A method and system for adaptive active flutter suppression of an airfoil

Abstract

The application discloses a kind of wing flutter risk self-adaptive active inhibition method and system.First, the wing flutter system training set is established, and the leading state transition relationship is learned using reserve pool calculation;Further, condition diffusion residual model is constructed to depict one-step prediction residual condition distribution, realize residual mean compensation and uncertainty scenario generation;Then, based on K nearest neighbor local residual covariance, the prediction reliability is quantified, and the normalized risk index is obtained using the quantification of the verification set quantification;Finally, the risk index is embedded into the stochastic model predictive control, and the control weight, the prediction scene number and the state constraint margin are adaptively adjusted, and the optimal control is solved and rolled out.The application overcomes the shortcomings of traditional methods, such as high dependence on fine models and poor adaptability of fixed control parameters, and improves the system robustness and convergence speed under parameter mismatch and high disturbance.

Description

Technical Field

[0001] This invention belongs to the fields of active control of aerospace structures, aeroelastic flutter suppression, data-driven predictive control, and stochastic model predictive control, and in particular, an adaptive active suppression method and system for wing flutter risk. Background Technology

[0002] As modern aircraft continue to evolve towards lighter weight, higher flexibility, longer endurance, and greater maneuverability, the flexibility of wing structures is constantly improving. The coupling effect between aerodynamic forces, inertial forces, and structural elastic forces is becoming increasingly significant, making wing flutter a key factor affecting flight safety and performance. Wing flutter is a typical aeroelastic instability phenomenon. When the flight speed reaches the critical flutter velocity, the amplitude of structural vibration increases rapidly, potentially leading to wing structural fatigue failure and even aircraft failure in severe cases. Therefore, how to actively suppress wing flutter in complex flight environments to improve aircraft operational safety, control stability, and environmental adaptability remains an important research direction in the field of aerospace aeroelastic control.

[0003] In recent years, Model Predictive Control (MPC) has been widely applied in the field of active flutter control for airfoils due to its ability to explicitly handle problems such as multivariable coupling, input constraints, and state constraints. For example, Darabseh et al. proposed a two-dimensional active flutter suppression method for airfoils based on discrete model predictive control. By establishing an aeroelastic state-space model of the airfoil and combining Kalman filtering and roll optimization strategies, active flutter control can be achieved, achieving good flutter suppression results under the condition of satisfying input constraints. However, this type of method still relies on a relatively accurate aeroelastic dynamic model. When model parameters change, model mismatch occurs, or flight conditions change, the model prediction accuracy is prone to decrease, thus affecting control performance and system stability. In addition, complex airfoil systems usually have strong nonlinear, time-varying, and random disturbance characteristics, making it difficult for traditional model predictive control to balance control accuracy and environmental adaptability.

[0004] To reduce the dependence of control systems on precise mathematical models, data-driven modeling methods have been gradually applied to the fields of wing aeroelastic prediction and active control in recent years. Lum et al. proposed a data-driven modeling method for wing flutter control, which establishes a reduced-order aeroelastic model through system identification and combines it with a robust control strategy to achieve active flutter control, verifying the application potential of data-driven modeling in complex aeroelastic systems. However, this type of method still relies on the system identification process, and its predictive ability is constrained by the quality of training data and model structure. When the system operating conditions change or enter a region with insufficient coverage of training samples, the prediction error tends to increase significantly, and its generalization ability for complex nonlinear and time-varying conditions remains limited.

[0005] With the development of artificial intelligence technology, data-driven prediction methods such as neural networks, deep learning, and Reservoir Computing (RC) have been gradually applied to the modeling of complex nonlinear dynamic systems. Reservoir Computing possesses advantages such as fast training speed, simple parameter optimization, and strong time-series modeling capabilities, showing promising application prospects in the field of complex dynamic system prediction. However, most existing data-driven prediction methods use minimum mean square error as the optimization objective, only outputting deterministic predictions of future states, making it difficult to effectively characterize the prediction error and the probability distribution characteristics brought about by random disturbances in the system. When the system enters unknown operating conditions or is subjected to strong external disturbances, the reliability of the prediction model usually decreases significantly, and existing methods generally lack online quantitative evaluation mechanisms for prediction errors and prediction reliability.

[0006] On the other hand, to further improve the control robustness of complex stochastic systems, stochastic model predictive control (SMPC) has received widespread attention from the academic community in recent years. This type of method generally predicts the future state of the system through stochastic disturbance models or probabilistic constraints, and optimizes the control input under the condition of satisfying the probabilistic constraints. However, most existing stochastic model predictive controls assume that stochastic disturbances follow a fixed probability distribution or use a fixed number of stochastic scenarios for optimization. The state weights, terminal weights, control weights, constraint margins, and the scale of the predicted scenarios in the controller are usually designed with fixed parameters. When the reliability of the prediction model changes, the controller cannot dynamically adjust the control parameters according to the current prediction risk, which can easily lead to a conservative or overly aggressive control strategy, making it difficult to simultaneously consider control performance, safety, and computational efficiency. Furthermore, existing research usually designs state prediction, uncertainty modeling, and control optimization as independent modules, lacking a unified coupling framework that covers prediction error probability modeling, online risk assessment, and stochastic model predictive control.

[0007] In summary, existing active flutter control methods for airfoils still have the following shortcomings: (1) Most methods rely on relatively accurate aeroelastic dynamics models, and the control performance is prone to decline when the model is mismatched or the system parameters change; (2) Existing data-driven prediction models usually only provide deterministic prediction results, which are difficult to accurately characterize prediction errors and system uncertainties; (3) There is a lack of online risk assessment mechanisms for prediction errors, which cannot reflect the current prediction credibility in real time; (4) In stochastic model predictive control, control parameters such as state weights, control weights, constraint margins, and the number of prediction scenarios are usually fixed and cannot be dynamically adjusted according to prediction risks, which limits the robustness, safety, and environmental adaptability of the system in complex flight environments. Therefore, how to construct an active flutter control method for airfoils that combines high-precision state prediction, prediction error probability modeling, online risk assessment, and risk-adaptive stochastic model predictive control capabilities remains an important technical problem that urgently needs to be solved in the field of aerospace aeroelastic control. Summary of the Invention

[0008] The purpose of this invention is to address the problems of existing active wing flutter suppression methods, such as strong model dependence, insufficient utilization of prediction error uncertainty, difficulty in adapting fixed control parameters to different risk conditions, and insufficient convergence performance under high disturbance and near-constraint conditions. This invention proposes an adaptive active wing flutter risk suppression method. The invention learns the dominant state transition relationship of the wing flutter system through a reservoir computational model, learns the conditional distribution of the one-step prediction error of the reservoir computation through a conditional diffusion residual model, constructs an online risk index through the K-nearest neighbor local residual covariance, and embeds the normalized risk index into a stochastic model predictive control framework. This enables adaptive adjustment of control weights, the number of prediction scenarios, and state constraint margins, thereby improving the convergence speed, robustness, and safety of wing flutter suppression under parameter mismatch, process disturbance, and near-constraint conditions.

