A vibration active suppression method based on sensor network and artificial intelligence

Abstract

The application discloses a kind of vibration active suppression method based on sensor network and artificial intelligence, it is related to precision servo control technical field, the method executes structure modal and dynamic characteristic acquisition, multi-source sensor network layout and data acquisition, deep neural network construction and vibration reverse decoupling, artificial intelligence decision module offline training, real-time vibration suppression closed-loop control step in turn, vibration data are collected by sensor network, equivalent vibration interference at actuator end is decoupled out using deep neural network reverse, optimal compensation strategy is generated by combining intelligent toilet module trained by deep reinforcement learning, and closed-loop control is realized by driving actuator.The application constructs complete vibration simulation technology system, realizes the nonlinear high-precision decoupling and intelligent adaptive control of vibration, engineering practicability is strong, robustness is good, effectively improves the tracking accuracy and operating stability of system.

Description

Technical Field

[0001] This invention relates to the field of precision servo control technology, specifically to a method for active vibration suppression based on sensor networks and artificial intelligence. Background Technology

[0002] In the field of precision servo control technology, high-precision tracking servo systems are core equipment for high-end applications such as precision optical tracking, laser communication, and astronomical observation. These systems often use fast-reflecting mirrors as the core actuators to achieve high-bandwidth servo control. However, their mechanical structures are complex and have multiple modes. Under external excitations such as the bumps and vibrations from aircraft, vehicles, and ships, as well as random vibrations from natural wind loads, the resonance of the system's mechanical structure is easily excited and transmitted to the actuator, coupling with the control signal. This leads to problems such as phase lag and resonance peaks in the servo loop, and in severe cases, it can even cause system instability and significantly reduce tracking and aiming accuracy. To address this issue, existing technologies primarily employ three methods: active and passive vibration isolation, adding notch filters, and feedforward compensation. However, none of these methods constitute a complete technical system for real-time vibration control based on the structural characteristics of the system, and they share common technical shortcomings: active and passive vibration isolation has limited effectiveness in suppressing low-frequency vibrations and vibrations near the structure's natural frequencies; notch filters rely on precise system parameter identification and have poor suppression capabilities for time-varying and multi-modal resonances; feedforward compensation struggles to handle nonlinear and multi-path coupled vibration transmission; and existing technologies all use linear methods to decouple vibration responses, failing to accurately quantify the equivalent vibration interference at the actuator end. Control strategy formulation also relies on human experience rather than intelligent algorithms, and real-time control only considers a single target error signal without taking vibration interference compensation into account. Overall, these technologies cannot solve the problem of accurately suppressing multi-modal, time-varying, and multi-path coupled vibrations under complex vibration environments, making it difficult to meet the high requirements of control accuracy and stability for high-precision tracking servo systems. Summary of the Invention

[0003] To address the shortcomings of existing technologies, the present invention aims to provide an active vibration suppression method based on sensor networks and artificial intelligence. It constructs a complete vibration simulation technology system, realizes nonlinear high-precision decoupling and intelligent adaptive control of vibration, has strong engineering applicability and good robustness, and effectively improves the tracking accuracy and operational stability of the system.

[0004] To achieve the above objectives, the embodiments of this invention provide the following technical solutions:

[0005] This application provides a vibration active suppression method based on sensor networks and artificial intelligence, applied to a high-precision tracking servo system. The method includes the following steps: S1. Conducting structural modal tests on the high-precision tracking servo system. A wideband excitation is applied to the high-precision tracking servo system using a vibrator. Vibration response data from the structural modal tests is collected. Modal analysis is performed based on the vibration response data to obtain the natural frequency, damping ratio, and mode shape of the high-precision tracking servo system as structural dynamic characteristic parameters, and a structural response database for the high-precision tracking servo system is constructed. S2. Based on the results of the modal analysis, the key coupling points of the high-precision tracking servo system are determined, and multiple acceleration sensors are deployed at the key coupling points to form a sensor network. Typical and boundary conditions are applied to the high-precision tracking servo system. Vibration response data from each acceleration sensor is collected in real time using a data acquisition system, and the collected vibration response data is preprocessed. S3. Constructing a deep neural network. The input of the deep neural network is the current and historical acceleration signal sequences from the sensor network. The output is the equivalent vibration disturbance coupled to the actuator; based on the preprocessed working condition vibration response data as offline data, the deep neural network is trained to minimize the error between the estimated disturbance output by the deep neural network and the actual disturbance. The trained deep neural network is used to reversely deconstruct the external vibration excitation characteristics and the resonance mode of the actuator; S4, an artificial intelligence decision module is constructed using a deep intensity learning algorithm, and a simulation model of the high-precision tracking servo system is constructed. Different boundary working conditions are used as the environmental states of the simulation model, controller commands are used as actions, and the tracking accuracy index of the high-precision tracking servo system is used as the reward function. The artificial intelligence decision module is trained offline; S5, when the high-precision tracking servo system is running, the equivalent vibration disturbance output by the deep neural network and the current working state of the high-precision tracking servo system are obtained. The equivalent vibration disturbance and the current working state are input to the trained artificial intelligence decision module, which generates the optimal compensation control strategy in real time. The control commands of the actuator are adjusted according to the optimal compensation control strategy to drive the actuator to perform closed-loop control.

[0006] Further, S1 specifically includes the following steps: S11, estimating the natural frequency range of the high-precision tracking servo system structure, dividing the measurement point network covering key parts, installing accelerometers at the measurement points, applying sinusoidal vibration with continuously changing frequency through an exciter, and synchronously acquiring the excitation signal of the exciter and the response signal of the accelerometer through a signal acquisition system; S12, fitting the acquired excitation signal and response signal using modal analysis software, identifying the natural frequency, damping ratio, and mode shape of the high-precision tracking servo system from the frequency response function, and constructing the structural response database of the high-precision tracking servo system through database software.

[0007] Further, S2 specifically includes the following steps: S21, based on the obtained high-precision tracking servo system structural response database and frequency response function, determine the installation points of the sensor network, and deploy a sensor network composed of triaxial MEMS accelerometers or piezoelectric accelerometers, wherein the working bandwidth of the triaxial MEMS accelerometers or piezoelectric accelerometers is 0-2000Hz; S22, load typical working conditions and boundary working conditions onto the high-precision tracking servo system using a vibration table, and collect the working condition vibration response data of the sensor network nodes in real time through a data acquisition system, and perform preprocessing to filter out power frequency interference and denoise the collected working condition vibration response data.

