Post-earthquake joint damage prediction method for fabricated steel structure building based on two-step method

Abstract

The application discloses a kind of two-step method-based assembled steel structure building post-earthquake node damage prediction method, it is related to post-earthquake node damage prediction technical field, comprising: obtaining the structural response data of assembled steel structure building under the action of earthquake, including magnitude, time stamp and ground acceleration;Based on the magnitude, time stamp and ground acceleration, the building bottom acceleration is predicted using RIME-LSTM-self-attention algorithm;Again based on magnitude, time stamp, ground acceleration and the building bottom acceleration, the target node acceleration is predicted using ELM algorithm;Park-Ang damage index parameters are determined, the Park-Ang damage index is calculated in combination with the target node acceleration, and the target node state is judged according to the Park-Ang damage index.The application realizes the post-earthquake node damage prediction of assembled steel structure building by two-step prediction method, first predicts the building bottom acceleration of assembled steel structure, then predicts the node acceleration, and finally calculates the damage index.

Description

Technical Field

[0001] This invention relates to the field of post-earthquake node damage prediction technology, and in particular to a two-step method for predicting post-earthquake node damage in prefabricated steel structure buildings. Background Technology

[0002] In the field of structural seismic damage assessment, the dynamic transmission relationship between ground acceleration and building nodal acceleration is the core physical mechanism for assessing damage. Existing techniques obtain structural response data of specific buildings under simulated seismic loading through shaking table tests, establishing a mapping relationship between ground motion input (i.e., base acceleration) and building nodal acceleration response. Based on this, combined with the Park-Ang damage model, the degree of nodal damage is quantitatively assessed. Alternatively, finite element simulations generate large-scale seismic-structural response datasets, and machine learning algorithms are used to construct direct prediction models from ground motion time histories (such as peak ground acceleration, PGA) to nodal damage. Furthermore, data-driven methods integrate ground motion intensity indices (such as CAV, Cumulative Absolute Velocity; ASI, Acceleration Spectrum Intensity) with structural response characteristics (inter-story drift angle, strain energy density) to establish end-to-end damage classification models. All these methods use ground motion characteristics as input, nodal dynamic responses as an intermediate bridge, and ultimately correlate them to the damage state output.

[0003] However, despite the progress made by existing methods in the field of structural seismic damage assessment, many limitations still exist:

[0004] First, data acquisition costs are high and diversity is insufficient: Shaking table tests are limited by the scale and cost of the test, and can usually only cover a single building type, making it difficult to fully reflect the response characteristics of different structures under seismic loading. While finite element simulation can simulate multiple structural types, it relies on high-performance computing resources and the modeling process is time-consuming and laborious, resulting in significant limitations in the scale and diversity of the generated datasets.

[0005] Secondly, the prediction accuracy is limited, making it difficult to break through the boundaries of physical simulation: The prediction accuracy of data-driven models largely depends on the accuracy of finite element simulation or experimental data. However, due to the many uncertainties and simplification assumptions in the physical simulation process, there is a certain error between the simulation data and the actual structural response, which limits the prediction accuracy of data-driven models and makes it difficult to break through the inherent accuracy boundaries of physical simulation.

[0006] Furthermore, the generalization ability is insufficient and it is difficult to adapt to changes in structural parameters: Traditional methods usually perform customized design for specific structural types during the modeling process, which is difficult to adapt to changes in parameters such as node connection form and component size of prefabricated steel structures. When structural parameters change, modeling and training need to be carried out again.

[0007] Furthermore, the utilization of features is insufficient: existing studies often directly map earthquake motion to damage results without effectively decoupling intermediate physical processes such as base acceleration transmission and nodal response evolution. Although this direct modeling approach simplifies the assessment process, it ignores the complex dynamic response mechanism of structures under seismic loading, which affects the reliability and interpretability of the assessment results. Summary of the Invention

[0008] To address the aforementioned technical problems, this invention provides a two-step method for predicting post-earthquake node damage in prefabricated steel structure buildings, thereby enabling the prediction of post-earthquake node damage in prefabricated steel structure buildings.

[0009] According to one aspect of the present invention, a two-step method for predicting post-earthquake node damage in prefabricated steel structure buildings is provided, comprising: simulating different magnitudes and seismic wave actions through shaking table tests to obtain structural response data of the prefabricated steel structure building under seismic loading, wherein the response data includes magnitude, timestamp, and ground acceleration; and predicting the acceleration at the base of the building using the RIME-LSTM-self-attention algorithm based on the magnitude, timestamp, and ground acceleration, wherein the RIME-LSTM-self-attention algorithm is a hybrid model combining optimization algorithms, deep learning components, and machine learning, and optimizes hyperparameters through the RIME algorithm. The LSTM layer captures time-series dependencies, and the Self-attention mechanism enhances key features and provides model interpretability. Then, based on magnitude, timestamp, ground acceleration, and the acceleration at the base of the building, the ELM algorithm predicts the acceleration of the target node. Park-Ang damage index parameters are determined, and combined with the target node acceleration, the Park-Ang damage index is calculated. The Park-Ang damage index parameters include yield strength, shear strength, cross-sectional area, story height, and node mass. The Park-Ang damage index is used to quantify the degree of damage to the target node. The state of the target node is determined based on the Park-Ang damage index.

[0010] Optionally, before predicting the building's base acceleration using the RIME-LSTM-self-attention algorithm based on the magnitude, timestamp, and ground acceleration, the method further includes constructing an LSTM-self-attention prediction network model. This includes: obtaining structural response data and corresponding building base acceleration of prefabricated steel structure buildings under different magnitudes and seismic waves through shaking table tests, constructing a first dataset containing magnitude, timestamp, ground acceleration, and building base acceleration; inputting the first dataset into a bidirectional LSTM network, generating forward and backward hidden state sequences through forward and backward propagation layers respectively, and concatenating them to form a bidirectional fused feature tensor; and then... The fusion feature tensor is input into the attention module, and the input features are mapped to query vectors, key vectors, and value vectors through a linear transformation layer, forming a multi-head attention computation space. The similarity matrix between the query vector and the key vector is calculated using the scaling dot product algorithm, and then normalized by Softmax to generate an attention weight matrix. The weight matrix represents the global dependency strength between features at different time steps. The attention weight matrix and the value vector are weighted and summed to generate a globally enhanced feature representation. A bottleneck compression layer is applied to the globally enhanced feature representation, using a 1×1 convolution kernel to reduce the channel dimension to 1 / 4 of the original input, and layer normalization is used to suppress redundant information. The compressed features are input into a fully connected regression layer to output the predicted acceleration value at the bottom of the building.

[0011] Optionally, after constructing the LSTM-self-attention prediction network model, the method further includes: using the adaptive LSTM hyperparameter configuration of the Frost Optimization algorithm, and jointly optimizing the number of LSTM neurons and the initial learning rate in the LSTM-self-attention prediction network model through a multi-objective fitness function to construct the RIME-LSTM-self-attention prediction model.

