Small and medium span bridge technical condition grade degradation prediction method and system

Abstract

The present application relates to a small and medium span bridge technical condition grade degradation prediction method and system, belongs to the technical field of bridge technical condition degradation analysis, solves the weak model generalization ability and low prediction accuracy problem existing in the prior art when predicting the technical condition degradation of small and medium span bridges. The method comprises: collecting detection data of multiple small and medium span bridges, establishing a bridge detection database; preprocessing the data in the database; constructing and training a prediction model that integrates attention mechanism and double-branch output; obtaining the detection data of the target small and medium span bridge, inputting the preprocessed data into the trained bridge technical condition grade degradation prediction model, outputting the bridge technical condition grade degradation prediction result, and simultaneously adding the obtained detection data and the bridge technical condition grade degradation prediction result to the bridge detection database for model iteration and updating. The present application can quickly predict the technical condition grade of small and medium span bridges, and has high prediction accuracy and strong model generalization ability.

Description

Technical Field

[0001] This invention relates to the field of bridge technical condition degradation analysis technology, specifically to a method and system for predicting the technical condition degradation level of small and medium-span bridges. Background Technology

[0002] Bridges are subjected to multiple degradation factors over long periods, including overloading, extreme weather, and material aging. Structural performance degradation can lead to major safety accidents. Numerous bridge collapses both domestically and internationally are directly related to inadequate maintenance and the failure to provide timely warnings of degradation.

[0003] Existing bridge technical condition assessment standards suffer from problems such as strong subjectivity and vague quantitative standards. For example, the classification of component defects is unclear, and the description of the degree of damage relies on experience-based judgment. Meanwhile, traditional linear or nonlinear degradation models struggle to accurately describe the complex dynamic process of bridge degradation and lack the ability to model the coupling effects of multiple factors. With the development of big data and machine learning technologies, the accumulation of large amounts of bridge inspection data provides a data foundation for predicting bridge technical condition degradation levels. However, for the numerous and widespread small- and medium-span bridges, extracting effective features from the vast amount of inspection data and constructing high-precision models remains a technical challenge.

[0004] Currently, some scholars have conducted research on predicting the degradation of bridge technical conditions based on bridge inspection or monitoring data, but there are still certain limitations.

[0005] For example, the master's thesis from Jilin University, "Prediction of Bridge Technical Condition Degradation and Optimization of Maintenance Scheme Based on Regular Inspection Data," proposed using a classic BP neural network to construct a bridge technical condition degradation model to predict the development trend of bridge technical condition. However, this study performed well for bridges with technical condition ratings of 1, 2, and 3 based on a large amount of data, but the prediction effect was poor for bridges of 4 and 5, which greatly limited the application of the model and its results.

[0006] For example, the invention patent with publication number CN108509656A proposes a "bridge inspection and evaluation system based on a data dictionary," which includes a web-based data management system and a mobile-based defect collection system. The web-based data management system and the mobile-based defect collection system synchronize, interact, and transmit information. This system uses on-site bridge defect data to evaluate the technical condition of the bridge. Although this system makes full use of bridge inspection data, data collection, transmission, and analysis require a certain amount of time, making real-time evaluation of the bridge's technical condition impossible and resulting in low evaluation efficiency.

[0007] For example, the invention patent with publication number CN111581697A proposes a "BIM-based bridge inspection information management method and system," which utilizes multiple drones, 3D lasers, and sensors to collect measured data of the bridge under inspection. Based on the bridge's BIM model and GIS technology, it obtains the inspection results of the bridge according to bridge evaluation standards. However, the process of constructing the BIM model and performing finite element analysis in this method and system can lead to a lack of timeliness in bridge evaluation.

[0008] For example, the invention patent with publication number CN117172098A proposes "a method for evaluating the technical condition of road network bridges based on an LSTM model." This method uses historical bridge information cards of the urban bridge network to construct and train an LSTM model that considers time series data using the PyTorch deep learning architecture, resulting in a model that can meet the prediction trend of actual structural performance degradation. However, this method requires continuous and long-term data support to construct the LSTM model, which is difficult to implement for small and medium-span bridges without health monitoring systems.

[0009] For example, the invention patent with publication number CN119646758A proposes "a comprehensive evaluation method for bridge condition based on multi-source data fusion." This method uses an autoregressive integral moving average model and a simple exponential smoothing method to construct models for predicting periodic and non-periodic indicators, respectively, and uses the analytic hierarchy process (AHP) to assign weights to different indicators to calculate a comprehensive evaluation value representing the bridge condition. However, this method relies on long-term continuous monitoring data of the bridge, and the regression and correlation analysis methods used cannot consider the nonlinear degradation process under the coupling effect of multiple factors.

