A method and system for operating an interference-resistant electrically tunable trap

Abstract

The application relates to the technical field of signal processing, and discloses an anti-interference electrically tunable wave trap operation control method and system, which comprises the following steps: collecting input signals of an adjustable wave trap group, constructing a reinforcement learning state vector by dynamically suppressing strong electromagnetic noise by using adaptive Kalman filtering, generating optimal action instructions of a tuning module by using a deep Q network reinforcement learning model, controlling the tuning module to adjust parameters of the adjustable wave trap group, and finally updating and adjusting the deep Q network reinforcement learning model and the adaptive Kalman filtering based on actual effect indexes. Thus, the application constructs an electrically tunable wave trap high-precision operation control method based on deep cooperation of adaptive Kalman filtering and a deep Q network, can help the electrically tunable wave trap to stably maintain high-precision wave trapping performance in a scene with complex dynamic interference and strong noise superposition, and completely solves many limitations of traditional schemes in interference adaptation, precision control, multi-frequency cooperation and stability.

Description

Technical Field

[0001] This invention relates to the field of signal processing technology, and in particular to an anti-interference electrically tunable notch filter operation control method and system. Background Technology

[0002] Notch filters, as core components for suppressing strong interference at specific frequencies, are widely used in fields such as communications, radar, and electronic measurement where signal purity is critical. Their core requirement is to maintain high-precision notch filtering performance stably under complex dynamic interference scenarios (such as interference frequency jumps, abrupt changes in intensity, multiple frequency points superimposed, and accompanied by strong electromagnetic noise with a signal-to-noise ratio ≤10dB).

[0003] Currently, mainstream technical solutions in the industry have significant limitations: traditional fixed-frequency notch filters are completely unable to adapt to the dynamic changes of interference signals and can only suppress a single fixed frequency; electrically tunable notch filters based on varactor diodes are limited by the power capacity of the devices and cannot withstand strong interference signals at the 40dBm level, and their tuning range and stability are limited; although some existing electromechanical integrated anti-strong interference electrically tunable notch filters have achieved notch filtering functions in a wide frequency band of 110-520MHz and 1-5 frequency points, they still have key technical shortcomings: weak noise suppression capability in the signal preprocessing stage, and strong electromagnetic noise can easily lead to distortion in interference feature extraction; they can only extract basic static features such as frequency and intensity of interference signals and cannot characterize dynamic characteristics such as frequency jump rate and duration, resulting in lag in notch filter response; tuning parameter decisions rely on fixed calibration relationships, resulting in poor adaptability and insufficient tuning accuracy when facing complex interference scenarios; multi-frequency notch filtering is easily affected by the coupling effect between notch filters, resulting in bandwidth exceeding the standard or abnormal passband attenuation; the feedback correction mechanism lacks self-learning capability, and the response speed does not improve when the same interference recurs, resulting in insufficient dynamic adaptability and stability of the system. The aforementioned technical bottlenecks mean that existing solutions cannot meet the application requirements of high-end electronic devices in terms of notch filtration accuracy, response speed, and stability under complex dynamic interference and strong noise superposition scenarios. There is an urgent need to build a full-link adaptive control method.

[0004] Therefore, improving the notch filtering accuracy, response speed, and stability of high-end electronic devices in complex dynamic interference and strong noise superposition scenarios is a technical problem that urgently needs to be solved. Summary of the Invention

[0005] This invention provides an anti-interference electrically adjustable notch filter operation control method and system, aiming to solve at least one of the above-mentioned technical problems.

[0006] To achieve the above objectives, the present invention provides an anti-interference electrically adjustable notch filter operation control method, comprising the following steps:

[0007] The control processing module acquires the input signal of the adjustable notch filter group and uses an adaptive Kalman filter to dynamically suppress strong electromagnetic noise to obtain the noise-reduced interference signal.

[0008] Extract the static and dynamic features from the interference signal to construct a reinforcement learning state vector;

[0009] Based on the constructed reinforcement learning state vector, the deep Q-network reinforcement learning model obtained after training is used to perform inference and generate the optimal action instructions for the tuning module.

[0010] The optimal action command is sent to the tuning module, which then parses the optimal action command, generates adaptive driving parameters, and uses the adaptive driving parameters to adjust the parameters of the adjustable notch filter group.

[0011] The control processing module calculates the real-time reward value based on the actual performance index set of the adjustable notch filter group, constructs a new state vector by combining the collected interference features after executing the optimal action, updates the deep Q-network reinforcement learning model, and adjusts the adaptive factor of the adaptive Kalman filter.

[0012] Optionally, the control processing module acquires the input signal of the adjustable notch filter group, dynamically suppresses strong electromagnetic noise using an adaptive Kalman filter, and obtains the denoised interference signal. This process specifically includes:

[0013] The control processing module uses two synchronous high-speed ADCs to acquire the original mixed signal and the pre-filtered reference signal from the input of the adjustable notch filter group, respectively, to obtain two discrete signal sequences.

[0014] Using the constructed adaptive Kalman filter model, the noise covariance and observation noise covariance of the system are estimated by recursive least squares method, and the adaptive factor adjustment covariance matrix is ​​determined.

[0015] Following the standard recursive process of adaptive Kalman filtering, the following steps are executed sequentially: state prediction, prediction error covariance calculation, Kalman gain solution, state update and covariance update. The denoised interference signal is then extracted from the updated state vector.

[0016] Optionally, the step of extracting static and dynamic features from the interference signal and constructing a reinforcement learning state vector specifically includes:

[0017] An N-point FFT transformation is performed on the denoised interference signal to obtain the frequency domain signal. The actual intensity and frequency of each frequency point are calculated, and the frequency points whose intensity meets the threshold are selected to obtain the static feature set.

[0018] The Db4 wavelet basis is used to perform wavelet decomposition on the denoised interference signal, extract the detail coefficients of the corresponding working frequency band, detect the modulus maxima of the detail coefficients to locate the interference transition time, calculate the interference frequency transition rate and duration, and obtain the dynamic feature set.

[0019] All features in the static and dynamic feature sets are normalized separately, and then combined to construct the reinforcement learning state vector.

[0020] Optionally, based on the constructed reinforcement learning state vector, inference is performed using the deep Q-network reinforcement learning model obtained after training to generate the optimal action instruction steps for the tuning module, specifically including:

[0021] Construct a fully connected deep Q-network structure; wherein the fully connected deep Q-network structure includes an input layer for receiving state vectors, a hidden layer using the ReLU activation function, and an output layer for outputting the Q-values ​​of each action, and an experience replay pool combined with a target network and an evaluation network in a dual network structure;

[0022] The fully connected deep Q network structure was pre-trained offline using a historical interference dataset, and the network parameters were updated and evaluated using a temporal difference algorithm until the loss value stabilized.

