A deep learning method suitable for predicting nonlinear floor acceleration response spectrum
By constructing a bidirectional convolutional long short-term memory network with an attention mechanism and a deep residual regression network, the problems of prediction accuracy and computational efficiency of nonlinear floor acceleration response spectrum are solved, and efficient and accurate prediction of floor acceleration response spectrum of arbitrary structures is achieved.
Patent Information
- Application Number
- CN202310182776.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-03-01
- Publication Date
- 2025-12-12
- Estimated Expiration
- 2043-03-01
AI Technical Summary
Existing deep learning methods suffer from gradient vanishing and data explosion problems when predicting nonlinear floor acceleration response spectra, leading to decreased prediction accuracy. Furthermore, multiple models need to be trained for different structures, increasing computational and modeling costs.
A model is constructed to predict the floor acceleration response spectrum of arbitrary structures by employing a bidirectional convolutional long short-term memory network with an attention mechanism and a deep residual regression network, combined with a multi-scale sliding time window. The attention mechanism extracts important information, and the deep residual regression network improves the prediction accuracy and efficiency.
It achieves accurate prediction of acceleration response spectrum of any floor in any structure, simplifies the modeling process, reduces computational cost, improves computational efficiency, and has good generalization ability and high accuracy.
Smart Images
Figure CN116415623B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of earthquake engineering, and particularly relates to a deep learning method suitable for predicting nonlinear floor acceleration response spectrum. BACKGROUND
[0002] For a long time, earthquakes as a kind of uncertainty and sudden natural disasters have always posed a great threat to the safety of buildings. In addition, with the increasing development of urban disaster resistance and safety, the focus of disaster prevention and mitigation research is shifting from single buildings to urban buildings. Therefore, at the present stage, the seismic analysis of urban areas is increasingly concerned, and the rapid prediction of the seismic response of urban buildings has become the main technical challenge at the present stage. This usually requires researchers to design buildings according to different seismic requirements, and the floor response spectrum method as the most commonly used method in building seismic calculation is widely studied.
[0003] At the present stage, the method for obtaining the floor response spectrum usually involves time history analysis method, response spectrum-floor response spectrum method, finite element method and random vibration theory method; however, such traditional models and static analysis often involve complex mathematical and physical models, which usually means high modeling or calculation cost and is difficult to achieve high-precision real-time calculation. Therefore, it is increasingly important to study an automatic prediction method for floor acceleration response spectrum to reduce the model calculation cost and achieve high-precision real-time calculation of the model.
[0004] At present, deep learning as a novel artificial intelligence method has the great advantage of automatically mining potential features from data; however, this method of combining seismic structural information data and response has the following problems when applied to nonlinear floor acceleration response spectrum:
[0005] 1) The existing deep learning methods such as one-dimensional convolutional neural network (1D-CNN) and RNN model, when predicting such nonlinear floor response spectrum data with time sequence information, will cause the prediction accuracy of the model to decrease significantly due to problems such as gradient disappearance and data explosion.
[0006] 2) In addition, the existing deep learning methods often need to train many models using the information of different structures when predicting the acceleration response spectrum of multiple different structures, which means that the model needs to be trained separately for specific structural information, which often wastes a lot of calculation and modeling cost.
[0007] Therefore, in order to overcome the above series of problems, it is necessary to design a high-precision nonlinear floor acceleration response spectrum prediction method that integrates multiple structures, so as to solve the limitations of nonlinear structural response modeling and prediction. SUMMARY
[0008] The purpose of the present application is to overcome the problems existing in the prior art, provide a deep learning method suitable for predicting nonlinear floor acceleration response spectrum, which utilizes a bidirectional convolutional long short-term memory network with attention mechanism to rely on attention mechanism, multi-scale sliding time window and deep residual regression network to achieve the purpose of precise prediction of any floor acceleration response spectrum in any structure by a model, and can simplify the complex modeling process and improve the calculation efficiency.
