An active vibration isolation method and system based on prediction and adaptive control

Abstract

The application belongs to the technical field of active vibration isolation, and specifically discloses an active vibration isolation method and system based on prediction and adaptive control. The method comprises the following steps: performing multi-degree-of-freedom dynamic modeling on a vibration isolation system, taking the pose of a load as the system output, taking the control quantity of an actuator as the system input, and uniformly representing unknown dynamics caused by structural nonlinearity and disturbance as a nonlinear term; constructing an adaptive neural network state observer based on the dynamic model to estimate states that cannot be directly measured; collecting historical operation data of the vibration isolation system, constructing input time series and output time series after preprocessing the data, training a long short-term memory neural network prediction model, and generating a prediction signal representing future position information; designing an adaptive control law according to the state estimation result and the prediction signal, obtaining a feedback control quantity, superimposing the feedback control quantity with a feedforward control quantity obtained based on the prediction signal, and forming a control instruction for driving the actuator.

Description

Technical Field

[0001] This invention belongs to the field of active vibration isolation technology, specifically relating to an active vibration isolation method and system based on prediction and adaptive control. Background Technology

[0002] Vibration, as a ubiquitous environmental disturbance, can significantly impact the performance of high-end equipment in precision manufacturing, nanoscale measurement, optical imaging, and aerospace testing. Especially in applications highly sensitive to micro-displacements, such as scanning electron microscopes and atomic force microscopes, even minute vibrations can lead to blurred images, measurement errors, or decreased machining accuracy. Therefore, isolating environmental vibrations is an indispensable and critical step in the operation of precision equipment.

[0003] Active vibration isolation technology, through a closed-loop system composed of sensors, controllers, and actuators, can theoretically achieve wide-band vibration suppression, and therefore has been widely studied and applied in precision equipment. Existing industrial systems generally employ proportional-integral-derivative (PID) linear control methods, although these are simple in structure and easy to implement.

[0004] While existing neural network-based or data-driven prediction methods can model complex dynamics to some extent, most only perform short-term predictions and cannot generate stable and controllable phase lead, nor do they have deep integration with real-time robust control strategies. Furthermore, in multi-degree-of-freedom vibration isolation systems, it is necessary to simultaneously address coupling relationships between multiple directions, estimation of unmeasurable states, and compensation for unknown nonlinear terms, and existing solutions often struggle to balance predictive capability, robustness, and real-time performance. Summary of the Invention

[0005] This invention addresses the problems in the prior art by providing an active vibration isolation method and system based on prediction and adaptive control. This solves the problem in the background art where, in multi-degree-of-freedom vibration isolation systems, it is necessary to simultaneously handle coupling relationships between multiple directions, estimation of unmeasurable states, and compensation for unknown nonlinear terms. Existing solutions often struggle to balance prediction capability, robustness, and real-time performance.

[0006] The technical solution adopted in this invention is as follows:

[0007] In a first aspect, this application provides an active vibration isolation method based on prediction and adaptive control, which includes the following steps:

[0008] Multi-degree-of-freedom dynamic modeling of the vibration isolation system is performed, with the load pose as the system output and the actuator control quantity as the system input. Unknown dynamics are uniformly represented as nonlinear terms, and the states in the system that cannot be directly measured are defined.

[0009] An adaptive neural network state observer is designed based on a multi-degree-of-freedom dynamic model. The neural network is used to approximate the nonlinear terms and obtain the estimation results of the unmeasurable state of the system.

[0010] A long short-term memory neural network prediction model is constructed, historical operating data of the vibration isolation system is collected, and the historical operating data is preprocessed to obtain the input timing sequence and output timing sequence.

[0011] The prediction model of the long short-term memory neural network is trained based on the input time series and the output time series, so that the prediction model can output a prediction signal to represent the future location information of the load.

[0012] Based on the state estimation results and prediction signals given by the state observer, an adaptive control strategy is adopted to construct a feedback control law, which compensates for nonlinear terms and parameter uncertainties, and obtains the feedback control quantity.

[0013] The feedback control quantity is superimposed with the feedforward control quantity obtained based on the prediction signal to generate a control command for driving the actuator of the vibration isolation system, thereby implementing active vibration isolation control on the load.

[0014] Furthermore, the steps for multi-degree-of-freedom dynamic modeling of the vibration isolation system include:

[0015] The position and attitude of the load in space are represented as the output variables of the multi-degree-of-freedom dynamic model, and the corresponding state dimension is determined according to the structural characteristics of the system.

[0016] The throttle valve opening, electromagnetic driving force, or air film adjustment amount are used as control inputs to the dynamic model to model and represent the force or torque exerted by the actuator on the load.

[0017] The unknown dynamics caused by changes in air film pressure, friction, changes in structural flexibility, and external disturbances in the vibration isolation system are uniformly grouped into nonlinear function terms and introduced into the model structure.

[0018] Define the velocity, attitude change rate or internal dynamic variables that cannot be directly obtained by sensors in the dynamic model, so that they together with the measurable output constitute the system state vector;

[0019] The multi-degree-of-freedom dynamic model is as follows:

[0020]

[0021]

[0022]

[0023] in, The system state, where For measurable load pose information; and These are control inputs and system outputs, respectively. For unknown smooth nonlinear functions; the multi-degree-of-freedom dynamic model has a total of One degree of freedom, by Each subsystem is represented separately.

[0024] Furthermore, the steps for designing an adaptive neural network state observer based on a multi-degree-of-freedom dynamic model include:

[0025] A state observation structure containing measurable output quantities and unmeasurable state quantities is constructed based on a multi-degree-of-freedom dynamic model;

[0026] Radial basis functions are selected as the basis function form of the neural network, and neural network input vectors for approximating nonlinear terms are constructed based on the system's input and measurable output quantities.

[0027] By setting the observer gain, the observer error, which is driven by both the measurable output and the neural network approximation result, can dynamically satisfy the convergence condition.

