A method for identifying and predicting damage of PC box girder under fire and a monitoring and early warning system thereof

Abstract

The application discloses a PC box girder damage identification and prediction method and a monitoring and early warning system thereof and belongs to the technical field of bridge safety risk assessment, which comprises the following steps: selecting features of PC box girder prestress loss and bending resistance capacity loss under fire; based on the selected features, prestress loss database and bending resistance capacity loss database of the PC box girder under fire are constructed; all kinds of original data in the prestress loss database and the bending resistance capacity loss database are preprocessed, all parameters are converted into a format available for machine learning model analysis, and the data set is divided into a training set and a test set; based on the training set and the test set, the hyperparameters of the prestress loss and the bending resistance capacity loss corresponding models are optimized respectively; the optimized models are subjected to performance evaluation, result visualization processing and SHAP explainability analysis, and are applied to PC box girder damage identification and prediction under fire. The application solves the problem of large prediction error in the prior art.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of bridge safety risk assessment, and particularly relates to a PC box girder damage identification and prediction method under fire and a monitoring and early warning system thereof. BACKGROUND

[0002] With the rapid development of infrastructure construction and transportation industry in China, a large number of bridges and vehicles are emerging, bringing great convenience to people's travel. At the same time, prestressed bridges play a dominant role in bridge engineering in China. However, the increase in the number of bridges and vehicles undoubtedly increases the probability of fire accidents. Prestressed concrete structures are seriously affected by high temperature in fire, and the material properties change significantly. It is very important to detect and predict the prestress loss of prestressed concrete bridges under fire. The fire assessment method can be divided into traditional damage assessment method and machine learning assessment method. The traditional damage assessment method usually relies on experience rules and manual inspection. These methods are easily affected by environmental and subjective factors, have limited data processing capacity, and result in inaccurate and inefficient results. At the same time, the detection process may cause different degrees of damage to the structure, increase the instability and risk of the bridge structure that has been on fire, and greatly threaten the safety of the detection and subsequent construction personnel. The damage assessment method based on machine learning uses advanced computer science and data analysis technology to improve the accuracy, reliability and efficiency of damage assessment. Compared with the traditional damage assessment method, the damage assessment method based on machine learning can realize fast and accurate damage assessment, improve the efficiency and accuracy of maintenance work, and also can reduce the investment of human resources and maintenance cost.

[0003] Among them, based on XGBoost, LightGBM, CatBoost, RandomForest, GradientBoosting, ExtraTrees

[0004] The damage identification method of these tree models needs to rely on a large amount of high-quality labeled data, but the bridge fire cases are relatively scarce and highly heterogeneous, which leads to insufficient training data, easy overfitting, and affects the model generalization ability. At the same time, the model extrapolation ability and explainability are poor, and it cannot provide confidence interval or probability output, while the structure safety assessment needs risk probability judgment, and the prediction may cover up the key uncertainty risk. Based on Lasso, Ridge, ElasticNet, BayesianRidge, these linear models cannot directly capture complex nonlinear relationships, are highly dependent on engineering features, and are very sensitive to outliers. SVR, KernelRidge, these kernel methods have large computational overhead, and their performance is highly dependent on the selection of kernel function type (such as RBF, polynomial) and its hyperparameters (such as gamma, C), and tuning requires a lot of calculation and experience. MLPRegressor neural network needs a large amount of data to avoid overfitting, and the training is unstable, with many hyperparameters. SUMMARY

[0005] The PC box girder damage identification prediction method and the monitoring and early warning system provided by the application solve the above problems in the prior art.

[0006] In order to achieve the above-mentioned purpose of the application, the technical scheme adopted by the application is as follows: a PC box girder damage identification prediction method under fire, comprising the following steps:

[0007] S1: selecting features for constructing PC box girder prestress loss and bending resistance capacity loss under fire, including bridge basic information, cross section geometric parameters, material indexes and fire scene;

[0008] S2: based on the selected features, constructing prestress loss database and bending resistance capacity loss database of PC box girder under fire;

[0009] S3: preprocessing various types of original data in the prestress loss database and the bending resistance capacity loss database, converting each parameter into a format available for machine learning model analysis, and dividing the data set into training set and test set;

[0010] S4: based on the training set and the test set, the hyperparameters of the prestress loss and the bending resistance capacity loss corresponding model are optimized respectively;

[0011] S5: performance evaluation, result visualization processing and SHAP explainability analysis are performed on the optimized model, and the model is applied to PC box girder damage identification prediction under fire.

[0012] Further, the hyperparameters of the prestress loss corresponding model are optimized, including the following steps:

[0013] A1: Based on the training set and the test set, the prestress loss original data features are processed, and the polynomial features are generated by using the original features;

[0014] A2: The feature distribution space of each sample is described from a statistical point of view, and statistical feature extraction is performed;

[0015] A3: Based on the statistical feature extraction result, the first ten important features are multiplied two by two to create interaction features, and the number of interaction features is generated;

[0016] A4: All generated features combined by polynomial features, statistical features and interaction features are combined into a matrix in order, and feature expansion is performed;

[0017] A5: Based on the feature expansion result, feature data standardization is performed;

[0018] A6: Based on the feature data standardization result, feature selection is performed, taking random forest as the evaluator, calculating the importance score of each feature, and selecting important features according to the threshold;

[0019] A7: Based on the important features, through multi-model initialization, the prestress damage prediction model is trained in parallel, and the evaluation calculation is performed to obtain the best model, and the hyperparameter optimization is performed;

[0020] A8: The model is integrated into a two-layer Stacking integrated structure, selects the best performing models GradientBoosting, RandomForest, ExtraTrees and XGBoost as base learners, and selects random forest, gradient boosting tree, ridge regression and Lasso regression four kinds of meta-learners for combination, through 3-fold cross-validation, selects the integrated model with the highest comprehensive evaluation score as the best integrated model;

[0021] A9: Compare the performance of single best model and integrated model, take the integrated model based on GradientBoosting, RandomForest, ExtraTrees and XGBoost as base learners, and Lasso regression as meta-learner as the prestress loss prediction model of PC box girder under fire.