[0009] To achieve the above purpose, the present application adopts the following technical scheme: a deep learning method suitable for predicting nonlinear floor acceleration response spectrum, comprising the following steps:
[0010] S1, data preprocessing is performed on the ground motion and floor acceleration response spectrum dataset GMS-FRS with timestamp:
[0011] The top floor acceleration response spectrum of the structure and the corresponding ground motion response spectrum are respectively normalized and pretreated on the time axis according to the structure period, and the normalized time series data is subjected to cubic spline interpolation operation; the floor height of the structure is also subjected to the same normalization pretreatment, and the proportion factor between the acceleration response spectrum value of each floor and the top floor acceleration response spectrum value under the structure modal period is calculated;
[0012] S2, a bidirectional convolutional long short-term memory network model with attention mechanism is constructed:
[0013] The bidirectional convolutional long short-term memory network structure with attention mechanism is constructed by a 1-D convolutional neural network layer, a Bi-LSTM network layer, an Attention attention layer and a full connection layer, and the four-layer network is arranged between the input layer and the output layer in the structure, wherein the 1-D convolutional neural network layer is composed of a plurality of one-dimensional convolutional neural units, a full connection layer is arranged at both ends of the 1-D convolutional neural network layer and the output layer, the front-end full connection layer is composed of a plurality of neural units, and the rear-end full connection layer is composed of one neural unit;
[0014] S3, the bidirectional convolutional long short-term memory network model with attention mechanism is trained by inputting the ground motion and floor acceleration response spectrum dataset processed by step S1:
[0015] The ground motion and floor acceleration response spectrum data pretreated in the above step S1 are regarded as multivariate time series features, and the i-th observed data x i is input into the model constructed in the above step S2 for training; first, a residual block and a full connection layer are built by a 1-D convolutional neural network layer, and the input data x iPerform feature extraction to obtain the hidden representation of the time series. With the context vector in the sequence The similarity information is used to calculate the importance weight α of the conditional probability distribution through time steps. i The summed vector C is then processed by the Attention layer. t Calculate for all hidden states h t to h t-1 A weighted sum based on weights, and the summation vector C is generated. t and the hidden state h at the current moment t Combined to generate attention vector h * Finally, the attention vector h is obtained. * Forward propagation is performed to the fully connected layer to predict the data for the next time step, ultimately obtaining the acceleration response spectrum F of the top floor in any structure. Total ;
[0016] in, In the formula, Mish represents the Mish activation function, W w b represents the weight coefficient of the hidden state. w The bias coefficient for the hidden state;
[0017] In the formula, exp represents the natural exponential function. Represented as h t The hidden representation;
[0018] In the formula, h i This represents the hidden state at the current moment;
[0019] h * =f(C t ,h t =Mish(Wc[C t h t In the formula, Wc is the summation vector C. t Weighting coefficients;
[0020] S4. Construct a deep residual regression network model:
[0021] The deep residual regression network structure is constructed by five deep residual blocks, one Flatten layer and three fully connected layers. The first convolutional layer uses ordinary one-dimensional convolution, and the remaining convolutional layers all use deep one-dimensional convolution. Each deep residual block contains two deep one-dimensional convolutional layers, two PReLU non-linear activation functions, one skip connection and one max pooling layer.
[0022] S5. Input the normalized floor heights to train the deep residual regression network model:
[0023] The normalized floor height and the acceleration response spectrum of each floor in step S1 above are input into the model as regression training of the scale factor of the top floor acceleration response spectrum, and several deep residual modules are used for automatic feature extraction and interactive fusion of the data. The data is flattened by the Flatten layer and then forward propagated to the fully connected layer to realize regression analysis of the normalized floor height and the scale factor. The inter-story response spectrum scale factor S T1 is obtained by selecting the first period at the modal period and the peak value at the second period in the non-linear floor acceleration response spectrum of the structure T2 , the initial value at the origin F T0 , and the minimum value between the peak points at the modal period Tmin The scale factor at the four points is t ;
[0024] The calculation formula of the inter-story response spectrum scale factor S t is as follows:
[0025] In the formula, represents the response spectrum value of any floor of the structure at period i, represents the response spectrum value of the top floor of the structure at period i;
[0026] S6, the top floor acceleration response spectrum of the structure and the inter-story response spectrum scale factor are combined and input into the deep residual regression network model for regression analysis:
[0027] The top floor acceleration response spectrum of any structure obtained in step S3 above is fitted into a three-section standardized response spectrum according to the following formula, and the specific formula is as follows:
[0028]
[0029] In the formula, C, α are constants, β max represents the maximum amplification factor, X represents the response spectrum vector of any floor of the structure, θ represents the weight coefficient corresponding to the response spectrum vector, t represents the period value, t0, t g , t m represent the maximum period value of each section of the three-section standardized response spectrum.
