A Method for Predicting Metal Fatigue of Long-Term Loads in Steel Structure Buildings Based on Machine Learning

Abstract

The present invention proposes a method for predicting metal fatigue of long-term loads in steel structure buildings based on machine learning, which relates to the field of data prediction. The present invention proposes a DWN prediction model, which is applied to the scenario of predicting metal fatigue of long-term loads in steel structure buildings, including a frequency domain feature extraction module, a time domain feature extraction module, and a learning and prediction module. Specifically, the frequency domain feature extraction module can extract frequency domain information and model frequency domain features, the time domain feature extraction module can extract time domain information and model time domain features, and the learning and prediction module is used to combine the frequency domain features and the time domain features and convert them into available prediction results for metal fatigue of long-term loads in steel structure buildings through contrastive learning. The modules cooperate with each other to achieve accurate prediction of metal fatigue of long-term loads in steel structure buildings.

Description

Technical Field

[0001] The present invention belongs to the field of data prediction, and particularly relates to a method for predicting long-term load metal fatigue of steel structure buildings based on machine learning. Background Art

[0002] Steel structure buildings are widely used in modern urban construction and are subjected to long-term loads, including self-weight, service loads, wind forces, and external forces such as earthquakes. The influence of long-term loads on steel structures accumulates gradually. As time goes by, the microstructure of steel changes, resulting in the occurrence of metal fatigue. The fatigue damage of steel structures is manifested as cracks gradually appearing in the material after repeated loads, and ultimately leading to structural failure. Especially in complex loads and non-uniform environments, the prediction of long-term load metal fatigue of steel structure buildings is an important issue in structural health monitoring.

[0003] In recent years, machine learning has made remarkable progress in the field of engineering structural health monitoring. In the aspect of steel structure fatigue prediction, machine learning methods can automatically learn and extract potential laws in a large amount of data. The process of steel structure fatigue damage is highly nonlinear and complex. Traditional analysis methods often cannot comprehensively consider the influence of various load and environmental factors. However, machine learning algorithms can, through the analysis of a large amount of historical data, identify key patterns in the steel structure fatigue process, be able to process and analyze data in real time in a changing actual environment, identify early signs of fatigue damage, and provide predictions for future fatigue development trends.

[0004] Currently, the problem faced by steel structure fatigue prediction is how to construct a model that can not only handle complex nonlinear relationships but also have high prediction accuracy in the presence of noise, sparse data, and variable working conditions. Facing these challenges, through multi-scale feature extraction, fatigue change patterns at different time scales can be effectively identified. By combining time-domain and frequency-domain features, multi-scale information can be effectively captured when processing long time series data, enhancing the robustness of the model to noise and sparse data. Summary of the Invention

[0005] The present invention provides a method for predicting metal fatigue of long-term loads in steel structure buildings based on machine learning. Aiming at the long-term load time series data of steel structure buildings with strong noise, non-linearity and complex dependency relationships, a DWN (Dynamic WaveNet) model based on machine learning is proposed. The model consists of a frequency domain feature extraction module, a time domain feature extraction module and a learning and prediction module. The frequency domain feature extraction module captures frequency domain features through wavelet transform and introduces a variational local adaptive wavelet transform method to extract the most representative frequency components. The time domain feature extraction module extracts effective time domain features from the metal fatigue data of long-term loads in steel structure buildings through a designed learnable convolution kernel. The learning and prediction module combines the above features, designs a learnable contrast loss weight to train the model and realizes the prediction of metal fatigue of long-term loads.

[0006] The technical solution adopted by the present invention to achieve the above object specifically includes the following steps:

[0007] S1. Collect metal fatigue data of long-term loads in steel structure buildings, including physical parameters and usage environment parameters, and preprocess the collected data;

[0008] S2. Standardize the preprocessed metal fatigue data using the mean normalization method and map it through dilated convolution, and divide the data into a training set and a test set;

[0009] S3. Construct a frequency feature extraction module, introduce a variational local adaptive wavelet transform, and construct frequency domain features. The specific steps are as follows:

[0010] S31. Dynamically calculate and select an adaptive scale for the time series features of the metal fatigue data of steel structure buildings through variational optimization;

[0011] S32. Calculate the local frequency domain energy of the adaptive scale and optimize the adaptive scale through path integration;

[0012] S33. Calculate and optimize the contributions of different frequency components using the inverse graph Laplacian weights, and reconstruct the wavelet features of the weighted metal fatigue data of steel structure buildings back to the time domain signal using the inverse wavelet transform;

[0013] S4. Construct a time domain feature extraction module, design a learnable convolution kernel, and construct time domain features through residual connection. The specific steps are as follows:

[0014] S41. Design a learnable convolution kernel Perform a convolution operation on the time series features of the metal fatigue data of steel structure buildings to obtain the output after the convolution operation;

[0015] S42. Apply an activation function and perform layer normalization on the convolution result;

[0016] S43. Perform a residual connection between the output after layer normalization and the input, concatenate the features of all time steps to obtain the final time-domain feature representation;

[0017] S5. Construct a learning prediction module, fuse the metal fatigue data features of the steel structure building in the time domain and frequency domain, design learnable weights , adjust the contrast loss weight, train the model through contrastive learning, and input the processed metal fatigue data into the model to obtain the metal fatigue prediction result.

