An off-line real-time hybrid test method based on boundary incremental residual force correction intelligent model

Abstract

The application discloses an offline real-time hybrid test method based on a boundary incremental residual force correction intelligent model and belongs to the technical field of structure hybrid test.The application adopts predefined compensation displacement to perform multitask servo loading on a test substructure, performs multi-round iterative training and update on an LSTM neural network based on measured time-history displacement, speed and restoring force data, accurately approximates the constitutive relation of the test piece and the boundary incremental residual relation, synchronously replaces the traditional test substructure and the physical constitutive force correction model with the trained LSTM model, embeds the structure motion equation for step-by-step integral solution, and realizes self-adaptive force correction through a convergence state symbol function, dynamic acceleration convergence and divergence suppression.The application significantly improves the test convergence efficiency and numerical stability under a strong nonlinear scene, reduces the test hardware threshold, and is suitable for the seismic dynamic response test of various engineering structures such as building frames, bridges, shock isolation and vibration reduction structures.

Description

Technical Field

[0001] This invention specifically relates to an offline real-time hybrid testing method based on a boundary incremental residual force correction intelligent model, and relates to the field of structural hybrid testing technology. Background Technology

[0002] Real-time hybrid testing is an advanced experimental method that divides the overall structure into numerical substructures and experimental substructures, and achieves synchronous loading and numerical calculation. This method can effectively simulate the rate dependence and inertial effects of structures under dynamic loads such as earthquakes and strong winds, and therefore has been increasingly widely used in the seismic performance evaluation of large and complex structures.

[0003] However, traditional online real-time hybrid testing relies on dedicated real-time communication hardware, highly synchronous control systems, and multi-actuator parallel loading devices. This places stringent demands on the hardware configuration, computational real-time performance, and communication stability of the test platform. The complexity of numerical substructure modeling is limited, resulting in high testing costs and deployment difficulties, making it difficult to popularize in conventional structural laboratories. To overcome these limitations, offline real-time hybrid testing methods have been developed in recent years. This method employs an iterative open-loop mechanism, offline inputting the measured restoring force of the substructure from the previous iteration into the numerical calculation system. This decouples physical loading from numerical calculation, completing the process sequentially and independently offline. Multiple iterations approximate the actual dynamic response of the structure, reducing reliance on dedicated real-time hardware and software. However, existing offline real-time hybrid testing methods have inherent drawbacks: their open-loop iterative mode lacks online real-time error feedback correction. When the test substructure contributes significant stiffness or damping to the overall structure, the iteration process is prone to slow convergence, oscillations, divergence, and even test instability.

[0004] Existing technologies mostly employ force correction methods based on physical constitutive models to optimize convergence. These methods rely on preset physical constitutive parameters, and the accuracy of parameter calibration directly determines the correction effect. When dealing with highly nonlinear specimens such as seismic isolation bearings, dampers, and energy dissipation nodes, the physical constitutive model is difficult to accurately match the actual mechanical behavior of the specimen, resulting in poor force correction adaptability and an inability to effectively solve the iterative instability and convergence difficulties caused by strong nonlinearity in multiple specimens.

[0005] Therefore, a method to solve the above-mentioned technical problems is urgently needed. Summary of the Invention

[0006] To address the problems mentioned in the background section, the present invention aims to provide an offline real-time hybrid testing method based on a boundary incremental residual force correction intelligent model, comprising the following steps:

[0007] S1. Multi-task loading is performed on the test substructure using predefined compensation displacement to obtain the measured time history displacement, velocity and measured time history restoring force data of each test substructure.

[0008] S2. Based on the measured data of each iteration, the LSTM neural network model is trained and updated in multiple iterations to make the model gradually approximate the nonlinear constitutive relationship and boundary incremental residual force relationship of the experimental substructure.

[0009] S3. The trained LSTM neural network model is used to replace the traditional experimental substructure and its corresponding force correction model based on physical constitutive model, and it is embedded into the structural motion equation to carry out step-by-step integration solution.

