A multi-task local restart real-time hybrid test method based on a multi-on-off state hybrid proxy model

Abstract

A multi-task local restart real-time hybrid test method based on a multi-on-off state hybrid agent model; model simplification processing will inevitably cause problems such as distortion of structural mechanical properties, boundary constraint deviation and dynamic response solving error, and further form a non-negligible model truncation error and parameter deviation, which directly reduces the reduction accuracy of the overall structure seismic response, and there is no standardized and accurate processing method at present; the present application is to establish a physically consistent constraint Transform network module after offline training of the data set, and sequentially and continuously generate, execute, judge and execute the reset command of the multi-task local loading command according to the Transform network module, simultaneously adopt the UKF parameter identification method, and combine the measured data of the test piece to online update and cyclic iteration of the offline trained LSTM agent model, until the process of outputting the structure response at the discrete time step i = k .

Description

Technical Field

[0001] Specifically, this invention is a real-time hybrid experimental method for multi-task local restart based on a multi-on / off state hybrid agent model. Background Technology

[0002] Real-time hybrid testing, combining physical testing with numerical simulation, has become one of the most effective and economical methods for evaluating the response of structures under dynamic loads such as earthquakes. In real-time hybrid testing, components with complex or uncertain behavior undergo physical testing in the laboratory, while the rest of the structure is simulated numerically. Its main advantage lies in its ability to consider the true nonlinear behavior of physical substructures without relying on potentially flawed simplified analytical models. Existing research has applied real-time hybrid testing to the performance evaluation of various structural systems and devices, including magnetorheological dampers, base-isolated structures, and non-ductile reinforced concrete frames.

[0003] Although real-time hybrid testing has been proven to be an effective scientific testing method in the field of structural seismic testing, its widespread application in engineering still faces many insurmountable fundamental technical bottlenecks. The main challenges lie in the real-time synchronous integration of numerical substructure simulation and physical substructure loading control. Real-time hybrid testing has strict millisecond-level synchronous control requirements for the transmission of force and displacement feedback signals, iterative solutions, and command output. The entire closed-loop calculation process must be perfectly time-matched with the physical loading process. To ensure real-time computational efficiency under dynamic loading, traditional testing methods can only use simplified numerical models with reduced degrees of freedom, idealized boundary conditions, simplified material constitutive models, and reduced nonlinear behavior for solution calculations, significantly reducing efficiency. Reducing the scale of finite element calculations inevitably leads to distortions in structural mechanical properties, deviations in boundary constraints, and errors in dynamic response solutions, resulting in non-negligible model truncation errors and parameter deviations, directly reducing the overall accuracy of seismic response reconstruction. On the other hand, if a refined finite element model, high-precision solid elements, complex contact nonlinearity, and refined damping model are used for simulation analysis, the mesh size, iterative degrees of freedom, and nonlinear convergence solution will significantly increase the computation time, easily leading to problems such as numerical solution lag, displacement and internal force iteration delays, and timing asynchrony. This cannot meet the real-time closed-loop control requirements of the entire dynamic loading process, thus creating an inherent technical contradiction that is difficult to reconcile between the accuracy of refined structural simulation and the efficiency of real-time dynamic response solution. Meanwhile, when conducting split tests on large-span, high-rise, and complex overall structures, it is often necessary to divide multiple physical test substructures to participate in collaborative loading. There are displacement coordination, internal force balance, and boundary deformation coupling relationships between the substructures. This places extremely high demands on the synchronous loading accuracy of actuators, the time delay consistency of servo control systems, and the closed-loop coordination of multi-channel force and displacement. Under complex working conditions, problems such as asynchronous loading channels, phase deviation, non-coordination of boundary deformation, and unbalanced internal forces at the interface are prone to occur. Long-distance collaborative loading across sites and laboratories will further amplify the engineering implementation difficulties of signal transmission delay, differences in software and hardware interface adaptation, and timing deviations of control systems. It places extremely stringent operational requirements on the accuracy of loading equipment, distributed collaborative control algorithms, and integrated software and hardware communication architecture. The on-site test setup is difficult, the system debugging cycle is long, and the stability is difficult to guarantee. The above-mentioned contradictions in timing control mechanisms, inherent conflicts in model accuracy, and challenges in multi-substructure collaborative control jointly restrict the development of real-time hybrid test theory and the practical application of seismic testing scenarios for engineering structures.

[0004] For large or complex structures containing a large number of test substructures, the test time costs are almost unacceptable, and the test efficiency and reliability cannot meet the engineering requirements. In short, the large number of test substructures in large or complex structures leads to too many multi-task loading times, making it difficult to guarantee test efficiency and reliability, and there is no corresponding standard and precise handling method. Summary of the Invention

[0005] This invention provides a real-time hybrid test method for multi-task local restart based on a multi-on / off state hybrid agent model to solve the above problems.

[0006] A real-time hybrid experimental method for multi-task local restart based on a multi-on / off state hybrid proxy model is proposed. The method involves training the dataset offline and establishing a Transformer network module with physical consistency constraints. Based on the Transformer network module, multi-task local loading command generation, loading command execution, accuracy determination, and reset command execution are performed sequentially. At the same time, the UKF parameter identification method is used, and the offline trained LSTM proxy model is updated and iterated online in combination with the test specimen data until the output of the structural response is reached at discrete time step i=k.

