residual connection and layer normalization processing. The main branch parallel fully-connected layer decodes to generate a first set of adjustment parameters, including capacitance, gain, phase, and dynamic range adjustment amounts. The sub-branch receives the hidden state vector of the main branch, constructs a Jacobian matrix to quantify the parameter side effects (such as the impact of gain adjustment on group delay), generates a second set of adjustment parameters through adversarial gradient backpropagation, combines the main and sub-branch parameters, and performs constrained optimization with the goal of minimizing side effects to generate a set of compensation parameters to ensure that the adjustment amount suppresses the passband ripple while avoiding causing secondary interference.

[0028] S104. Perform dynamic coupling mapping according to the compensation parameters and the pre-set physical constraint conditions to obtain a set of constrained parameters.

[0029] Exemplarily, impose physical constraints and hardware compatibility mapping on the compensation parameters. Adaptive adjustment needs to be performed for each parameter to prevent over-adjustment. The capacitance adjustment amount is constrained within the physical range of the adjustable capacitor array through clipping processing; the gain and phase adjustment amounts are fitted to the cross-coupling equation by the least squares method to enforce dynamic balance; the dynamic range adjustment amount is weighted and integrated in combination with the insertion loss frequency domain distribution and converted into a DAC control voltage signal. The environmental compensation module predicts the temperature change trend based on a Kalman filter and generates a heater power control signal to achieve prospective temperature drift compensation. After all parameters are verified in multiple dimensions (such as the compatibility check between capacitance adjustment and heating power), a set of hardware-executable constrained parameters is generated to ensure the stability and reliability of the system.

[0030] S105. Control the corresponding hardware based on the set of constrained parameters to perform passband ripple compensation on the filter.

[0031] Exemplarily, the constrained parameter set is converted into specific hardware control instructions: the digitally tunable capacitor array adjusts the capacitance value in steps of 0.1 pF; the 16-bit DAC module outputs a dynamic bias voltage to adjust the amplifier gain; the FPGA calculates the finite impulse response coefficients in real time to drive the phase rotator to correct the phase linearity; the distributed heater adjusts the temperature according to the predicted power. Multiple actuators synchronize the timing through the pre-emphasis algorithm to compensate for mechanical delays (such as relay switching) to ensure that the action synchronization error is ≤ 10 μs. Thus, the passband ripple is suppressed to ±0.1 dB, the group delay fluctuation is reduced by more than 50%, and the environmental adaptability is achieved.

[0032] The embodiment of the present application provides a method for automatically compensating the passband fluctuation of a filter. The method includes: obtaining the multi-dimensional tensor data of the data of the filter; extracting the features of the multi-dimensional tensor data to obtain the time-frequency domain joint feature matrix; inputting the time-frequency domain joint feature matrix into a preset dual-branch LSTM network. The preset dual-branch LSTM network includes: a main branch and a sub-branch. The first adjustment parameter set and the hidden state vector are generated through the main branch. Through the sub-branch, the first adjustment parameter set is reversely corrected according to the hidden state vector to generate the compensation parameter; dynamic coupling mapping is performed according to the compensation parameter and the preset physical constraint conditions to obtain the constrained parameter set; the corresponding hardware is controlled based on the constrained parameter set to perform passband fluctuation compensation on the filter. Through the above method, the data representation ability is improved through multi-dimensional tensor data fusion to accurately capture the passband fluctuation mode; the adjustment parameter set is generated through the main branch of the dual-branch LSTM network, and the sub-branch reversely corrects the compensation parameter based on the hidden state vector to effectively suppress the negative impact of parameter adjustment on the group delay and insertion loss; the compensation parameter is dynamically mapped through the physical constraint conditions to avoid over-adjustment or device damage, and the hardware is directly driven based on the constrained parameter set to achieve the real-time compensation ability with environmental parameter adaptability.

[0033] To more clearly introduce the technical solution of the present application, the technical solution of the present application will also be introduced through specific embodiments below. It should be noted that the specific embodiment is used to expand the description of the technical solution of the present application, rather than limiting the present application.

[0034] In some embodiments, obtaining the multi-dimensional tensor data of the data of the filter includes: S1011-S1016.

[0035] S1011. Collect the frequency response, phase response, insertion loss, group delay, and environmental parameters of the filter through multiple channels.

[0036] Exemplarily, a multi-sensor synchronous acquisition system is adopted to capture multi-dimensional physical quantity data of the filter in the operating state in real time. The frequency response is obtained by a vector network analyzer (VNA) in a sweep frequency mode, covering all frequency points within the passband, with a scanning step of 0.1 dB resolution to ensure the fineness of the amplitude-frequency characteristic curve. The phase response data synchronously acquires the I / Q two-channel baseband signals through a high-speed data acquisition card, and uses quadrature demodulation technology to extract the original phase information. The group delay data is indirectly calculated based on the frequency-domain differential characteristic of the phase response to avoid the hardware delay error caused by direct measurement. The insertion loss synchronously measures the RMS power values at the input and output ends by a high-precision power probe to ensure the instantaneous consistency of the power ratio calculation. The environmental parameters are collected through a distributed temperature and humidity sensor network, with a temperature measurement accuracy of ±0.1 °C and a humidity accuracy of ±2%RH to monitor the change of the filter working environment in real time. All data channels are strictly synchronized through a hardware trigger signal to eliminate the timestamp deviation.

