Self-adaptive physical information learning method for nonlinear seismic response prediction of RC structure

By introducing physical constraints and adaptive weight optimization algorithms into deep networks, combining automatic differential technology and Soft Adapt technology, the prediction inconsistency and data scarcity in the existing technology are solved, and high-precision and stable nonlinear seismic response of RC structures are achieved. predict.

CN118940610AActive Publication Date: 2024-11-12CHONGQING UNIV

Patent Information

Application Number
CN202410930727.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-11
Publication Date
2024-11-12
Estimated Expiration
2044-07-11

AI Technical Summary

Technical Problem

When the prior art predicts the nonlinear response of reinforced concrete structures under earthquake action, it is difficult to combine physical laws and data-driven methods, resulting in inconsistent prediction results and degradation of performance in the case of scarcity of data.

Method used

Using deep networks based on physical constraints and adaptive weight optimization algorithms, through the strategy of step-by-step training and freezing of parameters, combined with automatic differential technology and Soft Adapt technology, a model that can maintain high prediction accuracy under data scarcity is built.

Benefits of technology

The prediction accuracy of nonlinear seismic response of RC structure is significantly improved, the robustness and adaptability of the model is enhanced, the training time and resource requirements are reduced, and the stability and accuracy of the prediction are ensured.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118940610A_ABST
    Figure CN118940610A_ABST
Patent Text Reader

Abstract

The invention relates to the field of seismic wave prediction, and particularly discloses a self-adaptive physical information learning method for RC structure nonlinear seismic response prediction, which comprises the following steps: S1, establishing a deep network based on physical constraints; s2, establishing an adaptive weight optimization algorithm; s3, four networks are redefined according to an adaptive optimization algorithm, and each network corresponds to a corresponding output variable, namely u, v, z and epsilon; the method comprises the following steps: firstly, training the ith network to obtain network parameters theta i, 0, and then freezing the parameters to train the following i + 1 networks to obtain theta i + 1, 0; then, sequentially training the four networks, and re-training the four networks; and updating the parameters until the target loss functions of all networks reach a preset value, and obtaining the final network parameters theta < i, j > through the jth loop iteration. By adopting the technical scheme of the invention, the structure can be modeled and predicted by utilizing available measurement data and combining physical laws.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to the field of seismic wave prediction, and in particular to an adaptive physical information learning method for predicting nonlinear seismic response of RC structures. Background Art

[0002] In the field of structural engineering, accurate prediction of the response of reinforced concrete (RC) structures under earthquakes is crucial to ensure the safety and robustness of buildings. With the rapid development of machine learning techniques, data-driven methods have been widely used to learn and predict structural behavior from monitoring data. However, existing technologies have several limitations and challenges in dealing with this problem.

[0003] Traditional data-driven methods, such as certain types of neural networks, may perform well in fitting observed data, but often ignore the laws of physics, resulting in predictions that may be theoretically inconsistent or unreasonable. These intelligent models are somewhat like black boxes, and it is difficult to intuitively evaluate the accuracy of their predictions, which is unacceptable for safety-critical engineering applications.

[0004] In addition, current mainstream intelligent models rely heavily on abundant data to reveal underlying patterns. In system-level experiments, they usually face data scarcity problems due to the high cost and complexity of data collection, which limits the training and verification of models.

[0005] Therefore, there is an urgent need for an adaptive physics-informed learning method for predicting the nonlinear seismic response of RC structures that can utilize available measurement data and combine physical laws to model and predict the structure. Summary of the invention

[0006] The present invention provides an adaptive physical information learning method for predicting the nonlinear seismic response of RC structures, which can utilize available measurement data and combine physical laws to model and predict structures.

[0007] In order to solve the above technical problems, this application provides the following technical solutions:

[0008] An adaptive physical information learning method for predicting the nonlinear seismic response of RC structures includes the following:

[0009] S1. Building deep networks based on physical constraints;

[0010] S2. Establish an adaptive weight optimization algorithm;

[0011] S3. Four networks are redefined according to the adaptive optimization algorithm, each network corresponds to a corresponding output variable, namely u, v, z, ε; first train the i-th network to obtain the network parameter θ i,0, and then freeze the parameters to train the following i+1 networks to get θ i+1,0 ; Then, train the four networks in turn and retrain them again; update the parameters until the target loss function of all networks reaches the preset value, and the final network parameter θ is obtained after the jth cycle iteration i,j .

