Bridge technology condition rating prediction method and system fused with maintenance history information

Abstract

The present disclosure relates to a bridge technical condition rating prediction method and system fusing maintenance history information. The method comprises: constructing a bridge maintenance history information dataset based on the profile coefficient method and the K-means algorithm, and constructing a bridge inspection information database; discretizing the numerical data after data cleaning to convert it into classification data; calculating the Cramer's V coefficient between all classification data and bridge technical condition rating information, selecting feature data based on this, converting data with classification data attributes into one-hot encoding, and reducing the dimension using the principal component analysis method to obtain reduced classification data; normalizing and equalizing to obtain a dataset for model training; training a bridge technical condition rating prediction model using the PSO-LightGBM algorithm; and predicting the bridge technical state rating level using the prediction model. The present disclosure improves the accuracy of bridge technical condition rating prediction.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the field of bridge technical condition prediction, in particular to a bridge technical condition rating prediction method and system fusing maintenance history information. BACKGROUND

[0002] As a key component of transportation infrastructure, bridges play an important role in the transportation network. Bridges gradually decline in service performance during service due to factors such as material aging, environmental effects, and traffic loads. In China, the service performance of bridges is mainly reflected by bridge technical condition rating, so accurate prediction of bridge technical condition rating will help traffic management departments effectively grasp the changing trend of bridge service performance and facilitate scientific bridge maintenance decisions.

[0003] Bridge technical condition rating prediction models can be roughly divided into deterministic models, stochastic models, and artificial intelligence models.

[0004] Deterministic models use mathematical methods to describe the degradation state of bridge service performance, such as regression methods. For example, the Chinese patent with publication number CN114626119A and title "Maintenance intelligent decision-making method and system based on bridge big data" establishes a mathematical model for predicting bridge structural disease indicators through curve fitting. Deterministic models have high efficiency when facing large-scale road network prediction and evaluation, but their defects are also very obvious, such as lack of uncertainty, only predicting overall average results, etc.

[0005] Stochastic models are mathematical models that describe random processes or phenomena, such as Markov models. For example, the master's thesis "Liu Zhuocheng. Road network level bridge degradation prediction and auxiliary decision-making system design research [D]. Chongqing Jiaotong University, 2022" applies a Markov model that optimizes inverse matrix to solve transition matrix to actual road network level bridge service performance degradation prediction. Although the Markov model considers the uncertainty in degradation and the influence of the current state on the future, the assumptions of discrete transition time interval, constant sample size, and fixed transition probability are not scientific; and the Markov model believes that the future state of the structure is only related to the current state, which is not consistent with the actual situation.

[0006] Machine learning models can describe the bridge service performance degradation behavior using methods such as decision trees and neural networks. Compared with the previous models, they have more advantages and fewer limitations. For example, a Chinese patent with the title "A testing method for bridge technical condition" (CN109102016A) establishes an LSTM neural network to predict the technical condition rating of a bridge and verifies the rationality of the prediction results by setting the proportion of bridge grades in the prediction results. A Chinese patent with the title "Bridge technical condition prediction method and system based on machine learning" (CN118606847A) uses the XGBoost ensemble learning method to build a bridge technical condition rating prediction model, with a prediction accuracy of 87.3%.

[0007] However, during the establishment of machine learning models, differences in various aspects such as original data acquisition, data processing, feature selection, and model training can affect the training accuracy and prediction accuracy of the model. Existing machine learning models for predicting bridge technical condition ratings still have problems with incomplete consideration of influencing factors and their correlations, resulting in prediction accuracy below 90%. For example, a master's thesis "Jiang Jia-heng. Bridge component degradation prediction based on BP neural network and MCMC [D]. Guizhou University, 2022" builds a neural network model and predicts future bridge ratings based on the increase in bridge age and traffic volume. However, due to the lack of comprehensive consideration of influencing factors and feature selection correlations in this study, the prediction accuracy is relatively low.

[0008] Generally, during the long-term service of a bridge, the traffic management department will carry out corresponding maintenance and repair work based on the results of previous bridge detection and technical condition evaluation to improve the service performance of the bridge and improve the next detection results. Therefore, the maintenance and repair history information during the service of the bridge is an important factor that needs to be considered in the prediction of bridge technical condition ratings. For example, the paper "Xia Ye, Lei Xiaoming, Wang Peng, et al. Regional bridge degradation modeling and application for network-level evaluation [J]. Journal of Central South University (Natural Science Edition), 2021, 52(3): 828-838" published in the Journal of Central South University (Natural Science Edition) and the paper "Xia Ye, Wang Peng, Sun Limin. Bridge network-level evaluation method based on multi-source information [J]. Journal of Tongji University (Natural Science Edition), 2019, 47(11): 1574-1584" published in the Journal of Tongji University (Natural Science Edition) both use Boolean data to represent whether there was maintenance information on the bridge in the previous year, with 1 indicating that there was maintenance information and 0 indicating that there was no maintenance information. However, the processing of the above maintenance history information is too simple and does not take into account the different impacts of different maintenance levels on the service performance of the bridge.

[0009] In actual engineering, the bridge inspection information database obtained usually includes multiple characteristics, but not all of them are related to the bridge technical condition rating. Therefore, it is necessary to select the characteristics by a suitable method to improve the prediction accuracy of the bridge technical condition rating. In previous studies, Pearson correlation coefficient is often used to analyze the correlation between variables. For example, in the master's thesis "Li Wei. Prediction of Bridge Time-varying Technical Condition Based on Deep Learning Theory[D]. Sichuan University, 2023", Pearson correlation coefficient is used to analyze the correlation between bridge age, average daily traffic volume, structure type and bridge technical condition rating, which is used as the basis for feature selection. However, Pearson correlation coefficient is only suitable for analyzing numerical variables or ordered categorical variables. For some unordered categorical variables (such as structure type, bridge material type, etc.), Pearson correlation coefficient cannot correctly reflect the correlation between variables, and a correlation analysis method suitable for unordered categorical variables is needed for feature selection.

[0010] Currently, the bridge technical condition rating prediction research based on machine learning mainly has the following problems: ignoring the influence of maintenance history information or simply processing the maintenance history information, which cannot reflect the influence of different maintenance levels on the bridge technical condition rating results; and not considering the influence of unordered categorical variables on feature selection. This leads to low accuracy of bridge technical condition rating prediction, which needs to be improved. SUMMARY

[0011] Therefore, it is necessary to provide a bridge technical condition rating prediction method and system that integrates maintenance history information to solve the above problems.

