Deep learning-based structure anti-seismic performance evaluation system

Abstract

The invention relates to a deep learning-based structure anti-seismic performance evaluation system, which comprises a data acquisition module, a data preprocessing module, a feature extraction module, a numerical simulation module, a deep learning model training module, an evaluation and verification module, a visualization module and a model updating and maintenance module, the data preprocessing module is used for data preprocessing, the feature extraction module is used for data feature extraction, the numerical simulation module is used for simulating seismic response of a structure, the deep learning model training module is used for constructing a deep learning model, the evaluation and verification module is used for model evaluation and verification, and the visualization module is used for evaluation result display. And the model updating and maintaining module is used for updating and maintaining the model. According to the invention, an improved seismic data feature fusion extraction algorithm is proposed for feature extraction, an improved seismic perception depth state space network is proposed for model construction, and a better scheme is provided for a deep learning-based structure seismic performance evaluation system.

Description

Technical Field

[0001] This invention relates to the fields of feature extraction and deep learning, specifically to a deep learning-based structural seismic performance evaluation system. Background Technology

[0002] Feature extraction technology is the process of selecting and constructing a set of features that best represent the essential characteristics of raw data. Its purpose is to reduce the dimensionality of data, remove redundant information, and retain the most valuable information for the task, so as to facilitate subsequent data analysis and processing. Feature extraction technology includes statistical feature calculation, time-frequency domain analysis, feature selection algorithms, deep neural networks, and multivariate data fusion. The collaboration of these technologies has jointly constructed feature extraction technology for earthquake-related data, enabling more effective learning and expression of the complex and dynamic characteristics of earthquake-related data.

[0003] Deep learning technology automatically learns high-level features and abstract representations of data from a large amount of data by constructing neural network models with multiple hidden layers, thereby enabling the recognition and prediction of complex patterns. Deep learning technologies include Long Short-Term Memory (LSTM) networks, gated recurrent unit networks, multi-LSTM networks based on physical information, and physical-guided convolutional neural networks. The collaboration of these technologies jointly constructs the deep learning model of the structural seismic performance assessment system, making the structural seismic performance assessment system have a more stable deep learning training model. Summary of the Invention

[0004] To address the aforementioned problems, this invention aims to provide a deep learning-based structural seismic performance evaluation system.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a deep learning-based structural seismic performance evaluation system, comprising a data acquisition module, a data preprocessing module, a feature extraction module, a numerical simulation module, a deep learning model training module, an evaluation and verification module, a visualization module, and a model update and maintenance module. The data acquisition module is used to acquire earthquake-related data; the data preprocessing module is used to preprocess the acquired earthquake data; the feature extraction module proposes an improved earthquake data feature fusion extraction algorithm to extract structural dynamic features from the earthquake data; the numerical simulation module is used to numerically simulate the structural response under seismic loading; the deep learning model training module proposes an improved earthquake-sensing deep state space network to construct a deep learning model to learn the mapping relationship between structural response data and earthquake input data; the evaluation and verification module includes a model evaluation and verification unit and a structural performance evaluation unit; the model evaluation and verification unit is used to evaluate and verify the trained deep learning model and its prediction results; the structural performance evaluation unit is used to evaluate the seismic performance of the structure based on the structural response data predicted by the deep learning model; the visualization module is used to display the structural seismic performance evaluation results in an intuitive form; and the model update and maintenance module is used to update and maintain the deep learning model.

[0006] Furthermore, the data acquisition module collects various types of information about the structure and earthquake-related data through seismic network detection, satellite remote sensing technology, on-site survey sampling, UAV surveys, and historical data collection. This includes basic structural information, structural response data during historical earthquakes, and seismic wave data.

[0007] Furthermore, the data preprocessing module preprocesses the acquired earthquake-related data through baseline correction, time normalization, data interpolation and segmentation, and data standardization and normalization, so that the deep learning model can better process the earthquake data.

[0008] Furthermore, the feature extraction module proposes an improved seismic data feature fusion extraction algorithm to dynamically extract features from structural response data and seismic wave data, thereby simplifying model input and improving model training efficiency and prediction accuracy.

