Major-aftershock sequence seismic response prediction method based on knowledge enhancement deep learning model

Abstract

The invention discloses a main shock and aftershock sequence earthquake response prediction method based on a knowledge enhancement deep learning model, and belongs to the technical field of building structure earthquake response prediction. The method solves the problems that an existing deep learning model is scarce in training data, poor in capacity of capturing long sequence global correlation and incapable of processing input vibration with different lengths and sampling frequencies. The method specifically comprises the following steps: step 1, establishing a target structure finite element model, and selecting and adjusting seismic oscillation records of main and aftershocks; step 2, constructing a training data set based on the finite element model and the seismic oscillation record; 3, building a knowledge enhancement deep learning model comprising a consistency input coding module, an encoder, a decoder and a linear layer, designing a knowledge enhancement loss function, and training the model by using the training data set; and step 4, inputting the actual main shock and aftershock seismic oscillation sequence into the trained model to obtain a structure displacement response and structure restoring force prediction result. The method can be applied to earthquake response prediction.

Description

Technical Field

[0001] The present invention belongs to the technical field of earthquake response prediction of building structures, and specifically relates to a method for predicting earthquake response of mainshock and aftershock sequences based on a knowledge-enhanced deep learning model. Background Art

[0002] In recent years, earthquakes have occurred frequently. Earthquakes are extremely destructive and pose a significant threat to building structures and infrastructure. At the same time, post-earthquake structural damage can also induce secondary disasters such as fires and gas leaks, further exacerbating casualties and economic losses. Structural seismic response prediction can provide a scientific basis for post-earthquake structural damage assessment, thereby effectively helping to formulate emergency response plans such as building safety determination, rescue resource allocation, and personnel evacuation organization. In addition, rapid earthquake response prediction can efficiently identify high-risk areas and quickly assess potential safety hazards, which is crucial for post-disaster risk delineation and recovery and reconstruction plan formulation. Therefore, quickly and accurately predicting structural seismic response in earthquake events has become a key link in disaster reduction and rescue.

[0003] In a large number of earthquake events, the occurrence of the main shock usually induces multiple aftershocks. The harm of the main aftershock sequence to building structures is mainly reflected in the cumulative damage effect of the structure. The time interval between the main shock and the aftershock is short. The main shock-damaged structure is subjected to aftershocks without repair, which aggravates the nonlinear behavior of the structure and increases the potential collapse risk of the structure. In addition, the occurrence of aftershocks is uncertain, which significantly increases the safety hazards in post-disaster rescue and has an adverse impact on the risk classification of the disaster area. Therefore, in the current earthquake resistance and disaster reduction, the demand for the prediction of structural response under the action of the main aftershock sequence is constantly increasing. The rapid prediction of the long-sequence earthquake response of the structure is of great significance in improving the resilience of building structures and reducing post-earthquake losses.

[0004] Traditional structural seismic response assessment mainly relies on on-site sensor monitoring and structural finite element calculation. Sensors are usually installed in some important buildings, which can monitor the structural response in real time during earthquake events and provide an important basis for post-earthquake structural damage assessment. However, due to the limitations of cost, technology and site conditions, sensors are usually only set up at key locations of important buildings and cannot cover all buildings in the disaster area; in addition, under the action of multiple strong earthquakes, sensor equipment may be damaged or fail, thus failing to provide reliable real-time data. The finite element method is a classic method for calculating the seismic response of structures through numerical simulation, which can make high-precision predictions of the stress and deformation of structures. However, finite element analysis usually consumes a lot of time and computing resources, which seriously restricts the rapid post-earthquake assessment of structures. At the same time, finite element calculation also depends on the accuracy of the established model. Any deviation in material properties, boundary conditions or load input will affect the accuracy of the calculation results. It can be seen that the traditional seismic response assessment method cannot achieve rapid and accurate prediction of the seismic response of structures after earthquakes.

[0005] With the rapid development of artificial intelligence and machine learning technologies, deep learning models have empowered various industries. Deep learning has the advantages of capturing complex patterns, strong adaptability and automatic feature extraction in the prediction of structural seismic response. It can efficiently predict and quantify the response of buildings in earthquakes, overcoming the limitations of traditional earthquake response prediction methods. The prediction performance of deep learning models usually depends on a large amount of high-quality training data, but the epicenter structural monitoring data is scarce and the high-precision numerical model simulation calculation cost is high, resulting in the data available for model training is often very limited. At the same time, deep learning models are difficult to explain the working mechanism inside the model and cannot provide clear physical meaning or causal relationship of structural response. In addition, when dealing with ultra-long earthquake sequences such as main shocks and aftershocks, deep learning models face problems such as gradient vanishing and explosion, which seriously limits the ability of existing time series models to capture the global correlation of long sequences. Finally, when training traditional deep learning models, the input ground motion must have the same length and sampling frequency. However, due to the differences in accelerometers at different stations, actual earthquake records usually have irregular time intervals, which greatly inconveniences the practical application of deep learning models. Therefore, current deep learning models still face many challenges in predicting structural responses under long earthquake sequences. In order to solve the problems of scarce training data, poor ability to capture global correlations of long sequences, and inability to handle input ground motions with different lengths and sampling frequencies of existing deep learning models, it is urgent to propose an efficient, accurate and reliable long sequence structural seismic response prediction method. Summary of the invention

[0006] The purpose of the present invention is to solve the problems of scarcity of training data, poor ability to capture global correlation of long sequences and inability to process input seismic motions with different lengths and sampling frequencies in existing deep learning models, and to propose a mainshock-aftershock sequence seismic response prediction method based on knowledge-enhanced deep learning model.