[0007] As a preferred approach, the discrete training process for the LSTM surrogate model involves initializing and training the LSTM surrogate model using historical experimental data or high-fidelity simulation data. The input to the LSTM model is the loading displacement vector. The output is the corresponding restoring force vector. The network structure of the LSTM surrogate model consists of an input gate, a forget gate, an output gate, and cell states, as shown in the formula:

[0008] (1)

[0009] (2)

[0010] (3)

[0011] (4)

[0012] (5)

[0013] (6)

[0014] In equations (1) to (6) above, i is the input gate; f is the forget gate; o is the output gate; c is the cell state; g is the candidate memory; x is the input vector at the current time step; h is the hidden state, which is also the output y at the current time step; t represents element-wise multiplication; w is the weight matrix; b is the bias term; and the subscript t represents the time step.

[0015] As a preferred approach: the dataset consists of diverse historical datasets containing uncertainties in substructure parameters, test noise, and system time delays. The Transformer network module is built by training the dataset offline to form a surrogate model that conforms to embedded physical consistency constraints. The input to the Transformer network module is the measured restoring force vector. With earthquake excitation vector The output is the structural displacement response vector. ;

[0016] The surrogate model with embedded physical consistency constraints is a surrogate model that introduces both kinematic and dynamic physical constraints into the loss function. The calculation process consists of the following steps:

[0017] Acceleration is calculated using the central difference method, and the formula is as follows:

[0018] (7)

[0019] The displacement prediction error loss function is L1, and the calculation formula is:

[0020] (8)

[0021] The acceleration prediction error loss function is L2, and the calculation formula is:

[0022] (9)

[0023] The formula for calculating the normalized restoring force reference value is:

[0024] (10)

[0025] The loss function is L3, where L3 represents the normalized restoring power predicted by the network. The result calculated based on formula (10) The normalized error between them is calculated using the following formula:

[0026] (11)

[0027] The total loss function is defined as the weighted sum of all loss terms, and the formula for its calculation is as follows:

[0028] (12)

[0029] In the above formula, These are the mass matrix and damping matrix of the structure, respectively. These are the acceleration, velocity, and displacement vectors of the numerical substructure at the i-th integration step; It is the static restoring force vector of the numerical substructure at the i-th integration step; Δt is the time step; It is the seismic acceleration excitation of the i-th integration step; These are the weighting coefficients for each loss term, used to balance the importance of displacement prediction accuracy, kinematic consistency, and dynamic physical constraints in model training; the superscript t represents the reference solution, np represents the neural network prediction solution, and pp represents the physical constraint prediction solution; the superscript ~ represents the normalized solution; the subscript E represents the experimental substructure, and i represents the integration step;

[0030] The process of embedding physical consistency constraints into the loss function of the training process is completed by calculating equations (7) to (12) above.

[0031] As a preferred solution: The structural displacement response vector output by the Transformer network module... After saving and processing the data, the displacement vector is loaded into the input of the LSTM surrogate model. The process combines the generation of multi-task local loading commands, and performs two types of synchronous processing based on these commands. One type of processing is multi-task local restart loading processing, which generates multiple sets of local loading commands based on the displacement response calculated from the numerical substructure. Multiple sets of local loading commands Executed sequentially by the servo loading system After each loading cycle is completed, the precision control module compares the measured displacement vectors of adjacent loading steps. Consistency checks are performed. If the error exceeds a preset threshold, the number of local loading steps is increased and loading is re-executed until the accuracy requirements are met. s is not equal to 1 or 2, and the minimum value of s is 3. Another type of processing is LSTM parameter update and prediction processing, which involves updating and predicting the measured displacement vector of the experimental substructure. With measured force vector The feedback is sent to the LSTM proxy model, and the network parameters of the LSTM proxy model are updated online using UKF. The updated LSTM model then executes the next set of loading commands. The prediction is performed, and the system enters a waiting state for the next iteration, with the waiting command consisting of the initial displacement vector. After the LSTM parameter update and prediction processing are completed, the Transformer surrogate model prediction processing is performed. The Transformer surrogate model prediction processing process involves processing multiple measured force vectors. LSTM predicts the restoring force vector Combined into a complete experimental substructure restoring force matrix The latest step restoring force vector With seismic force The structural displacement response at the next integration time step is calculated by inputting the data into the Transformer surrogate model with physical consistency constraints. .

[0032] As a preferred solution: The Transformer network module continuously processes data through multi-task local loading commands, executing loading commands, accuracy determination, and executing reset commands. It also utilizes UKF parameter identification and online updates of the measured data of the offline-trained LSTM surrogate model, iterating until the output structural response is obtained at discrete time step i=k. The process involves synchronously repeating these two types of processing based on the multi-task local loading commands until all loading and prediction tasks for all integral time steps are completed, ultimately obtaining the complete structural time history response.