[0009] Furthermore, the improved seismic data feature fusion and extraction algorithm is as follows: Assume the original seismic data is... ,in, Raw data related to earthquakes. for The first piece of raw data, for The second piece of raw data, for The Middle First, the raw data, then the earthquake-related raw data. After sampling, the sampled earthquake-related data are: ,in, For the sampled earthquake-related data, The sampling period is For at any time Raw earthquake-related data at the time, For at any time Raw earthquake-related data at the time, For at any time The original earthquake-related data at that time was differentiated from the sampled earthquake-related data, i.e. ,in, For a moment The value at time is the differential value. The standardized coefficient is... It is a second-order difference operation, and ,in, For a moment The former Raw earthquake-related data at any given time. For a moment After Raw earthquake-related data at any given time. For a moment The former Raw earthquake-related data at any given time. For a moment After The raw earthquake data at time t, with a weighted average frequency response set to... ,in, For a moment The weighted average frequency response, For a moment Weighting coefficients for time data For a moment Weighting coefficients for time data For at any time Raw earthquake-related data at the time, For a moment The former The raw earthquake-related data at that time, and the characteristic information of the earthquake signal waveform are as follows: ,in, For a moment eigenvalues ​​at time, The number of signal values. For the first Weighting coefficients at time points, For earthquake signals at time forward The raw data at time t, with the average relative power set to change the signal spectral bandwidth, the average relative power is . ,in, The average relative power of the signal. The frequency of the signal is [value], and the frequency range is [value]. arrive , For signal Frequency domain information representation after time-frequency transformation For frequency domain information In frequency The square of the complex amplitude, For minute frequency changes;

[0010] To capture and describe the more complex and dynamic behavior in earthquake-related data, the interaction between multiple different seismic signals is further considered to optimize algorithm performance. It is assumed that the frequency domain information of the first seismic signal is... ,in, The first seismic signal at frequency The frequency domain components, The first seismic signal at frequency The frequency domain components, The first seismic signal at frequency The frequency domain components of the second seismic signal are: , The second seismic signal at frequency The frequency domain components, The second seismic signal at frequency The frequency domain components, The second seismic signal at frequency The frequency domain components, the first The frequency domain information of each seismic signal is , For the first A seismic signal at frequency The frequency domain components, For the first A seismic signal at frequency The frequency domain components, For the first A seismic signal at frequency The frequency domain components, therefore the average relative power Improved to ,Right now ,in, For the improved average relative power, for The frequency domain information of a signal in frequency The square of the complex amplitude of the interaction is then used to propose a method for evaluating frequency domain features using spectral width, i.e. ,in, For the spectral width, For frequency index, The number of signal values. The center frequency of the spectrum. For frequency domain information In frequency To further improve feature discrimination, a nonlinear attenuation factor based on a Gaussian function is proposed to improve the bandwidth, i.e., the square of the complex amplitude. ,in, For bandwidth, The center frequency of the spectrum. It is a non-linear decay factor. Since is the width parameter of the Gaussian function, the signal power spectrum varies with the length of the bandwidth, i.e. ,in, The power spectrum of the signal. For the expected operation, To improve the feature extraction algorithm for seismic data fusion, a nonlinear approach is proposed to extract features from seismic-related data. First, an average weighted frequency response is proposed to adjust the importance of different frequency components relative to the overall frequency response. Then, the interaction between multiple seismic signals is considered by improving the average relative power. At the same time, a nonlinear attenuation factor based on a Gaussian function is proposed to improve the bandwidth, so as to better capture and describe the more complex and dynamic behavior in seismic-related data, and realize the dynamic feature extraction of structural response data and seismic wave data.

[0011] Furthermore, the numerical simulation module uses finite element analysis software and boundary element method to simulate the structural response under seismic loading, generating a large amount of simulated structural response data to help deep learning models learn the behavioral characteristics of structures under different seismic conditions.

[0012] Furthermore, the deep learning model training module proposes an improved earthquake sensing deep state space network to construct the deep learning model in order to learn the mapping relationship between structural response data and earthquake input data.

[0013] Furthermore, the state-space network for improving earthquake sensing depth is specifically as follows: Let the state-space model of earthquake-related data be... , ,in, For time step, In time step The system state at that time, In time step The system state at that time, For time steps Input time, For the parameter set of the model, For time steps Time by parameters Parameterized state update function For time steps The system output at that time For time steps Time by parameters The parameterized output function is proposed because seismic data exhibits significant nonlinear characteristics and complex dynamic changes; therefore, a skip connection is used. Layered neural networks improve state-space models to enhance the flexibility of the model in learning the relationship between input data and output. Deep learning models are... , , ,in, Let the initial state vector be... For the initial state, The number of layers in the neural network. For the first in the network The output of the layer, For the first in the network The output of the layer, For activation function, For the first The weight matrix used by the layer for state transitions. For the first The layer is a weight matrix used for the transition between inputs and states. For the input data of this network layer, For the first Layer bias, For parameters A parameterized state update function is introduced to help the model focus on key features of the input earthquake-related data. An attention mechanism is also introduced to improve the model's ability to capture complex dynamic changes in the data. This improvement in the deep learning model is further refined to... , , ,in, For the improved network output, For element-wise multiplication, This is the attention weight vector;

