Real-time hybrid testing method and system based on parameter identification and deep learning surrogate model

Abstract

The real-time hybrid test method and system based on parameter identification and deep learning agent model belong to the technical field of structure test. In order to solve the problems that the calculation efficiency of the numerical substructure in the existing real-time hybrid test cannot meet the real-time requirement, and the error accumulation of the existing deep learning method for solving the calculation efficiency of the numerical substructure is easy to lead to test failure, the structure response time history under different constitutive parameters is obtained in advance, the deep learning agent model strucDL of the overall structure dynamic response is trained by using different constitutive parameters and the corresponding structure response time history; the constitutive parameters are obtained by parameter identification based on the test substructure restoring force and displacement of the i-th step in the test; the constitutive parameters are sent to strucDL, the subsequent displacement response time history of the test substructure is pre-tested, then the test substructure restoring force of the i+1-th step is obtained by loading, and the above steps are repeated until the test is completed.

Description

Technical Field

[0001] This invention belongs to the field of structural testing technology, specifically relating to a real-time hybrid testing method and system for structures under seismic loading. Background Technology

[0002] Hybrid testing methods combining substructure techniques provide a feasible experimental approach for assessing the seismic performance of large structures. By subjecting only key components to experimental loading while numerically simulating the rest of the structure, hybrid testing methods enable the creation of large-scale or even full-scale specimens.

[0003] When velocity-dependent components are present in the specimen, such as dampers and other energy dissipation and vibration reduction devices commonly found in modern structures, real-time loading of the specimen is required, i.e., real-time hybrid testing. In real-time hybrid testing, velocity-dependent components in the structure are extracted as test substructures for loading. Each loading command originates from the calculation of the numerical substructure, which requires that the numerical calculation be completed within a single time integration step. This requirement for the computational efficiency of the numerical substructure has meant that, despite the increasingly sophisticated theoretical foundation of real-time hybrid testing over the past two decades, it has been difficult to apply to large and complex structures, failing to align with the construction trends of practical engineering structures. Current research progress on the computational efficiency of numerical substructures in real-time hybrid testing can be broadly categorized into four types: simplifying numerical models, increasing the integration step size, improving hardware computational efficiency, and developing new, efficient computational methods. While these methods can improve the computational efficiency of numerical substructures, it is worth noting that the time of real-time hybrid testing is typically on the order of milliseconds or seconds, and in many cases, the single-step time step does not exceed 10 milliseconds. When facing finite element analysis of numerical substructures on a timescale of hours or days, completing the calculation within 10 milliseconds is virtually impossible.

[0004] For hybrid experiments, deep learning methods can provide a technical means to address the computational efficiency problem of numerical substructures. By using deep learning to build surrogate models of numerical substructures, the most time-consuming numerical computation and learning processes can be completed before the experiment, thus resolving the contradiction between computational efficiency and accuracy in real-time hybrid experiments. However, there are still many problems with using deep learning models for real-time hybrid experiments. For example, the displacement response of the numerical substructure in a real-time hybrid experiment is used for loading the experimental substructure, and the restoring force of the experimental substructure measured after loading is used as the input for the next step of the numerical substructure. Therefore, if the restoring force of the experimental substructure is used as the input of the deep learning model, the model prediction error will be propagated backward, leading to error accumulation. Generally, the predicted displacement will diverge after only a few steps, causing the experiment to fail. Summary of the Invention

[0005] This invention aims to address the problem that the computational efficiency of numerical substructures in existing real-time hybrid experiments cannot meet the real-time requirements, and the problem that existing methods using deep learning to improve the computational efficiency of numerical substructures are prone to error accumulation, which can easily lead to experimental failure.

[0006] A real-time hybrid experimentation method based on parameter recognition and deep learning agent models includes the following steps:

[0007] S1. Pre-set a series of constitutive parameter values ​​for the test substructures, establish and utilize the finite element model of the overall structure to obtain the structural response time history under different constitutive parameters;

[0008] S2. Establish a deep learning surrogate model strucDL for the overall structural dynamic response. The input of strucDL includes the constitutive parameters of the experimental substructure, and the output is the displacement response time history of the experimental substructure. Train strucDL using different constitutive parameters and the corresponding structural response time histories to obtain the trained strucDL.