[0028] During the operation of the observer, the weights of the neural network are adaptively updated based on the observation error in order to obtain a real-time estimate of the system's state that cannot be directly measured.

[0029] Furthermore, when Time state Given that the data is unmeasurable, design a neural network state observer:

[0030]

[0031] in It is an RBF neural network, used to approximate unknown nonlinear functions in a system. ;

[0032] By selecting the observer gain Make Given the Herwitz matrix, the total state of the system is asymptotically estimated using an observer. .

[0033] Furthermore, a backstepping framework method is adopted to analyze the output tracking error. To begin, for Design virtual control law Define the second error ,for Design virtual control law ;

[0034] This process is repeated recursively until the final step of designing the actual control law. ;

[0035] At each step, the unknown nonlinear combination term RBF neural network Approximation, Approximation error There is an upper boundary;

[0036] The control law and the weights of the neural network and The adaptive law design is as follows:

[0037]

[0038]

[0039]

[0040]

[0041] in , , , , , All are normal numbers.

[0042] Furthermore, the feature is that the construction of the long short-term memory neural network prediction model includes:

[0043] right Each subsystem is designed with an adaptive neural network controller. ,design A predictive controller.

[0044] Furthermore, the design of predictive controllers includes:

[0045] For the i-th subsystem standardization: ;

[0046] By employing quadratic regression based on Huber loss and solving iteratively reweighted least squares, outliers are automatically assigned low weights, resulting in a smooth curve that reflects the core trend of the data while remaining unaffected by noise. ;

[0047] Construction of training data with a lead effect:

[0048] enter For a piece of historical data ;

[0049] Output It is about the future Points Apply a Gaussian window for weighted summation: ;

[0050] The center of the Gaussian window is set at the back of the window;

[0051] Multiple constructed using the above method Yes, train a deep LSTM network.

[0052] Furthermore, the trained deep LSTM network receives the latest location sequence and outputs filtered location predictions with a lead time. ;

[0053] The location prediction controller is defined as .

[0054] Furthermore, the feedback control quantity is superimposed with the feedforward control quantity obtained based on the predicted signal to generate control commands for driving the actuator of the vibration isolation system, including:

[0055] The first virtual control signal of the adaptive neural network controller Compensation signal with position prediction controller The commands are then integrated to form the final control instructions. .

[0056] Secondly, this application provides an active vibration isolation system based on prediction and adaptive control, used to implement the active vibration isolation method based on prediction and adaptive control as described in the first aspect. The system includes:

[0057] The load position sensor is configured to acquire the position and orientation information of the load in space and output a load position and orientation measurement signal.

[0058] Vibration isolation device, including an air spring assembly and an actuator connected to the air spring, the actuator including a high-speed throttle valve and a Lorentz motor, configured to adjust the air film pressure or apply electromagnetic force according to a control signal;

[0059] The multi-degree-of-freedom dynamic model building unit is configured to establish a multi-degree-of-freedom dynamic model based on the load pose measurement signal. The spatial position and attitude of the load are defined as the model output, the high-speed throttle valve opening and electromagnetic driving force are used as control inputs, and the unknown dynamics formed by changes in air film pressure, friction, changes in structural flexibility and external disturbances are represented as nonlinear terms. The in directly measurable velocity, attitude change rate and internal dynamic variables are defined to form the system state vector.

[0060] The neural network state observer is configured to construct a state observation structure based on a multi-degree-of-freedom dynamic model, approximate nonlinear terms using a radial basis function neural network, and update the neural network weights online according to the observer error to obtain an estimate of the unmeasurable state of the system.

[0061] The long short-term memory neural network prediction module is configured to collect historical operating data of the vibration isolation system, standardize and smooth the historical operating data to form input and output time series, train the long short-term memory neural network to generate a prediction model that can output the future position signal of the load, and receive the latest position sequence during operation to generate position prediction values ​​with lead amount in real time.

[0062] The adaptive backstepping control module is configured to construct multi-layer virtual control quantities based on the state estimation results of the state observer and the prediction signal output by the prediction module, and to generate feedback control quantities by using a neural network to approximate unknown nonlinear combination terms and updating weights through an adaptive law.

[0063] The control signal fusion module is configured to superimpose the feedback control quantity with the feedforward control quantity generated by the prediction signal to form a composite control command for driving the actuator.

[0064] The real-time control unit, including the real-time target machine and the control model execution environment, is configured to receive load pose measurement signals, prediction signals and state estimation results, running state observer, adaptive backstepping control module and prediction module, and output composite control commands to the actuator to achieve active vibration isolation control.

[0065] As can be seen from the above technical solutions, the advantages of the present invention are:

[0066] In the dynamic model, the velocity, attitude change rate and internal dynamic variables that cannot be directly measured are defined, and a neural network state observer is constructed based on the multi-degree-of-freedom model. The nonlinear terms are approximated by the radial basis function neural network and the network weights are updated online in an adaptive manner, so that the observer can converge to obtain a real-time estimate of the unmeasurable state of the system.

[0067] Factors that are difficult to model precisely, such as changes in film pressure, friction, flexibility fluctuations, and external disturbances, are uniformly represented as nonlinear terms. A neural network is used within the backstepping control framework to approximate these nonlinear terms, and adaptive laws adjust the neural network parameters in real time to compensate for the system's unknown dynamics and parameter uncertainties. Compared to traditional control methods that rely on precise models, this invention maintains the correctness and stability of the control law even with incomplete models, parameter variations, and the presence of external disturbances.

[0068] By collecting historical operating data, constructing input and output timing sequences, and training a long short-term memory neural network prediction model, a displacement prediction signal with a fixed phase lead is obtained. The prediction result is used as a feedforward signal in the control strategy, enabling the control system to anticipate upcoming displacement changes and reduce tracking errors caused by actuator lag, transmission delay, and system inertia. This achieves an organic combination of feedforward prediction and adaptive feedback, improving vibration isolation performance under multi-source dynamic conditions.