[0012] Optionally, before predicting the target node acceleration using the ELM algorithm based on magnitude, timestamp, ground acceleration, and the acceleration at the base of the building, the method further includes constructing an ELM prediction network, including: constructing a second dataset containing magnitude, timestamp, ground acceleration, acceleration at the base of the building, and the acceleration of each corresponding monitoring node; independently constructing a single hidden layer feedforward neural network for each monitoring node, including an input layer, a hidden layer, and an output layer, wherein the input layer is used to receive the raw feature data collected by the sensor, and the dimension is consistent with the number of physical parameters; the hidden layer uses a nonlinear activation function to perform nonlinear mapping on the input signal, and the number of neurons is 2 to 5 times the dimension of the input layer; the output layer is a linear combination of the hidden layer output, used to generate the target prediction value; initial parameters are randomly generated from a uniform or normal distribution, and the initial parameters are used to lock the connection weight matrix and bias vector from the input layer to the hidden layer, which cannot be updated during training; the ELM prediction network is independently deployed to each monitoring node, receiving real-time data streams and synchronously outputting the target node acceleration prediction value.

[0013] Optionally, after constructing the ELM prediction network, the method further includes: using the adaptive ELM hyperparameter configuration of the frost optimization algorithm to jointly optimize the number of ELM neurons in the ELM prediction network through a multi-objective fitness function.

[0014] Optionally, determining the Park-Ang damage index parameters and calculating the Park-Ang damage index in conjunction with the target node acceleration includes:

[0015] Calculate the plastic bending moment based on the yield strength and plastic section modulus of the steel. :

[0016] ;

[0017] In the formula, and These are the yield strength and section modulus of steel, respectively.

[0018] Based on the plastic bending moment And the ultimate lateral load of the lateral force resisting members calculated based on the floor height. :

[0019] ;

[0020] In the formula, This refers to the total floor height of the prefabricated steel structure building.

[0021] According to the ultimate lateral load of the lateral force resisting member And the deformation limit for static loading in the stiffness calculation of components. :

[0022] ;

[0023] In the formula, For component stiffness;

[0024] The plastic deformation energy is quantified by integrating the product of the target node's acceleration and velocity.

[0025] ;

[0026] In the formula, This represents the dissipation of hysteresis energy, i.e., the energy generated during plastic deformation. For node quality, here we take the allocation quality of the target node; and They represent The acceleration and velocity at each time point, where, ;

[0027] According to the plastic deformation energy Deformation limits under static load Calculate the Park-Ang damage index of the target node:

[0028] ;

[0029] In the formula, Indicates the Park-Ang damage index of the target node; Indicates the action of an earthquake. The maximum deformation value before the specified time; Indicates the shear strength of the node. , The cross-sectional area; This represents the energy weighting coefficient.

[0030] Optionally, determining the target node status based on the Park-Ang damage index includes: if the Park-Ang damage index of the target node is greater than 1, then the target node is considered to be damaged.

[0031] According to another aspect of the present invention, a two-step method for predicting post-earthquake node damage in prefabricated steel structures is also provided. Based on the two-step method described above, the device includes: a data acquisition module for simulating different magnitudes and seismic wave actions through shaking table tests to acquire structural response data of the prefabricated steel structure under seismic loading; a building base acceleration prediction module for predicting the building base acceleration using the RIME-LSTM-self-attention algorithm based on the magnitude, timestamp, and ground acceleration; a target node acceleration prediction module for predicting the target node acceleration using the ELM algorithm based on the magnitude, timestamp, ground acceleration, and building base acceleration; and a node damage estimation module for determining Park-Ang damage index parameters, calculating the Park-Ang damage index in conjunction with the target node acceleration, and determining the target node status based on the Park-Ang damage index.

[0032] According to another aspect of the present invention, a computer device is also provided, the computer device including a memory, a processor and a computer program stored in the memory and executable on the processor, the computer program being executed by the processor to implement the steps of the two-step method for predicting post-earthquake node damage of prefabricated steel structure buildings as described above.

[0033] According to another aspect of the present invention, a storage medium is also provided, on which a computer program is stored, which, when executed by a processor, implements the steps of the two-step method for predicting post-earthquake node damage in prefabricated steel structure buildings as described above.

[0034] This invention collects structural response data of prefabricated steel structure CFS unit buildings through a series of shaking table tests, and establishes a node damage dataset based on this. Combining structural response characteristics and machine / deep learning algorithms, a two-step prediction method for node damage is constructed. This method decomposes node damage prediction into bottom acceleration prediction and node acceleration prediction. First, the bottom response of the entire building is predicted. Then, based on this response, the local response of key nodes is refined and predicted. Finally, by combining the Park-Ang and node acceleration prediction results, the degree of damage is quantified through a mechanical model, realizing the prediction of structural node damage from seismic input.

[0035] This invention addresses the high standardization of prefabricated steel structures by generating a high-precision full-node response dataset through controllable-cost shaking table tests. It proposes a two-stage decoupled prediction framework of "base acceleration prediction → node acceleration transmission → damage quantification," which breaks through the accuracy bottleneck of end-to-end mapping of a single model. At the same time, it utilizes structural consistency to achieve efficient generalization capability of "one test, multiple parameter reuse."

[0036] By decomposing the complex global prediction task into two simpler sub-problems and using intermediate physical quantities for information transfer, the learning difficulty and overall computational requirements of the model are effectively reduced, thus decreasing computational load and improving the model's robustness and accuracy. Specifically, directly predicting node acceleration from earthquake input requires the model to learn a highly nonlinear mapping relationship between ground motion and complex local node responses. Such tasks have high feature dimensions and large parameter scales, leading to a significant increase in computational burden during the training and inference phases. The two-step method proposed in this invention adopts a layer-by-layer transfer strategy. First, it predicts the acceleration at the base of the building (a relatively simple task with more obvious regularity), and then predicts the acceleration of the target node based on this intermediate result. This "compresses" the original high-dimensional input into a low-dimensional intermediate variable with clear physical meaning—the base acceleration—thereby significantly reducing the complexity of interactions between features. The number of parameters and computational load of each of the two sub-models are smaller than that of a single large end-to-end model, reducing the network structure size and not only lowering the overall computational cost but also accelerating the model convergence speed.

[0037] This invention enhances temporal feature extraction capabilities by introducing a self-attention mechanism; it also improves model generalization by optimizing the hyperparameters of LSTM and ELM using the RIME algorithm. Furthermore, this invention verifies the effectiveness of the RIME-LSTM-self-attention and ELM combined algorithm in nodal acceleration prediction, the effectiveness of incorporating a self-attention mechanism in reducing error propagation, and the effectiveness of nodal damage estimation based on the Park-Ang damage index. The nodal damage prediction method proposed in this invention improves prediction accuracy while rapidly assessing post-earthquake damage, demonstrating its significant engineering application value. Attached Figure Description

[0038] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings:

[0039] Figure 1 This is a flowchart of a two-step method for predicting post-earthquake node damage in prefabricated steel structure buildings, as described in Embodiment 1 of the present invention.

[0040] Figure 2 This is a schematic diagram of a prefabricated steel structure building.

[0041] Figure 3 The diagram shows the layout of measurement points. In the diagram, A1~A8 are nodes where accelerometers are installed, and L1~L8 are nodes where displacement gauges are installed.