[0018] Preferably, in S1, collect the metal fatigue data of the long-term load of the steel structure building, including the physical parameter data of the steel structure material and the usage environment parameter data. Among them, the physical parameter data includes the tensile strength data, compressive strength data, yield strength data, density data, elastic modulus data, fatigue life data, and hardness data of the steel structure material, and the installation environment parameter data includes long-term load data, environmental temperature data, environmental humidity data, environmental vibration data, and light condition data, and calculate the mean of the original sequence and the standard deviation . The specific formulas are:

[0019] ;

[0020] ;

[0021] In the formula, is the total number of time points, is the metal fatigue data of the steel structure building at time

[0022] Generate enhanced metal fatigue data of the steel structure building through scaling and offset operations . The specific formula is:

[0023] ;

[0024] In the formula, is the scaling factor, is the offset factor, is the hyperparameter controlling the scaling, is the hyperparameter controlling the offset.

[0025] Preferably, in S2, use the mean normalization method to standardize the preprocessed metal fatigue data of the steel structure building. The specific formula is:

[0026] ;

[0027] In the formula, is the metal fatigue data of the steel structure building after data augmentation, is the mean of the sequence, is the standard deviation of the sequence;

[0028] Use the dilated convolution module to extract the temporal features of the metal fatigue data of the steel structure building as the input signal , and the specific formula is:

[0029] ;

[0030] In the formula, is the linear layer, is the dilated convolution, is the projection layer.

[0031] Preferably, in S3 and S31, the input signal dynamically calculates and selects the adaptive scale through variational optimization , and the specific formula is:

[0032] ;

[0033] In the formula, is the wavelet scale, is the time offset, is the smoothing factor, is the gradient with respect to the wavelet scale;

[0034] is the wavelet transform, and the specific formula is:

[0035] ;

[0036] In the formula, is the Morlet basis function with respect to the wavelet scale , and the specific formula is:

[0037] ;

[0038] In the formula, is the wavelet scale, is the time offset.

[0039] Preferably, introduce the variational optimization algorithm to dynamically select the optimal adaptive scale, and automatically optimize the adaptive scale according to the change of the temporal features of the metal fatigue data of the steel structure building to improve the flexibility and accuracy of feature extraction, and accurately extract the frequency domain features of the metal fatigue data of the steel structure building at different scales.

[0040] Preferably, in S3 and S32, for the obtained adaptive scale Calculate the local frequency domain energy and optimize the adaptive scale through path integral. The specific formula is as follows:

[0041] ;

[0042] In the formula, is the optimal path obtained through variational optimization. The specific formula is as follows:

[0043] ;

[0044] In the formula, is the local signal energy. The specific formula is as follows:

[0045] ,

[0046] In the formula, is the Morlet basis function with respect to the adaptive scale .

[0047] Preferably, by combining the path integral optimization technique, the weights of different frequency components of the metal fatigue data of steel structure buildings can be dynamically adjusted according to the input local features. According to the changes in the input data, the optimal path is selected to calculate the energy of each frequency component, thereby improving the accuracy and efficiency of frequency domain feature extraction. Combining with the multi-scale characteristics of wavelet transform, the path integral further enhances the model's ability to capture detailed features, enabling effective frequency domain features to be extracted under scale conditions.

[0048] Preferably, in S3 and S33, the inverse graph Laplacian weight is used to calculate and optimize the contributions of different frequency components. The specific formula is as follows:

[0049] ;

[0050] In the formula, is the optimized adaptive scale, is the time offset, is the time series feature of the input metal fatigue data of steel structure buildings, is the Morlet basis function, is the smoothing factor for controlling the neighborhood smoothness, is the graph Laplacian matrix. The specific formula is as follows:

[0051] ;

[0052] In the formula, is the degree matrix, is the adjacency matrix, indicating the similarity between wavelet scales;

[0053] Degree matrix The specific formula is as follows:

[0054] ;

[0055] In the formula, is the number of nodes, is the node degree, and the specific formula is:

[0056] ;

[0057] Adjacency matrix The specific formula is:

[0058] ;

[0059] Among them, is the number of nodes. The in the adjacency matrix row and column indicate whether there is an edge between node and node . The specific formula is:

[0060] ;

[0061] Use the inverse wavelet transform to convert the wavelet features of the weighted steel structure building metal fatigue data into a time-domain signal. The specific formula is:

[0062] ;

[0063] In the formula, is the time-domain signal after the conversion of the steel structure building metal fatigue data. The specific formula is:

[0064] ;

[0065] In the formula, is the optimized adaptive scale, is the time offset.

