Full-angle mode starting optimization control method for hemispherical resonator gyroscope

Abstract

The invention relates to the technical field of inertial measurement, in particular to a hemispherical resonator gyroscope full-angle mode start optimization control method which comprises the following steps: S1, data collection and preprocessing, S2, deep learning model establishment, S3, random drift compensation, S4, delay compensation, S5, vibration asymmetry test and trimming, S6, integration and test, and S7, continuous optimization. The method comprises the steps of comprehensive data collection, deep learning model establishment, random drift compensation, delay compensation, vibration asymmetry test and trimming, integration and test, continuous optimization and the like. The technical problems that when the hemispherical resonator gyroscope is started in a full-angle mode, dynamic response prediction is inaccurate, random drift compensation is insufficient, the delay influence of a hemispherical resonator gyroscope control system is large, vibration is asymmetric, and the stability and accuracy of system performance are insufficient are solved. The integrated application of the series of steps improves the overall performance and reliability of the system.

Description

Technical Field

[0001] The invention relates to the technical field of inertial measurement, and in particular to a full-angle mode startup optimization control method for a hemispherical resonant gyroscope. Background Art

[0002] The hemispherical resonant gyroscope is a high-performance solid-state vibration gyroscope with the characteristics of long life, high reliability and high precision. In full-angle mode, it uses the vibration characteristics of the resonator to detect and measure angular velocity and angle changes. The hemispherical resonant gyroscope in full-angle mode can achieve full-angle sensitive output from 0° to 360°, which means that it can measure changes at any angle without being limited by the measurement range of traditional gyroscopes. Compared with the traditional force balance mode, the dynamic characteristics of the hemispherical resonant gyroscope in full-angle mode are greatly improved, and the angular rate integration function can be realized. The measured angular velocity measurement value exceeds 300° / s, and the angle measurement linearity is better than 10 -4 .

[0003] However, when the hemispherical resonant gyro is started in full-angle mode, due to the damping unevenness in full-angle mode, the random drift in the output angle needs to be dynamically compensated. In addition, the delay of the hemispherical resonant gyro control system and the gain mismatch at the signal detection end will cause angle gain fluctuations, affecting the accuracy of angle solution.

[0004] In the prior art, PI control is usually used to eliminate dynamic errors caused by uneven stiffness and asymmetric damping. However, as the size of the hemispherical resonator gradually decreases, the frequency mismatch problem increases. Moreover, as the size decreases, the noise increases, and the sensitivity (proportional factor) and driving force decrease. Summary of the invention

[0005] The present invention aims to provide a method for optimizing the control of starting a full-angle mode of a hemispherical resonant gyroscope to solve the above-mentioned problem.

[0006] The present invention is achieved through the following technical solutions: A method for optimizing the start-up control of a full-angle mode of a hemispherical resonant gyroscope comprises the following steps: S1: Data collection and preprocessing, collecting high-quality data of the hemispherical resonant gyroscope during the full-angle mode startup process, including frequency, amplitude, temperature, etc., and performing denoising and normalization processing; S2: Deep learning model establishment: Establish an LSTM network model that can accurately predict and compensate for the dynamic response of the gyroscope, reduce the risk of overfitting through dropout technology, and use the Adam optimizer and mean square error (MSE) loss function for efficient and accurate training; S3: Random drift compensation. An autoregressive moving average model ARMA based on time variation is established to describe the random drift of the gyroscope. The parameters are optimized through analytical methods to ensure that the model fitting accuracy reaches more than 99%. Then, the autoregressive moving average model ARMA based on time variation is applied to the Kalman filter for random drift compensation.

[0007] S4: Delay compensation, the Smith predictor compensates the control delay and sensor delay of the hemispherical resonant gyro control system, and reduces the impact of delay on system performance by adjusting the transfer functions of the forward channel and feedback channel; S5: Vibration asymmetry test and adjustment: Design a vibration asymmetry error test and adjustment system, identify vibration asymmetry errors through frequency sweep and multi-position free vibration tests, and achieve high-precision adjustment to adjust the frequency difference to below 0.2mHz; S6: Integration and testing: Integrate the LSTM network model, autoregressive moving average model ARMA, Smith predictor and frequency error adjustment system into a unified control platform, conduct tests to verify various performance indicators, obtain test results and test data, and adjust and optimize the control strategy based on the test results; S7: Continuous optimization, continuously collect operating data for further training and optimization of the LSTM network model and the autoregressive moving average model ARMA, and continuously adjust and optimize the parameters of the LSTM network model and the autoregressive moving average model ARMA based on new data and feedback.

[0008] The present invention solves the technical problems of inaccurate dynamic response prediction, insufficient random drift compensation, large delay effect of hemispherical resonant gyro control system, vibration asymmetry problem, and insufficient system performance stability and accuracy when the hemispherical resonant gyro is started in full-angle mode through comprehensive data collection, deep learning model establishment, random drift compensation, delay compensation, vibration asymmetry test and adjustment, integration and testing, and continuous optimization. The integrated application of this series of steps improves the overall performance and reliability of the system.

[0009] Preferably, in S1, the following steps are specifically included: S1.1: Collect data of the hemispherical resonant gyroscope during the full-angle mode startup process, including frequency, amplitude, temperature, etc.; S1.2: Preprocess the collected data, including denoising and normalization, to improve data quality. Specifically, use wavelet transform and Fourier transform to denoise the raw data to ensure that the signal-to-noise ratio of the data is above 40 dB. Normalization maps the data to the range of [0, 1] to improve data processability and model training effect. S1.3: Introduce multimodal data fusion technology to fuse multimodal data such as frequency, amplitude, and temperature to improve data integrity and accuracy. Specifically, use a combination of principal component analysis (PCA) and independent component analysis (ICA) to extract features and reduce dimensionality of multimodal data to ensure that high-dimensional features of the data are effectively extracted and utilized.