[0042] Figure 4 A schematic diagram of the ground acceleration selected for the shaking table test;

[0043] Figure 5 This is a schematic diagram of nodal damage under 0.4 g KKM GA.

[0044] Figure 6 The flowchart shows a two-step node damage prediction method.

[0045] Figure 7 This is a schematic diagram of the LSTM-self-attention network architecture;

[0046] Figure 8 A schematic diagram showing the predicted results of nodal acceleration and building base acceleration;

[0047] Figure 9 This is a diagram illustrating the degree of agreement between the predicted values ​​and the actual values ​​in the test set.

[0048] Figure 10 A comparison chart showing the original nodal acceleration and the nodal acceleration predicted by the two-step method;

[0049] Figure 11 This is a schematic diagram illustrating the accuracy improvement effect after adopting the two-step prediction method;

[0050] Figure 12 A schematic diagram of the Dpa index predicted using the two-step method;

[0051] Figure 13 This diagram illustrates the maximum damage prediction error for each node. Detailed Implementation

[0052] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, and not all of them. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present application. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of the present application can be combined with each other.

[0053] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0054] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0055] To predict node damage, accurate prediction of node acceleration is required. However, in the case of node acceleration (… In the complex process of earthquake prediction, the ground acceleration record is not a single, regular signal; its frequency components are extremely complex and exhibit significant differences. Simultaneously, the structural response under seismic loading is closely linked to the frequency components of the input ground acceleration, which is in turn influenced by a multitude of complex factors. Among these, site soil conditions play a crucial role. Different types of soil, such as sand and clay, exhibit drastically different characteristics in the propagation and absorption of seismic waves. For instance, sand may allow seismic waves to propagate at a faster speed while absorbing relatively less of certain frequency components, while clay may slow down the propagation speed of seismic waves and significantly absorb energy within a specific frequency range. This difference directly leads to a significant change in the frequency distribution of ground acceleration, thereby affecting the structural response and nodal acceleration.

[0056] Earthquake magnitude is a key indicator of the amount of energy released during an earthquake. Higher magnitudes release more complex energy, resulting in more diverse frequency components of ground acceleration. Furthermore, the distance between the structural site and the epicenter or seismogenic fault determines the attenuation and waveform changes of seismic waves during propagation, thus significantly influencing the frequency components of ground acceleration. Ground acceleration (…) ) and peak ground acceleration (PGA, the peak value of the ground acceleration time history) and nodal acceleration ( There is a high correlation between them, which provides certain clues and basis for the prediction of nodal acceleration.

[0057] Based on the above analysis, this invention designs a two-step prediction method for nodal damage. This method decomposes the earthquake-structural response-nodal damage relationship into two easily statistically and analytical stages: nodal damage prediction is broken down into two steps: building base acceleration prediction and nodal acceleration prediction. Dedicated prediction models are trained for each of these two stages. In the first step, the building base acceleration is predicted using known earthquake information. In the second step, the nodal acceleration is further predicted using the building base acceleration as input. Finally, based on the outputs of these two prediction models and the determined nodal damage parameter values, the degree of nodal damage is scientifically estimated. This step-by-step prediction method not only improves the accuracy and reliability of prediction but also provides more effective technical support for structural seismic design and damage assessment.

[0058] Specifically, refer to Figure 1 , Figure 1 This is a flowchart of a two-step method for predicting post-earthquake node damage in prefabricated steel structure buildings according to an embodiment of the present invention, as shown below. Figure 1 As shown, the method includes:

[0059] S1 uses shaking table tests to simulate different magnitudes and seismic wave effects to obtain structural response data of prefabricated steel structure buildings under seismic loading;

[0060] The response data includes magnitude, timestamp, and ground acceleration. Magnitude reflects the intensity of the earthquake, timestamp captures the temporal dynamics of the earthquake process, and ground acceleration characterizes the acceleration response at a specific location on the ground under earthquake action, and is one of the basic characteristics of an earthquake.

[0061] In order to accurately predict the acceleration at the base of a building ( In this embodiment of the invention, the input parameters are selected as magnitude (M), time (t), and ground acceleration (m). ).

[0062] S2, based on the magnitude, timestamp, and ground acceleration, the RIME-LSTM-self-attention algorithm is used to predict the acceleration at the bottom of the building;

[0063] The RIME-LSTM-self-attention algorithm is a hybrid model that integrates optimization algorithms, deep learning components, and machine learning models. This hybrid model uses RIME (Rime optimization algorithm) as the hyperparameter optimizer to automatically search for the optimal hyperparameter configuration in the model; it utilizes a Long Short-Term Memory (LSTM) network to capture long-term temporal dependencies in seismic response data, overcoming the gradient vanishing or exploding problems of traditional recurrent neural networks; and it introduces a self-attention mechanism to enhance the representation ability of key features while improving the model's interpretability.

[0064] LSTM, a special type of recurrent neural network (RNN), effectively processes long sequence data and is specifically designed to address the vanishing or exploding gradient problems encountered by standard RNNs when handling long sequences. Through convolutional layers or attention mechanisms, it sequentially scans the normalized input data to effectively extract spatial features, such as identifying anomalous fluctuation patterns in specific regions from the spatial distribution of seismic acceleration. Subsequently, LSTM layers capture general temporal features, and a forget gate filters out anomalous fluctuation patterns. In earthquake data prediction scenarios, this improves the model's robustness and accuracy, thus providing strong support for accurately predicting the current building base acceleration.

[0065] Self-attention is a type of attention mechanism. Its core idea is to allow the model to dynamically weigh the importance of different parts of the input sequence, enabling the model to capture global dependencies between distant features, enhance key features, and improve the model's interpretability. Using self-attention, the network can effectively learn global long-distance dependencies between geographically distant regions in the entire feature map of its inner layers.

[0066] In this invention, LSTM is combined with a self-attention mechanism to better learn nonlinear features with fewer parameters and lower computational cost. By processing sequences in both forward and backward directions and integrating these two representations, the LSTM-self-attention hybrid network can capture more complex dependencies.

[0067] Furthermore, before using the RIME-LSTM-self-attention algorithm to predict the acceleration at the base of a building, it is necessary to construct and train an LSTM-self-attention prediction network model, specifically including:

[0068] S201, Constructing the dataset:

[0069] Establish a system that includes magnitude (M), timestamp (t), and ground acceleration (t). ) and acceleration at the base of the building ( The first dataset;

[0070] S202, Bidirectional extraction of temporal features:

[0071] The first dataset is input into a bidirectional LSTM network, and forward and backward propagation layers are used to generate forward and backward hidden state sequences, respectively, which are then concatenated to form a bidirectional fused feature tensor.

[0072] S203, Global Dependency Dynamic Modeling:

[0073] Input the bidirectional fused feature tensor into the self-attention module and perform the following operations:

[0074] (a) Feature projection: The input features are mapped into query vector, key vector and value vector through a linear transformation layer, forming a multi-head attention computation space;

[0075] (b) Weight calculation: The similarity matrix between the query vector and the key vector is calculated using the scaling dot product algorithm, and the attention weight matrix is ​​generated by Softmax normalization. The weight matrix represents the global dependency strength between features at different time steps.