[0009] The technical solution for achieving the objective of this invention is as follows: On one hand, an adaptive active suppression method for wing flutter risk is provided, the method comprising:

[0010] Step 1: Obtain offline operating data of the wing flutter system and construct an offline training dataset containing system state variables and control variables;

[0011] Step 2: Construct and train the main dynamics prediction model to learn the dominant state transition law of the wing flutter system;

[0012] Step 3: Calculate the one-step prediction residual based on the main dynamics prediction model, and construct conditional residual modeling samples;

[0013] Step 4: Construct and train the conditional diffusion residual model to obtain the conditional probability distribution of the predicted residuals and the residual mean compensation.

[0014] Step 5: Construct an online residual risk index based on the local residual distribution characteristics, and normalize the online residual risk index to obtain a normalized risk index.

[0015] Step 6: Based on the normalized risk index, adaptively adjust the control parameters of the stochastic model predictive controller;

[0016] Step 7: Combining the main dynamics prediction model, the conditional diffusion residual model, and the adjusted control parameters, establish and solve the stochastic model predictive control optimization problem to obtain the optimal control sequence;

[0017] Step 8: Apply the optimal control sequence to the wing flutter system and continuously update the closed-loop control process to achieve active suppression of wing flutter.

[0018] Furthermore, in step 1, the system state variables include pitch angle, pitch rate, heave displacement, and heave velocity; the control variables include the control channel input acting on the wing section system.

[0019] Furthermore, in step 2, the main dynamic prediction model is a reservoir calculation model; the training objective of the reservoir calculation model is to predict the state increment of the wing flutter system, and then obtain a one-step state prediction value by adding the state increment to the current state.

[0020] Furthermore, the reservoir calculation model includes a fixed randomly generated input weight matrix and a sparse reservoir internal connection matrix, and the reservoir internal connection matrix is ​​normalized according to the spectral radius.

[0021] Furthermore, in step 3, the conditional residual modeling sample includes conditional input and prediction residual; the conditional input simultaneously integrates the current state variable, the current control variable, and the one-step state prediction value obtained from the master dynamics prediction model.

[0022] Furthermore, in step 4, the residual mean prediction is output using the trained conditional diffusion residual model, and the residual mean prediction is superimposed as a compensation term of the principal dynamics prediction model to form a combined prediction model.

[0023] Furthermore, in step 4, the conditional diffusion residual model also generates multiple mean-free residual perturbation scenarios based on the current conditional input sampling, for multi-scenario state prediction.

[0024] Furthermore, in step 5, the specific process of constructing the online residual risk index includes:

[0025] Store historical standardized conditional inputs and their corresponding residual mean prediction errors to construct a residual error library;

[0026] During the online control phase, a K-nearest neighbor search is performed in the residual error database based on the current condition input to obtain k neighborhood residual error samples.

[0027] The local residual covariance matrix is ​​calculated based on the k neighborhood residual error samples, and the trace of the local residual covariance matrix is ​​used as the original residual risk index.

[0028] Furthermore, in step 5, the specific process of normalizing the online residual risk index includes:

[0029] A set of validation set risk indicators is obtained using validation set samples, and the lower and upper quantiles of the set of validation set risk indicators are determined.

[0030] Using the lower quantile and the higher quantile, the original residual risk index calculated online is linearly mapped and truncated to a specified numerical range to obtain the normalized risk index.

[0031] Furthermore, in step 6, the adaptive adjustment of the control parameters of the stochastic model prediction controller includes at least one of the following:

[0032] The state weight matrix and / or terminal weight matrix are adjusted monotonically increasing according to the normalized risk index.

[0033] The control increment weight matrix is ​​adaptively adjusted based on the normalized risk index.

[0034] The number of random prediction scenarios is adjusted according to the monotonically increasing normalized risk index.

[0035] The state safety constraint boundary is dynamically tightened based on the normalized risk index.

[0036] Furthermore, in step 7, the objective function of the stochastic model predictive control optimization problem includes a soft constraint penalty term corresponding to the violation of state safety constraints.

[0037] Further, in step 8, the rolling update closed-loop control process is as follows: the first term of the optimal control sequence obtained by the solution is applied to the wing flutter system. After the system evolves to the next sampling time, the new system state is obtained and the online residual risk index and control parameters are recalculated. The optimization and control process is executed cyclically until the system state converges to the set region.

[0038] On the other hand, an adaptive active suppression system for wing flutter risk is provided, implemented based on the aforementioned adaptive active suppression method for wing flutter risk. The system includes:

[0039] The first module is used to: acquire offline operating data of the wing flutter system and construct an offline training dataset containing system state variables and control variables;

[0040] The second module is used to: construct and train the main dynamics prediction model, and learn the dominant state transition law of the wing flutter system;

[0041] The third module is used to: calculate the one-step prediction residual based on the main dynamics prediction model, and construct conditional residual modeling samples;

[0042] The fourth module is used to: construct and train the conditional diffusion residual model, obtain the conditional probability distribution of the predicted residuals and residual mean compensation;

[0043] The fifth module is used to: construct an online residual risk index based on the local residual distribution characteristics, and normalize the online residual risk index to obtain a normalized risk index;

[0044] The sixth module is used to: adaptively adjust the control parameters of the stochastic model prediction controller based on the normalized risk index;

[0045] The seventh module is used to: combine the main dynamics prediction model, the conditional diffusion residual model, and the adjusted control parameters to establish and solve the stochastic model predictive control optimization problem to obtain the optimal control sequence;

[0046] The eighth module is used to: apply the optimal control sequence to the wing flutter system and continuously update the closed-loop control process to achieve active suppression of wing flutter.

[0047] On the other hand, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the adaptive active suppression method for wing flutter risk.

[0048] Compared with the prior art, the significant advantages of this invention are:

[0049] (1) The dominant state transition law of the wing flutter system is learned online by using the reservoir calculation (RC) model, which reduces the strong dependence on the accurate aeroelastic analytical model and significantly improves the system's adaptability and anti-interference ability under model mismatch and parameter change conditions.