[0010] The present invention solves the technical problems of poor data quality, low signal-to-noise ratio, and high data unprocessability by collecting key data of the hemispherical resonant gyroscope during the full-angle mode startup process, and performing denoising and normalization processing. In particular, the denoising process is performed using wavelet transform and Fourier transform to ensure that the signal-to-noise ratio of the data reaches more than 40dB, and the normalization process maps the data to the range of [0, 1], which improves the processability of the data and the training effect of the model, and provides a high-quality data foundation for subsequent model training and control strategy optimization.

[0011] Preferably, in S2, the following steps are specifically included: S2.1: Select LSTM network as the deep learning model for modeling and estimating gyroscope data. LSTM network can effectively capture the long dependencies of time series data through its gating mechanism, which is suitable for data modeling of hemispherical resonant gyroscope.

[0012] S2.2: Introduce an adaptive learning rate mechanism in the LSTM network to improve the training efficiency and accuracy of the model by dynamically adjusting the learning rate. The adaptive learning rate mechanism automatically adjusts the learning rate according to the changes in the loss function during the training process. The initial learning rate is 0.001 and the maximum learning rate is 0.01, ensuring that the model can maintain the best training effect at different training stages.

[0013] S2.3: Introduce dropout technology into the model to reduce the risk of overfitting and reduce computational complexity. Specifically, the dropout rate is set to 0.2, that is, 20% of neurons are randomly discarded during training to enhance the generalization ability of the model.

[0014] S2.4: Use the preprocessed data to train the deep learning model so that it can accurately predict and compensate for the dynamic response of the gyroscope. The training process uses the Adam optimizer, the learning rate is set to 0.001, and the loss function uses the mean square error (MSE) to ensure the efficiency and accuracy of the model training.

[0015] The present invention solves the technical problems of insufficient capture of long-term dependencies, low model training efficiency and accuracy, and high risk of overfitting in hemispherical resonant gyroscope data modeling by selecting LSTM network as the deep learning model and introducing adaptive learning rate mechanism and dropout technology. The adaptive learning rate mechanism automatically adjusts the learning rate according to the change of loss function during training, and the dropout technology reduces the risk of overfitting and the computational complexity, thereby improving the model's ability to predict and compensate for dynamic responses.

[0016] Preferably, in S3, the following steps are specifically included: S3.1: A time-varying autoregressive moving average (ARMA) model is established to describe the random drift of the gyroscope. The ARMA model captures the autocorrelation and random noise of the data through the autoregressive and moving average parts, respectively, to accurately model the random drift. Specifically, an online parameter estimation method is used to update the parameters of the ARMA model every 100 sampling points to ensure that the model can maintain high-precision random drift compensation under different dynamic conditions.

[0017] S3.2: Optimize the parameters through analytical methods and establish the optimal random drift model. Specifically, use the optimization method combining the least squares method and the genetic algorithm to optimize the parameters of the autoregressive moving average model ARMA to ensure that the model fitting accuracy reaches more than 99%.

[0018] The optimization method combining the least square method and the genetic algorithm specifically includes the following steps: S3.2.1: Randomly generate a set of initial parameter vectors ; S3.2.2: For each parameter vector , using the least squares method to calculate the residual sum of squares As fitness; S3.2.3: Based on fitness Select parameter vectors for reproduction. The smaller the fitness, the higher the probability of being selected. S3.2.4: cross-combine the selected parameter vectors to generate new parameter vectors and perform random mutation on the newly generated parameter vectors; S3.2.5: For each newly generated parameter vector , use the least squares method for local optimization to further reduce the residual sum of squares ; S3.2.6: Replace the old parameter vector with the new parameter vector to form a new population; S3.2.7: Repeat steps S3.2.2-S.2.6 until the maximum number of iterations or the improvement in the residual sum of squares is less than a preset threshold; S3.3: Apply the time-varying autoregressive moving average model ARMA to the Kalman filter for random drift compensation. The Kalman filter updates the random drift of the gyroscope in real time through state estimation and prediction to ensure the accuracy of the output angle.

[0019] The present invention describes the random drift of the gyroscope by establishing an autoregressive moving average model ARMA based on time variation, and uses a Kalman filter to compensate for the random drift, thereby solving the technical problem that the random drift of the gyroscope is difficult to accurately model and compensate. The combined use of the online parameter estimation method and the analytical method ensures that the model can maintain high-precision random drift compensation under different dynamic conditions, thereby improving the stability and accuracy of the system.

[0020] Preferably, S3 also includes S3.4: using an anti-interference delay compensation method, through a dual closed-loop control strategy, to improve the anti-interference ability and robustness of the system, wherein the dual closed-loop control strategy includes an inner loop and an outer loop, the inner loop is used for fast response and delay compensation, and the outer loop is used for smooth control and anti-interference. Specifically, the inner loop adopts a Smith predictor, and the outer loop adopts a PID controller. Through the dual closed-loop control strategy, accurate compensation for control delay and anti-interference control are achieved, and its mathematical expression is as follows: ; ; ; in, is the inner loop transfer function, represents the gyro response time, that is, the time constant from when the gyro receives the control signal to when it generates a response; is the outer loop transfer function, Represents the external environment adaptation time of the gyroscope, that is, the time constant of the gyroscope's adaptation to external environment changes, including temperature, vibration, etc.; is the PID controller transfer function, is the proportional gain, which determines how quickly the gyro responds to deviations. is the integration time constant, which affects the cumulative response of the gyro to long-term deviations, is the differential time constant, which determines the gyro's ability to predict and respond to rapid changes; is the complex frequency variable in Laplace transform, which describes the relationship between the system output and input; The present invention improves the anti-interference ability and robustness of the system by introducing an anti-interference double closed-loop control strategy.

[0021] The inner loop adopts Smith predictor for fast response and delay compensation, and the outer loop adopts PID controller for smooth control and anti-interference, which realizes accurate compensation for control delay and anti-interference control, and solves the problem of performance degradation of hemispherical resonant gyro control system when facing external interference.

[0022] Preferably, S4 specifically includes the following steps: S4.1: Smith predictor is introduced to compensate for the control delay and sensor delay of the HRG control system. Smith predictor combines the delay model of the forward channel with the Kalman filter of the feedback channel to effectively reduce the delay effect of the HRG control system.