[0037] S1012. Perform a sweep of the amplitude-frequency characteristic curve within the passband for the frequency response through a vector network analyzer to generate a frequency response matrix.

[0038] Exemplarily, the vector network analyzer (VNA) performs a sweep frequency test within a preset passband range, measures the transmission coefficient (S21 parameter) of the filter point by point to generate an amplitude-frequency characteristic curve. The test frequency interval is dynamically adjusted according to the passband width. For example, it is scanned with a 1 MHz step within a 100 MHz passband to form discrete frequency response points. Within each time window, the VNA samples the same frequency point multiple times and takes the average value to suppress random noise interference, and then constructs a frequency response matrix according to the time-frequency dimension. The rows of the matrix correspond to the time series, the columns correspond to the frequency points, and the element values are the normalized amplitude values (unit: dB), providing a standardized input for subsequent feature extraction.

[0039] S1013. Perform IQ demodulation on the phase response data to determine the phase offset, calculate the derivative of the phase with respect to frequency of the group delay through Hilbert transform to obtain a group delay tensor, and generate a phase-group delay joint matrix according to the phase offset and the group delay tensor.

[0040] Exemplarily, perform unwrapping processing on the original phase data obtained by IQ demodulation to eliminate phase jumps and generate a continuous phase offset sequence. Based on the phase-frequency relationship, calculate the first derivative of the phase with respect to frequency through Hilbert transform to obtain a group delay tensor. Specifically, calculate the difference value of the phase data at each frequency point within each time window to generate an instantaneous group delay curve. Align the phase offset matrix and the group delay tensor according to the frequency-time dimension and merge them into a phase-group delay joint matrix. This matrix simultaneously retains the phase linearity and group delay stability information, providing a data basis for subsequent analysis of phase distortion and time delay fluctuation.

[0041] S1014. Measure the RMS power values at the input and output ends for insertion loss through a power probe, calculate the power ratio of the RMS power values at the input and output ends, and generate an insertion loss matrix.

[0042] Exemplarily, the power probe synchronously acquires the RMS power values at the input and output ends, and the sampling rate is synchronized with the frequency response scan to ensure data timing consistency. The insertion loss calculation adopts the form of logarithmic power ratio, that is, for the input power P in and the output power P out calculate the decibel value IL, specifically: , and generate a decibel value matrix. To eliminate instantaneous interference, perform a moving average filter on the insertion loss at the same frequency point within a time window. The insertion loss matrix reflects the change in the transmission efficiency of the filter at different frequencies and provides a key indicator for transient fluctuation analysis.

[0043] S1015. Perform temperature and humidity drift modeling and linear compensation on environmental parameters to generate an environmental compensation parameter matrix.

[0044] Exemplarily, after preprocessing the temperature and humidity sensor data, construct a linear compensation model. For example, the temperature drift compensation formula for capacitance value is ΔC = α(T - T0) + β(H - H0), where T0 and H0 are the calibration environmental reference values, and α, β are the material characteristic coefficients. Determine the coefficients by least squares fitting of historical data and perform compensation calculation on the real-time acquired temperature and humidity values to generate an environmental compensation parameter matrix. This matrix contains the pre-correction amounts of each component under different environmental conditions and is used to eliminate the basic offset of temperature and humidity on the filter performance.

[0045] S1016. Synthesize multi-dimensional tensor data according to the frequency response matrix, phase-group delay joint matrix, insertion loss matrix, and environmental compensation parameter matrix.

[0046] Exemplarily, align the frequency response matrix, phase-group delay joint matrix, insertion loss matrix, and environmental compensation parameter matrix according to a unified time-frequency reference to construct a five-dimensional tensor data structure. The tensor dimensions include time stamp, frequency point, physical quantities (amplitude-frequency response, phase, group delay, insertion loss), and environmental variables (temperature, humidity compensation value). Missing data is filled by cubic spline interpolation, and outliers are detected and replaced based on the Mahalanobis distance model. Thus, the tensor data is stored in a hierarchical form to support efficient access and analysis by subsequent feature extraction modules.

[0047] In some embodiments, extract the features of the multi-dimensional tensor data to obtain a time-frequency domain joint feature matrix, including: S1021 - S1027.

[0048] S1021. Calculate the range between the maximum amplitude and the minimum amplitude of the frequency response matrix within a preset passband to generate a ripple coefficient vector, and perform linear regression fitting on the amplitude-frequency characteristic curve corresponding to the frequency response matrix to generate a passband slope index vector.

[0049] For example, for each time window in the frequency response matrix, the amplitude data of all frequency points in the passband are traversed, the maximum and minimum values ​​are searched and the difference is calculated to generate a ripple coefficient vector to quantify the passband fluctuation amplitude. Subsequently, a linear regression analysis is performed on the amplitude-frequency characteristic curve, and the slope of the fitting line is used as the passband inclination index to characterize the overall offset trend of the passband (such as upward or downward tilt). The inclination calculation eliminates interference from abnormal frequency points to ensure that the fitting results reflect the real physical characteristics. The two indicators jointly describe the static and dynamic changes of the passband amplitude-frequency characteristics.

[0050] S1022. Calculate the phase nonlinearity index vector of the phase-group delay joint matrix by frequency domain differential entropy.