[0014] Meanwhile, to improve the model's stability, the deep learning model was first modified into an implicit model to better adapt to and process complex seismic data. , , ,in, Let be the transformation matrix of the initial state. Let the initial state vector be... For the process The transformed first The output of the layer, To make the first The hidden state of layer 1 is transitioned to the 1st layer. The weight matrix of the layer, For the first The hidden state of the layer This represents the total number of layers in the neural network. This is the transformation matrix of the last layer of the neural network. For the first To improve the convergence speed of the model and simplify its training and optimization, the output of each layer of the implicit model is proposed. Steps are divided into Sub-steps are used to achieve finer control over the model's behavior, i.e. ,in, In time step and the The output of the layer, In time step and the The output of the layer, For the first The activation function of the layer is determined, and then constraints are added to ensure the stability of the model. The constraints are as follows: ,in, For the first The transformation matrix of the layer, for The transpose of the matrix, For the first The diagonal positive definite matrix of the layer, for The diagonal positive definite matrix of the layer, for The transpose of the matrix is ​​used to define the model's loss function as minimizing the simulation error function, i.e. , ,in, The set of all possible parameters. To measure the simulation error, The total number of time steps. In time step The actual output at that time For time steps The model predicts the output at that time. For constraints, including and ,in, In time step Hidden layer output, In time step Hidden layer output, For time steps To prevent overfitting, an L2 regularization term is proposed to be added to the loss function for the measured input data. , ,in, For L2 regularization terms, The regularization coefficient is . Using the Frobenius norm, this method realizes the mapping relationship between learning structural response data and seismic input data, improving the deep state-space network for seismic sensing. It first proposes using a method with skip connections. Layered neural networks are used to improve the state-space model to enhance its flexibility in learning the relationship between seismic input data and output. Next, an attention mechanism is introduced to improve the model's ability to capture complex dynamic changes in the data. Then, it is proposed to divide each step of the implicit model into multiple sub-steps to improve the model's convergence speed and simplify model training and optimization. Finally, it is proposed to add an L2 regularization term to the loss function to improve the model's generalization ability, thereby constructing a stable deep learning model.

[0015] Furthermore, the model evaluation and validation unit compares the model's prediction accuracy and generalization ability with test data from the actual structure, thereby evaluating and validating the trained deep learning model and its prediction results.

[0016] Furthermore, the structural performance assessment unit quantitatively evaluates the seismic performance of the structure by combining the structure's performance indicators and seismic design codes. It assesses the seismic performance of the structure based on the structural response data predicted by the deep learning model, providing a basis for the seismic strengthening and repair of the structure.

[0017] Furthermore, the visualization module presents the structural seismic performance assessment results to users in an intuitive chart format, enabling users to more clearly understand the seismic performance of the structure and any existing problems.

[0018] Furthermore, the model update and maintenance module is used to update and maintain the deep learning model. Based on new earthquake-related data and evaluation results, the deep learning model is optimized and updated to improve the model's prediction accuracy and adaptability, enabling the model to continuously learn new knowledge and maintain good performance.

[0019] Compared with the prior art, the beneficial effects of the present invention are:

[0020] 1. An improved seismic data feature fusion extraction algorithm is proposed to dynamically extract features from structural response data and seismic wave data. The innovation of this invention lies in the fact that the improved seismic data feature fusion extraction algorithm first proposes an average weighted frequency response to adjust the importance of different frequency components relative to the overall frequency response, then considers the interaction between multiple seismic signals by improving the average relative power, and proposes a nonlinear attenuation factor based on a Gaussian function to improve the bandwidth, so as to better capture and describe the more complex and dynamic behavior in seismic related data, so as to realize the dynamic feature extraction of structural response data and seismic wave data.

[0021] 2. This invention proposes an improved deep state-space network for earthquake sensing to construct deep learning models. The innovation lies in the fact that the improved deep state-space network for earthquake sensing first proposes using a skip connection. Layered neural networks are used to improve the state-space model to enhance its flexibility in learning the relationship between seismic input data and output. Next, an attention mechanism is introduced to improve the model's ability to capture complex dynamic changes in the data. Then, it is proposed to divide each step of the implicit model into multiple sub-steps to improve the model's convergence speed and simplify model training and optimization. Finally, it is proposed to add an L2 regularization term to the loss function to improve the model's generalization ability, thereby constructing a stable deep learning model. Attached Figure Description

[0022] The invention will be further illustrated with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the invention. For those skilled in the art, other drawings can be obtained based on the following drawings without any creative effort.