[0050] (2) By using the conditional diffusion residual model to learn the conditional probability distribution of the prediction residual, the state prediction accuracy can be improved by using the residual mean. It can also make full use of the uncertainty information contained in the residual distribution, thus solving the defect that traditional data-driven models can only output deterministic prediction results.

[0051] (3) Based on the local residual covariance of K nearest neighbors, a residual error library is constructed, which can transform local prediction uncertainty into an intuitive scalar risk indicator in real time. Combined with the validation set quantiles, the normalization process is completed, which can stably characterize the prediction reliability of the current state-control region and provide support for the controller to dynamically adjust the control strategy according to the prediction reliability of the current state-control region.

[0052] (4) The controller can dynamically adjust the state weight, terminal weight, control increment weight, number of random prediction scenarios and state constraint margin based on the normalized risk index; reduce control conservatism in low-risk areas to improve computational efficiency, strengthen control strength and widen safety boundaries in high-risk or near-constraint areas, and simultaneously take into account control performance, control safety and computational efficiency.

[0053] (5) Compared with the fixed parameter control method, this method can significantly shorten the system response adjustment time and reduce the terminal residual error under normal operating conditions, high process disturbance conditions and near-constraint conditions close to the state safety boundary, and ensure the rapid and stable decay of wing flutter. It has a more obvious convergence advantage, especially under high disturbance and near-constraint conditions.

[0054] The present invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0055] Figure 1 This is a flowchart illustrating the overall implementation of the adaptive active suppression method for wing flutter risk of the present invention.

[0056] Figure 2 This is a structural diagram of the combination of the RC master prediction model and the conditional diffusion residual model in one embodiment.

[0057] Figure 3 This is a flowchart of the KNN local residual covariance risk estimation in one embodiment.

[0058] Figure 4 This is a mapping diagram from risk indicators to SMPC parameters in one embodiment.

[0059] Figure 5 This is a flowchart of risk-adaptive SMPC rolling optimization in one embodiment.

[0060] Figure 6 This is a comparison chart of the state response under high disturbance conditions in one embodiment, wherein... Figure 6 (a) in the figure is a comparison diagram of the state response under high disturbance pitch angle. Figure 6 (b) in the figure is a comparison diagram of the response to high disturbance buoyancy.

[0061] Figure 7 This is a comparison chart of near-constraint operating condition state responses in one embodiment, where... Figure 7(a) in the figure is a comparison diagram of the pitch angle state response under near-constraint conditions. Figure 7 (b) in the figure is a comparison diagram of the sinking and floating state response under near-constraint working conditions.

[0062] Figure 8 Here is a graph showing the changes in normalized risk indicators and the number of predicted scenarios in one embodiment, where Figure 8 In the diagram, (a) represents the normalized residual risk of the complete risk adaptive RC-Diffusion SMPC controller under different operating conditions. Figure 8 (b) in the figure represents the change in the number of adaptive prediction scenarios of the complete risk adaptive RC-Diffusion SMPC controller under different operating conditions. Detailed Implementation

[0063] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0064] It should be noted that if the embodiments of the present invention involve descriptions such as "first" and "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" and "second" may explicitly or implicitly include at least one of those features. Furthermore, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0065] In one embodiment, combined Figure 1 An adaptive active suppression method for wing flutter risk is provided, the method comprising:

[0066] Step 1: Obtain offline operating data of the wing flutter system and construct an offline training dataset containing system state variables and control variables;

[0067] Step 2: Construct and train the main dynamics prediction model to learn the dominant state transition law of the wing flutter system;

[0068] Step 3: Calculate the one-step prediction residual based on the main dynamics prediction model, and construct conditional residual modeling samples;

[0069] Step 4: Construct and train the conditional diffusion residual model to obtain the conditional probability distribution of the predicted residuals and the residual mean compensation.

[0070] Step 5: Construct an online residual risk index based on the local residual distribution characteristics, and normalize the online residual risk index to obtain a normalized risk index.

[0071] Step 6: Based on the normalized risk index, adaptively adjust the control parameters of the stochastic model predictive controller;

[0072] Step 7: Combining the main dynamics prediction model, the conditional diffusion residual model, and the adjusted control parameters, establish and solve the stochastic model predictive control optimization problem to obtain the optimal control sequence;

[0073] Step 8: Apply the optimal control sequence to the wing flutter system and continuously update the closed-loop control process to achieve active suppression of wing flutter.

[0074] Furthermore, in one embodiment, in step 1, the system state variables include pitch angle, pitch rate, heave displacement, and heave velocity; the control variables include control channel inputs acting on the wing section system.

[0075] Here, the state variables of the discrete-time wing flutter system are represented as:

[0076]

[0077] In the formula, For the first The pitch angle at each sampling time. The pitch angular velocity, For buoyancy displacement, For buoyancy, For the first The state variables of the wing flutter system at each sampling time.

[0078] Control variables (control inputs) are represented as :

[0079]

[0080] In the formula, and These are two control channels that act on the wing segment system.

[0081] The one-step state transition relationship of the wing flutter system is expressed as:

[0082]

[0083] In the formula, This is a mapping of state evolution within a sampling period. For process disturbances, it is used to represent factors such as external disturbances, unmodeled dynamics, numerical errors, and parameter mismatches; For the first The state variables of the wing flutter system at each sampling time;

[0084] Collect or generate offline state-control-next state samples to form the training dataset:

[0085]

[0086] In the formula, This indicates the number of offline samples.

[0087] Here, the condition for executing step 1 is that the simulation data or experimental response data of the wing flutter system has been obtained; its purpose is to provide a sample basis for subsequent data-driven prediction model training and residual uncertainty modeling; this step provides input data for steps 2 and 3.

[0088] Furthermore, in one embodiment, in step 2, the main dynamic prediction model is a reservoir calculation model; the training objective of the reservoir calculation model is to predict the state increment of the wing flutter system, and then obtain a one-step state prediction value by adding the state increment to the current state.

[0089] More preferably, in some embodiments, the reservoir calculation model includes a fixed randomly generated input weight matrix and a sparse reservoir internal connection matrix, and the reservoir internal connection matrix is ​​normalized according to the spectral radius.