[0009] S3. Using the restoring force and displacement of the test substructure as input, the constitutive parameters of the test substructure are obtained by using an identification algorithm. The process of obtaining the constitutive parameters of the test substructure by using the identification algorithm is called the test substructure constitutive parameter identification module, i.e., the expID module.

[0010] S4. Based on the established overall structural deep learning agent model strucDL and the experimental substructure parameter identification module, conduct real-time hybrid experiments;

[0011] In the experiment, the restoring force and displacement of the experimental substructure in step i are sent to the expID module for parameter identification to obtain the current constitutive parameters;

[0012] The constitutive parameters are then sent to strucDL to predict the displacement response time history of the experimental substructure at step i+1.

[0013] The predicted displacement response time history of step i+1 is sent to the actuator to load the test substructure and obtain the restoring force of the test substructure at step i+1.

[0014] S5. Repeat step S4 until the experiment ends.

[0015] Furthermore, the inputs to strucDL in S2 also include seismic motion.

[0016] Furthermore, the training of strucDL in S2 is performed using an offline learning method.

[0017] Furthermore, in the process of obtaining the constitutive parameters of the experimental substructure using the recognition algorithm in S3, the recognition algorithm used is the Kalman filter algorithm, which includes the unscented Kalman filter algorithm.

[0018] Furthermore, the deep learning agent model is an LSTM neural network model.

[0019] A real-time hybrid experimental system based on parameter recognition and deep learning agent models includes:

[0020] The overall structural dynamic response prediction module, namely the strucDL prediction module, uses the deep learning surrogate model strucDL for overall structural dynamic response to predict the displacement response time history of the experimental substructure. The input of strucDL includes the constitutive parameters of the experimental substructure, and the output is the displacement response time history of the experimental substructure.

[0021] The strucDL is pre-trained, and the training process includes the following steps:

[0022] A series of constitutive parameter values ​​for experimental substructures are preset in advance, and the structural response time history under different constitutive parameters is obtained using the finite element model of the overall structure.

[0023] The strucDL is trained using different constitutive parameters and the corresponding structural response time histories to obtain the trained strucDL.

[0024] The constitutive parameter identification module for the test substructure, namely the expID module, uses the restoring force and displacement of the test substructure as input and employs an identification algorithm to obtain the constitutive parameters of the test substructure.

[0025] Hybrid Test Control Module: Based on the overall structural dynamic response prediction module and the expID module, this module controls the real-time hybrid test until its completion. The process of controlling the real-time hybrid test includes the following steps:

[0026] In the experiment, the restoring force and displacement of the experimental substructure in step i are sent to the expID module for parameter identification to obtain the current constitutive parameters;

[0027] The constitutive parameters are then sent to the strucDL prediction module to predict the subsequent displacement response time history of the experimental substructure.

[0028] The value of step i+1 in the predicted time history is sent to the actuator to load the test substructure, and the restoring force of the test substructure at step i+1 is obtained.

[0029] Furthermore, the real-time hybrid experimental system based on parameter recognition and deep learning agent model also includes a loading system and an actuator. The process by which the hybrid experimental control module sends the value of the (i+1)th step in the predicted time history to the actuator is implemented based on the loading system. That is, the hybrid experimental control module sends the value of the (i+1)th step to the loading system through an interface. The loading system is used to control the actuator, and the actuator is used to perform loading on the experimental substructure.

[0030] Furthermore, the input to the deep learning proxy model strucDL for the overall structural dynamic response in the strucDL prediction module also includes seismic motion.

[0031] Furthermore, the identification algorithm used by the expID module to obtain the constitutive parameters of the experimental substructure is the Kalman filter algorithm, which includes the unscented Kalman filter algorithm.

[0032] Furthermore, the deep learning agent model is an LSTM neural network model.