[0023] Figure 1 This is a schematic diagram of the structure of the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] A deep learning-based structural seismic performance assessment system includes a data acquisition module, a data preprocessing module, a feature extraction module, a numerical simulation module, a deep learning model training module, an evaluation and verification module, a visualization module, and a model update and maintenance module. The data acquisition module collects earthquake-related data; the data preprocessing module preprocesses the collected earthquake data; the feature extraction module proposes an improved earthquake data feature fusion extraction algorithm to extract structural dynamic features from the earthquake data; the numerical simulation module performs numerical simulations of the structural response under seismic loading; the deep learning model training module proposes an improved earthquake-sensing deep state space network to construct a deep learning model to learn the mapping relationship between structural response data and earthquake input data; the evaluation and verification module includes a model evaluation and verification unit and a structural performance evaluation unit; the model evaluation and verification unit evaluates and verifies the trained deep learning model and its prediction results; the structural performance evaluation unit evaluates the seismic performance of the structure based on the structural response data predicted by the deep learning model; the visualization module displays the structural seismic performance assessment results in an intuitive manner; and the model update and maintenance module updates and maintains the deep learning model.

[0026] Preferably, the data acquisition module collects various types of information about the structure and earthquake-related data through seismic network detection, satellite remote sensing technology, on-site survey sampling, UAV surveys, and historical data collection. This includes basic information about the structure (such as the structure's geometry, dimensions, and material parameters), response data of the structure in historical earthquakes (such as acceleration, displacement, and strain data), and seismic wave data (such as records of seismic waves of different sites and intensities).

[0027] Preferably, the data preprocessing module preprocesses the acquired earthquake-related data through baseline correction, time normalization, data interpolation and segmentation, and data standardization and normalization to remove noise and outliers from the acquired earthquake data, fill in missing values, and convert data from different sources and with different dimensions into a unified format and dimension so that the deep learning model can better process the earthquake data.

[0028] Preferably, the feature extraction module proposes an improved seismic data feature fusion extraction algorithm to perform dynamic feature extraction on structural response data (such as peak acceleration, peak displacement, and energy dissipation characteristics) and seismic wave data (such as the spectral characteristics and duration characteristics of seismic waves). These features can concisely describe the dynamic characteristics of the structure, thereby simplifying the model input and improving the model training efficiency and prediction accuracy.

[0029] Specifically, the improved seismic data feature fusion and extraction algorithm is as follows: Assuming the original seismic data (basic structural information, structural response data, and seismic wave data) are... ,in, Raw data related to earthquakes. for The first piece of raw data, for The second piece of raw data, for The Middle First, the raw data, then the earthquake-related raw data. After sampling, the sampled earthquake-related data are: ,in, For the sampled earthquake-related data, The sampling period is For at any time Raw earthquake-related data at the time, For at any time Raw earthquake-related data at the time, For at any time The original earthquake-related data at that time was differentiated from the sampled earthquake-related data, i.e. ,in, For a moment The value at time is the differential value. The standardized coefficient is... It is a second-order difference operation, and ,in, For a moment The former Raw earthquake-related data at any given time. For a moment After Raw earthquake-related data at any given time. For a moment The former Raw earthquake-related data at any given time. For a moment After The raw earthquake data at time t, with a weighted average frequency response set to... ,in, For a moment The weighted average frequency response, For a moment Weighting coefficients for time data For a moment Weighting coefficients for time data For at any time Raw earthquake-related data at the time, For a moment The former The raw earthquake-related data at that time, and the characteristic information of the earthquake signal waveform are as follows: ,in, For a moment eigenvalues ​​at time, The number of signal values. For the first Weighting coefficients at time points, For earthquake signals at time forward The raw data at time t, with the average relative power set to change the signal spectral bandwidth, the average relative power is . ,in, The average relative power of the signal. The frequency of the signal is [value], and the frequency range is [value]. arrive , For signal Frequency domain information representation after time-frequency transformation For frequency domain information In frequency The square of the complex amplitude, For minute frequency changes;