[0008] Further, S3 specifically includes the following steps: S31, constructing a CNN-LSTM hybrid network or a Transformer temporal network as the deep neural network, setting the input of the deep neural network as the current and historical sensor network acceleration signal sequence, and the output as the equivalent disturbance torque or equivalent disturbance displacement coupled to the actuator; S32, using the preprocessed working condition vibration response data as offline data to train the deep neural network, with the training objective being to minimize the error between the estimated disturbance output by the deep neural network and the actual disturbance.

[0009] Furthermore, in step S32, before training the deep neural network, preprocessing of offline data is also included, including: normalizing the preprocessed vibration response data, reconstructing the one-dimensional time series signal into a format that the deep neural network can input through a sliding time window, and dividing the reconstructed dataset into a training set, a validation set, and a test set.

[0010] Furthermore, when the deep neural network is a CNN-LSTM hybrid network, the construction in S31 specifically includes: constructing an input layer, a CNN feature extraction layer, a feature transformation layer, an LSTM layer, and an output layer connected in sequence. The CNN feature extraction layer uses a 1D convolutional layer with a 1D pooling layer and the activation function is ReLU. The output layer is a fully connected layer with a linear activation function and the output dimension is the number of vibration interference prediction parameters.

[0011] Further, S4 specifically includes: S41, constructing a simulation model of the high-precision tracking servo system, taking different boundary conditions as the state input of the simulation model, taking the control commands of the controller as actions, and taking the RMS error of tracking accuracy, control bandwidth, and tracking jitter as the constituent indicators of the reward function; S42, substituting the simulation model, state input, actions, and reward function into a deep reinforcement learning algorithm to perform offline training on the artificial intelligence decision-making module.

[0012] Further, S5 specifically includes the following steps: S51, acquiring the vibration mode of the deep neural network inverse decoupling, the equivalent vibration interference of the actuator, and the current working state of the high-precision tracking servo system, the working state including target position, target velocity, target acceleration, and target error information; inputting the equivalent vibration interference, vibration mode, and current working state into the artificial intelligence decision module, which generates an optimal compensation control strategy in real time; S52, combining the optimal compensation control strategy with the target error signal, adjusting the control command of the actuator, and driving the fast-reflecting mirror to perform closed-loop control.

[0013] Furthermore, S5 also includes the following step: S53, updating the parameters of the deep neural network and the artificial intelligence decision-making module based on the actual data collected during the long-term operation of the high-precision tracking servo system.

[0014] The beneficial effects of this invention are as follows: By constructing a complete active vibration suppression technology system that integrates structural modal feature acquisition, multi-source sensor network data acquisition, deep neural network vibration inverse decoupling, deep reinforcement learning intelligent strategy training, and real-time closed-loop vibration suppression, this invention deeply combines the distributed data acquisition advantages of sensor networks with the nonlinear decoupling and intelligent decision-making capabilities of artificial intelligence. This overcomes the limitations of linear decoupling and the drawbacks of empirical control in existing technologies. It achieves precise quantification and inverse decoupling of equivalent vibration interference at the actuator end, providing a clear and precise target for vibration compensation control. Furthermore, the steps of the entire technical solution are closely connected and logically coherent, fully adapting to the working characteristics and control requirements of high-precision tracking servo systems. It can effectively suppress multimodal, time-varying, and multipath coupled vibration interference, fundamentally improving the tracking accuracy and operational stability of the system. At the same time, it lays a complete and solid technical foundation for the subsequent refinement and optimization of each step. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating an active vibration suppression method based on sensor networks and artificial intelligence, provided as an embodiment of this application. Detailed Implementation

[0016] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the scope of the present invention.

[0017] In this invention, the terms "system" and "network" are used interchangeably. "Multiple" refers to two or more; therefore, in this invention, "multiple" can also be understood as "at least two." "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. Additionally, the character " / ", unless otherwise specified, generally indicates that the preceding and following related objects have an "or" relationship. Furthermore, it should be understood that in the description of this invention, terms such as "first" and "second" are used only for descriptive purposes and should not be construed as indicating or implying relative importance or order.

[0018] like Figure 1As shown, this application provides a vibration active suppression method based on sensor networks and artificial intelligence, applied to a high-precision tracking servo system. The method includes the following steps: S1. Conducting structural modal tests on the high-precision tracking servo system. A wideband excitation is applied to the high-precision tracking servo system using a vibrator. Vibration response data from the structural modal tests is collected. Modal analysis is performed based on the vibration response data to obtain the natural frequency, damping ratio, and mode shape of the high-precision tracking servo system as structural dynamic characteristic parameters, and a structural response database for the high-precision tracking servo system is constructed. S2. Based on the results of the modal analysis, the key coupling points of the high-precision tracking servo system are determined, and multiple acceleration sensors are deployed at the key coupling points to form a sensor network. Typical and boundary conditions are applied to the high-precision tracking servo system. Vibration response data from each acceleration sensor is collected in real time using a data acquisition system, and the collected vibration response data is preprocessed. S3. Constructing a deep neural network. The input to the deep neural network is the current and historical acceleration signal sequences from the sensor network. The output is the equivalent vibration disturbance coupled to the actuator; based on the preprocessed vibration response data of the working condition as offline data, the deep neural network is trained to minimize the error between the estimated disturbance output by the deep neural network and the actual disturbance. The trained deep neural network is used to reversely deconstruct the external vibration excitation characteristics and the resonance mode of the actuator; S4, an artificial intelligence decision-making module is constructed using a deep intensity learning algorithm, and a simulation model of the high-precision tracking servo system is constructed. Different boundary conditions are used as the environmental states of the simulation model, controller commands are used as actions, and the tracking accuracy index of the high-precision tracking servo system is used as the reward function. The artificial intelligence decision-making module is trained offline; S5, when the high-precision tracking servo system is running, the equivalent vibration disturbance output by the deep neural network and the current working state of the high-precision tracking servo system are obtained. The equivalent vibration disturbance and the current working state are input to the trained artificial intelligence decision-making module, which generates the optimal compensation control strategy in real time. The control commands of the actuator are adjusted according to the optimal compensation control strategy to drive the actuator to perform closed-loop control.