[0033] As a preferred approach, UKF is used to update the LSTM network parameters online during the LSTM parameter update and prediction process. This online update involves constructing a state vector from the LSTM network's weight matrix and bias terms. A nonlinear state-space model is established, and the calculation formula is as follows:

[0034] (13)

[0035] (14)

[0036] In the above formula, For process noise, To measure noise, This is the nonlinear mapping function of the LSTM network.

[0037] As a preferred option, the Transformer surrogate model employs an encoder architecture, including a multi-head self-attention mechanism and a feedforward neural network. The input sequence of the Transformer surrogate model is composed of a restoring force matrix. Latest step recovery force vector With earthquake excitation vector The structure is composed of components, and the output is the structural displacement response vector for the next time step. .

[0038] As a preferred option, the real-time hybrid experimental method for multi-task local restart based on a multi-on / off state hybrid proxy model further includes a collaborative processing procedure between the Transformer proxy model and the online LSTM proxy model. The collaborative processing procedure between the Transformer proxy model and the online LSTM proxy model is as follows:

[0039] First, at each integration time step, the Transformer agent model is based on the restoring force matrix of the current combination. With earthquake excitation Predicting structural displacement response ;

[0040] Next, local restart loading vectors for multiple experimental substructures are generated. A local restart loading vector for loading a small number of specimens by a servo loading system. The measured data of the Transformer proxy model is used for UKF to update the LSTM parameters online, ensuring that the LSTM proxy model is in a state of continuously approximating the real mechanical behavior.

[0041] Then the LSTM surrogate model receives the local restart loading vectors of 80-90% of the total number of samples. Post-output predicted restoring force matrix ;

[0042] Finally, the restoring force matrix was measured. With LSTM predicting the restoring force matrix Combined into a new restoring force matrix Then the latest step's restoring force vector It can be used as the input vector for the next time step of the Transformer proxy model.

[0043] Compared with existing technologies, this invention provides a real-time hybrid experimental method for multi-task local restart based on a multi-on / off state hybrid agent model, which has the following beneficial effects:

[0044] This invention utilizes the Transformer network module to continuously process data through multi-task local loading commands, execution of loading commands, accuracy determination, and execution of reset commands. It also incorporates UKF parameter identification and online iterative updates based on measured data from the offline-trained LSTM surrogate model until the output structural response is reached at discrete time step i=k. This directly saves experimental time costs while simultaneously meeting experimental efficiency and reliability requirements. It is suitable for real-time hybrid experiments involving numerous substructures in large or complex structures. The implementation process is standardized and precise. Specific advantages include:

[0045] This invention is the first to propose introducing a Transformer proxy model with embedded physical consistency constraints into a real-time hybrid experiment based on restart loading technology, which is used to replace traditional finite element calculation as the solver for numerical substructures, thereby improving the computational efficiency of numerical substructures and ensuring that the refined model meets real-time requirements.

[0046] In the process of training the Transformer surrogate model, this invention introduces kinematic and dynamic dual physical constraints to construct a loss function to address the uncertainties of substructure parameters, measured noise, and system time delays. This ensures that the Transformer surrogate model can accurately capture the nonlinear dynamic response of the structure while maintaining high computational efficiency, thereby effectively improving experimental accuracy.

[0047] This invention uses an LSTM proxy model to replace a large number of specimens for multi-task local restart loading, and only performs physical experiments on a few representative specimens, which greatly reduces the number of multi-task loadings and the experimental time cost, and improves experimental efficiency and reliability.

[0048] Based on a small amount of measured data from physically loaded specimens, this invention uses UKF to update the parameters of the LSTM surrogate model online, ensuring that the surrogate model can adapt to the real nonlinear behavior of the experimental substructure in real time. While ensuring prediction accuracy, it effectively suppresses error accumulation and makes the fidelity of the experimental results as close as possible to the real behavior. Attached Figure Description

[0049] Figure 1 This is a schematic diagram of the process of the present invention;

[0050] Figure 2 This is a diagram of the LSTM module architecture in this invention;

[0051] Figure 3 This is a diagram of the Transformer module architecture in this invention;

[0052] Figure 4 This is a schematic diagram illustrating the implementation principle of the present invention when used in a high-rise frame vibration reduction / seismic isolation structure equipped with ten dampers;

[0053] Figure 5 This is a schematic diagram illustrating the implementation principle of the present invention when used in a high-rise frame vibration reduction / seismic structure equipped with eight seismic isolation bearings. Detailed Implementation

[0054] 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.

[0055] Specific implementation method one: Combining Figures 1 to 5 This embodiment describes a real-time hybrid experimental method for multi-task local restart based on a multi-on / off state hybrid proxy model. The method involves training the dataset offline and establishing a Transformer network module with physical consistency constraints. The Transformer network module continuously processes data through multi-task local loading commands, executing loading commands, accuracy determination, and executing reset commands. Simultaneously, it utilizes UKF parameter identification and online updates of the measured data of the offline-trained LSTM proxy model, iterating until the output structural response is reached at discrete time step i=k. The multi-on / off states include online and offline states.