[0007] The technical solution adopted by the present invention to solve the above technical problems is: a method for predicting earthquake response of main shock and aftershock sequence based on knowledge enhanced deep learning model, the method specifically comprises the following steps:

[0008] Step 1: Establish a finite element model of the target structure, select M main shock and aftershock ground motion records for the target structure, and adjust the ground motion intensity of the selected main shock and aftershock ground motion records to obtain M′ main shock and aftershock ground motion records after adjustment;

[0009] Step 2: Use the adjusted M′ mainshock and aftershock seismic records to perform nonlinear time history analysis on the target structure, and obtain the structural displacement response data and structural restoring force data of the target structure under each mainshock and aftershock seismic record, that is, obtain a training data set consisting of the mainshock and aftershock seismic sequence, structural displacement response data, and structural restoring force data;

[0010] Step 3: Build a knowledge-enhanced deep learning model including a consistent input encoding module, an encoder, a decoder, and a linear layer, and use the training data set to train the knowledge-enhanced deep learning model. During the training process, the main shock and aftershock seismic sequence is used as the input of the knowledge-enhanced deep learning model, and the structural displacement response data and structural resilience data are used as the output of the knowledge-enhanced deep learning model.

[0011] The loss function used in the training process is a knowledge enhancement loss function, and the training is stopped when the value of the knowledge enhancement loss function converges to obtain a trained knowledge enhancement deep learning model;

[0012] Step 4: Input the actual main shock and aftershock seismic motion sequence into the trained knowledge-enhanced deep learning model, and predict the structural displacement response and structural restoring force under the actual main shock and aftershock seismic motion sequence through the trained knowledge-enhanced deep learning model.

[0013] The beneficial effects of the present invention are:

[0014] (1) The method of the present invention utilizes a consistent input encoding module to cope with ground motion data with different sampling frequencies and lengths, ensuring that the model maintains stable performance when processing diverse inputs;

[0015] (2) The encoder-decoder framework based on the embedded attention mechanism realizes the parallel operation of ground motion input, which greatly improves the model prediction efficiency;

[0016] (3) By introducing the probabilistic sparse self-attention mechanism and distillation mechanism, memory usage can be effectively optimized and the model’s ability to capture ultra-long-term correlation relationships in sequence data is improved;

[0017] (4) The knowledge enhancement loss function that integrates the structural dynamics equation is adopted, so that the model still has high training efficiency and prediction accuracy with limited training data.

[0018] The method of the invention can efficiently, accurately and reliably predict the structural response under the action of ultra-long earthquake sequences, and can provide scientific guidance for the formulation of rescue response plans, structural risk assessment, etc. It is of great significance to enhance the society's ability to respond to disasters and improve the earthquake resistance and disaster reduction system. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1It is a flow chart of a method for predicting earthquake response of main shock and aftershock sequence based on knowledge enhanced deep learning model of the present invention;

[0020] Figure 2 The ground motion characteristic distribution diagram of the main shock and aftershock selected for the present invention;

[0021] Figure 3 It is a schematic diagram of the structure of the consistency input encoding module;

[0022] Figure 4 This is the structural diagram of the Phyformer deep learning model;

[0023] Figure 5 This is the displacement prediction result diagram of the deep learning model for the training set;

[0024] Figure 6 This is the displacement prediction result diagram of the deep learning model for the validation set;

[0025] Figure 7 This is the displacement prediction result diagram of the deep learning model for the test set;

[0026] Figure 8 Predict residual distribution plots for deep learning models on the training set;

[0027] Fig. 9 Predict residual distribution plots for deep learning models on a validation set;

[0028] Fig.10 Predict residual distribution plots for deep learning models on a test set;

[0029] Fig.11 Training curve diagrams for different time series prediction models;

[0030] Fig.12 Prediction of structural seismic response under main aftershock sequences. DETAILED DESCRIPTION

[0031] Specific implementation method 1: Combination Figure 1 This embodiment describes a method for predicting earthquake response of a main shock and aftershock sequence based on a knowledge-enhanced deep learning model, and the method specifically includes the following steps:

[0032] Step 1: Establish a finite element model of the target structure, select M main shock and aftershock ground motion records for the target structure, and adjust the ground motion intensity of the selected main shock and aftershock ground motion records to obtain M′ main shock and aftershock ground motion records after adjustment;

[0033] Step 2: Use the adjusted M′ mainshock and aftershock seismic records to perform nonlinear time history analysis on the target structure, and obtain the structural displacement response data and structural restoring force data of the target structure under each mainshock and aftershock seismic record, that is, obtain a training data set consisting of the mainshock and aftershock seismic sequence, structural displacement response data, and structural restoring force data;

[0034] Step 3: Build a knowledge-enhanced deep learning model including a consistent input encoding module, an encoder, a decoder, and a linear layer, and use the training data set to train the knowledge-enhanced deep learning model. During the training process, the main shock and aftershock seismic sequence is used as the input of the knowledge-enhanced deep learning model, and the structural displacement response data and structural resilience data are used as the output of the knowledge-enhanced deep learning model.

[0035] The loss function used in the training process is a knowledge enhancement loss function, and the training is stopped when the value of the knowledge enhancement loss function converges to obtain a trained knowledge enhancement deep learning model;

[0036] Step 4: Input the actual main shock and aftershock seismic motion sequence into the trained knowledge-enhanced deep learning model, and predict the structural displacement response and structural restoring force under the actual main shock and aftershock seismic motion sequence through the trained knowledge-enhanced deep learning model.

[0037] Specific implementation method 2: This implementation method is different from the specific implementation method 1 in that the main shock and aftershock earthquake records in step 1 are selected from the PEER database, and the selection rules are:

[0038] (1) The main shock ground motion and the corresponding aftershock ground motion are recorded by the same station, and the main shock and aftershock ground motion sequences are selected from free-field records;

[0039] (2) The PGA of the main shock and aftershocks are both greater than 0.1g;

[0040] (3) The shear wave velocity of the site recording the mainshock and aftershock seismic sequence is greater than 360 m / s.