[0023] S4.2: Adjust the closed-loop transfer function through the Smith predictor to reduce the effect of delay on the hemispherical resonant gyro control system. Specifically, adjust the transfer functions of the forward channel and the feedback channel of the Smith predictor so that the system can maintain stable performance in the presence of delay.

[0024] The closed-loop transfer function of the Smith predictor is:

[0025]

[0026] ; in, is the delay model, L is the propagation delay of the control signal of the HRG control system and the sensor response delay; S is a complex frequency variable; To describe the dynamic response of the HRG control system to the control input; It is a PID controller that adjusts the response of the hemispherical resonant gyro control system; is the forward channel transfer function; is the feedback channel transfer function; is the closed-loop transfer function; The invention effectively reduces the delay effect of the hemispherical resonant gyro control system by introducing the Smith predictor, solves the technical problems of large influence of system control delay and sensor delay, insufficient real-time performance and stability, and improves the response speed and overall performance of the hemispherical resonant gyro control system.

[0027] Preferably, S5 specifically includes the following steps: S5.1: Design a vibration asymmetry error test and adjustment system to optimize the frequency difference of the hemispherical resonator. The system identifies vibration asymmetry errors through frequency sweep and multi-position free vibration testing.

[0028] S5.2: Implement functions such as frequency sweep, multi-position free vibration and parameter identification to achieve high-precision adjustment. Specifically, the frequency sweep method is used to gradually scan in the range of 50Hz to 5000Hz, the multi-position free vibration test is performed at 0°, 90°, 180° and 270°, and the parameter identification determines the parameters of the asymmetric error by the least squares method.

[0029] S5.3: Adjust the frequency difference to below 0.2mHz to solve the frequency mismatch problem during the full-width mode startup process. Specifically, the frequency difference adjustment process is divided into three stages: coarse adjustment, fine adjustment and fine adjustment. Each stage adopts a different adjustment strategy. The coarse adjustment stage uses a large step adjustment, the fine adjustment stage uses a medium step adjustment, and the fine adjustment stage uses a small step adjustment to ensure that the frequency difference is finally reduced to below 0.2mHz, so that the frequency difference is reduced to below 0.2mHz, ensuring the stability of the full-width mode startup.

[0030] The present invention solves the technical problems of insufficient optimization of the frequency difference of the hemispherical resonant gyroscope and frequency mismatch during the full-angle mode startup by designing a vibration asymmetry error test and adjustment system. Vibration asymmetry errors are identified by frequency sweep and multi-position free vibration tests, and high-precision adjustment is achieved to adjust the frequency difference to below 0.2mHz, thereby improving the stability of the full-angle mode startup.

[0031] Preferably, S6 specifically includes the following steps: S6.1: Integrate the above solution into the hemispherical resonant gyro control system of the hemispherical resonant gyro. Specifically, integrate the LSTM network model, the autoregressive moving average model ARMA, the Smith predictor and the frequency error adjustment system into a unified control platform to achieve the coordinated work of the system.

[0032] S6.2: Conduct tests to verify the effects of the deep learning model, random drift compensation, and delay compensation. Specifically, conduct multiple full-width mode startup tests to verify various performance indicators, and obtain test results and test data to ensure the stability and accuracy of the system in actual applications.

[0033] S6.3: Adjust and optimize the control strategy according to the test results to ensure the stability and accuracy of full-width mode startup. Specifically, adjust the parameters of the LSTM network model, the parameters of the autoregressive moving average model ARMA, and the parameters of the Smith predictor according to the test data to optimize the control strategy.

[0034] The present invention integrates the LSTM network model, the autoregressive moving average model ARMA, the Smith predictor and the frequency difference adjustment system into a unified control platform, and conducts tests to verify various performance indicators, and obtains test results and test data, thereby solving the technical problems of insufficient coordination between independent models and adjustment systems, insufficient verification of performance indicators, and untimely optimization of control strategies, thereby ensuring the stability and accuracy of the system in practical applications.

[0035] Preferably, in S6.3, adjusting and optimizing the control strategy according to the test results specifically uses an adaptive control optimization algorithm to optimize the control parameters of the hemispherical resonant gyro control system in real time to improve the control effect and the robustness of the system, wherein the adaptive control optimization algorithm uses a reinforcement learning algorithm (RL).

[0036] ; Among them, S is the various operating parameters of the hemispherical resonant gyroscope, including amplitude A ,frequency f ,temperature T ; ; in, Refers to all the actions that a HRG control system can perform, including adjusting amplitude, changing frequency, and adjusting temperature; ; in, is the reward function, including response time, stability and accuracy, S represents the current working state of the hemispherical resonant gyro control system, For action space One of the actions in , and are the weights corresponding to response time, stability, and accuracy; is the response time, is stability, is accuracy; ; in, is the action value function, which means that In the case of , start from state S and take action Expected return; is the expected return, t is the time step, is the reward function at time step t, S is the current state, is the initial state, is the initial action; Indicates the selection of action in state S The probability of is the discount factor.

[0037] The present invention uses a reinforcement learning algorithm to perform adaptive control optimization, optimizes the control parameters of a hemispherical resonant gyro control system in real time, and improves the control effect and the robustness of the system. The reinforcement learning algorithm learns the optimal control strategy through interaction with the environment, solves the technical problems of insufficient real-time optimization of control parameters and poor adaptability of the system to changes and new challenges, and improves the adaptability and optimization ability of the hemispherical resonant gyro control system.

[0038] Preferably, S7 specifically includes the following steps: S7.1: Continuously collect operating data for further training and optimization of the LSTM network model and the autoregressive moving average model ARMA. Specifically, through the real-time data acquisition system, continuously collect data of the hemispherical resonant gyroscope during the full-angle mode startup process for further training and optimization of the LSTM network model and the autoregressive moving average model ARMA.

[0039] S7.2: Based on new data and feedback, continuously adjust and optimize the parameters of the LSTM network model and the autoregressive moving average model ARMA to adapt to possible changes and new challenges.

