Machine learning-based hollow slab bridge disease time sequence prediction method and system

Abstract

The invention provides a hollow slab bridge disease time sequence prediction method and system based on machine learning, and is applied to the technical field of bridge disease analysis and prediction.The method comprises the steps that hollow slab bridge detection data is acquired, and an original data set is constructed according to the disease data; performing structured preprocessing on the original data; feature engineering is carried out on the structured data, random forest screening is utilized, feature correlation analysis is carried out, and a low-redundancy high-correlation feature set is obtained; dividing the bridge detection data corresponding to the low-redundancy high-correlation features into training samples according to a time sliding window, and constructing a time sequence sample set; and carrying out model training and optimization, training a time sequence sample, and predicting continuous values of future diseases. According to the method, the accuracy, precision and stability of bridge disease prediction are improved through data acquisition, preprocessing, feature optimization, model establishment and output of the dynamic change prediction result of the disease continuous index, the efficiency is higher, and the method has a good application prospect.

Description

Technical Field

[0001] The present application relates to the technical field of bridge defect analysis and prediction, and specifically to a method and system for time series prediction of hollow slab bridge defects based on machine learning. Background Art

[0002] Hollow slab bridges, the primary structural form for small and medium-span bridges, are widely used in highway and municipal road systems. Under the combined effects of long-term environmental erosion, vehicle loads, and natural aging, these bridges can experience continuous deterioration, including crack propagation, increased spalling, increased steel corrosion, accumulated bearing deformation, and increased damage to expansion joints, in both the superstructure and substructure, as well as in attached components. In severe cases, these conditions can compromise the overall safety and durability of the bridge structure.

[0003] Currently, the method for predicting bridge defects can adopt statistical modeling methods based on regular inspection data. Although the cost is low, most of them only perform static assessments, and the prediction accuracy and stability are poor.

[0004] Based on this, a new technical solution is needed. Summary of the Invention

[0005] In view of this, the present application provides a method and system for time series prediction of hollow slab bridge defects based on machine learning.

[0006] This application provides the following technical solutions:

[0007] According to the present application, a method for predicting hollow slab bridge defects based on machine learning is provided, which includes the following steps:

[0008] Step S1: Obtain hollow slab bridge inspection data and construct an original data set based on the damage data of the inspection data;

[0009] Step S2: perform structured preprocessing on the original data;

[0010] Step S3: Perform feature engineering on the structured data, use random forest to screen, and perform feature correlation analysis to obtain a low-redundancy and high-correlation feature set;

[0011] Step S4: Divide the bridge inspection data corresponding to the low-redundancy and high-correlation features into training samples according to the time sliding window to construct a time series sample set;

[0012] Step S5: Perform model training and optimization, train the time series samples, and predict the continuous values ​​of future diseases.

[0013] Preferably, in step S1, the hollow slab bridge data in the regular bridge inspection report is obtained, and the inspection timestamp is retained for the disease data, and an original data set containing disease information and timestamp is constructed; wherein the hollow slab bridge data includes basic bridge information, geometric property data, material property data, environmental factor data and continuous disease indicators.

[0014] Preferably, in step S2, the original data is subjected to text keyword analysis, outlier removal, missing value filling, and unified data type and dimension, and converted into structured numerical input data;

[0015] Among them, missing value filling adopts a linear interpolation method based on time series, and interpolation operations are performed independently on each feature dimension to fill the missing continuous value features; statistical methods are used to identify and eliminate outliers in the data.

[0016] Preferably, in step S3, a random forest algorithm is used to calculate the Gini coefficient of each feature for importance ranking, a preset number of features are selected before ranking, and a correlation coefficient matrix is ​​calculated. Feature pairs with correlation greater than a preset threshold are screened, and features with high Gini coefficients are retained to form a low-redundancy, high-correlation feature set.

[0017] Preferably, in step S4, the bridge inspection data is divided into training samples according to a time sliding window, and the model is trained using the data of the previous several periods to predict the continuous value of the disease development in the next period, and a time series framework of a preset time is constructed;

[0018] Among them, a time series sample is constructed by setting a fixed time window. When constructing a time series sample, each input sample contains feature vectors of multiple historical time points, and the output is a continuous disease indicator at the target moment.

[0019] Preferably, in step S5, the optimal hyperparameter combination of the distributed gradient boosting library model is determined by random search, and regularization is introduced to control the complexity of the model during the model iteration process. At the same time, the changing trend of the validation set loss is judged by the early stopping mechanism, and the training is terminated early when there is no improvement for several consecutive rounds.