[0090] Furthermore, in some embodiments, step 2 specifically includes:

[0091] Step 2-1, change the current state With current control input The inputs for the reservoir calculation model are combined as follows:

[0092]

[0093] And on Standardization is performed to obtain standardized input variables. ;

[0094] Step 2-2, let the current state of the storage pool be... Update the internal state using the following formula:

[0095]

[0096]

[0097] In the formula, This represents the candidate internal state calculated using a nonlinear activation function based on the current standardized input and the previous state of the reservoir. The hyperbolic tangent activation function is used. For the input weight matrix, This is the internal connection matrix of the reserve pool. Leakage rate; This represents the next state of the storage pool. Here, A random sparse matrix is ​​used and normalized according to the spectral radius to control the memory capacity and nonlinear dynamic expression capacity of the reservoir.

[0098] Here, the reservoir computation model includes an input weight matrix, a reservoir internal connection matrix, and an output layer weight matrix. The input weight matrix and the reservoir internal connection matrix are randomly generated and kept fixed. The reservoir internal connection matrix adopts a sparse matrix form, and its dynamic memory capacity is adjusted by spectral radius normalization. Only the output layer weights are trained during model training.

[0099] Steps 2-3: Since the state changes between adjacent sampling times are usually small, this invention does not directly predict... Instead, it trains a reservoir to calculate the model and predict state increments. Specifically:

[0100] state increment Train the output layer weights for the training objective. The output of the main dynamics prediction model for the reservoir is The predicted state value for the first step is:

[0101]

[0102] In the formula, This represents the one-step predicted state given by the master model for calculating the reservoir.

[0103] Furthermore, in one embodiment, in step 3, the conditional residual modeling sample includes conditional input and prediction residual; the conditional input simultaneously integrates the current state variable, the current control variable, and the one-step state prediction value obtained from the master dynamics prediction model.

[0104] Here, step 3 mainly calculates the one-step prediction residual of the reservoir calculation model and constructs the training samples required for the conditional diffusion residual model. By analyzing the error between the predicted value and the actual state, the system prediction problem is decomposed into two parts: principal dynamics prediction and residual modeling, laying the foundation for the subsequent conditional diffusion model to learn the prediction error distribution.

[0105] Preferably, step 3 specifically includes:

[0106] Step 3-1: Perform a one-step prediction using a reservoir calculation for offline samples.

[0107] For each set of samples in the offline training dataset, the main dynamics prediction model is calculated using the reservoir trained in step 2 to obtain a one-step prediction value. .

[0108] Step 3-2: Calculate the predicted residuals for the reserve pool.

[0109] The predicted value of the reservoir at the next step is compared with the actual state at the next time step, and the prediction residual is calculated:

[0110]

[0111] in, This indicates the state prediction error that the calculated master dynamics prediction model for the reservoir cannot explain. This error includes model approximation error, process disturbance effects, parameter mismatch effects, and unmodeled dynamic effects.

[0112] Step 3-3: Construct the conditional inputs for the conditional diffusion residual model.

[0113] Constructing conditional input :

[0114]

[0115] This conditional input reflects the current state, control input, and reservoir calculation prediction results simultaneously, and therefore can be used to describe the local dynamic background of the current residual occurrence.

[0116] Steps 3-4 form residual modeling samples.

[0117] Enter the conditions Compared with the predicted residual These samples are used to form residual modeling samples for subsequent training of conditional diffusion residual models.

[0118] Furthermore, in one embodiment, in step 4, the residual mean prediction is output using the trained conditional diffusion residual model, and the residual mean prediction is superimposed as a compensation term of the principal dynamics prediction model to form a combined prediction model.

[0119] Furthermore, in step 4, the conditional diffusion residual model also generates multiple mean-free residual perturbation scenarios based on the current conditional input sampling, for multi-scenario state prediction.

[0120] Here, step 4 primarily trains the conditional diffusion residual model. The conditional probability distribution of the prediction error is calculated using the conditional sample learning pool constructed in step 3, simultaneously obtaining residual mean compensation capability and residual random sampling capability, providing probabilistic information for risk estimation and random prediction scenario generation.

[0121] Preferably, in some embodiments, step 4 specifically includes:

[0122] by As a condition variable, with Train a conditional diffusion residual model for the modeling object to obtain the residual conditional probability distribution under the current conditions. ;

[0123] Predict the mean residual value using the trained conditional diffusion residual model. The predicted residual mean is then embedded as a residual mean compensation term, resulting in the combined prediction model:

[0124]

[0125] In the formula, This indicates the combined prediction state after compensation by the residual mean.

[0126] Simultaneously, the conditional diffusion residual model generates the data based on the current conditional sampling. The scenario of perturbation of the mean residual. ,in Indicates the first A residual perturbation scenario.

[0127] More specifically, step 4 includes:

[0128] Step 4-1: Establish training samples for the conditional diffusion residual model.

[0129] Combination Figure 2 Input the conditions obtained in step 3 As a condition variable, the predicted residuals are calculated using the reservoir. The conditional diffusion residual model is trained as the modeling object. This model is used to learn the conditional distribution of the prediction residuals given the current state, control input, and reservoir prediction results.

[0130]

[0131] In the formula, This is the set of trainable parameters for the conditional diffusion residual model.

[0132] Step 4-2: Train the residual mean compensation network.

[0133] Conditional diffusion residual models can output residual mean predictions:

[0134]

[0135] The mean residual is used to compensate for the one-step prediction error of the master model for reservoir calculation.

[0136] Step 4-3: Construct a combined prediction model.

[0137] Using the residual mean compensation term, the combined prediction model is obtained:

[0138]

[0139] in, This represents the combined predicted state after residual mean compensation. Compared to using the reserve pool calculation model alone, this combined prediction model can not only obtain the state prediction mean, but also further obtain residual distribution information, providing a basis for subsequent risk assessment and multi-scenario stochastic predictive control.

[0140] Step 4-4: Train the conditional diffusion sampling network.

[0141] During diffusion training, noise is gradually added to the residual samples, causing them to diffuse from the original residual distribution to an approximate Gaussian noise distribution. Subsequently, the neural network is trained to learn the inverse denoising process, enabling the model to perform denoising under given conditions. Residual samples that conform to the residual conditional distribution are generated in time.

[0142] Steps 4-5 generate residual perturbation scenarios.

[0143] When multi-scenario prediction is required, the conditional diffusion residual model generates residual samples based on the current conditional input:

[0144]

[0145] in, Indicates the first One residual perturbation scenario. Preferably, in this embodiment, the diffusion model can have 15 steps. The conditional diffusion residual model is mainly used for residual mean compensation, residual sampling, and characterizing local prediction uncertainties.

[0146] Furthermore, in one embodiment, step 5, the specific process of constructing the online residual risk index includes:

[0147] Store historical standardized conditional inputs and their corresponding residual mean prediction errors to construct a residual error library;

[0148] During the online control phase, a K-nearest neighbor search is performed in the residual error database based on the current condition input to obtain k neighborhood residual error samples.