[0051] Exemplarily, based on the phase data in the phase-group delay joint matrix, the phase offset at each frequency point is differentiated in the frequency domain to obtain the phase change rate distribution. By statistically analyzing the distribution discreteness of the phase change rate (such as variance or entropy value), a phase nonlinearity index vector is generated. This index quantifies the degree of nonlinear distortion of the phase response. A high entropy value indicates that the phase changes irregularly with frequency, which may cause signal distortion and provide a key basis for subsequent phase compensation.

[0052] S1023. Perform sliding window standard deviation statistics on the group delay tensor to generate a group delay stability parameter vector.

[0053] For example, the group delay tensor is divided into sliding windows in the time dimension, and the standard deviation of the group delay value in each window is calculated to generate a stability parameter vector. The larger the standard deviation, the more drastic the fluctuation of the group delay over time, which may cause signal delay jitter. At the same time, the window with a sudden change in the standard deviation is marked as a transient anomaly interval, which triggers subsequent transient fluctuation analysis to ensure targeted processing of sudden anomalies.

[0054] S1024. Based on the sliding window, extract the transient response characteristics of the insertion loss matrix, calculate the root mean square value of the insertion loss change rate, and generate a transient fluctuation energy vector.

[0055] Exemplarily, within the transient abnormal interval marked in step S1023, the root mean square value (RMS) of the insertion loss change rate is calculated for the insertion loss matrix to generate a transient fluctuation energy vector. This energy value reflects the degree of violent fluctuation of signal power during the transient process, and high energy values ​​need to be suppressed by dynamic range adjustment. Overlapping sliding window smoothing is used during calculation to avoid misjudgment caused by noise interference.

[0056] S1025. Perform temperature and humidity coupling analysis on the environmental compensation parameter matrix to generate an environmental coupling feature matrix.

[0057] Exemplarily, analyze the temperature and humidity data in the environmental compensation parameter matrix, calculate the sensitivity coefficient of temperature to insertion loss (such as the change in insertion loss caused by a unit temperature change) and the sensitivity coefficient of humidity to phase shift. Generate an environmental coupling feature matrix through normalization to quantify the correlation strength between environmental parameters and physical quantities. For example, the temperature sensitivity coefficient increases significantly in a high-temperature environment, and relevant compensation strategies need to be activated preferentially.

[0058] S1026. Perform weight assignment and dimension alignment on the ripple coefficient vector, passband tilt index vector, phase nonlinearity index vector, group delay stability parameter vector, transient fluctuation energy vector, and environmental coupling feature matrix, and fuse them into a time-frequency domain joint feature preliminary matrix.

[0059] Exemplarily, input the ripple coefficient, passband tilt, phase nonlinearity index, group delay stability, transient fluctuation energy, and environmental coupling features into the feature fusion layer. Use the random forest algorithm to evaluate the importance of each feature, assign weights, and then align the feature dimensions through a dimensionality reduction algorithm (such as t-SNE) to eliminate redundant information and generate a time-frequency domain joint feature preliminary matrix. Its rows correspond to time windows, and its columns correspond to the fused feature dimensions, providing a standardized input for the LSTM network.

[0060] S1027. Perform dynamic feature selection on the time-frequency domain joint feature preliminary matrix based on the environmental compensation parameters to generate a time-frequency domain joint feature matrix.

[0061] Exemplarily, based on real-time environmental compensation parameters (such as the current temperature and humidity values), dynamically activate or inhibit specific channels in the feature preliminary matrix. For example, when the temperature exceeds the threshold, inhibit the phase nonlinearity feature that is insensitive to temperature and enhance the weight of the temperature and humidity coupling feature. At the same time, apply exponential decay weighting to the temperature-sensitive features to reduce the risk of overcompensation during environmental mutations, and output an optimized time-frequency domain joint feature matrix to ensure the adaptability of the feature set in different environments.

[0062] In some embodiments, a first set of adjustment parameters and a hidden state vector are generated through the main branch, and through the sub-branch, the first set of adjustment parameters is reversely corrected according to the hidden state vector to generate compensation parameters, including: S1031 - S1034.

[0063] S1031. Input the time-frequency domain joint feature matrix into the main branch for time-sequence dependence analysis to generate a first set of adjustment parameters and a hidden state vector.

[0064] Exemplarily, a deep temporal modeling is performed on the joint time-frequency domain feature matrix through a bidirectional long short-term memory network (LSTM) to generate a first set of adjustment parameters and a hidden state vector. The joint time-frequency domain feature matrix, as input data, has a dimension of time steps × feature dimension, reflecting the multi-dimensional characteristics of the passband fluctuations of the filter within different time windows. The bidirectional LSTM consists of a forward layer and a backward layer: the forward layer processes data in chronological order to capture historical temporal dependencies, such as the cumulative effect of passband fluctuations over time; the backward layer processes data in reverse order to capture future potential trends, such as the delayed impact caused by changes in environmental parameters. The outputs of the two layers are concatenated to form a preliminary set of hidden states, containing bidirectional temporal information. To further enhance the expressive ability of key features, an attention mechanism is introduced to dynamically allocate weights - adjusting the attention to features in different frequency bands according to real-time environmental parameters (such as temperature and humidity). For example, in a high-temperature environment, the features of temperature-sensitive frequency bands are weighted more heavily. The weighted hidden states are processed through residual connections and layer normalization to improve training stability while retaining the original information. The normalized hidden states are input into a multi-task decoding layer, and the capacitance adjustment amount, gain adjustment amount, phase correction amount, and dynamic range adjustment amount are respectively mapped and generated through parallel fully connected networks to form the first set of adjustment parameters. The hidden state vector serves as the input of the secondary branch, carrying a panoramic understanding of the system state by the main branch and providing a data basis for side effect evaluation.