[0076] (c) Feature aggregation: The attention weight matrix and the value vector are weighted and summed to generate a globally enhanced feature representation;

[0077] S204, Lightweight Feature Compression:

[0078] A bottleneck compression layer is applied to the global enhanced feature representation, a 1×1 convolution kernel is used to reduce the channel dimension to 1 / 4 of the original input, and redundant information is suppressed by layer normalization;

[0079] S205, Nonlinear Prediction Output:

[0080] The compressed features are input into a fully connected regression layer, and the output is the predicted structural response, i.e., the predicted acceleration at the bottom of the building.

[0081] S206, Construct an LSTM-self-attention prediction network model:

[0082] The above processes are integrated sequentially in MATLAB to complete the construction of the LSTM-self-attention prediction network model;

[0083] S207, RIME hyperparameter optimization:

[0084] By utilizing RIME's adaptive LSTM hyperparameter configuration, the number of LSTM neurons and the initial learning rate in S206 are jointly optimized through a multi-objective fitness function. The objective function is prediction accuracy.

[0085] RIME is a highly efficient optimization algorithm based on the physical phenomenon of rime ice. The RIME algorithm simulates the growth processes of soft and hard rime ice and constructs a soft rime ice search strategy and a hard rime ice penetration mechanism to realize exploration and utilization behaviors in the optimization method. RIME mainly consists of two stages: the first stage is soft rime ice growth, representing the exploration stage, where potential solutions are widely sampled. The second stage is hard rime ice penetration, reflecting the utilization stage, where local obstacles are overcome to optimize the solution. By simulating these natural processes, RIME can effectively balance global exploration and local exploitation during the optimization process, ensuring diverse exploration and targeted optimization.

[0086] "Adaptive LSTM hyperparameter configuration" refers to an automated and intelligent process. Its core is to use the RIME optimization algorithm to dynamically and data-drivenly find and set the most important hyperparameters in the LSTM network, rather than relying on researchers' experience for manual setting. Two key hyperparameters are configured:

[0087] ① Number of LSTM Units / Hidden Size: This refers to the dimension of the hidden states in the LSTM layer, i.e., how many LSTM units there are. It directly determines the model's ability and complexity to capture temporal patterns.

[0088] ② Initial Learning Rate: This is the step size that the optimizer (such as Adam) uses to update the network weights at the beginning of training. It is one of the most important hyperparameters in deep learning training.

[0089] This process is called "adaptive" because the RIME algorithm does not search blindly. It intelligently and dynamically adjusts its search strategy based on feedback from the fitness function (i.e., the model's performance on the validation set). High-performing hyperparameter combinations are retained and used, while low-performing ones are discarded. The algorithm "adapts" to the current problem and data characteristics, automatically evolving the most suitable hyperparameters, thus achieving "adaptive" configuration.

[0090] "Multi-objective" refers to the desire for the model to simultaneously optimize multiple performance metrics that may conflict with each other, including:

[0091] ①Accuracy Objective: To achieve the highest possible prediction accuracy, i.e., to minimize prediction error. Examples include mean squared error (MSE) and mean absolute error (MAE).

[0092] ② Complexity Objective: Pursuing simplicity and efficiency in the model. For example, fewer neurons, smaller model size, and faster training and inference speeds.

[0093] The goal of multi-objective optimization is to find an optimal balance between accuracy and complexity (Pareto optimal solution), that is, to achieve the highest possible accuracy with a relatively simple model.

[0094] The LSTM-self-attention model optimized by RIME to meet the accuracy requirements (i.e., the RIME-LSTM-self-attention model) will be used to predict the acceleration at the base of buildings. This model takes magnitude, timestamp, and ground acceleration as input to achieve an accurate estimate of the acceleration at the base.

[0095] S3. Based on the magnitude, timestamp, ground acceleration, and the acceleration at the bottom of the building, the ELM algorithm is used to predict the acceleration of the target node.

[0096] The prediction of (nodal acceleration) follows immediately after the prediction of the building's base acceleration. This step must be performed strictly in sequence, and each nodal acceleration requires a separately trained prediction model. The input parameters for this step, besides continuing to use M, t, and [other parameters], are the same as those used for predicting the building's base acceleration. In addition, the predicted value of the building bottom acceleration obtained from S2 is used as an input parameter to predict the nodal acceleration.

[0097] First, the building base acceleration prediction model (i.e., RIME-LSTM-self-attention) is invoked to predict the building base acceleration. Then, the nodal acceleration prediction model is used, taking the magnitude, timestamp, ground acceleration, and the building base acceleration as input parameters, and the output is the predicted nodal acceleration value.

[0098] Considering the high correlation between node acceleration and building base acceleration, and the fact that each node requires a separate training of its corresponding prediction model, this invention selects Extreme Learning Machine (ELM) as the node acceleration prediction algorithm to significantly improve training efficiency. ELM is a single-hidden-layer feedforward neural network, unique in that the connection weights w and bias parameters b between the input and hidden layers are initialized randomly and remain fixed during training. The model consists of three layers: the input layer receives the original data features, the hidden layer performs feature transformation on the random parameters using a non-linear activation function (such as Sigmoid or ReLU), and the output layer directly obtains the optimal connection weights β through analytical calculation. Unlike traditional neural networks, ELM does not require iterative adjustment of the hidden layer parameters; instead, it solves for the output layer weights in one step using the Moore-Penrose generalized inverse matrix. This unique mechanism gives it extremely fast training speed (typically reaching the second level) and excellent generalization performance. This model demonstrates significant advantages in classification, regression, and real-time prediction, especially when dealing with high-dimensional data and large-scale datasets. It can effectively overcome the problems of slow convergence speed and easy getting trapped in local optima that exist in traditional gradient descent algorithms.

[0099] Similarly, before using the ELM algorithm to predict the acceleration of the target node, it is necessary to construct and train the ELM prediction network, specifically including:

[0100] S301, Construct the dataset:

[0101] Establish a system that includes magnitude (M), timestamp (t), and ground acceleration (t). ), building base acceleration ( ) and the acceleration of a certain node ( The second dataset;

[0102] S302, a node-specific model architecture:

[0103] A single hidden layer feedforward neural network is independently constructed for each monitoring node, which includes: an input layer (receiving the raw feature data collected by the sensor, with the dimension consistent with the number of physical parameters), a hidden layer (using a non-linear activation function, such as the Sigmoid or ReLU function, to perform a non-linear mapping on the input signal, with the number of neurons being 2-5 times the dimension of the input layer), and an output layer (linearly combining the output of the hidden layer to generate the target prediction value).

[0104] S303, Random parameter initialization:

[0105] Perform the following operations on the connection weight matrix and bias vector from the input layer to the hidden layer:

[0106] (a) Randomly generate initial parameters from a uniform or normal distribution;

[0107] (b) The parameters are locked and cannot be updated during training;

[0108] S304, multi-node parallel deployment:

[0109] The model is independently deployed to each monitoring node to receive real-time data streams and synchronously output prediction results, including structural displacement, stress, and damage probability.