[0040] The present invention solves the technical problems of insufficient efficient operation capability and poor adaptability to changes and new challenges of the system in different environments and conditions by continuously collecting operation data and continuously optimizing the control algorithm. This ensures that the system can continuously adjust and optimize the parameters of the LSTM network model and the autoregressive moving average model ARMA according to new data and feedback, ensuring efficient operation and performance optimization of the system in different environments and conditions.

[0041] Compared with the prior art, the present invention has the following advantages and beneficial effects: The present invention provides a full-angle mode startup optimization control method for a hemispherical resonant gyroscope. By comprehensively using data collection and preprocessing, deep learning model establishment, random drift compensation, delay compensation, vibration asymmetry test and adjustment, integration and testing, and continuous optimization, the present invention effectively solves the technical problems faced by the hemispherical resonant gyroscope when starting in full-angle mode, such as inaccurate dynamic response prediction, insufficient random drift compensation, large delay influence of the hemispherical resonant gyroscope control system, vibration asymmetry problem, and insufficient system performance stability and accuracy. Through high-quality data preprocessing, long dependency capture, adaptive learning rate adjustment, dropout technology application, autoregressive moving average model ARMA combined with Kalman filter, dual closed-loop control strategy, Smith predictor and Q-Learning algorithm collaboration, vibration asymmetry error test and adjustment system design, and adaptive control optimization of reinforcement learning algorithm, the present invention not only improves the overall performance and reliability of the system, but also enhances the anti-interference ability and robustness of the system, ensures the efficient operation and performance optimization of the system under different environments and conditions, realizes accurate compensation and anti-interference control of control delay, and improves the stability and accuracy of full-angle mode startup. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The drawings described herein are used to provide a further understanding of the embodiments of the present invention, constitute a part of this application, and do not constitute a limitation of the embodiments of the present invention. In the drawings: Figure 1 It is a flow chart of the steps of the present invention. DETAILED DESCRIPTION

[0043] In order to make the purpose, technical scheme and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments and drawings. The schematic implementation modes and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention. It should be noted that the present invention is already in the actual development and use stage.

[0044] Embodiment: like Figure 1 As shown, this embodiment includes a hemispherical resonant gyroscope full-angle mode startup optimization control method, including the following steps: In this embodiment, the hemispherical resonator gyroscope is an inertial sensor that uses the vibration of a hemispherical resonator to sense angular velocity. It determines the rotation speed and direction of an object by detecting and measuring the changes in the vibration mode inside the hemispherical resonator.

[0045] Full-angle mode refers to the mode in which the gyroscope can measure 360-degree angular velocity, which can provide more comprehensive angular velocity information compared to the traditional half-angle or quarter-angle mode.

[0046] S1: Data collection and preprocessing, collecting high-quality data of the hemispherical resonant gyroscope during the full-angle mode startup process, including frequency, amplitude, temperature, etc., and performing denoising and normalization processing; S2: Deep learning model establishment: Establish an LSTM network model that can accurately predict and compensate for the dynamic response of the gyroscope, reduce the risk of overfitting through dropout technology, and use the Adam optimizer and mean square error (MSE) loss function for efficient and accurate training; In the present invention, the long dependency of time series data is captured through the LSTM network model, thereby improving the prediction accuracy of the dynamic response of the gyroscope; The dropout technique randomly discards some neurons during the training process to reduce the overfitting of the model and enhance the generalization ability of the model. S3: Random drift compensation. An autoregressive moving average model ARMA based on time variation is established to describe the random drift of the gyroscope. The parameters are optimized through analytical methods to ensure that the model fitting accuracy reaches more than 99%. Then, the autoregressive moving average model ARMA based on time variation is applied to the Kalman filter for random drift compensation.

[0047] In the present invention, the autoregressive moving average model ARMA describes the random drift of the gyroscope, and captures the autocorrelation and random noise of the data through the autoregressive and moving average parts.

[0048] S4: Delay compensation, the Smith predictor compensates the control delay and sensor delay of the hemispherical resonant gyro control system, and reduces the impact of delay on system performance by adjusting the transfer functions of the forward channel and feedback channel; S5: Vibration asymmetry test and adjustment: Design a vibration asymmetry error test and adjustment system, identify vibration asymmetry errors through frequency sweep and multi-position free vibration tests, and achieve high-precision adjustment to adjust the frequency difference to below 0.2mHz; S6: Integration and testing: Integrate the LSTM network model, autoregressive moving average model ARMA, Smith predictor and frequency error adjustment system into a unified control platform, conduct tests to verify various performance indicators, obtain test results and test data, and adjust and optimize the control strategy based on the test results; S7: Continuous optimization, continuously collect operating data for further training and optimization of the LSTM network model and the autoregressive moving average model ARMA, and continuously adjust and optimize the parameters of the LSTM network model and the autoregressive moving average model ARMA based on new data and feedback.

[0049] The present invention solves the technical problems of inaccurate dynamic response prediction, insufficient random drift compensation, large delay effect of hemispherical resonant gyro control system, vibration asymmetry problem, and insufficient system performance stability and accuracy when the hemispherical resonant gyro is started in full-angle mode through comprehensive data collection, deep learning model establishment, random drift compensation, delay compensation, vibration asymmetry test and adjustment, integration and testing, and continuous optimization. The integrated application of this series of steps improves the overall performance and reliability of the system.

[0050] In this embodiment, in S1, the following steps are specifically included: S1.1: Collect data of the hemispherical resonant gyroscope during the full-angle mode startup process, including frequency, amplitude, temperature, etc.; S1.2: Preprocess the collected data, including denoising and normalization, to improve data quality. Specifically, use wavelet transform and Fourier transform to denoise the raw data to ensure that the signal-to-noise ratio of the data is above 40dB, which means that after denoising, the effective signal strength in the data is at least 100 times the noise strength. Normalization maps the data to the [0, 1] range to improve the processability of the data and the training effect of the model; this means that all data will be scaled to values ​​between 0 and 1 for easy processing and comparison.

[0051] S1.3: Introduce multimodal data fusion technology to fuse multimodal data such as frequency, amplitude, and temperature to improve data integrity and accuracy. Specifically, use a combination of principal component analysis (PCA) and independent component analysis (ICA) to extract features and reduce dimensionality of multimodal data to ensure that high-dimensional features of the data are effectively extracted and utilized.