[0023] The state vector constructed in real time is input into the deep Q-network reinforcement learning model after training, and the action with the largest output Q value is selected as the optimal action instruction of the tuning module.

[0024] Optionally, the optimal action command is sent to the tuning module, which then parses the optimal action command, generates adaptive driving parameters, and uses the adaptive driving parameters to adjust the parameters of the adjustable notch filter group. This process specifically includes:

[0025] The optimal action command is sent to the tuning module, which then analyzes the optimal action command to obtain the target step angle and motor drive current of each notch filter.

[0026] The driving parameters are generated based on the analyzed motor drive current, and the stepper motor is driven to rotate toward the target step angle.

[0027] The actual step angle of the stepper motor is acquired in real time by an encoder, the deviation between the actual step angle and the target step angle is calculated, and the drive current is adjusted according to the magnitude of the deviation until the deviation meets the accuracy requirements.

[0028] Optionally, the method further includes:

[0029] The actual frequency and notch depth of each notch are acquired by the built-in sensor of the adjustable notch filter group, and the acquired data is fed back to the tuning module.

[0030] The tuning module compares the feedback data with the target parameters to determine whether the deviation between the feedback data and the target parameters exceeds the standard. If so, it returns to repeat the stepper motor adjustment action until the parameters of each notch filter meet the standard.

[0031] Optionally, the method further includes:

[0032] When in a multi-frequency notch filter scenario, the stepper motors of each notch filter are adjusted sequentially in order of interference jump rate from low to high. After each notch filter is adjusted, the step angle change of adjacent notch filters is collected by the encoder.

[0033] If the step angle change exceeds the coupling deviation threshold, the adjacent notch filter stepper motor is triggered to perform targeted adjustments based on the coupling deviation until the coupling deviation meets the requirements.

[0034] Optionally, the control processing module calculates the real-time reward value based on the actual performance index set of the adjustable notch filter group, constructs a new state vector by combining the collected interference features after executing the optimal action, and updates the deep Q-network reinforcement learning model and adjusts the adaptive factor of the adaptive Kalman filter, specifically including:

[0035] The input and output power of the adjustable notch filter group are collected by a directional coupler and a power meter, the output signal spectrum is collected by a spectrum analyzer, the standing wave ratio is collected by a standing wave ratio tester, and the response time is recorded by a timer. The results are integrated to obtain a set of actual performance indicators.

[0036] The interference features after the optimal action is executed are collected, and a new state vector is constructed by combining it with the real-time reward value. This vector is then used to update the deep Q-network reinforcement learning model and adjust the adaptive factor of the adaptive Kalman filter.

[0037] Optionally, the steps of collecting interference features after executing the optimal action, constructing a new state vector by combining it with the real-time reward value, updating the deep Q-network reinforcement learning model, and adjusting the adaptive factor of the adaptive Kalman filter specifically include:

[0038] Substitute the set of actual performance indicators into the reward function to calculate the real-time reward value, and collect the interference features after executing the optimal action to construct a new state vector;

[0039] The original state vector, optimal action, real-time reward value and new state vector are combined to form an experience sample and stored in the experience replay pool. Every time a preset number of samples are stored, the deep Q network is triggered to update online.

[0040] The system determines the range of the real-time reward value. If the reward value is below the first threshold, the adaptive factor of the adaptive Kalman filter is adjusted to optimize the noise suppression effect, features are re-extracted, and reinforcement learning is triggered to make a new decision. If the reward value is between the first and second thresholds, the adaptive Kalman filter parameters are kept unchanged, and the parameters of the Q-network reinforcement learning model are adjusted to accelerate the online update speed of the reinforcement learning network. If the reward value is above the second threshold, the current parameters are maintained, and the optimal stepper motor adjustment strategy is stored for reuse in similar interference scenarios.

[0041] Furthermore, in order to achieve the above objectives, the present invention also provides an anti-interference electrically tunable notch filter operation control system, comprising: a control processing module, an adjustable notch filter group, and a tuning module, wherein the control processing module, the adjustable notch filter group, and the tuning module are configured to execute the anti-interference electrically tunable notch filter operation control method described in any of the above embodiments.

[0042] The beneficial effects of this invention are as follows: It proposes an anti-interference electrically tunable notch filter operation control method and system. By acquiring the input signal of the tunable notch filter group, it uses an adaptive Kalman filter to dynamically suppress strong electromagnetic noise. Then, by constructing a reinforcement learning state vector, it uses a deep Q-network reinforcement learning model to perform inference and generate the optimal action command of the tuning module. This controls the tuning module to adjust the parameters of the tunable notch filter group. Finally, the deep Q-network reinforcement learning model and the adaptive Kalman filter are updated and adjusted based on the actual performance indicators. Therefore, this invention constructs a high-precision operation control method for electrically tunable notch filters based on the deep collaboration of adaptive Kalman filtering and deep Q-network. It utilizes adaptive Kalman filtering to dynamically adjust the noise covariance matrix, while employing a deep Q-network reinforcement learning model with a dual-network structure and a strategy of offline pre-training and online fine-tuning. The output stepper motor tuning commands enable the electrically tunable notch filter to stably maintain high-precision notch filtering performance in scenarios with complex dynamic interference and strong noise superposition. This completely solves many limitations of traditional solutions in terms of interference adaptation, precision control, multi-frequency collaboration, and stability, meeting the stringent requirements for strong interference suppression in fields such as communication and radar. Attached Figure Description

[0043] Figure 1 This is a flowchart illustrating the anti-interference electrically tunable notch filter operation control method according to an embodiment of the present invention.

[0044] Figure 2 This is a schematic diagram of the anti-interference electrically adjustable notch filter operation control system according to an embodiment of the present invention.

[0045] Explanation of reference numerals in the attached figures:

[0046] 10 - Control processing module; 20 - Adjustable notch filter group; 30 - Tuning module. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0048] This invention provides an anti-interference electrically tunable notch filter operation control method, referring to... Figure 1 As shown, it includes the following steps:

[0049] S1: The control processing module acquires the input signal of the adjustable notch filter group, uses adaptive Kalman filtering to dynamically suppress strong electromagnetic noise, and obtains the noise-reduced interference signal.