[0010] In summary, existing bridge technical condition degradation prediction technologies for small and medium-span bridges mainly suffer from the following problems:

[0011] (1) Insufficient data volume and continuity:

[0012] The lack of early bridge records led to significant errors in parameter calibration for technical condition degradation models built on limited monitoring data, resulting in insufficient prediction accuracy and making them unsuitable for the numerous and widespread small-to-medium span bridges. Furthermore, the low coverage of monitoring equipment, with health monitoring systems installed only on very large bridges, and the discontinuity of time-series data negatively impacted the training effectiveness of the prediction models.

[0013] (2) Inconsistent data format:

[0014] Bridge inspection data includes both numerical and textual data. Existing methods lack uniformity and standardization in terms of format and units for characteristic data, fail to establish correlations between multi-source data, and cannot reflect the comprehensive characteristics of structural degradation, thus affecting the compatibility and fusion efficiency of multi-source data.

[0015] (3) The generalization ability of models driven by a small amount of data is weak:

[0016] Commonly used linear or exponential degradation models are difficult to describe the nonlinear degradation process under the coupling of multiple factors. Machine learning-based prediction models (such as LSTM and random forest) require a large amount of labeled data for training, but the actual detection data often suffers from overfitting or poor generalization ability due to insufficient sample size or non-standard labeling.

[0017] (4) Low accuracy in assessing the technical condition of a single bridge:

[0018] Existing technical condition degradation prediction models are mostly designed for single bridges, creating "information silos" that cannot guarantee the accuracy of assessment and prediction models. Furthermore, the lack of utilization of a large amount of road network-level bridge inspection or monitoring data to achieve data sharing and integration makes it difficult to improve the accuracy and precision of bridge technical condition degradation prediction assessment and prediction models. Summary of the Invention

[0019] To address the aforementioned problems, this invention proposes a method and system for predicting the technical condition degradation of small- and medium-span bridges based on an MLP-dual-branch fully connected neural network. This method and system establish a detection database for small- and medium-span bridges, preprocess the data in the database, and construct and train a bridge technical condition degradation prediction model that integrates attention mechanisms and ensemble learning. This achieves accurate prediction of the technical condition degradation of small- and medium-span bridges at the road network level. This invention fully utilizes detection data within the road network, and the established intelligent prediction model exhibits strong generalization ability and high prediction accuracy. It balances accuracy with consistency in grade order and boasts fast inference efficiency, providing a scientific basis for bridge maintenance decisions and demonstrating promising application prospects and practical value.

[0020] To solve the above problems, the present invention adopts the following technical solution:

[0021] A method for predicting the degradation of the technical condition level of small and medium-span bridges, comprising the following steps:

[0022] Step 1: Collect bridge inspection data of multiple small and medium span bridges in the regional road network and establish a bridge inspection database. The bridge inspection database includes bridge construction information, bridge environment data, road network data, inspection data and maintenance management information.

[0023] Step 2: Preprocess the data in the bridge inspection database;

[0024] Step 3: Construct a bridge technical condition level degradation prediction model that integrates attention mechanism and ensemble learning, and train the bridge technical condition level degradation prediction model based on preprocessed data;

[0025] Step 4: Obtain the detection data of the target small-span bridge, preprocess the detection data, input the preprocessed data into the trained bridge technical condition level degradation prediction model, and output the bridge technical condition level degradation prediction result of the target small-span bridge; at the same time, add the detection data and the bridge technical condition level degradation prediction result to the bridge detection database for model iteration and update.

[0026] Accordingly, this invention also proposes a system for predicting the degradation of the technical condition level of small and medium-span bridges, the system comprising:

[0027] The database construction module is used to collect bridge inspection data of multiple small and medium span bridges in the regional road network and establish a bridge inspection database. The bridge inspection database includes bridge construction information, bridge environment data, road network data, inspection data and maintenance management information.

[0028] The data processing module is used to preprocess the data in the bridge inspection database;

[0029] The model building and training module is used to build a bridge technical condition level degradation prediction model that integrates attention mechanism and ensemble learning, and to train the bridge technical condition level degradation prediction model based on preprocessed data.

[0030] The prediction module is used to acquire the detection data of the target small-span bridge, preprocess the detection data through the data processing module, input the preprocessed data into the trained bridge technical condition level degradation prediction model, and output the bridge technical condition level degradation prediction result of the target small-span bridge.

[0031] The database update module is used to add the detection data and the bridge technical condition level degradation prediction results to the bridge detection database for model iteration and updates.

[0032] Compared with the prior art, the present invention has the following advantages:

[0033] (1) Currently, the safety of small and medium-span bridges is still based on periodic inspection data. However, the periodic inspection data of a single bridge is extremely limited, and it is impossible to establish and verify an accurate prediction model based on a small amount of data. This invention establishes a bridge inspection database, summarizes, classifies and filters the data of multiple bridges in the regional road network, realizes the sharing and complementarity of data with similar characteristics, and ensures the volume and continuity of data through outlier removal and missing value processing. This ensures the accuracy and reliability of the bridge technical condition level degradation prediction model built based on the bridge inspection database, and improves the accuracy and precision of the model in predicting bridge technical condition degradation.