[0012] The basic scheme principle and beneficial effects are as follows: In this scheme, physical laws and available measurement data are used to build a deep learning model that can learn data features and physical phenomena at the same time. Automatic differentiation technology is used to accurately calculate the time derivatives of the model solution to ensure that the model strictly follows the physical equations during training.

[0013] Develop an adaptive weight optimization algorithm that can dynamically adjust the weights of different task losses to solve the imbalance between physical loss and data loss. By introducing the measure of gradient imbalance and Soft Adapt technology, the algorithm's dependence on the initial parameters of the network is reduced, and the robustness and adaptability of the algorithm are improved.

[0014] Four independent networks are defined, each of which is optimized for a specific output variable. Each network is trained in turn using a step-by-step training and parameter freezing strategy to ensure that each network can focus on its corresponding output variable during training. The network parameters are continuously updated through iterative cycles until all networks reach the preset performance standard on the target loss function.

[0015] This scheme significantly improves the prediction accuracy of nonlinear seismic response of RC structures by combining physical information and data-driven methods. The adaptive weight optimization algorithm enables the model to generalize better to new data sets and maintain high prediction performance even when data is scarce. The application of iterative training strategy and transfer learning reduces the time and resources required for model training, especially when dealing with large-scale data sets. Through the adaptive weight optimization algorithm, the model has better robustness to different data features and noise, and can maintain stability in the variable seismic response prediction task. The adopted optimization algorithm and training strategy enable the model to converge quickly to the optimal solution, shortening the model training cycle.

[0016] In summary, this approach achieves the effect of being able to use available measurement data combined with the laws of physics to model and predict structures.

[0017] Furthermore, the deep network in S1 is an improved Bouc-Wen model.

[0018] Furthermore, the deep network in S1 is describing the following dynamic equations of a single degree of freedom system:

[0019]

[0020] in are the elastic-plastic displacement, velocity and acceleration of the system respectively; m is the mass of the system; F(t) is the restoring force; f(t) is the external exciting force; the structural restoring force of the improved Bouc-Wen model includes elastic restoring force and hysteresis restoring force, and its expression is

[0021] F(t)=F el +F ELA =αku(t)+(1-α)kz(t) (2)

[0022] F(t) represents the yield force of the system; where α is the ratio of the linear elastic stiffness after yield to the initial linear elastic stiffness, k is the initial linear elastic stiffness of the structure

[0023] The relationship between the nonlinear stiffness, z and the lateral displacement u is as follows:

[0024]

[0025] Where A, β, γ and n are shape hysteresis control functions; A is 1; β and γ are the control combination of hysteresis stiffness and strength change and smooth strength; n controls the sharpness of the yield point; h(z) is the pinching effect function; ν and η are strength degradation and stiffness degradation parameters, respectively;

[0026] v(ε)=1+δ v ε,η(ε)=1+δ η ε (4)

[0027] Indicates energy consumption

[0028] represents the pinching effect, sgn(-) is the sign function

[0029] ζ1(ε)=ζ s [1-exp(-pε)], ζ2(ε)=(ψ+δ ψ (ε))(λ+ζ1(ε)) (5)

[0030]

[0031] ζs represents the total slip; p is a constant that controls the initial rate of slope decrease; ψ is the pinching behavior degree parameter; δψ represents the dispersion rate of the pinching phenomenon; λ controls the change of parameters ζ1(εn) and ζ2(εn); Equation (5) can be converted into a set of first-order ODEs;

[0032] The goal of PINN or PI-LSTM is to learn the network parameters θ and the physical parameters by simultaneously training the neural network and solving the underlying physical equations L is the total loss function, L d is the data loss function, L i It is the composition of the physical loss function

[0033] part;

[0034]

[0035]

[0036] in

[0037]

[0038] The key step in calculating the physical loss function is to calculate the time derivative of the model solution and use automatic differentiation to solve the time derivative.

[0039] Furthermore, the adaptive weight optimization algorithm in S2 represents the gradient of the loss function, which represents the network parameters at the n+1th iteration.