[0030] The top floor standardized response spectrum and the inter-story response spectrum scale factor S t obtained in step S5 above are combined and input into the model for regression training, which obtains the acceleration response spectrum of any floor of any structure. The top floor standardized response spectrum curve is adjusted using the response spectrum scale factor S t , and finally the standardized acceleration floor response spectrum of any floor of any structure with double peak characteristics is obtained.
[0031] In step S1, the normalization calculation formula is as follows:
[0032]
[0033] In the formula, and The time vector representing a GMS data sample. and T represents the time vector in the FRS data sample. s This represents the basic cycle of architecture.
[0034] In step S3, before inputting the model, a multi-scale sliding time window method is used for data preprocessing. Specifically, sliding and convolution operations are performed on the input original time series information according to a given step size, and the original time series is divided into fixed-length subsequences [X1, X2, X3, ... X6]. After feature extraction is performed window by window according to the given sliding step size, the new features obtained from the original time series are arranged in natural order to obtain a new long-term time series vector [X1, X2, X3, ... X6]. w1 ,X w2 ,X w3 ,…X w6 The new long-term vector is then used as the input to the model.
[0035] In step S5, the interlayer reaction spectrum scaling factor S is obtained using GMS-FRS data. t And normalized floor height, where the scaling factor corresponding to the peak value at the structural modal period is: the scaling factor S corresponding to the first two modal periods. T1 and S T2 Based on structural modal eigenvalue analysis, the scaling factor S between the minimum values of peak points at the modal period is determined. Tmin Corresponding periodic point T min Obtained through linear search.
[0036] In step S6, the coefficients α and t0 to t0 in the reaction spectrum function relationship from 0 to t0 are... g Phase β max t is obtained by relating the scaling factors at 0 and the peak value to the corresponding time. g to t m The parameter θ in the stage function relation is obtained by training with the response spectrum data within that time period.
[0037] The beneficial effects of this invention are:
[0038] 1) This invention optimizes the bidirectional convolutional long short-term memory network using an attention mechanism, enabling the model to quickly lock in important information from long-term sequence learning and ignore irrelevant information, thus completing the task more efficiently. The proposed bidirectional convolutional long short-term memory network with attention mechanism can significantly enhance the changes between samples and suppress the cumulative error between adjacent time points by relying on the attention mechanism and multi-scale sliding time window, thereby enabling the model to exhibit good generalization ability. Furthermore, it can achieve excellent ability to accurately predict the acceleration response spectrum of the top floor in any structure by using time-stamped ground motion and floor response spectrum data.
[0039] 2) The deep residual regression model proposed in this invention adopts a deep one-dimensional convolution and a novel network structure design concept to increase the training depth of the network and improve the regression accuracy, which can achieve accurate prediction of acceleration response spectra of other floors.
[0040] 3) This invention inputs timestamped ground motion and floor response spectrum into a deep learning model for training, which solves the limitation that different models need to be built to predict the floor response spectrum in different structures. Only one model is needed to accurately predict the acceleration response spectrum of any floor in any structure. The entire training process does not require manual intervention, which reduces the complex modeling process and improves the computational efficiency. To a certain extent, it overcomes the limitations of high computational cost and complex modeling in traditional methods, and realizes the automated prediction of nonlinear floor acceleration response spectrum. Attached Figure Description
[0041] Figure 1 This is a flowchart of the design scheme of the present invention;
[0042] Figure 2 This is a diagram of the bidirectional convolutional long short-term memory network architecture with an attention mechanism constructed in this invention.
[0043] Figure 3 for Figure 2 A schematic diagram of the attention mechanism in the brain;
[0044] Figure 4 This is a schematic diagram of the multi-scale sliding time window used in this invention;
[0045] Figure 5 This is a diagram showing the predicted results of the top floor reaction spectrum in this invention;
[0046] Figure 6 This is a diagram of the deep residual regression network architecture built in this invention;
[0047] Figure 7 This is a comparison chart of the fitted real reaction spectrum and the standardized reaction spectrum in this invention. Detailed Implementation
[0048] The application will be further explained in connection with the accompanying drawings and specific embodiments.
[0049] Embodiments: As shown in the drawings, the application provides a deep learning method suitable for predicting nonlinear floor acceleration response spectrum, comprising the following steps: Figures 1-7
[0050] S1, data preprocessing is performed on the ground motion and floor acceleration response spectrum dataset GMS-FRS with time stamp:
[0051] The top floor acceleration response spectrum of the structure and the ground motion response spectrum corresponding to the floor acceleration response spectrum are respectively normalized on the time axis according to the structure period, and the calculation formula of the normalization processing is:
[0052] In the formula, and represent the time vector of the GMS data sample, and represent the time vector T in the FRS data sample. s represent the basic period of the building.
