A structural construction process explainable parameter prediction method based on a CBAM-GRU model

Abstract

This invention discloses a method for predicting interpretable parameters in the structural construction process based on a CBAM-GRU model, comprising: S1, identifying multiple parameters to be identified that influence the main parameters, acquiring data on the main parameters and parameters to be identified at each construction stage, and preprocessing them; S2, determining the correlation coefficients between each parameter to be identified and the main parameters, and identifying multiple main influencing parameters based on the correlation coefficients; S3, extracting data on the main parameters and each main influencing parameter to construct a dataset, and dividing it into a training set and a test set; S4, constructing a CBAM-GRU neural network model, training the model using the training set, and performing predictions on the test set using a rolling prediction mode to output the predicted values ​​of the main parameters for the next construction stage; S5, extracting the weight vectors of the channel attention layers in the CBAM module, and performing interpretability analysis on the contribution of each main influencing parameter to the predicted values ​​of the main parameters. This invention can predict multiple main parameters in the structural construction process and perform interpretability analysis on the prediction results.

Description

Technical Field

[0001] This invention relates to the field of structural construction technology. More specifically, this invention relates to a method for predicting interpretability parameters of structural construction processes based on the CBAM-GRU model. Background Technology

[0002] In the construction of bridges and other structures, predictions for the next stage of construction primarily focus on parameter identification and state prediction. These methods can predict the next construction state even without knowing the exact model or retrieving previous measurement data. However, these methods cannot provide the probability of each possibility occurring, nor can they provide interpretability analysis of the results. This makes it difficult for engineers to evaluate the predictions and adjust the next stage of construction. Therefore, a method that can provide interpretable predictions of the structural construction state would be more conducive to timely adjustments during the construction process and improve construction efficiency. Summary of the Invention

[0003] One object of the present invention is to solve at least the above-mentioned problems and to provide at least the advantages that will be described later.

[0004] To achieve these objectives and other advantages according to the present invention, a method for predicting interpretability parameters of structural construction process based on the CBAM-GRU model is provided, characterized by comprising the following steps: S1. Identify the multiple parameters to be identified that affect the main parameters, obtain the data of the main parameters and parameters to be identified at each construction stage, and perform preprocessing. S2. Analyze the influence of each parameter to be identified on the main parameter, determine the correlation coefficient between each parameter to be identified and the main parameter, and identify multiple main influencing parameters based on the correlation coefficient. S3. Extract the main parameters and the data of each major influencing parameter to construct the dataset, and divide it into training set and test set; S4. Construct a CBAM-GRU neural network model, train the model using the training set, and use the rolling prediction mode to make predictions on the test set, outputting the predicted values ​​of the main parameters for the next construction stage. S5. Extract the weight vector of the channel attention layer in the CBAM module, and perform interpretability analysis on the contribution of each major influencing parameter to the predicted value of the master parameter.

[0005] Preferably, in step S1, the preprocessing includes: completing the missing data of each parameter to be identified by using a polynomial interpolation method, and then normalizing each data.

[0006] Preferably, step S2 specifically includes: obtaining the correlation coefficient between each parameter to be identified and the principal parameter using the Spearman rank correlation coefficient method; and taking the parameters to be identified with a correlation coefficient greater than 0.5 as the main influencing parameters.

[0007] Preferably, in step S3, for the total number of samples N, the first M samples are used as the training set and the remaining NM samples are used as the test set in chronological order, with the ratio of M to (NM) between 7:3 and 8:2.

[0008] Preferably, the CBAM-GRU neural network model described in step S4 includes; The input layer is used to receive the training set data; The CBAM attention module includes a channel attention submodule and a spatial attention submodule; the channel attention submodule is used to generate attention weight vectors in the feature channel dimension; the spatial attention submodule is used to generate attention weight matrices in the time step dimension. The GRU temporal learning layer receives feature sequences weighted by the CBAM attention module and learns the long-term dependencies of the master parameter evolution through a gating mechanism. The output layer is used to output the predicted values ​​of the master parameters for the next construction stage.