[0065] S1032. Construct an initial Jacobian matrix based on the hidden state vector, and analyze the initial Jacobian matrix to obtain the parameter negative impact index.

[0066] Exemplarily, by analyzing the correlation between the hidden state vector and physical quantities, the negative impact of the main branch parameter adjustment on the system performance is quantified. The hidden state vector is subjected to feature decoupling, and independent component analysis (ICA) is used to separate the independent feature subspaces related to capacitance, gain, phase, and dynamic range adjustments, eliminating the coupling interference between parameters. Based on the decoupled feature subspaces, a Jacobian matrix is constructed in combination with a preset physical impact model - the rows of this matrix correspond to key performance indicators (such as group delay stability, transient fluctuation energy), the columns correspond to adjustment parameters, and the matrix elements represent the partial derivatives of specific parameters with respect to a certain performance indicator. For example, the influence coefficient of the gain adjustment amount on the group delay stability reflects the sensitivity of the group delay fluctuation caused by the gain change. To eliminate noise interference, a low-rank approximation is performed on the original Jacobian matrix, and the main components are retained through singular value decomposition, while the secondary noise components are removed. Then, the global influence weights of each parameter (such as the Frobenius norm quantifies the overall perturbation strength of the parameter on the system) are calculated and normalized to generate the parameter negative impact index. This index vector clearly identifies the high-side-effect parameters that need to be preferentially suppressed (such as the significant impact of gain adjustment on group delay), providing a quantitative basis for the reverse correction of the secondary branch.

[0067] S1033. Input the parameter negative impact index into the secondary branch and perform adversarial gradient backpropagation processing to generate a second set of adjustment parameters.

[0068] Exemplarily, the secondary branch takes the parameter negative impact index as the input and generates reverse adjustment parameters through an adversarial training mechanism to suppress the side effects of the main branch parameters. The core of adversarial training lies in the design of the gradient reversal layer (GRL): during forward propagation, the secondary branch receives the negative impact index and generates a second set of adjustment parameters; during backpropagation, the gradient sign is reversed, forcing the secondary branch network to learn a parameter adjustment strategy to counteract the negative impact of the main branch. Specifically, the secondary branch consists of a multi-layer fully connected network, with its input being the negative impact index vector and its output including the reverse adjustment amounts of capacitance, gain, and phase. For example, if the main branch suggests increasing the gain (ΔG), but this operation causes a decrease in the group delay stability, the secondary branch generates a reverse gain adjustment amount (ΔG') to partially offset the gain change of the main branch. The adversarial weight factor dynamically adjusts the reverse correction intensity. For example, in a high-humidity environment, the influence of humidity on the phase increases, and the secondary branch correspondingly increases the correction amplitude of the phase parameter. Through iterative training, the secondary branch gradually optimizes the parameter correction strategy to ensure minimizing the interference to the compensation effect of the main branch while suppressing the side effects.

[0069] S1034. Perform constrained optimization processing on the first set of adjustment parameters and the second set of adjustment parameters to generate compensation parameters.

[0070] Exemplarily, this step combines the output parameters of the main and secondary branches and generates an executable set of compensation parameters through multi-objective optimization and physical constraint mapping. Construct an optimization objective function to balance the passband ripple suppression effect and the side effect suppression requirement: the main branch parameters aim to minimize the passband ripple coefficient, and the secondary branch parameters aim to reduce the negative impact index. A regularization term is introduced during the optimization process to limit the sudden change of the parameter adjustment amount and avoid the instantaneous overload of the hardware actuator. Physical constraints are imposed to ensure parameter feasibility: the capacitance adjustment amount is limited within the physical range of the adjustable capacitor array (such as 0.1 pF step, total range ±20 pF); the gain and phase adjustment amounts are forced to be dynamically balanced through cross-coupling equations to prevent resonance or phase distortion caused by separate adjustments; the dynamic range parameter is weighted and integrated according to the frequency-domain distribution of the insertion loss and converted into a DAC control voltage signal to ensure that the amplifier operating point is linearly adjustable. The environmental compensation module predicts the future temperature change trend based on Kalman filtering and generates a pre-adjustment instruction for the heater power to shorten the response delay under environmental mutations. The optimized parameter set drives the hardware through a multi-actuator cooperative control protocol (such as timestamp-based instruction synchronization) to achieve high-precision real-time compensation of the passband ripple, with a ripple suppression accuracy of ±0.1 dB and a group delay fluctuation reduction of more than 50%.

[0071] In some embodiments, the time-frequency domain joint feature matrix is input into the main branch for temporal dependence analysis to generate a first set of adjustment parameters and a hidden state vector, including: S311-S314.

[0072] S311. Extract the forward dependence features and backward dependence features of the time-frequency domain joint feature matrix through the forward layer and the backward layer respectively to generate a preliminary set of hidden state vectors.