[0010] S4. Compare the time histories of the structural response in adjacent iterations. Determine whether the convergence condition is met based on the convergence index. If not, use the structural time histories displacement obtained in this round as the loading signal for the next round. Repeat the above steps until convergence. If the convergence condition is met, terminate the iteration.

[0011] Preferably, the iterative process of the method specifically includes:

[0012] A1. Establish a numerical model of the prototype structure, solve for the initial prototype structure response, and output the time-history displacement vector of each node. ;

[0013] A2. Time history displacement of each node in the prototype structure Perform time delay compensation, and then adjust the time history displacement after compensation. The data is sent to the servo loading system for multi-task loading of the experimental substructures, where the superscript j represents the iteration round and the subscript n represents the number of experimental substructures.

[0014] A3. Collect the measured displacement vectors of each test substructure. Measured velocity vector and measured force vector The network parameters of the LSTM neural network model are trained in multiple rounds and updated based on the training results;

[0015] A4. Replace the experimental substructure and boundary incremental residual force model with the trained LSTM neural network model respectively, and incorporate them into the motion equation to solve the structural response.

[0016] A5. Repeat steps A2 through A4 until the iteration termination condition determined by the accuracy threshold of the convergence index is met.

[0017] Preferably, the LSTM neural network model architecture is as follows:

[0018] (1)

[0019] (2)

[0020] (3)

[0021] (4)

[0022] (5)

[0023] (6)

[0024] Where i represents the input gate; f represents the forget gate; o represents the output gate; c represents the cell state; g represents the candidate memory; h represents the hidden state; x represents the input vector at the current time step; y represents the actual output vector; and tanh(·) represents the activation function. t represents element-wise multiplication; w represents the weight matrix; b represents the bias term; and the subscript t represents the time step.

[0025] Preferably, the equation of motion for the structure is:

[0026]

[0027] In the above formula, These represent the mass matrix and damping matrix of the structure, respectively. Let represent the acceleration, velocity, and displacement vectors of the numerical substructure at the ith integration step in the ith iteration; This represents the static restoring force vector of the numerical substructure at the i-th integration step in the j-th iteration; Let f(d) and f(i) represent the difference matrix between the restoring force predicted by the neural network model and the boundary force of the substructure at the i-th integration step in the j-th iteration, respectively; sign[f(d)] represents the sign function defined based on the numerical computation system f(d), and its value is determined by the convergence characteristics of f(d); the subscript N represents the numerical substructure, E represents the experimental substructure, and i represents the integration step; This represents the seismic acceleration excitation at the i-th integration step.

[0028] Preferably, the inputs for training the network parameters of the LSTM neural network model in multiple rounds are the measured time-history displacements. ,speed and boundary displacement difference Speed ​​difference The outputs are the measured time-history restoring forces. With boundary increment residual force ;in, , , , , and Each has n time history vectors.

[0029] Preferably, when the trained LSTM neural network model replaces the force correction model based on physical constitutive model for each experimental substructure, the inputs are the displacement vectors obtained in real time from the structural calculation. Velocity vector Difference between boundary displacements of substructure Boundary velocity difference Their outputs are the real-time predicted restoring force vectors. Boundary force difference vector with substructure ;in, , and Each has n time history vectors.

[0030] Preferably, the process of defining the convergence index includes:

[0031]

[0032]

[0033] In the above formula, q represents a specific experimental substructure, and the subscript m represents experimental measured data.

[0034] Preferably, determining whether the convergence condition is met based on the convergence index includes: when the convergence index... and If the accuracy threshold is less than 2%, the iteration is terminated.

[0035] Preferably, after comparing the time histories of adjacent iterative structural responses and determining whether the convergence condition is met based on the convergence index, the method further includes: applying a positive correction force when the structural response converges. Accelerate convergence; apply a negative correction force when the structural response diverges. To suppress dispersion.

[0036] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0037] I. This invention uses an LSTM neural network to simultaneously replace the experimental substructure and its corresponding force correction model based on physical constitutive model in offline and real-time hybrid experiments, and incorporates it into the equation of motion to solve the structural response. Unlike the existing technology that only uses neural networks to replace the internal interactions of numerical substructures, this invention directly performs integrated modeling of the nonlinear constitutive relations and correction terms of the experimental substructure, fundamentally solving the problem of insufficient accuracy of physical models and significantly improving nonlinear mapping capabilities.