[0052] PCA and ICA in this embodiment are two multivariate statistical analysis techniques used for data dimensionality reduction and feature extraction. PCA transforms the data into a new coordinate system through an orthogonal transformation so that the first largest variance of any projection of the data is on the first coordinate (called the first principal component), while ICA aims to separate multivariate signals into statistically independent non-Gaussian signal sources.

[0053] The present invention solves the technical problems of poor data quality, low signal-to-noise ratio, and high data unprocessability by collecting key data of the hemispherical resonant gyroscope during the full-angle mode startup process, and performing denoising and normalization processing. In particular, the denoising process is performed using wavelet transform and Fourier transform to ensure that the signal-to-noise ratio of the data reaches more than 40dB, and the normalization process maps the data to the range of [0, 1], which improves the processability of the data and the training effect of the model, and provides a high-quality data foundation for subsequent model training and control strategy optimization.

[0054] In this embodiment, in S2, the following steps are specifically included: S2.1: Select LSTM network as the deep learning model for modeling and estimating gyroscope data. LSTM network can effectively capture the long dependencies of time series data through its gating mechanism, which is suitable for data modeling of hemispherical resonant gyroscope.

[0055] S2.2: Introduce an adaptive learning rate mechanism in the LSTM network to improve the training efficiency and accuracy of the model by dynamically adjusting the learning rate. The adaptive learning rate mechanism automatically adjusts the learning rate according to the changes in the loss function during the training process. The initial learning rate is 0.001 and the maximum learning rate is 0.01, ensuring that the model can maintain the best training effect at different training stages.

[0056] S2.3: Introduce dropout technology into the model to reduce the risk of overfitting and reduce computational complexity. Specifically, the dropout rate is set to 0.2, that is, 20% of neurons are randomly discarded during training to enhance the generalization ability of the model.

[0057] The dropout in this embodiment is a regularization technique used to prevent overfitting of the neural network. During the training process, it randomly "drops" (i.e. temporarily removes) some neurons in the network, forcing the network to learn more robust features.

[0058] S2.4: Use the preprocessed data to train the deep learning model so that it can accurately predict and compensate for the dynamic response of the gyroscope. The training process uses the Adam optimizer, the learning rate is set to 0.001, and the loss function uses the mean square error (MSE) to ensure the efficiency and accuracy of the model training.

[0059] In this embodiment, the autoregressive moving average model ARMA is a statistical model used to analyze time series data, which combines autoregressive (AR) and moving average (MA) parts to describe the autocorrelation and random noise of data.

[0060] The present invention solves the technical problems of insufficient capture of long-term dependencies, low model training efficiency and accuracy, and high risk of overfitting in hemispherical resonant gyroscope data modeling by selecting LSTM network as the deep learning model and introducing adaptive learning rate mechanism and dropout technology. The adaptive learning rate mechanism automatically adjusts the learning rate according to the change of loss function during training, and the dropout technology reduces the risk of overfitting and the computational complexity, thereby improving the model's ability to predict and compensate for dynamic responses.

[0061] In this embodiment, in S3, the following steps are specifically included: S3.1: A time-varying autoregressive moving average (ARMA) model is established to describe the random drift of the gyroscope. The ARMA model captures the autocorrelation and random noise of the data through the autoregressive and moving average parts, respectively, to accurately model the random drift. Specifically, the online parameter estimation method is used to update the parameters of the ARMA model every 100 sampling points. This means that the parameters of the ARMA model will be updated in real time according to the latest data to adapt to possible changes, ensuring that the model can maintain high-precision random drift compensation under different dynamic conditions.

[0062] S3.2: Optimize the parameters through analytical methods and establish the optimal random drift model. Specifically, use the optimization method combining the least squares method and the genetic algorithm to optimize the parameters of the autoregressive moving average model ARMA to ensure that the model fitting accuracy reaches more than 99%.

[0063] S3.2.1: Randomly generate a set of initial parameter vectors ; S3.2.2: For each parameter vector , using the least squares method to calculate the residual sum of squares As fitness; S3.2.3: Based on fitness Select parameter vectors for reproduction. The smaller the fitness, the higher the probability of being selected. S3.2.4: cross-combine the selected parameter vectors to generate new parameter vectors and perform random mutation on the newly generated parameter vectors; S3.2.5: For each newly generated parameter vector , use the least squares method for local optimization to further reduce the residual sum of squares ; S3.2.6: Replace the old parameter vector with the new parameter vector to form a new population; S3.2.7: Repeat steps S3.2.2-S.2.6 until the maximum number of iterations or the improvement in the residual sum of squares is less than a preset threshold; S3.3: Apply the time-varying autoregressive moving average model ARMA to the Kalman filter for random drift compensation. The Kalman filter updates the random drift of the gyroscope in real time through state estimation and prediction to ensure the accuracy of the output angle.

[0064] The Kalman filter in this embodiment is an efficient recursive algorithm for estimating the state of a dynamic system from a series of measurements containing noise. It is applied to the time-varying autoregressive moving average model ARMA to perform random drift compensation and update the random drift of the gyroscope in real time to ensure the accuracy of the output angle.

[0065] The present invention describes the random drift of the gyroscope by establishing an autoregressive moving average model ARMA based on time variation, and uses a Kalman filter to compensate for the random drift, thereby solving the technical problem that the random drift of the gyroscope is difficult to accurately model and compensate. The combined use of the online parameter estimation method and the analytical method ensures that the model can maintain high-precision random drift compensation under different dynamic conditions, thereby improving the stability and accuracy of the system.