[0030] To capture and describe the more complex and dynamic behavior in earthquake-related data, the interaction between multiple different seismic signals is further considered to optimize algorithm performance. It is assumed that the frequency domain information of the first seismic signal is... ,in, The first seismic signal at frequency The frequency domain components, The first seismic signal at frequency The frequency domain components, The first seismic signal at frequency The frequency domain components of the second seismic signal are: , The second seismic signal at frequency The frequency domain components, The second seismic signal at frequency The frequency domain components, The second seismic signal at frequency The frequency domain components, the first The frequency domain information of each seismic signal is , For the first A seismic signal at frequency The frequency domain components, For the first A seismic signal at frequency The frequency domain components, For the first A seismic signal at frequency The frequency domain components, therefore the average relative power Improved to ,Right now ,in, For the improved average relative power, for The frequency domain information of a signal in frequency The square of the complex amplitude of the interaction is then used to propose a method for evaluating frequency domain features using spectral width, i.e. ,in, For the spectral width, For frequency index, The number of signal values. The center frequency of the spectrum. For frequency domain information In frequency To further improve feature discrimination, a nonlinear attenuation factor based on a Gaussian function is proposed to improve the bandwidth, i.e., the square of the complex amplitude. ,in, For bandwidth, The center frequency of the spectrum. It is a non-linear decay factor. Since is the width parameter of the Gaussian function, the signal power spectrum varies with the length of the bandwidth, i.e. ,in, The power spectrum of the signal. For the expected operation, To improve the feature extraction algorithm for seismic data fusion, a nonlinear approach is proposed to extract features from seismic-related data. First, an average weighted frequency response is proposed to adjust the importance of different frequency components relative to the overall frequency response. Then, the interaction between multiple seismic signals is considered by improving the average relative power. At the same time, a nonlinear attenuation factor based on a Gaussian function is proposed to improve the bandwidth, so as to better capture and describe the more complex and dynamic behavior in seismic-related data, and realize the dynamic feature extraction of structural response data and seismic wave data.

[0031] Preferably, the numerical simulation module uses finite element analysis software and boundary element method to simulate the response of the structure under seismic loading, generating a large amount of simulated structural response data to help the deep learning model learn the behavioral characteristics of the structure under different seismic conditions.

[0032] Preferably, the deep learning model training module proposes an improved earthquake sensing deep state space network to construct a deep learning model to learn the mapping relationship between structural response data and earthquake input data.

[0033] Specifically, the improvement of the earthquake sensing depth state space network is as follows: Let the state space model of earthquake-related data be... , ,in, For time step, In time step The system state at that time, In time step The system state at that time, For time steps Input time, For the parameter set of the model, For time steps Time by parameters A parameterized state update function, which describes how to update the state based on the current time step. Input and the previous time step status To calculate the current time step new state It is used to define how the state in a recurrent neural network evolves over time. For time steps The system output at that time For time steps Time by parameters A parameterized output function is proposed to map the state and input to the output. Since seismic data exhibits significant nonlinear characteristics and complex dynamic changes, a method with skip connections is suggested. Layered neural networks improve state-space models to enhance the flexibility of the model in learning the relationship between input data and output. Deep learning models are... , , ,in, Let the initial state vector be... For the initial state, The number of layers in the neural network. For the first in the network The output of the layer, For the first in the network The output of the layer, For activation function, For the first The weight matrix used by the layer for state transitions. For the first The layer is a weight matrix used for the transition between inputs and states. For the input data of this network layer, For the first Layer bias, For parameters A parameterized state update function is introduced to help the model focus on key features of the input earthquake-related data. An attention mechanism is also introduced to improve the model's ability to capture complex dynamic changes in the data. This improvement in the deep learning model is further refined to... , , ,in, For the improved network output, For element-wise multiplication, This is the attention weight vector;

[0034] Meanwhile, to improve the model's stability, the deep learning model was first modified into an implicit model to better adapt to and process complex seismic data. , , ,in, Let be the transformation matrix of the initial state. Let the initial state vector be... For the process The transformed first The output of the layer, To make the first The hidden state of layer 1 is transitioned to the 1st layer. The weight matrix of the layer, For the first The hidden state of the layer This represents the total number of layers in the neural network. This is the transformation matrix of the last layer of the neural network. For the first To improve the convergence speed of the model and simplify its training and optimization, the output of each layer of the implicit model is proposed. Steps are divided into Sub-steps are used to achieve finer control over the model's behavior, i.e. ,in, In time step and the The output of the layer, In time step and the The output of the layer, For the first The activation function of the layer is determined, and then constraints are added to ensure the stability of the model. The constraints are as follows: ,in, For the first The transformation matrix of the layer, for The transpose of the matrix, For the first The diagonal positive definite matrix of the layer, for The diagonal positive definite matrix of the layer, for The transpose of the matrix is ​​used to define the model's loss function as minimizing the simulation error function, i.e. , ,in, The set of all possible parameters. To measure the simulation error, The total number of time steps. In time step The actual output at that time For time steps The model predicts the output at that time. For constraints, including and ,in, In time step Hidden layer output, In time step Hidden layer output, For time steps To prevent overfitting, an L2 regularization term is proposed to be added to the loss function for the measured input data. , ,in, For L2 regularization terms, The regularization coefficient is . Using the Frobenius norm, this method realizes the mapping relationship between learning structural response data and seismic input data, improving the deep state-space network for seismic sensing. It first proposes using a method with skip connections. Layered neural networks are used to improve the state-space model to enhance its flexibility in learning the relationship between seismic input data and output. Next, an attention mechanism is introduced to improve the model's ability to capture complex dynamic changes in the data. Then, it is proposed to divide each step of the implicit model into multiple sub-steps to improve the model's convergence speed and simplify model training and optimization. Finally, it is proposed to add an L2 regularization term to the loss function to improve the model's generalization ability, thereby constructing a stable deep learning model.