[0073] Extract the forward and backward temporal dependence features of the time-frequency domain joint feature matrix through a bidirectional long short-term memory network (LSTM) to construct a preliminary set of hidden state vectors. The time-frequency domain joint feature matrix , where T is the number of time steps and D′ is the optimized feature dimension. The feature vector at each time step includes time-frequency domain joint features such as ripple coefficient, passband tilt, and phase nonlinearity index.

[0074] Forward LSTM processing: The forward layer processes the feature matrix in chronological order (t = 1 → T) to capture historical temporal dependence relationships, such as the cumulative effect of passband ripple over time, and outputs a forward hidden state sequence. The specific formula is: ; where L is the number of LSTM cells.

[0075] Backward LSTM processing: The backward layer processes the feature matrix in reverse order (t = T → 1) to capture future potential trends, such as the delayed impact of environmental parameter changes on passband fluctuations, and outputs a forward hidden state sequence , specifically: ; where L is the number of LSTM cells.

[0076] After the above forward and backward processing, Concatenate the bidirectional outputs along the feature dimension to generate a preliminary set of hidden states , specifically: ; This preliminary set of hidden states Contains both historical and future information and provides a basis for subsequent attention weighting.

[0077] S312. Generate an attention weight matrix according to the environmental compensation parameter matrix, and determine an enhanced set of hidden state vectors according to the preliminary set of hidden state vectors and the attention weight matrix.

[0078] Exemplarily, input the environmental compensation parameter matrix , specifically: , which includes a temperature compensation amount ΔT and a humidity compensation amount ΔH. The environmental compensation parameter matrix Integrate with the initial hidden state set to generate attention scores through a learnable weight matrix , specifically: ; Among them, , is a training parameter, and || represents the concatenation operation.

[0079] Normalize the attention scores , specifically: ; Among them, the softmax function ensures that the weights sum to 1 at each time step.

[0080] Apply the attention weights to weight the initial hidden state to generate an enhanced hidden state vector set , specifically: ; Among them, ⊙ is element-wise multiplication, highlighting the features of the environment-sensitive frequency bands.

[0081] S313. Add the initial hidden state vector and the enhanced hidden state vector set through a skip connection to generate a normalized hidden state vector.

[0082] Exemplarily, the original feature information is retained through a skip connection, and the training stability is improved through normalization. Add the initial hidden state set and the enhanced hidden state vector set to generate a residual vector set : ; Normalize the residual vector set so that its mean is 0 and variance is 1 to generate a normalized hidden state , specifically: ; ; Among them, μ, σ 2 are the mean and variance at time step t, γ, β are learnable parameters, and ϵ is a small constant to prevent division by zero.

[0083] S314. Perform multi-task decoding processing on the normalized hidden state vector, and generate a capacitance value adjustment amount, a gain adjustment amount, a phase correction amount, and a dynamic range adjustment amount respectively through parallel fully connected layers to obtain a first adjustment parameter set, and output the hidden state vector.

[0084] Decode the normalized hidden state vector into four types of hardware-executable adjustment parameters through parallel fully connected layers. Each fully connected layer independently learns the mapping relationship from the hidden state to a specific parameter.

[0085] For the normalized hidden state at each time step , independently decode and generate four types of adjustment parameters.

[0086] Capacitance adjustment decoding: Analyze the features related to the temperature drift and frequency response characteristics of the capacitance in the hidden state, generate the control signal for the digitally tunable capacitor array, with a step accuracy of 0.1 pF. The specific formula is: ; where is the capacitance decoding parameter.

[0087] Gain adjustment decoding: Analyze the variation pattern of the passband gain with frequency and time, output the voltage adjustment amount of the voltage-controlled amplifier, and compensate for the gain flatness. The specific formula is: ; where is the gain decoding parameter.

[0088] Phase correction decoding: Based on the phase nonlinearity index and the group delay stability characteristics, generate the coefficients of the finite impulse response (FIR) filter to correct the phase linearity ; where is the phase decoding parameter.

[0089] Dynamic range adjustment decoding: Combine the transient fluctuation energy of the insertion loss to generate a dynamic bias voltage signal and optimize the linear range of the operating point of the amplifier ; where is the dynamic range adjustment decoding parameter.

[0090] Then, fuse them into the first set of adjustment parameters , which is used for subsequent physical constraint mapping. The hidden state vector is passed to the secondary branch for side effect evaluation.

[0091] In some embodiments, construct an initial Jacobian matrix according to the hidden state vector, and analyze the initial Jacobian matrix to obtain the parameter negative impact indicators, including: S321 - S324.

[0092] S321. Decouple the features of the hidden state vector to obtain the parameter-related feature matrix.

[0093] Exemplarily, the hidden state vector is feature decoupled by Independent Component Analysis (ICA), and independent feature subspaces strongly related to each adjustment parameter (capacitance, gain, phase, dynamic range) are separated. The hidden state vector contains the temporal features extracted by the bidirectional LSTM, but there may be coupling interference between parameters inside it (such as the mutual influence between gain adjustment and phase adjustment). The ICA algorithm projects the hidden state into the independent feature space by finding the linear transformation matrix to obtain the parameter-related feature matrix , specifically: ; where is the ICA transformation matrix, whose dimension is determined by the hidden state dimension (2L) and the parameter "4". Each column of the parameter-related feature matrix corresponds to the feature subspace of an adjustment parameter (such as f C is related to capacitance adjustment), satisfying the statistical independence assumption. The objective function of ICA maximizes the non-Gaussianity to ensure that the parameter feature subspaces are decoupled as much as possible. The decoupled feature matrix provides a clear parameter-feature mapping relationship for the subsequent construction of the Jacobian matrix. For example, f G only reflects the independent influence of gain adjustment on the system state.