[0022] Further, the optimization target of the Lasso regression is:

[0023]

[0024]

[0025] wherein, is a column vector containing all sample true target values, The input feature matrix of the meta-learner. To find the coefficient vector, For regularization hyperparameters in Lasso regression, For GradientBoosting model The predicted value for each sample, For the RandomForest model The predicted value for each sample, For the ExtraTrees model The predicted value for each sample, For XGBoost model pair Predicted values ​​for each sample;

[0026] The final prediction model for Lasso regression is:

[0027]

[0028] in, For the final prediction model results, , , and To correspond to the weights of different base learners, For bias terms, These are the predicted values ​​from the Gradient Boosting model. These are the predicted values ​​from the RandomForest model. These are the predicted values ​​from the ExtraTrees model. These are the predicted values ​​from the XGBoost model.

[0029] Furthermore, the hyperparameters of the model corresponding to the loss of flexural bearing capacity are optimized, including the following steps:

[0030] B1: The Optuna framework is used to optimize the hyperparameters of XGBoost, SVR base learners, and Ridge meta learners using the TPE algorithm.

[0031] B2: Based on the optimization results, the prediction results of XGBoost and SVR base learners are concatenated into meta-features, the Ridge meta-learner is trained, and the MedianPruner pruning strategy is combined to dynamically eliminate inefficient parameter combinations to obtain the globally optimal parameters.

[0032] B3: Based on the globally optimal parameters, a preprocessing pipeline and a base model are constructed, and finally the Ridge meta-model is trained to form a two-layer stacked structure.

[0033] Further, the Ridge meta-learner introduces an L2 regularization term on the basis of ordinary least squares, and the optimization objective is:

[0034]

[0035]

[0036] wherein, is a regularization hyperparameter in ridge regression, is the predicted value of the SVR model for the i-th sample;

[0037] The final predicted value of the Ridge meta-learner is:

[0038]

[0039] wherein, is the final predicted value of the Ridge meta-learner, is the predicted value of the XGBoost model, is the predicted value of the SVR model, and is a weight coefficient.

[0040] A monitoring and early warning system for a PC box girder damage identification and prediction method under fire, comprising:

[0041] Temperature sensors are arranged at key parts of the bridge to monitor temperature changes and locate the fire area;

[0042] Infrared cameras are used to identify the location, height and range of the fire source and record the fire time;

[0043] Storage box, built-in data transceiver processing device and detection drone, when the data transceiver processing device receives temperature anomaly and detects fire characteristics, automatically opens the valve of the box body to release the detection drone;

[0044] Detection drone, equipped with infrared camera and data transmission device, used for multi-angle data acquisition;

[0045] Central processing unit, used for running damage identification and prediction model, outputting prediction results, including prestress damage and bending capacity loss identification results under fire;

[0046] Alarm system, including LED display screen, warning light and loudspeaker, used for fire warning and information broadcast;

[0047] Solar power generation device, providing power support for the system.

[0048] The beneficial effects of the present application are:

[0049] ​(1) The application can simultaneously process the prediction of prestress loss and the prediction of loss of bending resistance carrying capacity, both of which are predicted by integrated models to avoid the insufficient prediction of single models, and the advantages of each prediction model are exerted at the same time, and optimization strategies are developed for different tasks: prestress loss prediction integrated optimization scheme with GradientBoosting, RandomForest, ExtraTrees, XGBoost as base learners and Lasso as meta-learner, and bending resistance carrying capacity loss prediction integrated optimization scheme with XGBoost and SVR as base learners and Ridge as meta-learner. Two prediction models can be established at the same time, and the prestress loss and bending resistance carrying capacity loss of PC box girder after fire can be predicted with high precision. In the aspect of prestress loss prediction, the superiority of GradientBoosting, RandomForest, ExtraTrees and XGBoost in predicting the prestress loss of PC box girder under fire is successfully inherited, and Lasso is used as a meta-learner to assign weights to the above base learners and adjust L1 regularization, which takes into account the high precision of GB / XGB series and the strong robustness of RF / ET series, and has good prediction accuracy, efficiency and stability; in the aspect of bending resistance carrying capacity prediction, the short board of traditional single XGBoost or SVR model method in the aspect of generalization performance of unknown new samples in bending resistance carrying capacity prediction task is overcome, the core advantages of XGBoost and SVR as base learners and Ridge as meta-learner are exerted and combined with each other, and the prediction accuracy is significantly improved.

[0050] (2) High efficiency and high accuracy: the data of the fire position, temperature distribution, burst depth and fire time collected by the monitoring system are input into the trained model in real time, and the prestress loss and residual bending resistance carrying capacity are quickly predicted, compared with the traditional finite element analysis (which needs several minutes to calculate), the machine learning model can complete the evaluation in seconds to meet the emergency decision-making demand, and the prediction result has high accuracy.