[0020] Preferably, the distributed gradient boosting library model performs predictions by integrating multiple decision tree models. Each tree is constructed based on the residual of the previous tree. The goal is to continuously optimize the model performance through gradient boosting to obtain the minimized loss function.

[0021] Among them, multiple optimized decision tree models are obtained through training and integrated into the distributed gradient boosting library model framework; during the model training process, based on the gradient boosting strategy, with the goal of minimizing the loss function, several subtrees are iteratively constructed. In each round of iteration, the prediction residual of the current model is used to train a new regression tree to correct the error. In each round of training, the square error is minimized by updating the weight of the tree; in the model construction stage, the key hyperparameters of the distributed gradient boosting library model are tuned through random search or grid search methods to ensure that the model achieves the optimal balance between accuracy and generalization ability; after multiple rounds of iterative training and optimization, a distributed gradient boosting library integrated learning model that integrates multiple weak learners is obtained.

[0022] Preferably, the distributed gradient boosting library model evaluation uses mean square error, mean absolute error and coefficient of determination as evaluation indicators;

[0023] Among them, the mean square error is used to measure the average square difference between the predicted value and the true value; the mean absolute error is used to measure the average absolute difference between the predicted value and the true value; the coefficient of determination is used to measure the model's ability to explain the variability of the data, indicating the degree of fit between the model's predicted value and the actual observed value.

[0024] Preferably, the prediction method further includes step S6: outputting the bridge disease prediction results at future time points, feeding subsequent detection data back into the training set, building a closed-loop mechanism of prediction, verification, and updating, and continuously updating and optimizing the model.

[0025] According to the present application, a time series prediction system for hollow slab bridge defects based on machine learning is provided, comprising the following steps:

[0026] Module M1: Obtain hollow slab bridge inspection data and construct the original data set based on the disease data of the inspection data;

[0027] Module M2: structured preprocessing of raw data;

[0028] Module M3: Perform feature engineering on structured data, use random forest to screen, and perform feature correlation analysis to obtain a low-redundancy and high-correlation feature set;

[0029] Module M4: divide the bridge inspection data corresponding to low-redundancy and high-correlation features into training samples according to the time sliding window to construct a time series sample set;

[0030] Module M5: Perform model training and optimization, train time series samples, and predict the continuous values ​​of future diseases.

[0031] Compared with the prior art, the at least one technical solution adopted in this application can achieve the following beneficial effects:

[0032] This application is based on a time-series prediction method for hollow slab bridge defects based on historical inspection report data. Through data collection, preprocessing, and feature optimization, a model is established to output dynamic change prediction results of continuous defect indicators, thereby improving the accuracy, precision, and stability of bridge defect prediction, making it more efficient and having good application prospects. It fully combines machine learning modeling technology with bridge infrastructure operation and maintenance scenarios for bridge health monitoring, defect prediction, and maintenance management, and can realize dynamic modeling and prediction of bridge defect development trends, providing scientific and effective decision-making basis for preventive maintenance, repair priority sorting, and resource allocation of bridge management units. It has good engineering application prospects and social and economic value. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0034] Figure 1 is a flow chart of a bridge damage prediction method in an embodiment of the present application;

[0035] Figure 2 This is a flowchart of the bridge disease data preprocessing process in an embodiment of the present application;

[0036] Figure 3 This is a flow chart of feature engineering and redundancy elimination in the embodiment of this application;

[0037] Figure 4 This is a schematic diagram of constructing a time series sample in an embodiment of the present application;

[0038] Figure 5 This is a flowchart of the XGBoost model training process in the embodiment of this application. DETAILED DESCRIPTION

[0039] The embodiments of the present application are described in detail below with reference to the accompanying drawings.

[0040] The following describes the embodiments of the present application through specific examples, and those skilled in the art can easily understand other advantages and effects of the present application from the contents disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. The present application can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present application. It should be noted that, in the absence of conflict, the features in the following embodiments and embodiments can be combined with each other. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without making creative work are within the scope of protection of this application.

[0041] It should be noted that various aspects of the embodiments within the scope of the appended claims are described below. It should be apparent that the aspects described herein can be embodied in a wide variety of forms, and any specific structure and / or function described herein is merely illustrative. Based on this application, it should be understood by those skilled in the art that an aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number and aspect described herein can be used to implement an apparatus and / or practice a method. In addition, other structures and / or functionalities other than one or more of the aspects described herein can be used to implement this apparatus and / or practice this method.