[0053] According to the above method, the top floor acceleration response spectrum of 56 structures is normalized on the time axis according to the structure period, and after the normalization operation, the top floor acceleration response spectrum of 56 structures is independent of the dynamic characteristics of each structure; In addition, the floor height of the structure is also normalized and the proportion factor between the acceleration response spectrum value of each floor and the acceleration response spectrum value of the top floor is calculated.
[0054] Due to the difference in periodicity of each structure, the time interval of the normalized time column also differs, and considering the characteristics of the time interval of the time sequence required for algorithm training, the normalized time sequence data is subjected to cubic spline interpolation operation; Due to the reason that the higher the floor number, the greater the period, the time interval of the time column of the high-rise structure becomes smaller after normalization, therefore, for the interpolation operation, the time interval 0.003829s in the highest structure is taken as the value of the cubic spline interpolation, that is, the time interval between each data is 0.003829s, and 601 sample points are considered for each structure, and a total of 33656 data sets are obtained. The 33656 samples after data preprocessing are taken as the input of the bidirectional convolution long short-term memory network model with attention mechanism for training.
[0055] S2, constructing a bidirectional convolution long short-term memory network model with attention mechanism:
[0056] As shown in the drawings, Figure 2 As shown, the bidirectional convolutional long short-term memory network structure with attention mechanism is constructed by a 1-D convolutional neural network layer, a Bi-LSTM network layer, an Attention attention layer and a full connection layer, and the four-layer network is arranged between the input layer and the output layer in the structure; the 1-D convolutional neural network layer is composed of a plurality of one-dimensional convolutional neural units, one full connection layer is arranged at each end between the 1-D convolutional neural network layer and the output layer, the front-end full connection layer is composed of 512 neural units and converts the output of the 1-D convolutional neural network layer into an input vector of the Bi-LSTM network layer, the rear-end full connection layer is composed of one neural unit and reduces the feature dimension received by the Attention attention layer, and the Bi-LSTM network layer and the Attention attention layer are arranged between the two full connection layers.
[0057] S3, inputting the ground motion and floor acceleration response spectrum data set processed by step S1 to train the bidirectional convolutional long short-term memory network model with attention mechanism:
[0058] The ground motion and floor acceleration response spectrum data processed in step S1 are regarded as multivariate time series features, and the i-th observed data x i is input into the model constructed in step S2 for training.
[0059] As shown in Figure 3 , first, a residual block and a full connection layer are built by a 1-D convolutional neural network layer, and the input data x i is feature-extracted by the 1-D convolutional neural network to obtain hidden representations of the time series and similarity information of the context vector in the sequence, and the importance weight a i of the conditional probability distribution is calculated by a time step, which represents the importance of the i-th sliding time window to the prediction task; then, the summary vector C t is calculated by the Attention attention layer as a weighted sum of all hidden states h t to h t-1 based on the weight, and the summary vector C t and the hidden state h t at the current time are combined to generate an attention vector h * , finally, the obtained attention vector h * is forward propagated to the full connection layer to predict the data at the next time, and finally, the top floor acceleration response spectrum F Total in any structure is obtained.
[0060] wherein, In the formula, Mish represents the Mish activation function, W w b represents the weight coefficient of the hidden state. w The bias coefficient for the hidden state;
[0061] In the formula, exp represents the natural exponential function. Represented as h t The hidden representation;
[0062] In the formula, h i This represents the hidden state at the current moment;
[0063] h * =f(C t ,h t =Mish(Wc[C t h t In the formula, Wc is the summation vector C. t Weighting coefficients;
[0064] like Figure 4 As shown, before inputting the model, a multi-scale sliding time window method is needed for data preprocessing. Specifically, sliding and convolution operations are performed on the input original time series information according to a given step size, and the original time series is divided into fixed-length subsequences [X1, X2, X3, ... X6]. Taking the multi-scale sliding time window K1 as an example, the length of the entire window is 5, the width is 3, and the multi-scale scaling factor is set to 1 (representing that features are extracted once every time point). After feature extraction is performed window by window according to the given sliding step size, the new features obtained from the original time series are arranged in natural order to obtain a new long time series vector [X1, X2, X3, ... X6]. w1 ,X w2 ,X w3 ,…X w6 The new long-time vector is then used as the input to the model. The input shape of each Bi-LSTM unit is (1,t), and the sliding step is used to determine the next position in the detection time series. If the first subsequence starts at t0, the next position will be t0. 0+s , where S represents the step size of the sliding time window; after the above preprocessing operations, the original long time-series vector is converted into a smaller sub-vector and input into the Bi-LSTM unit for training, thereby achieving the task of predicting the data at the next time step.