[0033] Beneficial effects:

[0034] First, this invention uses deep learning to establish a surrogate model of structural dynamic response to conduct real-time hybrid experiments. All time-consuming computational work, including finite element calculations of the structure and training of the deep learning surrogate model, can be completed offline before the experiment. In the online phase, only the trained prediction model needs to be run, which can greatly improve computational efficiency. At the same time, it avoids the problem that the theory of real-time hybrid experiments using very simplified numerical models is difficult to apply to large and complex structures. This invention can be well applied to large and complex structures and can achieve very good experimental results.

[0035] Second, by combining the parameter identification method with the constitutive parameters of the experimental substructure instead of the restoring force as the input of the deep learning model, this invention can effectively avoid the problem of experimental failure caused by the accumulation of deep learning prediction errors. Attached Figure Description

[0036] Figure 1 This is a schematic diagram of a real-time hybrid experimental method based on parameter recognition and a deep learning agent model.

[0037] Figure 2 This is a flowchart of a real-time hybrid experiment method based on parameter recognition and a deep learning agent model.

[0038] Figure 3 This is a schematic diagram of the prototype frame structure with dampers and its substructures.

[0039] Figure 4 This is a schematic diagram illustrating how a dataset can be constructed based on finite element analysis.

[0040] Figure 5 This is a schematic diagram of a neural network architecture (using LSTM as an example) that takes constitutive parameters as input.

[0041] Figure 6 This is a schematic diagram of a real-time hybrid experiment system based on a deep learning agent model. Detailed Implementation

[0042] To address the insufficient computational efficiency of numerical substructures in current real-time hybrid experiments, this invention proposes a real-time hybrid experiment method based on parameter identification and a deep learning surrogate model. Using the constitutive parameters of the experimental substructure and seismic motion as model inputs, and the boundary displacement time histories of the experimental substructure as outputs, a deep learning surrogate model of the overall structural dynamic response is established. At each time step of the real-time hybrid experiment, the constitutive parameters of the experimental substructure are identified online using a parameter identification method, and then input into the deep learning surrogate model of the overall structural dynamic response to obtain or update the subsequent displacement loading time histories of the experimental substructure. Therefore, this invention also effectively solves the problem of error accumulation in deep learning surrogate models that use the restoring force of the experimental substructure as input.

[0043] Specific implementation method one: Combining Figure 1 and Figure 2 This implementation method is described below.

[0044] This implementation method is a real-time hybrid experiment method based on parameter recognition and deep learning proxy model. In this implementation method, the real-time hybrid experiment analyzes the following structural motion equilibrium equation.

[0045] Ma+Cv+R N +R E =-Ma g (1)

[0046] Where M and C represent the mass matrix and damping matrix of the overall structure, respectively; d, v, and a represent the displacement, velocity, and acceleration time histories of the structure, respectively; a g Represents the time history of ground motion acceleration; R N R E These are the restoring force time histories of the numerical substructure and the experimental substructure, respectively, where the subscripts N and E represent the numerical substructure and the experimental substructure, respectively.

[0047] In real-time hybrid tests, the models of the ground motion and the numerical substructure are known, and the constitutive relation type of the test substructure is usually determined. What is unknown is often only the values ​​of the constitutive parameters of the test substructure. Therefore, the restoring force time history of the test substructure can also be expressed as R E (d E ,θ E ), where θ EThe constitutive parameters representing the experimental substructure. Once θ is determined... E Equation (1) can be solved by explicit or implicit time integration methods, or by directly using finite element software to model and solve the overall structure.

[0048] Therefore, this method does not require the separate establishment of a computational model for the numerical substructure; instead, it only establishes a finite element model of the overall structure and a constitutive parameter identification module for the experimental substructure. The real-time hybrid experimental method based on parameter identification and a deep learning proxy model described in this embodiment includes the following steps:

[0049] S1. Based on engineering experience or preliminary component tests, determine the constitutive relation type of the test substructure and estimate the constitutive parameter θ. E The range of values ​​is determined, thereby pre-setting a series of constitutive parameter values ​​for the experimental substructures; a finite element model of the overall structure is established, and the specified seismic motion a is input. g We obtained the structural response time histories under a series of different experimental substructure constitutive parameters.