[0042] It should also be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present application. The illustrations only show components related to the present application and are not drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component can be changed at will, and the component layout type may also be more complicated.

[0043] Additionally, in the following description, specific details are provided to provide a thorough understanding of the examples, however, one skilled in the art will appreciate that the examples can be practiced without these specific details.

[0044] In view of this, the applicant conducted in-depth research and improved exploration on bridge defect prediction methods and found that the current bridge defect prediction methods mainly include two categories: (1) intelligent perception systems based on sensors and image recognition; (2) statistical modeling methods based on regular inspection data. The former requires the deployment of high-cost sensors and visual acquisition equipment, which is complex to maintain and has a limited scope of application. Although the latter is less expensive, most of them only perform static assessments and only implement static assessments, lack dynamic predictions for time evolution, and have not formed the ability to form a time series dynamic modeling of continuous disease indicators.

[0045] Based on this, the technical solutions provided by the various embodiments of the present application are described below in conjunction with the accompanying drawings.

[0046] The embodiment of this specification proposes a time series prediction method for hollow slab bridge defects based on machine learning, such as Figure 1 As shown, the following steps are included:

[0047] Step S1: Obtain hollow slab bridge inspection data and construct an original dataset based on the damage data from the inspection data. Hollow slab bridge data from historical bridge inspection reports is obtained. This data includes basic bridge information, geometric characteristics, material characteristics, environmental factors, and damage information. The damage information can include continuous damage indicators and detailed bridge damage information. Inspection timestamps are retained for the damage data to construct a time-series dataset covering multiple structural defects and components.

[0048] Step S2: Perform structured preprocessing on the raw data. The raw data is preprocessed, including text keyword analysis, outlier removal, missing value filling, and data type and dimension unification, and converted into structured numerical input data.

[0049] Step S3: Feature engineering is performed on the structured data, using random forests for screening and feature correlation analysis to obtain a low-redundancy, high-correlation feature set. The Gini coefficient of each feature is calculated using the random forest algorithm to rank the importance. The top-ranked features are selected and their Spearman correlation coefficient matrix is ​​calculated. For feature pairs with correlations greater than a preset threshold, only the one with higher importance is retained to form a low-redundancy, high-correlation feature set.

[0050] Step S4: Divide the bridge inspection data corresponding to low-redundancy, high-correlation features into training samples according to the time sliding window to construct a time series sample set. The bridge inspection data is divided into training samples according to the time sliding window. The model is trained using data from several previous periods to predict the continuous value of the disease development in the next period, constructing a time series framework from "tn to t+1".

[0051] Step S5: Perform model training and optimization. Train on time series samples to predict the continuous values ​​of future diseases. Use the XGBoost model for regression training to predict the continuous values ​​of future diseases. Use random search to determine the optimal hyperparameter combination for the model. L2 regularization is introduced during model iteration to control model complexity. Early stopping is used to determine the trend of validation set loss. Training is terminated early if there is no improvement after several rounds.

[0052] In one embodiment, in step S1, hollow slab bridge data from a regular bridge inspection report is obtained, and an inspection timestamp is retained for the damage data to construct an original data set containing damage information and timestamps; wherein the hollow slab bridge data includes basic bridge information, geometric property data, material property data, environmental factor data, and continuous damage indicators.

[0053] Defect information includes, but is not limited to, transverse cracks in the bottom slab, longitudinal cracks in the top slab, oblique cracks in the web, steel bar corrosion rate, hinged joint detachment area, bearing aging and cracking, pier cracks, expansion joint damage, and drainage system blockage. This example is based on regular inspection reports from 2012 to 2023 for over 100 hollow slab bridges in Guangzhou. Raw data for hollow slab beam bridges is extracted from these regular inspection reports, and timestamps are retained based on the inspection time. This constructs a dataset of time-series evolution paths for multiple defects and components in bridge structures.