[0056] In the above process, the LSTM surrogate model is trained before use. The discrete training process of the LSTM surrogate model involves initializing and training the LSTM surrogate model using historical experimental data or high-fidelity simulation data. The input of the LSTM model is the loading displacement vector. The output is the corresponding restoring force vector. The network structure of the LSTM surrogate model consists of an input gate, a forget gate, an output gate, and cell states, as shown in the formula:

[0057] (1)

[0058] (2)

[0059] (3)

[0060] (4)

[0061] (5)

[0062] (6)

[0063] In equations (1) to (6) above, i is the input gate; f is the forget gate; o is the output gate; c is the cell state; g is the candidate memory; x is the input vector at the current time step; h is the hidden state, which is also the output y at the current time step; t represents element-wise multiplication; w is the weight matrix; b is the bias term; and the subscript t represents the time step.

[0064] In addition, the dataset is created before use, specifically a diverse historical dataset containing uncertainties in substructure parameters, test noise, and system time delays. The Transformer network module is built by training the dataset offline to form a surrogate model that fits the embedded physical consistency constraints. The input to the Transformer network module is the measured restoring force vector. With earthquake excitation vector The output is the structural displacement response vector. ;

[0065] The surrogate model with embedded physical consistency constraints is a surrogate model that introduces both kinematic and dynamic physical constraints into the loss function. The calculation process consists of the following steps:

[0066] Acceleration is calculated using the central difference method, and the formula is as follows:

[0067] (7)

[0068] The displacement prediction error loss function is L1, and the calculation formula is:

[0069] (8)

[0070] The acceleration prediction error loss function is L2, and the calculation formula is:

[0071] (9)

[0072] The formula for calculating the normalized restoring force reference value is:

[0073] (10)

[0074] The loss function is L3, where L3 represents the normalized restoring power predicted by the network. The result calculated based on formula (10) The normalized error between them is calculated using the following formula:

[0075] (11)

[0076] The total loss function is defined as the weighted sum of all loss terms, and the formula for its calculation is as follows:

[0077] (12)

[0078] In the above formula, These are the mass matrix and damping matrix of the structure, respectively. These are the acceleration, velocity, and displacement vectors of the numerical substructure at the i-th integration step; It is the static restoring force vector of the numerical substructure at the i-th integration step; Δt is the time step; It is the seismic acceleration excitation of the i-th integration step; These are the weighting coefficients for each loss term, used to balance the importance of displacement prediction accuracy, kinematic consistency, and dynamic physical constraints in model training; the superscript t represents the reference solution, np represents the neural network prediction solution, and pp represents the physical constraint prediction solution; the superscript ~ represents the normalized solution; the subscript E represents the experimental substructure, and i represents the integration step;

[0079] The process of embedding physical consistency constraints into the loss function of the training process is completed by calculating equations (7) to (12) above.

[0080] In this embodiment, the Transformer proxy model adopts an encoder architecture. The Transformer proxy model includes a multi-head self-attention mechanism and a feedforward neural network. The input sequence of the Transformer proxy model is composed of the restoring force matrix. Latest step recovery force vector With earthquake excitation vector The structure is composed of components, and the output is the structural displacement response vector for the next time step. .

[0081] Specific Implementation Method Two: This implementation method further defines Specific Implementation Method One. Specifically, it is a multi-task local restart real-time hybrid experimental method based on an off-line / online hybrid proxy model. The main principle of this method is as follows: First, based on historical data containing different substructure parameters, test noise, and various time delays, a Transformer proxy model with embedded physical consistency constraints is trained offline to ensure it can replace traditional finite element calculations as a solver for numerical substructures. Second, during the real-time hybrid experiment, only a few representative specimens are physically loaded; the remaining specimens are replaced by LSTM proxy models pre-trained with historical data, and the LSTM model parameters are updated online using UKF based on the measured data of the physically loaded specimens. Finally, the trained Transformer, a small number of physical specimens, and a large number of online-updated LSTM proxy models work together to construct a refined local restart real-time hybrid experimental system. The Transformer completes the numerical substructure calculations, and the online LSTM proxy models provide the boundary conditions for the experimental substructures.

[0082] In this embodiment, the number of representative specimens subjected to physical loading accounts for 10-20% of the total number of specimens.

[0083] Specific Implementation Method Three: This implementation method is a further limitation of Specific Implementation Method One or Two. In this implementation method, the structural displacement response vector output by the Transformer network module is... After saving and processing the data, the displacement vector is loaded into the input of the LSTM surrogate model. The process combines the generation of multi-task local loading commands, and performs two types of synchronous processing based on these commands. One type of processing is multi-task local restart loading processing, which generates multiple sets of local loading commands based on the displacement response calculated from the numerical substructure. Multiple sets of local loading commands Executed sequentially by the servo loading system After each loading cycle is completed, the precision control module compares the measured displacement vectors of adjacent loading steps. Consistency checks are performed. If the error exceeds a preset threshold, the number of local loading steps is increased and loading is re-executed until the accuracy requirements are met. The parameter s is not set to 1 or 2, and its minimum value is 3. Another type of processing involves LSTM parameter updating and prediction. LSTM parameter updating and prediction involves updating the measured displacement vector of the experimental substructure. With measured force vector The feedback is sent to the LSTM proxy model, and the network parameters of the LSTM proxy model are updated online using UKF. The updated LSTM model then executes the next set of loading commands. The prediction is performed, and the system enters a waiting state for the next iteration, with the waiting command consisting of the initial displacement vector. After the LSTM parameter update and prediction processing are completed, the Transformer surrogate model prediction processing is performed. The Transformer surrogate model prediction processing process involves processing multiple measured force vectors. LSTM predicts the restoring force vector Combined into a complete experimental substructure restoring force matrix The latest step restoring force vector With seismic force The structural displacement response at the next integration time step is calculated by inputting the data into the Transformer surrogate model with physical consistency constraints. .