[0051] (3) Good generalization ability and intuitiveness, through early stopping mechanism, L1 regularization (Lasso regression), L2 regularization (Ridge regression), cross-validation (3-fold, 5-fold) and other technologies to better suppress overfitting. The effectiveness and advancement of the stacking ensemble framework of the prestress loss prediction model with Lasso as the meta-learner in improving the generalization ability of the model, and the core advantage comes from the built-in sparse feature selection ability and model complexity control ability given by L1 regularization. This mechanism enables the meta-learner to adaptively identify and integrate the most valuable base learner signals while suppressing noise and redundancy, and ultimately builds a strong learner with high prediction accuracy, high robustness and good generalization performance. The anti-bending capacity loss prediction stacking model (Stacking) combines the nonlinear fitting ability of XGBoost and the high-dimensional processing advantage of SVR, and further improves the generalization through the meta-learner (Ridge). The MedianPruner pruning is introduced in the Optuna framework to automatically terminate inefficient tests and speed up hyperparameter search. Finally, the two stacking models combine SHAP analysis to quantify feature contribution (such as the influence weight of fire time on prestress loss), providing clear improvement direction for fireproof design. Visualization tools (residual plot, bee plot) visually display model performance and error distribution, facilitating engineers to quickly verify and make decisions.

[0052] (4) Perfect system and engineering practicability, combined with intelligent monitoring and early warning device system, forming a complete "real-time monitoring-fire warning-damage assessment-emergency decision-making" whole-chain solution, further improving the engineering practicability and intelligent level of the system.

[0053] (5) Can help reduce the risk and structural damage caused by fire and shorten the emergency response time. According to the damage prediction results, adjust the monitoring focus and optimize resource allocation. The prediction results can quickly generate evaluation repair suggestions, shorten the post-disaster evaluation period; at the same time, in terms of safety, reduce the operation risk of detection personnel, and reduce the threat to the life safety of detection personnel due to unstable bridge structure after fire. BRIEF DESCRIPTION OF DRAWINGS

[0054] Figure 1 It is a flow chart of a PC box girder damage identification and prediction method under fire.

[0055] Figure 2 It is a schematic diagram of converting text features to numerical types through One-Hot encoding in the present application.

[0056] Figure 3 (a), (b), (c), (d), (e) are R 2Performance comparison chart, RMSE comparison chart, overfitting comparison chart, training time comparison chart, and stability comparison chart.

[0057] Figure 4 In Figures (a) and (b), the prestress loss of PC box girders under fire in this invention is based on GradientBoosting, RandomForest, ExtraTrees, XGBoost as the base learner and Lasso as the meta learner, and the training and testing scatter plots.

[0058] Figure 5 In the figure, (a), (d), and (e) are scatter plots, residual distribution scatter plots, and residual frequency distribution plots of the flexural bearing capacity loss of PC box girders under fire in this invention, respectively based on XGBoost and SVR as base learners and Ridge as meta learners. (b) and (c) are scatter plots based on XGBoost and SVR, respectively.

[0059] Figure 6 In this invention, (a) and (b) are respectively the SHAP global feature importance diagrams of prestress loss and flexural bearing capacity loss of PC box girders under fire conditions.

[0060] Figure 7 In this invention, (a) and (b) are SHAP swarm diagrams of prestress loss and flexural bearing capacity loss of PC box girders under fire conditions, respectively.

[0061] Figure 8 This is a schematic diagram of the overall structure of the monitoring and early warning system in this invention.

[0062] Figure 9 This is a schematic diagram of the operation and implementation process of the monitoring and early warning device system in this invention.

[0063] The components include: 1. LED display screen; 2. Infrared camera; 3. Solar power generation device; 4. Warning light; 5. Speaker; 6. Storage box; 601. Data transceiver and processing device; 7. Detection drone; 701. Small infrared camera; 702. Data storage and transmission device; 8. Temperature sensor; 9. Central processing unit. Detailed Implementation

[0064] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0065] Example 1, such as Figure 1 As shown, a method for damage identification and prediction of PC box girders under fire includes the following steps:

[0066] S1: Select the characteristics of prestress loss and flexural bearing capacity loss of PC box girder in fire, including basic bridge information, cross-sectional geometric parameters, material properties and fire scenario;

[0067] The selection criteria prioritize relevance and informational value to improve model performance, reduce computational costs, and prevent overfitting. Basic bridge information includes bridge type, number of spans, and span length; cross-sectional geometric parameters include bridge height, mid-span cross-section width, and mid-span cross-section thickness; material parameters include the location of steel strands at the mid-span cross-section, longitudinal reinforcement strength, protective layer thickness, and concrete spalling depth; and fire scenarios include the number of fire-exposed surfaces, fire-exposed locations, fire-exposed length, and fire duration. These parameters can be further subdivided into specific indicators.

[0068] S2: Based on the selected features, construct a database of prestress loss and flexural bearing capacity loss of PC box girders under fire conditions;

[0069] This embodiment first collects data, forming a database by using experimental data from existing literature and data obtained from finite element analysis.

[0070] The data obtained from finite element analysis is calculated using numerical methods such as computational fluid dynamics (CFD) and the finite element method (FEM). The steps for structural response analysis under fire using ANSYS finite element software generally consist of three steps: First, establishing a structural geometric model; second, inputting material thermal performance parameters, selecting thermal analysis elements with temperature degrees of freedom, meshing the geometric model to establish a finite element model, applying temperature loads in the form of temperature rise curves, and analyzing the temperature field distribution of the structure under fire; third, replacing the thermal analysis elements in the finite element model with corresponding structural analysis elements, inputting material properties such as stress-strain curves at different temperatures, setting self-weight and constraints, and applying external loads. The temperature field obtained at each time step in the thermal analysis is then applied to the structure as temperature loads, allowing for structural analysis. To better simulate the temperature field changes, deformation, and load-bearing capacity changes of steel plate composite beam bridges under fire, appropriate element selection and mesh generation are necessary.