[0040] Furthermore, the network parameters at the n+1th iteration in S2 are calculated using equations 10 and 11, and are assumed to remain constant as initial values ​​during the calculation process; the training is updated using the gradient descent method as follows:

[0041] is the learning rate;

[0042] Then, applying the second-order Taylor expansion to the loss function at n yields:

[0043]

[0044] Applying (8) to (9), we obtain

[0045] in:

[0046]

[0047] Furthermore, S2 also includes the following content:

[0048] An improved algorithm based on GradNorm; it is the weight of the nth epoch, with the initial value set to 1, including the following steps:

[0049] Step 1: We introduce the measurement effect of gradient imbalance

[0050]

[0051] Step 2: Quantify the impact of initialization by introducing Soft Adapt using the softmax function;

[0052]

[0053] β is a Bernoulli random variable

[0054] p(β=k)=p k (1-p) 1-k k∈{0,1}, randomly backtrack to decide whether to use the loss statistics of the previous step (L i,0 ,L d,0 corresponds to the initial parameters) to calculate the scaling ratio,

[0055] Step 3: Use the moving average twice;

[0056]

[0057] Dynamic adjustment is achieved through annealing factors, that is, softmax is used for approximation in the early stage, and the gradient algorithm is directly used in the later stage to reduce the impact of initialization parameters on the learning of neighboring areas while maintaining the accuracy of the gradient algorithm;

[0058] Step 4: Inspired by the GradNorm method, design an additional loss L grad To learn the weights of different task losses, but it does not participate in the reverse gradient update of the network layer parameters, and the gradients of other tasks can be changed to the same value through regularization:

[0059]

[0060] E task is the average gradient norm of all tasks during training, || ||2 is the L2-norm, and η is used to represent the learning speed. is the ratio of the loss of the ith task (step t) to the initial loss is the relative learning speed of the ith task;

[0061] Step 5: Update parameters by gradient descent:

[0062]

[0063] The process of the algorithm is that in each round of training, the weights of the losses of different tasks are first calculated through back propagation, and then the network parameters are updated. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1Schematic diagram of a five-degree-of-freedom (5DOF) system based on the Bouc-Wen model principle.

[0065] Figure 2a Figure 2. Performance of six networks (PI-NN, PhyLSTM3, PI-LSTM, PI-LSTM-G, and PI-LSTM-A) for predicting the nonlinear displacement of the SDOF Bouc-Wen model, including regression analysis boxplots of the predicted displacement time histories.

[0066] Figure 2b Figure 2 is a graphical representation of the loss functions comparing the predicted displacements.

[0067] Figure 3a Comparison diagram of the displacement predicted by LSTM and the reference value.

[0068] Figure 3b Comparative illustration of hysteresis dissipation.

[0069] Figure 3c Relevant performance display diagram of the PI-LSTM-A metamodel.

[0070] Figure 4 The comparison chart of the normalized nonlinear displacement prediction performance between PI-LSTM-A and PI-LSTM methods and PI-NN and PhyLSTM3 methods.

[0071] Figure 5 The figure shows a comparison of the hysteresis curves and numerical experimental results obtained by different machine learning methods. It can be seen that the method in this paper (PI-LSTM-A) is closer to the results of the numerical method.

[0072] Figure 6 Illustration of the applicability of the physically based LSTM network (PI-LSTM-A) framework in transfer learning.

[0073] Figure 7 Schematic diagram of the reduction in the number of iterations and calculation time under different combinations of structural hysteresis parameters after adopting transfer learning.

[0074] Figure 8 This is a graphic representation of the number of epochs required for retraining on a new dataset and the initial loss value.

[0075] Fig. 9 This is a comparison chart of the hysteresis curve and numerical implementation results of the (PI-LSTM-A) method under different hysteresis parameter combinations after transfer learning.

[0076] Fig.10 This is a graphic representation of the number of epochs required for retraining on a new dataset and the initial loss value.

[0077] Fig.11 Figure 2 is a correlation plot using synthetic white noise data.

[0078] Fig.12 Figure 2 is a related illustration using ground motion data from the PEER NGA database.

[0079] Fig.13 A schematic diagram of the structure used to verify the calculation effect of the model

[0080] Fig.14a Schematic diagram of the model loss function under different hyperparameter values.

[0081] Fig.14b Box plot of regression analysis of the normalized nonlinear displacement prediction performance of the PI-LSTM-A method and other benchmark models.

[0082] Fig.15 A comparison of the top floor displacement time histories predicted by different models under representative earthquakes.

[0083] Fig.16 This is a schematic diagram of the results of ground sports training.

[0084] Fig.17 This is a diagram of the pre-trained model.