[0019] In another possible embodiment, structural modal tests are first conducted on the high-precision tracking servo system. A wideband excitation is applied to the system using a vibrator; the wideband excitation is an excitation signal covering the system's estimated natural frequency range. Modal test vibration response data is collected using temporarily deployed accelerometers. Then, professional modal analysis is performed based on this vibration response data. Core structural dynamic characteristic parameters such as the system's natural frequency, damping ratio, and mode shape are identified and obtained from the frequency response function. Subsequently, database software is used to classify and store these parameters and test data, constructing a system structural response database to provide a basis for subsequent sensor network deployment. Then, based on the modal analysis results… Once the critical coupling points of the system are identified—locations where vibrations are easily generated and transmitted to the actuators, such as stress concentration points, actuator connection points, and peak points of structural mode shapes—multiple accelerometers are deployed at these critical coupling points to form a multi-source sensor network. The system is then subjected to typical and boundary conditions using a vibration table. Typical conditions include conventional turbulence vibrations from aircraft, vehicles, and ships, as well as conventional vibrations from natural wind loads. Boundary conditions include extreme wind speed vibrations, extreme turbulence vibrations, and vibrations under the system's extreme operating conditions. A high-speed data acquisition system is used to collect the vibration response data from each sensor in real time. Finally, the collected vibration response data is filtered to remove power frequency interference. The process involves several steps: 1) Noise reduction preprocessing to ensure data validity; 2) Constructing a deep neural network, using current and historical sensor network acceleration signal sequences as input and equivalent vibration disturbance coupled to the actuator control as output; 3) Using the preprocessed vibration response data as offline training data to train the deep neural network, with the core objective of minimizing the error between the network output's estimated disturbance and the actual system disturbance. The trained deep neural network can then reverse-engineer the external vibration excitation characteristics and the actuator's resonant modes. External vibration excitation characteristics include the frequency, amplitude, direction, and time-varying nature of the vibration; resonant modes include the resonant frequency, amplitude, direction, and time-varying nature of the vibration. The frequency, amplitude, and phase of the vibration are directly output as a precise equivalent vibration disturbance. Subsequently, a deep reinforcement learning algorithm is used to construct an artificial intelligence decision-making module. The deep reinforcement learning algorithm is a reinforcement learning algorithm applicable to continuous control. Then, simulation software is used to construct a simulation model consistent with the actual high-precision tracking servo system. Different boundary conditions are set as the environmental state of the simulation model, the control commands issued by the controller to the actuator are set as the intelligent agent actions, and the tracking accuracy index of the system is set as the reward function. Then, the environmental state, actions, and reward function are substituted into the simulation model to train the artificial intelligence decision-making module offline, so that the module learns the control strategy generation logic under different vibration conditions.Finally, during the actual operation of the high-precision tracking servo system, the equivalent vibration disturbance output by the previously trained deep neural network is acquired in real time. Simultaneously, the current operating state of the system is collected. The equivalent vibration disturbance and the current system operating state are input into the trained artificial intelligence decision-making module. The module then generates an optimal compensation control strategy in real time. Based on this optimal compensation control strategy, the original control commands of the actuator are adjusted to drive the actuator through closed-loop control. By compensating for the vibration, the coupling effect on the actuator is offset, thus achieving active suppression of system vibration.

[0020] By constructing a complete active vibration suppression technology system—from structural modal feature acquisition, multi-source sensor network data acquisition, deep neural network vibration inverse decoupling, deep reinforcement learning intelligent strategy training, to real-time closed-loop vibration suppression—this system deeply integrates the distributed data acquisition advantages of sensor networks with the nonlinear decoupling and intelligent decision-making capabilities of artificial intelligence. This overcomes the limitations of linear decoupling and the drawbacks of empirical control in existing technologies. It achieves precise quantification and inverse decoupling of equivalent vibration disturbances at the actuator end, providing a clear and precise target for vibration compensation control. Furthermore, the various steps of the entire technical solution are closely connected and logically coherent, fully adapting to the working characteristics and control requirements of high-precision tracking servo systems. It effectively suppresses multimodal, time-varying, and multipath-coupled vibration disturbances, fundamentally improving the system's tracking accuracy and operational stability. Simultaneously, it lays a complete and solid technical foundation for the subsequent refinement and optimization of each step.

[0021] In existing technologies, there is no standardized and unified implementation process for structural modal testing of high-precision tracking servo systems. The layout of the measuring point network is highly arbitrary and does not cover the key structural parts of the system, resulting in the inability to fully capture the vibration response of the system. Furthermore, the data acquisition process lacks a repeated verification step, which easily leads to distortion of the acquired test data due to random noise. At the same time, modal analysis only performs simple processing and calculation on the data and cannot accurately identify core parameters such as natural frequency, damping ratio, and mode shape from the frequency response function. This results in inaccurate acquisition of the dynamic characteristic parameters of the system structure. Subsequent steps such as sensor network deployment and vibration decoupling modeling lack reliable basic data support, ultimately directly affecting the implementation effect of the entire vibration suppression method.

[0022] In the embodiments of this application, S1 specifically includes the following steps: S11, estimating the natural frequency range of the high-precision tracking servo system structure, dividing the measurement point network covering key parts, installing accelerometers at the measurement points, applying sinusoidal vibration with continuously changing frequency through an exciter, and synchronously acquiring the excitation signal of the exciter and the response signal of the accelerometer through a signal acquisition system; S12, fitting the acquired excitation signal and response signal using modal analysis software, identifying the natural frequency, damping ratio, and mode shape of the high-precision tracking servo system from the frequency response function, and constructing the structural response database of the high-precision tracking servo system through database software.

[0023] In another possible embodiment, the natural frequency range of the system is first estimated based on the mechanical structure design drawings of the high-precision tracking servo system and engineering experience. Then, based on the estimation results, a network of measurement points covering key structural components is defined, including stress concentration points, vibration-sensitive areas, actuator connection points, and structural weak points. Accelerometers are installed at each measurement point. The exciter is rigidly connected to the system structure. A continuously varying sinusoidal vibration is applied to the system by the exciter to fully excite the various modes of the system. A high-speed signal acquisition system is then used to synchronously acquire the excitation signal from the exciter and the vibration response signals from each sensor. To reduce the impact of random noise on the data, this modal test... The process is repeated multiple times, and the valid data collected from each iteration is used as the basis for subsequent modal analysis. The previously collected multiple sets of measurement point data are then imported into modal analysis software, which is a professional software specifically designed for structural modal characteristic analysis. This software performs professional fitting processing on the measurement point data, accurately identifying core structural dynamic characteristic parameters such as the natural frequency, damping ratio, and mode shape of the high-precision tracking servo system from the frequency response function. The identified parameters, original test data, frequency response function curves, and other relevant information are then imported into database software. The information is categorized, organized, and stored according to the structural parts of the system and the number of tests, ultimately constructing a complete and standardized system structural response database.