[0050] S2. Using the data obtained from the finite element method in S1, a deep learning surrogate model (strucDL) is established for offline learning of the overall structural dynamic response. The input of this surrogate model is the constitutive parameter θ of the experimental substructure. E The output is the dynamic response time history of interest (e.g., the displacement response time history d of the degree of freedom of the experimental substructure). E Optionally, to accelerate network training, the input to the surrogate model can also include seismic motion a. g ;

[0051] S3. Establish a constitutive parameter identification module for the test substructure (expID for short). The least squares method, Kalman filter series method or other identification algorithms can be used. The input of the identification module is the restoring force and displacement of the test substructure, and the output is the constitutive parameters of the test substructure.

[0052] S4. Based on the established overall structural deep learning agent model strucDL and the experimental substructure parameter identification module expID, conduct real-time hybrid experiments;

[0053] In the experiment, the restoring force R of the experimental substructure in step i will be... E,i Displacement d E,i The parameter is sent to the expID module for parameter identification to obtain the current constitutive parameter identification value θ. E,i ;

[0054] Then identify the constitutive parameter value θ E,i Send the data to the strucDL model to predict the displacement response time history d of the experimental substructure. E ;

[0055] Predict time history d E The value d at step i+1 E,i+1 The load is sent to the actuator to apply load to the test substructure, and the restoring force R of the test substructure at step i+1 is obtained. E,i+1 ;

[0056] S5. Repeat step S4 until the experiment ends.

[0057] Specific Implementation Method Two: Combined Body Figure 6 This implementation method is described below.

[0058] This embodiment is a real-time hybrid experimental system based on parameter recognition and a deep learning proxy model, including an overall structural dynamic response prediction module, an experimental substructure constitutive parameter identification module, and a hybrid experimental control module. It may also include a loading system and actuators.

[0059] The overall structural dynamic response prediction module, namely the strucDL prediction module, uses the deep learning surrogate model strucDL for overall structural dynamic response to predict the displacement response time history of the test substructure. The input of strucDL includes the constitutive parameters of the test substructure, and the output is the displacement response time history of the test substructure. The input of strucDL can also include ground motion.

[0060] The strucDL is pre-trained, and the training process includes the following steps:

[0061] A series of constitutive parameter values ​​for experimental substructures are preset in advance, and the structural response time history under different constitutive parameters is obtained using the finite element model of the overall structure.

[0062] The strucDL is trained using different constitutive parameters and the corresponding structural response time histories to obtain the trained strucDL.

[0063] The constitutive parameter identification module for the test substructure, namely the expID module, uses the restoring force and displacement of the test substructure as input and employs an identification algorithm to obtain the constitutive parameters of the test substructure.

[0064] Hybrid Test Control Module: Based on the overall structural dynamic response prediction module and the expID module, this module controls the real-time hybrid test until its completion. The process of controlling the real-time hybrid test includes the following steps:

[0065] In the experiment, the restoring force and displacement of the experimental substructure in step i are sent to the expID module for parameter identification to obtain the current constitutive parameters;

[0066] The constitutive parameters are then sent to the strucDL prediction module to predict the subsequent displacement response time history of the experimental substructure.

[0067] The value of step i+1 in the predicted time history is sent to the actuator to load the test substructure and obtain the restoring force of the test substructure at step i+1.

[0068] Loading system and actuator: The process by which the hybrid test control module sends the value of step i+1 in the predicted time history to the actuator is based on the loading system. That is, the hybrid test control module sends the value of step i+1 to the loading system through an interface. The loading system is used to control the actuator, and the actuator is used to perform loading on the test substructure.

[0069] The Kalman filter algorithm is used in the process of obtaining the constitutive parameters of the experimental substructure by the expID module. The Kalman filter algorithm includes the unscented Kalman filter algorithm. In other embodiments, the least squares method or other recognition algorithms may also be used.

[0070] The deep learning agent model described is an LSTM neural network model; in other embodiments, other neural network models may also be used for the deep learning agent model.