[0085] Fig.18 This is an example diagram of the optimization algorithm details. DETAILED DESCRIPTION

[0086] The following is further described in detail through specific implementation methods:

[0087] An adaptive physical information learning method for predicting the nonlinear seismic response of RC structures includes the following:

[0088] Step 1: Deep network construction based on physical constraints

[0089] BWBN (Bouc-Wen-Baber-Noori) is a modified Bouc-Wen model designed to describe the following dynamic equations of a single degree of freedom system:

[0090]

[0091] in are the elastic-plastic displacement, velocity and acceleration of the system respectively; m is the mass of the system; F(t) is the restoring force; f(t) is the external excitation force. The structural restoring force of the BWBN model includes elastic restoring force and hysteresis restoring force, and its expression is

[0092] F(t)=F el +F ELA=αku(t)+(1-α)kz(t) (2)

[0093] F(t) represents the yield force of the system. In the formula, α is the ratio of the yield to the initial linear elastic stiffness, k is the initial linear elastic stiffness of the structure, z is the nonlinear hysteresis displacement, and the lateral displacement u has the following relationship:

[0094]

[0095] Where A, β, γ and n are shape hysteresis control functions. A is equal to 1; β and γ are the control combinations of hysteresis stiffness and strength change and smooth strength; n controls the sharpness of the yield point; h(z) is the pinching effect function; ν and η are strength degradation and stiffness degradation parameters, respectively.

[0096] v(ε)=1+δ v ε,η(ε)=1+δ η ε (4)

[0097] Indicates the amount of energy consumption;

[0098] represents the pinching effect, sgn(-) is the sign function;

[0099] ζ1(ε)=ζ s [1-exp(-pε)], ζ2(ε)=(ψ+δ ψ (b))(λ+ζ1(b)) (5)

[0100] in:

[0101] ζs represents the total slip; p is a constant that controls the initial rate of slope decrease; ψ is the pinching behavior degree parameter; δψ represents the dispersion rate of the pinching phenomenon; λ controls the change of parameters ζ1(εn) and ζ2(εn). Equation (1) can be converted into a set of first-order ODEs as follows:

[0102]

[0103] The goal of PINN or PI-LSTM is to learn the network parameters θ and the physical parameters by simultaneously training the neural network and solving the underlying physical equations L is the total loss function, L d is the data loss function, L i It is a component of the physical loss function.

[0104]

[0105] in

[0106]

[0107] A key step in computing the physical loss function is to compute the time derivative of the model solution. Finite difference methods are simple and general, but they introduce approximation errors due to sparse sampling or fast function changes, and performing multiple function sampling and difference calculations during the solution process consumes more computational resources. For these considerations, automatic differentiation (AD) is employed, which has been proven successful in physically informed neural networks (PINNs) for forward and inverse analysis of nonlinear ODEs. AD is part of various deep learning frameworks and enables physically informed neural networks to conveniently evaluate derivatives.

[0108] Step 2: Propose a new adaptive weight optimization algorithm

[0109] The adaptive weight optimization algorithm represents the gradient of the loss function, which can represent the network parameters at the n+1th iteration. We use equation (10-11) for calculation and assume that it remains constant as the initial value during the calculation process. The training uses the gradient descent method to update as follows: is the learning rate. Then, applying the second-order Taylor expansion to the loss function at n, we get:

[0110]

[0111] Applying (8) to (9), we obtain:

[0112] in:

[0113] Physics loss is often used to ensure that the model obeys the laws of physics, and usually has a much larger range of values ​​than data loss. When introduced directly, training is quickly dominated by it. Physics loss produces non-zero gradients across the entire domain, while data loss is only effective close to data points. This difference can cause gradient spread, complicating convergence. In addition, gradient calculation is affected by the loss value and the size of the domain. Therefore, even if the loss value is small, the gradient contribution of physics loss over a wide range of values ​​may still dominate training. This can cause the model to fall into a feasible but suboptimal local optimum where the laws of physics are satisfied but the performance of the data points is lacking. Previous researchers aimed to balance the contributions of the PINN loss function and its gradients, and proposed a series of different solutions from different perspectives, including learning rate annealing, GradNorm, SoftAdapt, and multi-objective loss balance. Learning rate annealing is too sensitive to gradient changes, causing some loss terms to fluctuate widely around the optimal value of the objective function and making it difficult to converge. Second, the algorithm generates networks to optimize results that are easier to achieve, ignoring potentially important but more difficult to optimize loss functions. GradNorm addresses specific challenges posed by traditional learning rate annealing. There are still some shortcomings. The first is initialization dependency: GradNorm's performance depends heavily on the initial parameters of the network, which are manually set based on experience. This is because it uses these initial parameters to calculate the initial training error, which is used for subsequent gradient normalization. If the initialization is not ideal, it may have an adverse effect on the final performance. The second is the inability to handle extreme class imbalance: Although GradNorm can handle the imbalance between tasks to a certain extent, if there is an extreme imbalance in the category distribution between tasks, GradNorm may not be enough to solve this problem.