[0024] By standardizing and refining the steps for acquiring structural modal and dynamic characteristics, the process is divided into two sub-steps: modal testing and modal analysis and database construction. This standardizes the complete implementation process of modal testing and the professional operation methods of modal analysis. By rationally dividing the measurement point network covering key parts of the system structure and synchronously acquiring excitation and response signals, the comprehensiveness and accuracy of the modal test data are ensured. By using professional modal analysis software to fit the test data, the identification accuracy of the system structure's dynamic characteristic parameters is significantly improved. The constructed system structure response database provides a reliable basis for the accurate and scientific deployment of the subsequent sensor network, ensuring the implementation accuracy and effectiveness of the entire vibration suppression technology solution from the source.

[0025] In existing technologies, the selection of sensor installation points for vibration data acquisition lacks clear scientific basis and relies heavily on manual experience. This results in sensors failing to capture key vibration signals transmitted to the actuator. Furthermore, the sensor type and operating bandwidth are mismatched with the system's vibration characteristics, making it impossible to accurately capture complex vibration signals involving a mixture of high and low frequencies. Additionally, the system is subjected to a single type of operating condition, failing to cover typical and boundary conditions in actual operation. Consequently, the acquired vibration data cannot accurately reflect the actual vibration state of the system. Moreover, data preprocessing only involves simple filtering operations without specifically filtering out power frequency interference or effectively reducing noise. This leads to high noise levels and limited effective information in the acquired vibration data, significantly reducing the training accuracy and reliability of subsequent artificial intelligence algorithms.

[0026] In the embodiments of this application, S2 specifically includes the following steps: S21, based on the obtained high-precision tracking servo system structural response database and frequency response function, determine the installation points of the sensor network, and deploy a sensor network composed of triaxial MEMS accelerometers or piezoelectric accelerometers, wherein the working bandwidth of the triaxial MEMS accelerometers or piezoelectric accelerometers is 0-2000Hz; S22, load typical working conditions and boundary working conditions onto the high-precision tracking servo system using a vibration table, and collect the working condition vibration response data of the sensor network nodes in real time through a data acquisition system, and perform preprocessing to filter out power frequency interference and denoise the collected working condition vibration response data.

[0027] In another possible embodiment, the system structural response database and frequency response function obtained from modal analysis are first retrieved. Combined with the location of key coupling points in the system, the specific installation points of the multi-source sensor network are scientifically determined. The selection of installation points is based on the core principle of comprehensively capturing all vibration signals transmitted to the actuator. Accelerometers are then deployed at the determined installation points. These accelerometers are either triaxial MEMS accelerometers or piezoelectric accelerometers. Triaxial MEMS accelerometers are suitable for acquiring low-frequency vibrations of the system, while piezoelectric accelerometers are suitable for acquiring high-frequency vibrations. Furthermore, the operating bandwidth of these accelerometers can cover the full frequency vibration range of the system. Then, all... The deployed accelerometers are connected to a high-speed data acquisition system to complete the overall construction of a multi-source sensor network. Then, a high-precision tracking servo system is fixed on a vibration table, and typical and boundary conditions are applied to the system sequentially through the vibration table. The high-speed data acquisition system collects the vibration response data of each sensor network node in real time. The collected raw vibration response data is then preprocessed in a targeted manner. First, a notch filter is used to filter out power frequency interference, which is a fixed-frequency interference signal introduced by the power grid. Then, a professional denoising method is used to denoise the data, removing random noise from the data, and finally obtaining clean and effective vibration response data.

[0028] The installation points of the sensor network were determined by modal analysis results and the system structural response database, ensuring the scientific, targeted, and rational placement of the sensors. The specific types and operating bandwidths of the sensors were clarified, enabling them to accurately adapt to the vibration characteristics of the system and achieve comprehensive and accurate acquisition of high and low frequency vibration signals. Typical and boundary conditions were applied using a vibration table, allowing the acquired vibration response data to comprehensively reflect all vibration states in the actual operation of the system. Targeted data preprocessing operations were developed to effectively filter out power frequency interference and random noise, improving the validity and purity of the data and providing a high-quality dataset for subsequent deep neural network training.

[0029] Existing technologies for decoupling vibration response all employ linear transfer function estimation methods. However, this method can only handle simple linear, single-path vibration transmission problems and cannot address the nonlinear, multi-path coupled vibration transmission problems that objectively exist in high-precision tracking servo systems. The decoupling accuracy is extremely low. Furthermore, existing technologies do not provide specific quantitative classifications of vibration interference at the actuator end, only describing it as "vibration interference" in a general sense. This makes it impossible to accurately determine the specific form of vibration's effect on the actuator, resulting in a lack of precise targets for subsequent vibration compensation control and a significant reduction in vibration suppression effectiveness.

[0030] In the embodiments of this application, S3 specifically includes the following steps: S31, constructing a CNN-LSTM hybrid network or a Transformer temporal network as the deep neural network, setting the input of the deep neural network as the current and historical sensor network acceleration signal sequence, and the output as the equivalent disturbance torque or equivalent disturbance displacement coupled to the actuator; S32, using the preprocessed working condition vibration response data as offline data to train the deep neural network, with the training objective being to minimize the error between the estimated disturbance output by the deep neural network and the actual disturbance.