[0041] The other steps and parameters are the same as those in the first embodiment.

[0042] Specific implementation method three: This implementation method is different from specific implementation method one or two in that the consistent input coding module includes a convolution layer, a position coding layer, and a first time coding layer to a Pth time coding layer; and the working process of the consistent input coding module is:

[0043] Take the tth main aftershock ground motion sequence as an example

[0044] Step 1: For the i-th element in the t-th mainshock / aftershock ground motion sequence The i-th element is transformed into The position pos is encoded as a position feature:

[0045]

[0046] Among them, L x is the length of the tth main aftershock ground motion sequence, PE (pos,i,2j) is the 2jth element in the position feature corresponding to the position pos of the mainshock and aftershock ground motion sequence, PE (pos,i,2j+1) is the 2j+1th element in the position feature corresponding to the position pos of the mainshock and aftershock earthquake motion sequence, where j is an integer and j=1,2,…,d model / 2,d model is the dimension of location characteristics;

[0047] Step 2: Use the first to Pth time coding layers to calculate the i-th element in the t-th main aftershock ground motion sequence The corresponding time is time-coded to obtain the coding result of the pth time coding layer [TE (pos,i) ] p , p = 1, 2, ..., P, and the encoding result [TE (pos,i) ] p The dimension is d model ;

[0048] Step 3: Substitute the i-th element in the t-th main aftershock ground motion sequence Through the convolution layer, the output of the convolution layer is of dimension d model Vector

[0049] Step 4: Combine the results of Step 1, Step 2, and Step 3 to obtain the i-th element in the t-th main aftershock ground motion sequence The corresponding splicing results

[0050] Step 5: Perform the process of steps 1 to 4 for each element in the tth main aftershock seismic sequence, and then connect the splicing results corresponding to each element in the tth main aftershock seismic sequence (that is, the splicing result corresponding to the first element in the tth main aftershock seismic sequence is used as the first to dth input vector of the encoder). model elements, and the concatenation result corresponding to the second element in the tth mainshock and aftershock ground motion sequence is used as the dth input vector of the encoder. model +1 to 2nd day model elements, and so on), and the connection result is used as the output of the tth main shock and aftershock earthquake motion sequence after the consistency input encoding module, that is, the connection result is used as the input vector of the encoder.

[0051] The other steps and parameters are the same as those in the first or second embodiment.

[0052] Specific implementation method 4: This implementation method is different from any one of the specific implementation methods 1 to 3 in that the specific process of step 4 is as follows:

[0053]

[0054] Among them, α is the weight parameter, PE (pos,i) is the i-th element The position pos corresponds to the position feature.

[0055] The other steps and parameters are the same as those in Specific Embodiments 1 to 3.

[0056] Specific implementation mode 5: This implementation mode is different from any one of specific implementation modes 1 to 4 in that the encoder includes a first encoding unit, a second encoding unit, a third encoding unit, a fourth encoding unit, a fifth encoding unit and a sixth encoding unit;

[0057] The first coding unit includes a probabilistic sparse self-attention mechanism layer and a distillation mechanism layer. In the first coding unit, the input of the first coding unit is first used as the input of the probabilistic sparse self-attention mechanism layer, and then the output of the probabilistic sparse self-attention mechanism layer is used as the input of the distillation mechanism layer, and the output of the distillation mechanism layer is used as the output of the first coding unit;

[0058] And the structures of the second coding unit, the third coding unit, the fourth coding unit, the fifth coding unit and the sixth coding unit are all the same as the structure of the first coding unit.

[0059] The other steps and parameters are the same as those in Specific Embodiments 1 to 4.

[0060] Specific implementation example 6: This implementation example is different from any one of specific implementation examples 1 to 5 in that the decoder includes a first decoding unit, a second decoding unit, a third decoding unit, a fourth decoding unit and a fully connected layer;

[0061] The first decoding unit includes a masked multi-head probabilistic sparse self-attention mechanism layer and a multi-head self-attention mechanism layer; and the structures of the second decoding layer, the third decoding layer, and the fourth decoding layer are the same as the structure of the first decoding layer.

[0062] The other steps and parameters are the same as those in Specific Implementation Methods 1 to 5.

[0063] Specific implementation method 7: This implementation method is different from any one of the specific implementation methods 1 to 6 in that the working process of the knowledge-enhanced deep learning model is:

[0064] Using the main shock and aftershock seismic motion sequence as the input of the consistency input encoding module, and then using the output of the consistency input encoding module as the input of the first encoding unit in the encoder;

[0065] Using the output of the first encoding unit as the input of the second encoding unit, and then using the output of the second encoding unit as the input of the third encoding unit;

[0066] Using the output of the third encoding unit as the input of the fourth encoding unit, and then using the output of the fourth encoding unit as the input of the fifth encoding unit;

[0067] The output of the fifth encoding unit is used as the input of the sixth encoding unit, and the key matrix K and the value matrix V in the feature map output by the sixth encoding unit are used as the input of the multi-head self-attention mechanism layer of each decoding unit in the decoder, that is, the input of the multi-head self-attention mechanism layer of each decoding unit includes the key matrix K and the value matrix V in the feature map output by the sixth encoding unit;

[0068] Mask the unpredicted positions of the structural data in the output vector of the consistency input encoding module, and use the masked vector as the input of the masked multi-head probabilistic sparse self-attention mechanism layer in the first decoding unit;

[0069] The query matrix output by the masked multi-head probabilistic sparse self-attention mechanism layer in the first decoding unit is used as the input of the multi-head self-attention mechanism layer in the first decoding unit; and the output of the multi-head self-attention mechanism layer in the first decoding unit is used as the input of the masked multi-head probabilistic sparse self-attention mechanism layer in the second decoding unit;