[0065] The length of the model input sequence is set to 6, the dimension is 3, and the prediction length is 1, that is, every 3 sequence lengths are predicted for 1 sequence length, wherein 80% of the data set is used for training and 20% is used for testing; the model predicts a total of 8408 time sequence samples, the network is trained for a total of 1000 epochs, the batch size is 2048, the loss function is mse, and the model is optimized using the Adam optimizer.
[0066] As shown in Figure 5 The prediction results of the model are shown, wherein the horizontal axis represents the number of prediction samples, and the vertical axis represents the acceleration response spectrum of the floor. As can be seen from the figure, the change trend of the response spectrum of the entire floor is very similar, and the prediction value of the model is in good agreement with the true value. In order to more clearly analyze the fitting of the prediction value and the true value, a detailed zoomed-in view of the prediction information is drawn in each subgraph. As can be seen in the zoomed-in view, the prediction result of the model is almost consistent with the true value, which clearly shows that the bidirectional convolutional long short-term memory network model with attention mechanism has excellent prediction ability and can predict the top floor response spectrum in any structure using the ground motion response spectrum with a timestamp.
[0067] S4, build a deep residual regression network model:
[0068] As shown in Figure 6 , a deep residual regression network structure is built by five deep residual blocks, a Flatten layer and three fully connected layers, wherein the first convolutional layer uses ordinary one-dimensional convolution and the remaining convolutional layers use deep one-dimensional convolution. Each deep residual block includes two deep one-dimensional convolutional layers, two PReLU nonlinear activation functions, a skip connection and a maximum pooling layer.
[0069] Each convolutional layer has 256 kernels with a size of 1, and all the pooling layers use maximum pooling with a size of 1 and a stride of 2. During the network building process, the shallow features of the data set processed by the convolutional layer are also connected to each deep residual block through a residual shortcut connection, increasing the depth of the network. Finally, the output of the convolutional layer is flattened using the Flatten layer, and then forward propagated to the fully connected layer for dimension reduction. The output of the last fully connected layer with one neuron is used as the prediction result.
[0070] S5, input normalized floor height for deep residual regression network model regression training:
[0071] The normalized floor height and the acceleration response spectrum of each floor in step S1 are input into the model as regression training of the proportional factor of the top floor acceleration response spectrum, and several deep residual modules are used for automatic feature extraction and interactive fusion of the data. The data is flattened by the Flatten layer and then forward propagated to the fully connected layer to realize the regression analysis of the normalized floor height and the proportional factor. The proportional factor of the inter-story response spectrum S is obtained by selecting the peak value F T1 at the first period, the peak value F T2 at the second period, the initial value F T0 at the 0 point, and the minimum value F Tmin between the two peak values in the nonlinear floor acceleration response spectrum of the structure. t ;
[0072] The calculation formula of the inter-story response spectrum proportional factor S t is as follows:
[0073] In the formula, represents the response spectrum value of any floor of the structure at period i, represents the response spectrum value of the top floor of the structure at period i; the inter-story response spectrum proportional factor S t and the normalized floor height are obtained by using GM-FRS data, wherein the proportional factor corresponding to the peak value at the modal period is: the proportional factor S T1 and S T2 corresponding to the first two modal periods. Tmin The proportional factor S min corresponding to the minimum value between the peak points at the modal period is determined according to the modal characteristic value analysis of the structure.
[0074] S6, the top floor acceleration response spectrum of the structure and the inter-story response spectrum proportional factor are combined and input into the deep residual regression network model for regression analysis.
[0075] The top floor acceleration response spectrum of any structure obtained in step S3 is fitted into a three-section standardized response spectrum (F Figure 7 ) according to the following formula, and the specific formula is as follows:
[0076]
[0077] In the formula, C, α are constants, β max represents the maximum amplification factor, X represents the response spectrum vector of a floor in any structure, θ represents the weight coefficient corresponding to the response spectrum vector, t represents the period value, t0, t g , and t m represent the maximum period value of each section of the three-section standardized response spectrum.