[0066] Preferably, the graph Laplacian matrix and the backpropagation method are used to eliminate the noise in the frequency-domain data of the steel structure building metal fatigue data, and optimize the accuracy of the frequency-domain signal during reconstruction, and perform weighted optimization on different frequency components to more accurately capture the frequency-domain information of the signal.

[0067] Preferably, in S4 and S41, for each time step in the time-series features of the steel structure building metal fatigue data , a learnable convolutional kernel is designed for the input of the time step Perform a convolution operation, and the specific formula is:

[0068] ;

[0069] In the formula, is the output after the convolution operation, is a 1D convolution operation, is the designed learnable convolution kernel;

[0070] The specific formula for the learnable convolution kernel is:

[0071] ;

[0072] In the formula, is the local curvature, representing the change speed of the time series characteristics of the metal fatigue data of the steel structure building at the time step , is the learnable parameter, is the maximum convolution kernel, is the minimum convolution kernel, is the Sigmoid function, is the rounding function;

[0073] The specific formula for the local curvature is:

[0074] ;

[0075] In the formula, is the second derivative;

[0076] The learnable parameter is obtained through training optimization, and the specific formula is:

[0077] ;

[0078] ;

[0079] In the formula, is the cross-entropy loss function, is the convolution kernel weight, is the learning rate;

[0080] The specific formula for updating the convolution kernel weight is:

[0081] ;

[0082] In the formula, is the cross-entropy loss function, is the learning rate;

[0083] The specific formula for the cross-entropy loss function is:

[0084] ;

[0085] In the formula, is the total sample number of the metal fatigue data of the steel structure building, is the true value, is the predicted value.

[0086] Preferably, by designing a learnable convolution kernel, the model can automatically adjust the convolution kernel according to the characteristics of the metal fatigue data of the steel structure building, enhancing the detailed learning ability of the temporal characteristics of the metal fatigue data of the steel structure building, enabling the model to flexibly adjust the feature extraction process according to the characteristics of each time step, thereby optimizing the expression of the temporal domain features and further improving the recognition ability of the detail changes in the steel structure fatigue prediction, ensuring efficient feature extraction and prediction results.

[0087] Preferably, for different convolution results in S4, S42 , apply the activation function to obtain the processed data , and the specific formula is:

[0088] ;

[0089] In the formula, is the time step after dynamic convolution, is the activation function, and the specific formula is:

[0090] ;

[0091] In the formula, and are predefined hyperparameters, ;

[0092] Perform layer normalization on , and the specific formula is:

[0093] ;

[0094] In the formula, is the layer normalization operation, and the specific formula is:

[0095] ;

[0096] In the formula, is the mean value of each time step , is the standard deviation of each time step , is a small constant set to prevent division by zero errors.

[0097] Preferably, the model applies an activation function to each convolution result for non-linear transformation, and then performs layer normalization on the output result, ensuring that the non-linear features of the data after passing through the activation function can be effectively normalized, maintaining the stability during the training process. Layer normalization reduces internal covariate shift, improves the stability and convergence speed of training, further enhances the model's ability to process features at different time steps, and thus improves the prediction accuracy.

[0098] Preferably, in S4 and S43, the output after layer normalization is connected with the input signal for residual connection, and the specific formula is:

[0099] ;

[0100] In the formula, is the feature representation after adding the residual connection;

[0101] Use feature concatenation to combine the features of all time steps to obtain the time-domain feature representation of the final steel structure building metal fatigue data. The specific formula is:

[0102] ;

[0103] In the formula, is the concatenation operation, is the time step is the feature representation after residual connection.

[0104] Preferably, the output after layer normalization is connected with the original steel structure building metal fatigue data for residual connection, and then the features of all time steps are combined through feature concatenation to obtain the final time-domain feature representation. The introduction of residual connection helps to avoid the problem of gradient disappearance, and at the same time improves the model's learning ability for complex time-series data. By concatenating the features of multiple time steps together, the model can more comprehensively understand and capture the changing trends in the time series, enhancing the prediction ability for the metal fatigue of long-term loads on steel structures, and further improving the prediction accuracy and stability of the model.