[0110] S305, Constructing the ELM prediction network:

[0111] Integrate the above process in MATLAB to complete the construction of the ELM feedforward neural network;

[0112] S306, RIME hyperparameter optimization:

[0113] By utilizing the adaptive ELM hyperparameter configuration of RIME (Frost Optimization Algorithm), the number of ELM neurons in S305 is jointly optimized through a multi-objective fitness function.

[0114] The relevant description of RIME hyperparameter optimization in this step can be found in S207 above, and will not be repeated here.

[0115] S4, determine the Park-Ang damage index parameters, and calculate the Park-Ang damage index in combination with the target node acceleration.

[0116] In this embodiment of the invention, the Park-Ang damage index parameters of prefabricated steel structure buildings include yield strength ( ), shear strength ( ), cross-sectional area ( ), floor height ( ) and node quality ( ).

[0117] To assess nodal damage, this invention employs the Park-Ang index, which combines displacement and accumulated hysteresis energy. Park and Ang define a damage index that combines ductility and accumulated hysteresis energy requirement:

[0118] ;

[0119] In the formula, This invention represents the deformation limit under static load. The calculation is based on the elastic limit, and the value is taken as... ,in, For the ultimate lateral load of the lateral force resisting member, This refers to the floor height (the total floor height of the prefabricated steel structure building). This refers to the stiffness of the component. Based on plastic bending moment Calculate, take as ,in , and The yield strength and plastic section modulus of steel are respectively measured. In addition, Indicates the action of an earthquake. The maximum deformation value before the specified time. Represents the shear strength of the node, by calculate, It refers to the cross-sectional area (i.e., the cross-sectional area of ​​the steel structure; the area of ​​the cross-section where the node is located is the cross-sectional area). This represents the energy weighting coefficient, which is set to 0.02518 due to the special characteristics of cold-formed thin-walled steel structures. The hysteresis energy dissipation is represented by the integral of the product of acceleration and velocity, which quantifies the energy of plastic deformation. The calculation formula is as follows:

[0120] ;

[0121] in, and They represent The acceleration and velocity at time points are taken as follows: , The node quality is simply the allocation quality of that node.

[0122] S5, determine the target node status based on the Park-Ang damage index.

[0123] The method for predicting post-earthquake node damage in prefabricated steel structure buildings in this embodiment of the invention consists of two steps: first, predicting the acceleration at the bottom of the building; then, based on the prediction result of the acceleration at the bottom of the building, further predicting the node acceleration; and finally, using the Park-Ang damage index to estimate the damage. If the calculated damage index is greater than 1, the target node is considered to be damaged.

[0124] The prefabricated steel structure building in this invention, as a new generation of green building system, has the following practical advantages:

[0125] ① High construction efficiency: The use of prefabrication significantly shortens the construction period;

[0126] ② Reusable assembly: Modular components facilitate transportation and quick disassembly and replacement;

[0127] ③ Widely used in earthquake zones: Its lightweight and high-strength material properties make it an important structural form in earthquake-prone areas, and it is particularly suitable for functional buildings such as schools, hospitals, and resettlement housing.

[0128] The practical application value of the two-step method for predicting post-earthquake joint damage in prefabricated steel structure buildings proposed in this invention includes:

[0129] ① Rapid post-earthquake assessment: Traditional manual assessment is inefficient and severely delayed after an earthquake; this method can rapidly predict and assess the post-earthquake damage status of nodes, providing data support for structural safety assessment and emergency decision-making.

[0130] ②Support intelligent urban renewal: Building a "digital twin" prediction model for prefabricated buildings to achieve automatic identification of structural health status is a key support for future intelligent construction and smart city development.

[0131] ③ Reduce manual inspection costs: Model prediction replaces the repeated deployment of sensors, reducing operation and maintenance and monitoring costs and improving the feasibility of engineering practice.

[0132] Compared to directly using sensors to measure (nodal acceleration), this invention has the following advantages:

[0133] ① Sensor deployment has physical and economic limitations:

[0134] In actual engineering projects, the total number of nodes far exceeds the number of sensors that can be deployed, especially in high-rise and long-span structures, where costs and deployment difficulties increase exponentially. Some critical connection points are difficult to physically place sensors due to structural complexity or space constraints. Under seismic conditions, sensors may fail, go offline, or experience power outages, therefore, we cannot rely entirely on measured data.

[0135] ②The two-step method of this invention has the advantages of predictive generalization ability and model simplification:

[0136] By collecting data using a small number of sensors deployed at the bottom or surface, the response of other nodes can be predicted, demonstrating good scalability and engineering applicability. The recursive approach of "predicting the bottom response first, then the node response" can significantly reduce data dimensionality and model complexity, which helps with training convergence and model robustness. By utilizing the temporal-spatial feature extraction capabilities of the LSTM-self-attention network, complex response relationships under different magnitudes and structural configurations can be accurately learned.

[0137] ③ This method embodies a technical approach of "controllable cost + comprehensive prediction + strong engineering adaptability":

[0138] Compared with direct sensor measurement, this method is a more scalable and maintainable intelligent structural health diagnosis strategy that can be applied to a large number of existing prefabricated buildings, and is particularly suitable for later renovation and post-disaster assessment scenarios.

[0139] In summary, the method of the present invention has high application value in terms of cost, practicality, adaptability and scalability.

[0140] Example 2: The following specific case provides a detailed description of the present invention and its effectiveness.

[0141] Precast structures ensure consistent manufacturing quality, reduce on-site labor, and accelerate structural deployment. Cold-formed steel (CFS) offers excellent strength-to-weight ratio and corrosion resistance; its use can improve performance in seismic zones.

[0142] The proposed structure in the embodiments of the present invention is as follows: Figure 2 As shown, Figure 2 This study evaluated the seismic performance of prefabricated double-layer cold-formed steel CFS units under various seismic excitations. This modular, lightweight structure is suitable for residential, agricultural facilities (such as farmhouses), or temporary shelters. High-strength cold-formed steel components are employed, including SHS 50×50×1.6 mm hot-dip galvanized columns, RHS 40×75×0.8 mm galvanized roof trusses, UB 127×76×13 kg / m hot-dip galvanized beams, and FLR CEE 50×100 / 150×1.2 mm galvanized steel. All components conform to JISG3302 SGC 570 and SGC 440 standards, with bolts of grade 8.8.

[0143] In structural studies such as earthquake research, full-scale models consistently offer significant advantages, providing reliable and realistic results for the structure. However, due to various limitations in laboratories, particularly when the size of experimental equipment (such as shaking tables) is often fixed and cannot be customized for each experiment, scaled-down models are typically used for experimental analysis. In this embodiment, the scaled-down 1:2 model (i.e., a 1:2 scaled-down version of the prototype for model testing) has a total planar area of ​​approximately 9.00 square meters (3.05 meters × 2.95 meters) and a height of approximately 2.95 meters. The model simulation rules are based on scaling theory. The three fundamental scaling factors for geometry, acceleration, and elastic modulus are chosen as ½, 1, and 1, respectively. Due to capacity limitations of the testing facility, a ½ length scale was chosen. Perfect similarity is maintained by setting the acceleration and elastic modulus scaling factors to 1, achieving true modeling accuracy. When using the same material, if the elastic modulus SE = 1, the stress and strain in the model and prototype must be identical. Other parameters, such as acceleration, force, and time, are scaled down according to the fundamental scaling factors. (These do not need to be listed here, as the parameters follow universally agreed-upon rules).