[0012] To solve the above problems, the present disclosure adopts the following technical solutions:

[0013] In a first aspect, a bridge technical condition rating prediction method integrating maintenance history information is provided, comprising the following steps:

[0014] Step 1: Obtain multiple bridge maintenance information records, construct a bridge maintenance history information dataset including bridge maintenance information record classification category information based on the contour coefficient method and the K-means algorithm, and the classification category types include maintenance cost clustering category, reconstruction category, and no maintenance category;

[0015] Step 2: Construct a bridge inspection information database including the bridge maintenance history information dataset, bridge basic information, bridge design information, bridge deck arrangement information, and traffic statistics information;

[0016] Step 3: Classify the data in the bridge inspection information database into numerical data and categorical data, and perform data cleaning;

[0017] Step 4, the numerical data after data cleaning is discretized to become categorical data; calculate the Cramer's V coefficient between all categorical data and the bridge technical condition rating information, and the data corresponding to the bridge inspection information database obtained in step 3 with a Cramer's V coefficient greater than the coefficient threshold are selected as the feature data;

[0018] Step 5, the data with categorical data attributes in the selected feature data is converted into one-hot encoding and the principal component analysis method is used to reduce the dimension to obtain the reduced dimension categorical data; the bridge inspection information data set is normalized, and the bridge inspection information data set includes the data with numerical data attributes in the selected feature data and the reduced dimension categorical data;

[0019] Step 6, the normalized bridge inspection information data set is balanced by using the SMOTE oversampling method and the NearMiss undersampling method to obtain the normalized balanced bridge inspection information data set ;

[0020] Step 7, the PSO-LightGBM algorithm and are used to train the bridge technical condition rating prediction model;

[0021] Step 8, the bridge technical condition rating prediction model is used to predict the bridge technical condition rating level.

[0022] In a preferred embodiment, part of the bridge maintenance information records include bridge technical condition rating information, and do not include maintenance cost information and reconstruction records; part of the bridge maintenance information records include bridge technical condition rating information and maintenance cost information, and do not include reconstruction records; part of the bridge maintenance information records include bridge technical condition rating information and reconstruction records;

[0023] The step 1 includes:

[0024] The number of maintenance cost clustering categories is determined based on the contour coefficient method and the K-means algorithm ;

[0025] The bridge maintenance information record with the reconstruction record is marked as the reconstruction class, and is recorded as the first The bridge maintenance information record without the reconstruction record and without the maintenance cost information is marked as the non-maintenance type, recorded as the 0th type; the bridge maintenance information record satisfying the conditions of no increase of the bridge technical condition rating compared with the bridge technical condition rating of the previous bridge maintenance information record, the maintenance cost being greater than or equal to the maintenance cost of the previous bridge maintenance information record, and no reconstruction record is marked as the non-maintenance type, recorded as the 0th type; the bridge maintenance information record without the reconstruction record and with the maintenance cost being less than the maintenance cost of the previous bridge maintenance information record is subjected to the determination of the maintenance cost clustering type, and is marked as the 0th type in the 0th type.

[0026] In a preferred embodiment, the bridge basic information includes at least one of the bridge age, the longitude, the latitude, the structure classification type, and the material classification type, the bridge design information includes at least one of the design load, the structure total length, the maximum span, the main bridge span number, the approach bridge span number, the skew angle, the minimum lateral clearance, the minimum vertical clearance, and the total horizontal clearance, the bridge deck arrangement information includes at least one of the bridge deck pavement structure type, the bridge deck pavement material type, the bridge deck total width, the road surface width of the vehicle lane, the number of upper vehicle lanes, and the number of lower vehicle lanes, and the traffic statistics information includes at least one of the average daily traffic volume, the average daily truck traffic volume ratio, and the traffic volume statistics year.

[0027] In a preferred embodiment, in the step 3, the linear interpolation method is used for filling in the missing values of the numerical data, the Z-score method is used for removing the outliers of the numerical data, and the mode filling method is used for filling in the missing values of the classification data.

[0028] In a preferred embodiment, the discretization processing of the numerical data after data cleaning to convert into classification data specifically includes:

[0029] The numerical data after data cleaning is arranged in ascending order to form a numerical data set ;

[0030] Based on , the method for determining the optimal segmentation point to divide the subsets based on the Gini index is used, and according to the judgment criterion that the difference between the Gini index of the subset and the Gini index of the optimal segmentation point of the subset is less than the preset difference threshold, a plurality of final subsets without intersection are determined, and the collection of all the final subsets is the .

[0031] In a preferred embodiment, the data with the classification data attribute in the selected feature data is converted into the one-hot encoding and the principal component analysis method is used to reduce the dimension to obtain the reduced dimension classification data specifically includes:

[0032] ​​Transforming the selected feature data with categorical data attribute into one-hot encoding matrix Subtracting the mean of each column from each data in the matrix to get the decentered matrix Calculating the covariance matrix of the matrix ;

[0033] Performing eigenvalue decomposition on the covariance matrix , where is the matrix composed of eigenvectors, is the diagonal matrix composed of eigenvalues, and the first eigenvectors corresponding to the largest eigenvalues are selected to form the matrix ; ;

[0034] According to , the matrix is calculated, which is the matrix composed of the reduced categorical data.

[0035] In a preferred embodiment, the step 6 comprises:

[0036] Dividing the normalized bridge inspection information dataset into a minority class sample dataset and a majority class sample dataset ;

[0037] Calculating the Euclidean distance between each sample dataset and all other sample datasets in ;

[0038] For each sample dataset , the nearest sample datasets are selected as its nearest neighbor sample dataset according to the Euclidean distance ;

[0039] From each nearest neighbor sample dataset , a sample dataset is randomly selected , and a new sample dataset is generated according to , is the new dataset generated by interpolation, is a random number in the interval (0, 1), and and all are merged to obtain the oversampled minority class sample dataset ; ​​​

[0040] computing for each sample data set to the minority sample data set Euclidean distance;

[0041] for each sample data set according to its Euclidean distance to select the nearest neighbor the smallest Euclidean distance as the nearest distance, calculate the average value of the Euclidean distance of its nearest neighbor ;

[0042] sort all sample data sets in ascending order of size, before sorting sample data sets constitute the undersampled majority sample data set ;

[0043] merge the oversampled minority sample data set and the undersampled majority sample data set to obtain a normalized and balanced bridge inspection information data set .

[0044] In a preferred embodiment, the formula of the bridge technical condition rating prediction model is:

[0045]

[0046] wherein, is the bridge technical condition rating prediction value, is the feature vector set of the normalized and balanced bridge inspection information data set, is the decision tree function, denotes the decision tree data space.

[0047] In a preferred embodiment, the input of the bridge technical condition rating prediction model includes the bridge basic information, bridge design information, bridge deck arrangement information, and traffic statistics information of the bridge to be predicted, and the input also includes the classification category information of the bridge maintenance information record of the bridge to be predicted. The output of the model is the bridge technical condition rating level and its probability.

[0048] In a second aspect, a bridge technical condition rating prediction system integrating maintenance history information is provided, comprising:

[0049] ​A dataset construction module is configured to obtain a plurality of bridge maintenance information records, and construct a bridge maintenance history information dataset including classified category information of the bridge maintenance information records based on a contour coefficient method and a K-means algorithm, wherein the classified categories include a maintenance cost clustering category, a reconstruction category, and a non-maintenance category.