[0070] The query matrix output by the masked multi-head probabilistic sparse self-attention mechanism layer in the second decoding unit is used as the input of the multi-head self-attention mechanism layer in the second decoding unit; and the output of the multi-head self-attention mechanism layer in the second decoding unit is used as the input of the masked multi-head probabilistic sparse self-attention mechanism layer in the third decoding unit;

[0071] Using the query matrix output by the masked multi-head probabilistic sparse self-attention mechanism layer in the third decoding unit as the input of the multi-head self-attention mechanism layer in the third decoding unit; and using the output of the multi-head self-attention mechanism layer in the third decoding unit as the input of the masked multi-head probabilistic sparse self-attention mechanism layer in the fourth decoding unit;

[0072] Using the query matrix output by the masked multi-head probabilistic sparse self-attention mechanism layer in the fourth decoding unit as the input of the multi-head self-attention mechanism layer in the fourth decoding unit;

[0073] The output of the multi-head self-attention mechanism layer in the fourth decoding unit is used as the input of the fully connected layer, and the output of the fully connected layer is used as the output of the decoder;

[0074] The output of the decoder is used as the input of the linear layer, and the structural displacement response data and structural restoring force data are output through the linear layer.

[0075] The other steps and parameters are the same as those in Specific Embodiments 1 to 6.

[0076] Specific implementation eight: This implementation differs from any one of specific implementations one to seven in that the knowledge enhancement loss function is:

[0077] L=α D L D +α P L P

[0078] Among them, L is the knowledge enhancement loss function, L D is the data loss, L P is the physical loss, α D and α P are all weighting factors.

[0079] The other steps and parameters are the same as those in Specific Embodiments 1 to 7.

[0080] Specific implementation method 9: This implementation method is different from any one of the specific implementation methods 1 to 8 in that the data loss L D yes:

[0081]

[0082] Where N is the number of training samples, is the structural displacement response value of the nth training sample predicted by the model, is the true value of the structural displacement response of the nth training sample, is the structural resilience value of the nth training sample predicted by the model, is the true value of the structural resilience of the nth training sample, and ||·||2 is the 2-norm.

[0083] The other steps and parameters are the same as those in Specific Embodiments 1 to 8.

[0084] Specific implementation method 10: This implementation method is different from the specific implementation methods 1 to 9 in that the physical loss L P for:

[0085]

[0086] in, is the predicted structural response acceleration of the nth training sample, is the earthquake acceleration of the nth training sample.

[0087] The other steps and parameters are the same as those in Specific Embodiments 1 to 9.

[0088] Example

[0089] The method of this embodiment first constructs a model training set that can take into account the dynamic equation, and uses a consistent input encoding module to enhance the representation capability of time series data; through the decoder-encoder framework and the improved parallel computing capability of the attention mechanism, the model's ability to capture global correlations is significantly improved, thereby improving the ability to process complex long-sequence seismic responses; combined with the knowledge enhancement loss function, the model has superior predictive capabilities and training efficiency.

[0090] This embodiment takes the prediction of the seismic response of the containment structure under the action of the main shock and aftershock sequence as an example, and the prediction method specifically includes the following steps:

[0091] Step 1: Establish a finite element model of the target structure (i.e., containment structure) by using ABAQUS software, then select M main shock and aftershock ground motion records for the target structure, and adjust the ground motion intensity of the selected main shock and aftershock ground motion records to obtain M′ main shock and aftershock ground motion records after adjustment;

[0092] The containment structure consists of three main parts: the cylinder, the dome and the foundation. The height of the cylinder is 44m, and the height of the foundation is 6m. The thickness of the cylinder and the dome are 1.07m and 0.76m respectively. The compressive strength of the concrete is 40MPa, and the yield strength of the steel bar is 350MPa. In the finite element model of the containment structure, the concrete is simulated by the C3D8R solid unit, and the steel bar is simulated by the T3D2 truss unit. Rayleigh damping is used in the numerical simulation, and the damping ratio is set to 5%.

[0093] The main shock and aftershock earthquake records in step 1 are selected from the PEER database, and the selection rules are as follows:

[0094] (1) To avoid the influence of the spatial variability of ground motion, the main shock ground motion and the corresponding aftershock ground motion were recorded by the same station, and in order to ignore the soil-structure interaction, the main shock and aftershock ground motion sequences were selected from free-field records;

[0095] (2) The PGA (peak acceleration) of the main shock and aftershocks is greater than 0.1g;

[0096] (3) The shear wave velocity of the site recording the mainshock and aftershock seismic sequence is greater than 360 m / s.

[0097] Based on the above selection rules, this embodiment finally selects 100 main shock and aftershock records from the PEER database, and the distribution of their ground motion characteristics is as follows: Figure 2As shown. Due to the differences in seismic recording instruments, the sampling frequencies of the selected seismic motions include 50Hz, 100Hz, 200Hz and 250Hz. Due to the limited number of actual earthquake records, the present invention modulates the seismic motions of the selected main and aftershock records in the amplitude modulation range of 0.1g to 3.0g, and finally obtains a total of 1000 main and aftershock sequences. Ensure that the input data is evenly distributed at each earthquake intensity level, so as to achieve a balanced data distribution to meet the diversity requirements of the training set, provide a good data foundation for deep learning model training, and ensure the generalization ability of the model.