[0038] Second, the neural network in this invention is not trained all at once, but through an iterative closed loop of "loading-testing-training-updating", the model parameters are continuously optimized as experimental data accumulates in each round. This iterative active learning strategy enables the surrogate model to continuously evolve during the experiment, accurately capturing the nonlinear evolution characteristics of the experimental substructure under cyclic loading, such as stiffness degradation and strength decay. This is significantly better than the one-time training mode based on predefined datasets in the prior art. The model is updated round by round using experimental data, dynamically adapting to strong nonlinear evolution and improving accuracy.

[0039] Third, this invention utilizes an LSTM model trained on accumulated iterative data for intelligent force correction. The correction term can dynamically adjust the direction and amplitude of the correction force through a sign function based on the real-time convergence state of the structural response. Compared to existing force correction methods that rely on fixed physical parameters, this invention possesses stronger nonlinear mapping capabilities and parameter adaptability, eliminating dependence on physical constitutive models. It effectively solves the difficulties in iterative convergence and experimental instability caused by the strong nonlinear behavior of multiple specimens, significantly improving convergence efficiency and numerical stability. This invention lowers the hardware threshold and is applicable to various structures such as frames, bridges, and seismic isolation systems. Attached Figure Description

[0040] For ease of explanation, the present invention will be described in detail below with reference to specific embodiments and accompanying drawings.

[0041] Figure 1 The flowchart shows the offline real-time hybrid test method based on the intelligent model of boundary incremental residual force correction.

[0042] Figure 2 The LSTM module architecture diagram is shown for the offline real-time hybrid experimental method based on the boundary incremental residual force correction intelligent model.

[0043] Figure 3 Flowchart of the principle of offline real-time hybrid test method for a seven-story frame vibration reduction / seismic structure with six seismic isolation bearings;

[0044] Figure 4 Flowchart of the principle of offline real-time hybrid test method for vibration reduction / seismic structure of a four-span bridge with three dampers based on a boundary incremental residual force correction intelligent model;

[0045] Figure 5 The flowchart illustrates the principle of an offline real-time hybrid test method based on a boundary incremental residual force correction intelligent model for a seven-story frame vibration reduction / seismic structure with seven dampers. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of this invention clearer, the invention is described below with reference to specific embodiments shown in the accompanying drawings. However, it should be understood that these descriptions are merely exemplary and not intended to limit the scope of the invention. Furthermore, descriptions of well-known structures and technologies are omitted in the following description to avoid unnecessarily obscuring the concept of the invention.

[0047] It should also be noted that, in order to avoid obscuring the invention with unnecessary details, only the structures and / or processing steps closely related to the solution according to the invention are shown in the accompanying drawings, while other details that are not closely related to the invention are omitted.

[0048] Specific implementation method one: Combining Figures 1 to 5 This embodiment describes an offline real-time hybrid testing method based on a boundary incremental residual force correction intelligent model, comprising the following steps:

[0049] S1. Multi-task loading is performed on the test substructure using predefined compensation displacement to obtain the measured time history displacement and measured time history restoring force data of each test substructure.

[0050] S2. Based on the measured data of each iteration, the LSTM neural network model is trained and updated in multiple iterations to gradually approximate the nonlinear constitutive relationship of the experimental substructure.

[0051] S3. The trained LSTM neural network model is used to replace the traditional experimental substructure and its corresponding force correction model based on physical constitutive model, and it is embedded into the structural motion equation to carry out step-by-step integration solution.

[0052] S4. Compare the time histories of the structural response in adjacent iterations. Determine whether the convergence condition is met based on the convergence index. If not, use the structural time histories displacement obtained in this round as the loading signal for the next round. Repeat the above steps until convergence. If the convergence condition is met, terminate the iteration.

[0053] This invention introduces an iteratively trained adaptive intelligent model to replace the traditional experimental substructure and force correction model in solving the equations of motion, which significantly enhances the nonlinear mapping capability and data adaptability, thereby effectively improving the convergence efficiency and numerical stability of the experiment.