[0069] Based on the feedback control quantity obtained from backstepping control, a feedforward compensation quantity based on the predicted signal is introduced. Through the superposition of composite control signals, the control system can simultaneously handle predictable trend dynamics and unpredictable disturbance factors. Feedback control is used to suppress nonlinearity and uncertainty, while feedforward control is used to compensate for the predicted trend, thereby reducing response lag and achieving faster response and higher precision control for multi-degree-of-freedom vibration isolation systems.

[0070] By establishing a multi-degree-of-freedom dynamic model, designing an adaptive neural network controller and a predictive controller for each degree of freedom, and fusing the final control signals, this invention can simultaneously achieve state estimation, predictive compensation, and adaptive control under conditions where the load has multiple coupled motion directions, and is applicable to precision equipment scenarios such as six-degree-of-freedom vibration isolation platforms. Attached Figure Description

[0071] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the description will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0072] Figure 1 This is a flowchart illustrating the steps of the active vibration isolation method based on prediction and adaptive control in the embodiment.

[0073] Figure 2 This is an architecture diagram of the active vibration isolation system based on prediction and adaptive control in the embodiment. Detailed Implementation

[0074] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0075] Please see Figure 1As shown, this application provides an active vibration isolation method based on prediction and adaptive control, including:

[0076] Step S1: Perform multi-degree-of-freedom dynamic modeling on the vibration isolation system, take the load pose as the system output, take the actuator control quantity as the system input, represent the unknown dynamics as nonlinear terms, and define the states in the system that cannot be directly measured.

[0077] In practical applications, vibration isolation systems typically include a load platform supporting precision equipment, a base structure supporting air springs, position sensors, and components such as throttle valves or electromagnetic actuators. Since the system possesses degrees of freedom in three translational and three rotational directions after loading, obtaining multi-degree-of-freedom dynamic equations is essential. In actual modeling, position sensors are placed on the load platform to continuously output the platform's displacement or attitude changes in each direction; these measurements are directly considered as the model's system output. Regarding the actuators, the air film pressure is adjusted by controlling the opening of the throttle valve, or electromagnetic force is applied by adjusting the current of the electromagnetic actuator, allowing these controllable quantities to naturally constitute the model's system input. Because air film stiffness fluctuates with pressure, the coefficient of friction changes with velocity, and structural flexibility is affected by environmental temperature and load variations, these dynamics, which cannot be precisely described, are abstracted into unified nonlinear terms. In numerical modeling, it is also necessary to define internal state quantities that cannot be directly measured, such as velocity and attitude change rate, and combine them with the pose quantities output by the sensors to form the system's state vector, so that the subsequent state observer can fully reconstruct the system state. By using the above methods, a multi-degree-of-freedom dynamic model suitable for subsequent control design can be obtained, so that the true characteristics of the system can be fully reflected in the modeling framework.

[0078] Step S2: Design an adaptive neural network state observer based on a multi-degree-of-freedom dynamic model, and use the neural network to approximate the nonlinear terms to obtain the estimation results of the unmeasurable state of the system.

[0079] In practical implementation, due to the existence of multiple unmeasurable state variables in the vibration isolation system, such as the platform's velocity in various directions, rotational angular velocity, or some internal coupled dynamics, a state observer is needed for estimation. Based on the dynamic model established in step S1, an observer structure can be constructed, using the pose measured by sensors as input. On this basis, a neural network is introduced as an approximator to approximate the unknown nonlinear terms in the model online. In engineering implementation, a radial basis function network (RBF) with low training difficulty and good convergence is typically used. The system's input, measurable output, and some estimated quantities are used as the neural network's input vector, and the network output compensates for deviations caused by nonlinearity. During operation, the observer continuously compares the difference between the predicted output and the actual measurement, and adjusts the network weights accordingly, so that the estimated value gradually approaches the true state. In this way, relatively accurate estimates of the system's velocity, attitude change rate, and other unmeasurable states can be obtained during continuous operation, providing reliable information support for the subsequent construction of the control law.

[0080] Step S3: Construct a long short-term memory neural network prediction model, collect historical operating data of the vibration isolation system, preprocess the historical operating data, and obtain the input timing sequence and output timing sequence.

[0081] During the actual operation of a vibration isolation platform, a large amount of time-series data from position sensors is continuously recorded. This data includes displacement changes in different directions and dynamic behaviors caused by environmental factors, load variations, and actuator responses. To build a prediction model, this raw data first needs to be organized, stored chronologically, and outliers and incomplete data segments identified. To improve the usability of the training data, the data from different channels needs to be standardized to ensure a uniform scale, facilitating neural network learning. Based on this, the standardized data is smoothed to filter out high-frequency noise, making the data sequence closer to the platform's actual motion trend. Subsequently, according to a set time window, a continuous historical data segment is extracted as the input time series, and the corresponding output time series is determined based on the time period following the window. These constitute the data pairs required for training the prediction model. In practical engineering, by combining data collected under multiple different operating conditions, a dataset covering a wider range of operations can be obtained, thereby improving the prediction model's generalization ability under various disturbance conditions.

[0082] Step S4: Train the long short-term memory neural network prediction model based on the input time series and output time series, so that the prediction model can output a prediction signal to represent the future location information of the load.

[0083] After obtaining the input and output time series, these are used as training samples to input into a Long Short-Term Memory (LSTM) neural network. This network structure can utilize historical information from a relatively distant time series, enabling it to perform well when dealing with objects like vibration isolation systems that possess dynamic memory characteristics. During training, data segments of different lengths are fed into the network. The network establishes an internal state representation of the historical sequence and updates its internal parameters based on the corresponding output time series, gradually enabling the model to extract trends from historical data. After training, the model can receive the latest position sequence during system operation and output predicted position values ​​for the near future. In engineering applications, when the vibration isolation platform is subjected to periodic or gradual disturbances, this predicted signal can reflect potential future displacement trends in advance, providing a basis for subsequent feedforward compensation and making the overall system response faster and more stable.

[0084] Step S5: Based on the state estimation results and prediction signals given by the state observer, construct a feedback control law using an adaptive control strategy to compensate for nonlinear terms and parameter uncertainties, and obtain the feedback control quantity.