[0105] Preferably, in S5, the features of the steel structure building metal fatigue data extracted from the time domain and the frequency domain are fused to form a unified feature representation , and the specific formula is

[0106] ;

[0107] In the formula, is the concatenation operation, is the linear layer, is the time-domain feature representation, is the frequency-domain feature representation;

[0108] Design learnable weights , adjust the weights of the time-domain contrast loss and the frequency-domain contrast loss, use contrastive learning to learn the contrastive features in the time domain and the frequency domain, and minimize the total loss to train the model. The specific formula is:

[0109] ;

[0110] In the formula, is the contrast loss in the time domain, is the contrast loss in the frequency domain;

[0111] The specific formula of the learnable weight is:

[0112] ;

[0113] In the formula, is the change rate of the time-domain loss, is the change rate of the frequency-domain loss, is the similarity measure of the time-domain and frequency-domain features, controlling the weights of the time domain and the frequency domain, is the learning rate;

[0114] The change rate of the time-domain loss The specific formula is:

[0115] ;

[0116] The change rate of the frequency-domain loss The specific formula is:

[0117] ;

[0118] The similarity measure of the time-domain and frequency-domain features The specific formula is:

[0119] ;

[0120] In the formula, is the L2 norm of the time-domain feature, is the L2 norm of the frequency-domain feature;

[0121] The contrast loss in the time domain The specific formula is:

[0122] ;

[0123] In the formula, is the feature obtained through the residual connection, is the final time-domain feature representation;

[0124] Frequency domain contrast loss The specific formula is as follows:

[0125] ;

[0126] In the formula, is the dynamic weight, is the amplitude loss, and the specific formula is:

[0127] ;

[0128] In the formula, is the total number of samples of the metal fatigue data of the steel structure building, is the predicted frequency domain amplitude, is the true frequency domain amplitude;

[0129] is the phase loss, and the specific formula is:

[0130] ;

[0131] In the formula, is the predicted frequency domain phase, is the true frequency domain phase.

[0132] Preferably, by fusing the characteristics of the metal fatigue data of the steel structure building extracted from the time domain and the frequency domain, a unified feature representation is formed, and the feature learning in the time domain and the frequency domain is further optimized through contrastive learning. A learnable loss weight is designed. By minimizing the total loss, the model can effectively perform contrastive learning on the time domain and frequency domain features, thereby improving the quality of the feature representation. Contrastive learning not only improves the discrimination ability of the features but also enhances the efficiency of the model in processing multi-source data, enabling the prediction model to make more accurate predictions in complex and variable working environments and improving the overall prediction effect.

[0133] In summary, due to the adoption of this technical solution, the beneficial effects of the present invention are as follows: The present invention proposes a DWN prediction model, which is applied to the metal fatigue prediction scenario of long-term loads of steel structure buildings, including a frequency domain feature extraction module, a time domain feature extraction module, and a learning and prediction module. Specifically, the frequency domain feature extraction module can extract frequency domain information and model frequency domain features, the time domain feature extraction module can extract time domain information and model time domain features, and the learning and prediction module is used to combine the frequency domain features and the time domain features and convert them into available metal fatigue prediction results for long-term loads of steel structure buildings through contrastive learning. The modules cooperate with each other to achieve accurate prediction of the metal fatigue of long-term loads of steel structure buildings. BRIEF DESCRIPTION OF THE DRAWINGS

[0134] Figure 1It is a flowchart of the method for predicting metal fatigue under long-term load of steel structure buildings.

[0135] Figure 2 It is a structural diagram of the DWN prediction model.

[0136] Figure 3 It is a structural diagram of the frequency-domain feature extraction module.

[0137] Figure 4 It is a structural diagram of the time-domain feature extraction module.

[0138] Figure 5 It is a fitting effect diagram of the DWN prediction model for predicting metal fatigue under long-term load of steel structure buildings. Specific implementation manner

[0139] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative work belong to the scope of protection of the present invention.

[0140] Please refer to Figures 1 - 5 , the present invention provides a technical solution: a method for predicting metal fatigue under long-term load of steel structure buildings based on machine learning. By constructing a frequency-domain feature extraction module, frequency-domain information can be extracted and frequency-domain features can be modeled. The time-domain feature extraction module can extract time-domain information and model time-domain features. The learning and prediction module is used to combine the frequency-domain features and time-domain features and convert them into available prediction results of metal fatigue under long-term load of steel structure buildings through contrast learning. The specific steps are as Figure 1 shown.

[0141] Construct a DWN prediction model, and its structure is as Figure 2 shown, and the specific steps are as follows.

[0142] S1. Collect data on metal fatigue under long-term load of steel structure buildings, including physical parameters and usage environment parameters, and preprocess the collected data.

[0143] Furthermore, collect data on metal fatigue under long-term load of steel structure buildings, including physical parameter data and usage environment parameter data of steel structure materials. Among them, the physical parameter data includes tensile strength data, compressive strength data, yield strength data, density data, elastic modulus data, fatigue life data, and hardness data of steel structure materials. The installation environment parameter data includes long-term load data, environmental temperature data, environmental humidity data, environmental vibration data, and lighting condition data. Collect data on metal fatigue under long-term load of steel structure buildings for a total of 140 days, and calculate the mean of the original sequence of and standard deviation , and the specific formula is:

[0144] ;

[0145] ;

[0146] In the formula, is the total number of time points, which is 140, is the metal fatigue data of the steel structure building at time , and the specific formula for generating the enhanced time series through scaling and offset operations is:

[0147] ;

[0148] In the formula, is the scaling factor, is the offset factor, is the hyperparameter controlling the scaling , and the initial value is set to 0.01, the hyperparameter controlling the offset, and the initial value is set to 0.01.