[0054] The disease data is based on the continuous quantitative indicators extracted from the "Highway Bridge Technical Condition Assessment Standard", such as the length of the bottom plate transverse crack (unit: mm), the width of the top plate longitudinal crack (unit: mm), the steel bar corrosion rate (unit: %), the hinge joint detachment area (unit: m 2 ) etc. That is, the defect information is extracted based on the "Highway Bridge Technical Condition Assessment Standard" to extract quantitative indicators, including but not limited to the type, length, width, spalling area, steel corrosion rate, hinged joint damage degree and other defect values, and a timestamp is attached to construct a time series data set. Unstructured text data is converted into quantitative indicators through keyword parsing. For example, "partial detachment of hinged joint" is parsed as "the proportion of hinged joint detachment area ≥5%", and "network cracks" is parsed as "crack density ≥0.2 / m 2 These quantitative indicators will serve as input features of the model.

[0055] In one embodiment, in step S2, the original data is subjected to text keyword analysis, outlier elimination, missing value filling, and unified data type and dimension, and converted into structured numerical input data; wherein, the missing value filling adopts a linear interpolation method based on time series, and the interpolation operation is independently performed on each feature dimension to fill the missing continuous value features; and a statistical method is used to identify and eliminate outliers in the data.

[0056] The acquired data is preprocessed by parsing text keywords, filling missing values, normalizing, and removing outliers. Disease evolution indicators are constructed based on multi-period data to enhance the model's sensitivity to temporal changes. Missing values ​​are filled using a linear interpolation method based on time series, performing independent interpolation operations on each feature dimension. Missing continuous-valued features are filled using interpolation methods. For numerical features, their original structure is retained and standardized or normalized to ensure data dimensional consistency and model input stability.

[0057] Specifically, the original data must be preprocessed before the data is obtained, including outlier removal, missing value processing, data type unification and dimension unification. After preprocessing, the bridge inspection data set is obtained. The data set contains multiple features of numerical and textual data. Therefore, feature selection is required before prediction, and the feature data must be processed, including unifying the data format and deciding whether to unify the dimension based on the nature of the algorithm. The specific data preprocessing process is as follows: Figure 2 shown.

[0058] Statistical methods are used to identify and remove outliers in the data. This example uses the Z-score method to detect outliers on continuous primitive features, such as disease magnitude, crack length, detachment area, and corrosion depth. Specifically, the mean of each feature is set to μ and the standard deviation is set to σ. For any data point X, the standard score is calculated as follows:

[0059]

[0060] Among them, Z is the standard score, which represents the distance between the data point and the mean in standard deviation units. If Z>α, where α is the set threshold, it is considered an outlier and is eliminated.

[0061] For bridge feature items with missing years, a linear interpolation method based on time series is used to fill the missing values. In the multidimensional feature space, the interpolation process is performed independently on each dimension. Assuming that the Kth feature is missing at time point t, the two adjacent valid values ​​y before and after it are used. t-1 and y t+1 Linear interpolation is performed to estimate missing values. The interpolation calculation formula is as follows:

[0062]

[0063] where y t is the estimated value at time point t, y t-1 and y t+1 are the actual observation values ​​at two valid time points before and after.

[0064] In cases where both numerical and categorical features coexist in scheduled inspection data, one-hot encoding and mapping encoding are used to unify the data type format. Numerical features, such as bridge length, age, traffic volume, crack length, and rating, are directly involved in model calculations and undergo standardization or normalization to ensure consistent scales across features, thereby minimizing the impact on model training. Categorical features, such as structural type, upper material type, and bridge location grade, are processed using one-hot encoding. Ordinal features, such as highway grade and technical condition rating, are mapped using mapping encoding, with integer mapping values ​​from 0 to N assigned according to grade.

[0065] In order to eliminate the offset effect of different dimensions on model training, all numerical features are standardized. The standardized feature value X ′ The calculation is as follows:

[0066]

[0067] Among them, X is the original eigenvalue, μ and σ are the mean and standard deviation of the feature samples in the training set, respectively.

[0068] After standardization, all features have a distribution characteristic with a mean of 0 and a standard deviation of 1, ensuring the comparability and stability of the model input dimensions.

[0069] In one embodiment, in step S3, a random forest algorithm is used to calculate the Gini coefficient of each feature to sort the importance, select a preset number of features before sorting, and calculate the correlation coefficient matrix. Feature pairs with correlation greater than a preset threshold are screened, and features with high Gini coefficients are retained to form a low-redundancy, high-correlation feature set.

[0070] Specifically, random forest is used to calculate the Gini coefficient of each feature to rank the importance. After retaining the top S features, their Spearman correlation coefficient matrix is ​​calculated. For feature pairs with correlation coefficients greater than the threshold, only those with higher Gini coefficients are retained to obtain a low-redundancy, high-correlation structural evolution feature set.