[0035] Preferably, the model evaluation and validation unit compares the model's prediction accuracy and generalization ability with test data from the actual structure, thereby evaluating and validating the trained deep learning model and its prediction results.

[0036] Preferably, the structural performance evaluation unit quantitatively evaluates the seismic performance of the structure by combining the structure's performance indicators and seismic design codes, and evaluates the seismic performance of the structure based on the structural response data predicted by the deep learning model, thus providing a basis for the seismic strengthening and repair of the structure.

[0037] Preferably, the visualization module presents the structural seismic performance assessment results to users in an intuitive chart format, so that users can more clearly understand the seismic performance of the structure and the existing problems.

[0038] Preferably, the model update and maintenance module is used to update and maintain the deep learning model. Based on new earthquake-related data and evaluation results, the deep learning model is optimized and updated to improve the model's prediction accuracy and adaptability. Through incremental learning and transfer learning techniques, the model can continuously learn new knowledge and maintain good performance.

[0039] This paper proposes a deep learning-based structural seismic performance evaluation system for assessing structural seismic performance. By integrating data acquisition, data preprocessing, feature extraction, numerical simulation, deep learning model training, evaluation and verification, visualization, and model updating and maintenance modules, a deep learning-based structural seismic performance evaluation system is provided. An improved seismic data feature fusion and extraction algorithm is proposed for dynamic feature extraction from structural response data and seismic wave data. The innovation of this invention lies in the following: the improved seismic data feature fusion and extraction algorithm first proposes an average weighted frequency response to adjust the importance of different frequency components relative to the overall frequency response; then, it considers the interaction between multiple seismic signals by improving the average relative power; and simultaneously, it proposes a Gaussian-based nonlinear attenuation factor to improve bandwidth, thereby better capturing and describing the more complex and dynamic behavior in seismic-related data, achieving dynamic feature extraction from structural response data and seismic wave data. An improved seismic sensing deep state space network is proposed for constructing the deep learning model. The innovation of this invention lies in the following: the improved seismic sensing deep state space network first proposes the use of skip connections. This invention improves the state-space model using layered neural networks to enhance its flexibility in learning the relationship between seismic input and output data. It then introduces an attention mechanism to improve the model's ability to capture complex dynamic changes in the data. Next, it proposes dividing each step of the implicit model into multiple sub-steps to improve convergence speed and simplify training and optimization. Finally, it proposes adding an L2 regularization term to the loss function to improve generalization ability, thereby constructing a stable deep learning model. This effectively improves the performance of a deep learning-based structural seismic performance assessment system, providing more comprehensive and accurate technical support for structural seismic performance assessment systems and better decision support for scientific and efficient structural seismic performance assessment. Furthermore, this invention relates to feature extraction and deep learning technologies, providing an efficient and stable structural seismic performance assessment system and contributing significant application value to structural seismic performance assessment.

[0040] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A deep learning-based structural seismic performance evaluation system, characterized in that, The system comprises a data acquisition module, a data preprocessing module, a feature extraction module, a numerical simulation module, a deep learning model training module, an evaluation and validation module, a visualization module, and a model update and maintenance module. The data acquisition module collects earthquake-related data; the data preprocessing module preprocesses the acquired earthquake data; the feature extraction module proposes an improved earthquake data feature fusion extraction algorithm to extract structural dynamic features from earthquake data; the numerical simulation module performs numerical simulations of the structural response under earthquake loading; the deep learning model training module proposes an improved earthquake-sensing deep state-space network to construct a deep learning model to learn the mapping relationship between structural response data and earthquake input data; the evaluation and validation module includes a model evaluation and validation unit and a structural performance evaluation unit; the model evaluation and validation unit evaluates and validates the trained deep learning model and its prediction results; the structural performance evaluation unit evaluates the seismic performance of the structure based on the structural response data predicted by the deep learning model; the visualization module displays the structural seismic performance evaluation results in an intuitive manner; and the model update and maintenance module updates and maintains the deep learning model.