[0144] Based on the similarity law, all models have dimensions of 3050mm × 2950mm × 2950mm and a 684mm roof. The experiment features a prefabricated double-layer pre-cooled bent steel (CFS) structure. For example... Figure 3As shown, this structure features an open-plan parking lot or storage room on the ground floor, with the first floor elevated as the main living area. The second-floor slab is 1100mm from the bottom of the structure and measures 1550mm × 2950mm. The model is bolted to the base plate and aligned to ensure accurate motion transmission. Sensors are placed at key locations to capture the structure's dynamic response to each input ground movement. In the diagram, A1~A8 are nodes where accelerometers are installed, and L1~L8 are nodes where displacement gauges are installed.

[0145] To analyze the seismic response mode of this structure, this embodiment uses five historical earthquake ground accelerations, including:

[0146] (a) El Centro (1940, USA) -- a slip fault with a strength of 0.40 gg;

[0147] (b) Kobe (1995, Japan) -- Crustal fault with a strength of 0.32g;

[0148] (c) Tohoku (2011, Japan) -- Subduction fault, strength tested up to 0.16g;

[0149] (d) Tabas (1978, Iran) -- High-strength fault with a strength of 0.16g;

[0150] (e) KKM (2015, Malaysia), strength up to 0.16g.

[0151] The original seismic acceleration wave, such as Figure 4 As shown, from top to bottom, the values ​​are 0.40g, 0.32g, 0.16g, 0.16g, and 0.16g. Due to the limitations of the vibration table, El Centro, Kobe, and Tohoku only tested up to the 0.16g level.

[0152] Each record was scaled at six different PGA levels: 0.08g, 0.12g, 0.16g, 0.24g, 0.32g, and 0.40g. Figure 5 The selected PGA range represents a broad spectrum of seismic intensity from low to high, consistent with common classifications in seismic design standards such as Eurocode 8 and ASCE 7.

[0153] This scale structure was fabricated using prefabricated commercial CFS components and assembled on-site in a structural laboratory. Steel plates were added to the structure to replicate actual loads representing furniture and occupied space. Testing was conducted on a 3 m × 3 m unidirectional shaking table driven by Shimadzu dynamic push rods with a travel of ±50 mm and a load capacity of 30 tons. White noise excitation was performed before each earthquake input to verify that the structure's fundamental frequency remained consistent. The instrumentation included eight Kyowa AS-2GB accelerometers and eight LVDTs (DT-50A, DT-100A, and DTS-A-100), mounted at key connections on the first, lower, and upper roofs, respectively. This embodiment of the invention only discusses acceleration; the accelerometer arrangement is as follows... Figure 3 As shown in (A1~A8), the Kyowa EDX-10 compact recording system is used to synchronize sensor readings with the drive system of the vibration table.

[0154] To assess structural damage after an earthquake, the focus should be on the damage at beam-column connections. Therefore, the damage at joints A2, A3, A6, A7, and A8 was calculated, and a joint damage dataset was established. Based on structural information, the Park-Ang index parameters for each node were analyzed and collected, as shown in Table 1. In the table, f... y D represents the ultimate strength of steel. pa-max This represents the maximum value in the Park-Ang damage index of the target node. In this embodiment, D represents the maximum damage index if it exceeds 1. pa-max If the value is greater than 1, the node is considered to be damaged.

[0155] Table 1: Park-Ang Index Parameter Table

[0156]

[0157] Figure 5 The acceleration and displacement (Dpa) of the nodes under the action of the 0.4g KKM wave are shown, representing the damage under the seismic wave. The results show that, despite local differences, the acceleration and displacement waveforms exhibit high similarity, and nodes A2, A3, A6, A7, and A8 fail under the action of the 0.4g KKM wave approximately 16 seconds after the seismic wave arrives.

[0158] Reference Figure 6 , Figure 6 The flowchart of the two-step node damage prediction method in this embodiment of the invention is as follows: Figure 6 As shown, firstly, the magnitude, timestamp, and ground acceleration are input into the first-step prediction model to predict the acceleration at the bottom of the building; then, the magnitude, timestamp, ground acceleration, and the acceleration at the bottom of the building obtained in the first step are input into the second-step prediction model to predict the acceleration of the target node, and thus predict the node damage.

[0159] The most commonly used metrics for evaluating the performance of predictive models include the coefficient of determination (R²). 2 The root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE) are also considered. 2 The values ​​range from 0 to 1, with values ​​closer to 1 indicating a better model fit. RMSE is a common metric for measuring the difference between observed and true values; a lower RMSE value indicates a smaller deviation between the observed and true values. MAE and MAPE calculate the mean and percentage of the absolute values ​​of the differences between predicted and actual values, respectively; smaller values ​​are better.

[0160] This invention employs two networks: CNN-SEattention-LSTM and LSTM-self-attention. LSTM-self-attention is based on the LSTM network and is primarily used for learning long-distance dependencies in temporal data. It processes temporal information through LSTM and then combines it with a self-attention mechanism to enhance feature representation capabilities. Figure 7 (Diagram of LSTM-self-attention network architecture); CNN-SEattention-LSTM introduces a convolutional neural network (CNN) and a compressed and activated self-attention mechanism (SEattention), which can extract local features and combine them with temporal information, improving the ability to process different types of input data. The RIME optimization algorithm is used as the hyperparameter optimizer to optimize the hyperparameters of both networks. The RIME-optimized LSTM-Self-attention is a network with 185 neurons and an initial learning rate of 10. -4 The neural network. Similarly, the RIME-optimized CNN-SEattention-LSTM has a neural network with 127 neurons and an initial learning rate of 0.057. Table 2 shows the prediction performance of these two algorithms. The results show that after RIME optimization of the hyperparameters, LSTM-Self-attention achieves better performance. In prediction problems, it demonstrates significantly higher accuracy than CNN-SEattention-LSTM. The RMSE results of LSTM-Self-attention show that... The mean absolute error of the prediction is 0.0464 m / s 2The results show that the RIME-LSTM-self-attention model exhibits significant advantages in evaluation metrics. Its performance improvement stems from two key technological innovations: First, the self-attention mechanism effectively captures the energy concentration phenomenon in the 3-5Hz frequency band during seismic wave propagation by dynamically calculating the correlation weights between different locations in the feature map. This frequency band has been confirmed to have a deterministic coefficient of 0.87 with node damage. Second, the RIME optimization algorithm achieves adaptive configuration of the LSTM network structure during the parameter tuning stage. Through global search during the frost growth stage and local optimization during the puncture stage, the model's prediction error in highly dynamic scenarios such as Tabas waves is reduced by 42%. Therefore, LSTM-Self-attention can be selected as a method for node damage prediction. Prediction algorithm.

