An encoder dynamic intelligent crosstalk adaptive suppression method

Abstract

This invention discloses a dynamic intelligent crosstalk adaptive suppression method for encoders. It involves real-time acquisition of multimodal state signals from the encoder and feature extraction to obtain sensing features. These sensing features are input into a crosstalk intensity prediction neural network, which outputs predicted crosstalk intensity values ​​for multiple future time steps. The sensing features and the predicted crosstalk intensity values ​​together form a state vector, and a soft actor-critic algorithm is used to output an optimal bias voltage control signal. Based on the optimal bias voltage control signal, the bias voltage of the varactor diode in the tunable repeater resonator is adjusted to change the resonant frequency of the tunable repeater resonator, minimizing the equivalent coupling coefficient between encoder signal channels and suppressing encoder crosstalk. By fusing a time-series deep learning prediction model with a reinforcement learning closed-loop optimization strategy and introducing a tunable repeater resonator, it achieves pre-prediction of crosstalk and active suppression of its physical sources, significantly improving the accuracy and reliability of the encoder in dynamic electromagnetic environments.

Description

Technical Field

[0001] This invention relates to a dynamic intelligent crosstalk adaptive suppression method for encoders, belonging to the field of electromagnetic compatibility and sensor technology. Background Technology

[0002] As the core position feedback unit of a closed-loop control system, encoders are widely used in various electronic devices and systems. However, with the current trend of miniaturization and high integration of devices, encoders face severe electromagnetic compatibility challenges. Within a limited physical space (typically smaller than 50mm×50mm×20mm), the multiple signal channels of the encoder form dense electromagnetic coupling paths. Crosstalk caused by near-field coupling directly degrades the signal-to-noise ratio and introduces angle measurement errors (typically exceeding 0.1°), which can lead to communication errors and control oscillations in severe cases. Currently, commonly used crosstalk suppression techniques mainly include passive shielding and isolation, fixed compensation networks, and signal post-processing filtering. Passive shielding and isolation techniques require the use of metal shielding covers and other devices, which occupy valuable space and have poor adaptability. Fixed compensation networks achieve crosstalk suppression by designing fixed capacitor or inductor compensation networks, but they are prone to failure when the environment changes. Signal post-processing filtering techniques use digital filters or adaptive notch filters, but they cannot effectively distinguish between crosstalk and valid signals, and their ability to suppress non-stationary and broadband crosstalk is insufficient. Summary of the Invention

[0003] The present invention provides a dynamic intelligent crosstalk adaptive suppression method for encoders to solve the problems existing in the prior art.

[0004] The technical solutions adopted in this invention are as follows:

[0005] A dynamic intelligent crosstalk adaptive suppression method for encoders includes the following steps:

[0006] S1: Acquire the multimodal state signals of the encoder and extract features to obtain perceptual features;

[0007] S2: Input the perceived features into the crosstalk intensity prediction neural network and output the predicted crosstalk intensity values ​​for multiple future time steps;

[0008] S3: Combine the perceived features with the predicted crosstalk intensity values ​​for multiple future time steps to form a state vector, and use the soft actor-commentator algorithm to output the optimal bias voltage control signal;

[0009] S4: Adjust the bias voltage of the varactor diode in the tunable relay resonator according to the optimal bias voltage control signal to change the resonant frequency of the tunable relay resonator, so as to minimize the equivalent coupling coefficient between encoder signal channels and suppress encoder crosstalk.

[0010] Further, in S2, the crosstalk intensity prediction neural network is a frequency-decoupled dual-path parallel network, including a frequency decoupling module, a low-pass filter branch, and a high-pass filter branch. The frequency decoupling module includes a low-pass filter with a cutoff frequency of 10kHz and a high-pass filter with a cutoff frequency of 50kHz. The low-pass filter branch is used to extract and predict features of the crosstalk components after filtering by the low-pass filter, and the high-pass filter branch is used to extract and predict features of the crosstalk components after filtering by the high-pass filter. The prediction results of the low-pass filter branch and the high-pass filter branch are fused to obtain the predicted crosstalk intensity values ​​for the next multiple time steps.