[0054] Specific Implementation Method Two: This embodiment only provides a preferred implementation method, specifically, as follows: Figure 3As shown, taking a high-rise frame vibration reduction structure with six seismic isolation bearings as an example, the basic principle and usage steps of the method of the present invention are explained. In this embodiment, when the prototype structure is a high-rise frame vibration reduction / seismic structure, the high-rise frame is taken as the numerical substructure, and finite element software is used for refined modeling. One of the seismic isolation bearings is taken as the test substructure, and a servo loading system is used to perform multi-task loading in sequence to carry out offline real-time hybrid test based on the boundary incremental residual force correction intelligent model. The specific process of the offline real-time hybrid test method based on the boundary incremental residual force correction intelligent model includes:

[0055] A1. In iteration 0, establish a numerical model of the prototype structure and solve for the initial prototype structure response. Use the time-history displacement vectors of each node. For example;

[0056] A2. During the j-th iteration of the multi-task servo loading phase, the time-history displacement of each node in the high-level framework of the prototype structure is calculated. Perform time delay compensation, and then adjust the time history displacement after compensation. The data is sent one by one to the servo loading system to perform multi-task loading on the test substructure, i.e., the seismic isolation bearing. Here, the superscript j represents the iteration round and the subscript n represents the number of test substructures.

[0057] A3. During the j-th iteration of surrogate model parameter training and update phase, the measured displacement vectors of each experimental substructure are collected. Measured velocity vector With measured force vector The LSTM neural network model's parameters are trained multiple times and updated based on the training results. The LSTM model architecture is as follows.

[0058]

[0059]

[0060] 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; h represents the hidden state; x represents the input vector at the current time step; y represents the actual output vector; and tanh(·) represents the activation function. t represents element-wise multiplication; w is the weight matrix; b is the bias term; and the subscript t represents the time step.

[0061] A4. In the online real-time stepwise solution stage, the trained LSTM neural network model is used to replace the experimental substructure and force correction model respectively, and is incorporated into the motion equations to solve the structural response; among them, the motion equations for the high-rise frame vibration reduction / seismic structure are:

[0062]

[0063] in, These are the mass matrix and damping matrix of a high-rise frame vibration reduction / seismic structure, respectively. These are the acceleration, velocity, and displacement vectors of the high-level frame at the i-th integration step in the j-th iteration, respectively. It is the static restoring force vector of the high-level frame at the i-th integration step in the j-th iteration; These are the difference matrices between the restoring force predicted by the neural network model and the boundary force of the substructure at the i-th integration step in the j-th iteration; sign[f(d)] is a sign function defined based on the numerical calculation system f(d), and its value is determined by the convergence characteristics of f(d); the subscript N represents the high-rise frame, E represents the seismic isolation bearing, and i represents the integration step; It is the seismic acceleration excitation of the i-th integration step;

[0064] A5. Repeat steps A2 through A4 until the iteration termination condition determined by the convergence index and the preset engineering accuracy threshold is met. Content not mentioned in this embodiment is the same as in Specific Implementation Method 1.

[0065] Specific Implementation Method 3: This embodiment only provides a preferred implementation method. The basic principle of this invention is similar when applied to other large and complex vibration reduction structures; specifically, as shown in the example below. Figure 4 As shown, taking a bridge vibration reduction / seismic structure as an example, the basic principle and usage steps of the method of this invention are explained. This embodiment provides a method for a four-span bridge vibration reduction / seismic structure where the bridge structure is taken as a numerical substructure, and refined modeling is performed using finite element software; one damper is taken as the test substructure, and a servo loading system is used to sequentially perform multi-task loading to conduct an offline real-time hybrid test based on a boundary incremental residual force correction intelligent model. The specific test procedure is as follows:

[0066] S1, Iteration Round 0. Establish a numerical model of the bridge structure and solve for the initial structural response. Use the time-history displacement vectors of each node... For example.

[0067] S2, Iterative multi-task servo loading stage in the j-th round. Time-history displacement of each node of the bridge structure. Perform time delay compensation, and then perform the compensated time history displacement. Each signal is sent to the servo loading system for multi-task loading of the damper.