[0149] The local residual covariance matrix is ​​calculated based on the k neighborhood residual error samples, and the trace of the local residual covariance matrix is ​​used as the original residual risk index.

[0150] Furthermore, in one embodiment, step 5, the specific process of normalizing the online residual risk index, includes:

[0151] A set of validation set risk indicators is obtained using validation set samples, and the lower and upper quantiles of the set of validation set risk indicators are determined.

[0152] Using the lower quantile and the higher quantile, the original residual risk index calculated online is linearly mapped and truncated to a specified numerical range to obtain the normalized risk index.

[0153] Here, step 5 mainly involves constructing an online residual risk index. By establishing a residual error library and using K-nearest neighbor local residual covariance analysis to analyze the degree of uncertainty near the current prediction conditions, the local prediction error distribution is transformed into a quantifiable risk index, providing a basis for subsequent adaptive adjustment of control parameters.

[0154] Preferably, in some embodiments, step 5 specifically includes:

[0155] Step 5-1, combined Figure 3 During the offline phase, input the offline conditions. Standardized to Calculate the mean residual prediction error of the combined prediction model:

[0156]

[0157] In the formula, Indicates the sample number in the offline dataset; Indicates the first RC one-step prediction residuals for each sample;

[0158] Or equivalently represented as:

[0159]

[0160] Construct a sample pair residual error library :

[0161]

[0162] Step 5-2, during the online control phase, based on the current status Candidate control input and the reserve pool's one-step prediction status Construct the current conditional input and standardized as ;

[0163] Step 5-3, in the residual error library China and Israel Perform a K-nearest neighbor search on the query point to obtain the nearest neighbor that is closest to the current condition. A number of historical residual error samples are used to obtain the current conditions. of A set of nearest neighbor samples ;

[0164] Step 5-4, calculate the local residual covariance matrix:

[0165]

[0166] In the formula, This indicates the operation for calculating sample covariance. Indicates standardization of current conditions Obtained for query point A set of nearest neighbor samples;

[0167] Step 5-5: Calculate the online residual risk index (take the trace of the local residual covariance matrix as the current original residual risk index):

[0168]

[0169] In the formula, Represented as the first The original residual risk index at each sampling time. Represents the trace of a matrix. The larger the value, the more dispersed the distribution of the predicted residuals near the current state-control region, and the higher the uncertainty of the model prediction. The smaller the value, the more concentrated the predicted residuals are in the current region, and the higher the reliability of the model prediction. In this embodiment, the number of K-nearest neighbors can be 60, and the risk indicator is the trace of the local residual covariance matrix.

[0170] Steps 5-6 involve normalizing the online residual risk index using the following formula:

[0171]

[0172] In the formula, As a normalized risk indicator, and These are the 5th and 99th percentiles of the set of online residual risk indicators calculated from the validation set samples, respectively. This represents a truncation function that cuts a variable to the interval [0, 1].

[0173] Steps 5-7: Identify risk areas.

[0174] when hour, This indicates a low-risk area; when hour, , indicating a high-risk area; when lie in and In between, Follow Linear change.

[0175] Preferably, in some embodiments, the low quantile is calculated based on the set of validation set risk indicators. and high quantile .

[0176] Here, due to the original risk indicators Due to the influence of state dimensions, residual scale, data distribution, and the K-nearest neighbor parameter, directly using these parameters for controller parameter adjustment may lead to instability. Therefore, this embodiment utilizes the validation set samples to repeat the local residual covariance calculation process in step 5 to obtain the validation set risk index set:

[0177]

[0178] In the formula, To verify the local residual risk index corresponding to the j-th sample in the set, j=1,2,..., , This represents the total number of samples in the validation set.

[0179] Here, step 5 mainly completes the normalization of online risk indicators. Risk quantiles are obtained using validation set statistics, and a unified scale mapping is performed on the risk indicators obtained under different operating conditions. This ensures that the risk indicators can stably reflect the current prediction reliability and provides standardized input for adaptive mapping of control parameters.

[0180] Furthermore, in one embodiment, step 6, the adaptive adjustment of the control parameters of the stochastic model prediction controller, includes at least one of the following:

[0181] The state weight matrix and / or terminal weight matrix are adjusted monotonically increasing according to the normalized risk index.

[0182] The control increment weight matrix is ​​adaptively adjusted based on the normalized risk index.

[0183] The number of random prediction scenarios is adjusted according to the monotonically increasing normalized risk index.

[0184] The state safety constraint boundary is dynamically tightened based on the normalized risk index.

[0185] Here, step 6 mainly involves adaptively adjusting the parameters of the stochastic model predictive controller based on the normalized risk index. The state weights, terminal weights, control weights, number of predicted scenarios, and state constraint margins are dynamically adjusted according to the current risk level, enabling the controller to balance control performance, safety, and computational efficiency.

[0186] Preferably, in some embodiments, step 6 specifically includes:

[0187] Step 6-1: Adjust the state weight matrix.

[0188] Combination Figure 4 Normalized risk indicators A stochastic model predictive controller is introduced. Based on... Adjusting the state weight matrix The higher the risk, the stronger the penalty for state deviation, making the controller pay more attention to the rapid convergence of pitch angle, heave displacement and velocity state.

[0189] Step 6-2: Adjust the terminal weight matrix.

[0190] according to Adjusting the terminal weight matrix The higher the risk, the stronger the requirement for convergence of the predicted end-time state.

[0191] Step 6-3: Adjust the control increment weight matrix.

[0192] according to Adjusting the control increment weight matrix When the risk is high, the penalty for control increment can be increased to suppress abrupt changes in control input, or stronger control intervention can be allowed as needed for engineering purposes.

[0193] Step 6-4: Adjust the number of predicted scenarios.

[0194] Adjust the number of predicted scenarios based on normalized risk indicators:

[0195]

[0196] The higher the risk, the more prediction scenarios are needed to better cover residual uncertainty; when the risk is low, the number of prediction scenarios is reduced to lower the computational burden. In this embodiment, the minimum number of prediction scenarios can be 4, the maximum number of prediction scenarios can be 10, and the scenario sensitivity coefficient can be 2.5.

[0197] Step 6-5: Adjust the state constraint boundaries.

[0198] according to Adjust state constraint boundaries. The higher the risk, the tighter the state safety boundaries used internally by the controller, thus reserving a larger safety margin for the actual system.

