Attention-based hybrid network for long-period earthquake structural nonlinear response prediction

Abstract

The present invention relates to the field of artificial intelligence and structural seismic design, and specifically discloses a method for predicting the nonlinear response of a long-period earthquake structure based on an attention-based hybrid network, including the following contents: S1. Under the assumption that the structural parameters have a random probability distribution, a transition Markov chain Monte Carlo based on a Bayesian update framework system identification scheme is used to identify the hysteresis behavior of the structural system; S2. A long-period earthquake nonlinear response database is established through historical records; S3. An adaptive window function is used to accurately calculate the precise local relationship, and the input data is divided into multiple windows, each of which is predicted using an independent long short-term memory model (LSTM), and then the predictions are connected to obtain a complete prediction sequence; S4. Earthquake characteristic parameters are extracted through a gradient boosting decision tree model, and hysteresis characteristic parameters are extracted in S3 through LSTM. The nonlinear response requirements of the structure under the action of long-period earthquake motion can be quickly and accurately evaluated and controlled.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and structural seismic design, and in particular to a method for predicting nonlinear responses of long-period earthquake structures using an attention-based hybrid network. Background Art

[0002] In order to develop structures that can cope with severe seismic events, models and tools are essential to quickly and accurately evaluate and control the performance of structures. Although nonlinear time-history analysis can accurately estimate structural responses, it usually requires a lot of computing resources. To overcome the above shortcomings, purely data-driven machine learning methods have attracted attention and have been widely used in practice, but machine learning methods often ignore the underlying physical laws and cannot provide satisfactory interpretability, which may lead to potentially misleading results.

[0003] The lack of explainability is not conducive to understanding the nature of the problem. In addition, the lack of connection between pure data-driven methods and physical laws may lead to specious errors that violate common sense. Such intelligent models are similar to black boxes, and it is difficult for engineers to evaluate their accuracy based on partial test results. This is unacceptable for safety-critical engineering applications.

[0004] Long-period ground motions and their impacts on buildings and the built environment have attracted increasing attention in recent years. Analytical studies of previous earthquake ground motions have shown that the combined effects of seismic wave directivity and geological basin effects may result in dominant periods of acceleration response spectra of two horizontal components exceeding 3 seconds or longer. This conclusion poses a major challenge to high-rise buildings, as the resulting huge displacement demands may exceed the spectral values ​​defined in the code, especially in the case of medium-height and high-rise buildings. In order to meet these displacement demands, structural engineers face numerous design and detail considerations. The displacement requirements of high-rise structures are related to the inter-story displacement requirements, and the results are strongly dependent on the structural nonlinearity and ground motion characteristics. Therefore, it is particularly important to quickly and accurately evaluate and control the nonlinear response requirements of structures under long-period ground motions. Summary of the invention

[0005] The present invention provides an attention-based hybrid network for predicting the nonlinear response of long-period earthquake structures, which can quickly and accurately evaluate and control the nonlinear response requirements of structures under long-period earthquake motions.

[0006] In order to solve the above technical problems, this application provides the following technical solutions:

[0007] The attention-based hybrid network is used for the prediction of nonlinear response of long-period earthquake structures, including the following contents:

[0008] S1. Under the assumption that the structural parameters have a random probability distribution, the probability parameter distribution of the structure is determined by transition Markov chain Monte Carlo simulation based on the system identification scheme in the Bayesian updating framework;

[0009] S2. Conduct time history analysis through a long-period earthquake record database and a probability parameter distribution based on a previously determined structure to establish a long-period earthquake nonlinear response database, wherein the long-period earthquake nonlinear response database includes a hysteresis curve database;

[0010] S3. A network structure of an attention mechanism is proposed, which uses an adaptive window function to accurately calculate the precise local relationship, divides the hysteresis curve into multiple windows, and each window is predicted using an independent LSTM model to output the maximum response displacement or other structural requirement parameters;

[0011] S4. Extract earthquake characteristic parameters from the long-period earthquake nonlinear response database through the gradient boosting decision tree model, and extract hysteresis characteristic parameters from the maximum response displacement or other structural demand parameters, and output nonlinear prediction results.