[0050] A database construction module is configured to construct a bridge inspection information database including the bridge maintenance history information dataset, bridge basic information, bridge design information, bridge deck arrangement information, and traffic statistical information.

[0051] A data cleaning module is configured to classify data in the bridge inspection information database into numerical data and classified data, and clean the data.

[0052] A feature data selection module is configured to discretize the numerical data after data cleaning to convert the numerical data into classified data, calculate Cramer's V coefficients between all classified data and bridge technical condition rating information, and select data in the bridge inspection information database after data cleaning corresponding to classified data with a Cramer's V coefficient greater than a coefficient threshold as selected feature data.

[0053] A data dimension reduction and normalization module is configured to convert data with classified data attributes in the selected feature data into one-hot encoding, reduce dimensions by using a principal component analysis method to obtain reduced classified data, and normalize the bridge inspection information dataset, wherein the bridge inspection information dataset includes data with numerical data attributes in the selected feature data and the reduced classified data.

[0054] A data balancing module is configured to balance the normalized bridge inspection information dataset by using a SMOTE oversampling method and a NearMiss undersampling method to obtain a normalized and balanced bridge inspection information dataset. ;

[0055] A model training module is configured to train a bridge technical condition rating prediction model by using a PSO-LightGBM algorithm and .

[0056] A prediction module is configured to predict a bridge technical condition rating level by using the bridge technical condition rating prediction model.

[0057] The bridge maintenance history information fusion bridge technical condition rating prediction method and system generates a bridge maintenance history information dataset including bridge maintenance information records and classification information of the bridge maintenance information records based on the profile coefficient method and the K-means algorithm, then constructs a bridge inspection information database, comprehensively considers the information of the bridge, especially fully considers the maintenance history information; the numerical data is discretized, then the Cramer's V coefficient of all data is calculated, the correlation index of all data including unordered classification data is unified, the influence of unordered classification variables on feature selection is considered, the selected feature data is more accurate based on this, and the accuracy of model prediction is further improved; the bridge technical condition rating prediction model is trained through the PSO-LightGBM algorithm and the normalized balanced bridge inspection information dataset, and the accuracy of predicting the bridge technical condition rating is improved. The disclosure improves the accuracy of predicting the bridge technical condition rating by considering the data of the maintenance history information, optimizing the data processing, and optimizing the model training. BRIEF DESCRIPTION OF DRAWINGS

[0058] Figure 1 The flowchart of the method in an embodiment of the disclosure is shown;

[0059] Figure 2 The flowchart of acquiring the bridge maintenance history information dataset in an embodiment of the disclosure is shown;

[0060] Figure 3 The profile coefficient method result graph in an embodiment of the disclosure is shown;

[0061] Figure 4 The Cramer's V coefficient calculation result graph in an embodiment of the disclosure is shown;

[0062] Figure 5 The model prediction confusion matrix graph in an embodiment of the disclosure is shown;

[0063] Figure 6 The bridge technical condition rating prediction result graph in an embodiment of the disclosure is shown;

[0064] Figure 7 The structure diagram of the system in an embodiment of the disclosure is shown. DETAILED DESCRIPTION

[0065] The technical solutions of the disclosure will be described in detail below with reference to the drawings and preferred embodiments.

[0066] Referring to Figure 1 The disclosure provides a bridge technical condition rating prediction method fusing maintenance history information, including:

[0067] Step 1, based on the contour coefficient method and K-means algorithm, a bridge maintenance history information dataset including classification category information of bridge maintenance information records is constructed, and the classification category includes reconstruction category, no maintenance category, or maintenance cost clustering category;

[0068] Step 2, a bridge inspection information database including the bridge maintenance history information dataset, bridge basic information, bridge design information, bridge deck arrangement information, and traffic statistics information is constructed;

[0069] Step 3, data in the bridge inspection information database is classified into numerical data and classification data, and data cleaning is performed;

[0070] Step 4, numerical data after data cleaning is discretized to convert into classification data; Cramer's V coefficient between all classification data and bridge technical condition rating information is calculated, and the bridge inspection information database obtained in step 3 (i.e., the bridge inspection information database after data cleaning) corresponding to classification data with a Cramer's V coefficient greater than a coefficient threshold is selected as selected feature data;

[0071] Step 5, data with classification data attributes in the selected feature data is converted into one-hot encoding and reduced in dimension by principal component analysis to obtain reduced classification data; the bridge inspection information dataset is normalized to obtain a normalized bridge inspection information dataset, which includes data with numerical data attributes in the selected feature data and also includes reduced classification data;

[0072] Step 6, the normalized bridge inspection information dataset is balanced by SMOTE oversampling method and NearMiss undersampling method to obtain a normalized and balanced bridge inspection information dataset ;

[0073] Step 7, a bridge technical condition rating prediction model is trained by using a PSO-LightGBM algorithm and the normalized and balanced bridge inspection information dataset ;

[0074] Step 8, the bridge technical condition rating prediction model is used to predict the bridge technical condition rating level.

[0075] In an embodiment, the bridge technical condition rating prediction method fusing maintenance history information specifically includes:

[0076] Step 1, based on the contour coefficient method and K-means algorithm, a bridge maintenance history information dataset is generated according to multiple bridge maintenance information records in a bridge management database; the bridge maintenance history information dataset includes bridge maintenance information records and classification information of the bridge maintenance information records.

[0077] It can be understood that all the bridge maintenance information records applied in step 1 include bridge technical condition rating information, maintenance cost information and reconstruction records.

[0078] The bridge maintenance information records are in the following three situations: part of the bridge maintenance information records include bridge technical condition rating information, do not include maintenance cost information and reconstruction records; part of the bridge maintenance information records include bridge technical condition rating information and maintenance cost information, do not include reconstruction records; and part of the bridge maintenance information records include bridge technical condition rating information and reconstruction records (not limited whether to include maintenance cost information, usually also include maintenance cost information).

[0079] The bridge maintenance information records include bridge technical condition rating information, or the bridge maintenance information records include bridge technical condition rating information and maintenance cost information, or the bridge maintenance information records include bridge technical condition rating information and reconstruction records, or the bridge maintenance information records include bridge technical condition rating information, maintenance cost information and reconstruction records. That is, usually, part of the bridge maintenance information records in the bridge management database only include bridge technical condition rating information among the three of bridge technical condition rating information, maintenance cost information and reconstruction records; another part of the bridge maintenance information records in the bridge management database include bridge technical condition rating information, also include maintenance cost information and / or reconstruction records.