[0114] Inspired by the above work, we propose an improved algorithm based on GradNorm. is the weight of the nth epoch, and its initial value is set to 1. Our proposed method is divided into five steps:

[0115] Step 1: We introduce the measurement effect of gradient imbalance

[0116]

[0117] Step 2: Quantify the impact of initialization by introducing Soft Adapt using the softmax function;

[0118]

[0119] β is a Bernoulli random variable:

[0120] p(β=k)=p k (1-p) 1-kk∈{0,1}, randomly backtrack to decide whether to use the loss statistics of the previous step (L i,0 ,L d,0 Corresponding to the initial parameters) to calculate the scaling ratio;

[0121] Step 3: Use two forms of moving average;

[0122]

[0123] Dynamic adjustment is achieved through the annealing factor, that is, softmax is used for approximation in the early stage, and the gradient algorithm is used directly in the later stage to reduce the impact of the initialization parameters on the learning of the adjacent area while maintaining the accuracy of the gradient algorithm

[0124] Step 4: Inspired by the GradNorm method, an additional loss L is designed grad To learn the weights of different task losses, but it does not participate in the reverse gradient update of the network layer parameters, and the gradients of other tasks can be changed to the same value through regularization:

[0125]

[0126] E task is the average gradient norm of all tasks during training, ‖‖2 is the L2-norm, and η is used to represent the learning speed. is the ratio of the loss of the ith task (step t) to the initial loss is the relative learning speed of the ith task;

[0127] Step 5: Update parameters by gradient descent:

[0128]

[0129] The process of the algorithm is that in each round of training, the weights of the losses of different tasks are first calculated through back propagation, and then the network parameters are updated.

[0130] Obviously, compared with the annealing algorithm, our method imposes a certain form of upper limit constraint on the gradient change by setting the relative learning rate, which effectively alleviates the optimal value oscillation problem that is common in the annealing algorithm. In addition, the combination of softmax and random parameters reduces the impact of initialization dependence, thereby enhancing the robustness of the method.

[0131] Based on the above adaptive optimization algorithm, we redefined four networks, each network corresponds to a corresponding output variable, namely u, v, z, ε. First, train the i-th network and obtain the network parameter θ according to the above process. i,0 , and then freeze the parameters to train the following i+1 networks to get θ i+1,0Then, train the four networks in turn and retrain them again. Update the parameters until the objective loss function of all networks reaches the preset value. In this way, the final network parameter θ is obtained after the jth cycle iteration. i,j .

[0132] To further

[0133] Numerical Verification 1: 5 Degrees of Freedom (5DOF) System

[0134] Using a five-degree-of-freedom (5DOF) system based on the Bouc-Wen model principle (such as Figure 1 ), and then a normalization procedure was performed. The accumulated dataset was divided into three subsets: a training set containing 37 samples, a validation set containing 13 samples, and a test set containing 50 samples.

[0135] Figure 2a and Figure 2b The performance of six networks (e.g., PI-NN, PhyLSTM3, PI-LSTM, PI-LSTM-G, and PI-LSTM-A) for predicting the nonlinear displacement of the SDOF Bouc-Wen model is shown. These prediction schemes include pure data-driven LSTM and four physical information models: the classic PI-NN, the PhyLSTM3 metamodel, the metamodel PI-LSTM proposed in this paper, the GradNorm optimization model (denoted as: PI-LSTM-G), and the optimization model using the adaptive weight mechanism in this paper (denoted as: PI-LSTM-A).

[0136] We perform regression analysis on the predicted displacement time histories of all test datasets and Figure 2a The correlation coefficients are summarized in the box plot of . It can be observed that most of the correlation coefficients (denoted as γ) of PI-LSTM-A, PI-LSTM-A, and PI-LSTM are greater than 0.8, indicating that the predictions are very accurate. Several key indicators that reflect the distribution range and dispersion of the data, such as median, shape, box height, etc., predicted by PI-LSTM-A are better than other methods. Looking at the outliers (representing the worst performance of the model corresponding to the sample γ), the value of the LSTM method is significantly lower, at 0.61, while PI-LSTM is 0.75 and PI-LSTM-A is 0.82. Predicted displacement time history at the lowest γ in the example, and box plots of PI-LSTM (γ = 0.75) and LSTM (γ = 0.61) ( Figure 2a ). Even for the case where the prediction is not very satisfactory (e.g., γ = 0.75), the PI-LSTM method can still reasonably predict the displacement time history using very limited training data.