[0012] Furthermore, the S1 further includes:

[0013] The MATLAB toolbox genetic algorithm (GA) and NSGA-II algorithm were used to determine the initial value of the BWBN parameter θ;

[0014] The optimal estimate is given by the following equation: Determine the MSE of the objective function;

[0015] The range of parameters is The surroundings are normally or evenly distributed, and their ranges are based on the capability curves in HAZUS-MH 2.1 and the pre-entered previous research records;

[0016] Use the observed data to update the prior distribution and determine the posterior probability distribution of the parameters;

[0017] Use the genetic algorithm in the MATLAB Global Optimization Toolbox to perform sensitivity analysis and determine the importance of parameters;

[0018] Based on the results of the sensitivity analysis, the four most important parameters were fixed at their median values;

[0019] The fixed parameter values ​​are used to construct the earthquake demand database while ensuring the acceptable range of mean square error.

[0020] Furthermore, the S2 further includes:

[0021] Extract seismic wave records that meet long-period characteristics from the NGA-West database;

[0022] The basin amplification effect and site amplification effect were evaluated using Z1.0, Z1.5, Z2.5 and vs30 parameters;

[0023] Construct parameter combinations to respond to different hysteresis states of the structure;

[0024] Apply seismic waves to structures with different hysteresis characteristics and perform time history analysis;

[0025] Nonlinear analysis was performed using the BWBN model and the Runge-Kutta numerical integration method. The results of the nonlinear analysis were extracted to form a database. Each data set included the acceleration time history of the ground motion as input and the response as output. The maximum displacement was extracted as the demand parameter.

[0026] The pushover analysis is performed by applying a gradually increasing lateral force pattern to the structure to obtain the hysteresis curve. The nonlinear analysis is performed using the BWBN model. The Runge-Kutta numerical integration method is used as the solver. The turning point from the linear region to the nonlinear region under different degradation and shrinkage parameters is described by finding the intersection between the pushover curve and a linear line starting from the origin with a slope of 90% of the initial stiffness.

[0027] The BWBN nonlinear model involves several parameter combinations of different hysteresis states of the structure. Each combination includes 240×2 input points to form a hysteresis database.

[0028] Furthermore, the S3 is implemented by an LSTM model relying on the attention mechanism.

[0029] Furthermore, the LSTM model of the attention mechanism includes an input layer, an LSTM layer, a fully connected layer, and an output layer.

[0030] Furthermore, the S3 further includes:

[0031] The attention mechanism calculates the final hidden state and attention score vector based on the input data from different channels of the LSTM model;

[0032] Normalize the scores using the SoftMax function; align and sum the vectors using equations;

[0033] Using empirical wavelet transform to decompose signals of different scales into sub-bands of multiple scales to extract feature information from signals of different scales;

[0034] Extracting different time scales and frequency features from the original signal enables the LSTM model to capture dynamic features and nonlinear relationships in the data;

[0035] Use wavelet coefficients to extract features and reduce the dimension of the original signal;

[0036] At each decoding time step t, the model receives as input the hidden state of the top layer of the stacked LSTM model;

[0037] Generate a sequence vector containing relevant information from the source to assist in predicting the current target;

[0038] A surface concatenation layer is used to merge the target hidden state with the sequence vector on the source side to produce the attention hidden state as shown below:

[0039] Initially predict the alignment position of each target token to generate a sequence vector; calculate the weighted average of the source hidden states within the specified window: w(i)∈[p t-D ,p t+D ], for a given input sequence X = [x1, x2, ..., x n ] and the target element y t =x j , first apply a linear transformation to the input sequence to obtain the query parameter Q, the field name K, and the specific data content V corresponding to K:

[0040] Q=X·W q

[0041] K=X·W k

[0042] V=X·W v

[0043] Here, W q , W k and W v is a learnable weight matrix; the self-attention mechanism is used to calculate the attention weight a ij and the context vector c j :

[0044] Similarity score: scor(e ij )=Q j ·K i

[0045] Attention weights:

[0046] Context vector: c j =Σ(a ij ·V i )

[0047] Use the context vector c j As the target element x j context information.

[0048] Further, S3 also includes: adaptive window function W(i)=a*L(i)+(1-a)*C(i), where W(i) represents the window function of the i-th element, L(i) is a window function based on the sequence length, C(i) is a window function based on context information, and a is a weighting coefficient; the sequence length part L(i) is defined as L(i)=min(w l ,max(1,ceil(λ*n))), where w l is the upper limit of the fixed window size, λ is a parameter that controls the window size (usually ranging from 0 to 1), and ceil is an upper limit function that rounds up to the nearest integer;

[0049] This part of the window function determines the window size based on the sequence length, limits the upper limit and ensures that the window contains at least one element;

[0050] C(i) is defined as C(i) = sum(c j *x j ) / sum(c j ), where c j Represents the i-th element x i The weight of the contextual relationship, x j represents the jth element in the sequence;

[0051] This part of the window function calculates weights based on contextual information to determine the position of the window in the sequence;

[0052] By adjusting the weight coefficient a and the control parameter λ, the effects of sequence length and context information on the window function can be balanced.