[0080] The number of cluster categories is determined based on the silhouette coefficient method and the K-means algorithm as The bridge maintenance information records with reconstruction records are marked as the first category, the first category is a reconstruction category, that is, marked as a reconstruction category; the bridge maintenance information records without reconstruction records and also without maintenance cost information are marked as the 0th category, the 0th category is a non-maintenance category; the bridge maintenance information records satisfying the three requirements of ① no reconstruction record, ② maintenance cost (usually the maintenance cost information corresponds to the maintenance cost, that is, the total maintenance cost) is greater than or equal to the maintenance cost of the previous bridge maintenance information record, and ③ bridge technical condition rating does not rise compared with the bridge technical condition rating of the previous bridge maintenance information record are marked as the 0th category, that is, marked as a non-maintenance category; the bridge maintenance information records satisfying the two requirements of ① no reconstruction record and ② maintenance cost is less than the maintenance cost of the previous bridge maintenance information record are determined according to the silhouette coefficient method and the K-means algorithm for maintenance cost cluster category, and are marked as one of the cluster categories, the number of cluster categories of the K-means algorithm is , and are marked as the first one of the classes. Here, the maintenance cost clustering class is a unit area maintenance cost clustering class.

[0081] Referring to Figure 2 , the specific flow of step 1 is:

[0082] Step 1.1, the bridge management database includes bridge maintenance information records, which can be understood as bridge maintenance information records are sequentially arranged, and the sorting method is not limited. The is the total number of bridge maintenance information records, which is a positive integer, and can be ; or In this embodiment, as an example, the bridge management database is the NBI database (National Bridge Inventory, USA).

[0083] The following operations are performed on each bridge maintenance information record in the bridge management database:

[0084] Select the th bridge maintenance information record;

[0085] Determine whether the th bridge maintenance information record has a reconstruction record. If yes, mark the bridge maintenance information record as "reconstruction" maintenance information; if no, determine whether the th bridge maintenance information record has maintenance cost;

[0086] Determine whether the th bridge maintenance information record has maintenance cost. If no, mark the bridge maintenance information record as "no maintenance" maintenance information; if yes, determine whether the maintenance cost in the th bridge maintenance information record is lower than the maintenance cost in the th bridge maintenance information record;

[0087] Determine whether the maintenance cost of the th bridge maintenance information record is lower than the maintenance cost of the th bridge maintenance information record. If yes, mark the bridge maintenance information record as the maintenance information of the unit area maintenance cost of the th bridge maintenance information record (i.e. the maintenance information of the unit area maintenance cost before the decrease); if no, determine whether the bridge technical condition rating of the i th bridge maintenance information record is lower than the bridge technical condition rating of the Whether the bridge technical condition rating recorded in the bridge maintenance information has increased; the maintenance cost per unit area is the maintenance cost (the cost recorded in the maintenance cost information) divided by the bridge area; the bridge area is the product of the total length of the bridge structure and the total width of the bridge deck; it can be understood that this includes the step of calculating the maintenance cost per unit area based on the bridge area and the maintenance cost information.

[0088] Judge the first The bridge technical condition rating recorded in the bridge maintenance information record is relative to the first bridge. If the status of a bridge maintenance record is not updated, then mark the record as "no maintenance"; if it is updated, then delete the record.

[0089] judge Is it greater than or equal to? If yes, then the acquisition of bridge maintenance information records will end, and the above judgment will conclude; if no, then the original... Add 1 to the value as the new Value (corresponding) Figure 2 of" The above judgment steps are repeated; this step is performed after each bridge maintenance information record is judged, and in some embodiments, it also proceeds to the step of deleting the bridge maintenance information record. For example, after the maintenance information marked as "reconstruction" is completed, the maintenance information marked as "no maintenance" is completed, and the record marked as "the first" is deleted. Once the maintenance information for the unit area maintenance cost of the item is completed.

[0090] In this embodiment, the above judgment is not performed for the first bridge maintenance information record. In other embodiments, the above judgment and classification may also be performed, that is... So when it is needed and the first If the comparison is not performed, it can be defined as not classifying the bridge maintenance information records, or directly clustering and classifying them based on the maintenance cost per unit area in step 1.2.

[0091] Understandably, this is an example and not a limitation, if the first If the bridge maintenance information record does not contain maintenance costs, then the process will end directly. The judgment of the bridge maintenance information record will directly jump to the original Add 1 to the value as the new Value steps, with the latest The above judgment steps are then repeated.

[0092] Step 1.2: Determine the optimal number of clusters for the bridge's maintenance cost per unit area using the profile coefficient method. The K-means algorithm was used to cluster the maintenance information of the unit area maintenance cost before the decrease into 1~ Class, corresponding Figure 2 The maintenance information based on the silhouette coefficient method and K-means algorithm is clustered into the first to third groups based on the maintenance cost per unit area. Class; Maintenance information marked as "no maintenance" and "reconstruction" are marked as Class 0 and Class 1, respectively. The class ultimately constructs a bridge maintenance history information dataset, corresponding to Figure 2 The "reconstruction" maintenance information is marked as the first. "+1 category" and "mark maintenance information with "no maintenance" as category 0".

[0093] See Figure 3 The figure shows the relationship between the number of clusters (i.e., the number of clusters) and the silhouette coefficient of the K-means algorithm in a specific embodiment. It can be seen that the silhouette coefficient is highest when the number of clusters for the maintenance cost per unit area of ​​the bridge is 5. Therefore, in this embodiment, =5, then the specific clustering categories of reconstruction, no maintenance, and maintenance cost are as follows:

[0094] The category of "no maintenance" is category 0 (Category 0). The specific clustering categories for maintenance costs are category 1 (Category 1), category 2 (Category 2), category 3 (Category 3), category 4 (Category 4), and category 5 (Category 5). The category of "reconstruction" is category 6 (Category 6). Therefore, any bridge maintenance information record can be classified into categories 0 to 6 (i.e., 0~). One of them.

[0095] Finally, a bridge maintenance history information dataset was constructed, and the classification results of the maintenance information are shown in Table 1.

[0096] Table 1

[0097]

[0098] Step 2: Construct a bridge scheduled inspection information database, which includes basic bridge information, bridge design information, bridge deck layout information, traffic statistics information, and the bridge maintenance history information dataset.

[0099] The bridge basic information in step 2 includes at least one of bridge age, longitude where the bridge is located, latitude where the bridge is located, structure classification type, material classification type, the bridge design information includes at least one of design load, structure full length, maximum span, main bridge span number, approach bridge span number, skew angle, minimum lateral clearance, minimum vertical clearance, total horizontal clearance, the bridge deck arrangement information includes at least one of bridge deck pavement structure type, bridge deck pavement material type, bridge deck total width, carriageway pavement width, upper carriageway number, lower carriageway number, the traffic statistics information includes at least one of average daily traffic volume, average daily truck traffic volume proportion and traffic volume statistics year, and the bridge inspection information database has s a plurality of records and u a plurality of fields;

[0100] In step 3, the data in the bridge inspection information database is classified into two data attributes of numerical value type data and classification data, data cleaning of the numerical value type data includes missing value filling and abnormal value removal, and data cleaning of the classification data includes missing value filling.