[0071] After completing preprocessing, in order to improve the efficiency of model training and reduce the redundancy between input variables, the feature space needs to be further screened and optimized.

[0072] The random forest algorithm is used to calculate the Gini importance index of each feature in the training set and sort them from high to low according to the score. The top S features are selected to enter the next step of correlation analysis. The Spearman rank correlation coefficient between the above S high-importance features is calculated. If the correlation coefficient of a feature pair is greater than the set threshold, the feature with more business interpretation or computational stability is retained, and redundant features are eliminated to reduce the risk of multicollinearity in the model. Figure 3 shown.

[0073] In one embodiment, in step S4, the bridge inspection data is divided into training samples according to a time sliding window, and the model is trained using the data of the previous several periods to predict the continuous value of the disease development in the next period and construct a time series framework of the preset time.

[0074] Among them, a time series sample is constructed by setting a fixed time window. When constructing a time series sample, each input sample contains feature vectors of multiple historical time points, and the output is a continuous disease indicator at the target moment.

[0075] Using data from several previous periods, we predict the continuous disease value for the next period and establish a time series sample framework from "tn to t+1". When constructing a time series sample, each input sample contains feature vectors from multiple historical time points, and the output is a continuous disease indicator at the target time.

[0076] Sample construction is carried out by setting a fixed time window to construct a time series sample. Each sample contains the historical disease records of several bridges. Through the sliding window technology, the corresponding training samples are constructed based on the annual inspection data. For each bridge, features are extracted from its inspection data in different years to form a time series input sample. Each sample input consists of a series of vectors describing the evolution characteristics of the disease, which are used for model training. Specific features include but are not limited to disease development trends, score changes, and other structural features. Figure 4 shown.

[0077] In one embodiment, in step S5, the optimal hyperparameter combination of the distributed gradient boosting library model is determined by random search, and regularization is introduced to control the complexity of the model during the model iteration process. At the same time, the early stopping mechanism is used to judge the changing trend of the validation set loss, and the training is terminated early when there is no improvement for several consecutive rounds.

[0078] Using XGBoost as the base model, the optimal hyperparameter combination is first determined through random search, followed by iterative training. In each round, the gradient residual is used as the target for fitting a new subtree, and multiple trees are integrated to complete the modeling of structural damage prediction. To avoid overfitting, L2 regularization is introduced to control model complexity. An early stopping mechanism is triggered when the validation set loss does not improve for several consecutive rounds, retaining the model structure with the best generalization ability.

[0079] This application enhances the generalization ability and stability of the model. Random search is used to optimize XGBoost hyperparameters, such as learning rate, subsampling rate, and tree depth. L2 regularization is introduced to suppress overfitting and ensure the stability and generalization ability of the model under different disease patterns and data distributions.

[0080] The hyperparameters of the XGBoost model include learning rate, subsampling rate, maximum tree depth, minimum split gain, and regularization coefficient. They are tuned through random search, with mean square error as the optimization target to ensure the accuracy and stability of the prediction.

[0081] Build and train an XGBoost model for time-series prediction of bridge damage. XGBoost is an efficient gradient boosting tree model widely used for classification and regression problems with structured data. Through multiple rounds of iterative training, this method aims to minimize the model's loss function and improve the accuracy of predicting damage trends.

[0082] In one embodiment, the distributed gradient boosting library model performs predictions by integrating multiple decision tree models. Each tree is constructed based on the residual of the previous round of trees. The goal is to continuously optimize the model performance through gradient boosting to obtain a minimized loss function.

[0083] Among them, multiple optimized decision tree models are obtained through training and integrated into the distributed gradient boosting library model framework; during the model training process, based on the gradient boosting strategy, with the goal of minimizing the loss function, several subtrees are iteratively constructed. In each round of iteration, the prediction residual of the current model is used to train a new regression tree to correct the error. In each round of training, the square error is minimized by updating the weight of the tree; in the model construction stage, the key hyperparameters of the distributed gradient boosting library model are tuned through random search or grid search methods to ensure that the model achieves the optimal balance between accuracy and generalization ability; after multiple rounds of iterative training and optimization, a distributed gradient boosting library integrated learning model that integrates multiple weak learners is obtained.

[0084] This application can improve prediction accuracy and robustness. It extracts key features in a targeted manner, uses XGBoost to integrate the prediction results of multiple decision trees to reduce the impact of noise, and applies adaptive weights to features that are susceptible to environmental interference, thereby improving the accuracy and stability of the model.