[0053] Furthermore, S3 further includes: the local alignment vector is restricted to a fixed dimension, in particular, at ∈ R 2D+1 At , the alignment position is predicted using the following mechanism:

[0054]

[0055] S is the length of the source sequence. Through the sigmoid function, pt∈[0,S] the network will learn model parameters to predict the position;

[0056] A Gaussian distribution centered around the edge is introduced to emphasize neighboring alignment points, and the alignment weight is defined as follows

[0057]

[0058] The principle and beneficial effects of the scheme are as follows: In this scheme, a bridge is built between explicit and implicit hysteresis models to learn hysteresis curves as representations of structural properties. A comprehensive database of long-term seismic nonlinear responses is established. Then, a novel attention mechanism is used to improve the prediction accuracy. The complex effects of complex hysteresis and seismic characteristics on peak seismic response are effectively captured through the design of a hybrid deep network. Numerical verification is performed to demonstrate the effectiveness of the proposed method.

[0059] Based on the transition Markov chain Monte Carlo of the Bayesian updating framework system identification scheme, the transition Markov chain Monte Carlo (TMCMC) based on the Bayesian updating framework system identification scheme is developed to identify the hysteretic behavior of the structural system, assuming that the structural parameters have a random probability distribution. A long-period earthquake nonlinear response database is established through historical records, including the acceleration time history and response of seismic waves. An adaptive window function is used to accurately calculate the precise local relationship. The input data is divided into multiple windows, each of which is predicted using an independent LSTM model, and the predictions are then connected to obtain a complete prediction sequence. The earthquake characteristic parameters are extracted by the gradient boosting decision tree model, and the hysteresis characteristic parameters are extracted in S3 by the long short-term memory model.

[0060] By building a comprehensive database of long-term earthquake nonlinear response, a novel attention mechanism is used to improve the prediction accuracy. The complex effects of complex hysteresis and earthquake characteristics on peak earthquake response are effectively captured through the design of a hybrid deep network. The results of genetic algorithm (GA) identification are used as the prior of the Bayesian method, and the posterior probability distribution is updated and determined using an update framework. After determining the parameter values, sensitivity analysis is performed using the MATLAB Global Optimization Toolbox (Genetic Algorithm). The following four most important parameters are fixed at the median value to construct the earthquake demand database, which significantly reduces the computational cost.

[0061] By using the empirical wavelet transform (EWT), feature information is extracted from signals of different scales by decomposing them into sub-bands of multiple scales. Extracting different time scales and frequency features in the original signal enables LSTM to capture dynamic features and nonlinear relationships in the data. By using wavelet coefficients to extract features and reduce the dimension of the original signal, the amount of data that the LSTM model needs to process can be reduced. This method improves the prediction accuracy of the model and improves its generalization ability.

[0062] The hybrid deep learning model proposed by the authors is composed of two components: extracting earthquake characteristic parameters based on the Light Gradient Boosting Machine (LGM) decision tree and extracting the hysteresis characteristic parameters obtained by steps one and two using the Bi-LSTM model in step three. Through mode decomposition and reorganization, the research method is extended to multi-degree-of-freedom systems and numerically verified through actual structures. By comparing with traditional methods, the feasibility and accuracy of the proposed method are verified again.

[0063] The empirical wavelet transform and local attention mechanism are embedded into the established framework to pursue better performance.

[0064] Based on the Bayesian updating framework system identification scheme, a long-period earthquake ground motion response dataset and a hysteresis model dataset are constructed, and a hybrid network is designed to study the complex interaction between long-period ground excitation and structural response.

[0065] The method of the present invention proposes a self-attention mechanism, which is based on the improvement of the local attention mechanism and proposes a method to improve the attention mechanism by predicting the position of the target time point and calculating the weighted average within an adaptive window.

[0066] The following is a detailed explanation of the improvements and principles.

[0067] (1) Adaptive window: An adaptive window function W(i)W(i) is introduced, which combines the sequence length and context information to determine the window size and position.

[0068] (2) Predicting alignment position: The model predicts the alignment position of each target time point through an adaptive mechanism, rather than simply assuming monotonic alignment.

[0069] (3) Gaussian distribution weighting: Gaussian distribution is introduced around the predicted alignment position to emphasize neighboring alignment points, thereby effectively reducing the computational cost and improving model performance.