[0066] In this embodiment, S3 also includes S3.4: using an anti-interference delay compensation method, through a dual closed-loop control strategy, to improve the anti-interference ability and robustness of the system, wherein the dual closed-loop control strategy includes an inner loop and an outer loop, the inner loop is used for fast response and delay compensation, and the outer loop is used for smooth control and anti-interference. Specifically, the inner loop adopts a Smith predictor, and the outer loop adopts a PID controller. Through the dual closed-loop control strategy, accurate compensation for control delay and anti-interference control are achieved, and its mathematical expression is as follows: ; ; ; in, is the inner loop transfer function, represents the gyro response time, that is, the time constant from when the gyro receives the control signal to when it generates a response; is the outer loop transfer function, Represents the external environment adaptation time of the gyroscope, that is, the time constant of the gyroscope's adaptation to external environment changes, including temperature, vibration, etc.; is the PID controller transfer function, is the proportional gain, which determines how quickly the gyro responds to deviations. is the integration time constant, which affects the cumulative response of the gyro to long-term deviations, is the differential time constant, which determines the gyro's ability to predict and respond to rapid changes; is the complex frequency variable in Laplace transform, which describes the relationship between the system output and input; The present invention improves the anti-interference ability and robustness of the system by introducing an anti-interference double closed-loop control strategy.

[0067] The inner loop adopts Smith predictor for fast response and delay compensation, and the outer loop adopts PID controller for smooth control and anti-interference, which realizes accurate compensation for control delay and anti-interference control, and solves the problem of performance degradation of hemispherical resonant gyro control system when facing external interference.

[0068] In this embodiment, S4 specifically includes the following steps: S4.1: Smith predictor is introduced to compensate for the control delay and sensor delay of the HRG control system. Smith predictor combines the delay model of the forward channel with the Kalman filter of the feedback channel to effectively reduce the delay effect of the HRG control system.

[0069] S4.2: Adjust the closed-loop transfer function through the Smith predictor to reduce the effect of delay on the hemispherical resonant gyro control system. Specifically, adjust the transfer functions of the forward channel and the feedback channel of the Smith predictor so that the system can maintain stable performance in the presence of delay.

[0070] The closed-loop transfer function of the Smith predictor is:

[0071]

[0072] ; in, is the delay model, L is the propagation delay of the control signal of the HRG control system and the sensor response delay; S is a complex frequency variable; To describe the dynamic response of the HRG control system to the control input; It is a PID controller that adjusts the response of the hemispherical resonant gyro control system; is the forward channel transfer function; is the feedback channel transfer function; is the closed-loop transfer function; The Smith predictor in this embodiment is a technology used to compensate for the influence of time delay in a hemispherical resonant gyro control system, which predicts and compensates for future outputs by estimating the dynamic behavior of the system, thereby reducing the influence of time delay on system performance; Q-Learning is a model-free reinforcement learning algorithm that learns a policy through interaction with the environment that maximizes the cumulative reward. It does not require prior knowledge about the environment and can learn directly from experience.

[0073] The invention effectively reduces the delay effect of the hemispherical resonant gyro control system by introducing the Smith predictor, solves the technical problems of large influence of system control delay and sensor delay, insufficient real-time performance and stability, and improves the response speed and overall performance of the hemispherical resonant gyro control system.

[0074] In this embodiment, S5 specifically includes the following steps: S5.1: Design a vibration asymmetry error test and adjustment system to optimize the frequency difference of the hemispherical resonator. The system identifies vibration asymmetry errors through frequency sweep and multi-position free vibration testing.

[0075] S5.2: Implement functions such as frequency sweep, multi-position free vibration and parameter identification to achieve high-precision adjustment. Specifically, the frequency sweep method is used to gradually scan in the range of 50Hz to 5000Hz, the multi-position free vibration test is performed at 0°, 90°, 180° and 270°, and the parameter identification determines the parameters of the asymmetric error by the least squares method.

[0076] S5.3: Adjust the frequency difference to below 0.2mHz to solve the frequency mismatch problem during the full-width mode startup process. Specifically, the frequency difference adjustment process is divided into three stages: coarse adjustment, fine adjustment and fine adjustment. Each stage adopts a different adjustment strategy. The coarse adjustment stage uses a large step adjustment, the fine adjustment stage uses a medium step adjustment, and the fine adjustment stage uses a small step adjustment to ensure that the frequency difference is finally reduced to below 0.2mHz, so that the frequency difference is reduced to below 0.2mHz, ensuring the stability of the full-width mode startup.

[0077] The present invention solves the technical problems of insufficient optimization of the frequency difference of the hemispherical resonant gyroscope and frequency mismatch during the full-angle mode startup by designing a vibration asymmetry error test and adjustment system. Vibration asymmetry errors are identified by frequency sweep and multi-position free vibration tests, and high-precision adjustment is achieved to adjust the frequency difference to below 0.2mHz, thereby improving the stability of the full-angle mode startup.

[0078] In this embodiment, S6 specifically includes the following steps: S6.1: Integrate the above solution into the hemispherical resonant gyro control system of the hemispherical resonant gyro. Specifically, integrate the LSTM network model, the autoregressive moving average model ARMA, the Smith predictor and the frequency error adjustment system into a unified control platform to achieve the coordinated work of the system.

[0079] S6.2: Conduct tests to verify the effects of the deep learning model, random drift compensation, and delay compensation. Specifically, conduct multiple full-width mode startup tests to verify various performance indicators, and obtain test results and test data to ensure the stability and accuracy of the system in actual applications.

[0080] S6.3: Adjust and optimize the control strategy according to the test results to ensure the stability and accuracy of full-width mode startup. Specifically, adjust the parameters of the LSTM network model, the parameters of the autoregressive moving average model ARMA, and the parameters of the Smith predictor according to the test data to optimize the control strategy.

[0081] The present invention integrates the LSTM network model, the autoregressive moving average model ARMA, the Smith predictor and the frequency difference adjustment system into a unified control platform, and conducts tests to verify various performance indicators, and obtains test results and test data, thereby solving the technical problems of insufficient coordination between independent models and adjustment systems, insufficient verification of performance indicators, and untimely optimization of control strategies, thereby ensuring the stability and accuracy of the system in practical applications.

[0082] In this embodiment, in S6.3, adjusting and optimizing the control strategy according to the test results specifically uses an adaptive control optimization algorithm to optimize the control parameters of the hemispherical resonant gyro control system in real time to improve the control effect and the robustness of the system, wherein the adaptive control optimization algorithm uses a reinforcement learning algorithm (RL).