[0085] In practical control implementation, information such as velocity and attitude change rate provided by the state observer, along with the position prediction value generated by the prediction model, jointly participate in the construction of the control law. The adaptive control strategy enables the system to maintain a stable operating state even when facing unknown parameters, model changes, and external disturbances. In the specific design, an output tracking error is set to reflect the difference between the platform's current position and the target position, and the evolution relationship of the error is constructed in conjunction with the internal state information provided by the observer. Subsequently, through multi-level control variable design, each level of control variable can gradually offset the deviation caused by nonlinear terms. The neural network continuously approximates the unknown nonlinear behavior, enabling the control law to be updated in real time to adapt to the continuous changes in the system's operating state. This approach exhibits significant robustness when dealing with complex vibration isolation platforms, maintaining the continuity and stability of output behavior even under load changes or when the platform is subjected to additional disturbances.

[0086] Step S6: Superimpose the feedback control quantity with the feedforward control quantity obtained based on the prediction signal to generate a control command for driving the actuator of the vibration isolation system, and implement active vibration isolation control on the load.

[0087] During the actual operation of the vibration isolation platform, the feedback control quantity can correct the system's response to disturbances in real time, while the feedforward control quantity generated by the predictive signal can make adjustments in advance for impending changes. The control system runs on the real-time target machine. By fusing the two types of control quantities, the generated final control command has both the ability to respond to the current state and the ability to compensate for future trends in advance. In actual scenarios, when the platform is affected by low-frequency disturbances or periodic external forces, the feedforward control quantity can adjust the force of the actuator in advance, so that the platform enters a better stress state before the disturbance arrives; at the same time, the feedback control quantity corrects instantaneous errors in a timely manner, ensuring the stability of the system in all directions. The final generated control command is sent to the throttle valve or electromagnetic actuator, causing it to output precise air film pressure or torque, thereby realizing active vibration isolation control of the load platform and ensuring that the precision equipment remains stable under various operating conditions.

[0088] In some embodiments, the steps of performing multi-degree-of-freedom dynamic modeling of a vibration isolation system include:

[0089] The position and attitude of the load in space are represented as the output variables of the multi-degree-of-freedom dynamic model, and the corresponding state dimension is determined according to the structural characteristics of the system.

[0090] The throttle valve opening, electromagnetic driving force, or air film adjustment amount are used as control inputs to the dynamic model to model and represent the force or torque exerted by the actuator on the load.

[0091] The unknown dynamics caused by changes in air film pressure, friction, changes in structural flexibility, and external disturbances in the vibration isolation system are uniformly grouped into nonlinear function terms and introduced into the model structure.

[0092] Define the velocity, attitude change rate or internal dynamic variables that cannot be directly obtained by sensors in the dynamic model, so that they together with the measurable output constitute the system state vector;

[0093] The multi-degree-of-freedom dynamic model is as follows:

[0094]

[0095]

[0096]

[0097] in, The system state, where For measurable load pose information; and These are control inputs and system outputs, respectively. For unknown smooth nonlinear functions; the multi-degree-of-freedom dynamic model has a total of One degree of freedom, by Each subsystem is represented separately.

[0098] In practical vibration isolation platforms, the load typically consists of precision instruments, optical components, or test units, supported by air springs on the platform and moving through multiple degrees of freedom. To accurately characterize the dynamic behavior of the load, position sensors are installed to acquire real-time displacement signals of the platform in various directions; these signals are directly considered as model outputs. In the actuator section, a high-speed throttle valve regulates the airflow entering the film chamber, thus affecting the supporting force of the air springs, while electromagnetic actuators can apply additional force or torque to the load; therefore, these adjustable quantities naturally constitute the model's control inputs. During system modeling, friction, film pressure fluctuations, and structural flexibility changes are difficult to quantify directly, so they are incorporated into a uniformly defined nonlinear function term, making the model structure more compact and scalable. Since only some states are measurable, the platform's velocity, angular velocity, and internal flexible response are defined and added to the state vector, estimated by the state observer in subsequent steps. The model constructed in this way can comprehensively reflect the force and response relationships of the platform across multiple degrees of freedom, enabling the control system to perform subsequent calculations and decisions based on the model.

[0099] In some embodiments, the steps of designing an adaptive neural network state observer based on a multi-degree-of-freedom dynamics model include:

[0100] A state observation structure containing measurable output quantities and unmeasurable state quantities is constructed based on a multi-degree-of-freedom dynamic model;

[0101] Radial basis functions are selected as the basis function form of the neural network, and neural network input vectors for approximating nonlinear terms are constructed based on the system's input and measurable output quantities.

[0102] By setting the observer gain, the observer error, which is driven by both the measurable output and the neural network approximation result, can dynamically satisfy the convergence condition.

[0103] During the operation of the observer, the weights of the neural network are adaptively updated based on the observation error in order to obtain a real-time estimate of the system's state that cannot be directly measured.

[0104] During system operation, position sensors alone cannot directly obtain the platform's velocity, angular velocity, and other implicit dynamic variables; therefore, a state observer is needed to estimate these quantities. Based on the dynamic model, the observer takes known measurable pose quantities as input and control quantities from the actuators as auxiliary inputs, generating state estimates through an internal predictive structure. To handle nonlinear factors in the model, the observer introduces a radial basis function neural network as a real-time approximator for the nonlinear terms, ensuring good performance even when the model is inaccurate. During operation, the observer continuously compares the deviation between the estimated output and the actual measurement, automatically adjusting the neural network weights based on changes in deviation to gradually approximate the actual nonlinear behavior. Through reasonable design of the observer gain, the overall observation error can gradually converge, enabling the system to obtain continuous and stable internal state estimation results in multi-degree-of-freedom conditions, ensuring reliable and consistent parameter inputs for subsequent control strategies.