[0149] S2. Standardize the preprocessed metal fatigue data of the steel structure building using the mean normalization method, and map it through dilated convolution. Divide the data into a training set and a test set in a ratio of 7:3.

[0150] Furthermore, standardize the preprocessed metal fatigue data of the steel structure building using the mean normalization method. The specific formula is:

[0151] ;

[0152] In the formula, is the metal fatigue data of the steel structure building after data augmentation, is the mean of the sequence, is the standard deviation of the sequence;

[0153] Extract the time series features of the metal fatigue data of the steel structure building using the dilated convolution module as the input signal , and the specific formula is:

[0154] ;

[0155] In the formula, is the linear layer, is the dilated convolution, is the projection layer. Then divide the steel structure building data into a training set and a test set according to a ratio.

[0156] S31. Dynamically calculate and select an adaptive scale for the input signal through variational optimization through variational optimization dynamic calculation to select an adaptive scale .

[0157] Furthermore, in S31, for the input signal dynamically calculate and select an adaptive scale through variational optimization , and the specific formula is:[[]]

[0158] ;

[0159] In the formula, is the wavelet scale, with an initial value set to 1, is the time offset, , is the time step, with a value of 1, is the smoothing factor controlling the scale change, with an initial value set to 0.001, is the time series feature of the metal fatigue data of the steel structure building extracted the gradient with respect to the scale;

[0160] The wavelet transform, the specific formula is:[[]]

[0161] ;

[0162] In the formula, is the Morlet basis function with respect to the wavelet scale , enabling the input data to be decomposed at different wavelet scales and time offsets , and the specific formula is:[[]]

[0163] ;

[0164] In the formula, is the wavelet scale, a smaller focuses on high-frequency features, and a larger focuses on low-frequency features, is the time offset.[[]]

[0165] S32. Calculate the local frequency domain energy for the obtained adaptive scale and optimize the adaptive scale through path integral . .

[0166] Furthermore, in S32, for the obtained adaptive scale calculate the local frequency domain energy and optimize the adaptive scale through path integral. The specific formula is:[[]]

[0167] ;

[0168] Wherein, is the optimal path obtained by variational optimization, and the specific formula is:

[0169] ;

[0170] Wherein, is the local signal energy, and the specific formula is:

[0171] ,

[0172] Wherein, is the Morlet basis function with respect to the adaptive scale .

[0173] S33. Calculate and optimize the contributions of different frequency components using the inverse graph Laplacian weight , and use the inverse wavelet transform to reconstruct the weighted wavelet features back to the time-domain signal .

[0174] Furthermore, calculating and optimizing the contributions of different frequency components using the inverse graph Laplacian weight in S33, the specific formula is:

[0175] ;

[0176] Wherein, is the optimized adaptive scale, is the time offset, is the input signal, is the Morlet basis function, is the smoothing factor for controlling the neighborhood smoothness, , the initial value is set to 0.05, is the graph Laplacian matrix, and the specific formula is:

[0177] ;

[0178] Wherein, is the degree matrix, is the adjacency matrix, representing the similarity between wavelet scales;

[0179] Degree matrix The specific formula is:

[0180] ;

[0181] Wherein, is the number of nodes, is the node The degree, and the specific formula is:

[0182] ;

[0183] Adjacency matrix The specific formula is:

[0184] ;

[0185] Among them, is the number of nodes. The in the adjacency matrix row and the column indicate whether there is an edge between node and node . The specific formula is:

[0186] ;

[0187] Use the inverse wavelet transform to convert the wavelet features of the weighted steel structure building metal fatigue data into a time-domain signal. The specific formula is:

[0188] ;

[0189] In the formula, is the time-domain signal of the steel structure building metal fatigue data after conversion. The specific formula is:

[0190] ;

[0191] In the formula, is the optimized adaptive scale, is the time offset, is the Morlet basis function.

[0192] S41. For each time step in the steel structure building metal fatigue data , design a learnable convolution kernel to perform a convolution operation on the input signal at time step to obtain the output after the convolution operation.