[0011] Furthermore, the low-pass filter branch includes a first dilated convolutional multi-scale module, a first 1×1 convolutional fusion layer, a unidirectional long short-term memory network, and a first fully connected output layer. The first dilated convolutional multi-scale module is used to extract multi-scale features from the crosstalk components after low-pass filtering. The first 1×1 convolutional fusion layer is used to perform channel fusion on the output of the first dilated convolutional multi-scale module. The unidirectional long short-term memory network is used to capture temporal dependencies. The first fully connected output layer is used to output the prediction results of the low-pass filter branch.

[0012] Furthermore, the high-pass filter branch includes a second dilated convolutional multi-scale module, a second 1×1 convolutional fusion layer, a temporal convolutional network, and a second fully connected output layer. The second dilated convolutional multi-scale module is used to extract multi-scale features from the crosstalk components after high-pass filtering. The second 1×1 convolutional fusion layer is used to perform channel fusion on the output of the second dilated convolutional multi-scale module. The temporal convolutional network is used to capture transient responses. The second fully connected output layer is used to output the prediction results of the high-pass filter branch.

[0013] Furthermore, the soft actor-critic algorithm aims to maximize the weighted sum of the expected cumulative reward and the policy entropy. The expected cumulative reward is calculated by a reward function, which is a weighted sum of crosstalk suppression, action smoothness, and power consumption terms. The formula for calculating the reward function is as follows:

[0014] ,

[0015] in, For a moment The reward value, For a moment The measured crosstalk intensity, For a moment The optimal bias voltage control signal. For a moment The optimal bias voltage control signal. This is the crosstalk suppression coefficient. The motion smoothness coefficient. This is the power consumption factor.

[0016] Furthermore, the tunable relay resonator adopts a structure of a planar spiral inductor and a varactor diode connected in parallel.

[0017] Furthermore, the resonant frequency of the tunable relay resonator Determined by the following formula:

[0018] ,

[0019] in, It is a planar spiral inductor. This represents the total capacitance.

[0020] Furthermore, the total capacitance is composed of a varactor diode capacitor and a parasitic capacitance connected in parallel. The varactor diode capacitor is adjusted by the bias voltage, and the capacitance-voltage characteristic of the varactor diode is expressed by the following formula:

[0021] ,

[0022] in, It is a varactor diode capacitor. Zero-bias junction capacitance Input voltage, For built-in potential, These are the gradient coefficients.

[0023] Furthermore, the equivalent coupling coefficient It can be expressed by the following formula:

[0024] ,

[0025] in, For direct coupling coefficients, Let be the coupling coefficient between signal channel A and the tunable repeater resonator. This represents the coupling coefficient between the B signal channel and the tunable repeater resonator. The detuned frequency, This refers to the relay loss rate. It is the imaginary unit.

[0026] Furthermore, the multimodal state signal includes electrical characteristics, environmental parameters, and mechanical parameters. The electrical characteristics include the time-domain waveforms of each signal channel of the encoder and the power rail noise. The environmental parameters include temperature, and the mechanical parameters include instantaneous rotational speed.

[0027] Furthermore, the feature extraction includes extracting the root mean square value and the crosstalk band energy ratio. The root mean square value is calculated based on the time-domain waveform, and the crosstalk band energy ratio is calculated based on the power spectral density of each signal channel of the encoder in the crosstalk-sensitive frequency band.

[0028] The present invention has the following beneficial effects:

[0029] (1) In view of the problem that passive shielding and isolation technologies occupy valuable space and have poor adaptability, this invention introduces a tunable relay resonator to actively control the electromagnetic coupling path from the physical source, so that direct coupling and indirect coupling will cause destructive interference. There is no need to set up additional isolation devices such as metal shielding covers, which effectively saves the limited physical space inside the encoder and realizes active physical suppression of crosstalk.

[0030] (2) To address the problem that fixed compensation networks are prone to failure when the environment changes, this invention adopts a soft actor-critic algorithm to incorporate the predicted crosstalk intensity values ​​of multiple future time steps into the state space, output the optimal bias voltage control signal online, and dynamically adjust the resonant frequency of the tunable relay resonator to minimize the equivalent coupling coefficient between encoder signal channels in real time. This enables the system to adapt to changes in temperature, speed and electromagnetic environment, and significantly improves the environmental adaptability of crosstalk suppression.