[0105] In some embodiments, when Time state Given that the data is unmeasurable, design a neural network state observer:

[0106]

[0107] in It is an RBF neural network, used to approximate unknown nonlinear functions in a system. ;

[0108] By selecting the observer gain Make Given the Herwitz matrix, the total state of the system is asymptotically estimated using an observer. .

[0109] In multi-degree-of-freedom structures of vibration isolation systems, apart from pose signals, other states are often difficult to measure directly by sensors, thus requiring estimation through an observer. In engineering applications, by introducing the measurable output into the observer's main structure, it can use measurable information to correct unmeasurable states in real time. The neural network's input typically includes estimated state variables and actually measured pose variables; its output is used to compensate for unknown nonlinear terms, making the observer output closer to the dynamic behavior of the real system. In actual operation, by adjusting the observer gain, it can converge quickly, ensuring that the observer does not diverge or oscillate under disturbances. As the running time increases, the neural network gradually learns and approximates the true nonlinear characteristics of the system, enabling the observer's state output to provide a reliable reference for subsequent control law construction.

[0110] In some embodiments, a backstepping framework method is employed to track the output error. To begin, for Design virtual control law Define the second error ,for Design virtual control law ;

[0111] This process is repeated recursively until the final step of designing the actual control law. ;

[0112] At each step, the unknown nonlinear combination term RBF neural network Approximation, Approximation error There is an upper boundary;

[0113] The control law and the weights of the neural network and The adaptive law design is as follows:

[0114]

[0115]

[0116]

[0117]

[0118] in , , , , , All are normal numbers.

[0119] In some embodiments, the method of constructing a long short-term memory neural network prediction model includes:

[0120] In practical control implementation, the backstepping method can construct control quantities in layers in a multi-degree-of-freedom nonlinear system, with each layer gradually constructed after the error of the previous layer has stabilized. In practical applications, the system first establishes the output tracking error based on the difference between the currently acquired load displacement signal and the desired position. Then, this error is transmitted to the internal state of the system through a virtual control law, stabilizing it layer by layer. As the error of each layer gradually decreases, the actual control quantity used for the actuator is finally obtained. Due to the presence of a large number of unknown or difficult-to-model nonlinear components in the system, the approximation of nonlinear combination terms by the neural network in each layer ensures that the controller maintains correctness under various operating conditions. By designing an adaptive law, the parameters in the controller can be automatically adjusted during operation, enabling the feedback control to adapt to changes in platform load, environmental disturbances, and structural uncertainties, thereby achieving a stable and reliable vibration isolation effect.

[0121] right Each subsystem is designed with an adaptive neural network controller. ,design A predictive controller.

[0122] In practical engineering, multi-degree-of-freedom vibration isolation systems may exhibit coupling between different directions. Therefore, when constructing a predictive model, it is necessary to establish a separate predictive structure for each degree of freedom to capture the dynamic trends in that direction. Before training the predictive model, a corresponding historical dataset is typically established for each direction so that the network can learn the dynamic patterns in that specific direction. By designing a separate predictive controller for each degree of freedom, the system can acquire the ability to predict future trends in different directions, making the overall control structure more flexible and scalable.

[0123] In some embodiments, the design of the predictive controller includes:

[0124] For the i-th subsystem standardization: ;

[0125] By employing quadratic regression based on Huber loss and solving iteratively reweighted least squares, outliers are automatically assigned low weights, resulting in a smooth curve that reflects the core trend of the data while remaining unaffected by noise. ;

[0126] Construction of training data with a lead effect:

[0127] enter For a piece of historical data ;

[0128] Output It is about the future Points Apply a Gaussian window for weighted summation: ;

[0129] The center of the Gaussian window is set at the back of the window;

[0130] Multiple constructed using the above method Yes, train a deep LSTM network.

[0131] In real-world scenarios, data collected by position sensors often contains noise spikes or is affected by external disturbances. Therefore, after standardizing the data, robust regression processing is required to obtain smooth sequences, thereby avoiding the impact of noise on model construction during training. In constructing the training samples, historical windows of different lengths are selected as inputs, and position sequences over a future period are aggregated into prediction targets through weighted windows, enabling the network to learn a mapping relationship that combines trend and anticipation. After training with a large amount of data, the LSTM network can capture the time-dependent characteristics of the vibration isolation platform under different operating states, allowing the predictive controller to generate future position trends in real time during system operation, providing accurate reference signals for subsequent feedforward control.

[0132] In some embodiments, the trained deep LSTM network receives the latest location sequence and outputs filtered location predictions with a lead. ;

[0133] The location prediction controller is defined as .

[0134] The trained predictive model is deployed in the real-time control unit. During operation, the system continuously provides the network with the latest position data. The network analyzes these inputs and outputs a pose prediction for the future timeframe. This prediction typically appears as a smooth signal with a fixed phase lead, allowing the controller to respond in advance. During control execution, the predictive controller compares the predicted value with the desired position and generates control variables to compensate for future error trends, enabling the system to approach the target position more quickly and reducing control lag caused by actuator delays or system inertia.

[0135] In some embodiments, superimposing the feedback control quantity with the feedforward control quantity obtained based on the predicted signal to generate control commands for driving the actuator of the vibration isolation system includes:

[0136] The first virtual control signal of the adaptive neural network controller Compensation signal with position prediction controller The commands are then integrated to form the final control instructions. .

[0137] In practical control systems, feedback control provides real-time suppression of disturbances and internal uncertainties, while feedforward control can compensate for trend changes in advance. By fusing the two, a control signal is ultimately formed that can both respond to the current state and predict future changes. This fused control quantity is sent to the actuator, such as a high-speed throttle valve or electromagnetic actuator, in real time, causing it to adjust its output according to the fused control value, thereby ensuring the vibration isolation platform remains stable under multiple directions and operating conditions. This method is particularly effective when facing periodic disturbances or rapidly changing load conditions, keeping the platform always within a controlled and stable operating range.

[0138] In one embodiment, this application employs the following steps:

[0139] Step 1: System Construction and Initialization;

[0140] Mechanical platform construction, the platform includes:

[0141] Base frame: weighing approximately 10,000 kg, fixed to the laboratory foundation with anchor bolts.