[0050] S2: Extract the static and dynamic features from the interference signal and construct a reinforcement learning state vector;

[0051] S3: Based on the constructed reinforcement learning state vector, use the deep Q-network reinforcement learning model obtained after training to perform inference and generate the optimal action instructions for the tuning module;

[0052] S4: Send the optimal action command to the tuning module, drive the tuning module to parse the optimal action command, generate adaptive drive parameters, and use the adaptive drive parameters to adjust the parameters of the adjustable notch filter group.

[0053] S5: The control processing module calculates the real-time reward value based on the actual performance index set of the adjustable notch filter group, constructs a new state vector by combining the collected interference features after executing the optimal action, updates the deep Q network reinforcement learning model, and adjusts the adaptive factor of the adaptive Kalman filter.

[0054] It should be noted that current mainstream technologies in the industry have significant limitations: traditional fixed-frequency notch filters are completely unable to adapt to the dynamic changes of interference signals and can only suppress a single fixed frequency; electrically tunable notch filters based on varactor diodes are limited by the power capacity of the devices and cannot withstand strong interference signals at the 40dBm level, and their tuning range and stability are also limited; although some existing electromechanical integrated anti-strong interference electrically tunable notch filters have achieved notch filtering functions in a wide frequency band of 110-520MHz and 1-5 frequency points, they still have key technical shortcomings: noise suppression in the signal preprocessing stage. The existing notch filter suffers from several technical bottlenecks. First, its control capabilities are weak, and strong electromagnetic noise easily distorts the extraction of interference features. It can only extract basic static features such as frequency and intensity of the interference signal, failing to characterize dynamic characteristics such as frequency jump rate and duration, resulting in lag in notch filter response. Second, its tuning parameter decisions rely on fixed calibration relationships, leading to poor adaptability and insufficient tuning accuracy in complex interference scenarios. Third, its multi-frequency notch filter is susceptible to coupling effects between notch filters, resulting in bandwidth exceeding limits or abnormal passband attenuation. Fourth, its feedback correction mechanism lacks self-learning capabilities, failing to improve response speed when the same interference recurs, leading to insufficient dynamic adaptability and stability. These technical limitations mean that existing solutions cannot meet the application requirements of high-end electronic equipment in complex dynamic interference and strong noise superposition scenarios, necessitating the development of a full-link adaptive control method.

[0055] To address the aforementioned issues, this embodiment acquires the input signal of the adjustable notch filter group, dynamically suppresses strong electromagnetic noise using an adaptive Kalman filter, constructs a reinforcement learning state vector, and uses a deep Q-network reinforcement learning model to perform inference, generating optimal action commands for the tuning module. This controls the tuning module to adjust the parameters of the adjustable notch filter group. Finally, the deep Q-network reinforcement learning model and the adaptive Kalman filter are updated and adjusted based on actual performance indicators.

[0056] Therefore, this invention constructs a high-precision operation control method for electrically tunable notch filters based on the deep collaboration of adaptive Kalman filtering and deep Q-network. It utilizes adaptive Kalman filtering to dynamically adjust the noise covariance matrix, while employing a deep Q-network reinforcement learning model with a dual-network structure and a strategy of offline pre-training and online fine-tuning. The output stepper motor tuning commands enable the electrically tunable notch filter to stably maintain high-precision notch filtering performance in scenarios with complex dynamic interference and strong noise superposition. This completely solves many limitations of traditional solutions in terms of interference adaptation, precision control, multi-frequency collaboration, and stability, meeting the stringent requirements for strong interference suppression in fields such as communication and radar.

[0057] In a preferred embodiment, the step of the control processing module acquiring the input signal of the adjustable notch filter group and dynamically suppressing strong electromagnetic noise using an adaptive Kalman filter to obtain the denoised interference signal specifically includes:

[0058] S11: The control processing module uses two synchronous high-speed ADCs to acquire the original mixed signal and the pre-filtered reference signal from the input of the adjustable notch filter group, respectively, to obtain two discrete signal sequences.

[0059] S12: Using the constructed adaptive Kalman filter model, the noise covariance and observation noise covariance of the system are estimated by recursive least squares method, and the adaptive factor adjustment covariance matrix is ​​determined.

[0060] S13: Following the standard recursive process of adaptive Kalman filtering, the state prediction, prediction error covariance calculation, Kalman gain solution, state update and covariance update are executed sequentially, and the denoised interference signal is extracted from the updated state vector.

[0061] In this embodiment, the control processing module employs two synchronous high-speed ADCs (with a sampling frequency 10 times the upper limit of the operating frequency band to ensure distortion-free signal acquisition) to acquire the original mixed signal (containing normal signal, interference signal, and electromagnetic noise) from the input of the adjustable notch filter group, and the reference signal after being filtered by a 5th-order Chebyshev type I low-pass filter (to suppress out-of-band noise and provide a noise comparison benchmark), thus obtaining two discrete signal sequences. Subsequently, based on the constructed AKF model, the recursive least squares method is used to estimate the system noise covariance Q and the observation noise covariance R in real time. The covariance matrix is ​​dynamically adjusted using adaptive factors α and β to adapt the filtering parameters to noise changes. Following the standard AKF recursive process, state prediction, prediction error covariance calculation, Kalman gain solution, state update, and covariance update are executed sequentially. The denoised interference signal is extracted from the updated state vector, achieving accurate separation of noise and interference signals.

[0062] In the constructed AKF model's state and observation vectors, the state equation describes the dynamic changes in the interference signal's state (frequency and intensity), while the observation equation establishes the mapping relationship between the acquired signal and the interference state. The combination of these two equations enables accurate estimation of the interference state. To meet the core requirement of real-time tracking of the interference frequency and intensity when adapting to a notch filter, this is expressed in the following form:

[0063]

[0064] In the formula, This represents the state vector, containing the interference frequency at time k. With interference intensity These correspond to the core parameters of notch filter control; The state transition matrix (identity matrix) is assumed to be stable over a short period of time, which conforms to the dynamic characteristics of most disturbance scenarios, simplifying the calculation while ensuring the estimation accuracy. This represents the system noise driving matrix (identity matrix), used to convey the influence of system noise on the state; It represents system noise and characterizes the minute fluctuations of the interference source itself, which is consistent with the randomness of actual interference scenarios.