[0070] The principle is explained as follows:

[0071] (1) Linear transformation of input sequence:

[0072] Perform a linear transformation on the input sequence X to obtain the query (Q), key (K) and value (V) representations:

[0073] Q=X·W q

[0074] K=X·W k

[0075] V=X·Wv

[0076] Where W q ,W k ,W v is a learnable weight matrix.

[0077] (2) Use the linearly transformed query (Q) and key (K) representations to calculate the similarity score:

[0078] score(e ij )=Q i ·K j

[0079] This similarity score reflects the correlation between each position xi and xj in the input sequence.

[0080] Then, the attention weights are calculated by normalizing the similarity scores:

[0081] Here d is the dimension of the key, and the softmax function ensures that the sum of the attention weights is 1.

[0082] Finally, use these attention weights to weight the sum value (V) to get the context vector c j :

[0083]

[0084] This context vector c j Combines the source sequence with the target position x j All relevant information.

[0085] (3) Adaptive window function:

[0086] To further optimize the computation, an adaptive window function W(i) is introduced, which combines the sequence length and context information:

[0087] W(i)=a·L(i)+(1-a)·C(i)

[0088] Among them, a is the weight coefficient, L(i) is based on the sequence length, and C(i) is based on the context information.

[0089] The sequence length part L(i) is defined as:

[0090] This ensures that the window contains at least one element and has a fixed upper limit on the window size.

[0091] C(i) is defined as:

[0092]

[0093] Here, c j is obtained in step (2) above and represents the i-th element x i The context weight of is used to determine the position of the window in the sequence.

[0094] (4) Prediction distribution:

[0095] Finally, the prediction distribution of the model is obtained by inputting the attention vector into the softmax layer.

[0096] In summary, this scheme achieves the goal of being able to quickly and accurately evaluate and control the nonlinear response requirements of the structure under long-period seismic motion. BRIEF DESCRIPTION OF THE DRAWINGS

[0097] Figure 1 A framework diagram of the novel method based on long short-term memory of the present invention for predicting the nonlinear response of structures.

[0098] Figure 2 This is a graph showing the experimental data verification results of the inversion hysteresis parameter characteristics of the Bayesian updating framework system of the present invention.

[0099] Figure 3 This is a graph showing the experimental data verification results of the inversion hysteresis parameter characteristics of the Bayesian updating framework system of the present invention.

[0100] Figure 4 This is a graph showing the experimental data verification results of the inversion hysteresis parameter characteristics of the Bayesian updating framework system of the present invention.

[0101] Figure 5 This is a graph showing the experimental data verification results of the inversion hysteresis parameter characteristics of the Bayesian updating framework system of the present invention.

[0102] Figure 6 This is a comparison chart of the actual earthquake response record of a six-story hotel of the present invention and the calculation result of the method of the present invention.

[0103] Figure 7 This is a comparison chart of the actual earthquake response record of a six-story hotel of the present invention and the calculation result of the method of the present invention.

[0104] Figure 8 It is a comparison diagram of the prediction results of the present invention for single-degree-of-freedom structures under different periods.

[0105] Fig. 9 It is a comparison diagram of the prediction results of the present invention for single-degree-of-freedom structures under different periods.

[0106] Fig.10 It is a comparison diagram of the prediction results of the present invention for single-degree-of-freedom structures under different periods.

[0107] Fig.11 This is a numerical verification result diagram of the prediction results of the present invention for a single-degree-of-freedom structure under different periods.

[0108] Fig.12 This is a diagram showing the verification results of the prediction results of the present invention for a single degree of freedom structure under different periods.

[0109] Fig.13 It is a comparison diagram of the prediction results of the present invention for single-degree-of-freedom structures under different periods.

[0110] Fig.14 It is a comparison diagram of the prediction results of the present invention for single-degree-of-freedom structures under different periods.

[0111] Fig.15 The following is a schematic diagram for listing some combination examples of the present invention.

[0112] Fig.16 Schematic diagram of the network structure of the attention mechanism in the present invention.

[0113] Fig.17 Schematic diagram of the relationship between all models in the present invention. DETAILED DESCRIPTION

[0114] The following is further described in detail through specific implementation methods:

[0115] Attention-based hybrid network for long-period earthquake structural nonlinear response prediction method (such as Figure 1 ), including the following:

[0116] S1. Under the assumption that the structural parameters have a random probability distribution, a transition Markov chain Monte Carlo simulation based on the system identification scheme in a Bayesian updating framework is used to determine the probability parameter distribution of the structure.