[0071] Compared with existing deep learning-based real-time hybrid experiment methods, the main advantages of this invention are:

[0072] (1) This invention does not require the separate establishment of a computational model for the numerical substructure, but only establishes a computational model for the dynamic response of the overall structure. Its advantage lies in solving the problem of incomplete boundary conditions for the experimental substructure and improving experimental accuracy. Due to limitations imposed by laboratory loading conditions, the boundary conditions of the experimental substructure often cannot be fully applied; that is, only displacements and forces on some boundary degrees of freedom can be achieved, while other boundary degrees of freedom are ignored. If the traditional method of separating the numerical and experimental substructures is used to establish a computational model for the numerical substructure, the displacements of the numerical and experimental substructures will be inconsistent on the ignored boundary degrees of freedom, thus introducing errors into the mixed experiment. However, by using the method proposed in this invention to establish a computational model for the overall structure, those degrees of freedom ignored during experimental loading can be considered in the computational model, thereby reducing the impact of incomplete boundary conditions for the experimental substructure.

[0073] (2) In this invention, the calculation model for the overall structural dynamic response adopts a deep learning surrogate model, and the input of the surrogate model is the constitutive parameters of the experimental substructure, rather than the restoring force of the experimental substructure. The advantage is that it can solve the error accumulation problem encountered when the surrogate model performs multi-step predictions, significantly improving experimental accuracy. This is because in real-time hybrid experiments, numerical calculation and experimental loading are interactive. The displacement obtained by the numerical calculation model is used for loading the experimental substructure, and the restoring force of the experimental substructure measured after loading participates in the next step of the numerical model's calculation. If the restoring force of the experimental substructure is directly used as the input of the deep learning surrogate model, the model prediction error will be propagated backward. That is, if there is an error in the single-step predicted displacement of the experimental substructure, this error will lead to an error in the restoring force of that step through experimental loading. Inputting this erroneous restoring force into the deep learning surrogate model will result in a larger error in the next step's predicted displacement of the experimental substructure. Therefore, there is a problem of multi-step error accumulation. Generally, after only a few steps, the predicted displacement of the surrogate model will diverge, causing experimental failure. The method proposed in this invention uses the constitutive parameters of the experimental substructure as input to the surrogate model. Even if the predicted displacement has errors, it will not affect the identification of the constitutive parameters, and the errors will not be propagated backward, thus avoiding the problem of error accumulation.

[0074] (3) In this invention, the output of the deep learning proxy model at each step is the entire displacement time history of the experimental substructure, rather than the displacement of a single time step. This has the advantage of not restricting the proxy model's prediction to be completed within a single time step, ensuring the real-time performance of the experimental loading; and it allows the use of larger deep learning models with better prediction performance, making it suitable for real-time hybrid experiments of large and complex structures. If each step can only predict the displacement of a single time step, the deep learning proxy model must complete the prediction within a single time step (usually on the order of milliseconds). This is feasible for small deep learning models with fewer layers and nodes (such as two-layer LSTM), but cannot guarantee the real-time performance of large deep learning models with more layers and nodes (such as ResNet). The computational models of actual engineering structures are often quite complex. However, using the method proposed in this invention, the entire displacement time history of the experimental substructure can be predicted each time the deep learning proxy model is called. Even if the prediction time exceeds the time step length, the previously predicted displacement can be used for loading, ensuring the smooth progress of the experiment.

[0075] Example:

[0076] The following is combined Figures 3-5 The technical solutions in the embodiments of the present invention are described below. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. 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.

[0077] This embodiment illustrates the basic principles and key technologies of the method of the present invention through a real-time hybrid test of a frame structure with a damper. The prototype structure is a two-story frame structure, with a damper installed on one floor, such as... Figure 3 As shown. The damper of one layer was used as the test substructure for real-time experimental loading, while the remaining frame structure was used as the numerical substructure for numerical simulation. Online data interaction was conducted during the experiment, enabling real-time hybrid testing. The substructure was divided as follows: Figure 3 As shown. Since real-time hybrid experiments require the calculation of numerical substructures to be completed within a single time integration step, it is almost impossible to perform high-precision finite element analysis during the experiment. Therefore, the deep learning surrogate model method proposed in this invention is needed.