[0065]

[0066] In the formula, This represents the observation vector, corresponding to the original signal and reference signal acquired by the dual-channel ADC, thus improving the completeness and reliability of the observation. It represents the observation matrix (identity matrix), establishes a direct mapping between observations and state vectors, and simplifies the observation model; Indicates observation noise (Gaussian distribution) ), characterizing the noise during the ADC acquisition process, and dynamically estimating Achieve precise suppression.

[0067] The noise covariance and observation noise covariance of the system are estimated using the recursive least squares method, and dynamically updated using the recursive least squares method. and This allows the AKF model to adapt to dynamic changes in noise in real time, avoiding the attenuation of filtering effect caused by a fixed covariance matrix. It is expressed in the following form:

[0068]

[0069]

[0070] In the formula, Let k represent the noise covariance matrix at time k. Let the noise covariance matrix at time k-1 be represented. , This represents the adaptive factor, used to control the update rate of the covariance matrix. A smaller β value can avoid model oscillations caused by sudden noise changes and is suitable for the slow noise changes in notch filtering scenarios. The covariance of the prediction error reflects the uncertainty in state prediction. It is determined by its relationship with... The difference adjusts the system noise covariance, making Adapt to the actual situation of fluctuations in the fit; This represents the observation residual, reflecting the deviation between the observed and predicted values. The observation noise covariance is adjusted by the sum of its squares. ,make Adapts to real-time changes in ADC acquisition noise.

[0071] Therefore, this embodiment adopts a combination of dual-channel synchronous high-speed ADC acquisition and adaptive Kalman filtering. By acquiring dual signals (the original mixed signal and the preliminary filtered reference signal), signal integrity is improved. Then, the noise covariance is dynamically estimated using the recursive least squares method to achieve adaptive adjustment of the noise suppression parameters. Furthermore, the state vector of the adaptive Kalman filter is defined as a comprehensive consideration of the interference frequency and interference intensity, accurately adapting to the core control requirements of the notch filter, thus solving the problem that the traditional AKF state vector has strong generalization but insufficient scenario adaptability.

[0072] In a preferred embodiment, the step of extracting static and dynamic features from the interference signal and constructing a reinforcement learning state vector specifically includes:

[0073] S21: Perform an N-point FFT transform on the denoised interference signal to obtain the frequency domain signal, calculate the actual intensity and frequency of each frequency point, filter out the frequency points whose intensity meets the threshold, and obtain the static feature set.

[0074] S22: The Db4 wavelet basis is used to perform wavelet decomposition on the denoised interference signal, extract the detail coefficients of the corresponding working frequency band, detect the modulus maxima of the detail coefficients to locate the interference transition time, calculate the interference frequency transition rate and duration, and obtain the dynamic feature set.

[0075] S23: Normalize all features in the static feature set and the dynamic feature set respectively, and then combine the features to construct the reinforcement learning state vector.

[0076] In this embodiment, an N-point FFT transform is performed on the denoised interference signal to convert the time-domain signal into a frequency-domain signal. The actual frequency and intensity of each frequency point are calculated, and frequency points with an intensity ≥10dBm (the threshold for distinguishing interference from normal signals) are selected to form a static feature set. Subsequently, the denoised signal is decomposed into 5-level wavelet decompositions using the Db4 wavelet basis to extract the detail coefficients d3 (high-frequency detail coefficients reflect the dynamic changes of interference) corresponding to the 130-520MHz working frequency band. The interference transition time is located by detecting the modulus maxima of the detail coefficients, and the frequency transition rate and duration are calculated to form a dynamic feature set. Finally, the static and dynamic features are normalized by min-max respectively to eliminate the difference in dimensions, and then combined in the order of frequency, intensity, transition rate, and duration to construct the reinforcement learning state vector.

[0077] In practical applications, the execution process includes the following:

[0078] right Perform an N-point FFT transform to calculate the frequency domain signal. ( ), calculate the actual intensity at each frequency point With frequency ,filter The frequency points are used to obtain the static feature set. ( );

[0079] in, , , Indicates the strength of the selected samples The frequency value of the frequency point, , , This represents the actual intensity value of these frequency points, where k indicates which frequency point is selected; only 5 frequency points are selected here.

[0080] Using Db4 wavelet basis pairs Perform 5-level wavelet decomposition and analyze the detail coefficients. (Corresponding to the 130-520MHz frequency band) Perform modulus maxima detection to locate the interference transition moment. Calculate the interference frequency hopping rate With duration To obtain the dynamic feature set ;

[0081] Min-max normalization is applied to static and dynamic features to eliminate dimensional differences, and state vectors for reinforcement learning are constructed. ( , , The features are normalized. (For normalized duration).

[0082] Performing an N-point FFT transform on the denoised interference signal converts the time-domain interference signal into a frequency-domain signal, visually presenting the amplitude and phase of each frequency component, thus providing a foundation for static feature extraction. Its core is to utilize the frequency-domain analysis characteristics of the Fourier transform to accurately locate the interference frequency points, expressed in the following form:

[0083]

[0084] In the formula, This represents the noise-reduced interference signal, ensuring the purity of the input signal for the FFT transform and avoiding frequency misjudgment caused by noise. Indicates the number of FFT transform points ( The number of sampling points is consistent with that of the ADC (sampling frequency 5200MHz × sampling time 5ms), ensuring that the frequency domain resolution meets the frequency point identification requirements of the operating frequency band (110-520MHz). ); Frequency point index ( ), corresponding to the actual frequency This enables precise mapping between frequency indexes and actual frequencies. It represents a complex exponential function, characterizes the frequency domain decomposition kernel of the Fourier transform, and realizes the conversion of time-domain signals to the frequency domain.

[0085] Normalizing all features in both the static and dynamic feature sets involves standardizing the value ranges of features from different dimensions to the [0,1] interval, thus eliminating the interference of dimensional differences (such as frequency units MHz and intensity units dBm) on the model. The core principle is to maintain the relative relationships of features through linear transformation while unifying the feature magnitudes, expressed in the following form:

[0086]

[0087] In the formula, Indicates the original characteristic values ​​(e.g., frequency 110-520MHz, intensity 10-40dBm); , The minimum / maximum value of the feature is represented (obtained through statistics from historical interference datasets), ensuring that the feature value is stable in the [0,1] interval after normalization; This represents the normalized feature value. After eliminating the difference in dimensions, the DQN model can treat features of each dimension fairly and avoid the dominance of a feature of a large magnitude (such as frequency) in the model's decision.