[0094] S322. Perform dynamic Jacobian matrix construction on the parameter-related feature matrix to generate the initial Jacobian matrix.

[0095] Exemplarily, based on the parameter-related feature matrix and the preset physical performance indicators (group delay stability τ std , transient fluctuation energy E trans ), construct the dynamic Jacobian matrix J to quantify the side effects of adjustment parameters on the performance indicators. The matrix element J ij represents the partial derivative of the jth parameter with respect to the ith performance indicator, and the initial Jacobian matrix J is obtained, specifically: ; where is obtained by regression fitting of historical data, is the sensitivity of the feature subspace to the parameter (calculated by backpropagation of the network gradient). The dynamically constructed Jacobian matrix reflects the side effect intensity of parameter adjustment under the current working conditions. For example, the influence coefficient of gain adjustment ΔG on the group delay stability τ std .

[0096] S323. Extract the first k principal components of the initial Jacobian matrix through singular value decomposition and eliminate the noise components to generate a denoised Jacobian matrix.

[0097] Exemplarily, perform singular value decomposition (SVD) on the initial Jacobian matrix J to eliminate noise interference and retain the principal components: ; where U ∈ R 2×2 and V ∈ R 4×4 are the left and right singular vector matrices respectively, and Σ ∈ R 2×4 is a diagonal matrix, and the element σ i is the singular value (arranged in descending order). By setting an energy threshold (such as the cumulative energy ratio ≥ 95%), select the first k principal components (usually k = 2) to reconstruct the denoised Jacobian matrix : ; where , , are the truncated submatrices. This operation eliminates the noise components corresponding to small singular values (such as measurement errors or instantaneous interferences), retains the main effects of parameter side effects, and improves the reliability of the negative impact index.

[0098] S324. Calculate the global impact weight of each adjustment parameter based on the Frobenius norm and the denoised Jacobian matrix to generate a parameter negative impact index.

[0099] Based on the denoised Jacobian matrix , for each column of the Jacobian matrix corresponding to each parameter, calculate the Frobenius norm to obtain the global impact weight , specifically: ; Normalize the global impact weight to ensure that the sum is 1 to obtain the parameter negative impact index , specifically: ; For example, if W G accounts for 60%, then the side branch needs to preferentially suppress the side effects of gain adjustment. The generated parameter negative impact index is directly input into the side branch to guide the generation of adversarial correction parameters.

[0100] In some embodiments, the compensation parameters include: capacitance value adjustment amount, gain adjustment amount, phase correction amount, dynamic range adjustment amount. Dynamic coupling mapping is performed according to the compensation parameters and preset physical constraint conditions to obtain a set of constrained parameters, including: Generate the constrained capacitance adjustment amount according to the preset capacitance boundary value and the capacitance value adjustment amount.

[0101] Perform cross-coupling calculation on the gain adjustment amount and the phase correction amount, and generate a joint compensation parameter pair by fitting the gain-phase coupling equation based on the least squares method.

[0102] Perform frequency-domain integration on the dynamic range adjustment amount, and generate a dynamic range weight vector in combination with the frequency-domain distribution characteristics of the insertion loss matrix.

[0103] Perform power-voltage conversion on the dynamic range weight vector, and generate a DAC control voltage matrix based on the preset amplifier bias curve.

[0104] Predict the temperature change amount of the environmental compensation parameter matrix in the future time window through a Kalman filter, and generate a heater power control signal according to the temperature change amount.

[0105] Perform multi-parameter collaborative verification processing on the constrained capacitance adjustment amount, the joint compensation parameter pair, the DAC control voltage matrix, and the heater power control signal, verify the logical consistency between the parameters according to the preset physical constraint conditions, and generate a set of constrained parameters.

[0106] In this step, the compensation parameters are mapped into hardware-executable instructions through physical constraints and dynamic coupling strategies to ensure parameter feasibility and system stability. First, impose boundary constraints on the capacitance value adjustment amount: according to the physical limits of the adjustable capacitance array (such as the minimum capacitance value C min and the maximum capacitance value C max),(The constrained capacitance adjustment amount is generated through a clipping function to avoid device damage caused by overshoot. Secondly, for the coupling relationship between the gain and phase parameters, the gain-phase joint equation is fitted by the least squares method to enforce the dynamic balance between the gain adjustment amount and the phase correction amount, preventing resonance or phase distortion caused by separate adjustment. The dynamic range adjustment amount is then integrated in the frequency domain by combining the frequency domain distribution characteristics of the insertion loss matrix, calculating the frequency band weight vector, assigning a larger adjustment weight to the high-loss frequency band, and optimizing the frequency response consistency of the dynamic range. This weight vector is converted into a DAC control voltage matrix through a preset amplifier bias curve to drive the linear adjustment of the amplifier operating point. The environmental compensation module predicts the temperature change amount in the future time window based on the Kalman filter and generates a heater power control signal to achieve forward-looking compensation for temperature drift. Multi-parameter collaborative verification is performed on the constrained capacitance adjustment amount, joint compensation parameter pairs, DAC control voltage, and heater power to check the logical consistency between the parameters (such as the temperature drift compatibility between capacitance adjustment and heating power, the stability of the gain-phase coupling equation, and the matching between the DAC voltage and the amplifier linear range), eliminating conflicting parameter combinations and generating a constrained parameter set to ensure the safety and effectiveness of hardware execution.)