[0009] Preferably, the hyperparameters of the CBAM-GRU neural network model include the sliding window length, the number of hidden neurons in the GRU hidden neurons, the learning rate, and the number of training epochs; step S4 further includes optimizing the hyperparameters of the CBAM-GRU neural network model using the control variable method, specifically including: A portion of the data in the training set is extracted as the validation set. Starting with a set of empirical hyperparameters, optimization is performed in the order of sliding window length, number of hidden neurons in the GRU hidden neuron set, learning rate, and number of training epochs. The optimization process is as follows: fix other hyperparameters, change the hyperparameter to be optimized, and select the value that makes the model achieve the best performance on the validation set as the final value of the hyperparameter to be optimized.

[0010] Preferably, the hyperparameters of the CBAM-GRU neural network model are set to their final values, and the model is trained using the training set. On the test set, a rolling prediction mode is adopted, using the data from the first L time steps to predict the deflection at the (L+1)th time step, and then using the predicted value as part of the input to predict the deflection at the (L+2)th time step, and so on, to complete the prediction for the entire test set. The root mean square error, mean square error, and fitting coefficient between the predicted values ​​of the principal parameters and the true values ​​of the test set are calculated to evaluate the model.

[0011] Preferably, in step S5, the average channel attention weight of each major influencing parameter is calculated based on all samples in the test set: when the prediction error of the main parameter exceeds the limit, fault diagnosis is performed based on the average channel attention weight.

[0012] The present invention has at least the following beneficial effects: 1. The structural construction process interpretability parameter prediction method based on the CBAM-GRU model provided by this invention adaptively focuses on key features and key time points through the CBAM attention mechanism, and combines the GRU network to capture long-term temporal dependencies, which significantly improves the prediction accuracy compared with the traditional LSTM model.

[0013] 2. The structural construction process interpretability parameter prediction method based on the CBAM-GRU model provided by this invention quantifies the CBAM attention weights as the contribution of the main influencing parameters to the principal parameters, forming a two-layer interpretability framework that combines Spearman's global monotonic correlation analysis and CBAM local dynamic attention weight analysis.

[0014] 3. The structural construction process interpretability parameter prediction method based on the CBAM-GRU model provided by this invention uses Spearman's rank correlation coefficient for feature selection, does not rely on data distribution assumptions, is insensitive to outliers, can capture nonlinear monotonic relationships, and ensures the scientific nature and robustness of input features.

[0015] 4. The structural construction process interpretability parameter prediction method based on the CBAM-GRU model provided by this invention can predict a variety of main parameters in the structural construction process and perform interpretability analysis on the prediction results, and has wide applicability.

[0016] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method for predicting interpretability parameters of the structural construction process based on the CBAM-GRU model as described in this invention; Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings, so that those skilled in the art can implement it based on the description.

[0019] It should be noted that, unless otherwise specified, the experimental methods described in the following embodiments are all conventional methods, and the reagents and materials described are all commercially available unless otherwise specified. In the description of this invention, the terms "lateral", "longitudinal", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", and "outer" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0020] like Figure 1 As shown, this invention provides a method for predicting interpretability parameters of a structure construction process based on the CBAM-GRU model, comprising the following steps: S1. Identify the multiple parameters to be identified that affect the main parameters, obtain the data of the main parameters and parameters to be identified at each construction stage, and perform preprocessing. S2. Analyze the influence of each parameter to be identified on the main parameter, determine the correlation coefficient between each parameter to be identified and the main parameter, and identify multiple main influencing parameters based on the correlation coefficient. S3. Extract the main parameters and the data of each major influencing parameter to construct the dataset, and divide it into training set and test set; S4. Construct a CBAM-GRU neural network model, train the model using the training set, and use the rolling prediction mode to make predictions on the test set, outputting the predicted values ​​of the main parameters for the next construction stage. S5. Extract the weight vector of the channel attention layer in the CBAM module, and perform interpretability analysis on the contribution of each major influencing parameter to the predicted value of the master parameter.

[0021] The following is a detailed explanation using the construction of a cable-stayed bridge as an example. The main parameter is set as the deflection value of the main beam segment. In step S1, the parameters to be identified are first defined, including: geometric and material parameters (weight of the main beam segment, elastic modulus of the main beam, elastic modulus of the stay cables), environmental parameters (ambient temperature, daily average temperature difference), and construction process parameters (tension force, construction stage number). This cable-stayed bridge has a main span of 350 meters and is constructed using the cantilever casting method. It consists of 30 standard segments, each containing 10 main construction conditions, totaling 300 construction stages. Data on the main parameters and parameters to be identified for each construction stage are obtained and preprocessed. The preprocessing includes: using polynomial interpolation to complete the missing data for each parameter to be identified, and then normalizing the data to scale all parameters to the [0,1] interval.