[0098] Step 2: Use the adjusted M′ mainshock and aftershock earthquake motion records to perform nonlinear time history analysis on the target structure (i.e., containment structure), and obtain the structural displacement response data and structural restoring force data of the target structure under each mainshock and aftershock earthquake motion record, that is, obtain a training data set consisting of the mainshock and aftershock earthquake motion sequence, structural displacement response data, and structural restoring force data;

[0099] Specifically, a unidirectional earthquake motion is applied to the containment structure, and only the structural seismic response in the load direction is considered. A 10-second time interval is set between the main shock and the aftershock, during which the acceleration is set to zero. The structural displacement response and corresponding restoring force data in the nonlinear time history analysis are recorded separately. The earthquake acceleration is set as the model input sequence, and the structural displacement and restoring force are set as the model output data to construct a model data set that can consider the structural dynamic equation. The data set is randomly divided into a training set, a validation set, and a test set in a ratio of 6:2:2.

[0100] Step 3: Figure 4 As shown in the figure, a knowledge-enhanced deep learning model (Phyformer model) including a consistent input encoding module, an encoder, a decoder and a linear layer is built, and the knowledge-enhanced deep learning model is trained using a training data set. During the training process, the main shock and aftershock seismic sequences are used as the input of the knowledge-enhanced deep learning model, and the structural displacement response data and structural resilience data are used as the output of the knowledge-enhanced deep learning model;

[0101] The details of the knowledge-enhanced deep learning model proposed in the present invention are described below:

[0102] like Figure 3 As shown, the consistent input coding module includes a convolutional layer, a position coding layer, and a first time coding layer to a Pth time coding layer; and the working process of the consistent input coding module is:

[0103] Take the tth main aftershock ground motion sequence as an example

[0104] Step 1: For the i-th element in the t-th mainshock / aftershock ground motion sequence The i-th element is transformed into The position pos is encoded as a position feature:

[0105]

[0106] Among them, L x is the length of the tth main aftershock ground motion sequence, PE (pos,i,2j) is the 2jth element in the position feature corresponding to the position pos of the mainshock and aftershock ground motion sequence, PE (pos,i,2j+1) is the 2j+1th element in the position feature corresponding to the position pos of the mainshock and aftershock earthquake motion sequence, where j is an integer and j=1,2,…,d model / 2,d model is the dimension of location characteristics;

[0107] Step 2: Use the first to Pth time coding layers to calculate the i-th element in the t-th main aftershock ground motion sequence The corresponding time is time-coded to obtain the coding result of the pth time coding layer [TE (pos,i) ] p , p = 1, 2, ..., P, and the encoding result [TE (pos,i) ] p The dimension is d model ;

[0108] Step 3: Substitute the i-th element in the t-th main aftershock ground motion sequence Through the convolution layer, the output of the convolution layer is of dimension d model Vector

[0109] Step 4: Combine the results of Step 1, Step 2, and Step 3 to obtain the i-th element in the t-th main aftershock ground motion sequence The corresponding splicing results

[0110]

[0111] Among them, α is a weight parameter used to adjust the contribution of the scalar mapping vector and the position / time encoding;

[0112] Step 5. Execute the process of steps 1 to 4 for each element in the t-th main aftershock seismic motion sequence, and then connect the splicing results corresponding to each element in the t-th main aftershock seismic motion sequence, and use the connection result as the output of the t-th main aftershock seismic motion sequence after the consistency input encoding module, that is, use the connection result as the input vector of the encoder.

[0113] The position coding layer of the present invention can capture the relative position information in the sequence, and the time coding layer can enhance the extraction of time information in the sequence data. It adopts p kinds of learnable time coding TE(pos) To obtain global time information, the seismic acceleration data is processed through the consistent input coding module, and the input vector is constructed through scalar mapping, position coding and time coding respectively, so as to optimize the representation of acceleration information. Moreover, the consistent input coding module in the present invention can process seismic data with arbitrary length and sampling frequency, which solves the problem that the seismic input must have the same sampling frequency and input length when training the traditional time series prediction model.

[0114] The encoder includes a first encoding unit, a second encoding unit, a third encoding unit, a fourth encoding unit, a fifth encoding unit and a sixth encoding unit;

[0115] The first coding unit includes a probabilistic sparse self-attention mechanism layer and a distillation mechanism layer. In the first coding unit, the input of the first coding unit is first used as the input of the probabilistic sparse self-attention mechanism layer, and then the output of the probabilistic sparse self-attention mechanism layer is used as the input of the distillation mechanism layer, and the output of the distillation mechanism layer is used as the output of the first coding unit;

[0116] And the structures of the second coding unit, the third coding unit, the fourth coding unit, the fifth coding unit and the sixth coding unit are all the same as the structure of the first coding unit.

[0117] The probabilistic sparse self-attention mechanism layer can focus on important features in the input sequence, significantly reducing the overall space complexity and enhancing the ability to process long input sequences.

[0118] In different attention mechanisms, the attention value can be calculated by the following formula:

[0119]

[0120] Among them, Q is the query matrix, K is the key matrix, and V is the value matrix. The three matrices are obtained by linearly changing the input data. k is the dimension of the key matrix; softmax(·) is the normalized exponential function.

[0121] The probabilistic sparse self-attention mechanism usually generates a large number of redundant value combinations in the feature map of the coding unit. In order to strengthen the main features, the method of the present invention uses a distillation mechanism layer to generate a more concentrated feature map and use it for the next coding unit. This operation can significantly reduce the complexity of the self-attention mechanism while retaining important information in the input data. The distillation process from layer j to layer j+1 is performed as follows:

[0122]

[0123] in,[·] ABrepresents an attention module with a multi-head self-attention mechanism; Conv1d(·) represents a one-dimensional convolution operation with a kernel width of 3. After the convolution operation, the data is activated by the ELU function. Finally, the maximum pooling technique is used for downsampling, and the input vector The size of is reduced by half, and the total memory consumption is reduced to O((2-ε)L x logL x ), where ε is a negligible constant.

[0124] The decoder includes a first decoding unit, a second decoding unit, a third decoding unit, a fourth decoding unit and a fully connected layer;

[0125] The first decoding unit includes a masked multi-head probabilistic sparse self-attention mechanism layer and a multi-head self-attention mechanism layer; and the structures of the second decoding layer, the third decoding layer, and the fourth decoding layer are the same as the structure of the first decoding layer.