[0137] The PI-LSTM predicted displacement of γ = 0.75 matches the reference well in terms of phase, amplitude and residual drift. Figure 3a As shown in Figure 3, although the predicted displacement peak amplitude and phase match the reference values ​​relatively well, the LSTM cannot accurately predict the residual drift (e.g. plastic deformation). These characteristics hidden in the laws of physics are complex and difficult to discover through purely data-driven training.

[0138] Figure 3b It shows the hysteresis dissipation. It can be seen that the traditional LSTM has obvious deviations, while the PI-LSTM with physical information is consistent with the actual situation. Figure 3c and Figure 3d Similar prediction performance is observed for the PI-LSTM-A meta-model in .

[0139] This demonstrates that learning complex hysteresis behaviors purely from training data is difficult, especially when the available dataset is limited.

[0140] Figure 2b The loss functions for predicting displacement are compared. It can be seen that the proposed adaptive scheme (PI-LSTM-A) and GradNorm (PI-LSTM-G) show a faster decrease in loss, which means that the gradient-based weight optimization algorithm can reduce the training cost and improve the prediction accuracy to a certain extent. During training, we observed that the loss function initially decreased smoothly from the random initial state of the network. However, it later exhibited oscillations. These oscillations can be partially alleviated by reducing the learning rate, applying parameter regularization, or increasing the batch size. However, the non-convexity of the loss function makes all these strategies significantly slow down the training and often lead to reduced accuracy.

[0141] exist Figure 4 In this paper, PI-NN and PhyLSTM3 are used as benchmark models to compare the prediction performance of PI-LSTM-A and PI-LSTM methods for normalized nonlinear displacement. Figure 4 It shows that when epch = 200, there is a significant deviation between the prediction results of all four methods (epch = 1000) and the numerical result curve. Despite the specific deviation from the numerical results, PI-LSTM and PI-LSTM-A show better prediction performance than PI-NN and PhyLSTM3. The results are shown in Table 1.

[0142]

[0143]

[0144] Table 1 Comparison of loss functions (MSE and MAE) and operation time of the five algorithms

[0145] In this context, we explored the Physically Based LSTM Network (PI-LSTM-A) framework ( Figure 6 ) is suitable for transfer learning: using a pre-trained neural network to perform inference on a new dataset. The expectation is that since the initial state of the neural network is no longer randomly chosen, the training should converge to the solution and parameters of the data faster. To this end, we utilize a pre-trained model ( Fig.17 ) is trained on datasets with degradation factors and degradation shrinkage factors. We then study how the loss evolves and how the training converges when the data features are different: with four different strength degradation factors, stiffness degradation factors: under the same squeeze effect parameters, pinch slope, pinch strength,

[0146] As shown in the figure, retraining on the new dataset only takes a few hundred epochs with a small initial loss value. This shows the advantage of PI-LSTM, that the cost of retraining on similar data is much lower than the classical method that relies on forward simulation. This is an important result that illustrates the robustness of the proposed method.

[0147] Figure 8 , Fig. 9 The analysis shows that the training convergence speed is significantly accelerated when combined with transfer learning. Similar observations have been made in previous studies using various physical information models, indicating that the PINN method is well adapted to transfer learning. This finding suggests that transfer learning can be used to improve efficiency and cost-effectiveness, increase domain adaptability, enhance adaptability to scarce or cost-effective training data, and cultivate better generalization capabilities.

[0148] We further validate the framework on synthetic data generated by numerical methods; the data are divided into two categories. Fig.11 ) Using synthetic white noise data, the second type ( Fig.12 ) using ground motions from the PEER NGA database.

[0149] We use four numerical algorithms, including the explicit Runge-Kutta method of order 5(4) (RK45), the explicit Runge-Kutta method of order 3(2) (RK23), the implicit Runge-Kutta method of family IIA of order 5 (Radau), and the implicit multi-step variable order (1 to 5) method based on the backward differentiation formula of the derivative approximation (BDF). For ground motion training ( Fig.16 ), training converges faster on datasets generated by high-order numerical methods. However, for band-limited white noise ( Fig.15), it contains a lot of high-frequency components and random fluctuations, and high-order numerical algorithms such as RK45 may be more susceptible to these high-frequency noises, resulting in unstable results. On the other hand, low-order algorithms such as RK23 may show better noise suppression capabilities to a certain extent, resulting in smoother and more stable results. These phenomena highlight the importance of using appropriate numerical algorithms and sequences to pre-train ANNs according to the data type.