[0085] The XGBoost model makes predictions by integrating multiple decision tree models. Each tree is constructed based on the residual of the previous tree, and the goal is to continuously optimize the model performance through gradient boosting. The training sample set is set to {(x1,y1),(x2,y2),...,(x n ,y n )}, where x i is the input feature vector, y i is the target label. The output of each tree is Represents the prediction result of the t-th round model. The final output of the model is the weighted sum of the prediction results of all trees. The formula is:

[0086]

[0087] in represents the model's predicted value for the i-th sample, T is the total number of decision trees, and f t (x i ) is the predicted value of the t-th tree.

[0088] The goal of XGBoost is to minimize the total loss function, which consists of two parts: prediction error and regularization term. The formula is:

[0089]

[0090] Where L is the total loss function, n is the number of samples, y i is the true value of the i-th sample, is the predicted value of the model, λ is the regularization coefficient, w j,t is the weight of the jth leaf node of the tth tree, and J is the total number of leaf nodes in a single tree.

[0091] The common form of L2 regularization is:

[0092]

[0093] Where Ω(f t ) is the regularization penalty term of the t-th tree.

[0094] Optimize the objective function. In each iteration, the model trains a new tree based on the residual and updates the predicted value. The formula is:

[0095]

[0096] in Represents the predicted value of the model after the tth iteration, Represents the predicted value of the model after the t-1th iteration, represents the learning rate, and controls the contribution weight of each tree.

[0097] Through training, multiple optimized decision tree models were obtained and integrated into the XGBoost model framework. During the model training process, based on the gradient boosting strategy, several subtrees were iteratively constructed with the goal of minimizing the loss function. In each round of iteration, the prediction residual of the current model was used to train a new regression tree to correct the error, thereby gradually improving the model performance. Each round of training minimized the squared error by updating the tree weights. During the model construction phase, the key hyperparameters of XGBoost were tuned through random search or grid search methods to ensure that the model achieved the optimal balance between accuracy and generalization ability. Finally, after multiple rounds of iterative training and optimization, an enhanced model integrating multiple weak learners was obtained, which is the final XGBoost ensemble learning model. This model can effectively predict the development trend of bridge diseases in the next few years and has good stability and engineering adaptability.

[0098] In one embodiment, the distributed gradient boosting library model evaluation uses mean square error, mean absolute error and determination coefficient as evaluation indicators; the evaluation indicators are used to evaluate and verify the integrated learning model, and the main indicators of the evaluation model are mean square error and mean absolute error, etc.

[0099] The mean square error (MSE) is used to measure the average square difference between the predicted value and the true value. The formula is:

[0100]

[0101] The mean absolute error (MAE) is used to measure the average absolute difference between the predicted value and the true value. The formula is:

[0102]

[0103] Coefficient of determination R 2 It measures the model's ability to explain data variability. It indicates the degree of fit between the model's predicted values ​​and the actual observed values. The value range is 0 to 1. The closer the value is to 1, the better the model's fit. The formula is:

[0104]

[0105] where y i is the true value of the i-th observation, is the predicted value of the ith observation, is the average of all observations n is the total number of samples.

[0106] After model evaluation is complete, the XGBoost model outputs predictions of bridge damage trends over the next several years. Specifically, the model predicts a continuous value of damage for each target year. The prediction for each bridge consists of the model's predicted value at a given point in time and its confidence interval. The specific output format is as follows:

[0107]

[0108] in, is the model's predicted value for the disease at time t, δ is the predicted confidence interval, and represents the uncertainty of the model regarding the predicted value. Predicted Value represents the predicted value, and ConfidenceInterval represents the confidence interval.

[0109] In one embodiment, the prediction method further includes step S6: outputting the bridge disease prediction results at a future time point, feeding subsequent detection data back into the training set, building a closed-loop mechanism of "prediction-verification-update", and continuously updating and optimizing the model.

[0110] The model outputs predicted values ​​for future defects and feeds subsequent annual inspection data back into the training set, establishing a closed-loop "prediction-verification-update" mechanism for continuous iterative optimization of the model. Ultimately, this integrated learning model enables the prediction of defects in hollow slab bridges.

[0111] The embodiments of this specification also disclose a time series prediction system for hollow slab bridge defects based on machine learning, comprising the following steps:

[0112] Module M1: Obtain hollow slab bridge inspection data and construct the original data set based on the disease data of the inspection data.