[0199] Furthermore, in one embodiment, in step 7, the objective function of the stochastic model predictive control optimization problem includes a soft constraint penalty term corresponding to the violation of state safety constraints.

[0200] Here, step 7 mainly establishes and solves the risk adaptive stochastic model predictive control optimization problem. The prediction model of the integrated reserve pool calculation, the conditional diffusion residual model, and the risk adaptive control parameters are used to solve for the optimal control sequence within the finite prediction time domain under the conditions of input constraints and state constraints, thereby achieving active suppression control of wing flutter.

[0201] Preferably, in some embodiments, the risk adaptive stochastic model predictive control optimization problem established in step 7 is expressed as:

[0202] At the current sampling time k, solve for the control increment or control input sequence to minimize the objective function. :

[0203]

[0204] in,

[0205]

[0206] In the formula, For the first The finite-time domain control cost corresponding to each prediction scenario To predict the number of scenarios; , and These are the risk adaptive state weight matrix, the control increment weight matrix, and the terminal weight matrix, respectively. The control input weight matrix is ​​fixed. This is a soft constraint penalty item. To predict the length of the time domain, For the current moment Based on what has been obtained, regarding future moments The Predicted values ​​for each scene state. For at any time The optimized result is planned for future moments. The applied control input, The control input increment between adjacent prediction steps. For the current moment , for the A random scenario at the end of the prediction time domain The predicted system state value;

[0207] Limited by constraints:

[0208] Predicting multi-scenario evolution: , Indicates the predicted scene number. Indicates the prediction step number. This is a scenario with mean-free residual perturbation obtained by sampling from a conditional diffusion residual model. The current sampling time , No. In a random prediction scenario, the future Predicted values ​​of step states, Calculate the master forecasting model for the reserve pool. This is a prediction of the residual mean output by the conditional diffusion residual model.

[0209] Input constraints: ,in, and These represent the upper and lower limits of the control input, respectively.

[0210] Dynamically tightened state constraints: ,in, , This is the risk tightening coefficient for state constraints. This is the original state safety boundary vector before risk tightening.

[0211] Specifically, step 7 includes the following steps:

[0212] Step 7-1: Construct the predictive control sequence.

[0213] Combination Figure 5 At each sampling time k, the current state is... As the initial state for prediction, construct a prediction time domain with a length of... Control sequence:

[0214]

[0215] Step 7-2: Perform multi-scenario state prediction.

[0216] Within the prediction time domain, the master model and conditional diffusion residual model are calculated using a reservoir to predict states across multiple scenarios. The state recursion under each prediction scenario can be represented as:

[0217]

[0218] Here, to improve the smoothness of the optimization process, residual mean propagation can also be used, that is, utilizing... Complete state prediction, while residual random sampling is mainly used for online risk assessment and adjustment of the number of prediction scenarios.

[0219] Step 7-3: Construct the risk adaptive objective function.

[0220] The optimization objective comprehensively considers state deviation, control input amplitude, control increment, terminal state error, and soft constraint penalty term. The risk adaptive stochastic model predictive control objective function is expressed as:

[0221]

[0222] in,

[0223] In the formula, To predict the number of scenarios; , and These are the risk adaptive state weight matrix, the control increment weight matrix, and the terminal weight matrix, respectively. The control input weight matrix is ​​fixed. This is a soft constraint penalty item.

[0224] Step 7-4: Set control input constraints.

[0225] Control inputs satisfy:

[0226]

[0227] In the formula, and These represent the upper and lower limits of the control input, respectively.

[0228] Step 7-5: Set state constraints after risk tightening.

[0229] Based on the current normalized risk level, the state safety boundary is dynamically tightened:

[0230]

[0231] Therefore, each predicted state satisfies:

[0232]

[0233] The higher the risk, the smaller the state constraint boundary, thus reserving a larger safety margin for the actual system.

[0234] Step 7-6: Construct soft constraint penalty terms.

[0235] To prevent the optimization problem from becoming infeasible due to model errors or external perturbations, a state constraint violation quantity is introduced:

[0236]

[0237] In the formula, For the first At the sampling time, the first The predicted scenario is in the... The number of state constraints violated in each prediction step;

[0238] The corresponding soft constraint penalty term is expressed as follows:

[0239]

[0240] in, This is the soft constraint penalty coefficient. By penalizing the predicted amount of violation that exceeds the safety boundary, the safety and robustness of the control system are improved.

[0241] Step 7-7: Solve the finite-time optimization problem.

[0242] By considering the adaptive risk weights, the number of prediction scenarios, and the state constraints after risk tightening, a finite prediction time-domain stochastic optimization problem is solved to obtain the optimal control sequence:

[0243]

[0244] The first control input in the optimal control sequence is applied to the actual system, and the system state is reacquired at the next sampling time. The above prediction optimization process is repeated to achieve rolling optimization control.

[0245] Furthermore, in one embodiment, in step 8, the rolling update closed-loop control process is as follows: the first term of the optimal control sequence obtained by the solution is applied to the wing flutter system. After the system evolves to the next sampling time, a new system state is obtained and the online residual risk index and control parameters are recalculated. The optimization and control process is executed cyclically until the system state converges to the set region (until the pitch angle, pitch velocity, heave displacement and heave velocity of the wing flutter system converge to the neighborhood of the equilibrium point).

[0246] Here, step 8 mainly implements closed-loop rolling optimization control. The first term of the currently optimized optimal control sequence is applied to the wing flutter system. The risk indicators and control strategies are continuously updated according to the real-time state of the system. The predictive optimization process is executed cyclically until the wing flutter state converges, thereby achieving adaptive and active risk suppression.

[0247] Specifically, step 8 includes:

[0248] Step 8-1: Apply the first term of the optimal control sequence.

[0249] Combination Figure 5 The first term of the optimal control sequence obtained in step 7 This is the current actual control input applied to the wing flutter system.

[0250] Step 8-2: Obtain the system state at the next sampling time.

[0251] Under the influence of this control input, the system operates until the next sampling time and obtains a new system state. .

[0252] Step 8-3: Recalculate online risk indicators and control parameters.

[0253] by As a new initial state, steps 5 to 7 are re-executed to recalculate the current conditional input, online risk index, normalized risk index, control weight, number of prediction scenarios, state constraint boundary, and optimal control input.

[0254] Step 8-4, repeat the rolling optimization process.

[0255] Repeat the rolling optimization process described above until the pitch angle, pitch velocity, heave displacement, and heave velocity of the wing flutter system converge to the neighborhood of the equilibrium point, thereby achieving active suppression of wing flutter.