[0150] Numerical verification 2, three-layer MRF structure.

[0151] We use PI-NN, PhyLSTM3 and LSTM as benchmark models to compare the prediction performance of PI-LSTM-A (PI-LSTM-G) method for normalized nonlinear displacement.

[0152] || ||2 is the L2 norm

[0153] Calculated by automatic differentiation (AD)

[0154] We perform regression analysis on the predicted displacement time histories of all test datasets and use Fig.14b The box plots in illustrate the correlation coefficients. It is clear that the correlation coefficients of PI-LSTM-A are mostly greater than 0.9. Several key metrics, such as the median and the shape and height of the box, which reflect the range and dispersion of the data distribution, outperform the other methods. These results emphasize their robustness and excellent prediction accuracy. Looking at the outliers corresponding to the γ of the sample, which represents the worst model performance, the value of the LSTM method is significantly lower than 0.62, while the value of PI-LSTM-A is 0.75.

[0155] To further illustrate, we choose Fig.14b This discrete point with γ = 0.62 is taken as sample A. The points below the first quartile (Q1) for the LSTM and PI-LSTM-A methods are randomly selected as sample B, and the predicted displacement time history of the top floor under earthquake action is plotted.

[0156] Fig.15 The predicted displacement time histories of the top floor for two representative earthquakes are shown, highlighting the worst-case LSTM correlation coefficients (γ = 0.65 for (a)-(c) and γ = 0.62 for (d)-(f)), corresponding to PI-LSTM-G (γ = 0.80 for (a)-(c) and γ = 0.74 for (d)-(f), and corresponding to PI-LSTM-A (γ = 0.83 for (a)-(c) and γ = 0.78 for (d)-(f). The predicted curves of PI-LSTM-A with γ = 0.83 for (a)-(c) are very close to the experimental data that illustrate the amplitude, phase, and residual drift of plastic deformation.

[0157] Even in the worst case, such as γ = 0.78 in (d)-(f), PI-LSTM-A reasonably demonstrates its ability to predict displacement time histories given limited training data. The PI-LSTM-A meta-model also achieves a similar level of performance. Fig.15 The normalized structural displacement predicted by the LSTM shown shows reasonable predictions of the displacement peak amplitude and phase; however, it is unable to accurately predict the residual drift (e.g., plastic deformation), which deviates significantly from experimental results, indicating the difficulty of learning complex hysteretic behavior purely from limited training data.

[0158] The above are only embodiments of the present invention. The invention is not limited to the fields involved in this implementation case. The common sense such as the known specific structures and characteristics in the scheme is not described in detail here. The ordinary technicians in the relevant field know all the common technical knowledge in the technical field to which the invention belongs before the application date or the priority date, can obtain all the existing technologies in the field, and have the ability to apply the conventional experimental means before that date. The ordinary technicians in the relevant field can improve and implement this scheme in combination with their own abilities under the enlightenment given by this application. Some typical known structures or known methods should not become obstacles for ordinary technicians in the relevant field to implement this application. It should be pointed out that for those skilled in the art, without departing from the structure of the present invention, several deformations and improvements can be made, which should also be regarded as the protection scope of the present invention, which will not affect the effect of the implementation of the present invention and the practicality of the patent. The protection scope required by this application shall be based on the content of its claims, and the specific implementation methods and other records in the specification can be used to explain the content of the claims.

Claims

1. An adaptive physical information learning method for predicting the nonlinear seismic response of RC structures, characterized by: It includes the following: S builds deep networks based on physical constraints; S establishes an adaptive weight optimization algorithm; S redefined four networks according to the adaptive optimization algorithm, each network corresponds to a corresponding output variable, namely u, v, z, ε; first train the i-th network to obtain the network parameter θ i,0 , and then freeze the parameters to train the following i+1 networks to get θ i+1,0 ; Then, train the four networks in turn and retrain them again; update the parameters until the target loss function of all networks reaches the preset value, and the final network parameter θ is obtained after the jth cycle iteration i,j .

2. The adaptive physical information learning method for predicting nonlinear seismic response of RC structures according to claim 1, characterized in that: The deep network in S1 is a modified Bouc-Wen model.