2. The structural seismic performance evaluation system based on deep learning according to claim 1, characterized in that, The data acquisition module collects various types of information about the structure and earthquake-related data through seismic network detection, satellite remote sensing technology, on-site survey sampling, UAV surveys, and historical data collection. This includes basic information about the structure, response data of the structure in historical earthquakes, and seismic wave data.

3. The structural seismic performance evaluation system based on deep learning according to claim 1, characterized in that, The data preprocessing module preprocesses the acquired earthquake-related data through baseline correction, time normalization, data interpolation and segmentation, and data standardization and normalization, so that the deep learning model can better process the earthquake data.

4. The structural seismic performance evaluation system based on deep learning according to claim 1, characterized in that, The feature extraction module proposes an improved feature fusion extraction algorithm for seismic data to dynamically extract features from structural response data and seismic wave data, thereby simplifying model input and improving model training efficiency and prediction accuracy.

5. The structural seismic performance evaluation system based on deep learning according to claim 4, characterized in that, The improved seismic data feature fusion and extraction algorithm is as follows: Assume the original seismic data is... ,in, Raw data related to earthquakes. for The first piece of raw data, for The second piece of raw data, for The Middle First, the raw data, then the earthquake-related raw data. After sampling, the sampled earthquake-related data are: ,in, For the sampled earthquake-related data, The sampling period is For at any time Raw earthquake-related data at the time, For at any time Raw earthquake-related data at the time, For at any time The original earthquake-related data at that time was differentiated from the sampled earthquake-related data, i.e. ,in, For a moment The value at time is the differential value. The standardized coefficient is... It is a second-order difference operation, and ,in, For a moment The former Raw earthquake-related data at any given time. For a moment After Raw earthquake-related data at any given time. For a moment The former Raw earthquake-related data at any given time. For a moment After The raw earthquake data at time t, with a weighted average frequency response set to... ,in, For a moment The weighted average frequency response, For a moment Weighting coefficients for time data For a moment Weighting coefficients for time data For at any time Raw earthquake-related data at the time, For a moment The former The raw earthquake-related data at that time, and the characteristic information of the earthquake signal waveform are as follows: ,in, For a moment eigenvalues ​​at time, The number of signal values. For the first Weighting coefficients at time points, For earthquake signals at time forward The raw data at time t, with the average relative power set to change the signal spectral bandwidth, the average relative power is . ,in, The average relative power of the signal. The frequency of the signal is [value], and the frequency range is [value]. arrive , For signal Frequency domain information representation after time-frequency transformation For frequency domain information In frequency The square of the complex amplitude, For minute frequency changes; To capture and describe the more complex and dynamic behavior in earthquake-related data, the interaction between multiple different seismic signals is further considered to optimize algorithm performance. It is assumed that the frequency domain information of the first seismic signal is... ,in, The first seismic signal at frequency The frequency domain components, The first seismic signal at frequency The frequency domain components, The first seismic signal at frequency The frequency domain components of the second seismic signal are: , The second seismic signal at frequency The frequency domain components, The second seismic signal at frequency The frequency domain components, The second seismic signal at frequency The frequency domain components, the first The frequency domain information of each seismic signal is , For the first A seismic signal at frequency The frequency domain components, For the first A seismic signal at frequency The frequency domain components, For the first A seismic signal at frequency The frequency domain components, therefore the average relative power Improved to ,Right now ,in, For the improved average relative power, for The frequency domain information of a signal in frequency The square of the complex amplitude of the interaction is then used to propose a method for evaluating frequency domain features using spectral width, i.e. ,in, For the spectral width, For frequency index, The number of signal values. The center frequency of the spectrum. For frequency domain information In frequency To further improve feature discrimination, a nonlinear attenuation factor based on a Gaussian function is proposed to improve the bandwidth, i.e., the square of the complex amplitude. ,in, For bandwidth, The center frequency of the spectrum. It is a non-linear decay factor. Since is the width parameter of the Gaussian function, the signal power spectrum varies with the length of the bandwidth, i.e. ,in, The power spectrum of the signal. For the expected operation, To improve the feature extraction algorithm for seismic data fusion, a nonlinear approach is proposed to extract features from seismic-related data. First, an average weighted frequency response is proposed to adjust the importance of different frequency components relative to the overall frequency response. Then, the interaction between multiple seismic signals is considered by improving the average relative power. At the same time, a nonlinear attenuation factor based on a Gaussian function is proposed to improve the bandwidth, so as to better capture and describe the more complex and dynamic behavior in seismic-related data, and realize the dynamic feature extraction of structural response data and seismic wave data.

6. The structural seismic performance evaluation system based on deep learning according to claim 1, characterized in that, The numerical simulation module uses finite element analysis software and boundary element method to simulate the response of structures under seismic loading, generating a large amount of simulated structural response data to help deep learning models learn the behavioral characteristics of structures under different seismic conditions.