[0088] Therefore, this embodiment proposes a dual-dimensional feature extraction scheme that combines static and dynamic features. Static features such as interference frequency and intensity are extracted through FFT transformation, while dynamic features such as frequency jump rate and duration are extracted through Db4 wavelet decomposition. This solves the problem that existing technologies can only extract basic static features of interference and cannot characterize dynamic characteristics. This enables reinforcement learning models to fully perceive the static attributes and dynamic change patterns of interference, providing sufficient basis for accurate decision-making.

[0089] In a preferred embodiment, based on the constructed reinforcement learning state vector, inference is performed using the deep Q-network reinforcement learning model obtained after training to generate the optimal action command steps for the tuning module, specifically including:

[0090] S31: Construct a fully connected deep Q-network structure; wherein, the fully connected deep Q-network structure includes an input layer for receiving state vectors, a hidden layer using the ReLU activation function, and an output layer for outputting the Q-values ​​of each action, and a dual-network structure consisting of an experience replay pool, a target network, and an evaluation network;

[0091] S32: The fully connected deep Q network structure is pre-trained offline using a historical interference dataset, and the network parameters are updated and evaluated using a temporal difference algorithm until the loss value is stable.

[0092] S33: Input the real-time constructed state vector into the deep Q-network reinforcement learning model obtained after training, and select the action with the largest output Q value as the optimal action instruction of the tuning module.

[0093] In this embodiment, a fully connected DQN structure is first constructed: the input layer ensures that the number of neurons is consistent with the dimension of the state vector; the hidden layers (2 layers, 64 / 32 neurons) use the ReLU activation function to introduce nonlinear fitting capability; and the number of neurons in the output layer is consistent with the dimension of the action space (step angle correction and motor drive current). Then, an experience replay pool is designed to store experience samples (state vector, action, reward value, new state vector) to avoid correlation in training data. A dual-network structure based on the target network and the evaluation network is designed, synchronizing parameters every 100 steps to avoid training oscillations. After this, the model is pre-trained offline using a dataset of 10,000 historical interference sets. The parameters of the evaluation network are updated using a temporal difference algorithm until the loss value is ≤0.01 (model convergence). During real-time inference, the constructed state vector is input into the pre-trained model, and the action with the largest output Q value is selected as the optimal tuning instruction.

[0094] The loss function of the constructed fully connected deep Q-network structure is used to update the evaluation network parameters sequentially by minimizing the mean squared error between the evaluated Q-value and the target Q-value, enabling the model to learn the mapping relationship between state, action, and reward. Its core is to utilize the idea of ​​temporal difference learning, taking into account both immediate and future rewards, and is expressed in the following form:

[0095]

[0096] In the formula, This indicates the batch size. 32 samples are randomly selected from the experience replay pool for batch updates, balancing training efficiency and gradient stability. The Q-value represents the evaluation of the network output, reflecting the current state. Next action Expected rewards; This represents the immediate reward value, calculated based on the notch effect metric, and directly reflects the action. The execution effect; This represents the discount factor, which weighs immediate rewards against future rewards. This indicates a greater emphasis on the cumulative rewards over multiple future steps, aligning with the long-term stability requirements of notch filter systems. This represents the Q-value output by the target network, providing a stable target Q-value and avoiding target fluctuations caused by updates to the network parameters. (Evaluate network parameters) (Target network parameters) represent updates achieved by minimizing the loss function. Every 100 steps will Sync to To ensure stable training.

[0097] The reward function of the constructed fully connected deep Q-network structure is a weighted reward function based on the core metrics of notch filtering (depth, bandwidth, passband attenuation, and response delay). It is used to quantify the performance of tuning actions and guide the model to learn the optimal tuning strategy. Its core is to reflect the priority of metrics through weight allocation, expressed in the following form:

[0098]

[0099] In the formula, Indicates the depth bonus of the notch ( ), The reward is 1, and the incentive model prioritizes notch depth (the core indicator for strong interference suppression). Indicates notch bandwidth bonus ( ), The time reward is 1, which constrains the notch bandwidth to not exceed the limit to avoid affecting the normal signal; Indicates passband decay reward ( ), The time reward is 1 to ensure minimal signal attenuation in the non-notch band. Indicates a response delay penalty ( ), The penalty is set to 0 to suppress actions that cause slow response in the model output. The weighting coefficients are calibrated through orthogonal experiments to reflect the priority of notch depth > bandwidth > passband attenuation > response delay, thus meeting the core requirement of strong interference suppression.

[0100] Therefore, this embodiment constructs a dual-network deep Q-network structure consisting of a target network and an evaluation network, coupled with an experience replay pool. This avoids data correlation and training oscillations during reinforcement learning training, improving model stability. The two-stage training strategy of offline pre-training and online fine-tuning ensures the initial decision accuracy of the model by pre-training it on a large-scale historical interference dataset, and then fine-tunes it online using real-time experience samples, allowing the model to continuously adapt to new scenarios. This design solves the problems of existing technologies where tuning parameter decisions rely on fixed calibration relationships and have poor adaptability. It enables tuning actions to autonomously adapt to different interference scenarios (e.g., single-frequency / multi-frequency, slow / fast transitions, low / high intensity), balancing the initial accuracy of the model with dynamic adaptability, ensuring the system's decision response speed and accuracy in complex interference scenarios.

[0101] In a preferred embodiment, the optimal action command is sent to the tuning module, which then parses the optimal action command to generate adaptive driving parameters. The adaptive driving parameters are then used to adjust the parameters of the adjustable notch filter group. Specifically, this includes:

[0102] S41: Send the optimal action command to the tuning module, and drive the tuning module to parse the optimal action command to obtain the target step angle and motor drive current of each notch filter.

[0103] S42: Generate driving parameters based on the analyzed motor drive current, and drive the stepper motor to rotate toward the target step angle;

[0104] S43: The actual step angle of the stepper motor is acquired in real time through the encoder, the deviation between the actual step angle and the target step angle is calculated, and the drive current is adjusted according to the magnitude of the deviation until the deviation meets the accuracy requirements.

[0105] In this embodiment, the tuning module receives the optimal action command via the EtherCAT bus, parses it after CRC32 verification, and obtains the target step angle (e.g., 9°) and motor drive current (e.g., 1.2A) for each notch filter. Based on the parsed drive current, pulse drive parameters are generated to drive the stepper motor to rotate towards the target step angle. The motor adjusts the core position of the passive resonant cavity within the adjustable notch filter group through a mechanical structure, synchronously changing the resonant frequency (notch frequency) and resonant intensity (notch depth). Afterward, the actual step angle of the motor is acquired in real time via an encoder (resolution 0.001°) to calculate the deviation. Target step angle - actual step angle, if (If the deviation is large) then increase the drive current by 0.2A (to speed up the correction). If the deviation is small, maintain the current (to ensure accuracy) until the deviation is ≤0.001°.