[0113] Module M2: Structured preprocessing of raw data.

[0114] Module M3: Perform feature engineering on structured data, use random forest screening, and perform feature correlation analysis to obtain a low-redundancy and high-correlation feature set.

[0115] Module M4: Divide the bridge inspection data corresponding to low-redundancy and high-correlation features into training samples according to the time sliding window to construct a time series sample set.

[0116] Module M5: Perform model training and optimization, train time series samples, and predict the continuous values ​​of future diseases.

[0117] In this embodiment, the specific implementation of each functional module can be found in the various implementation steps of the aforementioned bridge defect time series prediction method, and will not be detailed here. It should be noted that the system structure provided in this embodiment is only logically divided based on the above functional modules for illustrative purposes. In actual applications, the functions of each module can be reconfigured and combined according to engineering needs. In other words, different system components can assume some or all of the above functions to implement the prediction logic and processing flow proposed in this application.

[0118] This application provides a method and system for predicting hollow slab bridge defects using XGBoost. Through data cleaning, feature engineering, time series sample construction, and ensemble learning modeling, a machine learning model for bridge defect prediction is constructed. This model leverages structured and unstructured data extracted from periodic bridge inspection reports over the years, fully integrating machine learning modeling techniques with bridge infrastructure operation and maintenance scenarios. This model boasts lightweight, low-cost, high accuracy, and high adaptability, enabling numerical predictions of future bridge defect trends, including indicators like defect severity and crack propagation, providing accurate defect prediction results.

[0119] The present application relates to the technical field of bridge defect analysis and prediction. The method includes: constructing a structural defect dataset containing timestamps based on bridge inspection reports over the years, performing data cleaning, time series linear interpolation, structural transformation, and standardization processing; using random forests and Spearman coefficients to complete feature selection and construct a low-redundancy evolution feature set; constructing time series samples through a sliding time window, using the XGBoost model for iterative training and introducing regularization and Early Stopping mechanisms to control overfitting, realizing dynamic modeling of future defect values, and constructing a "prediction-verification-update" closed-loop optimization process. This method has the characteristics of controllable quantitative prediction error and strong dynamic evolution modeling capabilities, and can provide effective support for bridge defect prediction modeling and maintenance decision-making.

[0120] This application addresses the existing problem of bridge defect prediction and assessment relying primarily on traditional statistical modeling methods, which only provide static assessments and lack dynamic predictions based on time evolution. This application presents a machine learning-based method for predicting hollow slab bridge defects over time. Through data collection, preprocessing, and feature optimization, a model is established to output dynamic predictions of continuous defect indicators. This application improves the accuracy and efficiency of bridge defect prediction and has promising application prospects.

[0121] This application is a lightweight prediction method for inspection reports, which can integrate structured and unstructured disease information, establish a time-series evolution model, and provide quantitative data support and precise decision-making basis for the analysis of bridge disease evolution mechanisms and the formulation of preventive maintenance strategies. This application has the predictability of disease trends, constructs bridge inspection data into time-series samples, introduces time evolution characteristics, and outputs dynamic prediction results of continuous indicators to provide precise support for quantitative maintenance decisions. In response to the heterogeneity of hollow slab bridge data and its dependence on engineering experience, this application constructs a customized preprocessing and feature engineering process, converts the inspection text into quantitative indicators through a structured rule engine, and screens features that are highly correlated with diseases based on mechanical principles and eliminates redundant items.

[0122] In this specification, the same or similar parts between the various embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the embodiments described later, the description is relatively simple, and the relevant parts can be referred to the partial description of the previous embodiments.

[0123] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A time series prediction method for hollow slab bridge defects based on machine learning, characterized in that: The steps include: Step S1: Obtain hollow slab bridge inspection data and construct an original data set based on the damage data of the inspection data; Step S2: perform structured preprocessing on the original data; Step S3: Perform feature engineering on the structured data, use random forest to screen, and perform feature correlation analysis to obtain a low-redundancy and high-correlation feature set; Step S4: Divide the bridge inspection data corresponding to the low-redundancy and high-correlation features into training samples according to the time sliding window to construct a time series sample set; Step S5: Perform model training and optimization, train the time series samples, and predict the continuous values ​​of future diseases.