[0078] The top-level normalized reaction spectrum and the interlayer reaction spectrum scaling factor S obtained in step S5 above are used. t Regression training is performed using the input model to obtain the acceleration response spectrum of any floor in any structure, and the response spectrum scaling factor S is used. t The standardized response spectrum curve of the top floor is adjusted to obtain the standardized acceleration floor response spectrum of any floor in any structure with bimodal characteristics. Figure 7 );
[0079] The coefficients α and t0 in the reaction spectrum function relationship from 0 to t0 g Phase β max t is obtained by relating the scaling factors at 0 and the peak value to the corresponding time. g to t m The parameter θ in the stage function relation is obtained by training with the response spectrum data within that time period.
[0080] The standardized floor response spectrum results fitted using the above method are as follows: Figure 7 As shown, the spectral pattern variation of the actual seismic response spectrum curve within the frequency bands where it appears in the rising and falling segments is complex and irregular. However, the standardized response spectrum calibration method using a bidirectional convolutional long short-term memory network with attention mechanism and a deep residual regression network model can describe the true spectral characteristics of the relevant response spectrum to a certain extent.
[0081] The method of this invention utilizes a bidirectional convolutional long short-term memory network with an attention mechanism. By relying on the attention mechanism, multi-scale sliding time windows, and deep residual regression network, it achieves the goal of accurately predicting the acceleration response spectrum of any floor in any structure with a single model. It can also simplify the complex modeling process and improve computational efficiency.
[0082] The above description is only used to illustrate the technical solution of the present invention and is not intended to limit it. Any other modifications or equivalent substitutions made by those skilled in the art to the technical solution of the present invention, as long as they do not depart from the spirit and scope of the technical solution of the present invention, should be covered within the scope of the claims of the present invention.
Claims
1. A deep learning method suitable for predicting nonlinear floor acceleration response spectra, characterized by: The method comprises the following steps: S1, data preprocessing is performed on the ground motion and floor acceleration response spectrum dataset GMS-FRS with a timestamp: The top floor acceleration response spectrum of the structure and the corresponding ground motion response spectrum are respectively normalized and pretreated on the time axis according to the structure period, and the normalized time series data is subjected to a cubic spline interpolation operation; The floor height of the structure is also normalized and pretreated, and the ratio factor between the acceleration response spectrum value of each floor and the top floor acceleration response spectrum value under the structure modal period is calculated; S2, a bidirectional convolutional long short-term memory network model with an attention mechanism is constructed: The bidirectional convolutional long short-term memory network structure with an attention mechanism is constructed by a 1-D convolutional neural network layer, a Bi-LSTM network layer, an Attention attention layer and a full connection layer, four network layers are arranged between the input layer and the output layer in the structure, the 1-D convolutional neural network layer is composed of a plurality of one-dimensional convolutional neural units, a full connection layer is arranged at both ends of the 1-D convolutional neural network layer and the output layer, the front end full connection layer is composed of a plurality of neural units, and the rear end full connection layer is composed of one neural unit; S3, the bidirectional convolutional long short-term memory network model with an attention mechanism is trained by inputting the ground motion and floor acceleration response spectrum dataset processed in step S1: The preprocessed ground motion and floor acceleration response spectrum data from step S1 are considered as multivariate time series features. Data observed within a multi-scale sliding window are then used to... i Observed data x i The input data is fed into the model constructed in step S2 above for training; firstly, residual blocks and fully connected layers are built through 1-D convolutional neural network layers, and then the 1-D convolutional neural network is used to process the input data. x i Perform feature extraction to obtain the hidden representation of the time series. With the context vector in the sequence Similarity information is used to calculate the importance weights of the conditional probability distribution through time steps. The aggregated vectors are then processed by the Attention layer. C t Calculate for all hidden states h t arrive h t-1 A weighted sum based on weights, and a summary vector. C t and the hidden state at the current moment h t Combined to generate attention vectors h * Finally, the attention vector obtained h * The data is forward propagated to the fully connected layer to predict the data at the next moment, ultimately obtaining the acceleration response spectrum of the top floor in any structure. F Total ; wherein, =Mish ( W w h t + b w ), in which, Mish is represented as a Mish activation function, W w is a weight coefficient of a hidden state, b w is a bias coefficient of a hidden state; wherein, exp is expressed as a