[0142] Load: As a mounting platform for precision equipment, its mass is approximately 1000 kg.

[0143] Vibration isolators: Three high-performance air bearing vibration isolators are arranged in an equilateral triangle configuration between the base frame and the fretting frame, with a side length of [missing information]. Distance from the centroid to each vertex Each vibration isolator integrates:

[0144] Capacitive position sensor: measures the displacement of the micro-motion frame relative to the vibration isolator body.

[0145] Lorentz motor: As an auxiliary actuator, it can generate precise electromagnetic force.

[0146] High-speed throttle valve: As the main actuator, it regulates the airflow supplied to the air bearing, thereby changing the air film pressure and stiffness.

[0147] Control system integration:

[0148] Real-time target unit: Receives sensor signals in real time and transmits control signals to the throttle valve and Lorentz motor.

[0149] System modeling and parameter identification:

[0150] Based on the mechanical drawings and materials, the mass matrix and moment of inertia matrix of the micro-motion frame were initially determined.

[0151] Frequency response function tests were conducted using hammer impact or a vibrator to identify the initial values ​​of stiffness and damping of each vibration isolator in six directions, as well as the equivalent stiffness and damping of the base frame.

[0152] All these parameters are filled into the system dynamics equations, and a simulation model of the system is built in Simulink for preliminary verification of the subsequent control algorithm.

[0153] Step 2: Implementation of the LSTM position prediction controller:

[0154] Data Acquisition and Preprocessing:

[0155] Switch the platform to a simple PID control mode and allow it to run stably. Apply a wideband white noise disturbance signal to excite all modes of the system. Continuously acquire position sensor data for all degrees of freedom for at least 1 hour, with the sampling frequency set to 1kHz.

[0156] Standardization: Calculate the mean value for the data from each channel separately. and standard deviation and conduct Process. Save these. and Used for online destandardization.

[0157] Robust Quadratic Regression: Calls the robust quadratic regression algorithm. Sets the threshold. (Adjusted according to noise level), convergence threshold For standardized data Perform fitting to obtain smoothed data. .

[0158] Training data construction and network training:

[0159] Setting parameters: Window length (i.e., 0.1s history), prediction layer (i.e., 0.02s in the future), Gaussian window length Standard deviation (The control window is relatively wide, the smoothing effect is strong, and the advance amount is moderate).

[0160] Network input / output generation: Traverse the entire dataset and generate input vectors for each time point. (of length l) (sequence) and output scalar (future indivual (Weighted sum of points and Gaussian window). Use 80% of the data as the training set and 20% as the test set.

[0161] Network Training: The `trainNetwork` function is used in MATLAB's Deep Learning Toolbox. The network layers are configured as follows: input layer, one to multiple LSTM layers (each containing 128 or 256 hidden units to capture temporal dependencies), Dropout layers (dropout rate 0.2-0.5 to prevent overfitting), fully connected layers (integrating features), and a regression output layer (outputting predicted values). Training options: optimizer is adam, maximum number of epochs is 200, initial learning rate is 0.005, and MiniBatchSize is 64. The test set loss is monitored during training to prevent overfitting. After training, the network model is saved.

[0162] Online deployment:

[0163] Build a real-time control model in Simulink. Perform the same standardization and smoothing on the new data as on the offline data (using the mean and variance calculated online, or using the offline values); input the smoothed data points into the `predict` function to call the loaded network model; destandardize the output to obtain the predicted location. .

[0164] Calculate the feedforward control quantity: ,in It is a set value (usually 0) that controls the gain. The initial value is set to 1000, and can be fine-tuned later.

[0165] Step 3: Implementation of the adaptive neural network controller:

[0166] State observer implementation:

[0167] Implement the state observer differential equation using MATLAB Function Blocks in Simulink.

[0168] Initializing the RBF neural network: Center Uniformly distributed in the input space, width Neural network weights , Initialize to a zero vector or a small random number.

[0169] Observer Gain Setting it to 300 makes the observer poles much faster than the system dynamics, ensuring fast convergence.

[0170] Adaptive backstepping controller implementation:

[0171] In Simulink, the computational modules for the virtual and actual control laws are constructed step-by-step according to the backstepping method design process. Simultaneously, the neural network weights are updated online based on the adaptive law. and parameters .

[0172] Step 4: System Integration and Performance Verification

[0173] Control Fusion: In the Simulink model, the final control force is set as... This control force is distributed to the throttle valve (main) and the Lorentz motor (auxiliary) as needed.

[0174] Code generation and deployment: Use Simulink to link the entire control model to the real-time target machine for execution.

[0175] Performance testing:

[0176] Tracking performance: Given a displacement setpoint, perform a step or sinusoidal change and observe the response.

[0177] Error analysis: Under steady-state conditions, record the control error.

[0178] Vibration isolation performance: A sweeping vibration is applied to the base frame using an additional exciter, while the vibrations of the base frame and the micro-motion frame are measured using a high-precision accelerometer. Transmissibility is calculated. .

[0179] Robustness test: Add / remove weights from the micro-motion frame (simulating load changes) and observe the system response. Adaptive controllers can quickly adjust parameters and restore performance, while traditional PID controllers will exhibit steady-state errors or performance degradation.

[0180] like Figure 2 As shown, in some embodiments, this application provides an active vibration isolation system based on prediction and adaptive control to implement an active vibration isolation method based on prediction and adaptive control. The system includes:

[0181] The load position sensor is configured to acquire the position and orientation information of the load in space and output a load position and orientation measurement signal.

[0182] Vibration isolation device, including an air spring assembly and an actuator connected to the air spring, the actuator including a high-speed throttle valve and a Lorentz motor, configured to adjust the air film pressure or apply electromagnetic force according to a control signal;

[0183] The multi-degree-of-freedom dynamic model building unit is configured to establish a multi-degree-of-freedom dynamic model based on the load pose measurement signal. The spatial position and attitude of the load are defined as the model output, the high-speed throttle valve opening and electromagnetic driving force are used as control inputs, and the unknown dynamics formed by changes in air film pressure, friction, changes in structural flexibility and external disturbances are represented as nonlinear terms. The in directly measurable velocity, attitude change rate and internal dynamic variables are defined to form the system state vector.