[0084] In the above process, the Transformer network module continuously processes the data through multi-task local loading commands, executing loading commands, accuracy determination, and executing reset commands. It also utilizes UKF parameter identification and online updates of the measured data of the offline-trained LSTM surrogate model, iterating until the output structural response is obtained at discrete time step i=k. The process involves synchronously repeating two types of processing based on the multi-task local loading commands until all loading and prediction tasks for all integral time steps are completed, ultimately obtaining the complete structural time history response.

[0085] Specific Implementation Method Four: This implementation method is a further limitation of Specific Implementation Methods One, Two, or Three. In this implementation method, during the LSTM parameter update and prediction process, UKF is used to perform online update processing of the LSTM network parameters. The online update processing is to construct a state vector using the weight matrix and bias terms of the LSTM network. A nonlinear state-space model is established, and the calculation formula is as follows:

[0086] (13)

[0087] (14)

[0088] In the above formula, For process noise, To measure noise, This is the nonlinear mapping function of the LSTM network.

[0089] Specific Implementation Method Five: This implementation method is a further limitation of Specific Implementation Methods One, Two, Three, or Four. In this implementation method, the real-time hybrid experimental method for multi-task local restart based on a multi-on / off state hybrid proxy model further includes a collaborative processing procedure between the Transformer proxy model and the online LSTM proxy model. The collaborative processing procedure between the Transformer proxy model and the online LSTM proxy model is as follows:

[0090] First, at each integration time step, the Transformer agent model is based on the restoring force matrix of the current combination. With earthquake excitation Predicting structural displacement response ;

[0091] Next, local restart loading vectors for multiple experimental substructures are generated. A local restart loading vector for loading a small number of specimens by a servo loading system. The measured data of the Transformer proxy model is used for UKF to update the LSTM parameters online, ensuring that the LSTM proxy model is in a state of continuously approximating the real mechanical behavior.

[0092] Then the LSTM surrogate model receives the local restart loading vectors of 80-90% of the total number of samples. Post-output predicted restoring force matrix ;

[0093] Finally, the restoring force matrix was measured. With LSTM predicting the restoring force matrix Combined into a new restoring force matrix Then the latest step's restoring force vector It can be used as the input vector for the next time step of the Transformer proxy model.

[0094] Specific Implementation Method Six: This implementation method is a further limitation of Specific Implementation Methods One, Two, Three, Four, or Five. Specifically, the real-time hybrid experimental method for multi-task local restart based on a multi-on / off state hybrid proxy model is a real-time hybrid experimental method for multi-task local restart based on a Transformer and online LSTM proxy model. The specific steps for its implementation are as follows:

[0095] S1. Offline Training Phase of the Transformer Proxy Model. A refined finite element model is established, generating diverse historical datasets that include uncertainties in substructure parameters, test noise, and system time delays. Based on this, the Transformer proxy model with embedded physical consistency constraints is trained offline, with the measured restoring force vector as input. With earthquake excitation vector The output is the structural displacement response vector. .

[0096] S2. Offline training phase of the LSTM surrogate model. The LSTM surrogate model is initialized and trained using historical experimental data or high-fidelity simulation data. The input to the LSTM model is the loading displacement vector. The output is the corresponding restoring force vector. Its network structure consists of an input gate, a forget gate, an output gate, and cell states, and its specific mathematical expression is as follows:

[0097] (1)

[0098] (2)

[0099] (3)

[0100] (4)

[0101] (5)

[0102] (6)

[0103] Where i is the input gate; f is the forget gate; o is the output gate; c is the cell state; g is the candidate memory; x is the input vector at the current time step; and h is the hidden state, which is also the output y at the current time step. t represents element-wise multiplication; w is the weight matrix; b is the bias term; and the subscript t represents the time step.

[0104] S3. Multi-task local restart loading phase. Based on the displacement response calculated by the Transformer proxy model, multiple sets of local loading commands are generated. Where 2≤s≤9, s is the specimen number, and the loading is performed sequentially by the servo loading system. After each loading cycle, the precision control module compares the measured displacement vectors from adjacent loading steps. Consistency is ensured. If the error exceeds a preset threshold, the number of local loading steps is increased and loading is re-executed until the accuracy requirements are met.

[0105] In this step, the 2nd to 9th dampers, i.e., when 2≤s≤9, are replaced by LSTM proxy models and no physical loading is performed.