[0117] For a set of BWBN parameters θ, the genetic algorithm (GA) identification results are used as the priors for the Bayesian method: First, the initial values ​​are determined using the MATLAB toolbox genetic algorithm (GA) and the NSGA-II algorithm. The optimal estimate is determined by the following equation: The objective function is MSE. Then, assuming that the range of parameters is Normal or uniform distribution around. The range of the parameter prior distribution is based on the capacity curve in HAZUS-MH 2.1 and previous research assumptions, and the present invention assumes a uniform prior distribution. The update framework is then used to update and determine the posterior probability distribution. After determining the parameter values, a sensitivity analysis is performed using the MATLAB Global Optimization Toolbox (Genetic Algorithm). Based on the above analysis, the results of Table 1 are obtained, including the sensitivity levels and recommended parameter values. The sensitivity analysis sorts the parameters according to their importance, fixes the following four most important parameters at the median, and constructs a seismic demand database. This significantly reduces the computational cost, and previous studies have shown that it only slightly increases the mean square error. The parameters considered in the analysis include three stages of stiffness degradation, three stages of strength degradation, and nine different degrees of pinching effects.

[0118]

[0119]

[0120] Table 1

[0121] S2. Through the long-period earthquake record database and based on the previously determined probability parameter distribution according to the determined structure, a time history analysis is performed to establish a long-period earthquake nonlinear response database, which includes a hysteresis curve database.

[0122] The seismic waves used in this embodiment are extracted from the NGA-West database. These records are all located within the basin and may produce long-period seismic motions. 1014 seismic waves that meet the long-period characteristics are selected. Z1.0, Z1.5 and Z2.5 are indicators used in earthquake engineering to indicate the depth at which the shear wave velocity reaches a specific value (1.0km / s, 1.5km / s and 2.5km / s, respectively). They are generally used as characterization parameters in the ground motion prediction equation (GMPE) to evaluate basin effects and earthquake amplification. Z1.0, Z1.5 and Z2.5 are respectively related to the basin amplification effect, while vs30 is related to the site amplification effect. Fig.11 The distribution relationships of Mw with Z1.0, Z1.5 and Z2.5 are shown respectively.

[0123] A total of 30*10*4*3*3*9=97200 parameters are used to respond to different hysteresis states of the structure using the parameter combinations in Table 1. The structural period is 0.05 to 10 seconds, divided into 30 equal parts.

[0124] A total of 1104 seismic waves were applied to 3240 structures with different hysteresis characteristics, resulting in 1104*97200=107308800 time history analyses. Only 5% damping was used. The BWBN model was used for nonlinear analysis, using the Runge-Kutta numerical integration method (especially the ode45 solver). The results of the nonlinear analysis were extracted to form a database, and each data set included the acceleration time history of the ground motion as input and the response as output. The present invention only extracted the maximum displacement (m) as the required parameter. Some physical properties of the selected GM such as Figure 13-14 shown.

[0125] In addition to the seismic properties, the hysteresis curves are also used as input parameters of the model. Pushover analysis is performed by applying a gradually increasing lateral force pattern to the structure to obtain the hysteresis curves. The nonlinear analysis is performed using the BWBN model with the Runge–Kutta numerical integration method (ode45) as the solver. The turning point from the linear region to the nonlinear region under different degradation and shrinkage parameters is described by finding the intersection between the pushover curve and a linear line starting from the origin with a slope of 90% of the initial stiffness. The BWBN nonlinear model in Table 2 involves 97,200 parameter combinations for different hysteresis states of the structure, each with 240 × 2 input points to form the hysteresis database. Fig.13 Typical examples from a subset of samples are shown; in these examples, the yield strength of the structure is set to 0.5g according to HAZUS-MH2.1 (FEMA2012). Most earthquake input ground motions can enter the nonlinear stage. When the structure does not enter the nonlinear stage, only the linear elastic model is used to calculate the results. In the linear elastic state, the corresponding hysteresis parameter: α is set to 1, and other parameters such as pinch severity parameter, strength / stiffness degradation factor: are set to 0.

[0126] In other studies of the invention, various combinations of hysteresis parameters of the structure were investigated, including four post-yield stiffness ratios, three degrees of stiffness degradation, three types of strength degradation, and nine shrinkage effects, resulting in a total of 324 combinations ( Fig.15 A total of 286,416 input and output samples of 1,104 long-period waves were used as the training set, consisting of 884*324 samples with different lag parameters. In addition, 71,280 samples were used as the test set.