3. The adaptive physical information learning method for predicting nonlinear seismic response of RC structures according to claim 2, characterized in that: The deep network in S1 is describing the following dynamic equations of a single degree of freedom system: in are the elastic-plastic displacement, velocity and acceleration of the system respectively; m is the mass of the system; F(t) is the restoring force; f(t) is the external exciting force; the structural restoring force of the improved Bouc-Wen model includes elastic restoring force and hysteresis restoring force, and its expression is F(t)=F el +F ELA =aku(t)+(1-α)kz(t) (2) F(t) represents the yield force of the system; where α is the ratio of the yield to the initial linear elastic stiffness, k is the initial linear elastic stiffness of the structure, z is the nonlinear hysteresis displacement, and the lateral displacement u has the following relationship: Where A, β, γ and n are shape hysteresis control functions; A is 1; β and Y are the control combination of hysteresis stiffness and strength change and smooth strength; n controls the sharpness of the yield point; h(z) is the pinching effect function; ν and n are strength degradation and stiffness degradation parameters, respectively; v(e)=1+d v e, n(e)=1+d η e (4) Indicates energy consumption represents the pinching effect, sgn() is the sign function ζ1(e)=ζ s [1-exp(-pε)],ζ2(ε)=(ψ+δ ψ (e))(λ+ζ1(e)) (5) ζs represents the total slip; p is a constant that controls the initial rate of slope decrease; ψ is the pinching behavior degree parameter; δψ represents the dispersion rate of the pinching phenomenon; λ controls the change of parameters ζ1(εn) and ζ2(εn); Equation (5) can be converted into a set of first-order ODEs; The goal of PINN or PI-LSTM is to learn the network parameters θ and the physical parameters by simultaneously training the neural network and solving the underlying physical equations L is the total loss function, L d is the data loss function, L i It is a component of the physical loss function; in The key step in calculating the physical loss function is to calculate the time derivative of the model solution and use automatic differentiation to solve the time derivative.

4. The adaptive physical information learning method for predicting nonlinear seismic response of RC structures according to claim 3, characterized in that: The adaptive weight optimization algorithm in S2 represents the gradient of the loss function and represents the network parameters at the n+1th iteration.

5. The adaptive physical information learning method for predicting nonlinear seismic response of RC structures according to claim 4, characterized in that: The network parameters at the n+1th iteration in S2 are calculated using equations 10 and 11, and are assumed to remain constant as initial values ​​during the calculation process; the training is updated using the gradient descent method as follows: is the learning rate; Then, applying the second-order Taylor expansion to the loss function at n yields: Applying (8) to (9), we obtain: in:

6. The adaptive physical information learning method for predicting nonlinear seismic response of RC structures according to claim 5, characterized in that: The S2 also includes the following contents: An improved algorithm based on GradNorm; it is the weight of the nth epoch, with the initial value set to 1, and includes the following steps: Step 1: We introduce the measurement effect of gradient imbalance Step 2: Quantify the impact of initialization by introducing Soft Adapt using the softmax function; β is a Bernoulli random variable p(β=k)=o k (1-p) 1-k k∈{0, 1}, randomly backtrack to decide whether to use the loss statistics of the previous step (L i,0 , L d,0 corresponds to the initial parameters) to calculate the scaling ratio, Step 3: Use the moving average twice; Dynamic adjustment is achieved through annealing factors, that is, softmax is used for approximation in the early stage, and the gradient algorithm is directly used in the later stage to reduce the impact of initialization parameters on the learning of neighboring areas while maintaining the accuracy of the gradient algorithm; Step 4: Inspired by the GradNorm method, design an additional loss L grad to learn the weights of different task losses, but it does not participate in the reverse gradient update of the network layer parameters, and the gradients of other tasks can be changed to the same value through regularization: E task is the average gradient norm of all tasks during training, || ||2 is the L2-norm, and η is used to represent the learning speed. is the ratio of the loss of the i-th task, i.e., the loss at step t to the initial loss is the relative learning speed of the ith task; Step 5: Update parameters by gradient descent: The process of the algorithm is that in each round of training, the weights of the losses of different tasks are first calculated through back propagation, and then the network parameters are updated.

Citation Information

Patent Citations

  • Reservoir prediction method based on bidirectional recurrent neural network

    CN109799533A

  • Multivariate parameter coupling data-knowledge dual-drive seismic inversion method

    CN118295016A

  • Model-driven deep learning-based seismic super-resolution inversion method

    US11226423B1

  • Methods, systems, and computer readable media for causal training of physics-informed neural networks

    WO2023172408A2

Cited By

  • Hysteresis prediction credibility evaluation method based on physical loss real-time monitoring

    CN122527545A