[0193] Furthermore, in S41, for each time step in the time series features of the steel structure building metal fatigue data , use the learnable convolution kernel to perform a convolution operation on the input at time step . The specific formula is:

[0194] ;

[0195] In the formula, is the output after the convolution operation, is a one-dimensional convolution operation, is a learnable convolution kernel;

[0196] The specific formula for the learnable convolution kernel is:

[0197] ;

[0198] In the formula, is the local curvature, representing the change speed of the time series characteristics of the metal fatigue data of the steel structure building at the time step ; is a learnable parameter, is the maximum convolution kernel, set to 9, is the minimum convolution kernel, set to 3, is the Sigmoid function, is the rounding function;

[0199] The specific formula for the local curvature is:

[0200] ;

[0201] In the formula, is the second derivative;

[0202] The learnable parameter is obtained through training optimization, and the specific formula is:

[0203] ;

[0204] ;

[0205] In the formula, is the cross-entropy loss function, is the convolution kernel weight, is the learning rate, set to 0.01;

[0206] The specific formula for updating the convolution kernel weight is:

[0207] ;

[0208] In the formula, is the cross-entropy loss function, is the learning rate, set to 0.01;

[0209] The specific formula for the cross-entropy loss function is:

[0210] ;

[0211] In the formula, is the total number of samples of the metal fatigue data of the steel structure building, which is set to 140, is the true value, is the predicted value.

[0212] S42. For each convolution result , apply activation function to obtain the processed data , and perform layer normalization on to obtain the normalized data .

[0213] Furthermore, for different convolution results in S42, apply activation function to obtain the processed data , and the specific formula is:

[0214] ;

[0215] In the formula, is the time step after the output of the dynamic convolution, is the activation function, and the specific formula is:

[0216] ;

[0217] In the formula, and are predefined hyperparameters, , with the initial value set to 0.1, , with the initial value set to 1.6733;

[0218] Perform layer normalization on , and the specific formula is:

[0219] ;

[0220] In the formula, is the layer normalization operation, and the specific formula is:

[0221] ;

[0222] In the formula, is the mean value of each time step , is the standard deviation of each time step , is a small constant set to prevent division by zero errors, .

[0223] S43. The output after layer normalization With the time series characteristics of the metal fatigue data of steel structure buildings Perform residual connection, and use feature splicing to merge the features of all time steps to obtain the final time domain feature representation .

[0224] Furthermore, in S43, the output after layer normalization With the time series characteristics of the metal fatigue data of steel structure buildings Perform residual connection, and the specific formula is:

[0225] ;

[0226] In the formula, Is the feature representation after adding residual connection;

[0227] Use feature splicing to merge the features of all time steps to obtain the final time domain feature representation, and the specific formula is:

[0228] ;

[0229] In the formula, Is the splicing operation, Is the time step Is the feature representation after performing residual connection.

[0230] S5. Construct a learning prediction module to fuse the features of the metal fatigue data of steel structure buildings extracted from the time domain and frequency domain to form a unified feature representation , Design a learnable weight , Adjust the weights of the time domain contrast loss and the frequency domain contrast loss, and train the model through contrast learning and input the preprocessed data into the model for prediction.

[0231] Furthermore, in S5, the features extracted from the time domain and frequency domain are fused to form a unified feature representation , And the specific formula is

[0232] ;

[0233] In the formula, Is the splicing operation, Is the linear layer, Is the time domain feature representation, Is the frequency domain feature representation;

[0234] Design a learnable weight , Adjust the weights of the time domain contrast loss and the frequency domain contrast loss, use contrast learning to learn the contrast features of the time domain and frequency domain, and train the model by minimizing the total loss The specific formula is:

[0235] ;

[0236] wherein, is the contrast loss in the time domain, is the contrast loss in the frequency domain;

[0237] learnable weight has the specific formula as:

[0238] ;

[0239] wherein, is the change rate of the time domain loss, is the change rate of the frequency domain loss, is the similarity measure of the time domain and frequency domain features, controlling the weights of the time domain and frequency domain, is the learning rate, set to 0.01;

[0240] change rate of the time domain loss has the specific formula as:

[0241] ;

[0242] change rate of the frequency domain loss has the specific formula as:

[0243] ;

[0244] similarity measure of the time domain and frequency domain features has the specific formula as:

[0245] ;

[0246] wherein, is the L2 norm of the time domain features, is the L2 norm of the frequency domain features;

[0247] time domain contrast loss has the specific formula as:

[0248] ;

[0249] wherein, is the feature obtained through the residual connection, is the final time domain feature representation;

[0250] frequency domain contrast loss has the specific formula as:

[0251] ;

[0252] wherein, is the dynamic weight, is the amplitude loss, and the specific formula is:

[0253] ;

[0254] In the formula, is the total sample number of the metal fatigue data of the steel structure building, and the value is 140, is the predicted frequency domain amplitude, is the true frequency domain amplitude;

[0255] is the phase loss, and the specific formula is:

[0256] ;

[0257] In the formula, is the predicted frequency domain phase, is the true frequency domain phase.

[0258] Furthermore, the DWN prediction model is written in Python language. The experiment runs on the Windows operating system. Pytorch is selected as the framework in the CUDA11.27 environment, and it is trained on the GeForce RTX3090. The optimizer is selected, the initial learning rate is set to 0.001, the training batch is set to 64, and the data set is the metal fatigue data of the long-term load of the steel structure building for 140 days. After preprocessing, it is input into the DWN prediction model.