[0083] ; Among them, S is the various operating parameters of the hemispherical resonant gyroscope, including amplitude A ,frequency f ,temperature T ; ; in, Refers to all the actions that a HRG control system can perform, including adjusting amplitude, changing frequency, and adjusting temperature; ; in, is the reward function, including response time, stability and accuracy, S represents the current working state of the hemispherical resonant gyro control system, For action space One of the actions in , and are the weights corresponding to response time, stability, and accuracy; is the response time, is stability, is accuracy; ; in, is the action value function, which means that In the case of , start from state S and take action Expected return; is the expected return, t is the time step, is the reward function at time step t, S is the current state, is the initial state, is the initial action; Indicates the selection of action in state S The probability of is the discount factor.

[0084] The present invention uses a reinforcement learning algorithm to perform adaptive control optimization, optimizes the control parameters of a hemispherical resonant gyro control system in real time, and improves the control effect and the robustness of the system. The reinforcement learning algorithm learns the optimal control strategy through interaction with the environment, solves the technical problems of insufficient real-time optimization of control parameters and poor adaptability of the system to changes and new challenges, and improves the adaptability and optimization ability of the hemispherical resonant gyro control system.

[0085] In this embodiment, S7 specifically includes the following steps: S7.1: Continuously collect operating data for further training and optimization of the LSTM network model and the autoregressive moving average model ARMA. Specifically, through the real-time data acquisition system, continuously collect data of the hemispherical resonant gyroscope during the full-angle mode startup process for further training and optimization of the LSTM network model and the autoregressive moving average model ARMA.

[0086] S7.2: Based on new data and feedback, continuously adjust and optimize the parameters of the LSTM network model and the autoregressive moving average model ARMA to adapt to possible changes and new challenges.

[0087] The present invention solves the technical problems of insufficient efficient operation capability and poor adaptability to changes and new challenges of the system in different environments and conditions by continuously collecting operation data and continuously optimizing the control algorithm. This ensures that the system can continuously adjust and optimize the parameters of the LSTM network model and the autoregressive moving average model ARMA according to new data and feedback, ensuring efficient operation and performance optimization of the system in different environments and conditions.

[0088] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for optimizing the start-up control of a full-angle mode of a hemispherical resonant gyroscope, characterized in that: The steps include: S1: Data collection and preprocessing, collecting data of the hemispherical resonant gyroscope during the full-angle mode startup process, including frequency, amplitude, temperature, and performing denoising and normalization processing; S2: deep learning model establishment, establishing an LSTM network model, the LSTM network model is used to predict and compensate for the dynamic response of the gyroscope, the overfitting reduction operation is performed through the dropout technology, and the Adam optimizer and mean square error loss function are used to train the LSTM network model; S3: Random drift compensation, establishing an autoregressive moving average model ARMA based on time variation, the autoregressive moving average model ARMA is used to describe the random drift of the gyroscope, and performing parameter optimization through an analytical method, and then applying the autoregressive moving average model ARMA based on time variation to the Kalman filter for random drift compensation; S4: Delay compensation, which compensates the control delay and sensor delay of the HRG control system through the Smith predictor by adjusting the transfer functions of the forward channel and the feedback channel; S5: Vibration asymmetry test and adjustment, design vibration asymmetry error test and adjustment system, identify vibration asymmetry error through frequency sweep and multi-position free vibration test, and adjust the frequency difference to below 0.2mHz; S6: Integration and testing: Integrate the LSTM network model, autoregressive moving average model ARMA, Smith predictor and frequency error adjustment system into the control platform, conduct tests to verify various performance indicators, obtain test results and test data, and adjust and optimize the control strategy based on the test results; S7: Continuous optimization, continuously collect operating data for further training and optimization of the LSTM network model and the autoregressive moving average model ARMA, and continuously adjust and optimize the parameters of the LSTM network model and the autoregressive moving average model ARMA based on new data and feedback.

2. The method for optimizing the start-up control of a full-angle mode of a hemispherical resonant gyroscope according to claim 1, characterized in that: In S1, the following steps are specifically included: S1.1: Collect data of the HRG during the full-angle mode startup process, including frequency, amplitude, and temperature; S1.2: Preprocess the collected data, including denoising and normalization. Specifically, use wavelet transform and Fourier transform to denoise the raw data until the signal-to-noise ratio of the data reaches above 40 dB, and normalize the data to map it to the range of [0, 1]; S1.3: Introduce multimodal data fusion technology to fuse frequency, amplitude, and temperature multimodal data, and use a combination of principal component analysis and independent component analysis to extract features and reduce the dimension of multimodal data.

3. The method for optimizing the start-up control of a full-angle mode of a hemispherical resonant gyroscope according to claim 1, characterized in that: In S2, the following steps are specifically included: S2.1: Select LSTM network as the deep learning model for modeling and estimating gyroscope data; S2.2: Introduce an adaptive learning rate mechanism in the LSTM network to dynamically adjust the learning rate. The adaptive learning rate mechanism automatically adjusts the learning rate according to the change of the loss function during the training process. The initial learning rate is 0.001 and the maximum learning rate is 0.

01. S2.3: Introduce dropout technology into the model, set the dropout rate to 0.2, and randomly discard 20% of neurons during training; S2.4: Use the preprocessed data to train a deep learning model to predict and compensate for the dynamic response of the gyroscope; the training process uses the Adam optimizer, the learning rate is set to 0.001, and the loss function uses the mean square error.