[0256] In this embodiment, the sampling period can be 0.05s, the closed-loop simulation time can be 8s, the maximum number of optimizer iterations can be 30, and the optimization tolerance can be 10. -4 Measurement noise is turned off, and process noise is turned on, to highlight the role of residual modeling and risk adaptive control mechanisms under parameter mismatch and process disturbance conditions.

[0257] In one embodiment, an adaptive active suppression system for wing flutter risk is provided, implemented based on the adaptive active suppression method for wing flutter risk. The system includes:

[0258] The first module is used to: acquire offline operating data of the wing flutter system and construct an offline training dataset containing system state variables and control variables;

[0259] The second module is used to: construct and train the main dynamics prediction model, and learn the dominant state transition law of the wing flutter system;

[0260] The third module is used to: calculate the one-step prediction residual based on the main dynamics prediction model, and construct conditional residual modeling samples;

[0261] The fourth module is used to: construct and train the conditional diffusion residual model, obtain the conditional probability distribution of the predicted residuals and residual mean compensation;

[0262] The fifth module is used to: construct an online residual risk index based on the local residual distribution characteristics, and normalize the online residual risk index to obtain a normalized risk index;

[0263] The sixth module is used to: adaptively adjust the control parameters of the stochastic model prediction controller based on the normalized risk index;

[0264] The seventh module is used to: combine the main dynamics prediction model, the conditional diffusion residual model, and the adjusted control parameters to establish and solve the stochastic model predictive control optimization problem to obtain the optimal control sequence;

[0265] The eighth module is used to: apply the optimal control sequence to the wing flutter system and continuously update the closed-loop control process to achieve active suppression of wing flutter.

[0266] Specific limitations regarding the adaptive active suppression system for wing flutter risk can be found in the limitations of the adaptive active suppression method for wing flutter risk described above, and will not be repeated here. Each module in the aforementioned adaptive active suppression system for wing flutter risk can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0267] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements:

[0268] Step 1: Obtain offline operating data of the wing flutter system and construct an offline training dataset containing system state variables and control variables;

[0269] Step 2: Construct and train the main dynamics prediction model to learn the dominant state transition law of the wing flutter system;

[0270] Step 3: Calculate the one-step prediction residual based on the main dynamics prediction model, and construct conditional residual modeling samples;

[0271] Step 4: Construct and train the conditional diffusion residual model to obtain the conditional probability distribution of the predicted residuals and the residual mean compensation.

[0272] Step 5: Construct an online residual risk index based on the local residual distribution characteristics, and normalize the online residual risk index to obtain a normalized risk index.

[0273] Step 6: Based on the normalized risk index, adaptively adjust the control parameters of the stochastic model predictive controller;

[0274] Step 7: Combining the main dynamics prediction model, the conditional diffusion residual model, and the adjusted control parameters, establish and solve the stochastic model predictive control optimization problem to obtain the optimal control sequence;

[0275] Step 8: Apply the optimal control sequence to the wing flutter system and continuously update the closed-loop control process to achieve active suppression of wing flutter.

[0276] For specific limitations on each step, please refer to the limitations of the adaptive active suppression method for wing flutter risk mentioned above, which will not be repeated here.

[0277] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program being implemented when executed by a processor:

[0278] Step 1: Obtain offline operating data of the wing flutter system and construct an offline training dataset containing system state variables and control variables;

[0279] Step 2: Construct and train the main dynamics prediction model to learn the dominant state transition law of the wing flutter system;

[0280] Step 3: Calculate the one-step prediction residual based on the main dynamics prediction model, and construct conditional residual modeling samples;

[0281] Step 4: Construct and train the conditional diffusion residual model to obtain the conditional probability distribution of the predicted residuals and the residual mean compensation.

[0282] Step 5: Construct an online residual risk index based on the local residual distribution characteristics, and normalize the online residual risk index to obtain a normalized risk index.

[0283] Step 6: Based on the normalized risk index, adaptively adjust the control parameters of the stochastic model predictive controller;

[0284] Step 7: Combining the main dynamics prediction model, the conditional diffusion residual model, and the adjusted control parameters, establish and solve the stochastic model predictive control optimization problem to obtain the optimal control sequence;

[0285] Step 8: Apply the optimal control sequence to the wing flutter system and continuously update the closed-loop control process to achieve active suppression of wing flutter.

[0286] For specific limitations on each step, please refer to the limitations of the adaptive active suppression method for wing flutter risk mentioned above, which will not be repeated here.

[0287] As a specific example, the invention will be further verified and illustrated in one embodiment.

[0288] To verify the technical effectiveness of the method of this invention, three types of test conditions can be set: normal test condition, high disturbance test condition, and near-constraint test condition. Among them, the normal test condition is used to verify the convergence performance of the method of this invention under ordinary closed-loop control conditions; the high disturbance test condition is used to verify the robustness of the method of this invention under strong process disturbances; and the near-constraint test condition is used to verify the ability of the method of this invention to quickly escape the boundary and the safety margin when approaching the state safety boundary.

[0289] In a specific simulation embodiment, the initial states for normal operating conditions and high-disturbance operating conditions can be set as follows:

[0290]

[0291] The initial state of the near-constraint condition can be set as follows:

[0292]

[0293] State safety boundaries can be set as follows:

[0294] The absolute value of the pitch angle shall not exceed 0.35.

[0295] The absolute value of the pitch angular velocity does not exceed 1.80;

[0296] The absolute value of the floating and sinking displacement does not exceed 0.25;

[0297] The absolute value of the buoyancy velocity does not exceed 1.80.

[0298] For comparative verification, four types of controllers can be set up: the first type is RC-onlySMPC that only uses the reserve pool for calculation and prediction; the second type is fixed RC-Diffusion SMPC that uses the reserve pool for calculation and conditional diffusion residual compensation but does not perform risk adaptive adjustment; the third type is weight adaptive RC-Diffusion SMPC that only adjusts the control weights according to the risk; and the fourth type is the full-risk adaptive RC-Diffusion SMPC proposed in this invention, which simultaneously adjusts the control weights, the number of prediction scenarios, and the state constraint margin.

[0299] Combination Figure 6 Under high-disturbance conditions, the method of this invention can cause the pitch angle and heave displacement to decay to the neighborhood of the equilibrium point more quickly; combined with Figure 7 Under near-constraint conditions, the method of this invention can reduce pitch angle and heave displacement response more quickly and move the system state away from the safety boundary; combined with Figure 8 The normalized residual risk of the method of this invention can change with the current state-control conditions and drive the number of predicted scenarios to adjust adaptively. Combined with... Figure 8 The normalized residual risk of the method of the present invention can change with the current state-control conditions and drive the number of predicted scenarios to increase from the base value to a higher level, thus demonstrating that the risk adaptive mechanism of the present invention is effectively activated in the closed-loop control process.