[0031] In another possible embodiment, based on the vibration data characteristics of the high-precision tracking servo system, a CNN-LSTM hybrid network or a Transformer temporal network is first constructed as the deep neural network of this invention. Both of these networks can accurately adapt to the processing and decoupling requirements of temporal vibration data. Then, the input of this deep neural network is set to the current and historical sensor network acceleration signal sequences, and the network output is set to the equivalent disturbance torque or equivalent disturbance displacement coupled to the actuator control end. The equivalent disturbance torque is the disturbance torque that causes angular vibration in the actuator, suitable for fast-reflecting mirror angular displacement actuators; the equivalent disturbance displacement is the disturbance displacement that causes linear vibration in the actuator, suitable for linear motion actuators. This achieves precise quantification of vibration interference at the actuator end. Then, preprocessed vibration response data is retrieved and used as offline training data for the deep neural network. The training objective of the network is to minimize the error between the network's estimated interference and the actual system interference. The actual system interference is accurately obtained in two ways: first, by back-calculating from the finite element model of a high-precision tracking servo system; second, by indirectly observing the actual vibration state at the actuator end using high-precision sensors. The offline training data is then input into the constructed deep neural network, and a professional optimization algorithm is used to train the network until the error between the network's estimated interference and the actual interference reaches a preset ideal threshold, thus completing the offline training of the deep neural network.

[0032] By clarifying the specific type of deep neural network, it can accurately adapt to the processing and decoupling requirements of time-series vibration data; by quantifying the equivalent vibration disturbance at the actuator end into equivalent disturbance torque or equivalent disturbance displacement, the specific action form of vibration on the actuator is accurately determined, giving subsequent compensation control a clear target; by using deep neural networks to achieve nonlinear inverse decoupling of vibration response, it breaks through the accuracy limitations of traditional linear transfer function estimation, accurately handles the vibration transmission problem of multi-path and multi-modal coupling in the system, and significantly improves decoupling accuracy; at the same time, the core training objective of deep neural networks is clarified, ensuring that the estimated disturbance output by the network is highly matched with the actual disturbance of the system, providing accurate and reliable vibration disturbance data for the generation of subsequent intelligent control strategies.

[0033] If offline training data is not standardized and restructured before training a deep neural network, inconsistencies in the dimensions and numerical ranges of the original data will lead to slow network training convergence and problems such as vanishing and exploding gradients. Furthermore, the format of one-dimensional temporal vibration data does not match the input requirements of deep neural networks and cannot be directly input into the network for training. In addition, if the dataset is not properly divided and all data is used directly for network training, it is impossible to effectively verify the network's generalization ability and prediction accuracy, which can easily lead to overfitting or underfitting. Ultimately, this will significantly reduce the vibration decoupling accuracy and operational reliability of the deep neural network.

[0034] In the embodiments of this application, S32 further includes preprocessing of offline data before training the deep neural network, including: normalizing the preprocessed working condition vibration response data, reconstructing the one-dimensional time series signal into a format that the deep neural network can input through a sliding time window, and dividing the reconstructed dataset into a training set, a validation set, and a test set.

[0035] In another possible embodiment, before performing the offline training step of the network, the preprocessed vibration response data is first standardized. Then, the processed dataset is input into the deep neural network for training. First, a professional normalization method is used to normalize the vibration response data to eliminate differences in the dimensions and numerical ranges of the data, mapping all data to a unified numerical range to ensure data standardization. Then, the sliding time window method is used to reconstruct the one-dimensional time-series vibration response signal into a format that the deep neural network can directly input. The sliding time window method divides the continuous one-dimensional time-series signal into multiple network-recognizable sample data by setting a fixed window length and step size. Then, the reconstructed normalized dataset is divided into training set, validation set, and test set according to a reasonable ratio. The training set is used for fitting training of the deep neural network, the validation set is used for network parameter tuning and overfitting monitoring during the training process, and the test set is used for final performance evaluation of the network after training to ensure that the network has good generalization ability. After completing the above data preprocessing operations, the processed training set and validation set data are input into the deep neural network for offline training.

[0036] By supplementing and optimizing the offline training steps of deep neural networks, a standardized offline data preprocessing step was added before network training. Normalization eliminated the differences in the dimensions and numerical ranges of the training data, significantly improving the training convergence speed of deep neural networks. By using a sliding time window to reconstruct one-dimensional time-series signals into a format that can be directly input into deep neural networks, the adaptability of data and networks was ensured. By reasonably dividing the reconstructed dataset, the phased verification and evaluation of network training effects were achieved, which effectively avoided overfitting or underfitting problems in the network and significantly improved the generalization ability of deep neural networks and the prediction accuracy of vibration decoupling.

[0037] When using neural networks to process time-series vibration data, if 2D convolutional layers are used for feature extraction, there will be obvious dimensionality redundancy problems, making it impossible to accurately capture the local correlation features of time-series data. At the same time, unreasonable network structure design, without the use of pooling layers for feature dimensionality reduction, leads to excessive network computation, low training efficiency, and inappropriate selection of activation functions, which can easily cause gradient vanishing problems, resulting in insufficient feature extraction of vibration data by the network.

[0038] In the embodiments of this application, when the deep neural network is a CNN-LSTM hybrid network, the construction in S31 specifically includes: constructing an input layer, a CNN feature extraction layer, a feature transformation layer, an LSTM layer and an output layer connected in sequence. The CNN feature extraction layer uses a 1D convolutional layer and a 1D pooling layer, and the activation function is ReLU. The output layer is a fully connected layer, uses a linear activation function, and the output dimension is the number of vibration interference prediction parameters.

[0039] In another possible embodiment, when the deep neural network is a CNN-LSTM hybrid network, a complete network structure is constructed in a sequential manner, consisting of an input layer, a CNN feature extraction layer, a feature transformation layer, an LSTM layer, and an output layer. First, the input layer is constructed, with its input dimension matching the format of the reconstructed temporal vibration data, allowing it to directly receive the reconstructed temporal vibration response data. Then, a CNN is constructed after the input layer. The feature extraction layer employs a 1D convolutional layer for temporal feature extraction. A reasonable kernel size and number of filters are set based on the characteristics of the vibration data. The convolutional kernels capture the correlation of vibration features between adjacent time steps, while the filters extract different features from the data. A 1D pooling layer follows the 1D convolutional layer to reduce the dimensionality of the extracted features, thus reducing network computation. Furthermore, the activation function of the CNN feature extraction layer uses the ReLU function, effectively alleviating the gradient vanishing problem during network training. A feature transformation layer is then built after the CNN feature extraction layer. A specialized function converts the feature data output by the CNN feature extraction layer into a format directly acceptable to the LSTM layer, achieving the conversion from convolutional features to temporal features. Subsequently, an LSTM layer is constructed, with a reasonable number of LSTM units set according to the complexity of the vibration data. Multiple LSTM layers can be stacked as needed to fully capture the long-term temporal correlation features of the vibration data, improving the sufficiency of feature extraction. Finally, an output layer is built after the LSTM layer. This output layer uses a fully connected layer with a linear activation function. The output dimension of the output layer is consistent with the number of predicted parameters for vibration interference, allowing direct output of the equivalent vibration interference estimate at the actuator end.