[0034] (2) This invention filters the data in the detection database, constructs derived features such as span ratio and cumulative traffic load by combining feature engineering, and then processes the category features by one-hot encoding, standardization to eliminate dimensions, and PCA dimensionality reduction to reduce redundant dimensions. At the same time, the SMOTE oversampling balance processing method is used to optimize the data distribution. These preprocessing operations effectively reduce noise interference and data dimensionality while retaining key features. This not only improves the calculation speed and enhances the real-time performance of the prediction, but also directly improves the accuracy of the model prediction through high-quality input features.

[0035] (3) Commonly used linear or exponential degradation models are difficult to describe the nonlinear degradation process under the coupling effect of multiple factors. The input feature set considered in this invention not only includes basic parameters such as span, bridge age, traffic volume, and total bridge length, but also covers deep features such as span variation coefficient and material-load interaction mined through feature engineering. By constructing a bridge technical condition level degradation prediction model that integrates attention mechanism and ensemble learning, and by optimizing it through five-fold cross-validation and dynamic weight loss function, the complex law of bridge technical condition degradation is effectively captured, the generalization ability of the model is significantly improved, and the rapid and accurate prediction of degradation of small and medium span bridges at the road network level is realized. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without exceeding the scope of protection claimed by the present invention.

[0037] Figure 1 This is a flowchart of the method for predicting the degradation level of technical condition of small and medium-span bridges according to an embodiment of the present invention;

[0038] Figure 2 This is an entity relationship diagram for the bridge inspection database. In the diagram, "1" and "n" are used to represent the relationship type between entities. "1" represents the "one" party in a "one-to-one" or "one-to-many" relationship, and "n" represents the "many" party in a "one-to-many" or "many-to-many" relationship.

[0039] Figure 3 A schematic diagram of the predictive model structure that integrates the attention mechanism and the dual-branch output;

[0040] Figure 4 Here is a flowchart of the data preprocessing process;

[0041] Figure 5 This is a schematic diagram of the one-hot processing results for some labeled data;

[0042] Figure 6This is a schematic diagram of the processing results for some ordinal data.

[0043] Figure 7 This is a schematic diagram showing the dimensionality reduction results of some labeled data using principal component analysis.

[0044] Figure 8 This is a schematic diagram of the results of dimensional normalization.

[0045] Figure 9 A confusion matrix diagram of the model's prediction results;

[0046] Figure 10 A graph of the ROC curve for the model's prediction results;

[0047] Figure 11 This is a schematic diagram of a bridge technical condition level degradation prediction system. Detailed Implementation

[0048] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0049] like Figure 1 As shown in the figure, this embodiment provides a method for predicting the degradation of the technical condition level of small and medium-span bridges. The method includes the following steps:

[0050] Step 1: Collect bridge inspection data of multiple small and medium span bridges in the regional road network and establish a bridge inspection database.

[0051] In this step, the established bridge inspection database specifically includes: bridge construction information, bridge environmental data, road network data, inspection data, and maintenance management information. Bridge construction information includes the bridge name, center pile number, bridge design parameters, construction period, and ancillary facilities; bridge environmental data refers to the environmental conditions during bridge maintenance; road network data includes route attributes, route names, and traffic volume; inspection data includes defects found in previous inspections, technical condition scores, rating levels, and the inspection implementing unit; and maintenance management information includes past maintenance history and maintenance decisions based on inspection results.

[0052] Furthermore, the process of establishing a bridge inspection database specifically includes the following steps:

[0053] Step 1.1: Construct an Entity-Relationship (ER) graph for the bridge detection database to describe the relationships between the database tables.

[0054] Step 1.2: Based on the constructed entity relationship diagram, using Python, connect to Navicat through the PyCharm compiler, call the bridge to detect data in the database, and design and build the database tables.

[0055] In step 1.2, the focus of database table construction is on clarifying the table names and functions of each information table. In this embodiment, the database tables are mainly divided into three core data tables: basic information, inspection data, and maintenance. The database tables constructed in this step include: a bridge basic information table, a bridge route information table, a bridge construction unit information table, a bridge basic attribute correspondence table, a bridge inspection unit information table, a bridge component scoring table, a bridge part scoring table, a defects table, a historical maintenance information table, and a maintenance decision table.

[0056] Step 2: Preprocess the data in the bridge inspection database to obtain the input data for the prediction model.

[0057] Specifically, the preprocessing of data in the bridge inspection database includes the following steps:

[0058] Step 2.1: Based on the bridge inspection database established in Step 1, filter out some factors that have a significant impact on the bridge technical condition level degradation prediction model, and determine the input feature set used for training the bridge technical condition level degradation prediction model.

[0059] Furthermore, the input feature set includes basic attribute features, derived features, and category interaction features. Basic attribute features include traffic volume, highway grade, design load, bridge type, total bridge length, span, maximum single-span span, service life, bridge deck width, and materials directly extracted from the bridge inspection database. Derived features include span ratio, cumulative traffic load, bridge deck area, and span variation coefficient constructed through feature engineering. Category interaction features include material-load interaction, length grade, and aging degree. The input feature set also includes a target feature, which is the bridge technical condition grade (or bridge technical condition rating).