7. The structural seismic performance evaluation system based on deep learning according to claim 1, characterized in that, The deep learning model training module proposes an improved earthquake sensing deep state space network to construct a deep learning model to learn the mapping relationship between structural response data and earthquake input data.

8. The structural seismic performance evaluation system based on deep learning according to claim 7, characterized in that, The improvement of the earthquake sensing depth state space network is as follows: Let the state space model of earthquake-related data be... , ,in, For time step, In time step The system state at that time, In time step The system state at that time, For time step Input time, For the parameter set of the model, For time step Time by parameters Parameterized state update function For time steps The system output at that time For time step Time by parameters The parameterized output function is proposed because seismic data exhibits significant nonlinear characteristics and complex dynamic changes; therefore, a skip connection is used. Layered neural networks improve state-space models to enhance the flexibility of the model in learning the relationship between input data and output. Deep learning models are... , , ,in, Let be the initial state vector. For the initial state, This represents the number of layers in the neural network. For the first in the network The output of the layer, For the first in the network The output of the layer, For activation function, For the first The weight matrix used by the layer for state transitions. For the first The layer is a weight matrix used for the transition between input and state. For the input data of this network layer, For the first Layer bias, For parameters A parameterized state update function is introduced to help the model focus on key features of the input earthquake-related data. An attention mechanism is also introduced to improve the model's ability to capture complex dynamic changes in the data. This improvement in the deep learning model is further refined. , , ,in, For the improved network output, For element-wise multiplication, This is the attention weight vector; Meanwhile, to improve the model's stability, the deep learning model was first modified into an implicit model to better adapt to and process complex seismic data. , , ,in, Let be the transformation matrix of the initial state. Let be the initial state vector. For the process The transformed first The output of the layer, To make the first The hidden state of layer 1 is transitioned to the 1st layer. The weight matrix of the layer, For the first The hidden state of the layer This represents the total number of layers in the neural network. This is the transformation matrix of the last layer of the neural network. For the first To improve the convergence speed of the model and simplify its training and optimization, the output of each layer of the implicit model is proposed. Steps are divided into Sub-steps are used to achieve finer control over the model's behavior, i.e. ,in, In time step and the The output of the layer, In time step and the The output of the layer, For the first The activation function of the layer is determined, and then constraints are added to ensure the stability of the model. The constraints are as follows: ,in, For the first The transformation matrix of the layer, for The transpose of the matrix, For the first The diagonal positive definite matrix of the layer, for The diagonal positive definite matrix of the layer, for The transpose of the matrix is ​​used to define the model's loss function as minimizing the simulation error function, i.e. , ,in, The set of all possible parameters. To measure the simulation error, The total number of time steps. In time step The actual output at that time For time step The model predicts the output at that time. For constraints, including and ,in, In time step Hidden layer output, In time step Hidden layer output, For time step To prevent overfitting, an L2 regularization term is proposed to be added to the loss function for the measured input data. , ,in, For L2 regularization terms, The regularization coefficient is . Using the Frobenius norm, this method realizes the mapping relationship between learning structural response data and seismic input data, improving the deep state-space network for seismic sensing. It first proposes using a method with skip connections. Layered neural networks are used to improve the state-space model to enhance its flexibility in learning the relationship between seismic input data and output. Next, an attention mechanism is introduced to improve the model's ability to capture complex dynamic changes in the data. Then, it is proposed to divide each step of the implicit model into multiple sub-steps to improve the model's convergence speed and simplify model training and optimization. Finally, it is proposed to add an L2 regularization term to the loss function to improve the model's generalization ability, thereby constructing a stable deep learning model.

9. The structural seismic performance evaluation system based on deep learning according to claim 1, characterized in that, The model evaluation and validation unit compares the model's prediction accuracy and generalization ability with test data from the actual structure, in order to evaluate and validate the trained deep learning model and its prediction results. The structural performance assessment unit quantifies the seismic performance of a structure by combining its performance indicators and seismic design codes. It evaluates the seismic performance of a structure based on the structural response data predicted by a deep learning model, providing a basis for seismic reinforcement and repair of the structure.

10. The structural seismic performance evaluation system based on deep learning according to claim 1, characterized in that, The visualization module presents the structural seismic performance assessment results to users in an intuitive chart format, enabling users to more clearly understand the seismic performance of the structure and any existing problems. The model update and maintenance module is used to update and maintain the deep learning model. Based on new earthquake-related data and assessment results, the deep learning model is optimized and updated to improve the model's prediction accuracy and adaptability, allowing the model to continuously learn new knowledge and maintain good performance.

Images