[0031] (3) In view of the problem that signal post-processing filtering technology cannot effectively distinguish crosstalk from valid signals and has insufficient ability to suppress non-stationary broadband crosstalk, the present invention uses a crosstalk intensity prediction neural network to predict the future crosstalk intensity in advance, and uses the predicted crosstalk intensity values ​​of multiple future time steps and the current sensing features to form a state vector, so that the soft actor-critic algorithm can perceive the crosstalk trend in advance and make pre-compensation actions, effectively distinguish and suppress crosstalk components, and significantly improve the crosstalk suppression accuracy and robustness of the encoder under non-stationary broadband electromagnetic interference conditions.

[0032] (4) This invention achieves intelligent adaptive suppression of the sensing layer, prediction layer, decision layer and execution layer by real-time acquisition of encoder multimodal state signals, pre-prediction of future crosstalk intensity, online optimization of bias voltage control action and dynamic adjustment of the resonant frequency of tunable relay resonator. This significantly reduces the encoder angle measurement error and improves the accuracy and reliability of the encoder in complex dynamic electromagnetic environment. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the overall structure of the present invention.

[0034] Figure 2 A schematic diagram of the neural network structure for predicting crosstalk intensity.

[0035] Figure 3 This is a schematic diagram of the LSTM structure.

[0036] Figure 4 This is a schematic diagram of the Temporal Convolutional Network (TCN).

[0037] Figure 5 This is a schematic diagram of a tunable repeater resonator. Detailed Implementation

[0038] The invention will now be further described with reference to the accompanying drawings.

[0039] This invention proposes a dynamic intelligent crosstalk adaptive suppression method for encoders. By fusing a temporal deep learning prediction model with a reinforcement learning closed-loop optimization strategy and introducing a tunable relay resonator, it continuously suppresses crosstalk in confined spaces and dynamic environments, significantly improving the accuracy and reliability of the encoder. The proposed method adopts a hierarchical closed-loop adaptive suppression architecture, consisting of a perception layer, a prediction layer, a decision layer, and an execution layer. The perception layer acquires multimodal state signals from the encoder; the prediction layer uses a deep temporal network to predict future crosstalk intensity; the decision layer uses a deep reinforcement learning agent to output the optimal control action; and the execution layer changes the electromagnetic coupling path through the tunable relay resonator. The overall system structure is as follows: Figure 1 As shown. The method of the present invention will be described in detail below.

[0040] S1: Real-time sensing of encoder multimodal status.

[0041] Real-time multimodal state perception provides comprehensive real-time data for crosstalk prediction and control, and is fundamental to achieving environmental adaptation and pre-compensation, improving suppression accuracy and robustness. This invention acquires the encoder's internal state parameter set in real time, mainly including:

[0042] Electrical characteristics, mainly including the time-domain waveforms of each signal channel: Phase A signal B-phase signal Index signal and power rail noise ;

[0043] Environmental parameters, mainly temperature ;

[0044] Mechanical parameters, mainly instantaneous speed By analyzing the A-phase signal Real-time calculation of zero-crossing detection.

[0045] Based on a certain period, the following features are extracted from the above signals:

[0046] (1) Root mean square value ,in The root mean square value is calculated as follows:

[0047] ,

[0048] in, For the feature extraction period, in this invention , The sampling point number, This represents the sampling time interval.

[0049] (2) Crosstalk band energy ratio The calculation method is shown in the following formula:

[0050] ,

[0051] in, , These are the A-phase signals. B-phase signal The power spectral density, This refers to the pre-defined crosstalk-sensitive frequency band.

[0052] Concatenate the above features into a feature vector to obtain:

[0053] ,

[0054] Standardize each component, and let the standardized perceptual feature be . ,but:

[0055] ,

[0056] in, For feature component index, , These are the mean and variance, respectively, obtained from the dataset statistics.

[0057] S2: Neural network for predicting crosstalk intensity.

[0058] This invention uses algorithms such as convolutional neural networks and unidirectional long short-term memory networks in series to design a frequency-decoupled dual-path parallel network structure. It processes crosstalk components filtered by low-pass filters and crosstalk components filtered by high-pass filters, respectively. Both paths use dilated convolution to construct multi-scale feature extraction modules, maintaining a sufficient temporal receptive field while controlling the amount of computation.