[0127] S3. A network structure of an attention mechanism is proposed, which uses an adaptive window function to accurately calculate the precise local relationship, divides the hysteresis curve into multiple windows, and each window is predicted using an independent LSTM model to output the maximum response displacement or other structural requirement parameters.

[0128] The self-attention mechanism captures global dependencies but has high computational complexity. On the other hand, the local attention mechanism provides efficient computation and adaptability to different sequence lengths but may ignore global information. Considering these trade-offs, a novel approach combining adaptive and local attention mechanisms is proposed, employing an adaptive window function to accurately compute the precise local relations.

[0129] The LSTM model with the proposed attention mechanism (AM) is shown in Fig.16 As shown in the figure, the model consists of an input layer, an LSTM layer, a fully connected layer, and an output layer. The proposed attention mechanism attention calculates the final hidden state and attention score vector based on the input data from different channels of the LSTM network. The scores are then normalized using the SoftMax function. The vectors are aligned and summed using the equation. LSTM uses the above-mentioned recurrent neural network to selectively recall the immediate history or long-term patterns of the input sequence. The empirical wavelet transform (EWT) is used to extract feature information from signals of different scales by decomposing them into sub-bands of multiple scales. Extracting different time scales and frequency features from the original signal enables LSTM to capture dynamic features and nonlinear relationships in the data. By using wavelet coefficients to extract features and reduce the dimensionality of the original signal, the amount of data that the LSTM model needs to process can be reduced. This approach improves the prediction accuracy of the model and improves its generalization ability. In addition, the input data is divided into multiple windows. Each window is predicted using an independent LSTM model, and the predictions are then connected to obtain a complete prediction sequence.

[0130] At each decoding time step t, the model receives as input the hidden state of the top layer of the stacked LSTM.

[0131] Next, the model aims to generate a sequence vector containing relevant information from the source to assist in predicting the current target.

[0132] Next, the model uses a surface concatenation layer to merge the target hidden state (denoted as) with the sequence vector on the source side to produce the attention hidden state, as shown below

[0133] The model initially predicts the alignment position of each target token to generate a sequence vector. Subsequently, it calculates the alignment position of each target token in the specified window w(i)∈[p t-D ,p t+D ] is the weighted average of the endogenous hidden states,

[0134] For a given input sequence X = [x1, x2, ..., xn] and a target element y t =x j , we first apply a linear transformation on the input sequence to obtain query (Q), key (K), and value (V) representations:

[0135] Q=X·W q

[0136] K=X·W k

[0137] V=X·W v

[0138] Here, W q , W k and W v is a learnable weight matrix. Next, the self-attention mechanism is used to calculate the attention weight a ij and the context vector c j :

[0139] Similarity score: scor(e ij )=Q j ·K i

[0140] Attention weights:

[0141] Context vector: c j =Σ(a ij ·V i )

[0142] Finally, using the context vector c j As the target element x j context information.

[0143] Now, an adaptive window function W(i)=a*L(i)+(1-a)*C(i) is introduced, where W(i) represents the window function of the i-th element, L(i) is the window function based on the sequence length, C(i) is the window function based on the context information, and a is the weighting coefficient. The sequence length part L(i) is defined as L(i)=min(wl,max(1,ceil(λ*n))), where wl is the upper limit of the fixed window size, λ is a parameter that controls the window size (usually ranging from 0 to 1), and ceil is an upper limit function rounded up to the nearest integer. This part of the window function determines the window size based on the sequence length, limits the upper limit, and ensures that the window contains at least one element.

[0144] The context part C(i) is defined as C(i) = sum(c j *x j ) / sum(c j ), where c j represents the weight of the contextual relationship with the i-th element xi, x jrepresents the jth element in the sequence. This part of the window function calculates a weight based on the context information to determine the position of the window in the sequence.

[0145] By adjusting the weight coefficient a and the control parameter λ, the effects of sequence length and context information on the window function can be balanced.

[0146] The local alignment vector is constrained to have a fixed dimension, specifically at ∈ R 2D+1 This is different from the usual global approach. The model predicts the alignment position using the following mechanism:

[0147] S is the length of the source sequence. Through the sigmoid function, pt∈[0,S] the network will learn the model parameters to predict the position.

[0148] A Gaussian distribution centered around is introduced to emphasize neighboring alignment points. Specifically, the alignment weight is defined as follows

[0149] Fig.16 Shows the network structure after the introduction of the attention mechanism.