[0184] The neural network state observer is configured to construct a state observation structure based on a multi-degree-of-freedom dynamic model, approximate nonlinear terms using a radial basis function neural network, and update the neural network weights online according to the observer error to obtain an estimate of the unmeasurable state of the system.

[0185] The long short-term memory neural network prediction module is configured to collect historical operating data of the vibration isolation system, standardize and smooth the historical operating data to form input and output time series, train the long short-term memory neural network to generate a prediction model that can output the future position signal of the load, and receive the latest position sequence during operation to generate position prediction values ​​with lead amount in real time.

[0186] The adaptive backstepping control module is configured to construct multi-layer virtual control quantities based on the state estimation results of the state observer and the prediction signal output by the prediction module, and to generate feedback control quantities by using a neural network to approximate unknown nonlinear combination terms and updating weights through an adaptive law.

[0187] The control signal fusion module is configured to superimpose the feedback control quantity with the feedforward control quantity generated by the prediction signal to form a composite control command for driving the actuator.

[0188] The real-time control unit, including the real-time target machine and the control model execution environment, is configured to receive load pose measurement signals, prediction signals and state estimation results, running state observer, adaptive backstepping control module and prediction module, and output composite control commands to the actuator to achieve active vibration isolation control.

[0189] During the actual operation of the system, capacitive position sensors mounted on the vibration isolation platform continuously collect displacement information of the load in various directions. The pose signals output by the sensors are transmitted to the real-time control platform as real-time measurements. The real-time control platform consists of a target machine and a control algorithm execution environment. It can read sensor data at a fixed sampling frequency and use it as the basic input for subsequent state estimation and control calculations.

[0190] After receiving the displacement measurement signal, the real-time control platform first uses a neural network state observer to estimate the unmeasurable internal state of the system. The observer combines the dynamic model, the control input of the actuator, and the currently sampled pose signal, using its internal radial basis function neural network to approximate the unknown nonlinearity. It continuously updates the network parameters using observation errors, enabling real-time reconstruction of the platform's internal states such as velocity and angular velocity. After processing, the complete state estimation result is sent to the adaptive backstepping control module.

[0191] The system simultaneously runs a Long Short-Term Memory (LSTM) neural network prediction module. This module acquires the latest historical displacement sequence from the real-time control platform, preprocesses it, and then inputs it into the prediction network. Based on the learned temporal characteristics, the prediction network outputs a predicted position value for a future period. This predicted signal, after filtering, reflects the system's future dynamic trend. The real-time control platform uses this prediction information as input to the control feedforward section, enabling the system to respond in advance to potential displacement changes.

[0192] The backstepping control module constructs error variables based on the state estimation results provided by the observer and generates intermediate virtual control quantities through a layer-by-layer recursive approach. In each control layer, the neural network's approximation of the unknown nonlinear term is used to compensate for model uncertainties, enabling the module to output stable feedback control quantities. Subsequently, the system fuses the feedforward compensation quantity generated by the predictive control module with the feedback control quantity generated by the backstepping control to obtain the final composite control command used to drive the actuator.

[0193] The composite control commands are sent in real time to the actuators of the vibration isolation platform, including a high-speed throttle valve and a Lorentz force actuator. The high-speed throttle valve adjusts the air film pressure in the air spring according to the control signal, changing the air film's supporting force; the Lorentz force actuator applies additional force or torque according to the control signal, enabling the vibration isolation platform to respond to the control input in multiple degrees of freedom. The force generated by the actuator is directly applied to the load, causing the load's motion state to converge towards the target position, thereby achieving active vibration isolation.

[0194] During system operation, the real-time control platform continuously monitors the displacement signals output by the sensors and compares them with the desired position to determine whether the vibration isolation control effect meets the requirements. The system can also update the predictive model and neural network approximator based on long-term operating data, ensuring the control system maintains stable performance under different operating conditions. This entire process is repeated continuously in real-time loops, enabling the vibration isolation platform to maintain the stable operation of precision equipment under external disturbances, load changes, and structural nonlinearities.

[0195] The above description is merely a preferred embodiment of one or more embodiments of this specification and is not intended to limit the scope of one or more embodiments of this specification. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of one or more embodiments of this specification should be included within the protection scope of one or more embodiments of this specification.

Claims

1. An active vibration isolation method based on prediction and adaptive control, characterized in that, Includes the following steps: Multi-degree-of-freedom dynamic modeling of the vibration isolation system is performed, with the load pose as the system output and the actuator control quantity as the system input. Unknown dynamics are uniformly represented as nonlinear terms, and the states in the system that cannot be directly measured are defined. An adaptive neural network state observer is designed based on a multi-degree-of-freedom dynamic model. The neural network is used to approximate the nonlinear terms and obtain the estimation results of the unmeasurable state of the system. A long short-term memory neural network prediction model is constructed, historical operating data of the vibration isolation system is collected, and the historical operating data is preprocessed to obtain the input timing sequence and output timing sequence. The prediction model of the long short-term memory neural network is trained based on the input time series and the output time series, so that the prediction model can output a prediction signal to represent the future location information of the load. Based on the state estimation results and prediction signals given by the state observer, an adaptive control strategy is adopted to construct a feedback control law, which compensates for nonlinear terms and parameter uncertainties, and obtains the feedback control quantity. The feedback control quantity is superimposed with the feedforward control quantity obtained based on the prediction signal to generate a control command for driving the actuator of the vibration isolation system, thereby implementing active vibration isolation control on the load. Using a backstepping framework, the tracking error is analyzed from the output. To begin, for Design virtual control law Define the second error ,for Design virtual control law ; This process is repeated recursively until the final step of designing the actual control law. ; At each step, the unknown nonlinear combination term RBF neural network Approximation, Approximation error There is an upper boundary; The control law and the weights of the neural network and The adaptive law design is as follows: in , , , , , All are normal numbers.