[0071] The temperature field at each time step is obtained using CFD software, standards, or empirical models to obtain the adiabatic surface temperature or surface heat flux of the target structure, and then used as input for FEM heat conduction calculation to calculate the structure temperature.

[0072] Repeat the finite element simulation multiple times by changing the features described in step S1 to populate the database content.

[0073] When the dataset of fire temperature field obtained by the model has many variable features, it is impossible to use formula derivation methods commonly used in machine learning, such as genetic algorithms, to obtain an expression that can match the accuracy of the model. Instead, formula fitting methods are used to establish an expression that meets the accuracy requirements of the model.

[0074] S3: Preprocess each type of raw data in the prestress loss database and the bending resistance capacity loss database, convert each parameter into a format available for machine learning model analysis, and divide the data set into a training set and a test set;

[0075] In this embodiment, One-Hot encoding is performed on the category variable to generate an orthogonal basis feature space.

[0076] Among them, the bridge type is processed by label encoding, and the simply supported beam is encoded as 0, and the continuous beam is encoded as 1. The concrete burst depth and the fire location are processed by one-hot encoding, such as the top plate fire in the fire location, which can be encoded as (1, 0, 0, 0, 0, 0), and the bottom plate fire can be encoded as (0, 1, 0, 0, 0, 0). The detailed encoding results can be seen in Figure 2 .

[0077] The data samples are normalized and cleaned, and the missing values (NaN is replaced by 0, and infinite value is replaced by a large number) are processed by np.nan_to_num to ensure data integrity.

[0078] Z-score standardization is the most commonly used method, which converts data to a standard normal distribution with a mean of 0 and a standard deviation of 1. For the original feature x , the normalized feature z can be calculated by the following formula:

[0079]

[0080] In the formula, is the mean of the feature, is the standard deviation of the feature.

[0081] Min-max normalization, also known as maximum and minimum normalization, linearly maps data to a given range, usually between [0, 1]. For the original feature x , the normalized feature z can be calculated by the following formula:

[0082]

[0083] In the formula, is the minimum value of the feature, is the maximum value of the feature.

[0084] Decimal scaling normalization converts data to the range [-1, 1] by dividing by a fixed base, usually the maximum absolute value of the feature. For the original feature x , the decimal scaling normalized feature z can be calculated by the following formula:

[0085]

[0086] wherein, d is an integer such that the maximum absolute value of the feature is less than 1.

[0087] After data preprocessing, the training set and the test set are divided according to 8.5:1.5 for the prestress loss data set and 8:2 for the flexural capacity loss data set. The training set is used to learn the certain potential distribution rule of the data, and the validation set is used to verify the discriminant ability of the learned model parameters to new samples.

[0088] S4: Based on the training set and the test set, the hyperparameters of the prestress loss and the flexural capacity loss corresponding model are optimized for the prestress loss, respectively;

[0089] (1) The hyperparameters of the prestress loss corresponding model are optimized, including the following steps:

[0090] A1: Use the original features to generate polynomial features to capture the interaction effect between the features, wherein is the number of original features, i , j is a certain original feature, , represent two different features in the original feature set;

[0091] A2: Describe the feature distribution characteristics of each sample from a statistical point of view, and perform the following statistical feature extraction, 1. Mean: , 2. Standard deviation: , 3. Maximum: , 4. Minimum: , 5. Median: , 6. 25% quantile: , 7. 75% quantile: , 8. Range: ;

[0092] A3: Create interaction features, multiply the first ten important features two by two, further enhance the non-linear relationship capture ability between the features, and generate the number of interaction features: ;

[0093] A4: Combine all the generated features of the polynomial features, statistical features, and interaction features in sequence into a matrix to improve the model performance, solve the problem of insufficient original feature information, and at the same time, perform controllable feature expansion to avoid the problem of dimension disaster caused by unlimited increase of features;

[0094] A5: Standardization of feature data is performed to make features of different scales comparable and improve model convergence. Z-score standardization is the most commonly used method, which converts data into a standard normal distribution with a mean of 0 and a standard deviation of 1.

[0095] A6: Feature selection is performed, and random forest is selected as the evaluator, which has the advantages of handling high-dimensional features, providing reliable feature importance evaluation, and being less prone to overfitting. The importance score of each feature is calculated, and important features are selected according to the threshold. The data is converted, and only the selected features are retained.

[0096] The random forest feature importance calculation method is based on the average impurity reduction:

[0097]

[0098] Where: is the importance coefficient of feature j , is the number of decision trees (100), is the number of nodes split using feature j , is the number of samples in node s , is the total number of samples, is the impurity reduction brought by node s split.

[0099] The calculation of impurity reduction is usually based on variance reduction: , where is the variance, is the parent node (before splitting), is the variance of all sample target values in the parent node, is the number of samples assigned to the left child node after splitting, is the number of samples assigned to the right child node after splitting.

[0100] For threshold setting, the median threshold is selected as the threshold setting, which has the characteristics of self-adaptation and insensitivity to outliers. Features with importance scores higher than the median are selected.

[0101] A7: Perform stress loss prediction model training and optimization, adopt multi-model initialization, including 13 models: including 6 tree models: XGBoost, LightGBM, CatBoost, RandomForest, GradientBoosting, ExtraTrees, 4 linear models: Lasso, Ridge, ElasticNet, BayesianRidge, 2 kernel methods: SVR, KernelRidge, 1 neural network: MLPRegressor.