[0101] The data set composed of the numerical value type data is classified as numerical value type data , and the data set composed of the classification data is classified as classification data

[0102] The linear interpolation method is used to fill the missing values in the data set with numerical value type data attribute in the field of the bridge inspection information database:

[0103]

[0104] wherein, is the position sequence number of the missing value in the data set , the filling value after linear interpolation is the numerical value of the first element in the data set , is the position sequence number of the previous valid element with the numerical value of in the data set, and is the position sequence number of the next valid element with the numerical value of in the data set.

[0105] The Z-score method is used to remove the abnormal values in the data set with numerical value type data attribute in the field of the bridge inspection information database, and the threshold is set to 3 times the standard deviation:

[0106]

[0107] wherein, represents the data set ​​​​​order number of the middle element, representing the mean of the data set , representing the standard deviation of the data set , is a standard score, representing the distance of the th element in the data set from the mean in units of the standard deviation , if , then the element is determined to be an outlier and is removed from the data set ;

[0108] The mode filling method is used to fill in the missing values in the data set with classification data attribute in the field of the bridge inspection information database, and the mode filling method is to take the mode of the data set in the field as the filling of the missing values.

[0109] Step 4: First, the numerical data after data cleaning in step 3 is discretized to convert into classification data, and then the Cramer's V coefficient between all classification data (original classification data after classification in step 3 and classification data after discretization in step 4) and the bridge technical condition rating is calculated. The classification data in step 4 is referred to as first classification data, and feature selection is performed by setting a threshold value of the Cramer's V coefficient. The data corresponding to the first classification data with a coefficient greater than the threshold value is selected as the selected feature data. The corresponding data is the data in the bridge inspection information database after final data cleaning (i.e., after data cleaning) in step 3.

[0110] The specific steps for converting the numerical data after data preprocessing in step 3 into classification data in step 4 are as follows:

[0111] Step 41: Discretize the numerical data after data cleaning to convert into classification data: arrange the numerical data after data cleaning in ascending order to form an dimensional numerical data set .

[0112] Based on , the method for determining the optimal split point to divide the subsets based on the Gini index is used. According to the judgment criterion that the difference between the Gini index of a subset and the Gini index of the optimal split point of the subset is less than a preset threshold, a plurality of final subsets without intersection are determined. The collection of all the final subsets is the .

[0113] The method for determining the optimal segmentation point based on the Gini index comprises: calculating the average value of each two adjacent elements in the set as a candidate segmentation point, dividing the set into two subsets with the candidate segmentation point, calculating the Gini index of the candidate segmentation point, selecting the candidate segmentation point corresponding to the minimum Gini index as the optimal segmentation point, and dividing the two subsets with the optimal segmentation point. The two subsets divided by the optimal segmentation point are used as the set for the next execution of the method for determining the optimal segmentation point based on the Gini index or as the final subsets.

[0114] Step 4.1 specifically comprises:

[0115] Step 4.1.1, using the data set of the bridge technical condition rating information in the bridge inspection information database as the label for decision tree training, and using the data set of the numerical data (numerical data after data cleaning) finally obtained in step 3 arranging in ascending order to form dimensional numerical data set , denoted as , the is the number of data in the data set ;

[0116] Step 4.1.2, in the sorted numerical data set , a candidate segmentation point is generated for each two adjacent elements , the candidate segmentation point of each two adjacent elements is the average value of the two adjacent elements, wherein, , indicates the th candidate segmentation point; each candidate segmentation point divides the numerical data set into two subsets (numerical ) and (numerical ), and the Gini index of the candidate segmentation points in the data set is calculated one by one:

[0117]

[0118] wherein, indicates the number of elements of the data set , indicates the number of elements of the data set , indicates the number of elements of the data set , indicates the Gini index of the subset , Representing a subset The Gini index is calculated using the following formula: , , The total number of levels for rating the technical condition of bridges. This indicates the bridge's technical condition rating level number, typically corresponding to the [number missing]. The bridge's technical condition level is referred to as the [number]. Level of technical condition rating For subset The Middle The number of elements in the technical condition rating as a subset The proportion of the total number of elements For subset The Middle The number of elements in the technical condition rating as a subset The proportion of the total number of elements; based on calculations The Gini index of each candidate split point is used to select the candidate split point with the minimum Gini index as the optimal split point. (First optimal split point) to divide the dataset Divided into and Two subsets;

[0119] Step 4.1.3, for the partitioned subsets and ( and Using the average of every two adjacent elements in the computation set as the "set" of candidate split points, the steps described above for determining the optimal split point based on the Gini index are repeated to determine the subset. Optimal split point Determine the subset Optimal split point Determine subsets Gini index with subset Optimal split point Gini index If the difference between the two Gini indices is greater than a preset difference threshold (threshold is 0.05), and if it is, the optimal split point selection step is repeated to continue splitting the subset, i.e., the subset is further divided using the method of determining the optimal split point based on the Gini index; the subset is then judged. Gini index with subset Optimal split point Gini index whether the difference is greater than a preset difference threshold (the difference threshold is 0.05), if greater than the preset difference threshold, repeating the above operation, continue to divide the subset by the method of determining the optimal split point based on the Gini index; if less than or equal to the preset difference threshold, the purity of the classification of the subset has reached the satisfaction and stops the subsequent partition of the subset, realizing the discretization processing into classification data.

[0120] Taking the bridge span as an example, the data range of the bridge span is 6-170m, and the 7 split points obtained by discretization processing the bridge span into classification data are 7.15m, 11.1m, 19.35m, 24.2m, 37.95m, 41.25m, and 120.85m. The bridge span is divided into 8 groups.

[0121] Step 4.2, calculate the Cramer's V coefficient between all classification data and the bridge technical condition rating information, and the data in the bridge inspection information database corresponding to the classification data with the Cramer's V coefficient greater than the coefficient threshold is selected as the selected feature data.

[0122] All classification data includes classification data after data cleaning in step 3 and classification data obtained by discretization processing in step 4.1. All classification data can form a contingency table, and then the Cramer's V coefficient is calculated.

[0123] Specifically, a contingency table is constructed for the classification data row column, and then the Cramer's V coefficient is calculated based on the contingency table.

[0124] In an embodiment, the Cramer's V coefficient between the classification data after data cleaning in step 3 and the bridge technical condition rating information is calculated, the Cramer's V coefficient between the classification data obtained by discretization processing in step 4.1 and the bridge technical condition rating information is calculated, and the threshold value of the Cramer's V coefficient is set to select the features of the bridge inspection information database, and the bridge inspection information database is selected from the bridge inspection information database. The specific steps are as follows:

[0125] Step 4.2.1, taking the data set of the bridge technical condition rating field in the bridge inspection information database as the target of association analysis. Based on the original classification data in step 2 and the classification data obtained by discretization processing, a contingency table is constructed row column, hereinafter referred to as the original classification data contingency table for the classification data after data cleaning, and the new classification data contingency table for the classification data obtained by transformation. Among them, the total number of categories of the classification data, the total number of categories (total levels) of the bridge technical condition rating; it can be understood that it is only used to represent the number of rows and columns of any contingency table.