[0106] S4, LSTM parameter update and prediction stage. The measured displacement vectors of the 1st and 10th dampers are... With measured force vector The feedback is fed back to the LSTM model, and its network parameters are updated online using UKF. The value of 's' ranges from 2 to 9, meaning the updated LSTM model immediately applies the loading commands for the 2nd to 9th dampers. Perform restoring force prediction, and simultaneously load the system into a waiting state, where the waiting command consists of the initial displacement vector.

[0107] S5, Transformer surrogate model prediction stage. Multiple measured force vectors are used in this stage. LSTM predicts the restoring force vector Combined into a complete experimental substructure restoring force matrix The latest step restoring force vector With seismic force The structural displacement response at the next integration time step is calculated by inputting the data into the Transformer surrogate model with physical consistency constraints. .

[0108] S6, repeat S3 to S5 until all loading and prediction tasks for all integration time steps are completed, and finally obtain the complete structural time history response.

[0109] Specific implementation method seven: Combination Figures 1 to 4 This implementation method describes a real-time hybrid test method for multi-task local restart based on a multi-on / off state hybrid proxy model. The method is implemented using a high-rise frame vibration reduction structure with ten dampers. The prototype structure is a high-rise frame vibration reduction / seismic structure, with the high-rise frame as the numerical substructure. A Transformer proxy model with embedded physical consistency constraints is established. The first and tenth dampers are taken as the test substructures. A servo loading system is used to sequentially perform multi-task local restart loading to achieve a precise and standardized real-time hybrid test process. The real-time hybrid test method for multi-task local restart specifically includes the following steps:

[0110] S1. Offline Training Phase of the Transformer Proxy Model. A refined finite element model of the frame vibration reduction / seismic structure is established, generating diverse historical datasets that include uncertainties in structural modeling parameters, test noise, and system time delays. Based on this, the Transformer proxy model with embedded physical consistency constraints is trained offline, with the measured restoring force vector as its input. With earthquake excitation vector The output is the structural displacement response vector. .

[0111] S2. Offline training phase of the LSTM surrogate model. Offline experiments are conducted on the damper to establish a historical test database, and the LSTM surrogate model is initialized and trained. The input to the LSTM model is the loading displacement vector. The output is the corresponding restoring force vector. Its network structure consists of an input gate, a forget gate, an output gate, and cell states, and its specific mathematical expression is as follows:

[0112] (15)

[0113] (16)

[0114] (17)

[0115] (18)

[0116] (19)

[0117] (20)

[0118] Where i is the input gate; f is the forget gate; o is the output gate; c is the cell state; g is the candidate memory; x is the input vector at the current time step; and h is the hidden state, which is also the output y at the current time step. t represents element-wise multiplication; w is the weight matrix; b is the bias term; and the subscript t represents the time step.

[0119] S3. Multi-task local restart loading stage. The value of 's' ranges from 2 to 9, where 's' is the specimen number. Based on the displacement response calculated using the Transformer surrogate model, multiple sets of local loading commands are generated. And executed sequentially by the server loading system. After each loading cycle, the precision control module compares the measured displacement vectors from adjacent loading steps. Consistency is ensured. If the error exceeds a preset threshold, the number of local loading steps is increased and loading is re-executed until the accuracy requirements are met. In this embodiment, the value of s ranges from 2 to 9, meaning that the 2nd to 9th dampers are replaced by LSTM surrogate models, eliminating the need for physical loading.

[0120] S4, LSTM parameter update and prediction stage. The measured displacement vectors of the 1st and 10th dampers are... With measured force vector The feedback is fed back to the LSTM model, and its network parameters are updated online using UKF. The value of 's' ranges from 2 to 9. The updated LSTM model then applies the loading commands for the 2nd to 9th dampers. Perform restoring force prediction, and simultaneously load the system into a waiting state, where the waiting command consists of the initial displacement vector.

[0121] S5, Transformer surrogate model prediction stage. This involves the 1st and 10th dampers. LSTM predicts the restoring force vector Combined into a complete damper restoring force matrix The latest step restoring force vector With seismic force The structural displacement response at the next integration time step is calculated by inputting the data into the Transformer surrogate model with physical consistency constraints. .

[0122] S6. Repeat steps S3 to S5 until all loading and prediction tasks for all integration time steps are completed, ultimately obtaining the complete structural time history response. Other unmentioned content and implementation steps are the same as in specific implementation methods one, two, three, four, five, or six.

[0123] Specific implementation method eight: Combination Figure 1 , Figure 2 , Figure 3 and Figure 5 This implementation method describes a real-time hybrid test method based on a multi-task local restart based on a hybrid proxy model. The method uses a high-rise frame vibration reduction structure with eight seismic isolation bearings as the implementation object. The prototype structure is a high-rise frame vibration reduction / seismic structure, with the high-rise frame as the numerical substructure. A Transformer proxy model with embedded physical consistency constraints is established. The first and second seismic isolation bearings are taken as the test substructure. A servo loading system is used to sequentially perform multi-task local restart loading to achieve a precise and standardized real-time hybrid test process. The real-time hybrid test method for multi-task local restart specifically includes the following steps:

[0124] S1. Offline Training Phase of the Transformer Proxy Model. A refined finite element model of the frame vibration reduction / seismic structure is established, generating diverse historical datasets that include uncertainties in structural modeling parameters, test noise, and system time delays. Based on this, the Transformer proxy model with embedded physical consistency constraints is trained offline, with the measured restoring force vector as its input. With earthquake excitation vector The output is the structural displacement response vector. .