[0102] Training takes parallel training and evaluation calculation, evaluation indicators are training set R²: model fitting degree on training set; validation set R²: model generalization ability on validation set; training set MSE: training set mean square error; validation set MSE: validation set mean square error; cross-validation R²: average score of 3-fold cross-validation; training time: model training time, wherein the main selection basis is validation set R², auxiliary standards are overfitting degree, cross-validation stability, and training time efficiency. The model data table is shown in Table 1 below.

[0103] Table 1 Comparison of multi-model training results

[0104]

[0105] It is concluded that the GradientBoosting model performs best, and its hyperparameters are optimized. The optimization method is RandomizedSearchCV, which randomly explores 20 points in the preset parameter space. The parameter space is: number of trees (500-1200): controls model complexity, learning rate (0.005-0.05): adjusts learning step size to avoid missing the optimal solution, tree depth (4-10 layers): balances model expression ability and overfitting risk, sample sampling rate (70%-100%): proportion of data used each time, split condition: stopping condition for tree growth, evaluate the effect of each parameter group through 3-fold cross-validation, select the parameter combination with the highest average R² score, and then retrain the best model using the complete training set.

[0106] A8: Perform ensemble learning on the model, adopt a two-layer Stacking ensemble structure, the first layer of base learners selects the four models with the best performance in S5 as base learners, which are Gradient Boosting Tree (GradientBoosting), Random Forest (RandomForest), Extremely Randomized Trees (ExtraTrees), and Extremely Gradient Boosting Tree (XGBoost), ensuring the balance between technical diversity and superior performance.

[0107] Four meta-learners are combined, which are random forest, gradient boosting tree, ridge regression, and Lasso regression. Through 3-fold cross-validation, the meta-learner with the highest R² score is selected: the best meta-learner .

[0108] where, : find the parameter value that makes the function take the maximum value, m : a single meta-learner, M : the set of all meta-learner candidates, CV-R 2 : cross-validation R² score.

[0109] Through out-of-fold prediction, the training data is divided into 3 mutually exclusive subsets. For each base learner, 2 / 3 of the data is used for training, and the remaining 1 / 3 is used for prediction. This process is repeated 3 times to ensure that each sample has a prediction value. These predictions are used as the training features for the meta-learner, avoiding data leakage that can cause severe overfitting.

[0110] Final prediction formula: where, is the combined function learned by the meta-learner, is the prediction of the ensemble model.

[0111] For linear meta-learners: .

[0112] where , , and are the weights corresponding to different base learners, is the bias term, is the prediction value of the GradientBoosting model, is the prediction value of the RandomForest model, is the prediction value of the ExtraTrees model, is the prediction value of the XGBoost model.

[0113] A9: Perform final evaluation and output on the above model results. By comparing the R² determination coefficient and R² difference of the test set and training set of the single best model and the ensemble model, while considering RMSE, MAE, stability, and overfitting degree, the comparison data of each model is shown in Table 2:

[0114] Table 2 Comparison data of each model

[0115]

[0116] The performance of the Lasso ensemble model is very close to that of the GradientBoosting single model, but the R² difference of the former is less than that of the latter, indicating that the Lasso ensemble model has a lower risk of overfitting and better stability and reliability in generalizing to unknown data. The larger gap in GradientBoosting means that it may be more affected by the noise of the training data. When facing new data slightly different from the training data distribution, the performance of the Lasso ensemble model may be more robust.

[0117] Finally, GradientBoosting, RandomForest, ExtraTrees, and XGBoost are selected as base learners, and Lasso is selected as a meta-learner as the prediction model for prestress loss of PC box girder under fire.

[0118] The best ensemble model Lasso meta-learner introduces an L1 regularization term based on ordinary least squares, and its optimization objective can be expressed as: , where the regularization term is composed of the L1 norm, i.e. . is the regularization hyperparameter in Lasso regression, used to control the penalty strength; is the coefficient vector to be solved; is a column vector containing all sample true target values; X is the input feature matrix of the meta-learner, which has the following structure:

[0119]

[0120] In this matrix, each column represents the predicted value of a base learner (such as GradientBoosting, RandomForest, etc.) for n samples.

[0121] The solution process of Lasso regression tends to produce a sparse solution. The final prediction model can be represented as:

[0122]

[0123] After optimization, the optimal parameter vector is mapped to the parameter name, and the best parameter combination is output, obtaining the optimized ensemble model as the optimal algorithm for predicting prestress loss of PC box girder under fire.

[0124] S8: Perform performance evaluation, result visualization processing, and SHAP explainability analysis on the optimized model, and apply it to PC box girder prestress damage identification and prediction under fire.

[0125] As shown in Tables 1 and 2, the performance of the test set is quantified by the R² index, and the prediction effect is visually displayed by means of double-panel visualization. As shown inFigure 3 , Figure 4 As shown, Figure 3 (a), (b), (c), (d), and (e) are the R values ​​of the prestress loss of PC box girders under fire conditions based on various models in this invention. 2 Performance comparison chart, RMSE comparison chart, overfitting comparison chart, training time comparison chart, stability comparison chart; Figure 4 In Figures (a) and (b), the prestress loss of PC box girders under fire in this invention is based on GradientBoosting, RandomForest, ExtraTrees, XGBoost as the base learner and Lasso as the meta learner. The distribution trends of the predicted values ​​and the true values ​​and the ideal reference line are compared through the scatter plots.

[0126] SHAP interpretability analysis was performed. An interpreter was built using the SHAP library. SHAP values ​​were calculated on the trained model to analyze the contribution of each feature to fire-induced prestress loss and flexural capacity loss, and plotted as follows: Figure 6 , Figure 7 SHAP bee colony graph and feature dependency graph.