[0126] Masking the multi-head probabilistic sparse self-attention mechanism layer can prevent the current position from referencing information about future positions when predicting the current position, ensuring that the current prediction only depends on historical information.

[0127] The knowledge-enhanced deep learning model of the present invention is built on an encoder-decoder architecture. The encoder module processes the input seismic sequence and converts it into a fixed-length vector, and the decoder module uses the vector to generate a structural response sequence. This architecture performs well in processing sequence-sequence problems. The working process of the overall knowledge-enhanced deep learning model is:

[0128] Using the main shock and aftershock seismic motion sequence as the input of the consistency input encoding module, and then using the output of the consistency input encoding module as the input of the first encoding unit in the encoder;

[0129] Using the output of the first encoding unit as the input of the second encoding unit, and then using the output of the second encoding unit as the input of the third encoding unit;

[0130] Using the output of the third encoding unit as the input of the fourth encoding unit, and then using the output of the fourth encoding unit as the input of the fifth encoding unit;

[0131] The output of the fifth encoding unit is used as the input of the sixth encoding unit, and the key matrix K and the value matrix V in the feature map output by the sixth encoding unit are used as the input of the multi-head self-attention mechanism layer of each decoding unit in the decoder;

[0132] Mask the unpredicted positions of the structural data (structural displacement response data and structural resilience data) in the output vector of the consistency input encoding module, and use the masked vector as the input of the masked multi-head probabilistic sparse self-attention mechanism layer in the first decoding unit;

[0133] The masking process of the vector is described in detail as follows: when the model needs to predict the structural displacement response data and structural restoring force data corresponding to the ath element in the seismic motion sequence, it is necessary to mask the part corresponding to the elements after the ath in the seismic motion sequence in the vector output by the consistency input encoding module.

[0134] The query matrix output by the masked multi-head probabilistic sparse self-attention mechanism layer in the first decoding unit is used as the input of the multi-head self-attention mechanism layer in the first decoding unit; and the output of the multi-head self-attention mechanism layer in the first decoding unit is used as the input of the masked multi-head probabilistic sparse self-attention mechanism layer in the second decoding unit;

[0135] The query matrix output by the masked multi-head probabilistic sparse self-attention mechanism layer in the second decoding unit is used as the input of the multi-head self-attention mechanism layer in the second decoding unit; and the output of the multi-head self-attention mechanism layer in the second decoding unit is used as the input of the masked multi-head probabilistic sparse self-attention mechanism layer in the third decoding unit;

[0136] Using the query matrix output by the masked multi-head probabilistic sparse self-attention mechanism layer in the third decoding unit as the input of the multi-head self-attention mechanism layer in the third decoding unit; and using the output of the multi-head self-attention mechanism layer in the third decoding unit as the input of the masked multi-head probabilistic sparse self-attention mechanism layer in the fourth decoding unit;

[0137] Using the query matrix output by the masked multi-head probabilistic sparse self-attention mechanism layer in the fourth decoding unit as the input of the multi-head self-attention mechanism layer in the fourth decoding unit;

[0138] The output of the multi-head self-attention mechanism layer in the fourth decoding unit is used as the input of the fully connected layer, and the output of the fully connected layer is used as the output of the decoder;

[0139] The output of the decoder is used as the input of the linear layer, and the structural displacement response data and structural restoring force data are output through the linear layer.

[0140] In the self-attention mechanism of the knowledge-enhanced deep learning model encoder of the present invention, matrices Q, K, and V are all obtained from the output of the previous encoding unit. In the masked self-attention mechanism layer of the decoder unit, by assigning the dot product of the unpredicted position to infinity, the masking of information that does not appear in the sequence is achieved, so that each position prediction can only focus on historical information. In the multi-head self-attention mechanism layer of the decoder unit, the matrix Q obtained from the previous decoding unit is combined with the matrices K and V output by the last encoding unit of the encoder, so that the decoder can use the complete input information when generating the output sequence.

[0141] The loss function used in the training process of the knowledge-enhanced deep learning model is:

[0142] L=α D L D +α P L P

[0143]

[0144] Among them, L is the loss function, L D is the data loss, L P is the physical loss, α D and α P are weighted factors (all assigned a value of 1), N is the number of training samples, is the structural displacement response value of the nth training sample predicted by the model, is the true value of the structural displacement response of the nth training sample, is the structural resilience value of the nth training sample predicted by the model, is the true value of the structural resilience of the nth training sample, ||·||2 is the 2-norm, is the predicted structural response acceleration of the nth training sample, is the earthquake acceleration of the nth training sample.

[0145] The training of traditional deep learning models relies entirely on identifying intrinsic features in the data, which will lead to extremely high demands on the amount of training data and produce predictions that do not conform to physical laws. Knowledge related to the prediction problem can be used as a regularization mechanism for deep learning models, effectively solving the overfitting problem, reducing dependence on large amounts of training data, and improving the robustness of the model. Therefore, the method of the present invention selects structural dynamics equations as physical constraints and uses knowledge-enhanced loss functions to train the Phyformer model. The dynamics equations can guide the learning process in model training and limit the model output to a feasible solution set, thereby improving the training efficiency and prediction accuracy of the model, and reducing the data dependence of the model and the physical inconsistency of the prediction results.