4. The method for optimizing the start-up control of a full-angle mode of a hemispherical resonant gyroscope according to claim 1, characterized in that: In S3, the specific steps include: S3.1: Establish an autoregressive moving average model ARMA based on time variation to describe the random drift of the gyroscope. The autoregressive moving average model ARMA captures the autocorrelation and random noise of the data through the autoregressive and moving average parts respectively. The parameters of the autoregressive moving average model ARMA are updated every 100 sampling points using an online parameter estimation method. S3.2: Optimize the parameters of the autoregressive moving average model ARMA by using the analytical method and the optimization method combining the least square method and the genetic algorithm; The optimization method combining the least square method and the genetic algorithm specifically includes the following steps: S3.2.1: Randomly generate a set of initial parameter vectors ; S3.2.2: For each parameter vector , using the least squares method to calculate the residual sum of squares As fitness; S3.2.3: Based on fitness Select parameter vectors for reproduction. The smaller the fitness, the higher the probability of being selected. S3.2.4: cross-combine the selected parameter vectors to generate new parameter vectors and perform random mutation on the newly generated parameter vectors; S3.2.5: For each newly generated parameter vector , use the least squares method for local optimization to further reduce the residual sum of squares ; S3.2.6: Replace the old parameter vector with the new parameter vector to form a new population; S3.2.7: Repeat steps S3.2.2-S.2.6 until the maximum number of iterations or the improvement in the residual sum of squares is less than a preset threshold; S3.3: Apply the time-varying autoregressive moving average model ARMA to the Kalman filter for random drift compensation; the Kalman filter updates the random drift of the gyroscope in real time through state estimation and prediction.

5. The method for optimizing the start-up control of a full-angle mode of a hemispherical resonant gyroscope according to claim 4, characterized in that: S3 also includes S3.4: using an anti-interference delay compensation method and a dual closed-loop control strategy, wherein the dual closed-loop control strategy includes an inner loop and an outer loop, the inner loop uses a Smith predictor, and the outer loop uses a PID controller. The dual closed-loop control strategy has the following mathematical expression: ； ； ； in, is the inner loop transfer function, represents the gyro response time, which is the time constant from when the gyro receives the control signal to when it generates a response; is the outer loop transfer function, Represents the external environment adaptation time of the gyroscope, that is, the time constant of the gyroscope's adaptation to external environment changes, including temperature and vibration; is the PID controller transfer function, is the proportional gain, is the integration time constant, is the differential time constant; is the complex frequency variable in the Laplace transform.

6. The method for optimizing the start-up control of a full-angle mode of a hemispherical resonant gyroscope according to claim 1, characterized in that: S4 specifically includes the following steps: S4.1: Introduce the Smith predictor to compensate for the control delay and sensor delay of the HRG control system. The Smith predictor combines the delay model of the forward channel with the Kalman filter of the feedback channel. S4.2: adjusting the closed-loop transfer function by the Smith predictor, and adjusting the transfer functions of the forward channel and the feedback channel of the Smith predictor; The closed-loop transfer function of the Smith predictor is: ； ； ； in, is the delay model, L is the propagation delay of the control signal of the HRG control system and the sensor response delay; S is a complex frequency variable; To describe the dynamic response of the HRG control system to the control input; It is a PID controller that adjusts the response of the hemispherical resonant gyro control system; is the forward channel transfer function; is the feedback channel transfer function; is the closed loop transfer function.

7. The method for optimizing the start-up control of a full-angle mode of a hemispherical resonant gyroscope according to claim 1, characterized in that: S5 specifically includes the following steps: S5.1: Design a vibration asymmetry error test and adjustment system; the system identifies vibration asymmetry errors through frequency sweep and multi-position free vibration tests; S5.2: realize frequency sweep, multi-position free vibration and parameter identification functions, use frequency sweep method to scan step by step in the range of 50Hz to 5000Hz, multi-position free vibration test is carried out at 0°, 90°, 180° and 270° positions, and parameter identification determines the parameters of asymmetric error by least square method; S5.3: Adjust the frequency difference to below 0.2mHz. Divide the frequency difference adjustment process into three stages: coarse adjustment, fine adjustment and fine adjustment. Each stage adopts a different adjustment strategy. The coarse adjustment stage uses a large step adjustment, the fine adjustment stage uses a medium step adjustment, and the fine adjustment stage uses a small step adjustment to reduce the frequency difference to below 0.2mHz.

8. The method for optimizing the start-up control of a full-angle mode of a hemispherical resonant gyroscope according to claim 1, characterized in that: S6 specifically includes the following steps: S6.1: Integrate the LSTM network model, autoregressive moving average model ARMA, Smith predictor and frequency error adjustment system into the control platform; S6.2: Conduct tests to verify the effects of the deep learning model, random drift compensation, and delay compensation. Perform multiple full-width mode startup tests to verify various performance indicators and obtain test results and test data. S6.3: Adjust and optimize the control strategy according to the test results, and adjust the parameters of the LSTM network model, the parameters of the autoregressive moving average model ARMA, and the parameters of the Smith predictor according to the test data.

9. The method for optimizing the start-up control of a full-angle mode of a hemispherical resonant gyroscope according to claim 8, characterized in that: In S6.3, the control strategy is adjusted and optimized according to the test results, specifically using an adaptive control optimization algorithm to optimize the control parameters of the hemispherical resonant gyro control system in real time, wherein the adaptive control optimization algorithm adopts a reinforcement learning algorithm (RL); ； Among them, S is the various operating parameters of the hemispherical resonant gyroscope, including amplitude A ,frequency f ,temperature T ; ； in, Refers to all the actions that a HRG control system can perform, including adjusting amplitude, changing frequency, and adjusting temperature; ； in, is the reward function, including response time, stability and accuracy, S represents the current working state of the hemispherical resonant gyro control system, For action space One of the actions in , and are the weights corresponding to response time, stability, and accuracy; is the response time, is stability, is accuracy; ； in, is the action value function, which means that In the case of , start from state S and take action Expected return; is the expected return, t is the time step, is the reward function at time step t, S is the current state, is the initial state, is the initial action; Indicates the selection of action in state S probability; is the discount factor.

10. The method for optimizing the start-up control of a full-angle mode of a hemispherical resonant gyroscope according to claim 1, characterized in that: S7 specifically includes the following steps: S7.1: Through the real-time data acquisition system, continuously collect the data of the hemispherical resonant gyroscope during the full-angle mode startup process for further training and optimization of the LSTM network model and the autoregressive moving average model ARMA; S7.2: Continuously adjust and optimize the parameters of the LSTM network model and the autoregressive moving average model ARMA based on new data and feedback.

Images