[0078] This embodiment of conducting a real-time hybrid experiment based on parameter recognition and a deep learning agent model specifically includes the following steps:

[0079] (1) Using the constitutive parameter θ of the experimental substructure E Earthquake a g As model input, the boundary displacement time histories d of the numerical substructure and the experimental substructure are used. E As output, a deep learning proxy model of the overall structure is trained. Specifically, this includes the following steps:

[0080] a) Constructing the training dataset: Establish, for example... Figure 3 The finite element model of the overall structure is shown, where the numerical substructure is a two-layer frame and the experimental substructure is a damper. The constitutive model of the experimental substructure adopts a bilinear model, assuming a series of different initial stiffnesses E and yield strengths f. y Using the strain hardening ratio b, a series of displacement response time histories of the test substructure under a given seismic motion were calculated. The constitutive model parameters θ of the test substructure were then used. E =(E,f y b) Earthquake motion a g As input, the time history of the displacement response of the experimental substructure d E As output, a training dataset is constructed. The dataset construction method based on finite element calculation is as follows: Figure 4 As shown.

[0081] b) Design of the neural network architecture as a deep learning proxy model: In this embodiment, an LSTM neural network model suitable for processing time-series data is selected. The constitutive model parameters of the experimental substructure and the seismic motion are used as inputs, and the displacement response time history of the experimental substructure is used as the output. The neural network architecture consists of two LSTM layers and two fully connected layers, with 50 nodes in each layer. The designed network architecture is as follows: Figure 5 As shown.

[0082] c) Train and deploy the deep learning proxy model: Train the aforementioned network to obtain a deep learning proxy model of the overall structural dynamic response, and deploy it to the host computer of the real-time hybrid experiment. The deep learning proxy model can be deployed via LabVIEW. The experimental system design is as follows: Figure 6 As shown.

[0083] (2) Conduct real-time hybrid experiments, in which the overall structure adopts the deep learning agent model established in step (1). Specifically, it includes:

[0084] A) Overall structural pre-calculation: First, estimate the initial stiffness E0 and yield strength f of the assumed damper. y0 The strain hardening ratio b0 was used to preliminarily predict the displacement time history of the experimental substructure for the first 800 time steps using a deep learning surrogate model.

[0085] B) Preloading of the test substructure: The test substructure is preloaded using the 800 displacement time histories obtained from the preliminary prediction. Preloading can be stopped when the test substructure begins to enter nonlinearity.

[0086] C) Parameter identification during the preloading stage of the test substructure: During the loading process, the restoring force and displacement of the test substructure are tested at each step. In this embodiment, the unscented Kalman filter (UKF) method is used to identify constitutive parameters and obtain the constitutive parameters θ of the test substructure. E,i .

[0087] The preloading described in step C) can also be performed on the identified θ. E,i Stop when it tends to stabilize.

[0088] D) Deep learning predictions for subsequent loading stages of the experimental substructure: After preloading, the experiment is not paused, but immediately uses the θ identified in the last step of preloading. E,i Using seismic motion as input, a deep learning model is used to predict the subsequent displacement time history of the experimental substructure, which is then used for loading.

[0089] E) Parameter identification in subsequent loading stages of the experimental substructure: In the subsequent loading stages, constitutive parameters are still identified at each step using the measured restoring force and displacement of the experimental substructure to obtain more accurate constitutive parameters. Once the constitutive parameters change, deep learning prediction is performed using the new constitutive parameters to update the displacement time history of the experimental substructure in subsequent time steps in step D), and loading continues until the experiment is completed.

[0090] This invention may have other embodiments. Without departing from the spirit and essence of this invention, those skilled in the art can make various corresponding changes and modifications according to this invention, but these corresponding changes and modifications should all fall within the protection scope of the appended claims.