[0259] Furthermore, the fitting effect of the DWN model for predicting the metal fatigue of the long-term load of the steel structure building is as Figure 5 shown. The abscissa in the figure is the load cycle (days), and the ordinate is the metal fatigue degree (%). The gray dotted line and dots are the real data, and the black solid line and squares are the predicted data. It can be seen from the figure that the overall trend of the predicted curve and the real curve is highly consistent. Especially in the early stage of the load cycle, the two curves basically coincide, indicating that the model can accurately capture the change characteristics of the short-term load metal fatigue and accurately model the initial fatigue condition. As the load cycle increases, the metal fatigue degree gradually rises, which indicates that the steel structure gradually bears a greater fatigue load. Generally speaking, the overall prediction effect of the model is good, it can effectively capture the long-term load metal fatigue characteristics of the steel structure building, and has good long-term prediction ability.

Claims

1. A long-term load metal fatigue prediction method for steel structure buildings based on machine learning, characterized in that, It includes the following steps: S1. Collect the metal fatigue data of long-term loads of steel structure buildings, including physical parameters and usage environment parameters, and preprocess the collected data; S2. Standardize the preprocessed metal fatigue data using the mean normalization method, and map it through dilated convolution, and divide the data into a training set and a test set; S3. Construct a frequency feature extraction module, introduce variational local adaptive wavelet transform, and construct frequency domain features. The specific steps are as follows: S31. Dynamically calculate and select the adaptive scale for the input sequence through variational optimization. The specific steps are as follows: Dynamically calculate and select an adaptive scale s for the input signal r through variational optimization * , and the specific formula is as follows: where s is the wavelet scale, τ is the time offset, λ is the smoothing factor, is the gradient of the input signal r with respect to the wavelet scale; W(r, s, τ) is the wavelet transform, and the specific formula is: W(r, s, T) = ∫r(t)ψ s，τ (t)dt; where ψ s，τ (t) is the Morlet basis function with respect to the wavelet scale s, and the specific formula is: In the formula, s is the wavelet scale, and τ is the time shift; S32. Calculate the local frequency domain energy of the adaptive scale, and optimize the adaptive scale through path integral. The specific steps are as follows: For the obtained adaptive scale s * Calculate the local frequency domain energy, and optimize the adaptive scale through path integration. The specific formula is as follows: where γ(s * , τ) is the optimal path obtained by variational optimization, and the specific formula is: where L(X, s * , τ) is the local signal energy, and the specific formula is as follows: wherein, is the Morlet basis function with respect to the adaptive scale s * ; S33. Calculate and optimize the contributions of different frequency components using the inverse graph Laplacian weights, and reconstruct the weighted wavelet features back to the time domain signal using the inverse wavelet transform; S4. Construct a time domain feature extraction module, design a learnable convolution kernel, and construct time domain features through residual connection. The specific steps are as follows: S41. Design a learnable convolution kernel K t Perform a convolution operation on the input signal; S42. Apply the SeLU activation function to the convolution result and perform layer normalization; S43. Perform a residual connection between the output after layer normalization and the input signal, splice the features of all time steps, and obtain the final time domain feature representation; S5. Construct a learning prediction module, design a learnable weight γ to adjust the contrast loss weight, train the model through contrastive learning, and input the preprocessed data into the model to obtain the metal fatigue prediction result.

2. The long-term load metal fatigue prediction method for steel structure buildings based on machine learning according to claim 1, characterized in that In step S33, the inverse graph Laplacian weights are used to calculate and optimize the contributions of different frequency components. The specific formula is: W f (s * , τ) = (I - αL G ) -1 W(r, s * , τ); where s * is the optimized adaptive scale, τ is the time offset, r is the input signal, α is the smoothing factor for controlling the neighborhood smoothness, and L G is the graph Laplacian matrix, and the specific formula is as follows: L G = D - A; In the formula, D is the degree matrix, A is the adjacency matrix, representing the similarity between wavelet scales; The specific formula of the degree matrix D is: D = {D ii | i ∈ {1, 2,..., n}}; where n is the number of nodes, and D ii is the degree of node i, and the specific formula is: D ii = ∑ j A ij ; The specific formula of the adjacency matrix A is: where n is the number of nodes, and the element in the i-th row and j-th column of the adjacency matrix A, denoted as A ij indicates whether there is an edge between node i and node j. The specific formula is as follows: Use the inverse wave transform to convert the weighted wavelet features into a time domain signal. The specific formula is: where X freq (t) is the converted time-domain signal, and the specific formula is: where s * is the optimized adaptive scale, and τ is the time offset.