2. The active vibration isolation method based on prediction and adaptive control according to claim 1, characterized in that, The steps for performing multi-degree-of-freedom dynamic modeling of a vibration isolation system include: The position and attitude of the load in space are represented as the output variables of the multi-degree-of-freedom dynamic model, and the corresponding state dimension is determined according to the structural characteristics of the system. The throttle valve opening, electromagnetic driving force, or air film adjustment amount are used as control inputs to the dynamic model to model and represent the force or torque exerted by the actuator on the load. The unknown dynamics caused by changes in air film pressure, friction, changes in structural flexibility, and external disturbances in the vibration isolation system are uniformly grouped into nonlinear function terms and introduced into the model structure. Define the velocity, attitude change rate or internal dynamic variables that cannot be directly obtained by sensors in the dynamic model, so that they together with the measurable output constitute the system state vector; The multi-degree-of-freedom dynamic model is as follows: in, The system state, where For measurable load pose information; and These are control inputs and system outputs, respectively. For unknown smooth nonlinear functions; the multi-degree-of-freedom dynamic model has a total of One degree of freedom, by Each subsystem is represented separately.

3. The active vibration isolation method based on prediction and adaptive control according to claim 2, characterized in that, The steps involved in designing an adaptive neural network state observer based on a multi-degree-of-freedom dynamics model include: A state observation structure containing measurable output quantities and unmeasurable state quantities is constructed based on a multi-degree-of-freedom dynamic model; Radial basis functions are selected as the basis function form of the neural network, and neural network input vectors for approximating nonlinear terms are constructed based on the system's input and measurable output quantities. By setting the observer gain, the observer error, which is driven by both the measurable output and the neural network approximation result, can dynamically satisfy the convergence condition. During the operation of the observer, the weights of the neural network are adaptively updated based on the observation error in order to obtain a real-time estimate of the system's state that cannot be directly measured.

4. The active vibration isolation method based on prediction and adaptive control according to claim 3, characterized in that, when Time state Given that the data is unmeasurable, design a neural network state observer: in It is an RBF neural network, used to approximate unknown nonlinear functions in a system. ; By selecting the observer gain Make Given the Herwitz matrix, the total state of the system is asymptotically estimated using an observer. .

5. The active vibration isolation method based on prediction and adaptive control according to any one of claims 1-4, characterized in that, The construction of a long short-term memory neural network prediction model includes: right Each subsystem is designed with an adaptive neural network controller. ,design A predictive controller.

6. The active vibration isolation method based on prediction and adaptive control according to claim 5, characterized in that, When designing a predictive controller, the following should be included: For the i-th subsystem standardization: ; By employing quadratic regression based on Huber loss and solving iteratively reweighted least squares, outliers are automatically assigned low weights, resulting in a smooth curve that reflects the core trend of the data while remaining unaffected by noise. ; Construction of training data with a lead effect: enter For a piece of historical data ; Output It is about the future Points Apply a Gaussian window for weighted summation: ; The center of the Gaussian window is set at the back of the window; Multiple constructed using the above method Yes, train a deep LSTM network.

7. The active vibration isolation method based on prediction and adaptive control according to claim 6, characterized in that, The trained deep LSTM network receives the latest location sequence and outputs filtered location predictions with a lead time. ; The location prediction controller is defined as .

8. The active vibration isolation method based on prediction and adaptive control according to claim 7, characterized in that, The feedback control quantity is superimposed with the feedforward control quantity obtained based on the predicted signal to generate control commands for driving the actuator of the vibration isolation system, including: The first virtual control signal of the adaptive neural network controller Compensation signal with position prediction controller The commands are then integrated to form the final control instructions. .

9. An active vibration isolation system based on predictive and adaptive control, used to implement the active vibration isolation method based on predictive and adaptive control as described in claim 1, characterized in that, The system includes: The load position sensor is configured to acquire the position and orientation information of the load in space and output a load position and orientation measurement signal. Vibration isolation device, including an air spring assembly and an actuator connected to the air spring, the actuator including a high-speed throttle valve and a Lorentz motor, configured to adjust the air film pressure or apply electromagnetic force according to a control signal; The multi-degree-of-freedom dynamic model building unit is configured to establish a multi-degree-of-freedom dynamic model based on the load pose measurement signal. The spatial position and attitude of the load are defined as the model output, the high-speed throttle valve opening and electromagnetic driving force are used as control inputs, and the unknown dynamics formed by changes in air film pressure, friction, changes in structural flexibility and external disturbances are represented as nonlinear terms. The in directly measurable velocity, attitude change rate and internal dynamic variables are defined to form the system state vector. The neural network state observer is configured to construct a state observation structure based on a multi-degree-of-freedom dynamic model, approximate nonlinear terms using a radial basis function neural network, and update the neural network weights online according to the observer error to obtain an estimate of the unmeasurable state of the system. The long short-term memory neural network prediction module is configured to collect historical operating data of the vibration isolation system, standardize and smooth the historical operating data to form input and output time series, train the long short-term memory neural network to generate a prediction model that can output the future position signal of the load, and receive the latest position sequence during operation to generate position prediction values ​​with lead amount in real time. The adaptive backstepping control module is configured to construct multi-layer virtual control quantities based on the state estimation results of the state observer and the prediction signal output by the prediction module, and to generate feedback control quantities by using a neural network to approximate unknown nonlinear combination terms and updating weights through an adaptive law. The control signal fusion module is configured to superimpose the feedback control quantity with the feedforward control quantity generated by the prediction signal to form a composite control command for driving the actuator. The real-time control unit, including the real-time target machine and the control model execution environment, is configured to receive load pose measurement signals, prediction signals and state estimation results, running state observer, adaptive backstepping control module and prediction module, and output composite control commands to the actuator to achieve active vibration isolation control.

Images