[0068] S3, Iterative training and update phase of the proxy model parameters in the j-th round. Collect the measured displacement vectors of each damper. Measured velocity vector With measured force vector The network parameters of the LSTM model are trained and updated multiple times. The LSTM model architecture is as follows:

[0069]

[0070]

[0071]

[0072]

[0073]

[0074]

[0075] 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; h represents the hidden state; x represents the input vector at the current time step; y represents the actual output vector; and tanh(·) represents the activation function. t represents element-wise multiplication; w is the weight matrix; b is the bias term; and the subscript t represents the time step.

[0076] S4. Online Real-Time Stepwise Solution Stage. The trained LSTM surrogate model replaces each damper and force correction model, and is incorporated into the equations of motion to solve for the structural response. The equations of motion for the bridge vibration reduction / seismic structure are as follows:

[0077]

[0078] in, These are the mass matrix and damping matrix of the bridge vibration reduction / seismic structure, respectively. These are the acceleration, velocity, and displacement vectors of the bridge structure at the i-th integration step in the j-th iteration; It is the static restoring force vector of the bridge structure at the i-th integration step in the j-th iteration; These are the difference matrices between the restoring force predicted by the neural network model and the boundary force of the substructure at the i-th integration step in the j-th iteration; sign[f(d)] is a sign function defined based on the numerical calculation system f(d), and its value is determined by the convergence characteristics of f(d); the subscript N represents the bridge structure, E represents the damper, and i represents the integration step; It is the seismic acceleration excitation of the i-th integration step;

[0079] S5. Repeat steps S2 to S4 until the iteration termination condition determined by the convergence index and the preset engineering accuracy threshold is met. Content not mentioned in this embodiment is the same as in specific embodiment one or two.

[0080] Specific Implementation Method Four: This embodiment only provides a preferred implementation method. The basic principle of this invention is similar when applied to other large and complex vibration reduction structures; specifically, as shown in... Figure 5 As shown, taking a high-rise frame structure with seven dampers for vibration reduction / seismic testing as an example, the basic principle and usage steps of the method of this invention are explained. In this embodiment, when the prototype structure is a high-rise frame structure with seven dampers for vibration reduction / seismic testing, the high-rise frame is taken as the numerical substructure, and refined modeling is performed using finite element software; one damper is taken as the test substructure, and a servo loading system is used to perform multi-task loading sequentially to conduct offline real-time hybrid tests based on the boundary incremental residual force correction intelligent model. The specific test procedure is as follows:

[0081] S1, Iteration Round 0. Establish a numerical model of the high-level framework and solve for the initial structural response. Use the time-history displacement vectors of each node... For example.

[0082] S2, Iteration j-th round of multi-task servo loading phase. Time-history displacement of each node in the high-level framework. Perform time delay compensation, and then perform the compensated time history displacement. Each signal is sent to the servo loading system for multi-task loading of the damper.

[0083] S3, Iterative training and update phase of the proxy model parameters in the j-th round. Collect the measured displacement vectors of each damper. Measured velocity vector With measured force vector The network parameters of the LSTM model are trained and updated multiple times. The LSTM model architecture is as follows:

[0084]

[0085]

[0086]

[0087]

[0088]

[0089]

[0090] 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; h represents the hidden state; x represents the input vector at the current time step; y represents the actual output vector; and tanh(·) represents the activation function. t represents element-wise multiplication; w is the weight matrix; b is the bias term; and the subscript t represents the time step.

[0091] S4. Online Real-Time Stepwise Solution Stage. The trained LSTM surrogate model replaces each damper and force correction model, and is incorporated into the equations of motion to solve for the structural response. The equations of motion for the high-rise frame vibration reduction / seismic structure are as follows:

[0092]

[0093] in, These are the mass matrix and damping matrix of a high-rise frame vibration reduction / seismic structure, respectively. These are the acceleration, velocity, and displacement vectors of the high-level frame at the i-th integration step in the j-th iteration, respectively. It is the static restoring force vector of the high-level frame at the i-th integration step in the j-th iteration; These are the difference matrices between the restoring force predicted by the neural network model and the boundary force of the substructure at the i-th integration step in the j-th iteration; sign[f(d)] is a sign function defined based on the numerical calculation system f(d), and its value is determined by the convergence characteristics of f(d); the subscript N represents the high-rise frame, E represents the damper, and i represents the integration step; It is the seismic acceleration excitation of the i-th integration step;

[0094] S5. Repeat S2 to S4 until the iteration termination condition determined by the convergence index and the preset engineering accuracy threshold is met; the contents not mentioned in this embodiment are the same as those in specific embodiments one, two or three.