3. The long-term load metal fatigue prediction method for steel structure buildings based on machine learning according to claim 2, characterized in that In the step S41, for each time step t in the input sequence r = {r1, r2,...., r T}, a learnable convolutional kernel K t is designed to perform a convolution operation on the input signal r at time step t. The specific formula is as follows: h t = Convld(r t , K t ); where h t is the output after the convolution operation, Convld is a one-dimensional convolution operation, and K t is the designed learnable convolution kernel; The specific formula of the learnable convolution kernel is: k t = round(σ(C t + θ t )·(k max - k min ) + k min ); where C t is the local curvature, θ t is a learnable parameter, k max is the maximum convolution kernel, k min is the minimum convolution kernel, σ(·) is the Sigmoid function, and round(·) is the rounding function; The specific formula of the local curvature is: In the formula, is the second derivative; Learnable parameter θ t Obtained by training optimization, and the specific formula is: where L is the cross-entropy loss function, W i is the convolutional kernel weight, and η is the learning rate; The specific formula for updating the convolution kernel weights is: In the formula, L is the cross-entropy loss function, and η is the learning rate; The specific formula of the cross-entropy loss function is: where N is the total number of samples, y t is the true value, is the predicted value.

4. A method for predicting long-term load metal fatigue of steel structure buildings based on machine learning according to claim 3, characterized in that, In step S42, for the convolution results h at different times t , the SeLU activation function is applied to obtain the processed data The specific formula is as follows: where h t is the output after dynamic convolution at time step t, and SeLU is the activation function. The specific formula is as follows: SeLU(x) = λ·max(0, x)+(1 - λ)·(α·(exp(x)-1)); In the formula, λ and α are predefined hyperparameters; For perform layer normalization, and the specific formula is: In the formula, LayerNorm is the layer normalization operation, and the specific formula is: where μ t is the mean at each time step t, σ t is the standard deviation at each time step t, and ∈ is a small constant set to prevent division by zero errors.

5. A method for predicting long-term load metal fatigue of a steel structure building based on machine learning according to claim 4, characterized in that The output after layer normalization in step S43 is subjected to residual connection with the input signal r t The specific formula is as follows: In the formula, is the feature representation after adding the residual connection; Use feature splicing to merge the features of all time steps to obtain the final time-domain feature representation. The specific formula is as follows: where Concat is the concatenation operation, is the feature representation after residual connection at time step t.

6. A method for predicting long-term load metal fatigue of a steel structure building based on machine learning according to claim 5, characterized in that In step S5, the features extracted from the time domain and the frequency domain are fused to form a unified feature representation h. The specific formula is Wherein, Concat is a splicing operation, and Linear is a linear layer. is the time-domain feature representation. is the frequency-domain feature representation. Design a learnable weight γ to adjust the weights of the temporal contrast loss and the frequency domain contrast loss. Use contrastive learning to learn the contrastive features in the temporal and frequency domains, and train the model by minimizing the total loss L total The specific formula is as follows: L total = γ·L time + (1 - γ)·L freq ; where L time is the contrastive loss in the time domain, and L freq is the contrastive loss in the frequency domain; The specific formula of the learnable weight γ is: where, Δ time is the change rate of the time-domain loss, Δ freq is the change rate of the frequency-domain loss, S time，freq is the similarity measure of the time-domain and frequency-domain features, controlling the weights of the time domain and the frequency domain, and η is the learning rate; The change rate Δ of the time-domain loss time The specific formula is as follows: The change rate Δ of the frequency-domain loss freq The specific formula is as follows: Similarity measure S of time-domain and frequency-domain features time，freq The specific formula is as follows: In the formula, is the L2 norm of the time-domain feature, is the L2 norm of the frequency-domain feature; The contrastive loss L in the time domain time The specific formula is as follows: In the formula, is the feature obtained through the residual connection, is the final time-domain feature representation; Frequency domain contrastive loss L freq The specific formula is as follows: L freq = λL amp + (1 - λ)L phase ; where λ is the dynamic weight, and L amp is the amplitude loss, and the specific formula is: Where N is the number of samples, and |F pred (i)| is the predicted frequency-domain amplitude, and F true (i)| is the true frequency-domain amplitude; L phase is the phase loss, and the specific formula is: where arg(F pred (i)) is the predicted phase in the frequency domain, and arg(F true (i)) is the true phase in the frequency domain.

7. A method for predicting metal fatigue of long-term loads of steel structure buildings based on machine learning according to claim 1, characterized in that, Regarding the problem of predicting metal fatigue under long-term loads in steel structure buildings, the collected data for predicting metal fatigue under long-term loads in steel structure buildings includes physical parameter data of steel structure materials and usage environment parameter data. Among them, the physical parameter data includes tensile strength data, compressive strength data, yield strength data, density data, elastic modulus data, fatigue life data, and hardness data of steel structure materials. The installation environment parameter data includes long-term load data, environmental temperature data, environmental humidity data, environmental vibration data, and lighting condition data. Preprocess the collected relevant data to ensure that there are no outliers in the data, and then divide the processed data into a training set and a test set for training and evaluating the performance of the prediction model for metal fatigue under long-term loads in steel structure buildings.

Images