[0022] In step S2: the correlation coefficient between each parameter to be identified and the principal parameter is obtained using the Spearman rank correlation coefficient method; parameters with a correlation coefficient exceeding 0.5 are selected as the main influencing parameters. The specific process of screening using the Spearman rank correlation coefficient method is as follows: For each parameter (including the main parameter and each parameter to be identified), its N observations are sorted from smallest to largest, and the original values ​​are replaced with the rank (1,2,...,N); if there are ties, the average rank is taken. For each parameter to be identified, the Spearman rank correlation coefficient ρ between its rank sequence and the deflection rank sequence is calculated. s : (1) In equation (1), This represents the difference in grade between this parameter and the deflection during the i-th construction stage.

[0023] according to The values ​​are sorted among all parameters to be identified, and then selected. Parameters with a value greater than 0.5 are considered as the main influencing parameters. In this embodiment, the main influencing parameters obtained are: elastic modulus of the stay cable, weight of the main beam segment, elastic modulus of the main beam, and cable force of the stay cable.

[0024] In step S3, the main parameters and the main influencing parameters are extracted from the data preprocessed in step S1 to construct a dataset, and the dataset is divided into a training set and a test set. For the total number of samples N, N=300, the first X samples are used as the training set and the remaining NX samples are used as the test set in chronological order. The ratio of X to (NX) is between 7:3 and 8:2.

[0025] In step S4, the CBAM-GRU neural network model includes; The input layer is used to receive training set data; the input vector is (m,L,K), where m is the batch size, L is the sliding window length, and K is the number of features (including the main parameters and the main influencing parameters).

[0026] The CBAM attention module includes a channel attention submodule and a spatial attention submodule; The channel attention submodule is used to generate attention weight vectors along the feature channel dimensions; it calculates weights for K feature channels and identifies the most important features for the current prediction. Channel information is extracted through global average pooling and global max pooling, and then channel attention weights are generated via a shared multilayer perceptron.

[0027] The spatial attention submodule generates an attention weight matrix along the time step dimension; it calculates weights for L time steps, focusing on the most critical historical moments for prediction. Average pooling and max pooling are applied along the channel dimension, followed by convolutional layers to generate spatial attention weights. The CBAM module output is the element-wise product of the channel and spatial attention weights, i.e., the feature sequence weighted by the CBAM attention module.

[0028] The GRU temporal learning layer receives feature sequences weighted by the CBAM attention module. This layer contains hidden layer neurons and learns long-term dependencies in the evolution of the master parameters through a gating mechanism. The output layer outputs the predicted master parameters for the next construction stage. The output layer maps the final hidden state of the GRU to the predicted deflection value for the next construction stage.

[0029] Furthermore, the key hyperparameters of the CBAM-GRU neural network model include the sliding window length, the number of hidden neurons in the GRU hidden neurons, and the learning rate; step S4 also includes optimizing the hyperparameters of the CBAM-GRU neural network model using the control variable method, specifically including: A portion of the data in the training set is extracted as the validation set. Starting with a set of empirical hyperparameters, optimization is performed in the order of sliding window length, number of hidden neurons in the GRU hidden neuron set, and learning rate. The optimization process is as follows: keep other hyperparameters fixed, change the hyperparameter to be optimized, and select the value that makes the model achieve the best performance on the validation set as the final value of the hyperparameter to be optimized.

[0030] ① Optimize the sliding window length L: Fixed parameters: GRU hidden layer neuron count H=64, learning rate η=0.01, m=32, number of training epochs=300 Candidate values ​​for L: L∈[3,5,7,10] Evaluation: The root mean square error (RMSE) of the predicted deflection values ​​corresponding to each L was calculated on the validation set. The values ​​were 2.35, 2.05, 2.18, and 2.42 mm, respectively. The L value with the smallest RMSE was selected as the final value of the sliding window length, L=5.