[0040] By employing a 1D convolutional layer to adapt to the one-dimensional structure of temporal vibration data, the local temporal correlation features of the data can be accurately captured, effectively avoiding the dimensionality redundancy problem of 2D convolutional layers. The 1D pooling layer is used to reduce the dimensionality of the extracted features, significantly reducing the computational load of the network and improving its training efficiency. The ReLU activation function effectively alleviates the gradient vanishing problem, ensuring that the network fully extracts the spatiotemporal features of the vibration data. The output layer is designed according to the number of predicted vibration interference parameters, ensuring that the network output highly matches the actual vibration suppression requirements, significantly improving the vibration decoupling accuracy and training efficiency of the CNN-LSTM hybrid network.

[0041] If the reward function is designed in a simplistic way, only considering a single tracking accuracy metric without taking into account core operational metrics such as system control bandwidth and tracking jitter, the generated control strategy will be unable to balance the multi-dimensional system control requirements. Furthermore, if the control strategy training does not clearly define the specific actions of the agent and the environmental state, and the training process lacks standardization, the resulting control strategy will ultimately have extremely poor adaptability and effectiveness.

[0042] In the embodiments of this application, S4 specifically includes: S41, constructing a simulation model of the high-precision tracking servo system, taking different boundary conditions as the state input of the simulation model, taking the control commands of the controller as actions, and taking the RMS error of tracking accuracy, control bandwidth, and tracking jitter as the constituent indicators of the reward function; S42, substituting the simulation model, state input, actions, and reward function into a deep reinforcement learning algorithm to perform offline training on the artificial intelligence decision module.

[0043] In another possible embodiment, a simulation model matching the actual system at a 1:1 scale is first constructed using specialized simulation software based on the mechanical structure, electrical parameters, and control logic of the high-precision tracking servo system. This simulation model can accurately simulate the actual working state of the system under different vibration conditions. Then, different boundary conditions of the high-precision tracking servo system are set as the state input of the simulation environment, and the control commands issued by the controller to the actuator are set as the actions of the reinforcement learning agent. The reward function is designed with the RMS error of the system tracking accuracy, control bandwidth, and tracking jitter as the core indicators. The RMS error, or root mean square error, is the core indicator for measuring the system tracking deviation, and the control bandwidth is the core indicator for measuring the real-time performance of the system control. Tracking jitter is a core indicator for measuring system stability. The smaller the RMS error, the wider the control bandwidth, and the smaller the tracking jitter, the larger the reward function value. The aforementioned simulation model, defined environmental inputs, agent actions, and designed reward function are all substituted into a deep reinforcement learning algorithm. These elements serve as the training environment for the AI ​​decision-making module, which is then trained offline. During training, the agent continuously selects corresponding actions based on the current environmental state. These actions, applied to the simulation model, result in new environmental states and reward values. Through continuous interactive learning, the agent gradually masters the logic for generating optimal control strategies under different vibration conditions.

[0044] By constructing a simulation model that matches the actual high-precision tracking servo system at a 1:1 scale, the training environment is highly consistent with the actual operating conditions of the system, ensuring the practical adaptability of the control strategy after training. Based on multiple core indicators such as RMS error, control bandwidth, and tracking jitter, a reward function is designed to enable the control strategy generated by the AI ​​decision-making module after training to take into account multiple core control requirements of the system and achieve globally optimal control performance. The specific definitions of the state input and agent actions in the simulation environment are clarified, the training process of deep reinforcement learning is standardized, and the training efficiency of the AI ​​decision-making module and the effectiveness of the control strategy are significantly improved.

[0045] If the calculation of the target Q value is only vaguely described as being based on future reward calculation, the calculation results will have a large deviation. In addition, the sampling of empirical data adopts a sequential sampling method, and there is a strong correlation between samples, which leads to low efficiency in updating network parameters and easy overfitting. Ultimately, this results in low training accuracy of the artificial intelligence decision-making module and poor optimization effect of the generated control strategy.

[0046] In the embodiments of this application, in step S42, the agent selects a control strategy based on the input state of the policy network and applies it to the simulation environment to obtain the next state and immediate reward, and then uses the interaction tuple. Stored in the experience replay pool, where This is the current state. For the current action, For instant rewards, The state for the next time step; random sampling of experience data from the experience replay pool, based on... The target Q value is calculated by discounting the maximum future reward, and the mean square error between the predicted Q value and the target Q value is defined as the loss function. The network parameters are updated using the gradient descent algorithm.

[0047] In another possible embodiment, the deep reinforcement learning agent first selects the corresponding control policy based on the input state of the policy network. The policy network, which is a reinforcement learning decision network composed of deep neural networks, is used to select the optimal action based on the environmental state. This control policy is then applied to the simulation environment of the high-precision tracking servo system. After receiving the control policy, the simulation environment outputs the next state and an immediate reward. The immediate reward is calculated by a pre-designed reward function. The interaction tuples during the interaction process are then stored in an experience replay pool. The experience replay pool is a specialized queue structure used to store system interaction data. Each interaction tuple includes the current state, the current action, the immediate reward, and the next state, with the specific meaning of each parameter clearly defined. The current state is the real-time state of the simulation environment, the current action is the control policy selected by the agent, and the immediate reward is the reward output by the simulation environment. The reward value is the new state of the simulation environment after the control strategy takes effect. When the interaction data in the experience replay pool reaches a preset amount, a batch of experience data is randomly sampled from the experience replay pool to avoid correlation between samples. This data is used as training data for updating network parameters. The target Q value is then calculated based on the maximum future reward discount of the next time step state in the sampled data. A reasonable discount of the future reward is achieved through a professional calculation formula. The predicted Q value output by the policy network is then compared with the calculated target Q value. The mean square error between the two is defined as the network's loss function. Finally, a professional gradient descent algorithm is used to optimize the loss function. The parameters of the policy network are updated through backpropagation to achieve optimized training of the network. All the above operations are repeated until the value of the reward function tends to stabilize, thus completing the offline training of the artificial intelligence decision-making module.