[0060] Step 2.2: Process the missing and outlier values ​​in the input feature set to obtain mixed data containing numerical and nominal features.

[0061] In this step, the processing of the input feature set is divided into two categories: filling in missing values ​​and removing outliers.

[0062] 1) For missing values: use the median to fill in for numerical features; use the mode to fill in for categorical features, to avoid reducing the amount of data due to deleting samples;

[0063] 2) For outliers: Identify outliers in numerical features using box plots, and directly remove outlier samples to ensure the reasonableness of data distribution.

[0064] After the above processing, we finally obtain mixed feature data containing both numerical and nominal features, which serves as the basis for subsequent processing.

[0065] Step 2.3: Normalize the data for different types of features (numerical features and nominal features).

[0066] In this step, numerical features refer to features directly represented by numbers, such as traffic volume, total bridge length, span, service life, and span ratio. To eliminate the influence of dimensions, numerical feature data is normalized through standardization. Nominal feature data can be divided into two categories: labeled data and ordinal data. They require normalization processing according to their respective characteristics as follows:

[0067] 1) For labeled data: Use one-hot encoding to convert it into an n-dimensional 0-1 vector.

[0068] 2) For ordinal data: A numerical mapping method is used to map its n unique values ​​to consecutive integers in the interval [0, n-1], preserving their inherent order and priority relationship, thus realizing the conversion to numerical variables.

[0069] Step 2.4: Principal Component Analysis (PCA) is used to reduce the dimensionality of the normalized nominal feature data. This simplifies the multiple variables representing the internal structure of the data into a few independent principal components, thereby reducing the data dimensionality while retaining the main information and reducing the number of neurons in the input layer of the prediction model.

[0070] In this step, based on the labeled data after one-hot encoding and the ordinal data after numerical mapping, the principal components with the largest variance contribution are extracted using the PCA method. The principal components with a cumulative variance contribution rate of ≥95% are retained, thereby mapping the original high-dimensional features to the low-dimensional principal component space. This reduces the feature dimension while retaining the core information, lowers the computational complexity of the model, and avoids overfitting.

[0071] Step 2.5: Use the range method to map the numerical feature data and the dimensionality-reduced data to the range [0-1] to eliminate the influence of the units on the importance of the features and improve the balance and stability of the model input.

[0072] In this step, the range method is used to perform a linear transformation on the data after dimensionality reduction by PCA, mapping the data to the [0,1] interval. The specific calculation is shown in formula (1).

[0073] (1)

[0074] in, To eliminate dimensional data, For data generated after principal component analysis, The minimum value of the dataset after principal component analysis. This represents the maximum value of the dataset after processing with principal component analysis.

[0075] Step 3: Construct a bridge technical condition level degradation prediction model that integrates attention mechanism and ensemble learning, and train the bridge technical condition level degradation prediction model based on preprocessed data. The trained model is used to achieve accurate prediction of the technical condition level of small and medium span bridges at the road network level.

[0076] Furthermore, a bridge technology status degradation prediction model that integrates attention mechanisms and ensemble learning is constructed and trained through the following steps:

[0077] Step 3.1: Based on the deep learning framework, construct a prediction model that integrates the attention mechanism and the dual-branch output, and use the five-fold cross-validation method to optimize the model hyperparameters when training the prediction model that integrates the attention mechanism and the dual-branch output.

[0078] like Figure 3 As shown, the prediction model constructed in this step, which integrates the attention mechanism and the dual-branch output, comprises an input layer, a hidden layer, and a dual-branch output layer connected in sequence. The input layer receives the bridge data feature vector composed of the input data formed after preprocessing in step 2. Furthermore, the preprocessing in step 2 also includes SMOTE oversampling balancing of the training set samples.

[0079] The hidden layer is a deep neural network with an attention mechanism, including an attention module and a backbone network. The attention module is used to assign weights to input features (such as traffic volume, service life, span ratio, etc.) and automatically discover key features that significantly affect the degradation of bridge technical condition (such as bridge deck width, cumulative traffic load, etc.), thereby enhancing the model's ability to capture important features. The backbone network adopts a multilayer perceptron (MLP) structure, which contains three fully connected layers. High-order features are extracted through these three fully connected layers. Each fully connected layer contains a batch normalization layer and a Dropout regularization layer, and the activation function is the GELU function.