[0150] S4. Extract earthquake characteristic parameters from the long-period earthquake nonlinear response database through the gradient boosting decision tree model, and extract hysteresis characteristic parameters from the maximum response displacement or other structural demand parameters, and output nonlinear prediction results.

[0151] The displacement requirements of high-rise structures are related to the inter-story displacement requirements, and the results strongly depend on the structural nonlinearity and ground motion characteristics. In order to characterize the impact of earthquake characteristics on structural response, a novel hybrid model based on the original framework is proposed. The proposed hybrid deep learning model consists of two components: the extraction of earthquake characteristic parameters based on the gradient boosting decision tree (LightGradient Boosting Machine) and the step three Bi-LSTM model to extract the hysteresis characteristic parameters obtained by steps one and two. Fig.17 It is a flow chart of the hybrid model (the whole method model).

[0152] In specific use: Through numerical and experimental examples, the feasibility of the invention method is demonstrated from multiple aspects. Preliminary verification (two examples): First, the invention adopts the Bayesian updating framework system to invert the hysteresis parameter characteristics, and verifies the accuracy of the method through experimental data (such as Figure 2 ,3,4,5), and secondly, through a six-story hotel (such as Figure 6 The comparison between the real earthquake response records and the calculation results of the proposed method preliminarily verifies the effectiveness and accuracy of the proposed method (as shown in Figure 7 shown).

[0153] In-depth verification (two examples): First, for single-degree-of-freedom structures (such as Figure 8 The proposed method is compared with the traditional capacity spectrum method, correlation coefficient method (ASCE 41-172) and convolutional neural network (CNN), which verifies that the proposed method is more accurate than other methods in predicting the impact of medium- and long-period structures on long-period seismic waves (such as Fig. 9 At the same time, the method of the present invention has good interpretability (as shown in Fig.10 Secondly, by decomposing and reorganizing vibration modes, this research method is extended to multi-degree-of-freedom systems and numerically verified by actual structures (as shown in Fig.11 By comparing with the traditional method, the feasibility and accuracy of the method proposed by the present invention are verified again (as shown in Fig.12 As shown in Table 2).

[0154]

[0155] Table 2

[0156] 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. A hybrid network based on attention is used for the prediction of nonlinear response of long-period earthquake structures, characterized by: include: S1. Under the assumption that the structural parameters have a random probability distribution, the probability parameter distribution of the structure is determined by transition Markov chain Monte Carlo simulation based on the system identification scheme in the Bayesian updating framework; S2. Conduct time history analysis through a long-period earthquake record database and a probability parameter distribution based on a previously determined structure to establish a long-period earthquake nonlinear response database, wherein the long-period earthquake nonlinear response database includes a hysteresis curve database; S3. A network structure of an attention mechanism is proposed, which uses an adaptive window function to accurately calculate the precise local relationship, divides the hysteresis curve into multiple windows, and each window is predicted using an independent LSTM model to output the maximum response displacement or other structural requirement parameters; S4. extracting earthquake characteristic parameters from the long-period earthquake nonlinear response database and hysteresis characteristic parameters from the maximum response displacement or other structural demand parameters through a gradient boosting decision tree model, and outputting nonlinear prediction results; Among them, the network structure of the attention mechanism consists of an input layer, an LSTM layer, a fully connected layer, and an output layer. The proposed attention mechanism attention calculates the final hidden state and attention score vector according to the input data from different channels of the LSTM network; then the score is normalized using the SoftMax function; the vectors are aligned and summed using equations; LSTM uses a recurrent neural network to selectively recall the direct history or long-term patterns of the input sequence; the empirical wavelet transform is used to extract feature information from signals of different scales by decomposing them into sub-bands of multiple scales; extracting different time scales and frequency features in the original signal enables LSTM to capture dynamic features and nonlinear relationships in the data.

2. The method for predicting the nonlinear response of long-period earthquake structures using an attention-based hybrid network according to claim 1, characterized in that: The S1 further comprises: The MATLAB toolbox genetic algorithm (GA) and NSGA-II algorithm were used to determine the initial value of the BWBN parameter θ; The optimal estimate is given by the following equation: Determine the MSE of the objective function; The range of parameters is The surroundings are normally or evenly distributed, and their ranges are based on the capability curves in HAZUS-MH 2.1 and the pre-entered previous research records; Use the observed data to update the prior distribution and determine the posterior probability distribution of the parameters; Use the genetic algorithm in the MATLAB Global Optimization Toolbox to perform sensitivity analysis and determine the importance of parameters; Based on the results of the sensitivity analysis, the four most important parameters were fixed at their median values; The fixed parameter values ​​are used to construct the earthquake demand database while ensuring the acceptable range of mean square error.