Claims

1. A real-time hybrid experimental method based on parameter recognition and deep learning agent model, characterized in that: Includes the following steps: S1. Pre-set a series of constitutive parameter values ​​for the test substructures, establish and utilize the finite element model of the overall structure to obtain the displacement response time history under different constitutive parameters; S2. Establish a deep learning surrogate model strucDL for the overall structural dynamic response. The input of strucDL includes the constitutive parameters of the experimental substructure, and the output is the displacement response time history of the experimental substructure. Train strucDL using different constitutive parameters and corresponding displacement response time histories to obtain a trained strucDL. S3. Using the restoring force and displacement of the test substructure as input, the constitutive parameters of the test substructure are obtained by using an identification algorithm. The process of obtaining the constitutive parameters of the test substructure by using the identification algorithm is called the test substructure constitutive parameter identification module, i.e., the expID module. S4. Based on the established overall structure deep learning proxy model strucDL and the experimental substructure constitutive parameter identification module, conduct real-time hybrid experiments; In the experiment, the restoring force and displacement of the experimental substructure in step i are sent to the expID module for parameter identification to obtain the current constitutive parameters; The constitutive parameters are then sent to strucDL to predict the displacement response time history of the experimental substructure at step i+1. The predicted displacement response time history of step i+1 is sent to the actuator to load the test substructure and obtain the restoring force of the test substructure at step i+1. S5. Repeat step S4 until the experiment ends; The inputs to strucDL in S2 also include ground motion; In S2, the training of strucDL is performed using an offline learning method.

2. The real-time hybrid experimental method based on parameter recognition and deep learning agent model according to claim 1, characterized in that: The identification algorithm used in S3 to obtain the constitutive parameters of the experimental substructure is the Kalman filter algorithm, which includes the unscented Kalman filter algorithm.

3. The real-time hybrid experimental method based on parameter recognition and deep learning agent model according to claim 2, characterized in that: The deep learning agent model mentioned is an LSTM neural network model.

4. A real-time hybrid experimental system based on parameter recognition and deep learning agent model, characterized in that: The system includes: The overall structural dynamic response prediction module, namely the strucDL prediction module, uses the deep learning surrogate model strucDL for overall structural dynamic response to predict the displacement response time history of the experimental substructure. The input of strucDL includes the constitutive parameters of the experimental substructure, and the output is the displacement response time history of the experimental substructure. The strucDL is pre-trained, and the training process includes the following steps: A series of constitutive parameter values ​​for the experimental substructures were preset in advance, and the displacement response time history under different constitutive parameters was obtained using the finite element model of the overall structure. The strucDL is trained using different constitutive parameters and corresponding displacement response time histories to obtain the trained strucDL. The constitutive parameter identification module for the test substructure, namely the expID module, uses the restoring force and displacement of the test substructure as input and employs an identification algorithm to obtain the constitutive parameters of the test substructure. Hybrid Test Control Module: Based on the overall structural dynamic response prediction module and the expID module, this module controls the real-time hybrid test until its completion. The process of controlling the real-time hybrid test includes the following steps: In the experiment, the restoring force and displacement of the experimental substructure in step i are sent to the expID module for parameter identification to obtain the current constitutive parameters; The constitutive parameters are then sent to the strucDL prediction module to predict the subsequent displacement response time history of the experimental substructure. The value of step i+1 in the predicted time history is sent to the actuator to load the test substructure and obtain the restoring force of the test substructure at step i+1. The inputs to strucDL also include ground motion; The training of strucDL is performed offline.

5. The real-time hybrid experimental system based on parameter recognition and deep learning agent model according to claim 4, characterized in that: It also includes a loading system and an actuator. The process by which the hybrid test control module sends the value of the (i+1)th step in the predicted time history to the actuator is based on the loading system. That is, the hybrid test control module sends the value of the (i+1)th step to the loading system through an interface. The loading system is used to control the actuator, and the actuator is used to perform loading on the test substructure.

6. The real-time hybrid experimental system based on parameter recognition and deep learning agent model according to claim 5, characterized in that: The Kalman filter algorithm is used in the process of obtaining the constitutive parameters of the experimental substructure by the expID module. The Kalman filter algorithm includes the unscented Kalman filter algorithm.

7. The real-time hybrid experimental system based on parameter recognition and deep learning agent model according to claim 6, characterized in that: The deep learning agent model mentioned is an LSTM neural network model.

Images