[0048] By standardizing the calculation method of the target Q-value, the target Q-value is calculated based on the maximum future reward discount of the next time step state, ensuring the accuracy and rationality of the target Q-value. Random sampling is used to obtain empirical data from the experience replay pool, effectively avoiding correlation between samples and improving the efficiency of network parameter updates. The definition of the loss function and the network parameter update algorithm are clarified, ensuring the standardization and effectiveness of the deep reinforcement learning training process, and significantly improving the training accuracy of the artificial intelligence decision-making module and the optimization effect of the control strategy.

[0049] In existing technologies, the real-time closed-loop control of high-precision tracking servo systems relies solely on single-target error signals collected by optical sensors, without considering vibration interference at the actuator end and the full-dimensional operating status of the system. This results in a lack of specificity in the generation of control strategies, which cannot effectively counteract the impact of vibration on the system.

[0050] In the embodiments of this application, S5 specifically includes the following steps: S51, obtaining the vibration mode of the deep neural network inverse decoupling and the equivalent vibration interference of the actuator, as well as the current working state of the high-precision tracking servo system, wherein the working state includes target position, target velocity, target acceleration, and target error information; inputting the equivalent vibration interference, vibration mode, and current working state into the artificial intelligence decision module, wherein the artificial intelligence decision module generates the optimal compensation control strategy in real time; S52, combining the optimal compensation control strategy with the target error signal, adjusting the control command of the actuator, and driving the fast-reflecting mirror to perform closed-loop control.

[0051] In another possible embodiment, during the actual operation of the high-precision tracking servo system, the vibration modes obtained from the inverse decoupling of the trained deep neural network and the equivalent vibration interference at the actuator end are acquired in real time. The vibration modes are the frequency, amplitude, direction, and time-varying nature of the external vibration. Simultaneously, the current operating state of the system is collected through various sensors, which specifically include the target's position, velocity, and acceleration, as well as target error information collected by optical sensors. The optical sensors are specifically CCD cameras, laser rangefinders, infrared imagers, and other sensors used for target detection. The target error information is the deviation between the actual position of the target and the tracking position of the system. Then, the aforementioned vibration modes, equivalent vibration interference, and the current operating state of the system are all input into the system. The trained AI decision-making module generates logic based on the pre-trained control strategy and outputs the optimal compensation control strategy adapted to the current vibration conditions in real time. This optimal compensation control strategy is then fused with the target error signal collected by the optical sensor. Based on the fusion result, the original control commands sent by the controller to the fast-reflecting mirror are adjusted. The fast-reflecting mirror, the core actuator of the high-precision tracking servo system, is either a piezoelectric or electromagnetic fast-reflecting mirror. The adjusted control commands drive the fast-reflecting mirror to perform corresponding actions. Through precise compensation of angle or position, the fast-reflecting mirror offsets the deviation caused by vibration coupling and accurately tracks the actual position of the target, achieving closed-loop control of the high-precision tracking servo system and actively suppressing system vibration.

[0052] By clarifying the complete input information for generating the real-time control strategy, and combining vibration modes, equivalent vibration disturbances, and the system's full-dimensional operating state, the optimal compensation control strategy generated by the artificial intelligence decision-making module becomes highly targeted, accurately offsetting the coupling effect of vibration on the actuator. The specific composition of the system's current operating state is clarified, ensuring the comprehensiveness of the input information for the control strategy generation. By combining the optimal compensation control strategy with the target error signal, the fast-reflecting mirror is driven for closed-loop control, achieving dual control of vibration compensation and target tracking. This significantly improves the closed-loop control accuracy and vibration suppression effect of the high-precision tracking servo system.

[0053] In the embodiments of this application, S5 further includes the following step: S53, updating the parameters of the deep neural network and the artificial intelligence decision module based on the actual data collected during the long-term operation of the high-precision tracking servo system.

[0054] In another possible embodiment, an online algorithm fine-tuning sub-step is added and executed synchronously with the aforementioned two sub-steps (S51, S52). During the long-term operation of the high-precision tracking servo system, actual operating condition vibration response data of the system is continuously collected through a multi-source sensor network. Simultaneously, control effect data of the system is collected through the system's professional detection module. The control effect data includes core indicators such as the system's RMS error, control bandwidth, and tracking jitter. The collected actual operating condition vibration response data and control effect data are summarized and filtered according to a preset period to remove abnormal data, obtaining effective actual system operation data. The filtered effective actual operation data is then used as small-batch training data and input into the deep neural network and artificial intelligence decision-making module. Mini-batch gradient descent is used to update the parameters of the two models in small increments, with strict control over the update magnitude to avoid system instability caused by large updates. The core objective of parameter updates is to ensure that the model output closely matches the actual operating state of the system, guaranteeing the vibration decoupling accuracy of the deep neural network and the effectiveness of the control strategy of the artificial intelligence decision-making module. Finally, after the parameter updates are completed, the control effect data of the system is monitored in real time. If all the core control effect indicators of the system remain within the preset optimal range, the parameter updates are confirmed to be effective. If the control effect of the system fluctuates significantly, the model parameters before the update are immediately restored, and the update cycle or magnitude is adjusted according to the actual operating conditions of the system, and small updates are performed again.

[0055] By adding a sub-step for online algorithm fine-tuning, a dedicated online fine-tuning mechanism was designed. This allows the parameters of the deep neural network and artificial intelligence decision-making module to be updated slightly based on the actual operating data of the system, effectively adapting to the slow time-varying nature of system structural parameters and the continuous changes in the external vibration environment. The parameter updates are conducted in small increments, avoiding system instability caused by large updates and ensuring the continuity and stability of system control. Through online parameter updates, the technical solution of this invention can maintain high vibration decoupling accuracy and control effect at all times, ensuring the tracking accuracy and vibration suppression stability of the high-precision tracking servo system during long-term operation.

[0056] The optional embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the embodiments of the present invention are not limited to the specific details in the above embodiments. Within the scope of the technical concept of the embodiments of the present invention, various simple modifications can be made to the technical solutions of the embodiments of the present invention, and these simple modifications all fall within the protection scope of the embodiments of the present invention.

[0057] It should also be noted that the various specific technical features described in the above embodiments can be combined in any suitable manner without contradiction. To avoid unnecessary repetition, the embodiments of the present invention will not describe the various possible combinations separately.

[0058] Furthermore, various different implementations of the present invention can be combined arbitrarily, as long as they do not violate the spirit of the present invention, they should also be regarded as the content disclosed in the present invention.