[0059] S21: Network structure.

[0060] Network structure such as Figure 2As shown, it consists of a frequency decoupling module, a low-pass filter branch, a high-pass filter branch, and a fusion output module. The frequency decoupling module consists of a low-pass filter and a high-pass filter. The cutoff frequency of the low-pass filter is 10kHz, and the cutoff frequency of the high-pass filter is 50kHz. Both filters adopt an FIR equiripple design and have an order of 32.

[0061] The low-pass filter branch is responsible for extracting and predicting crosstalk components after low-pass filtering, such as changes in coupling coefficients caused by temperature drift. Its structure consists of, in sequence, a first dilated convolutional multi-scale module, a first 1×1 convolutional fusion layer, a unidirectional long short-term memory network (unidirectional LSTM), and a first fully connected output layer.

[0062] Combination Figure 3 The low-pass filter branch contains two branches. To expand the receptive field without increasing the kernel size, the low-pass filter branch uses dilated convolution. Branch 1 has a kernel size of 3 and a dilation rate of 1; Branch 2 has a kernel size of 5 and a dilation rate of 2, achieving an equivalent receptive field of 9. Both branches have a convolution stride of 1 and 16 output channels. The ReLU activation function is used, and each convolution is followed by batch normalization. Then, the outputs of the two branches are concatenated along the channel dimension. The first 1×1 convolutional fusion layer linearly fuses the concatenated data using a 1×1 convolution, with 32 output channels, no activation function, and adaptively adjusting the weights at each scale. A unidirectional long short-term memory network is used to capture long-term crosstalk factors such as temperature drift and aging. Its update follows the formula:

[0063] ,

[0064] in, The features output by the first 1×1 convolutional fusion layer, The hidden state output at the current moment. This is the hidden state from the previous moment. This refers to the current state of memory cells. This represents the state of memory cells from the previous moment. This is the candidate memory vector for the current moment, representing the new information to be added; To offset the forget gate, For input gate bias, Output for the forget gate. For input gate output, , , These are weight matrices for the forget gate, input gate, and candidate memory, respectively, used to map the current input to the previous hidden state as gating signals to control information retention, updating, and memory generation. It is the Sigmoid activation function. The hyperbolic tangent activation function is used. For output gate output, This is used to bias the output gate.

[0065] The first fully connected output layer consists of two layers. The first layer has an input dimension of 64 and an output dimension of 32, using ReLU as the activation function. The second layer has an input dimension of 32 and an output dimension of 10.

[0066] The high-pass filtering branch is responsible for extracting and predicting crosstalk components after high-pass filtering, such as edge spikes and power supply glitches. Its structure is similar to the low-pass filtering branch, including a second dilated convolutional multi-scale module, a second 1×1 convolutional fusion layer, a temporal convolutional network (TCN), and a second fully connected output layer. The second dilated convolutional multi-scale module also contains two branches: branch 1 has a kernel size of 3 and a dilation rate of 1; branch 2 has a kernel size of 3 and a dilation rate of 2, resulting in an equivalent receptive field of 5. Feature fusion is then achieved using the second 1×1 convolutional fusion layer.

[0067] Temporal convolutional networks employ dilated causal convolutions, enabling parallel processing of the entire sequence and exhibiting fast response to high-frequency transients. They consist of two dilated convolutional layers, each with a kernel size of 3 and dilation rates of 1 and 2, respectively. ReLU is used as the activation function, and residual structures are added. The structure of the Temporal Convolutional Network (TCN) is as follows: Figure 4 As shown.

[0068] The second fully connected output layer consists of two layers. The first layer has an input dimension of 64 and an output dimension of 32, using ReLU as the activation function. The second layer has an input dimension of 32 and an output dimension of 10.

[0069] The prediction results of the low-pass filter branch and the high-pass filter branch are added element by element, and the final output is the predicted value of crosstalk intensity for the next 10 time steps.

[0070] S22: Network training.