[0146] It should be noted that the knowledge enhancement loss function is only used in the training process of the deep learning model. The mean square error (MSE) is selected to measure the data loss and physical loss. The total loss is back-propagated from the decoder output to the entire model, and finally the weights and bias parameters of the entire network are optimized so that the Phyformer deep learning model can predict the structural displacement response that satisfies the dynamic equation. The optimization function uses the Adam algorithm, and the learning rate is set to 0.001. The prediction model is trained for 2000 iterations, and the batch size used for each iteration is 128. Figure 5 , Figure 6 and Figure 7 is the predicted displacement of the Phyformer model in the training set, validation set, and test set, and Figure 8 , Fig. 9 and Fig.10 This is the distribution of model prediction residuals for different data sets. It can be seen that the model has good prediction performance in different data sets.

[0147] Step 4: Input the actual main shock and aftershock seismic motion sequence into the trained knowledge-enhanced deep learning model. The trained knowledge-enhanced deep learning model is used to predict the structural displacement response and structural resilience under the actual main shock and aftershock seismic motion sequence, providing an effective numerical basis for structural damage assessment.

[0148] In order to verify the effectiveness of the method of the present invention in predicting earthquake response, the prediction results of the method of the present invention are compared with those of traditional LSTM and CNN models. Fig.11 The training curves of different models show that LSTM and CNN require 440 and 310 iterations of training respectively to converge, while the model built by the present invention only needs 120 trainings to converge, proving that the Phyformer model has higher training efficiency. A main aftershock record recorded in the Chi-Chi earthquake event was selected. The sampling frequency of the earthquake sequence is 200Hz, the duration of the earthquake sequence is 85 seconds, and it contains a total of 17001 data points. The main aftershock sequence record is input into the three trained models for structural seismic response prediction. The prediction results are shown in Figure 2. Fig.12 As shown. The displacement generated by LSTM has good consistency with the actual structural response in the first 20 seconds, but it cannot accurately predict the structural response under the subsequent earthquake. In contrast, CNN is better than LSTM in prediction accuracy, but still shows reduced accuracy in aftershock displacement prediction. The model in the method of the present invention provides the most accurate results in the prediction of structural response under the main shock and aftershocks. At the same time, it takes 8.6 hours for ABAQUS software to calculate the structural displacement response under the earthquake record, while the method of the present invention only takes 0.32 seconds to obtain the corresponding seismic response. It can be seen that the method of the present invention can efficiently, accurately and reliably predict the structural seismic response under the action of a long earthquake sequence.

[0149] The above calculation examples of the present invention are only used to explain the calculation model and calculation process of the present invention in detail, and are not intended to limit the implementation methods of the present invention. For ordinary technicians in the relevant field, other different forms of changes or modifications can be made based on the above description. It is impossible to list all the implementation methods here. All obvious changes or modifications derived from the technical solution of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for predicting earthquake response of mainshock and aftershock sequences based on knowledge-enhanced deep learning model, characterized in that: The method specifically comprises the following steps: Step 1: Establish a finite element model of the target structure, select M main shock and aftershock ground motion records for the target structure, and adjust the ground motion intensity of the selected main shock and aftershock ground motion records to obtain M′ main shock and aftershock ground motion records after adjustment; Step 2: Use the adjusted M′ mainshock and aftershock seismic records to perform nonlinear time history analysis on the target structure, and obtain the structural displacement response data and structural restoring force data of the target structure under each mainshock and aftershock seismic record, that is, obtain a training data set consisting of the mainshock and aftershock seismic sequence, structural displacement response data, and structural restoring force data; Step 3: Build a knowledge-enhanced deep learning model including a consistent input encoding module, an encoder, a decoder, and a linear layer, and use the training data set to train the knowledge-enhanced deep learning model. During the training process, the main shock and aftershock seismic sequence is used as the input of the knowledge-enhanced deep learning model, and the structural displacement response data and structural resilience data are used as the output of the knowledge-enhanced deep learning model. The loss function used in the training process is a knowledge enhancement loss function, and the training is stopped when the value of the knowledge enhancement loss function converges to obtain a trained knowledge enhancement deep learning model; Step 4: Input the actual main shock and aftershock seismic motion sequence into the trained knowledge-enhanced deep learning model, and predict the structural displacement response and structural restoring force under the actual main shock and aftershock seismic motion sequence through the trained knowledge-enhanced deep learning model.

2. According to claim 1, a method for predicting earthquake response of mainshock and aftershock sequences based on a knowledge-enhanced deep learning model is characterized in that: The main shock and aftershock earthquake records in step 1 are selected from the PEER database, and the selection rules are as follows: (1) The main shock ground motion and the corresponding aftershock ground motion are recorded by the same station, and the main shock and aftershock ground motion sequences are selected from free-field records; (2) The PGA of the main shock and aftershocks are both greater than 0.1g; (3) The shear wave velocity of the site recording the mainshock and aftershock seismic sequence is greater than 360 m / s.

3. The method for predicting earthquake response of mainshock and aftershock sequences based on knowledge-enhanced deep learning model according to claim 2 is characterized in that: The consistent input coding module includes a convolutional layer, a position coding layer, and a first time coding layer to a Pth time coding layer; and the working process of the consistent input coding module is: Take the tth main aftershock ground motion sequence as an example Step 1: For the i-th element in the t-th mainshock / aftershock ground motion sequence The i-th element is transformed into The position pos is encoded as a position feature: Among them, L x is the length of the tth main aftershock ground motion sequence, PE (pos,i,2j) is the 2jth element in the position feature corresponding to the position pos of the mainshock and aftershock ground motion sequence, PE (pos,i,2j+1) is the 2j+1th element in the position feature corresponding to the position pos of the mainshock and aftershock earthquake motion sequence, where j is an integer and j=1,2,…,d model 2, d model is the dimension of location characteristics; Step 2: Use the first to Pth time coding layers to calculate the i-th element in the t-th main aftershock ground motion sequence The corresponding time is time-coded to obtain the coding result of the pth time coding layer [TE (pos,i) ] p , p = 1, 2, ..., P, and the encoding result [TE (pos,i) ] p The dimension is d model ; Step 3: Substitute the ith element x in the tth main aftershock ground motion sequence i Through the convolution layer, the output of the convolution layer is of dimension d model Vector Step 4: Combine the results of Step 1, Step 2, and Step 3 to obtain the i-th element in the t-th main aftershock ground motion sequence The corresponding splicing results Step 5. Execute the process of steps 1 to 4 for each element in the t-th main aftershock seismic motion sequence, and then connect the splicing results corresponding to each element in the t-th main aftershock seismic motion sequence, and use the connection result as the output of the t-th main aftershock seismic motion sequence after the consistency input encoding module, that is, use the connection result as the input vector of the encoder.