[0080] The dual-branch output layer, connected to the backbone network, comprises two parallel branches: a classification branch and an ordered regression branch. The classification branch outputs the classification probability of the bridge's technical condition level, while the ordered regression branch outputs the level order score. The dual-branch output layer balances classification accuracy with level order consistency; therefore, it primarily includes the classification branch outputting the degradation level prediction result, and the ordered regression branch optimizing the prediction result using the level order. Furthermore, a dynamic weighted loss layer is added to the prediction model that integrates the attention mechanism and the dual-branch output. By adjusting the dynamic weight α (ranging from 0.5 to 0.9), the results after each round of testing are backpropagated with the new weight α until the fifth fold calculation is completed. The dynamic weighted loss layer employs a dynamic weight loss function, optimizing classification accuracy through cross-entropy loss, constraining the level order relationship with ordered loss, and dynamically adjusting the weights of the classification loss and ordered loss in each fold of training during the five-fold cross-validation (e.g., initial α = 0.7). The dimensions of the input, hidden, and output layers can be set based on the amount of data. The dropout rate is 0.3 (to suppress overfitting), and the GELU activation function is used (to improve non-linear fitting ability). During the model training and validation phases, five independent models are constructed using five-fold cross-validation. The class imbalance problem is solved by using a weighted random sampler. Each model is trained on different subsets of data and the optimal weights are saved.

[0081] Step 3.2: The prediction model of the fusion attention mechanism and the dual-branch output after training is comprehensively evaluated by multi-dimensional evaluation indicators to ensure prediction accuracy and generalization ability, and to obtain the bridge technical condition level degradation prediction model.

[0082] In this step, multi-dimensional evaluation metrics include confusion matrix, F1 score, precision, recall, quadratic weighted kappa (QWK), mean squared error (MSE), root mean squared error (RMSE), and mean absolute error (MAE). These metrics are selected to comprehensively evaluate the performance of the established prediction model.

[0083] A confusion matrix reflects the number of correctly predicted and incorrectly predicted data. Table 1 shows a binary confusion matrix. The first letter, T, and F represent correct and incorrect predictions, respectively, and the second letter indicates the predicted outcome. For example, TP (TruePositive) represents the number of correctly predicted data that also turned out to be correct, and TN (TrueNegative) represents the number of incorrectly predicted data that also turned out to be incorrect.

[0084] Table 1 Confusion Matrix

[0085]

[0086] Precision (or accuracy) refers to the percentage of samples that a classifier predicts to be "correct". The number of samples predicted as "correct" accounts for the total number of samples predicted as "correct". The proportion is shown in formula (2). The higher the precision value, the lower the false alarm rate of the prediction model.

[0087] (2)

[0088] Recall (or recall) is the probability that a sample is considered "correct" by the model and is indeed "correct" out of all samples that are actually "correct", as shown in formula (3). The higher the recall value, the lower the false negative rate of the prediction model.

[0089] (3)

[0090] The F1 score can characterize the performance of the learner, and its calculation is shown in formula (4).

[0091] (4)

[0092] in, To achieve precision, This refers to the recall rate.

[0093] The weighted Kappa coefficient (QWK) measures the consistency between the order of the technical condition levels predicted by the model and the actual levels, and solves the problem that ordinary accuracy ignores the inherent order of the levels. Its calculation is shown in formula (5).

[0094] (5)

[0095] in, In this embodiment, the number of categories represents the bridge's technical condition level. ;

[0096] To confuse the matrix "actually a level" Predicted as a level The number of samples;

[0097] For "actually a level" Predicted as a level The expected sample size is calculated as follows: ;

[0098] This is a weight matrix used to represent the "actual level". With prediction level The degree of difference is determined by a quadratic weight matrix in this model, i.e. .

[0099] Mean square error (MSE) and root mean square error (RMSE) can assess the deviation between the model prediction results and the actual values ​​from a regression perspective. The root mean square error is the square root of the mean square error. Here, only the formula for mean square error (MSE) is listed, and its calculation is shown in formula (6).

[0100] (6)

[0101] in, This represents the total number of samples in the test set. For the first The actual grade code value of each sample; For the first The predicted grade code value for each sample.

[0102] The mean absolute error (MAE) can reflect the "average deviation level of most samples", and its calculation is shown in formula (7).

[0103] (7)

[0104] The meanings of the parameters in formula (7) are the same as those of the mean square error (MSE), and will not be repeated here.

[0105] Step 4: Obtain the detection data of the target small-span bridge, preprocess the acquired detection data, and input the preprocessed data into the trained bridge technical condition level degradation prediction model. The model outputs the bridge technical condition level degradation prediction result of the target small-span bridge. At the same time, the acquired detection data and the bridge technical condition level degradation prediction result output by the model are added to the bridge detection database for iterative updates of the bridge technical condition level degradation prediction model.

[0106] In this step, after obtaining the detection data of the target small-span bridge, the detection data is preprocessed using the same preprocessing method as in step 2. Then, the bridge technical condition level degradation prediction model, which has been trained, is used to quickly predict and output the bridge technical condition level degradation prediction result of the target small-span bridge. On the other hand, the detection data of the target small-span bridge and the predicted bridge technical condition level degradation result are imported into the bridge detection database as new data. This is used to dynamically update and iteratively optimize the existing bridge technical condition level degradation prediction model, thereby continuously improving the accuracy and generalization ability of the model in predicting bridge technical condition.