[0126] Step 4.2.2, calculating the Pearson chi-square value between the classification data and the bridge technical condition rating

[0127]

[0128] wherein, is the cross frequency of the classification data of the category and the bridge technical condition rating data of the j level, , , is the expected frequency, , represents the row total of the classification data of the category, represents the column total of the bridge technical condition rating of the level, is the total number of samples of the contingency table;

[0129] Step 4.2.3, calculating the Cramer's V coefficient:

[0130]

[0131] wherein, represents the Cramer's V coefficient, represents the minimum value under the comparison of the number of rows and the number of columns of the contingency table, ; if the Cramer's V coefficient is greater than or equal to a set coefficient threshold value, it indicates that the classification data of the field has strong correlation with the bridge technical condition rating information, and the data corresponding to the classification data in the bridge inspection information database is selected as the feature data selected from the bridge inspection information database.

[0132] The strength (correlation strength) of the association between the value of the Cramer's V coefficient and the bridge technical condition rating information is shown in Table 2, and when calculating, ≥ 0.1 can indicate that the feature has correlation with the bridge technical condition rating. In the embodiment, the threshold value of the Cramer's V coefficient is set to 0.1, and the data with a threshold value greater than 0.1 is selected as the input feature of the model. The calculation result of the Cramer's V coefficient is shown in Figure 4

[0133] Table 2​​

[0134]

[0135] Step 5, all selected feature data constitute bridge inspection information dataset, the data with classification data attribute in the bridge inspection information dataset constitute classification dataset, the classification dataset is converted into one-hot encoding and the dimension is reduced by principal component analysis method to obtain reduced classification data; the bridge inspection information dataset is normalized by Min-Max method to obtain normalized bridge inspection information dataset, the bridge inspection information dataset includes reduced classification data and data with numerical data attribute in the selected feature data. The specific steps are as follows:

[0136] Step 5.1, the matrix composed of the data set with classification data attribute in the field of the bridge inspection information dataset after being converted by one-hot encoding is denoted as , the matrix after being decentered by subtracting the mean value of the column where each data is located is denoted as , the covariance matrix of the matrix is denoted as :

[0137]

[0138] Wherein, represents the number of columns of the matrix , also represents the number of rows and columns of the covariance matrix denoted as ; represents the function for calculating the covariance matrix, ; and are random variables, is the expectation of , is the expectation of , represents the number of samples, represents the total number of samples, represents the th sample value in , represents the th sample value in , and respectively represent the sample mean of and .

[0139] Step 5.2, the eigenvalue decomposition of the covariance matrix is performed, wherein, represents the matrix composed of eigenvectors, denotes a diagonal matrix composed of eigenvalues, according to the eigenvalues in the descending order of size, the eigenvectors corresponding to the first eigenvalues are selected to form a matrix , denotes the eigenvector corresponding to the first largest eigenvalue in the denotes the eigenvector corresponding to the second largest eigenvalue in the denotes the eigenvector corresponding to the third largest eigenvalue in the denotes the eigenvector corresponding to the third largest eigenvalue in the denotes the eigenvector corresponding to the third largest eigenvalue in the denotes the eigenvector corresponding to the third largest eigenvalue in the According to the calculation matrix , the matrix is composed of the reduced classification data;

[0140] Step 5.3, the Min-Max method is used for normalization processing on the bridge inspection information data set, and a normalized bridge inspection information data set is obtained, and the calculation formula of the Min-Max method is:

[0141]

[0142] wherein, denotes the normalized bridge inspection information data set, is the bridge inspection information data set before normalization, denotes the minimum value in denotes the maximum value in denotes the maximum value in

[0143] Step 6, the SMOTE oversampling method (Synthetic Minority Over-sampling Technique, synthetic minority over-sampling technique) and the NearMiss undersampling method are used for equalization processing on the normalized bridge inspection information data set, and a normalized and balanced bridge inspection information data set is obtained

[0144] The NearMiss undersampling method (NearMiss algorithm based on undersampling) is an undersampling scheme based on KNN algorithm, and random undersampling is used for majority class sample data to achieve data balance.

[0145] The specific steps of the step 6 are:

[0146] Step 6.1, the normalized bridge inspection information data set is divided into a minority class sample data set and a majority class sample data set ;​​

[0147] Step 6.2, Calculation Each (minority class) sample dataset and The Euclidean distance to all other sample datasets is calculated using the following formula:

[0148]

[0149] in, and express The number of the sample dataset, express The Middle A sample dataset, express The Middle A sample dataset, express and European distance, and , This refers to the total number of fields in the normalized bridge inspection information dataset, i.e., the number of datasets with different feature dimensions. Indicates the first records (i.e.) ) No. Minority class sample data for each field, Indicates the first records (i.e.) ) No. Minority class sample data for each field;

[0150] Step 6.3, for each sample dataset The nearest neighbor is selected based on Euclidean distance. Each sample dataset is considered as the set of its nearest neighbor samples. ;

[0151] Sort the data in ascending order (from smallest to largest) according to the calculated Euclidean distance, and select the first... Using each sample dataset as its nearest neighbor, we obtain the set of nearest neighbor samples. ;

[0152] Step 6.4, from each sample dataset The set of nearest neighbor samples A sample dataset is randomly selected from all datasets. Based on randomly selected sample datasets Generate new sample dataset All new sample datasets All included The minority sample data set after oversampling is finally obtained , comprises and all ;

[0153] The formula for generating the new sample data set is as follows:

[0154]

[0155] wherein, is the new data set generated by interpolation, is a random number in the interval (0, 1);

[0156] Step 6.5, calculate the Euclidean distance of each (majority class) sample data set to the minority sample data set The calculation formula is as follows:

[0157]

[0158] wherein, is the calculated Euclidean distance, n is the number of data sets of different feature dimensions in the bridge inspection information data set, denotes the majority class sample data in the th record (i.e. ) and the th field;

[0159] Step 6.6, for each sample data set , the nearest neighbor Euclidean distances are selected according to the Euclidean distance, and the average value of the nearest neighbor Euclidean distances is calculated , wherein denotes the average value of the nearest neighbor Euclidean distances of the sample data set ;

[0160] Step 6.7, sort all sample data sets in ascending order according to , and retain the first sample data sets , and the first sample data sets constitute the majority class sample data set after undersampling ;

[0161] Step 6.8, based on the minority sample data set after oversampling And the under-sampling majority class sample data set The merging is performed to obtain a normalized balanced bridge inspection information data set .

[0162] Table 3 illustrates the amount of bridge technical condition rating data before and after data balancing.