[0095] Specific Implementation Method Five: This implementation method is a further limitation of Specific Implementation Methods One, Two, Three, or Four. The inputs for multiple rounds of training of the LSTM neural network model's network parameters are the measured time-history displacements. ,speed and boundary displacement difference Speed ​​difference The outputs are the measured time-history restoring forces. With boundary increment residual force ;in, , , , , and All are n time history vectors. The contents not mentioned in this embodiment are the same as those in specific embodiments one, two, three or four.

[0096] Specific Implementation Method Six: This implementation method is a further limitation of Specific Implementation Methods One, Two, Three, Four, or Five. When the trained LSTM neural network model replaces each experimental substructure and the force correction model based on physical constitutive model, the inputs are the displacement vectors obtained in real time from the structural calculation. Velocity vector Difference between boundary displacements of substructure Boundary velocity difference Their outputs are the real-time predicted restoring force vectors. Boundary force difference vector with substructure ;in, , , , , and All are n time history vectors. The contents not mentioned in this embodiment are the same as those in specific embodiments one, two, three, four or five.

[0097] Specific Implementation Method Seven: This implementation method is a further limitation of Specific Implementation Methods One, Two, Three, Four, Five, or Six. The definition process of the convergence index includes:

[0098]

[0099]

[0100] In the above formula, q represents a specific experimental substructure, and the subscript m represents experimental measured data. Contents not mentioned in this embodiment are the same as those in specific embodiments one, two, three, four, five, or six.

[0101] Specific Implementation Method Eight: This implementation method is a further limitation of Specific Implementation Methods One, Two, Three, Four, Five, Six, or Seven. Determining whether the convergence condition is met based on the convergence index includes: when the convergence index... and If the accuracy threshold is less than 2%, the iteration is terminated. Content not mentioned in this embodiment is the same as that in specific embodiments one, two, three, four, five, six or seven.

[0102] Specific Implementation Method Nine: This implementation method is a further limitation of Specific Implementation Methods One, Two, Three, Four, Five, Six, Seven, or Eight. In the motion equation, sign[f(d)] is used to monitor the convergence state of the structural response in real time, and corresponding force correction terms are dynamically applied. After comparing the time histories of adjacent iterations of the structural response and determining whether the convergence condition is met based on the convergence index, the method further includes: when the structural response converges, applying a positive correction force. Accelerate convergence; apply a negative correction force when the structural response diverges. To suppress dispersion, any content not mentioned in this embodiment is the same as in specific embodiments one, two, three, four, five, six, seven, or eight.

[0103] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. An offline real-time hybrid experimental method based on a boundary incremental residual force correction intelligent model, characterized in that, Includes the following steps: S1. Multi-task loading is performed on the test substructure using predefined compensation displacement to obtain the measured time history displacement, velocity and measured time history restoring force data of each test substructure. S2. Based on the measured data of each iteration, the LSTM neural network model is trained and updated in multiple iterations to make the model gradually approximate the nonlinear constitutive relationship and boundary incremental residual force relationship of the experimental substructure. S3. The trained LSTM neural network model is used to replace the traditional experimental substructure and its corresponding force correction model based on physical constitutive model, and it is embedded into the structural motion equation to carry out step-by-step integration solution. S4. Compare the time histories of the structural response in adjacent iterations. Determine whether the convergence condition is met based on the convergence index. If not, use the structural time histories displacement obtained in this round as the loading signal for the next round. Repeat the above steps until convergence. If the convergence condition is met, terminate the iteration.