4. The method for predicting earthquake response of mainshock and aftershock sequences based on knowledge-enhanced deep learning model according to claim 3 is characterized in that: The specific process of step 4 is as follows: Among them, α is the weight parameter, PE (pos,i) is the i-th element The position pos corresponds to the position feature.

5. The method for predicting earthquake response of mainshock and aftershock sequences based on knowledge-enhanced deep learning model according to claim 4 is characterized in that: The encoder includes a first encoding unit, a second encoding unit, a third encoding unit, a fourth encoding unit, a fifth encoding unit and a sixth encoding unit; The first coding unit includes a probabilistic sparse self-attention mechanism layer and a distillation mechanism layer. In the first coding unit, the input of the first coding unit is first used as the input of the probabilistic sparse self-attention mechanism layer, and then the output of the probabilistic sparse self-attention mechanism layer is used as the input of the distillation mechanism layer, and the output of the distillation mechanism layer is used as the output of the first coding unit; And the structures of the second coding unit, the third coding unit, the fourth coding unit, the fifth coding unit and the sixth coding unit are all the same as the structure of the first coding unit.

6. A method for predicting earthquake response of mainshock and aftershock sequences based on knowledge-enhanced deep learning model according to claim 5, characterized in that: The decoder includes a first decoding unit, a second decoding unit, a third decoding unit, a fourth decoding unit and a fully connected layer; The first decoding unit includes a masked multi-head probabilistic sparse self-attention mechanism layer and a multi-head self-attention mechanism layer; and the structures of the second decoding layer, the third decoding layer, and the fourth decoding layer are the same as the structure of the first decoding layer.

7. A method for predicting earthquake response of mainshock and aftershock sequences based on knowledge-enhanced deep learning model according to claim 6, characterized in that: The working process of the knowledge-enhanced deep learning model is: Using the main shock and aftershock seismic motion sequence as the input of the consistency input encoding module, and then using the output of the consistency input encoding module as the input of the first encoding unit in the encoder; Using the output of the first encoding unit as the input of the second encoding unit, and then using the output of the second encoding unit as the input of the third encoding unit; Using the output of the third encoding unit as the input of the fourth encoding unit, and then using the output of the fourth encoding unit as the input of the fifth encoding unit; The output of the fifth encoding unit is used as the input of the sixth encoding unit, and the key matrix K and the value matrix V in the feature map output by the sixth encoding unit are used as the input of the multi-head self-attention mechanism layer of each decoding unit in the decoder; Mask the unpredicted positions of the structural data in the output vector of the consistency input encoding module, and use the masked vector as the input of the masked multi-head probabilistic sparse self-attention mechanism layer in the first decoding unit; The query matrix output by the masked multi-head probabilistic sparse self-attention mechanism layer in the first decoding unit is used as the input of the multi-head self-attention mechanism layer in the first decoding unit; And the output of the multi-head self-attention mechanism layer in the first decoding unit is used as the input of the masked multi-head probabilistic sparse self-attention mechanism layer in the second decoding unit; The query matrix output by the masked multi-head probabilistic sparse self-attention mechanism layer in the second decoding unit is used as the input of the multi-head self-attention mechanism layer in the second decoding unit; And the output of the multi-head self-attention mechanism layer in the second decoding unit is used as the input of the masked multi-head probabilistic sparse self-attention mechanism layer in the third decoding unit; The query matrix output by the masked multi-head probabilistic sparse self-attention mechanism layer in the third decoding unit is used as the input of the multi-head self-attention mechanism layer in the third decoding unit; And the output of the multi-head self-attention mechanism layer in the third decoding unit is used as the input of the masked multi-head probabilistic sparse self-attention mechanism layer in the fourth decoding unit; Using the query matrix output by the masked multi-head probabilistic sparse self-attention mechanism layer in the fourth decoding unit as the input of the multi-head self-attention mechanism layer in the fourth decoding unit; The output of the multi-head self-attention mechanism layer in the fourth decoding unit is used as the input of the fully connected layer, and the output of the fully connected layer is used as the output of the decoder; The output of the decoder is used as the input of the linear layer, and the structural displacement response data and structural restoring force data are output through the linear layer.

8. The method for predicting earthquake response of mainshock and aftershock sequences based on knowledge-enhanced deep learning model according to claim 7, characterized in that: The knowledge enhancement loss function is: L=a D L D +a P L P Among them, L is the knowledge enhancement loss function, L D is the data loss, L P is the physical loss, α D and α P are all weighting factors.

9. The method for predicting earthquake response of mainshock and aftershock sequences based on knowledge-enhanced deep learning model according to claim 8, characterized in that: The data loss L D yes: Where N is the number of training samples, is the structural displacement response value of the nth training sample predicted by the model, is the true value of the structural displacement response of the nth training sample, is the structural resilience value of the nth training sample predicted by the model, is the true value of the structural resilience of the nth training sample, and ·2 is the 2-norm.

10. A method for predicting earthquake response of mainshock and aftershock sequences based on knowledge-enhanced deep learning model according to claim 9, characterized in that: The physical loss L P for: in, is the predicted structural response acceleration of the nth training sample, is the earthquake acceleration of the nth training sample.

Images