[0106] Therefore, this embodiment proposes a stepper motor drive strategy based on adaptive adjustment of deviation and current. It dynamically adjusts the motor drive current based on the deviation between the actual step angle and the target step angle, achieving a balance between tuning accuracy and response speed. This solves the problems of insufficient tuning accuracy (step angle deviation ≥ 0.03°) in existing stepper motor open-loop control technology, enabling step angle deviation to be controlled within ≤ 0.001°, corresponding to a frequency deviation ≤ 0.0001MHz and a notch depth deviation ≤ 0.025dB. Simultaneously, the design of synchronously adjusting two parameters with a single motor avoids the complex logic and failure risks of multi-component coordination, improving the stability and response speed of the tuning link and ensuring that notch parameters are quickly and accurately met.

[0107] In a preferred embodiment, the method further includes:

[0108] S44: Obtain the actual frequency and notch depth of each notch from the built-in sensor of the adjustable notch filter group, and feed the collected data back to the tuning module.

[0109] S45: The tuning module compares the feedback data with the target parameters and determines whether the deviation between the feedback data and the target parameters exceeds the standard. If so, it returns to repeat the stepper motor adjustment action until the parameters of each notch filter meet the standard.

[0110] In this embodiment, the actual frequency and notch depth of each notch filter are collected using the built-in sensors of the adjustable notch filter group (frequency measurement accuracy 0.0001MHz, depth measurement accuracy 0.01dB), and fed back to the tuning module via a communication link. Then, the tuning module compares the feedback data with the target parameters (target frequency corresponds to target step angle, such as 200MHz corresponding to 9°; target notch depth is calibrated by step angle association, such as 9° corresponding to 40dB) to determine whether the deviation exceeds the standard (frequency deviation > 0.001MHz or depth deviation > 0.025dB). If it exceeds the standard, the aforementioned stepper motor adjustment action is returned, and the process of driving, collecting, deviation calculation, and current adjustment is repeated until the deviation meets the standard.

[0111] Therefore, this embodiment uses the built-in sensors of the adjustable notch filter group to collect the actual notch filter parameters (frequency and depth) in real time, compare them with the target parameters, and perform repeated adjustments for deviations exceeding the standard. This solves the problem of no feedback calibration and uncorrectable deviations after the tuning action in the prior art, further reducing tuning errors and ensuring that the core indicators (frequency and depth) of each notch filter are stably met. At the same time, the feedback calibration mechanism improves the system's robustness to sudden factors such as mechanical wear and electromagnetic interference, and can quickly correct even small tuning deviations, maintaining the stability of the notch filter performance.

[0112] In a preferred embodiment, the method further includes:

[0113] S46: When in a multi-frequency notch filter scenario, adjust the stepper motors of each notch filter in order of interference jump rate from low to high. After each notch filter is adjusted, the step angle change of adjacent notch filters is collected by the encoder.

[0114] S47: Determine whether the change in step angle exceeds the coupling deviation threshold. If so, trigger the adjacent notch filter stepper motor to perform targeted adjustments based on the coupling deviation until the coupling deviation meets the requirements.

[0115] In this embodiment, when in a multi-frequency notch filter scenario (e.g., 3 frequencies: 150MHz, 250MHz, and 350MHz, with jump rates of 2MHz / ms, 5MHz / ms, and 3MHz / ms respectively), the stepper motors of each notch filter are adjusted sequentially according to the jump rate from low to high (150MHz→350MHz→250MHz). Then, after each notch filter adjustment is completed (e.g., 150MHz notch filter adjustment completed), the stepper motors of the adjacent notch filter (350MHz) are collected by the encoder. The change in the angle (e.g., 0.003°, i.e., coupling deviation) is then determined. Next, it is determined whether the coupling deviation exceeds the threshold of 0.002°. If so, the fine adjustment is calculated according to "correction amount = coupling deviation × 0.8" (0.003° × 0.8 = 0.0024°). The correction direction is opposite to the coupling offset direction (e.g., if coupling causes the step angle to increase, the fine adjustment amount is -0.0024°). The stepper motor of the adjacent notch is driven to perform the correction action with a microstep size of 0.0005° / step until the coupling deviation is ≤0.002°.

[0116] Therefore, this embodiment proposes a multi-frequency notch filter control strategy based on jump rate sorting adjustment and targeted fine-tuning of coupling deviation. The notch filters are adjusted in order of interference jump rate from low to high to reduce mutual interference during the adjustment process. Then, by collecting the step angle change (coupling deviation) of adjacent notch filters, reverse targeted fine-tuning is performed to suppress the coupling effect. This solves the problems of bandwidth overlap and excessive passband attenuation caused by the coupling effect between frequency points in the existing multi-frequency notch filter technology. The coupling deviation can be controlled within ≤0.002° (corresponding to a frequency deviation of 0.0002MHz and a depth deviation of 0.01dB).

[0117] Meanwhile, by adopting a combination strategy of orderly adjustment and targeted fine-tuning, the efficiency and accuracy of multi-frequency tuning are taken into account, ensuring that the performance indicators of each notch filter can be stably met in scenarios with 1-5 frequency points superimposed, thereby improving the system's multi-scenario adaptability.

[0118] In a preferred embodiment, the control processing module calculates the real-time reward value based on the actual performance index set of the adjustable notch filter group, constructs a new state vector by combining the collected interference features after executing the optimal action, and updates the deep Q-network reinforcement learning model and adjusts the adaptive factor of the adaptive Kalman filter. Specifically, this includes:

[0119] S51: The input and output power of the adjustable notch filter group are collected by a directional coupler and a power meter, the output signal spectrum is collected by a spectrum analyzer, the standing wave ratio is collected by a standing wave ratio tester, the response time is recorded by a timer, and the actual performance indexes are integrated to obtain the set of actual performance indicators.

[0120] S52: Collect the interference features after executing the optimal action, combine them with the real-time reward value to construct a new state vector, update the deep Q-network reinforcement learning model, and adjust the adaptive factor of the adaptive Kalman filter.