Claims

1. A vibration active suppression method based on sensor networks and artificial intelligence, applied to a high-precision tracking servo system, characterized in that, Includes the following steps: S1. Conduct structural modal tests on the high-precision tracking servo system. Apply broadband excitation to the high-precision tracking servo system through a vibrator, collect vibration response data of the structural modal test, perform modal analysis based on the vibration response data of the structural modal test, obtain the natural frequency, damping ratio, and mode shape of the high-precision tracking servo system as structural dynamic characteristic parameters, and construct a structural response database of the high-precision tracking servo system. S2. Based on the results of the modal analysis, determine the key coupling points of the high-precision tracking servo system, and deploy multiple acceleration sensors at the key coupling points to form a sensor network. The high-precision tracking servo system is subjected to typical and boundary conditions. The vibration response data of each acceleration sensor is collected in real time through the data acquisition system, and the collected vibration response data is preprocessed. S3. Construct a deep neural network. The input of the deep neural network is the current and historical sensor network acceleration signal sequence, and the output is the equivalent vibration disturbance coupled to the actuator. Based on the preprocessed working condition vibration response data as offline data, the deep neural network is trained to minimize the error between the estimated disturbance output by the deep neural network and the actual disturbance. The trained deep neural network is used to reversely deconstruct the external vibration excitation characteristics and the resonance mode of the actuator. S4. An artificial intelligence decision-making module is constructed using a deep intensity learning algorithm, and a simulation model of the high-precision tracking servo system is constructed. Different boundary conditions are used as the environmental states of the simulation model, controller commands are used as actions, and the tracking accuracy index of the high-precision tracking servo system is used as the reward function. The artificial intelligence decision-making module is then trained offline. S5. When the high-precision tracking servo system is running, the equivalent vibration disturbance output by the deep neural network and the current working state of the high-precision tracking servo system are obtained. The equivalent vibration disturbance and the current working state are input to the trained artificial intelligence decision module. The artificial intelligence decision module generates the optimal compensation control strategy in real time. The control command of the actuator is adjusted according to the optimal compensation control strategy to drive the actuator to perform closed-loop control.

2. The vibration active suppression method based on sensor networks and artificial intelligence according to claim 1, characterized in that, S1 specifically includes the following steps: S11. Estimate the natural frequency range of the high-precision tracking servo system structure, divide the measurement point network covering key parts, install acceleration sensors at the measurement points, apply sinusoidal vibration with continuously changing frequency through the exciter, and synchronously collect the excitation signal of the exciter and the response signal of the acceleration sensor through the signal acquisition system. S12. Modal analysis software is used to fit the collected excitation signal and response signal, and the natural frequency, damping ratio and mode shape of the high-precision tracking servo system are identified from the frequency response function. The structural response database of the high-precision tracking servo system is constructed using database software.

3. The vibration active suppression method based on sensor networks and artificial intelligence according to claim 2, characterized in that, S2 specifically includes the following steps: S21. Based on the obtained high-precision tracking servo system structural response database and frequency response function, determine the installation points of the sensor network and deploy a sensor network composed of triaxial MEMS accelerometers or piezoelectric accelerometers, wherein the working bandwidth of the triaxial MEMS accelerometers or piezoelectric accelerometers is 0-2000Hz. S22. Apply typical and boundary conditions to the high-precision tracking servo system using a vibration table, and collect the vibration response data of the sensor network nodes in real time using a data acquisition system. Perform preprocessing on the collected vibration response data to filter out power frequency interference and reduce noise.

4. The vibration active suppression method based on sensor networks and artificial intelligence according to claim 1, characterized in that, S3 specifically includes the following steps: S31. Construct a CNN-LSTM hybrid network or a Transformer temporal network as the deep neural network, set the input of the deep neural network as the current and historical sensor network acceleration signal sequence, and the output as the equivalent disturbance torque or equivalent disturbance displacement coupled to the actuator. S32. Using the preprocessed vibration response data as offline data, the deep neural network is trained. The training objective is to minimize the error between the estimated disturbance output by the deep neural network and the actual disturbance.

5. The vibration active suppression method based on sensor networks and artificial intelligence according to claim 4, characterized in that, In step S32, before training the deep neural network, the offline data is preprocessed, including: normalizing the preprocessed vibration response data, reconstructing the one-dimensional time series signal into a format that the deep neural network can input through a sliding time window, and dividing the reconstructed dataset into a training set, a validation set, and a test set.

6. The vibration active suppression method based on sensor networks and artificial intelligence according to claim 4, characterized in that, When the deep neural network is a CNN-LSTM hybrid network, the construction in S31 specifically includes: constructing an input layer, a CNN feature extraction layer, a feature transformation layer, an LSTM layer and an output layer connected in sequence. The CNN feature extraction layer uses a 1D convolutional layer and a 1D pooling layer, and the activation function is ReLU. The output layer is a fully connected layer, uses a linear activation function, and the output dimension is the number of vibration interference prediction parameters.

7. The vibration active suppression method based on sensor networks and artificial intelligence according to claim 1, characterized in that, S4 specifically includes: S41. Construct a simulation model of the high-precision tracking servo system, take different boundary conditions as the state input of the simulation model, take the control command of the controller as the action, and take the RMS error of tracking accuracy, control bandwidth, and tracking jitter as the constituent indicators of the reward function. S42. Substitute the simulation model, state input, action, and reward function into a deep reinforcement learning algorithm to train the artificial intelligence decision-making module offline.

8. The vibration active suppression method based on sensor networks and artificial intelligence according to claim 1, characterized in that, S5 specifically includes the following steps: S51. Obtain the vibration mode of the deep neural network inverse decoupling and the equivalent vibration interference of the actuator, as well as the current working state of the high-precision tracking servo system. The working state includes target position, target velocity, target acceleration, and target error information. Input the equivalent vibration interference, vibration mode, and current working state into the artificial intelligence decision module, and the artificial intelligence decision module generates the optimal compensation control strategy in real time. S52. Combine the optimal compensation control strategy with the target error signal, adjust the control command of the actuator, and drive the fast-reflecting mirror to perform closed-loop control.

9. The vibration active suppression method based on sensor networks and artificial intelligence according to claim 8, characterized in that, S5 further includes the following steps: S53. Based on the actual data collected during the long-term operation of the high-precision tracking servo system, update the parameters of the deep neural network and the artificial intelligence decision-making module.

Images