[0161] Table 2: Evaluation Indicators for Step 1 Prediction

[0162]

[0163] In the second prediction step, after RIME optimization, the optimal model obtained is an ELM feedforward neural network with 96 neurons. Figure 8 This is a schematic diagram illustrating the predicted results of nodal acceleration and building base acceleration. Figure 8 In this context, LSTM-Selfattention refers to the prediction result of the LSTM-self-attention network for the acceleration at the bottom of the building. Figure 9 This demonstrates the effect of 0.4g KKM seismic waves in the test set. and Comparison of predicted and actual values. R 2 The mean reached 0.98, and the mean RMSE and MAE reached 0.068 and 0.038 respectively. Furthermore, ELM performed well at each node. In terms of prediction, the calculated results of the evaluation indicators are highly similar, but the prediction accuracy of nodes A6 and A7 shows a significant decrease. This indicates that ELM has generally high prediction accuracy and consistency in nodal acceleration prediction. Furthermore, a comparison chart of predicted value-true value fit is introduced, such as... Figure 9 As shown. Figure 9 The red dashed line represents the 3σ limit of the prediction error. It can be observed that the vast majority of predicted points are within the 3σ limit. However, only the prediction of node A1 is observed, which may lead to prediction bias in the prediction of local acceleration peaks. Therefore, ELM can be used... The prediction.

[0164] After completing the step-by-step verification of the algorithm, a two-step method for predicting node acceleration was implemented, namely, the prediction of node A1's acceleration. Predicted value, as The input parameters for prediction are shown in Table 3. The prediction accuracy results for each node are also shown in Table 3. Thanks to the combined effect of ELM's random weight initialization mechanism and generalized inverse solving, the model can meet the requirements of high-precision damage assessment with a small sample size. The two-step prediction accuracy R of the RIME-LSTM-self-attention+ELM model is shown in Table 3. 2 The average value reached 0.98. The MAE metric indicates that although the two-step method can capture peak abrupt changes in acceleration relatively well, there is still a slight systematic underestimation. This may be due to the insufficient proportion of high-intensity samples of Tabas waves (PGA=0.32g) and KKM waves (PGA=0.40g) in the training data (accounting for only 17% of the dataset).

[0165] Table 3: Prediction accuracy of nodal acceleration under the two-step method

[0166]

[0167] The self-attention mechanism introduced in this invention effectively captures the global correlation of time-frequency features during seismic wave propagation and enhances the modeling ability of high-frequency components by dynamically adjusting feature weights. Meanwhile, the RIME optimization algorithm successfully avoids local optima traps during parameter tuning, achieving optimal configuration of the number of neurons and learning rate in the hidden layers of the LSTM network. Furthermore, as... Figure 10 As shown in the figure, the independent LSTM-self-attention The predictions are based on the "Original" data in the legend, while the two-step framework predicts data based on the "Two-step" data. The graph shows that the "Original" data has a wider bandwidth and higher data dispersion, while the "Two-step" data has a significantly narrower bandwidth and significantly higher prediction accuracy. Therefore, compared to independent LSTM-self-attention... The two-step prediction showed a significant improvement in accuracy. This indicates that the two-step framework significantly improves the accuracy of nodal predictions, and the error in the first step is mitigated in the second step, highlighting the importance of separating ground acceleration prediction from nodal acceleration prediction. Figure 11 This is a schematic diagram illustrating the improved accuracy after using the two-step prediction method.

[0168] Subsequently, based on the two-step prediction results, the node damage of each node was calculated, such as... Figure 12 As shown in the figure. The results indicate that this method can capture the dynamic process of damage evolution over time and maintain consistency with... The method achieves consistent prediction accuracy. Furthermore, it accurately distinguishes the degree of damage to different nodes, reflecting the differences between local and global damage, thus providing a basis for targeted node maintenance. Figure 13The maximum damage prediction error for each node is shown, with an average prediction error of approximately 9% and a maximum prediction error reaching 15%. The prediction errors mainly occur at points of abrupt changes in the Dpa value. Combined with the damage index calculation formula, it can be seen that the error is based on the integral obtained from the prediction results. The value showed a significant error. However, in order to obtain... The value requires integration twice. Figure 9 The results show that there is a large error in the prediction of the peak abrupt change in acceleration. Figure 9 The red dashed line represents the 3σ limit of the prediction error. It can be observed that the vast majority of predicted points are within the 3σ limit. Predictions for only nodes A6 and A7 are observed, but prediction biases are prone to occur in the prediction of local acceleration peaks (the gray-white spheres in the figure indicate these biases). Therefore, the conversion... The process of calculating values ​​will greatly amplify the prediction error. This indicates that further research is needed on nodal velocity prediction.

[0169] The two-step prediction framework significantly reduces the error accumulation effect of traditional end-to-end models by decoupling ground motion feature extraction from local nodal response mapping. Experimental results show that the combined model based on RIME-LSTM-self-attention+ELM maintains robustness under complex seismic wave inputs (such as near-fault pulse-type ground motions), and its nodal acceleration prediction accuracy is improved by 58% compared to the single-stage model. This breakthrough achievement is attributed to the physics-guided attention mechanism design—the self-attention layer not only enhances the energy feature extraction of key frequency bands (3~5Hz) through channel weight calibration, but also dynamically analyzes the cross-scale correlation between nodal acceleration and damage indicators. This framework shows significant improvement in prediction stability under extreme conditions such as the Malaysian KKM wave (PGA=0.40g), providing high-confidence decision support for post-earthquake damage assessment of prefabricated steel structures.

[0170] Example 3: This example provides a two-step method-based device for predicting post-earthquake joint damage in prefabricated steel structures, used to implement the above examples and preferred embodiments, including:

[0171] The data acquisition module is used to simulate different magnitudes and seismic wave effects through shaking table tests to acquire structural response data of prefabricated steel structure buildings under seismic loading. The response data includes magnitude, timestamp, and ground acceleration.

[0172] The building base acceleration prediction module is used to predict the building base acceleration based on the magnitude, timestamp, and ground acceleration using the RIME-LSTM-self-attention algorithm.

[0173] The target node acceleration prediction module is used to predict the target node acceleration based on magnitude, timestamp, ground acceleration, and the acceleration at the bottom of the building, using the ELM algorithm.

[0174] The node damage estimation module is used to determine the Park-Ang damage index parameters, calculate the Park-Ang damage index in combination with the target node acceleration, and determine the state of the target node based on the Park-Ang damage index.

[0175] This invention also provides a computer device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, it implements the steps in any of the above method embodiments.

[0176] This invention also provides a storage medium storing a computer program, wherein the computer program is configured to execute the steps in any of the above method embodiments when it is run.

[0177] Specific examples in this embodiment can be found in the examples described in the above embodiments and optional implementations, and those already described will not be repeated here.

[0178] Optionally, in this embodiment, the storage medium may include, but is not limited to, various media capable of storing computer programs, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0179] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0180] In the above embodiments of the present invention, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0181] The steps in the methods of the above embodiments of the present invention can be adjusted, combined, or deleted according to actual needs. The technical features can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the embodiments are described. However, as long as the combinations of these technical features do not contradict each other, they should all be considered within the scope of the present invention.