[0163] Table 3

[0164]

[0165] The bridge technical condition rating index in Table 3 is 0-9 categories, 10 levels, which is recorded in the NBI database of the United States National Bridge. Category 9 represents perfect condition, category 8 represents very good condition without disease, category 7 represents good condition with minor problems, category 6 represents satisfactory condition with slight structural loss, category 5 represents good condition with slight disease, category 4 represents poor condition, category 3 represents serious condition with serious disease affecting the structure, category 2 represents critical condition (the bridge needs to be closed), category 1 represents a state of imminent serious damage, and category 0 represents a state of failure. It can be understood that it is not limited to this bridge technical condition rating index, for example, a bridge technical condition rating level of 1 to 5 levels can be used, or other ways, the 1 to 5 levels specifically refer to: category 1 represents brand-new state, fully functional, category 2 represents slight defect, no impact on bridge function, category 3 represents moderate defect, still maintains normal use function, category 4 represents major component with large defect, seriously affects bridge use function; or affects carrying capacity, cannot guarantee normal use, category 5 represents major component with serious defect, cannot be used normally, endangers bridge safety, and the bridge is in a dangerous state.

[0166] Step 7, first, the parameter optimization of the LightGBM algorithm is performed based on the particle swarm optimization algorithm to form a PSO-LightGBM algorithm, which is a hybrid optimization method for automatically searching the optimal parameter combination of LightGBM through the particle swarm optimization (PSO) algorithm. The PSO-LightGBM algorithm and the normalized balanced bridge inspection information data set in step 6 are used to iteratively train and construct a bridge technical condition rating prediction model based on the PSO-LightGBM algorithm. That is, this step establishes a bridge technical condition rating prediction model through the PSO-LightGBM algorithm, and uses the normalized balanced bridge inspection information data set to train the model.

[0167] The specific steps of step 7 are as follows:

[0168] Step 7.1: Based on the Particle Swarm Optimization (PSO) algorithm, optimize the parameters of the LightGBM (Light Gradient Boosting Machine, or LGBM) algorithm to form the PSO-LightGBM algorithm, which is the parameter-optimized LightGBM algorithm.

[0169] The objective function of the LightGBM algorithm is:

[0170]

[0171] in, Represents the objective function value. This indicates the number of sample data points for the LightGBM algorithm. Indicates the first The actual value of each sample data point Indicates the first Predicted values ​​for each sample data point Represents the loss function. Indicates the total number of decision trees. Indicates the number of the decision tree. Indicates the first Decision tree function, This represents the penalty term of the model. This represents the threshold for leaf node splitting. Represents the L2 regularization coefficient. Indicates the number of leaf nodes. These are the weight vectors for the leaf nodes. The model's fitting ability and generalization ability are balanced through "loss minimization + regularization".

[0172] The parameters optimized by the particle swarm optimization algorithm include the number of leaf nodes, learning rate, subsample ratio, number of decision trees, and feature sampling ratio. The formula for the particle swarm optimization algorithm is:

[0173]

[0174] in, For the first The particle in the first The iteration of the ... Dimensional speed, No. The particle in the first The iteration of the ... Dimensional position, For inertial weights, and All are learning factor constants. and A random number between (0, 1) This is the optimal position in the particle's history. The optimal position for the entire group;

[0175] Step 7.2: Combining the normalized and balanced bridge inspection information dataset from Step 5, a bridge technical condition rating prediction model based on the PSO-LightGBM algorithm is constructed through iterative training.

[0176] The formula for the bridge technical condition rating prediction model is as follows:

[0177]

[0178] in, This is a predicted value for the bridge's technical condition rating. This is the set of feature vectors in the normalized and balanced bridge inspection information dataset. For the first Decision tree function, This represents the decision tree data space.

[0179] The formula for calculating the evaluation metric during iterative training is:

[0180]

[0181] in, For model accuracy, For model accuracy, For model recall, For the model's F1 score, This represents the number of samples that were actually classified as positive and were correctly predicted as positive. This represents the number of samples that were actually classified as negative and were correctly predicted as negative. This represents the number of samples that were actually classified as negative but were incorrectly predicted as positive. The number of samples that were actually positive but were incorrectly predicted as negative.

[0182] The confusion matrix predicted by the bridge technical condition rating prediction model is as follows: Figure 5 As shown, the horizontal axis represents the predicted values, and the vertical axis represents the accurate values, i.e., the true values. The final calculated accuracy rate is 98.24%, the precision rate is 98.24%, the recall rate is 98.24%, and the F1 score is 98.24%, indicating that the model training effect is good.

[0183] Step 8: Utilize the bridge technical condition rating prediction model obtained in Step 7 to predict the bridge's technical condition rating level. The model's input includes basic bridge information, bridge design information, bridge deck layout information, traffic statistics, and the classification information of the bridge maintenance information records. The model's output is the bridge technical condition rating level, specifically, the bridge technical condition rating level and its probability. The classification information of the bridge maintenance information records is one of the following: reconstruction, no maintenance, or several maintenance cost clustering categories. The maintenance cost clustering categories are... The category information for the bridge maintenance information records of the bridge to be predicted is classified as reconstruction, reconstruction, etc. One of the classes.

[0184] Specifically, the basic information of the bridge to be predicted, the bridge design information, the bridge deck layout information, the traffic statistics information, the maintenance cost information, and the reconstruction records are obtained, and the bridge technical condition rating prediction model obtained in step 7 is used to predict the level of the technical condition rating of the bridge to be predicted.

[0185] When using NBI data for rating prediction, the traffic volume at the time of the first inspection of each bridge is used as the baseline. For subsequent inspections, the traffic volume for future years is calculated based on the bridge's age at the time of inspection. The bridge inspection dataset updated with bridge age and traffic volume is input into the constructed model to test the accuracy of this method for rating prediction. The prediction results are as follows: Figure 6 As shown, the numbers 1 to 30 on the horizontal axis represent the bridges to be predicted from the initial inspection year (the first inspection year) to the 30th inspection year. The "prediction year" on the horizontal axis means the prediction of which year's data is used. Specifically, the bridge technical condition rating of bridges in Georgia is based on the NBI database, and the average accuracy rate exceeds 0.953, indicating that this disclosure has high accuracy in predicting the technical condition rating of bridges.

[0186] See Figure 7 This publicly disclosed bridge technical condition rating and prediction system integrates maintenance history information, including:

[0187] The dataset construction module is used to obtain multiple bridge maintenance information records. Based on the profile coefficient method and K-means algorithm, a bridge maintenance history information dataset is constructed, which includes the classification category information of bridge maintenance information records. The classification categories include maintenance cost clustering, reconstruction, and no maintenance.

[0188] The database construction module is used to construct a bridge scheduled inspection information database, which includes basic bridge information, bridge design information, bridge deck layout information, traffic statistics information, and the bridge maintenance history information dataset.

[0189] The data cleaning module is used to classify the data in the bridge inspection information database into numerical data and categorical data, and to perform data cleaning.