[0121] In this embodiment, a directional coupler and a power meter (accuracy 0.1dBm) are used to acquire the input power of the adjustable notch filter array. With output power Used to calculate notch depth ( Then, the spectrum of the output signal is acquired using a spectrum analyzer (resolution bandwidth 10kHz), and the notch bandwidth is extracted. (Frequency range corresponding to 3dB attenuation) and passband attenuation (Maximum attenuation in the non-notch band); then, the input and output standing wave ratios are acquired using a standing wave ratio (SWR) meter (accuracy 0.01). The signal matching degree was evaluated; and the response time from interference identification to tuning completion was recorded using a timer (accuracy 1μs). Finally, the above indicators are integrated to form a set of actual effect indicators, providing data support for subsequent reward value calculation and model updates.

[0122] Therefore, this embodiment adopts a multi-dimensional, comprehensive notch filter performance indicator acquisition scheme, integrating five core indicators: input / output power, notch bandwidth, passband attenuation, VSWR, and response time, to form a set of actual performance indicators, providing a comprehensive basis for model updates and parameter adjustments. This ensures that model updates and parameter adjustments take into account all dimensions of notch filter performance, avoiding the overshooting of other indicators due to optimization of a single indicator. It also provides a reliable foundation for calculating reinforcement learning reward values, ensuring that the reward values ​​accurately reflect the notch filter performance and guiding the model to learn the optimal tuning strategy.

[0123] Based on this, the interference features after executing the optimal action are collected, and a new state vector is constructed by combining it with the real-time reward value. This process updates the deep Q-network reinforcement learning model and adjusts the adaptive factor of the adaptive Kalman filter. Specifically, this includes:

[0124] S521: Substitute the set of actual performance indicators into the reward function to calculate the real-time reward value, and collect the interference features after executing the optimal action to construct a new state vector;

[0125] S522: The original state vector, optimal action, real-time reward value and new state vector are combined to form an experience sample and stored in the experience replay pool. Every time a preset number of samples are stored, the deep Q network is triggered to update online.

[0126] S523: Determine the range of the real-time reward value. If the reward value is lower than the first threshold, adjust the adaptive factor of the adaptive Kalman filter to optimize the noise suppression effect, re-extract features, and trigger reinforcement learning to make a new decision. If the reward value is between the first and second thresholds, keep the adaptive Kalman filter parameters unchanged, adjust the parameters of the Q-network reinforcement learning model, and speed up the online update of the reinforcement learning network. If the reward value is higher than the second threshold, maintain the current parameters and store the optimal stepper motor adjustment strategy for reuse in similar interference scenarios.

[0127] In this embodiment, the set of actual performance indicators is substituted into the reward function to calculate the real-time reward value. Collect the interference features (new static and dynamic features) after executing the optimal action to construct a new state vector. Then the original state vector Optimal action Reward Value New state vector Experience samples are compiled and stored in the experience replay pool. Every 32 samples stored triggers an online update of the DQN model; then, a judgment is made. Location within the interval: If (Poor notch filtering effect), adjust the adaptive factors α and β of AKF (α increased to 0.015, β increased to 0.008) to optimize noise suppression, re-extract features and trigger DQN to re-determine; if (Notch filtering works well), keeping the AKF parameters constant, the DQN update speed is accelerated by increasing the learning rate, shortening the target network synchronization cycle, and increasing the batch sample size; if (Excellent notch filtering effect), maintain the current parameters, and store the corresponding tuning strategy (step angle correction, motor drive current) for direct reuse in similar interference scenarios.

[0128] Therefore, this embodiment, through a dual-technology collaborative optimization strategy based on reward value ranges, can dynamically adjust the AKF adaptive factor and DQN model parameters according to the real-time reward value range (below 0.7, 0.7-0.9, above 0.9), achieving a synergistic improvement in noise suppression and decision-making capabilities. This dual-technology collaborative optimization strategy avoids the limitations of single-technology optimization; the noise suppression effect of AKF and the decision-making accuracy of DQN mutually reinforce each other. It also addresses the problems of existing technologies lacking self-learning capabilities and failing to improve response speed when the same interference recurs, enabling the system to continuously optimize performance through real-time experience samples.

[0129] Reference Figure 2 , Figure 2 This is a schematic diagram of the anti-interference electrically tunable notch filter operation control system according to an embodiment of the present invention. Figure 2As shown, in an optional embodiment, the present invention also proposes an anti-interference electrically tunable notch filter operation control system, comprising: a control processing module 10, an adjustable notch filter group 20, and a tuning module 30, wherein the control processing module 10, the adjustable notch filter group 20, and the tuning module 30 are configured to execute the anti-interference electrically tunable notch filter operation control method as described above.

[0130] Specifically, the control processing module 10 acquires the input signal and status data of the adjustable notch filter group 20, namely frequency and depth, and generates an optimal tuning command including step angle and drive current. The drive tuning module 30 adjusts the parameters of the adjustable notch filter group 20 based on the stepper motor drive signal. The tuning module 30 feeds back the stepper motor operating status to the control processing module 10, and the adjustable notch filter group 20 also feeds back the actual notch filter status to the control processing module 10, which then performs adaptive adjustment of the model parameters.

[0131] Other embodiments or specific implementations of the anti-interference electrically adjustable notch filter operation control system of the present invention can be referred to the above-described method embodiments, and will not be repeated here.

[0132] It is understood that in the description of this specification, references to terms such as "one embodiment," "another embodiment," "other embodiments," or "first embodiment to Nth embodiment," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0133] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0134] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for controlling the operation of an anti-interference electrically tunable notch filter, characterized in that, Includes the following steps: The control processing module acquires the input signal of the adjustable notch filter group and uses an adaptive Kalman filter to dynamically suppress strong electromagnetic noise to obtain the noise-reduced interference signal. Extract the static and dynamic features from the interference signal to construct a reinforcement learning state vector; Based on the constructed reinforcement learning state vector, the deep Q-network reinforcement learning model obtained after training is used to perform inference and generate the optimal action instructions for the tuning module. The optimal action command is sent to the tuning module, which then parses the optimal action command, generates adaptive driving parameters, and uses the adaptive driving parameters to adjust the parameters of the adjustable notch filter group. The control processing module calculates the real-time reward value based on the actual performance index set of the adjustable notch filter group, constructs a new state vector by combining the collected interference features after executing the optimal action, updates the deep Q-network reinforcement learning model, and adjusts the adaptive factor of the adaptive Kalman filter.