[0125] S2. Offline training phase of the LSTM surrogate model. Offline experiments were conducted on the seismic isolation bearings to establish a historical test database, and the LSTM surrogate model was initialized and trained. The input to the LSTM model is the loading displacement vector. The output is the corresponding restoring force vector. Its network structure consists of an input gate, a forget gate, an output gate, and cell states, and its specific mathematical expression is as follows:

[0126] (twenty one)

[0127] (twenty two)

[0128] (twenty three)

[0129] (twenty four)

[0130] (25)

[0131] (26)

[0132] Where i is the input gate; f is the forget gate; o is the output gate; c is the cell state; g is the candidate memory; x is the input vector at the current time step; and h is the hidden state, which is also the output y at the current time step. t represents element-wise multiplication; w is the weight matrix; b is the bias term; and the subscript t represents the time step.

[0133] S3. Multi-task local restart loading stage. 's' ranges from 3 to 8, where 's' is the specimen number. Based on the displacement response calculated using the Transformer surrogate model, multiple sets of local loading commands are generated. And executed sequentially by the server loading system. After each loading cycle, the precision control module compares the measured displacement vectors from adjacent loading steps. Consistency is ensured. If the error exceeds a preset threshold, the number of local loading steps is increased and loading is re-executed until the accuracy requirements are met. In this embodiment, the 3rd to 8th seismic isolation bearings are replaced by LSTM surrogate models, eliminating the need for physical loading.

[0134] S4, LSTM parameter update and prediction stage. The measured displacement vectors of the first and second seismic isolation bearings are... With measured force vector The feedback is fed back to the LSTM model, and its network parameters are updated online using UKF. The updated LSTM model then applies the loading commands to the 3rd to 8th seismic isolation bearings. Perform restoring force prediction, and simultaneously load the system into a waiting state, where the waiting command consists of the initial displacement vector.

[0135] S5, Transformer surrogate model prediction stage. The first and second seismic isolation bearings... LSTM predicts the restoring force vector These are combined to form a complete restoring force matrix for seismic isolation bearings. The latest step restoring force vector With seismic force The structural displacement response at the next integration time step is calculated by inputting the data into the Transformer surrogate model with physical consistency constraints. .

[0136] S6. Repeat steps S3 to S5 until all loading and prediction tasks for all integration time steps are completed, ultimately obtaining the complete structural time history response. Other unmentioned content and implementation steps are the same as those in specific implementation methods one, two, three, four, five, or seven.

Claims

1. A real-time hybrid experimental method for multi-task local restart based on a multi-on / off state hybrid proxy model, characterized in that: The multi-task local restart real-time hybrid experimental method involves training the dataset offline and establishing a Transformer network module with physical consistency constraints. Based on the Transformer network module, multi-task local loading command generation, loading command execution, accuracy determination, and reset command execution are performed sequentially and continuously. At the same time, the UKF parameter identification method is used, and the offline trained LSTM surrogate model is updated and iterated online in combination with the test specimen data until the output of the structural response is reached at discrete time step i=k.

2. The real-time hybrid experimental method for multi-task local restart based on a multi-on / off state hybrid proxy model according to claim 1, characterized in that: The process of discrete training of the LSTM surrogate model involves initializing and training the LSTM surrogate model using historical experimental data or high-fidelity simulation data. The input to the LSTM model is the loading displacement vector. The output is the corresponding restoring force vector. The network structure of the LSTM surrogate model consists of an input gate, a forget gate, an output gate, and cell states, as shown in the formula: （1） （2） （3） （4） （5） （6） In equations (1) to (6) above, i is the input gate; f is the forget gate; o is the output gate; c is the cell state; g is the candidate memory; x is the input vector at the current time step; h is the hidden state, which is also the output y at the current time step; This represents element-wise multiplication; w is the weight matrix; b is the bias term; and the subscript t represents the time step.

3. The real-time hybrid experimental method for multi-task local restart based on a multi-on / off state hybrid proxy model according to claim 2, characterized in that: The dataset consists of diverse historical datasets containing uncertainties in substructure parameters, test noise, and system time delays. The Transformer network module is built by training the dataset offline to form a surrogate model that meets the embedded physical consistency constraints. The input to the Transformer network module is the measured restoring force vector. With earthquake excitation vector The output is the structural displacement response vector. ; The surrogate model with embedded physical consistency constraints is a surrogate model that introduces both kinematic and dynamic physical constraints into the loss function. The calculation process consists of the following steps: Acceleration is calculated using the central difference method, and the formula is as follows: （7） The displacement prediction error loss function is L1, and the calculation formula is: （8） The acceleration prediction error loss function is L2, and the calculation formula is: （9） The formula for calculating the normalized restoring force reference value is: （10） The loss function is L3, where L3 represents the normalized restoring power predicted by the network. The result calculated based on formula (10) The normalized error between them is calculated using the following formula: （11） The total loss function is defined as the weighted sum of all loss terms, and the formula for its calculation is as follows: （12） In the above formula, These are the mass matrix and damping matrix of the structure, respectively. These are the acceleration, velocity, and displacement vectors of the numerical substructure at the i-th integration step; It is the static restoring force vector of the numerical substructure at the i-th integration step; Δt is the time step; It is the seismic acceleration excitation of the i-th integration step; These are the weighting coefficients for each loss term, used to balance the importance of displacement prediction accuracy, kinematic consistency, and dynamic physical constraints in model training; the superscript t represents the reference solution, np represents the neural network prediction solution, and pp represents the physical constraint prediction solution; the superscript ~ represents the normalized solution; the subscript E represents the experimental substructure, and i represents the integration step; The process of embedding physical consistency constraints into the loss function of the training process is completed by calculating equations (7) to (12) above.