[0031] ② Optimize the number of neurons H in the GRU hidden layer: Fixed parameters: L = final value, learning rate η = 0.01, m = 32, number of training epochs = 300 Candidate values ​​for H: H∈[32,64,128,256] Evaluation: The root mean square error (RMSE) of the predicted deflection values ​​corresponding to each H was calculated on the validation set. The values ​​were 2.25, 2.08, 2.05, and 2.07 mm, respectively. The H value with the smallest RMSE was selected as the final value of the number of neurons in the GRU hidden layer, H=128.

[0032] ③ Optimize the learning rate η: Fixed parameters: L = final value, H = final value, m = 32, number of training rounds = 300 Candidate values ​​for η: η∈[0.001,0.01,0.1] Evaluation: Observe the training loss convergence curve and select the η value as the final value of the learning rate, which is the fastest and most stable convergence. When η=0.01, the convergence is the fastest and most stable.

[0033] ④ Optimize batch size m: Fixed parameters: L = final value, H = final value, η = final value, number of training rounds = 300 Candidate values ​​for m: m∈[16,32,64] Evaluation: Considering both training efficiency (time / round) and validation set RMSE, the optimal value of m was selected as the final batch size, m=32. The hyperparameters of the CBAM-GRU neural network model were set to their final values ​​(L=5, H=128, η=0.01, m=32), and the model was trained using the training set. During training, the validation set loss was monitored, and training was stopped when the validation loss no longer decreased after P consecutive rounds (P being the patience value, which can be set to P=20).

[0034] A rolling prediction mode is adopted on the test set. The deflection at step L+1 is predicted using the data from the first L time steps. Then, the predicted value is used as part of the input to predict the deflection at step L+2, and so on, to complete the prediction for the entire test set. The root mean square error (RMSE), mean square error (MSE), and fitting coefficient are calculated between the predicted values ​​of the principal parameters and the actual values ​​of the test set for model evaluation. Based on the values ​​of the RMSE, MSE, and fitting coefficient, it is determined whether the model prediction accuracy meets the structural construction control requirements. In this embodiment, the RMSE is 1.98 mm, the MSE is 3.92, and the fitting coefficient R² is 0.963. All indicators are better than the preset thresholds (RMSE≤3.0mm, MSE≤9.0, R²≥0.90), indicating that the model prediction accuracy meets the construction control requirements of the cable-stayed bridge.

[0035] In step S5, based on all samples in the test set, the average channel attention weights of each major influencing parameter are calculated. When the prediction error of the main parameter exceeds the limit, fault diagnosis is performed based on the average value of the channel attention weights.

[0036] (2) In equation (2), Let K be the attention weight of the k-th feature of the i-th sample, where K is the number of features and M is the number of test samples.

[0037] Furthermore, based on the average channel attention weights of each key influencing parameter, a tiered construction control strategy can be formulated: First-level key parameters (weight > 0.3) are monitored in real time with strict early warning thresholds. Second-level important parameters (0.15 < weight ≤ 0.3) are reviewed daily with regular early warning thresholds. Third-level attention parameters (weight ≤ 0.15) are tested in a trial phase as reference indicators. The temporal changes in attention weights during key construction stages (such as the maximum cantilever stage and the closure stage) are then tracked to identify the dominant influencing parameters for different construction stages. For example, in the early stages of cantilever construction, the weight of the main beam is the highest (0.35-0.40); in the maximum cantilever stage, the weight of the elastic modulus of the stay cables significantly increases to 0.45-0.50, indicating that the dominant influencing factors differ across construction stages. When the prediction error of a main parameter exceeds the threshold, the distribution of attention weights at that moment is analyzed. If the weight of a major influencing parameter is significantly higher than the historical average (e.g., exceeding the mean + 2 standard deviations), it indicates that the parameter may be abnormal. For example, the predicted deflection for a certain construction stage is 42.5 mm, while the measured deflection is 46.8 mm, an error of 4.3 mm (exceeding the limit). In this construction phase, the weight of the elastic modulus of the stay cables in the attention weighting was 0.52 (significantly higher than the average of 0.32). On-site verification revealed that the cable force of cable No. 28 was 12% lower than the design value; after adjusting the cable force, the prediction error in subsequent stages returned to within 2.1 mm.