[0107] To address the current practice of assessing the safety status of numerous small- and medium-span bridges using a "one bridge, one inspection" approach, and the lack of full utilization of extensive regional survey data on small- and medium-span bridges, coupled with the absence of artificial intelligence methods for rapid and accurate prediction of bridge technical conditions, this invention provides a method for predicting the degradation of technical conditions of small- and medium-span bridges. This method fully utilizes a large amount of survey data on small- and medium-span bridges within a regional road network. Based on a trained bridge technical condition degradation prediction model, it can quickly and accurately predict the degradation of the technical condition of target small- and medium-span bridges, demonstrating promising engineering applications and practical value.

[0108] The following uses a group of small and medium-span bridges in a certain region as an example to illustrate the implementation process and prediction effect of the prediction method of the present invention.

[0109] First, this embodiment uses the test data of 1,968 small and medium span bridges in a certain region. Table 2 shows the technical condition rating of these bridges.

[0110] Table 2. Screening of Bridge Technical Condition Rating

[0111]

[0112] Then, based on the specific characteristics of the bridges in the region, suitable model input features and output parameters are selected to form an input feature set, as shown in Table 3.

[0113] Table 3 Model Input Features and Output Parameters

[0114]

[0115] Next, the input and output feature data are preprocessed. This embodiment uses the `OneHotEncoder()` function from the sklearn library to process four labeled data types: "Structure Type," "Structure Material," "Length Grade," and "Material-Load Interaction." The processing results for "Structure Type," "Structure Material," and "Length Grade" are shown in Tables 4, 5, and 6, respectively. The Material-Load Interaction feature is generated by cross-combining "Structure Material" and "Design Load Grade," resulting in 28 combinations. Its processing method is the same as the one-hot encoding method for Structure Type, Structural Material, and Length Grade, and will not be repeated here. Some One-Hot processing results for labeled data are shown below. Figure 5 As shown, the dimensions of the data after one-hot encoding of the labeled data are [1968, 40]. The mapping tables for highway grade, design load, and aging degree are shown in Tables 7, 8, and 9, respectively, based on the LabelEncoder() function for processing ordinal data. Some ordinal data processing results are shown below. Figure 6 As shown.

[0116] Table 4. One-hot coding processing for different bridge structural types

[0117]

[0118] Table 5 Thermal coding processing for different structural materials

[0119]

[0120] Table 6. One-hot encoding processing for different length levels

[0121]

[0122] Table 7 Mapping Processing for Different Highway Grades

[0123]

[0124] Table 8. Design Load Mapping Treatment under Different Standards

[0125]

[0126] Table 9 Aging Degree Mapping Process

[0127]

[0128] Subsequently, principal component analysis was used to perform dimensionality reduction and dimensional normalization (i.e., dimensionless elimination) on the nominal data. The data processing flow is as follows: Figure 4 As shown, the PCA dimensionality reduction results and dimensionality normalization results for nominal data are respectively as follows: Figure 7 and Figure 8 As shown.

[0129] Next, this embodiment combines the feature dimensions (approximately 50-100 dimensions after preprocessing) and data distribution characteristics of the bridge technical condition prediction task, and determines the hyperparameters of the prediction model through five-fold cross-validation and grid search: the number of hidden layer neurons is calculated based on the empirical formula "input feature dimension × 1.5 + number of output categories" and optimized to 256 (first layer), 256 (middle layer), and 128 (last layer), which is suitable for the information extraction requirements of high-dimensional input features (such as span ratio, cumulative traffic load, and other derived features); the activation function of the hidden layer is GELU (Gaussian error linear unit); the output layer adopts the Softmax activation function, combined with the ordered regression branch to output the probability distribution of the level; other hyperparameters are tested and the optimal configuration is determined by the control variable method, and the test results are shown in Table 10.

[0130] Table 10 Hyperparameter Settings Table

[0131]

[0132] Finally, this embodiment uses a stratified sampling mechanism to divide the data after dimensionality reduction and normalization processing using principal component analysis in step 2 into training, validation, and test sets in a ratio of 6:2:2. First, the dataset processed by SMOTE oversampling balancing is divided into a "training + validation" set and a test set in an 8:2 ratio. Then, the "training + validation" set is further split into training and validation sets in a 7.5:2.5 ratio. Throughout the process, the stratify parameter ensures a consistent distribution of bridge rating levels across all datasets. When the sample size for each level meets the stratification requirements, the division is strictly based on the level ratio; if the sample size for some levels is insufficient, the sampling strategy is automatically adjusted to maintain distribution stability. The test set ultimately contains bridge data with a distribution consistent with the training and validation sets: 85 bridges in category 1, 261 in category 2, 105 in category 3, 8 in category 4, and 1 in category 5. Finally, the test set is tested based on the trained bridge technical condition level degradation prediction model to obtain a confusion matrix, as shown below. Figure 9 As shown in Table 11, recall, precision, QWK, MSE, RMSE, MAE, and F1 score were calculated for different categories.