4. The real-time hybrid experimental method for multi-task local restart based on a multi-on / off state hybrid proxy model according to claim 3, characterized in that: The structural displacement response vector output by the Transformer network module After saving and processing the data, the displacement vector is loaded into the input of the LSTM surrogate model. The process combines the generation of multi-task local loading commands, and performs two types of synchronous processing based on these commands. One type of processing is multi-task local restart loading processing, which generates multiple sets of local loading commands based on the displacement response calculated from the numerical substructure. Multiple sets of local loading commands Executed sequentially by the servo loading system After each loading cycle is completed, the precision control module compares the measured displacement vectors of adjacent loading steps. Consistency checks are performed. If the error exceeds a preset threshold, the number of local loading steps is increased and loading is re-executed until the accuracy requirements are met. s is not equal to 1 or 2, and the minimum value of s is 3. Another type of processing is LSTM parameter update and prediction processing, which involves updating and predicting the measured displacement vector of the experimental substructure. With measured force vector The feedback is sent to the LSTM proxy model, and the network parameters of the LSTM proxy model are updated online using UKF. The updated LSTM model then executes the next set of loading commands. The prediction is performed, and the system enters a waiting state for the next iteration, with the waiting command consisting of the initial displacement vector. After the LSTM parameter update and prediction processing are completed, the Transformer surrogate model prediction processing is performed. The Transformer surrogate model prediction processing process involves processing multiple measured force vectors. LSTM predicts the restoring force vector Combined into a complete experimental substructure restoring force matrix The latest step restoring force vector With seismic force The structural displacement response at the next integration time step is calculated by inputting the data into the Transformer surrogate model with physical consistency constraints. .

5. The real-time hybrid experimental method for multi-task local restart based on a multi-on / off state hybrid proxy model according to claim 4, characterized in that: The Transformer network module continuously processes data through multi-task local loading commands, executing loading commands, accuracy determination, and executing reset commands. It also utilizes UKF parameter recognition and online updates of the measured data of the offline-trained LSTM surrogate model, iterating until the output structural response is obtained at discrete time step i=k. The process involves synchronously repeating two types of processing based on the multi-task local loading commands until all loading and prediction tasks for all integral time steps are completed, ultimately obtaining the complete structural time history response.

6. A real-time hybrid experimental method for multi-task local restart based on a multi-on / off state hybrid proxy model according to any one of claims 1 to 5, characterized in that: In the LSTM parameter update and prediction process, UKF is used to update the LSTM network parameters online. The online update process involves constructing a state vector from the weight matrix and bias terms of the LSTM network. A nonlinear state-space model is established, and the calculation formula is as follows: （13） （14） In the above formula, For process noise, To measure noise, This is the nonlinear mapping function of the LSTM network.

7. The real-time hybrid experimental method for multi-task local restart based on a multi-on / off state hybrid proxy model according to claim 6, characterized in that: The Transformer surrogate model employs an encoder architecture, incorporating a multi-head self-attention mechanism and a feedforward neural network. The input sequence of the Transformer surrogate model consists of a restoring force matrix. Latest step recovery force vector With earthquake excitation vector The structure is composed of components, and the output is the structural displacement response vector for the next time step. .

8. A real-time hybrid experimental method for multi-task local restart based on a multi-on / off state hybrid proxy model according to any one of claims 1 to 5, characterized in that: The real-time hybrid experimental method for multi-task local restart based on a multi-on / off state hybrid proxy model also includes a collaborative processing procedure between the Transformer proxy model and the online LSTM proxy model. The collaborative processing procedure between the Transformer proxy model and the online LSTM proxy model is as follows: First, at each integration time step, the Transformer agent model is based on the restoring force matrix of the current combination. With earthquake excitation Predicting structural displacement response ; Next, local restart loading vectors for multiple experimental substructures are generated. A local restart loading vector for loading a small number of specimens by a servo loading system. The measured data of the Transformer proxy model is used for UKF to update the LSTM parameters online, ensuring that the LSTM proxy model is in a state of continuously approximating the real mechanical behavior. Then the LSTM surrogate model receives the local restart loading vectors of 80-90% of the total number of samples. Post-output predicted restoring force matrix ; Finally, the restoring force matrix was measured. With LSTM predicting the restoring force matrix Combined into a new restoring force matrix Then the latest step's restoring force vector It can be used as the input vector for the next time step of the Transformer proxy model.

Images