[0038] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. For those skilled in the art, other modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and illustrations shown and described herein.

Claims

1. A method for predicting interpretability parameters of structural construction process based on the CBAM-GRU model, characterized in that, Includes the following steps: S1. Identify the multiple parameters to be identified that affect the main parameters, obtain the data of the main parameters and parameters to be identified at each construction stage, and perform preprocessing. S2. Analyze the influence of each parameter to be identified on the main parameter, determine the correlation coefficient between each parameter to be identified and the main parameter, and identify multiple main influencing parameters based on the correlation coefficient. S3. Extract the main parameters and the data of each major influencing parameter to construct the dataset, and divide it into training set and test set; S4. Construct a CBAM-GRU neural network model, train the model using the training set, and use the rolling prediction mode to make predictions on the test set, outputting the predicted values ​​of the main parameters for the next construction stage. S5. Extract the weight vector of the channel attention layer in the CBAM module, and perform interpretability analysis on the contribution of each major influencing parameter to the predicted value of the master parameter.

2. The method for predicting interpretability parameters of structural construction process based on the CBAM-GRU model as described in claim 1, characterized in that, In step S1, the preprocessing includes: completing the missing data of each parameter to be identified by using a polynomial interpolation method, and then normalizing each data.

3. The method for predicting interpretability parameters of structural construction process based on the CBAM-GRU model as described in claim 1, characterized in that, Step S2 specifically includes: obtaining the correlation coefficient between each parameter to be identified and the principal parameter using the Spearman rank correlation coefficient method; and taking the parameters to be identified with a correlation coefficient greater than 0.5 as the main influencing parameters.

4. The method for predicting interpretability parameters of structural construction process based on the CBAM-GRU model as described in claim 1, characterized in that, In step S3, for the total number of samples N, the first X samples are used as the training set and the remaining NX samples are used as the test set in chronological order. The ratio of X to (NX) is between 7:3 and 8:

2.

5. The method for predicting interpretability parameters of structural construction process based on the CBAM-GRU model as described in claim 1, characterized in that, The CBAM-GRU neural network model described in step S4 includes: The input layer is used to receive the training set data; The CBAM attention module includes a channel attention submodule and a spatial attention submodule; The channel attention submodule is used to generate attention weight vectors along the feature channel dimensions; The spatial attention submodule is used to generate an attention weight matrix along the time step dimension; The GRU temporal learning layer receives feature sequences weighted by the CBAM attention module and learns the long-term dependencies of the master parameter evolution through a gating mechanism. The output layer is used to output the predicted values ​​of the master parameters for the next construction stage.

6. The method for predicting interpretability parameters of structural construction process based on the CBAM-GRU model as described in claim 1, characterized in that, The hyperparameters of the CBAM-GRU neural network model include the sliding window length, the number of hidden neurons in the GRU hidden neurons, the learning rate, and the number of training epochs; step S4 also includes optimizing the hyperparameters of the CBAM-GRU neural network model using the control variable method, specifically including: A portion of the data in the training set is extracted as the validation set. Starting with a set of empirical hyperparameters, optimization is performed in the order of sliding window length, number of hidden neurons in the GRU hidden neuron set, learning rate, and number of training epochs. The optimization process is as follows: fix other hyperparameters, change the hyperparameter to be optimized, and select the value that makes the model achieve the best performance on the validation set as the final value of the hyperparameter to be optimized.

7. The method for predicting interpretability parameters of structural construction process based on the CBAM-GRU model as described in claim 6, characterized in that, The hyperparameters of the CBAM-GRU neural network model are set to their final values, and the model is trained using the training set. On the test set, a rolling prediction mode is used to predict the deflection at step L+1 using the data from the first L time steps. The predicted value is then used as part of the input to predict the deflection at step L+2, and so on, to complete the prediction for the entire test set. Model evaluation is performed by calculating the root mean square error, mean square error, and fit coefficient between the predicted values ​​of the main parameters and the true values ​​of the test set.

8. The method for predicting interpretability parameters of structural construction process based on the CBAM-GRU model as described in claim 7, characterized in that, In step S5, the average channel attention weight of each major influencing parameter is calculated based on all samples in the test set. When the prediction error of the main parameter exceeds the limit, fault diagnosis is performed based on the average channel attention weight.

Images