[0071] To address the characteristics of encoder crosstalk, this invention employs a combination of Huber loss and L2 regularization as the network loss function. Huber's penalty for large errors is linear, preventing outliers (such as abnormal crosstalk spikes caused by electromagnetic pulse interference) from dominating the gradient, thus improving training robustness. The loss function is shown in the following equation:

[0072] ,

[0073] in, To predict the number of time steps, i.e. the number of future crosstalk values ​​output by the model at one time, in this invention... , For time step index, For the actual crosstalk strength, To predict crosstalk strength, For Huber's losses, The Huber threshold parameter is used in this invention. , , These are the network weights for the low-pass filter branch and the high-pass filter branch, respectively. , The regularization coefficient is used in this invention. , For L2 norm squaring operations, This refers to the current moment.

[0074] Training data consisted of 500,000 samples generated using co-simulation with COMSOL and MATLAB, covering temperatures from -20°C to 85°C, rotation speeds from 0 to 6000 rpm, and manufacturing tolerances of ±10%. The Adam optimizer was employed. , The values ​​were set to 0.9 and 0.999 respectively, with an initial learning rate of 1×10⁻⁶. -3 The batch size is 256, and the number of training rounds is 50.

[0075] To verify the performance of the crosstalk intensity prediction neural network of this invention, the same dataset was used, and a comparison was made with neural network architectures such as LSTM, TCN, and BiLSTM. The results are shown in Table 1.

[0076] Table 1 Comparison of prediction performance of different network models

[0077]

[0078] The experimental results show that the root mean square error of the prediction network of this invention is reduced by 32.2% and 23.9% compared with pure LSTM and CNN-LSTM (ordinary convolution), respectively, indicating that the dilated convolution and frequency decoupling design has substantial gains in capturing multi-scale temporal patterns of crosstalk. A score of 0.953 means the model can explain over 95% of the crosstalk intensity variation. Compared to BiLSTM, although BiLSTM utilizes future information, the prediction network of this invention still achieves a lower prediction error.

[0079] S3: Encoder crosstalk compensation action decision.

[0080] In the encoder dynamic intelligent crosstalk adaptive suppression method, the compensation action decision layer is responsible for mapping the current system state and crosstalk prediction values ​​to the optimal control action, thereby driving the tunable relay resonator to achieve dynamic control of the electromagnetic coupling path. This invention uses a Markov decision process to model this problem and employs a soft actor-critic algorithm to solve for the optimal strategy in the continuous action space.

[0081] S31: Markov decision process modeling.

[0082] First, define a quadruple. , which are the state space, action space, state transition probability, and reward function, respectively.

[0083] state space The state vector at time t It is composed of the current perceived features and the predicted value of future crosstalk intensity, which is:

[0084] ,

[0085] in, The perceptual features at the current moment, This represents the predicted future crosstalk value output by the crosstalk intensity prediction neural network in S2.

[0086] The action in the action space is a single continuous variable, representing the normalized bias voltage control signal. Let time... action The actual bias voltage is ,but:

[0087] ,

[0088] in, , These are the minimum voltage and the maximum voltage, respectively.

[0089] This invention uses model-free reinforcement learning, where state transition probabilities do not directly participate in the update calculation of policy or value functions.

[0090] reward function The reward function at time t is calculated as a weighted sum of three terms, guiding the agent to simultaneously optimize crosstalk suppression, action smoothness, and power consumption.

[0091] ,

[0092] in, Indicates time The measured crosstalk intensity, , , These are the crosstalk suppression coefficient, the smoothness coefficient, and the power consumption coefficient, respectively, and in this invention, their values ​​are 1.0, 0.1, and 0.05, respectively.

[0093] S32: Soft Actor-Critic (SAC) control.

[0094] The Soft Actor-Critic Algorithm (SAC algorithm) learns a stochastic optimal policy by maximizing the weighted sum of the expected cumulative reward and the policy entropy, thus achieving adaptive closed-loop control of the bias voltage of the encoder-tunable relay resonator. The objective function of the Soft Actor-Critic Algorithm is:

[0095] ,

[0096] in, Indicates the current moment. As a strategy, For state, For action, For expectation operator, This refers to the temperature parameter. For instant rewards, Let be the policy entropy. The soft actor-critic algorithm employs a maximum entropy framework, which offers better exploration capabilities and robustness to hyperparameters compared to traditional deterministic policy gradients.