[0127] (2) Optimize the hyperparameters of the model corresponding to the loss of bending bearing capacity, including the following steps:

[0128] For the flexural bearing capacity prediction model, its integrated architecture is Stacking (two-layer integration). Its base learner layer adopts XGBoost + SVR, and its meta learner layer adopts Ridge Regression. The base model captures different data patterns, and the meta model integrates the prediction results of the base model to improve the generalization ability.

[0129] B1: The Optuna framework is used to optimize the hyperparameters of XGBoost, SVR base learners, and Ridge meta learners using the TPE algorithm.

[0130] Hyperparameter optimization of the algorithm corresponding to the flexural bearing capacity loss model was performed using the Optuna framework for multi-objective joint search. Optuna employed the TPE algorithm for Bayesian optimization, modeling the hyperparameter probability distribution based on historical data after each trial: for trials with good performance ( l ( x The test of difference ()) and difference () g ( x Fit probability density functions respectively, and select l ( x ) / g ( x) to perform the next round of experiments. Overall, the optimization range covers preprocessing strategies (degree of polynomial feature expansion), base model parameters (number of trees, depth, learning rate for XGBoost, regularization coefficient, kernel function selection for SVR) and meta-model parameters (regularization strength for Ridge regression). At the same time, 5-fold cross-validation is nested in the optimization process: within each fold, non-linear features are generated by polynomial expansion and standardized, and then XGBoost and SVR models are trained respectively. Further, when training XGBoost, an early stopping mechanism is equipped to prevent overfitting.

[0131] The Ridge meta-learner introduces an L2 regularization term based on ordinary least squares, and the optimization objective is:

[0132]

[0133]

[0134] where, is the regularization hyperparameter in Ridge regression, is the coefficient vector to be solved in Ridge regression model, is a column vector containing all sample true target values, is the input feature matrix of the meta-learner, is the prediction value of the XGBoost model for samples, is the prediction value of the SVR model for samples;

[0135] The final prediction value of the Ridge meta-learner is:

[0136]

[0137] where, is the final prediction value of the Ridge meta-learner, is the prediction value of the XGBoost model, is the prediction value of the SVR model, and are weight coefficients representing the contribution proportion of XGBoost and SVR prediction values in the meta-learner.

[0138] Optuna performs Bayesian optimization through the TPE algorithm, modeling the probability distribution of hyperparameters after each experiment: fitting probability density functions for good experiments ( l ( x )) and poor experiments ( g ( x )), and selecting l the hyperparameters with the highest probability.x ) / g ( x ) most hyperparameters to the next round of experiments;

[0139] Objective: find the hyperparameters such that: where f ( x ) is the score of the model on the validation set x under hyperparameters ;

[0140] Mathematical expression: ,

[0141] where l ( x )) is the probability density of good experiments, ( g ( x )) is the probability density of bad experiments, is the current best score, and we select P ( x | y ) most hyperparameters x to the next round of experiments.

[0142] Pruning condition: if the score of the current experiment satisfies: terminate the experiment, where is the score of the current experiment at step t , and is the set of scores of all historical experiments at step t ;

[0143] Add early stopping mechanism to define dynamic convergence criteria: terminate training when the validation loss changes less than a threshold m = 1e-4 for = 5 consecutive iterations; where is the loss function value on the validation set at the k th iteration, is the loss function value on the validation set at the th iteration, m is the window size for early stopping detection.

[0144] B2: Based on the optimization results, the prediction results of XGBoost and SVR base learners are spliced into meta features, a Ridge meta-learner is trained, and the MedianPruner pruning strategy is used to dynamically eliminate inefficient parameter combinations to obtain the globally optimal parameters.

[0145] The prediction results of the two are spliced as meta-features, and then a Ridge meta-model is trained. The R² score in the validation set is used as the optimization target, and the MedianPruner pruning strategy is used to dynamically eliminate inefficient parameter combinations. Finally, the global optimal parameters are determined through a specified number of parallel experiments (200 parallel experiments are performed in this model).

[0146] B3: Based on the global optimal parameters, a preprocessing pipeline and a base model are constructed, and a Ridge meta-model is finally trained to form a two-layer stacked structure.

[0147] The best parameter configuration obtained from the model construction stage is reused for optimization, and the preprocessing pipeline and base model are reconstructed. Finally, the Ridge meta-model is trained on the fused meta-features, and L2 regularization is used to constrain the complexity, forming a two-layer stacked structure.

[0148] To avoid data leakage during meta-feature generation, the 5-fold cross-validation strategy is still used: within each fold, the base model only uses the training subset to fit, and the validation subset is used to predict and fill the meta-feature matrix; the prediction of the test set is achieved by weighted average of multiple fold results to improve stability.

[0149] S5: Performance evaluation, result visualization, and SHAP explainability analysis are performed on the optimized model, and it is applied to PC box girder damage identification prediction under fire.

[0150] As shown in Table 3, the performance of the test set is quantified by R², MAE, and RMSE, and the prediction effect is visually displayed by means of double-panel visualization. As shown in Figure 5 , Figure 5 Figures (a), (d), and (e) are scatter plots, residual distribution scatter plots, and residual frequency distribution plots of the bending capacity loss of PC box girders under fire based on XGBoost and SVR as base learners and Ridge as meta-learners, (b) and (c) are scatter plots based on XGBoost and SVR. By comparing the distribution trend of the predicted value and the true value with the ideal reference line, the overall fitting ability of the model is reflected; the residual plot reveals the distribution pattern of the prediction error, which helps to diagnose the model bias.