[0107] Please refer to Figure 2 , Figure 2 FIG. is a schematic block diagram of a filter passband ripple automatic compensation device provided by an embodiment of the present application. The filter passband ripple automatic compensation device 200 is used to execute the aforementioned filter passband ripple automatic compensation method. Among them, the filter passband ripple automatic compensation device 200 can be configured in a server.

[0108] Among them, the server can be an independent server, a server cluster, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, Content Delivery Network (CDN), and big data and artificial intelligence platforms.

[0109] As Figure 2 shown, the filter passband ripple automatic compensation device 200 includes: a data acquisition module 201, a feature extraction module 202, a compensation calculation module 203, a constraint mapping module 204, and a hardware control module 205.

[0110] The data acquisition module 201 is used to acquire multi-dimensional tensor data of the filter for data.

[0111] The feature extraction module 202 is used to extract the features of the multi-dimensional tensor data to obtain a time-frequency domain joint feature matrix.

[0112] A compensation calculation module 203 is configured to input a time-frequency domain joint feature matrix into a preset dual-branch LSTM network. The preset dual-branch LSTM network includes a main branch and a secondary branch. The main branch is used to generate a first set of adjustment parameters and a hidden state vector, and the secondary branch is used to reversely correct the first set of adjustment parameters according to the hidden state vector to generate compensation parameters.

[0113] A constraint mapping module 204 is configured to perform dynamic coupling mapping according to the compensation parameters and preset physical constraint conditions to obtain a set of constrained parameters.

[0114] A hardware control module 205 is configured to control corresponding hardware based on the set of constrained parameters to perform passband ripple compensation on the filter.

[0115] An embodiment of the present application provides an electronic device, which includes a memory and a processor. The memory is used to store a computer program. The processor is configured to execute the computer program and implement the filter passband ripple automatic compensation method according to any one of the embodiments of the present application when executing the computer program.

[0116] An embodiment of the present application provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the processor is caused to implement the filter passband ripple automatic compensation method according to any one of the embodiments of the present application.

[0117] As described above, the above is only the specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present application can easily think of various equivalent modifications or substitutions, and these modifications or substitutions should be covered by the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the protection scope of the claims.

Claims

1. An automatic compensation method for filter passband ripple, characterized in that, The method includes: Obtaining multi-dimensional tensor data of the filter for data; Extracting features of the multi-dimensional tensor data to obtain a joint time-frequency domain feature matrix; Inputting the joint time-frequency domain feature matrix into a preset dual-branch LSTM network, where the preset dual-branch LSTM network includes a main branch and a sub-branch. Through the main branch, a first set of adjustment parameters and a hidden state vector are generated. Through the sub-branch, according to the hidden state vector, the first set of adjustment parameters is reversely corrected to generate compensation parameters; Performing dynamic coupling mapping according to the compensation parameters and preset physical constraint conditions to obtain a constrained parameter set; Based on the constrained parameter set, controlling the corresponding hardware to perform passband ripple compensation on the filter.

2. The automatic compensation method for filter passband ripple according to claim 1, characterized in that, The obtaining of the multi-dimensional tensor data of the filter for data includes: Collecting the frequency response, phase response, insertion loss, group delay, and environmental parameters of the filter through multiple channels; Using a vector network analyzer to perform an in-band amplitude-frequency characteristic curve scan on the frequency response to generate a frequency response matrix; Performing IQ demodulation on the phase response data to determine the phase offset. Calculating the derivative of the phase of the group delay with respect to frequency through Hilbert transform to obtain a group delay tensor. Generating a phase-group delay joint matrix according to the phase offset and the group delay tensor; Measuring the RMS power values of the input and output ends of the insertion loss through a power probe, and calculating the power ratio of the input and output ends RMS power values to generate an insertion loss matrix; Performing temperature and humidity drift modeling and linear compensation on the environmental parameters to generate an environmental compensation parameter matrix; Synthesizing the multi-dimensional tensor data according to the frequency response matrix, the phase-group delay joint matrix, the insertion loss matrix, and the environmental compensation parameter matrix.

3. The automatic compensation method for filter passband ripple according to claim 2, characterized in that, The extracting of the features of the multi-dimensional tensor data to obtain a joint time-frequency domain feature matrix includes: Calculating the range between the maximum amplitude and the minimum amplitude of the frequency response matrix within a preset passband to generate a ripple coefficient vector, and performing linear regression fitting on the amplitude-frequency characteristic curve corresponding to the frequency response matrix to generate a passband tilt index vector; Calculating the phase non-linearity index vector of the phase-group delay joint matrix through frequency domain differential entropy; Performing sliding window standard deviation statistics on the group delay tensor to generate a group delay stability parameter vector; Based on a sliding window, extracting the transient response features of the insertion loss matrix and calculating the root mean square value of the insertion loss change rate to generate a transient fluctuation energy vector; Performing temperature and humidity coupling analysis and processing on the environmental compensation parameter matrix to generate an environmental coupling feature matrix; Performing weight assignment and dimension alignment on the ripple coefficient vector, the passband tilt index vector, the phase non-linearity index vector, the group delay stability parameter vector, the transient fluctuation energy vector, and the environmental coupling feature matrix, and fusing them into a preliminary joint time-frequency domain feature matrix; Performing dynamic feature selection on the preliminary joint time-frequency domain feature matrix based on the environmental compensation parameters to generate the joint time-frequency domain feature matrix.