[0133] Table 11 Values ​​of Each Evaluation Indicator

[0134]

[0135] The precision and QWK values ​​show that bridges with technical condition ratings of 4 and 5 have the best prediction performance, followed by bridges in categories 1 and 3, and finally bridges in category 2. For all five categories of bridges, the precision values ​​are 80% or higher, demonstrating that the prediction model can effectively classify each category with high accuracy. Furthermore, considering the QWK results, which generally explore the model's classification ability for a specific category, a QWK value closer to 1 indicates better classification performance. In this embodiment, the bridge technical condition degradation prediction model achieves a QWK value of 70% or higher for each category of bridges. The final overall model precision is 90.53%, and the QWK value is 95.10%, demonstrating the model's good classification ability for different categories of bridges. Therefore, this bridge technical condition degradation prediction model is reliable.

[0136] To further verify the model's ability to distinguish between bridges of different technical condition levels, this embodiment also plotted ROC curves for each technical condition rating, such as... Figure 10 As shown in the figure, the horizontal axis represents the false positive rate (FPR), and the vertical axis represents the true positive rate (TPR). The AUC values ​​for each category and the baseline for random guessing are also labeled. Figure 10It can be seen that the area under the ROC curve (AUC) for the fifth type of bridge reaches 1.0000, and the AUC for the fourth type of bridge reaches 0.9998, indicating that the prediction model of this invention has a strong ability to identify these two types of bridges. Meanwhile, the AUC for the first type of bridge is 0.9844, the AUC for the third type is 0.9664, and the AUC for the second type is 0.9199, all significantly higher than the baseline of AUC=0.5 for random guessing. This demonstrates that the prediction model constructed in this invention has excellent discriminative ability and prediction accuracy for all technical condition levels, verifying the reliability and generalization ability of the model in predicting bridge degradation in actual road networks.

[0137] In another embodiment, such as Figure 11 As shown, a system for predicting the degradation of technical condition levels of small and medium-span bridges is also provided. This system is based on the same inventive concept as the above-described method embodiments and specifically includes a database construction module, a data processing module, a model construction and training module, a prediction module, and a database update module.

[0138] Specifically, the database construction module is used to collect bridge inspection data from multiple small- and medium-span bridges within the regional road network, establishing a bridge inspection database. This database includes bridge construction information, environmental data, road network data, inspection data, and maintenance management information. Bridge construction information includes the bridge name, center pile number, bridge design parameters, construction period, and ancillary facilities; environmental data describes the environmental conditions under which the bridge is maintained; road network data includes route attributes, route name, and traffic volume; inspection data includes defects found in previous inspections, technical condition scores, rating levels, and the inspection implementing unit; and maintenance management information includes past maintenance history and maintenance decisions based on inspection results.

[0139] The data processing module is used to preprocess the data in the bridge inspection database, specifically including: determining the input feature set based on the bridge inspection database, which includes basic attribute features, derived features, and category interaction features; processing missing and outlier values ​​in the input feature set to obtain mixed feature data containing numerical and nominal features; normalizing different types of feature data, where numerical features are normalized through standardization, and nominal features are processed using one-hot encoding or numerical mapping methods; using principal component analysis to reduce the dimensionality of the normalized nominal features; and using the range method to map the dimensionality-reduced data to the [0,1] interval.

[0140] The model building and training module is used to construct a prediction model that integrates an attention mechanism and a two-branch output, and to train the prediction model based on preprocessed data. The prediction model integrating the attention mechanism and the two-branch output consists of an input layer, a hidden layer, and a two-branch output layer connected in sequence. The specific structure can be found in the description in the aforementioned method embodiments, and will not be repeated here. During training, five-fold cross-validation is used to optimize hyperparameters, and a weighted random sampler is used to alleviate the class imbalance problem.

[0141] The prediction module is used to acquire the detection data of the target small-span bridge. The data processing module preprocesses the detection data and inputs the preprocessed data into the trained bridge technical condition level degradation prediction model, outputting the bridge technical condition level degradation prediction result of the target small-span bridge.

[0142] The database update module is used to add the detection data of small-span bridges in the target area and the bridge technical condition level degradation prediction results obtained by the model to the bridge detection database. This allows the model building and training module to dynamically update and iteratively optimize the existing bridge technical condition level degradation prediction model, thereby continuously improving the accuracy and generalization ability of the model in predicting bridge technical condition.

[0143] The specific implementation methods and working principles of the above modules correspond one-to-one with steps 1 to 4 in the aforementioned method embodiments, and will not be repeated here.