[0300] Simulation results show that, compared with RC-only SMPC and fixed RC-Diffusion SMPC, the method of this invention can reduce terminal state error, shorten settling time, and exhibit a more significant convergence advantage under high disturbance and near-constraint conditions. Although the method of this invention increases control energy and computation time in some operating conditions, this increase is compensated by stronger prediction uncertainty coverage, faster state convergence speed, and better adaptability to high-risk operating conditions. Therefore, this invention is applicable to active flutter suppression scenarios with parameter mismatch, process disturbance, model prediction error, and state safety constraints.

[0301] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention without departing from its spirit and scope should be included within the protection scope of the present invention.

Claims

1. An adaptive active suppression method for wing flutter risk, characterized in that, The method includes: Step 1: Obtain offline operating data of the wing flutter system and construct an offline training dataset containing system state variables and control variables; Step 2: Construct and train the main dynamics prediction model to learn the dominant state transition law of the wing flutter system; Step 3: Calculate the one-step prediction residual based on the main dynamics prediction model, and construct conditional residual modeling samples; Step 4: Construct and train the conditional diffusion residual model to obtain the conditional probability distribution of the predicted residuals and the residual mean compensation. Step 5: Construct an online residual risk index based on the local residual distribution characteristics, and normalize the online residual risk index to obtain a normalized risk index. Step 6: Based on the normalized risk index, adaptively adjust the control parameters of the stochastic model predictive controller; Step 7: Combining the main dynamics prediction model, the conditional diffusion residual model, and the adjusted control parameters, establish and solve the stochastic model predictive control optimization problem to obtain the optimal control sequence; Step 8: Apply the optimal control sequence to the wing flutter system and continuously update the closed-loop control process to achieve active suppression of wing flutter.

2. The adaptive active suppression method for wing flutter risk according to claim 1, characterized in that, In step 1, the system state variables include pitch angle, pitch rate, heave displacement, and heave velocity; the control variables include the control channel inputs acting on the wing section system. In step 2, the main dynamics prediction model is a reservoir calculation model; the training objective of the reservoir calculation model is to predict the state increment of the wing flutter system, and then obtain a one-step state prediction value by adding the state increment to the current state. The reservoir calculation model includes a fixed randomly generated input weight matrix and a sparse reservoir internal connection matrix, and the reservoir internal connection matrix is ​​normalized according to the spectral radius.

3. The adaptive active suppression method for wing flutter risk according to claim 1, characterized in that, In step 3, the conditional residual modeling sample includes conditional input and prediction residual; the conditional input simultaneously integrates the current state variable, the current control variable, and the one-step state prediction value obtained from the master dynamics prediction model.

4. The adaptive active suppression method for wing flutter risk according to claim 1, characterized in that, In step 4, the residual mean prediction is output using the trained conditional diffusion residual model, and the residual mean prediction is superimposed as a compensation term of the principal dynamics prediction model to form a combined prediction model. In step 4, the conditional diffusion residual model also generates multiple mean-free residual perturbation scenarios based on the current conditional input sampling, which are used for multi-scenario state prediction.

5. The adaptive active suppression method for wing flutter risk according to claim 1, characterized in that, Step 5, the specific process of constructing the online residual risk index includes: Store historical standardized conditional inputs and their corresponding residual mean prediction errors to construct a residual error library; During the online control phase, a K-nearest neighbor search is performed in the residual error database based on the current condition input to obtain k neighborhood residual error samples. The local residual covariance matrix is ​​calculated based on the k neighborhood residual error samples, and the trace of the local residual covariance matrix is ​​used as the original residual risk index. Step 5, the specific process of normalizing the online residual risk index includes: A set of validation set risk indicators is obtained using validation set samples, and the lower and upper quantiles of the set of validation set risk indicators are determined. Using the lower quantile and the higher quantile, the original residual risk index calculated online is linearly mapped and truncated to a specified numerical range to obtain the normalized risk index.

6. The adaptive active suppression method for wing flutter risk according to claim 1, characterized in that, In step 6, the adaptive adjustment of the control parameters of the stochastic model prediction controller includes at least one of the following: The state weight matrix and / or terminal weight matrix are adjusted monotonically increasing according to the normalized risk index. The control increment weight matrix is ​​adaptively adjusted based on the normalized risk index. The number of random prediction scenarios is adjusted according to the monotonically increasing normalized risk index. The state safety constraint boundary is dynamically tightened based on the normalized risk index.

7. The adaptive active suppression method for wing flutter risk according to claim 6, characterized in that, In step 7, the objective function of the stochastic model predictive control optimization problem includes a soft constraint penalty term corresponding to the violation of state safety constraints.

8. The adaptive active suppression method for wing flutter risk according to claim 1, characterized in that, In step 8, the rolling update closed-loop control process is as follows: the first term of the optimal control sequence obtained by the solution is applied to the wing flutter system. After the system evolves to the next sampling time, the new system state is obtained and the online residual risk index and control parameters are recalculated. The optimization and control process is executed cyclically until the system state converges to the set region.

9. An adaptive active suppression system for wing flutter risk, implemented based on the method described in any one of claims 1 to 8, characterized in that, The system includes: The first module is used to: acquire offline operating data of the wing flutter system and construct an offline training dataset containing system state variables and control variables; The second module is used to: construct and train the main dynamics prediction model, and learn the dominant state transition law of the wing flutter system; The third module is used to: calculate the one-step prediction residual based on the main dynamics prediction model, and construct conditional residual modeling samples; The fourth module is used to: construct and train the conditional diffusion residual model, obtain the conditional probability distribution of the predicted residuals and residual mean compensation; The fifth module is used to: construct an online residual risk index based on the local residual distribution characteristics, and normalize the online residual risk index to obtain a normalized risk index; The sixth module is used to: adaptively adjust the control parameters of the stochastic model prediction controller based on the normalized risk index; The seventh module is used to: combine the main dynamics prediction model, the conditional diffusion residual model, and the adjusted control parameters to establish and solve the stochastic model predictive control optimization problem to obtain the optimal control sequence; The eighth module is used to: apply the optimal control sequence to the wing flutter system and continuously update the closed-loop control process to achieve active suppression of wing flutter.

10. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method according to any one of claims 1 to 8.

Images