2. The offline real-time hybrid experimental method based on the boundary incremental residual force correction intelligent model according to claim 1, characterized in that, The iterative process of the method specifically includes: A1. Establish a numerical model of the prototype structure, solve for the initial prototype structure response, and output the time-history displacement vector of each node. ; A2. Time history displacement of each node in the prototype structure Perform time delay compensation, and then adjust the time history displacement after compensation. The data is sent to the servo loading system for multi-task loading of the experimental substructures, where the superscript j represents the iteration round and the subscript n represents the number of experimental substructures. A3. Collect the measured displacement vectors of each test substructure. Measured velocity vector and measured force vector The network parameters of the LSTM neural network model are trained in multiple rounds and updated based on the training results; A4. Replace the experimental substructure and boundary incremental residual force model with the trained LSTM neural network model respectively, and incorporate them into the motion equation to solve the structural response. A5. Repeat steps A2 through A4 until the iteration termination condition determined by the accuracy threshold of the convergence index is met.

3. The offline real-time hybrid test method based on the boundary incremental residual force correction intelligent model according to claim 2, characterized in that, The architecture of the LSTM neural network model is as follows: （1） （2） （3） （4） （5） （6） Where i represents the input gate; f represents the forget gate; o represents the output gate; c represents the cell state; g represents the candidate memory; h represents the hidden state; x represents the input vector at the current time step; y represents the actual output vector; and tanh(·) represents the activation function. t represents element-wise multiplication; w represents the weight matrix; b represents the bias term; and the subscript t represents the time step.

4. The offline real-time hybrid test method based on the boundary incremental residual force correction intelligent model according to claim 3, characterized in that, The equation of motion for the structure is: ； In the above formula, These represent the mass matrix and damping matrix of the structure, respectively. Let represent the acceleration, velocity, and displacement vectors of the numerical substructure at the ith integration step in the ith iteration; This represents the static restoring force vector of the numerical substructure at the i-th integration step in the j-th iteration; Let f(d) and e(d) represent the difference matrix between the restoring force predicted by the neural network model and the boundary force of the substructure at the i-th integration step in the j-th iteration, respectively; sign[f(d)] represents the sign function defined by the value calculation system f(d), the value of which is determined by the convergence characteristics of f(d); subscript N represents the numerical substructure, E represents the experimental substructure, and i represents the integration step; This represents the seismic acceleration excitation at the i-th integration step.

5. The offline real-time hybrid testing method based on the boundary incremental residual force correction intelligent model according to claim 4, characterized in that, The inputs for training the network parameters of the LSTM neural network model in multiple rounds are the measured time-history displacements. ,speed and boundary displacement difference Speed ​​difference The outputs are the measured time-history restoring forces. With boundary increment residual force ;in, , , , , and Each has n time history vectors.

6. The offline real-time hybrid testing method based on the boundary incremental residual force correction intelligent model according to claim 5, characterized in that, When the trained LSTM neural network model is used to replace the experimental substructures and the force correction model based on physical constitutive model, the inputs are the displacement vectors obtained from the real-time solution of the structure. Velocity vector Difference between boundary displacements of substructure Boundary velocity difference Their outputs are the real-time predicted restoring force vectors. Boundary force difference vector with substructure ;in, , and Each has n time history vectors.

7. The offline real-time hybrid testing method based on the boundary incremental residual force correction intelligent model according to claim 6, characterized in that, The process of defining the convergence metric includes: ； ； In the above formula, q represents a specific experimental substructure, and the subscript m represents experimental measured data.

8. The offline real-time hybrid testing method based on the boundary incremental residual force correction intelligent model according to claim 7, characterized in that, The step of determining whether the convergence condition is met based on the convergence index includes: when the convergence index... and If the accuracy threshold is less than 2%, the iteration is terminated.

9. The offline real-time hybrid testing method based on the boundary incremental residual force correction intelligent model according to claim 8, characterized in that, The process of comparing the time histories of adjacent iterative structural responses and determining whether the convergence condition is met based on the convergence index further includes: applying a positive correction force when the structural response converges. Accelerate convergence; apply a negative correction force when the structural response diverges. To suppress dispersion.

Images