2. The method for predicting hollow slab bridge defects based on machine learning according to claim 1 is characterized in that: In step S1, the hollow slab bridge data in the regular bridge inspection report is obtained, and the inspection timestamp is retained for the disease data, and an original data set containing disease information and timestamp is constructed; wherein the hollow slab bridge data includes basic bridge information, geometric property data, material property data, environmental factor data and continuous disease indicators.

3. The method for predicting hollow slab bridge defects based on machine learning according to claim 1 is characterized in that: In step S2, the original data is parsed for text keywords, outliers are removed, missing values ​​are filled, and data types and dimensions are unified to convert them into structured numerical input data; Among them, missing value filling adopts a linear interpolation method based on time series, and interpolation operations are performed independently on each feature dimension to fill the missing continuous value features; statistical methods are used to identify and eliminate outliers in the data.

4. The method for predicting hollow slab bridge defects based on machine learning according to claim 1 is characterized in that: In step S3, the random forest algorithm is used to calculate the Gini coefficient of each feature to sort the importance, select a preset number of features before sorting, and calculate the correlation coefficient matrix. Feature pairs with correlation greater than a preset threshold are screened, and features with high Gini coefficients are retained to form a low-redundancy high-correlation feature set.

5. The method for predicting hollow slab bridge defects based on machine learning according to claim 1 is characterized in that: In step S4, the bridge inspection data is divided into training samples according to the time sliding window, and the model is trained using the data of the previous several periods to predict the continuous value of the disease development in the next period and construct a time series framework of the preset time; Among them, a time series sample is constructed by setting a fixed time window. When constructing a time series sample, each input sample contains feature vectors of multiple historical time points, and the output is a continuous disease indicator at the target moment.

6. The method for predicting hollow slab bridge defects based on machine learning according to claim 1 is characterized in that: In step S5, the optimal hyperparameter combination of the distributed gradient boosting library model is determined by random search, and regularization is introduced to control the complexity of the model during the model iteration process. At the same time, the early stopping mechanism is used to judge the changing trend of the validation set loss, and the training is terminated early when there is no improvement for several consecutive rounds.

7. The method for predicting hollow slab bridge defects based on machine learning according to claim 6 is characterized in that: The distributed gradient boosting library model makes predictions by integrating multiple decision tree models. Each tree is constructed based on the residual of the previous tree. The goal is to continuously optimize the model performance through gradient boosting to obtain the minimized loss function. Among them, multiple optimized decision tree models are obtained through training and integrated into the distributed gradient boosting library model framework; during the model training process, based on the gradient boosting strategy, with the goal of minimizing the loss function, several subtrees are iteratively constructed. In each round of iteration, the prediction residual of the current model is used to train a new regression tree to correct the error. In each round of training, the square error is minimized by updating the weight of the tree; in the model construction stage, the key hyperparameters of the distributed gradient boosting library model are tuned through random search or grid search methods to ensure that the model achieves the optimal balance between accuracy and generalization ability; after multiple rounds of iterative training and optimization, a distributed gradient boosting library integrated learning model that integrates multiple weak learners is obtained.

8. The method for predicting hollow slab bridge defects based on machine learning according to claim 7 is characterized in that: The distributed gradient boosting library model evaluation uses mean square error, mean absolute error and coefficient of determination as evaluation indicators; Among them, the mean square error is used to measure the average square difference between the predicted value and the true value; the mean absolute error is used to measure the average absolute difference between the predicted value and the true value; the coefficient of determination is used to measure the model's ability to explain the variability of the data, indicating the degree of fit between the model's predicted value and the actual observed value.

9. The method for predicting hollow slab bridge defects based on machine learning according to claim 1 is characterized in that: The prediction method further includes step S6: outputting the bridge disease prediction results at a future time point, feeding subsequent detection data back into the training set, building a closed-loop mechanism of prediction, verification, and updating, and continuously updating and optimizing the model.

10. A time series prediction system for hollow slab bridge defects based on machine learning, characterized by: The steps include: Module M1: Obtain hollow slab bridge inspection data and construct the original data set based on the disease data of the inspection data; Module M2: structured preprocessing of raw data; Module M3: Perform feature engineering on structured data, use random forest to screen, and perform feature correlation analysis to obtain a low-redundancy and high-correlation feature set; Module M4: divide the bridge inspection data corresponding to low-redundancy and high-correlation features into training samples according to the time sliding window to construct a time series sample set; Module M5: Perform model training and optimization, train time series samples, and predict the continuous values ​​of future diseases.

Images