[0151] Table 3 Training results of PC box girder bending capacity loss under fire

[0152]

[0153] SHAP explainability analysis is performed, and the SHAP library is used to build an interpreter to calculate the SHAP values of the trained model. The contribution of each feature to the fire-induced prestress loss and bending capacity loss is analyzed, and the SHAP value is plotted as Figure 6 、 Figure 7SHAP swarm plot and feature dependence plot.

[0154] The above is the procedure of the damage assessment operation processing. Further, a monitoring and early warning system as shown in Figure 8 、 Figure 9 synchronously used.

[0155] Embodiment 2, a monitoring and early warning system for damage identification and prediction method of PC box girder under fire, comprising:

[0156] Temperature sensor, laid in the key parts of the bridge, used for monitoring temperature change, positioning the fire area;

[0157] Infrared camera, used for identifying the location, height and range of the fire source, and recording the time of fire;

[0158] Storage box, built-in data transceiver processing device and detection drone, when the data transceiver processing device receives temperature anomaly and detects fire characteristics, automatically opens the valve of the box to release the detection drone;

[0159] Detection drone, equipped with infrared camera and data transmission device, used for multi-angle data collection;

[0160] Central processing unit, used for running the damage identification and prediction model, outputting the prediction results, including the prestressed damage and bending capacity loss identification results under fire;

[0161] Alarm system, including LED display screen, warning light and loudspeaker, used for fire warning and information broadcast;

[0162] Solar power generation device, providing power support for the system.

[0163] As shown in Figure 8 and Figure 9As shown, the monitoring and early warning system includes a temperature sensor 8, an infrared camera 2, a solar power generation device 3, a storage box 6, a detection drone 7, and an alarm system. The alarm system includes an LED display screen 1, a warning light 4, and a loudspeaker 5, which further locates the fire location, the fire range, the explosion situation, the fire type, and the fire situation. The temperature sensor 8 is arranged on the bridge and is responsible for detecting abnormal temperature and locating the bridge position information of the temperature change part caused by fire. The infrared camera 2 can rotate 360° and is responsible for detecting and identifying the fire source, providing the fire source position, height, and range information, and recording the time effect, cooperating with the temperature sensor 8 to record the fire time and the fire source position. The storage box 6 stores the detection drone 7, and is arranged with a data transceiving and processing device 601, which facilitates data transmission between the detection device and the central processor 9. The above monitoring equipment can be but is not limited to a detection drone 7, which is automatically started by the temperature sensor 8 detecting abnormal temperature data, flies to the specific fire location according to the fire source position information, and detects and reports the situation.

[0164] The detection drone 7 carries four small infrared cameras 701, records the fire situation, and facilitates subsequent identification of the fire type and auxiliary positioning of the fire source position. The small infrared camera 701 can rotate 360°, and multiple cameras can faster shoot and cover the disaster area, shoot the explosion situation, and facilitate subsequent identification of the explosion degree. The detection drone 7 carries a data storage and transmission device 702.

[0165] The solar power generation device 3 provides economic, environmentally friendly, and sustainable power for the system.

[0166] The alarm system includes an LED display screen 1, a warning light 4, and a loudspeaker 5. After the above devices detect the fire characteristics, the fire situation is quickly uploaded to the fire department through the data transceiving and processing device 601, and the warning light 4 and the loudspeaker 5 are turned on to prompt the occurrence of the fire and remind passing vehicles and pedestrians to avoid causing secondary disasters. The content played by the LED display screen 1 can include but is not limited to the fire warning prompt wordings in front, the distance from the fire point, and the fire time.

[0167] In one embodiment of the present application, the operation process of the above monitoring and early warning system is as follows:

[0168] 1) In the system initialization stage, the initial information of the bound bridge is input, including but not limited to the bridge type, the number of spans and the bridge span, the bridge height, the cross-sectional width and thickness in the middle of the span, the cross-sectional steel beam position in the middle of the span, the longitudinal reinforcement strength, and the protection layer thickness;

[0169] 2) The monitoring and early warning system is arranged, and the number of system devices can be arranged as needed according to the bridge length;

[0170] 3) When the temperature sensor 8 arranged on the bridge detects abnormal temperature, the data transceiver processing device 601 in the storage box 6 receives the signal, and the infrared camera 2 above rotates to the specified fire position to assist in verifying whether the temperature anomaly is caused by fire;

[0171] 4) When the temperature anomaly caused by fire is detected, the fire time is recorded synchronously, the valve of the storage box 6 is opened to release the detection drone 7 wearing the small infrared camera 701 and the data storage and transmission device 702, and the warning light 4, the loudspeaker 5 and the LED display screen 1 are opened synchronously to prompt the vehicles and pedestrians passing by the bridge that the fire accident has occurred;

[0172] 5) The detection drone 7 synchronously transmits the detection data information to the central processor 9, and the detection information includes but is not limited to the fire range, the burst condition, the fire type and the fire time;

[0173] 6) The central processor 9 predicts the prestress and bending capacity loss under fire, provides reference for rescue under fire and post-fire repair evaluation, and reports the related information to the fire department.

[0174] The central processor 9 is equipped with a trained image recognition processing system, which can be realized by models such as Faster R-CNN, YOLO series and SSD, so as to efficiently and quickly identify the burst degree, fire type and range. At the same time, it carries the above-mentioned PC box girder fire damage identification and prediction method, which can efficiently and accurately predict the prestress damage identification and bending capacity loss under fire, provide reference for bridge fire extinguishing and rescue measures under fire, avoid property loss and personnel casualty when the bridge structure is about to be damaged, and also provide reference value for post-fire bridge repair work and assist in post-fire bridge disaster evaluation work.