[0097] The soft actor-critic algorithm internally maintains five multilayer perceptrons to sample actions and perform operations such as experience replay pooling, critic updates, and action updates. After algorithm training is complete, only the Actor network needs to be deployed for online decision-making and inference deployment. The inference process is as follows:

[0098] (1) Collect the current status ;

[0099] (2) Calculate the mean of motion during one forward propagation and the logarithmic standard deviation of the action ;

[0100] (3) Calculate deterministic actions ;

[0101] (4) Linearly map at to The output is sent to a digital-to-analog converter.

[0102] S4: Execution of crosstalk suppression.

[0103] To suppress encoder crosstalk, this invention introduces a tunable repeater resonator to actively control the electromagnetic coupling path. Crosstalk suppression is achieved by adjusting the tunable repeater resonator to dynamically adjust the phase and amplitude of the equivalent coupling coefficient, causing destructive interference between direct and indirect coupling, thereby physically blocking the propagation path of crosstalk.

[0104] The relay resonator adopts a structure of a planar spiral inductor and a varactor diode connected in parallel, such as... Figure 5 As shown. Its resonant frequency is:

[0105] ,

[0106] in, It is a planar spiral inductor. This is the total capacitance. , It is a varactor diode capacitor. This is the parasitic capacitance. The capacitance-voltage characteristic of a varactor diode is approximately:

[0107] ,

[0108] in, Zero-bias junction capacitance For built-in potential, These are the gradient coefficients. This is the input voltage.

[0109] After introducing the relay resonator, the equivalent coupling coefficient between the main signal channels A and B is: :

[0110] ,

[0111] in, Let be the direct coupling coefficient between channels A and B. , These are the coupling coefficients between signal channels A and B and the tunable repeater resonator, respectively. The detuned frequency, This refers to the relay loss rate. It is the imaginary unit.

[0112] By adjusting the bias voltage The capacitance of the varactor diode can be changed. This changes the total capacitance. and resonant frequency Due to detuning frequency , Since the center frequencies of signal channels A and B are given, the equivalent coupling coefficient can be adjusted by changing the bias voltage Vbias. When minimized, the energy coupled from signal channel A to signal channel B approaches zero, and crosstalk is significantly suppressed.

[0113] To verify the effectiveness of the method and system of this invention, in the same scenario, this invention was compared with traditional crosstalk suppression schemes such as optical shielding, fixed relays plus digital filters, etc. The results are shown in Table 2:

[0114] Table 2. Experimental Comparison Results of Crosstalk Suppression Effect

[0115]

[0116] As shown in Table 2, the suppression depth of the proposed solution remains above -35dB across the entire temperature range of -20°C to 85°C, which is approximately 5dB better than the fixed relay solution. The standard deviation of the angle measurement error, 0.033°, is superior to the 0.07° of the fixed relay solution, demonstrating the significant advantage of the proposed solution in suppressing encoder crosstalk.

[0117] The above description is only a preferred embodiment of the present invention. It should be noted that those skilled in the art can make several improvements without departing from the principle of the present invention, and these improvements should also be considered within the scope of protection of the present invention.

Claims

1. A dynamic intelligent crosstalk adaptive suppression method for encoders, characterized in that: Includes the following steps: S1: Acquire the multimodal state signals of the encoder and extract features to obtain perceptual features; S2: Input the perceived features into the crosstalk intensity prediction neural network and output the predicted crosstalk intensity values ​​for multiple future time steps; S3: Combine the perceived features with the predicted crosstalk intensity values ​​for multiple future time steps to form a state vector, and use the soft actor-commentator algorithm to output the optimal bias voltage control signal; S4: Adjust the bias voltage of the varactor diode in the tunable relay resonator according to the optimal bias voltage control signal to change the resonant frequency of the tunable relay resonator, so as to minimize the equivalent coupling coefficient between encoder signal channels and suppress encoder crosstalk.

2. The encoder dynamic intelligent crosstalk adaptive suppression method as described in claim 1, characterized in that: In S2, the crosstalk intensity prediction neural network is a frequency-decoupled dual-path parallel network, including a frequency decoupling module, a low-pass filter branch, and a high-pass filter branch. The frequency decoupling module includes a low-pass filter with a cutoff frequency of 10kHz and a high-pass filter with a cutoff frequency of 50kHz. The low-pass filter branch is used to extract and predict features of the crosstalk components after filtering by the low-pass filter, and the high-pass filter branch is used to extract and predict features of the crosstalk components after filtering by the high-pass filter. The prediction results of the low-pass filter branch and the high-pass filter branch are fused to obtain the predicted crosstalk intensity values ​​for the next multiple time steps.