4. The automatic compensation method for filter passband ripple according to claim 3, characterized in that, Generating a first set of adjustment parameters and a hidden state vector through the main branch, and through the secondary branch, reversely correcting the first set of adjustment parameters according to the hidden state vector to generate compensation parameters, including: Inputting the time-frequency domain joint feature matrix into the main branch to perform temporal dependence analysis, and generating a first set of adjustment parameters and a hidden state vector; Constructing an initial Jacobian matrix according to the hidden state vector, and analyzing the initial Jacobian matrix to obtain a parameter negative impact index; Inputting the parameter negative impact index into the secondary branch to perform adversarial gradient backpropagation processing, and generating a second set of adjustment parameters; Performing constraint optimization processing on the first set of adjustment parameters and the second set of adjustment parameters to generate the compensation parameters.

5. The automatic compensation method for filter passband ripple according to claim 4, characterized in that, The step of inputting the time-frequency domain joint feature matrix into the main branch to perform temporal dependence analysis and generating a first set of adjustment parameters and a hidden state vector includes: Extracting the forward dependence features and the backward dependence features of the time-frequency domain joint feature matrix through a forward layer and a backward layer respectively, and generating a preliminary set of hidden state vectors; Generating an attention weight matrix according to the environmental compensation parameter matrix, and determining an enhanced set of hidden state vectors according to the preliminary set of hidden state vectors and the attention weight matrix; Adding the preliminary hidden state vectors to the enhanced set of hidden state vectors through skip connections to generate a normalized hidden state vector; Performing multi-task decoding processing on the normalized hidden state vector, and respectively mapping through parallel fully connected layers to generate a capacitance value adjustment amount, a gain adjustment amount, a phase correction amount, and a dynamic range adjustment amount, obtaining the first set of adjustment parameters, and outputting the hidden state vector.

6. The automatic compensation method for filter passband ripple according to claim 4, characterized in that, The step of constructing an initial Jacobian matrix according to the hidden state vector and analyzing the initial Jacobian matrix to obtain a parameter negative impact index includes: Performing feature decoupling on the hidden state vector to obtain a parameter-related feature matrix; Performing dynamic Jacobian matrix construction on the parameter-related feature matrix to generate the initial Jacobian matrix; Extracting the first k principal components of the initial Jacobian matrix through singular value decomposition and removing noise components to generate a denoised Jacobian matrix; Calculating the global impact weight of each adjustment parameter based on the Frobenius norm and the denoised Jacobian matrix to generate the parameter negative impact index.

7. The automatic compensation method for filter passband ripple as claimed in claim 3, wherein, The compensation parameters include: a capacitance value adjustment amount, a gain adjustment amount, a phase correction amount, and a dynamic range adjustment amount. Dynamically coupling and mapping according to the compensation parameters and preset physical constraint conditions to obtain a set of constrained parameters, including: Generating a constrained capacitance adjustment amount according to a preset capacitance boundary value and the capacitance value adjustment amount; Performing cross-coupling calculation on the gain adjustment amount and the phase correction amount, and fitting a gain-phase coupling equation based on the least squares method to generate a pair of joint compensation parameters; Performing frequency domain integration on the dynamic range adjustment amount, and combining the frequency domain distribution characteristics of the insertion loss matrix to generate a dynamic range weight vector; Performing power-voltage conversion on the dynamic range weight vector, and generating a DAC control voltage matrix based on a preset amplifier bias curve; Predict the temperature change amount of the environmental compensation parameter matrix in the future time window through a Kalman filter, and generate a heater power control signal according to the temperature change amount; Perform multi-parameter collaborative verification processing on the constrained capacitance adjustment amount, the joint compensation parameter pair, the DAC control voltage matrix, and the heater power control signal, verify the logical consistency between parameters according to the preset physical constraint conditions, and generate the constrained parameter set.

8. An automatic compensation device for filter passband ripple, wherein, The filter passband ripple automatic compensation device is used to execute the filter passband ripple automatic compensation method according to any one of claims 1-7. The filter passband ripple automatic compensation device includes: A data acquisition module, configured to acquire multi-dimensional tensor data of the filter for data; A feature extraction module, configured to extract features of the multi-dimensional tensor data to obtain a time-frequency domain joint feature matrix; A compensation calculation module, configured to input the time-frequency domain joint feature matrix into a preset dual-branch LSTM network. The preset dual-branch LSTM network includes: a main branch and a sub-branch. Generate a first adjustment parameter set and a hidden state vector through the main branch, and through the sub-branch, reverse-correct the first adjustment parameter set according to the hidden state vector to generate compensation parameters; A constraint mapping module, configured to perform dynamic coupling mapping according to the compensation parameters and preset physical constraint conditions to obtain a constrained parameter set; A hardware control module, configured to control the corresponding hardware based on the constrained parameter set to perform passband ripple compensation on the filter.

Images