2. The anti-interference electrically tunable notch filter operation control method as described in claim 1, characterized in that, The control processing module acquires the input signal of the adjustable notch filter group, dynamically suppresses strong electromagnetic noise using an adaptive Kalman filter, and obtains the denoised interference signal. The specific steps include: The control processing module uses two synchronous high-speed ADCs to acquire the original mixed signal and the pre-filtered reference signal from the input of the adjustable notch filter group, respectively, to obtain two discrete signal sequences. Using the constructed adaptive Kalman filter model, the noise covariance and observation noise covariance of the system are estimated by recursive least squares method, and the adaptive factor adjustment covariance matrix is ​​determined. Following the standard recursive process of adaptive Kalman filtering, the following steps are executed sequentially: state prediction, prediction error covariance calculation, Kalman gain solution, state update and covariance update. The denoised interference signal is then extracted from the updated state vector.

3. The anti-interference electrically tunable notch filter operation control method as described in claim 1, characterized in that, The steps for extracting static and dynamic features from the interference signal and constructing a reinforcement learning state vector specifically include: An N-point FFT transformation is performed on the denoised interference signal to obtain the frequency domain signal. The actual intensity and frequency of each frequency point are calculated, and the frequency points whose intensity meets the threshold are selected to obtain the static feature set. The Db4 wavelet basis is used to perform wavelet decomposition on the denoised interference signal, extract the detail coefficients of the corresponding working frequency band, detect the modulus maxima of the detail coefficients to locate the interference transition time, calculate the interference frequency transition rate and duration, and obtain the dynamic feature set. All features in the static and dynamic feature sets are normalized separately, and then combined to construct the reinforcement learning state vector.

4. The anti-interference electrically tunable notch filter operation control method as described in claim 1, characterized in that, Based on the constructed reinforcement learning state vector, the deep Q-network reinforcement learning model obtained after training is used to perform inference and generate the optimal action command steps for the tuning module, specifically including: Construct a fully connected deep Q-network structure; wherein the fully connected deep Q-network structure includes an input layer for receiving state vectors, a hidden layer using the ReLU activation function, and an output layer for outputting the Q-values ​​of each action, and an experience replay pool combined with a target network and an evaluation network in a dual network structure; The fully connected deep Q network structure was pre-trained offline using a historical interference dataset, and the network parameters were updated and evaluated using a temporal difference algorithm until the loss value stabilized. The state vector constructed in real time is input into the deep Q-network reinforcement learning model after training, and the action with the largest output Q value is selected as the optimal action instruction of the tuning module.

5. The anti-interference electrically tunable notch filter operation control method as described in claim 1, characterized in that, The optimal action command is sent to the tuning module, which then parses the optimal action command, generates adaptive drive parameters, and uses these adaptive drive parameters to adjust the parameters of the adjustable notch filter group. Specifically, this includes: The optimal action command is sent to the tuning module, which then analyzes the optimal action command to obtain the target step angle and motor drive current of each notch filter. The driving parameters are generated based on the analyzed motor drive current, and the stepper motor is driven to rotate toward the target step angle. The actual step angle of the stepper motor is collected in real time by the encoder, the deviation between the actual step angle and the target step angle is calculated, and the drive current is adjusted according to the magnitude of the deviation until the deviation meets the accuracy requirements.

6. The anti-interference electrically tunable notch filter operation control method as described in claim 5, characterized in that, The method further includes: The actual frequency and notch depth of each notch are acquired by the built-in sensor of the adjustable notch filter group, and the acquired data is fed back to the tuning module. The tuning module compares the feedback data with the target parameters to determine whether the deviation between the feedback data and the target parameters exceeds the standard. If so, it returns to repeat the stepper motor adjustment action until the parameters of each notch filter meet the standard.

7. The anti-interference electrically tunable notch filter operation control method as described in claim 6, characterized in that, The method further includes: When in a multi-frequency notch filter scenario, the stepper motors of each notch filter are adjusted sequentially in order of interference jump rate from low to high. After each notch filter is adjusted, the step angle change of adjacent notch filters is collected by the encoder. If the step angle change exceeds the coupling deviation threshold, the adjacent notch filter stepper motor is triggered to perform targeted adjustments based on the coupling deviation until the coupling deviation meets the requirements.

8. The anti-interference electrically tunable notch filter operation control method as described in claim 1, characterized in that, The control processing module calculates the real-time reward value based on the actual performance index set of the adjustable notch filter group, constructs a new state vector by combining the collected interference features after executing the optimal action, and updates the deep Q-network reinforcement learning model and adjusts the adaptive factor of the adaptive Kalman filter. Specifically, this includes: The input and output power of the adjustable notch filter group are collected by a directional coupler and a power meter, the output signal spectrum is collected by a spectrum analyzer, the standing wave ratio is collected by a standing wave ratio tester, and the response time is recorded by a timer. The results are integrated to obtain a set of actual performance indicators. The interference features after the optimal action is executed are collected, and a new state vector is constructed by combining it with the real-time reward value. This vector is then used to update the deep Q-network reinforcement learning model and adjust the adaptive factor of the adaptive Kalman filter.

9. The anti-interference electrically tunable notch filter operation control method as described in claim 8, characterized in that, The steps include: collecting interference features after executing the optimal action, constructing a new state vector by combining it with the real-time reward value, updating the deep Q-network reinforcement learning model, and adjusting the adaptive factor of the adaptive Kalman filter. Substitute the set of actual performance indicators into the reward function to calculate the real-time reward value, and collect the interference features after executing the optimal action to construct a new state vector; The original state vector, optimal action, real-time reward value and new state vector are combined to form an experience sample and stored in the experience replay pool. Every time a preset number of samples are stored, the deep Q network is triggered to update online. The system determines the range of the real-time reward value. If the reward value is below the first threshold, the adaptive factor of the adaptive Kalman filter is adjusted to optimize the noise suppression effect, features are re-extracted, and reinforcement learning is triggered to make a new decision. If the reward value is between the first and second thresholds, the adaptive Kalman filter parameters are kept unchanged, and the parameters of the Q-network reinforcement learning model are adjusted to accelerate the online update speed of the reinforcement learning network. If the reward value is above the second threshold, the current parameters are maintained, and the optimal stepper motor adjustment strategy is stored for reuse in similar interference scenarios.

10. An anti-interference electrically tunable notch filter operation control system, characterized in that, include: The control processing module, the adjustable notch filter group, and the tuning module are configured to execute the anti-interference electrically adjustable notch filter operation control method as described in any one of claims 1-9.

Images