3. The method for predicting the nonlinear response of long-period earthquake structures using an attention-based hybrid network according to claim 2, characterized in that: The S2 further includes: Extract seismic wave records that meet long-period characteristics from the NGA-West database; The basin amplification effect and site amplification effect were evaluated using Z1.0, Z1.5, Z2.5 and vs30 parameters; Construct parameter combinations to respond to different hysteresis states of the structure; Apply seismic waves to structures with different hysteresis characteristics and perform time history analysis; Nonlinear analysis was performed using the BWBN model and the Runge-Kutta numerical integration method. The results of the nonlinear analysis were extracted to form a database. Each data set included the acceleration time history of the ground motion as input and the response as output. The maximum displacement was extracted as the demand parameter. The pushover analysis was performed by applying a gradually increasing lateral force pattern to the structure to obtain the hysteresis curve. The nonlinear analysis was performed using the BWBN model, and the Runge-Kutta integration method was used as the solver. The turning point from the linear region to the nonlinear region under different degradation and shrinkage parameters was described by finding the intersection between the pushover curve and a linear line starting from the origin with a slope of 90% of the initial stiffness. The BWBN nonlinear model involves several parameter combinations of different hysteresis states of the structure. Each combination includes 240×2 input points to form a hysteresis database.

4. The method for predicting the nonlinear response of long-period earthquake structures using an attention-based hybrid network according to claim 3, characterized in that: The S3 further includes: The network structure of the attention mechanism calculates the final hidden state and attention score vector based on the input data from different channels of the LSTM model; Normalize the scores using the SoftMax function; align and sum the vectors using equations; Using empirical wavelet transform to decompose signals of different scales into sub-bands of multiple scales to extract feature information from signals of different scales; Extracting different time scales and frequency features from the original signal enables the LSTM model to capture dynamic features and nonlinear relationships in the data; Use wavelet coefficients to extract features and reduce the dimension of the original signal; At each decoding time step t, the model receives as input the hidden state of the top layer of the stacked LSTM model; Generate a sequence vector containing relevant information from the source to assist in predicting the current target; A surface concatenation layer is used to merge the target hidden state with the sequence vector on the source side to produce the attention hidden state as shown below: Initially predict the alignment position of each target token to generate a sequence vector; calculate the weighted average of the source hidden states within the specified window: w(i)∈[p t-D ,p t+D ], for a given input sequence X = [x1, x2, ..., x n ] and the target element y t =x j , first apply a linear transformation to the input sequence to obtain the query parameter Q, the field name K, and the specific data content V corresponding to K: Q=X·W q K=X·W k V=X·W v Here, Wq, Wk, and Wv are learnable weight matrices; the self-attention mechanism is used to calculate the attention weight a ij and the context vector c j : Similarity score: scor(e ij )=Q j ·K i Attention weights: Context vector: c j =Σ(a ij ·V i ) Use the context vector c j As the target element x j context information.

5. The method for predicting the nonlinear response of long-period earthquake structures using an attention-based hybrid network according to claim 4, characterized in that: S3 also includes: an adaptive window function W(i)=a*L(i)+(1-a)*C(i), where W(i) represents the window function of the i-th element, L(i) is a window function based on the sequence length, C(i) is a window function based on context information, and a is a weighting coefficient; the sequence length part L(i) is defined as L(i)=min(wl,max(1,ceil(λ*n))), where wl is the upper limit of the fixed window size, λ is a parameter that controls the window size, ranging from 0 to 1, and ceil is an upper limit function rounded up to the nearest integer; This part of the window function determines the window size based on the sequence length, limits the upper limit and ensures that the window contains at least one element; The context part C(i) is defined as C(i) = sum(c j *x j ) / sum(c j ), where c j Represents the i-th element x i The weight of the contextual relationship, x j represents the jth element in the sequence; This part of the window function calculates weights based on contextual information to determine the position of the window in the sequence; By adjusting the weight coefficient a and the control parameter λ, the effects of sequence length and context information on the window function can be balanced.

6. The method for predicting the nonlinear response of long-period earthquake structures using an attention-based hybrid network according to claim 5, characterized in that: The S3 also includes: the local alignment vector is restricted to a fixed dimension, at ∈ R 2D+1 At , the alignment position is predicted using the following mechanism: S is the length of the source sequence.

Images