natural exponential function, is expressed as h t hidden representation; wherein h i denotes the hidden state at the current time instant; h * = f ( C t , h t ) =Mish ( W c [ C t ; h t In the formula, W c For summary vector C t Weighting coefficients; S4, a deep residual regression network model is built: The deep residual regression network structure is built by five deep residual blocks, one Flatten layer and three full connection layers, wherein the first convolutional layer adopts ordinary one-dimensional convolution, the remaining convolutional layers all adopt deep one-dimensional convolution, each deep residual block comprises two deep one-dimensional convolutional layers, two PReLU nonlinear activation functions, a skip connection and a maximum pooling layer; S5, the deep residual regression network model is regression trained by inputting the normalized floor height: The normalized floor height in step S1 above and the acceleration response spectrum of each floor are taken as the input of the model for regression training, and several deep residual modules are used for automatic feature extraction and interactive fusion of the data. After the data is flattened by the Flatten layer, it is forward propagated to the fully connected layer to realize the regression analysis of the normalized floor height and the scale factor. The first period and the peak value at the second period in the modal period of the non-linear floor acceleration response spectrum in the structure are selected to obtain the inter-story response spectrum scale factor F T1 With F T2 The initial value at 0 point F T0 The minimum value between the modal period and the peak point F Tmin The scale factor at the four points S t ; where the interlayer reaction spectrum scaling factor S t The formula for calculating is: , In the formula, This indicates that any floor of the structure has a period of i The reaction spectrum value at that location, The top floor of the structure has a period of i The reaction spectrum value at the location; S6, the top floor acceleration response spectrum and the inter-story response spectrum ratio factor in the structure are combined to input the deep residual regression network model for regression analysis: The top floor acceleration response spectrum in any structure obtained in step S3 is fitted into a three-section standardized response spectrum according to the following formula, and the specific formula is: ， wherein C , is a constant, β max represents the maximum amplification factor, X represents a response spectrum vector of a certain floor in an arbitrary structure, θ represents a weight coefficient corresponding to the response spectrum vector, t represents a period value, t 0, t g , t m represents the maximum period value of each segment in the three-segment standardized response spectrum; using the top floor normalized response spectrum and the inter-story response spectrum scaling factor obtained in step S5 above S t In combination with the regression training in the input model, the input model is trained to obtain the normalized acceleration response spectrum of any floor in any structure using the response spectrum scaling factor S t Adjusting the top floor normalized response spectrum curve, the normalized acceleration response spectrum of any floor in any structure with a bimodal characteristic is finally obtained.
2. The deep learning method suitable for predicting nonlinear floor acceleration response spectrum according to claim 1, wherein: In step S1, the calculation formula of the normalization processing is: ， ， wherein and a time vector representing the GMS data samples, and a time vector representing the FRS data samples, T s represents the fundamental period of the building.
3. The deep learning method for predicting nonlinear floor acceleration response spectrum according to claim 1, wherein: In the step S3, a multi-scale sliding time window method is used for data preprocessing before model input. Specifically, sliding and convolution operations are performed on the input original time series information according to a given step, and the original time series is divided into fixed-length subsequences. X 1, X 2, X 3, … X 6] After feature extraction according to a given sliding step for each window, the new features obtained from the original time series are arranged in natural order to obtain a new long sequence vector. X w1 , X w2 , X w3 , … X w6 ] The new long sequence vector is used as the input of the model.
4. The deep learning method for predicting nonlinear floor acceleration response spectrum according to claim 1, wherein: In step S5, the inter-story response spectrum scaling factor is obtained using the GMS-FRS data S t and the normalized story height, where the scaling factor corresponding to the peak at the structural modal period: the scaling factor corresponding to the first two modal periods S T1 and S T2 The scaling factor at the minimum between the peak points at the modal periods is determined from the structural modal eigenvalue analysis S Tmin corresponding to the period point T min is obtained by linear search.
5. The deep learning method for predicting nonlinear floor acceleration response spectrum according to claim 1, wherein: The step S6, 0 to t The coefficient in the function relationship of the 0 stage reaction spectrum 、 t The step S6, 0 to t g The function relationship of the 0 stage β max The corresponding time relationship is obtained by the proportional factor at the 0 point and the peak value respectively, t g The step S6, 0 to t m The parameter in the function relationship of the 0 stage θ In step S1, the calculation formula of the normalization processing is: It is obtained by training the reaction spectrum data in this time period.