[0182] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A two-step method for predicting post-earthquake joint damage in prefabricated steel structure buildings, characterized in that, include: By simulating different magnitudes and seismic waves through shaking table tests, structural response data of prefabricated steel structure buildings under seismic loading is obtained. The response data includes magnitude, timestamp, and ground acceleration. Based on the magnitude, timestamp, and ground acceleration, the RIME-LSTM-self-attention algorithm is used to predict the acceleration at the bottom of the building. The RIME-LSTM-self-attention algorithm is a hybrid model that combines optimization algorithms, deep learning components, and machine learning. The RIME algorithm optimizes hyperparameters, the LSTM layer captures time-series dependencies, and the self-attention mechanism enhances key features and provides interpretability of the model. Based on the magnitude, timestamp, ground acceleration, and the acceleration at the base of the building, the ELM algorithm is used to predict the acceleration of the target node. Determine the Park-Ang damage index parameters and calculate the Park-Ang damage index in combination with the target node acceleration. The Park-Ang damage index parameters include yield strength, shear strength, cross-sectional area, story height, and node mass. The Park-Ang damage index is used to quantify the damage degree of the target node. The state of the target node is determined based on the Park-Ang damage index. The determination of Park-Ang damage index parameters, and the calculation of the Park-Ang damage index in conjunction with the target node acceleration, includes: Calculate the plastic bending moment based on the yield strength and plastic section modulus of the steel. : In the formula, and These are the yield strength and section modulus of steel, respectively. Based on the plastic bending moment And the ultimate lateral load of the lateral force resisting members calculated based on the floor height. : In the formula, This refers to the total floor height of the prefabricated steel structure building. According to the ultimate lateral load of the lateral force resisting member And the deformation limit for static loading in the stiffness calculation of components. : In the formula, For component stiffness; The plastic deformation energy is quantified by integrating the product of the target node's acceleration and velocity. In the formula, This represents the dissipation of hysteresis energy, i.e., the energy generated during plastic deformation. For node quality, here we take the allocation quality of the target node; and They represent The acceleration and velocity at each time point, where, ; According to the plastic deformation energy Deformation limits under static load Calculate the Park-Ang damage index of the target node: In the formula, Indicates the Park-Ang damage index of the target node; Indicates the action of an earthquake. The maximum deformation value before the specified time; Indicates the shear strength of the node. , The cross-sectional area; This represents the energy weighting coefficient.

2. The method for predicting post-earthquake node damage in prefabricated steel structure buildings based on a two-step approach according to claim 1, characterized in that, Based on the magnitude, timestamp, and ground acceleration, before predicting the acceleration at the base of a building using the RIME-LSTM-self-attention algorithm, the process also includes constructing an LSTM-self-attention prediction network model, including: Through shaking table tests, structural response data and corresponding building base acceleration of prefabricated steel structure buildings under different magnitudes and seismic waves were obtained, and a first dataset containing magnitude, timestamp, ground acceleration and building base acceleration was constructed. The first dataset is input into a bidirectional LSTM network, and forward and backward propagation layers are used to generate forward and backward hidden state sequences, respectively, which are then concatenated to form a bidirectional fused feature tensor. The bidirectional fused feature tensor is input into the attention module, and the input features are mapped into query vectors, key vectors, and value vectors through a linear transformation layer, forming a multi-head attention computation space. The similarity matrix between the query vector and the key vector is calculated using the scaling dot product algorithm, and then normalized by Softmax to generate an attention weight matrix. The weight matrix represents the global dependency strength between features at different time steps. The attention weight matrix and the value vector are weighted and summed to generate a globally enhanced feature representation. A bottleneck compression layer is applied to the global enhanced feature representation, a 1×1 convolution kernel is used to reduce the channel dimension to 1 / 4 of the original input, and redundant information is suppressed by layer normalization; The compressed features are input into a fully connected regression layer, which outputs the predicted acceleration value at the bottom of the building.

3. The method for predicting post-earthquake node damage in prefabricated steel structure buildings based on a two-step method according to claim 2, characterized in that, After constructing the LSTM-self-attention prediction network model, the following is also included: By utilizing the adaptive LSTM hyperparameter configuration of the Frost Optimization Algorithm, the number of LSTM neurons and the initial learning rate in the LSTM-self-attention prediction network model are jointly optimized through a multi-objective fitness function to construct the RIME-LSTM-self-attention prediction model.

4. The method for predicting post-earthquake node damage in prefabricated steel structure buildings based on a two-step approach according to claim 1, characterized in that, Before predicting the target node acceleration using the ELM algorithm based on magnitude, timestamp, ground acceleration, and the acceleration at the base of the building, the process also includes constructing an ELM prediction network, including: Construct a second dataset containing magnitude, timestamp, ground acceleration, building base acceleration, and the corresponding acceleration of each monitoring node; A single hidden layer feedforward neural network is independently constructed for each monitoring node, comprising an input layer, a hidden layer, and an output layer. The input layer receives the raw feature data collected by the sensor, and its dimension is consistent with the number of physical parameters. The hidden layer uses a non-linear activation function to perform a non-linear mapping on the input signal, and the number of neurons is 2 to 5 times the dimension of the input layer. The output layer is a linear combination of the hidden layer outputs, used to generate the target prediction value. Initial parameters are randomly generated from a uniform or normal distribution. These initial parameters are used to lock the connection weight matrix and bias vector from the input layer to the hidden layer, and they cannot be updated during training. The ELM prediction network is independently deployed to each monitoring node, receiving real-time data streams and synchronously outputting the target node's acceleration prediction values.

5. The method for predicting post-earthquake node damage in prefabricated steel structure buildings based on a two-step method according to claim 4, characterized in that, After constructing the ELM prediction network, the method further includes: By utilizing the adaptive ELM hyperparameter configuration of the frost optimization algorithm, the number of ELM neurons in the ELM prediction network is jointly optimized through a multi-objective fitness function.

6. The method for predicting post-earthquake node damage in prefabricated steel structure buildings based on a two-step method according to claim 1, characterized in that, Determining the target node status based on the Park-Ang damage index includes: If the Park-Ang damage index of the target node is greater than 1, the target node is considered to be damaged.

7. A two-step method for predicting post-earthquake joint damage in prefabricated steel structures, based on the two-step method for predicting post-earthquake joint damage in prefabricated steel structures according to any one of claims 1 to 6, characterized in that, include: The data acquisition module is used to simulate different magnitudes and seismic wave effects through shaking table tests to obtain structural response data of prefabricated steel structure buildings under seismic loading. The building base acceleration prediction module is used to predict the building base acceleration based on the magnitude, timestamp, and ground acceleration using the RIME-LSTM-self-attention algorithm. The target node acceleration prediction module is used to predict the target node acceleration based on magnitude, timestamp, ground acceleration, and the acceleration at the bottom of the building, using the ELM algorithm. The node damage estimation module is used to determine the Park-Ang damage index parameters, calculate the Park-Ang damage index in combination with the target node acceleration, and determine the state of the target node based on the Park-Ang damage index.

8. A computer device, characterized in that, The computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When executed by the processor, the computer program implements the steps of the two-step method for predicting post-earthquake node damage in prefabricated steel structure buildings as described in any one of claims 1 to 6.

9. A storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the steps of the two-step method for predicting post-earthquake node damage in prefabricated steel structure buildings as described in any one of claims 1 to 6.

Images