[0144] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0145] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A method for predicting the degradation of the technical condition level of small and medium-span bridges, characterized in that, Includes the following steps: Step 1: Collect bridge inspection data of multiple small and medium span bridges in the regional road network and establish a bridge inspection database. The bridge inspection database includes bridge construction information, bridge environment data, road network data, inspection data and maintenance management information. The process of establishing the bridge inspection database in step 1 includes the following steps: Step 1.1: Based on the inspection data of the small and medium span bridges, construct an entity relationship diagram of the bridge inspection database to describe the relationship between the database tables; Step 1.2: Based on the entity relationship diagram, design and construct database tables, including a bridge basic information table, a bridge route information table, a bridge construction unit information table, a bridge basic attribute correspondence table, a bridge inspection unit information table, a bridge component scoring table, a bridge part scoring table, a defect table, a historical maintenance information table, and a maintenance decision table. Step 2: Preprocess the data in the bridge inspection database; Step 3: Construct a bridge technical condition level degradation prediction model that integrates attention mechanism and ensemble learning, and train the bridge technical condition level degradation prediction model based on preprocessed data; Step 3 includes the following steps: Step 3.1: Construct a prediction model that integrates attention mechanism and dual-branch output based on deep learning, and optimize the model hyperparameters by using five-fold cross-validation when training the prediction model that integrates attention mechanism and dual-branch output; Step 3.2: The prediction model of the fusion attention mechanism and the dual-branch output is comprehensively evaluated by multi-dimensional evaluation indicators to obtain the bridge technical condition level degradation prediction model; The prediction model that integrates the attention mechanism and the dual-branch output includes: The input layer is used to receive the feature vectors of the bridge data; The hidden layer is used to assign weights to the input features through the attention module, automatically mining key features that significantly affect the degradation of bridge technology. Then, a backbone network containing three fully connected layers is used to extract high-order features. The fully connected layers include batch normalization layers and Dropout regularization layers, and the activation function is the GELU function. The dual-branch output layer, connected after the backbone network, includes two parallel branches: The classification branch is used to output the classification probability of the bridge's technical condition level; The ordinal regression branch is used to output the rank order score to constrain the order consistency of the predicted ranks; Step 4: Obtain the detection data of the target small-span bridge, preprocess the detection data, input the preprocessed data into the trained bridge technical condition level degradation prediction model, and output the bridge technical condition level degradation prediction result of the target small-span bridge; at the same time, add the detection data and the bridge technical condition level degradation prediction result to the bridge detection database for model iteration and update.

2. The method for predicting the degradation of technical condition level of small and medium-span bridges according to claim 1, characterized in that, The preprocessing process in step 2 specifically includes: Step 2.1: Based on the bridge detection database, determine the input feature set for prediction, which includes basic attribute features, derived features, and category interaction features; Step 2.2: Process the missing and outlier values ​​in the input feature set to obtain mixed feature data containing numerical and nominal features; Step 2.3: Normalize the feature data of different types: For numerical feature data, normalization is achieved through standardization processing; For nominal feature data, one-hot encoding is used to normalize the label data and numerical mapping is used to normalize the ordinal data. Step 2.4: Principal component analysis is used to reduce the dimensionality of the normalized nominal features; Step 2.5: Use the range method to map the dimension-reduced data to the interval [0, 1].

3. The method for predicting the degradation of technical condition level of small and medium-span bridges according to claim 1, characterized in that, The prediction model that integrates the attention mechanism and the dual-branch output also includes a dynamic weighted loss layer. The dynamic weighted loss layer uses a dynamic weight loss function, optimizes the classification accuracy through cross-entropy loss, and combines ordered loss to constrain the order relationship. In each fold of training of the five-fold cross-validation, the weights of the classification loss and the ordered loss are dynamically adjusted.

4. The method for predicting the degradation of technical condition level of small and medium-span bridges according to claim 1, characterized in that, The hyperparameters of the prediction model that integrates the attention mechanism and the bi-branch output include the learning rate, number of iterations, number of hidden layers, number of hidden layer neurons, early stopping patience value, dropout rate, and initial value of dynamic weights.

5. The method for predicting the degradation of technical condition level of small and medium-span bridges according to claim 1, characterized in that, In step 3.1, when training the prediction model that integrates the attention mechanism and the dual-branch output, a weighted random sampler is used to process the training samples to alleviate class imbalance.

6. The method for predicting the degradation of technical condition level of small and medium-span bridges according to claim 1, characterized in that, The multi-dimensional evaluation metrics include confusion matrix, F1 score, precision, recall, weighted Kappa coefficient, mean square error, root mean square error, and mean absolute error.

7. A system for predicting the degradation of technical condition levels of small and medium-span bridges, characterized in that, include: The database construction module is used to collect bridge inspection data of multiple small and medium span bridges in the regional road network and establish a bridge inspection database. The bridge inspection database includes bridge construction information, bridge environment data, road network data, inspection data and maintenance management information. The data processing module is used to preprocess the data in the bridge inspection database; The model building and training module is used to build a bridge technical condition level degradation prediction model that integrates attention mechanism and ensemble learning, and to train the bridge technical condition level degradation prediction model based on preprocessed data. The prediction module is used to acquire the detection data of the target small-span bridge, preprocess the detection data through the data processing module, input the preprocessed data into the trained bridge technical condition level degradation prediction model, and output the bridge technical condition level degradation prediction result of the target small-span bridge. The database update module is used to add the detection data and the bridge technical condition level degradation prediction results to the bridge detection database for model iteration and updates.

Images