[0175] Those skilled in the art will appreciate that the embodiments described herein are presented to assist the reader in understanding the principles of the present application and should be understood as not limiting the scope of protection of the present application to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations according to the technical inspiration disclosed in the present application without departing from the essence of the present application, and these modifications and combinations are still within the scope of protection of the present application.

Claims

1. A method for damage identification and prediction of PC box girders under fire conditions, characterized in that, Includes the following steps: S1: Select the characteristics of prestress loss and flexural bearing capacity loss of PC box girder in fire, including basic bridge information, cross-sectional geometric parameters, material properties and fire scenario; S2: Based on the selected features, construct a database of prestress loss and flexural bearing capacity loss of PC box girders under fire conditions; S3: Preprocess various raw data in the prestress loss database and flexural bearing capacity loss database, convert each parameter into a format that can be analyzed by machine learning models, and divide the dataset into training set and test set; S4: Based on the training set and the test set, optimize the hyperparameters of the models corresponding to prestress loss and flexural bearing capacity loss respectively; The hyperparameters of the model corresponding to prestress loss are optimized, including the following steps: A1: Based on the training set and test set, the original data features of prestress loss are processed, and polynomial features are generated using the original features; A2: Describe the feature distribution space of each sample from a statistical perspective and extract statistical features; A3: Based on the statistical feature extraction results, multiply the top ten important features in pairs to create interactive features and generate the number of interactive features; A4: Combine all generated features from the combination of multinomial features, statistical features, and interactive features into a matrix in sequence, while simultaneously expanding the features; A5: Based on the feature expansion results, perform feature data standardization; A6: Based on the standardized results of the feature data, feature selection is performed. Random forest is used as the evaluator to calculate the importance score of each feature and select important features according to the threshold. A7: Based on key features, the prestress damage prediction model is trained in parallel through multi-model initialization and evaluated to obtain the best-performing model, and its hyperparameters are optimized. A8: The model is integrated for learning. A two-layer stacking ensemble structure is adopted. The best-performing models GradientBoosting, RandomForest, ExtraTrees and XGBoost are selected as base learners. Four meta-learners, Random Forest, Gradient Boosting Tree, Ridge Regression and Lasso Regression, are selected and combined. Through 3-fold cross-validation, the ensemble model with the highest comprehensive evaluation score is selected as the best ensemble model. A9: Comparing the performance of single best models and ensemble models, the ensemble model based on GradientBoosting, RandomForest, ExtraTrees and XGBoost as base learners and Lasso regression as meta learner is used as the prediction model for prestress loss of PC box girders under fire. The hyperparameters of the model corresponding to the loss of bending bearing capacity are optimized, including the following steps: B1: The Optuna framework is used to optimize the hyperparameters of XGBoost, SVR base learners, and Ridge meta learners using the TPE algorithm. B2: Based on the optimization results, the prediction results of XGBoost and SVR base learners are concatenated into meta-features, the Ridge meta-learner is trained, and the MedianPruner pruning strategy is combined to dynamically eliminate inefficient parameter combinations to obtain the globally optimal parameters. B3: Based on the globally optimal parameters, a preprocessing pipeline and base model are constructed, and finally the Ridge meta-model is trained to form a two-layer stacked structure; S5: The optimized model is evaluated for performance, the results are visualized and analyzed using SHAP interpretability, and then applied to damage identification and prediction of PC box girders under fire conditions.

2. The method for identifying and predicting damage to PC box girders under fire according to claim 1, characterized in that, The optimization objective of the Lasso regression is: in, This is a column vector containing the true target values ​​of all samples. The input feature matrix of the meta-learner. To find the coefficient vector, For regularization hyperparameters in Lasso regression, For GradientBoosting model The predicted value for each sample, For the RandomForest model The predicted value for each sample, For the ExtraTrees model The predicted value for each sample, For XGBoost model pairs Predicted values ​​for each sample; The final prediction model for Lasso regression is: in, For the final prediction model results, , , and To correspond to the weights of different base learners, For bias terms, These are the predicted values ​​from the Gradient Boosting model. These are the predicted values ​​from the RandomForest model. These are the predicted values ​​from the ExtraTrees model. These are the predicted values ​​from the XGBoost model.

3. The method for identifying and predicting damage to PC box girders under fire according to claim 2, characterized in that, The Ridge meta-learner introduces an L2 regularization term based on ordinary least squares, with the optimization objective being: in, For ridge regression, the regularization hyperparameter, For SVR model pair Predicted values ​​for each sample; The final prediction value of the Ridge meta-learner is: in, This represents the final prediction value of the Ridge meta-learner. These are the predicted values ​​from the XGBoost model. These are the predicted values ​​from the SVR model. and These are the weighting coefficients.

4. A monitoring and early warning system for the PC box girder fire damage identification and prediction method as described in any one of claims 1-3, characterized in that, include: Temperature sensors are deployed at key parts of the bridge to monitor temperature changes and locate fire-affected areas. Infrared cameras are used to identify the location, height, and range of fire sources and record the time the fire has lasted; The storage box contains a built-in data transceiver and processing device and a detection drone. When the data transceiver and processing device receives a temperature abnormality and detects fire characteristics, it automatically opens the box valve and releases the detection drone. The drone is equipped with an infrared camera and a data transmission device for multi-angle data acquisition. The central processing unit is used to run the damage identification and prediction model and output the prediction results, including the identification results of prestress damage and flexural bearing capacity loss under fire. Alarm systems, including LED displays, warning lights, and speakers, are used for fire warning and information broadcasting; Solar power generation devices provide electricity to the system.

Images