3. The encoder dynamic intelligent crosstalk adaptive suppression method as described in claim 2, characterized in that: The low-pass filter branch includes a first dilated convolutional multi-scale module, a first 1×1 convolutional fusion layer, a unidirectional long short-term memory network, and a first fully connected output layer. The first dilated convolutional multi-scale module is used to extract multi-scale features from the crosstalk components after low-pass filtering. The first 1×1 convolutional fusion layer is used to perform channel fusion on the output of the first dilated convolutional multi-scale module. The unidirectional long short-term memory network is used to capture temporal dependencies. The first fully connected output layer is used to output the prediction results of the low-pass filter branch.

4. The encoder dynamic intelligent crosstalk adaptive suppression method as described in claim 2, characterized in that: The high-pass filter branch includes a second dilated convolutional multi-scale module, a second 1×1 convolutional fusion layer, a temporal convolutional network, and a second fully connected output layer. The second dilated convolutional multi-scale module is used to extract multi-scale features from the crosstalk components after high-pass filtering. The second 1×1 convolutional fusion layer is used to perform channel fusion on the output of the second dilated convolutional multi-scale module. The temporal convolutional network is used to capture transient responses. The second fully connected output layer is used to output the prediction results of the high-pass filter branch.

5. The encoder dynamic intelligent crosstalk adaptive suppression method as described in claim 1, characterized in that: The soft actor-critic algorithm aims to maximize the weighted sum of the expected cumulative reward and the policy entropy. The expected cumulative reward is calculated by a reward function, which is a weighted sum of crosstalk suppression, action smoothness, and power consumption terms. The formula for calculating the reward function is as follows: ， in, For a moment The reward value, For a moment The measured crosstalk intensity, For a moment The optimal bias voltage control signal. For a moment The optimal bias voltage control signal. This is the crosstalk suppression coefficient. The motion smoothness coefficient. This is the power consumption factor.

6. The encoder dynamic intelligent crosstalk adaptive suppression method as described in claim 1, characterized in that: The tunable relay resonator adopts a structure of a planar spiral inductor and a varactor diode connected in parallel.

7. The encoder dynamic intelligent crosstalk adaptive suppression method as described in claim 6, characterized in that: The resonant frequency of the tunable relay resonator Determined by the following formula: ， in, It is a planar spiral inductor. This represents the total capacitance.

8. The encoder dynamic intelligent crosstalk adaptive suppression method as described in claim 7, characterized in that: The total capacitance is composed of a varactor diode capacitor and a parasitic capacitance connected in parallel. The varactor diode capacitor is adjusted by the bias voltage, and the capacitance-voltage characteristic of the varactor diode is expressed by the following formula: ， in, It is a varactor diode capacitor. Zero-bias junction capacitance Input voltage, For built-in potential, These are the gradient coefficients.

9. The encoder dynamic intelligent crosstalk adaptive suppression method as described in claim 1, characterized in that: The equivalent coupling coefficient It can be expressed by the following formula: ， in, For direct coupling coefficients, Let be the coupling coefficient between signal channel A and the tunable repeater resonator. This represents the coupling coefficient between the B signal channel and the tunable repeater resonator. The detuned frequency, This refers to the relay loss rate. It is the imaginary unit.

10. The encoder dynamic intelligent crosstalk adaptive suppression method as described in claim 1, characterized in that: The multimodal state signal includes electrical characteristics, environmental parameters, and mechanical parameters. The electrical characteristics include the time-domain waveforms of each signal channel of the encoder and the power rail noise. The environmental parameters include temperature, and the mechanical parameters include instantaneous rotational speed.

11. The encoder dynamic intelligent crosstalk adaptive suppression method as described in claim 10, characterized in that: The feature extraction includes extracting the root mean square value and the crosstalk band energy ratio. The root mean square value is calculated based on the time-domain waveform, and the crosstalk band energy ratio is calculated based on the power spectral density of each signal channel of the encoder in the crosstalk-sensitive frequency band.

Images