[0190] The feature data selection module is used to discretize the cleaned numerical data and transform it into categorical data; calculate Cramer's V coefficient between all categorical data and bridge technical condition rating information; and select the data in the cleaned bridge inspection information database (i.e., the bridge inspection information database finally obtained by the data cleaning module) corresponding to categorical data with Cramer's V coefficient greater than the coefficient threshold as the selected feature data.

[0191] The data dimensionality reduction and normalization module is used to transform data with categorical attributes in the selected feature data into one-hot encoding and use principal component analysis to reduce dimensionality to obtain dimensionality-reduced categorical data; and to normalize the bridge inspection information dataset, which includes dimensionality-reduced categorical data and data with numerical attributes in the selected feature data.

[0192] The data equalization module is used to equalize the normalized bridge inspection information dataset using the SMOTE oversampling method and the NearMiss undersampling method, resulting in a normalized equalized bridge inspection information dataset. ;

[0193] The model training module is used to utilize the PSO-LightGBM algorithm and the... Training a bridge technical condition rating prediction model;

[0194] The prediction module is used to predict the bridge's technical condition rating level using the bridge technical condition rating prediction model.

[0195] In this embodiment, the data cleaning module is used to fill in missing values ​​in numerical data using linear interpolation, to remove outliers in numerical data using the Z-score method, and to fill in missing values ​​in categorical data using the mode imputation method.

[0196] In this embodiment, some of the bridge maintenance information records include bridge technical condition rating information but exclude maintenance cost information and reconstruction records; some of the bridge maintenance information records include bridge technical condition rating information and maintenance cost information but exclude reconstruction records; and some of the bridge maintenance information records include bridge technical condition rating information and reconstruction records.

[0197] The dataset construction module is specifically used to determine the number of clusters for maintenance costs based on the silhouette coefficient method and the K-means algorithm. ; used to mark bridge maintenance information records with reconstruction records as reconstruction class, denoted as the first a bridge maintenance information record which does not have a reconstruction record and has a maintenance cost information equal to that of the previous bridge maintenance information record is marked as the no-maintenance class, recorded as the 0th class; a bridge maintenance information record which satisfies the following conditions simultaneously, i.e., the bridge technical condition rating is not higher than that of the previous bridge maintenance information record, the maintenance cost is equal to or higher than that of the previous bridge maintenance information record, and has no reconstruction record, is marked as the no-maintenance class, recorded as the 0th class; a bridge maintenance information record which does not have a reconstruction record and has a maintenance cost lower than that of the previous bridge maintenance information record is subjected to the determination of the maintenance cost clustering category and marked as one of the 1st to 9th classes. a bridge maintenance information record which does not have a reconstruction record and has a maintenance cost information equal to that of the previous bridge maintenance information record is marked as the no-maintenance class, recorded as the 0th class; a bridge maintenance information record which satisfies the following conditions simultaneously, i.e., the bridge technical condition rating is not higher than that of the previous bridge maintenance information record, the maintenance cost is equal to or higher than that of the previous bridge maintenance information record, and has no reconstruction record, is marked as the no-maintenance class, recorded as the 0th class; a bridge maintenance information record which does not have a reconstruction record and has a maintenance cost lower than that of the previous bridge maintenance information record is subjected to the determination of the maintenance cost clustering category and marked as one of the 1st to 9th classes.

[0198] In this embodiment, the data dimension reduction and normalization module is specifically configured to divide the normalized bridge inspection information data set into a minority class sample data set and a majority class sample data set . ​​​​​​​​​​​​​​​​​​​​​​​​​​; for combining oversampled minority class sample data sets and under-sampled majority class sample data sets to obtain a normalized balanced bridge inspection information data set .

[0199] The bridge technology condition rating prediction system fusing maintenance history information can realize bridge technology condition rating prediction in the specific implementation of any of the above embodiments, and the specific implementation steps will not be repeated.

[0200] The method according to the present disclosure can realize an electronic device, which comprises: a memory; one or more processors; one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the one or more programs comprise instructions for executing the bridge technology condition rating prediction method fusing maintenance history information according to any of the above embodiments.

[0201] The present disclosure also provides a computer-readable storage medium comprising instructions that, when executed on a computer, cause the computer to perform the steps of the bridge technology condition rating prediction method fusing maintenance history information according to any of the above embodiments.

[0202] The effect of the bridge technology condition rating prediction method and system fusing maintenance history information according to the present disclosure is:

[0203] The present disclosure generates bridge maintenance history information dataset including bridge maintenance information record and classification information of bridge maintenance information record based on profile coefficient method and K-means algorithm, and reconstructs bridge inspection information database including bridge basic information, bridge design information, bridge deck arrangement information, traffic statistics information and the bridge maintenance history information dataset, comprehensively considers the information of the bridge, especially fully considers the maintenance history information; the present disclosure discretizes the numerical data, and then calculates Cramer's V coefficient of all data, realizes the uniformity of the correlation index of all data (numerical data and classification data) and the bridge technical condition rating information, that is, realizes the uniformity of the correlation index of numerical data, ordered classification data and unordered classification data, considers the influence of unordered classification variable on feature selection, and the selected feature data is more accurate based on this, and the prediction accuracy of the model is further improved; the present disclosure trains the bridge technical condition rating prediction model through the PSO-LightGBM algorithm and the normalized balanced bridge inspection information dataset, can avoid model overfitting, improve the generalization ability, and also improve the accuracy of predicting the bridge technical condition rating. In summary, the present disclosure predicts the bridge technical condition rating, improves the accuracy of predicting the bridge technical condition rating due to the consideration of the data of maintenance history information, the optimization of data processing and the optimization of model training, that is, the present disclosure considers the relationship between maintenance history information, correctly reflects the relationship between unordered classification variable and bridge technical condition rating, so that unordered classification variable can also participate in feature selection, trains the bridge technical condition rating prediction model based on the PSO-LightGBM algorithm, and improves the accuracy of predicting the bridge technical condition rating.

[0204] Specifically, the present disclosure refines the maintenance history information, and reflects different degrees of maintenance. The unit bridge deck area maintenance cost is classified by using the K-means method, different maintenance levels are set in combination with the cases of no maintenance and reconstruction, and the maintenance levels are used as one of the characteristics of the bridge inspection database. This processing method improves the factors considered in the bridge technical condition rating prediction. At the same time, the refinement of the maintenance history information is also beneficial to designing different maintenance schemes in the bridge maintenance decision-making work.

[0205] Specifically, the present disclosure optimizes the feature selection method. The prior art mostly directly selects features through engineering experience, which has defects in rigor; or selects features based on Pearson correlation coefficient or Spearman correlation coefficient, but these methods are not suitable for analyzing unordered classification variables. The present disclosure discretizes numerical data by using the decision tree method, and then calculates Cramer's V coefficient, realizes the uniformity of the correlation index of numerical data, ordered classification data and unordered classification data, and is more scientific and reasonable. The selected feature data is more accurate based on this, and the normalized balanced bridge inspection information dataset based on the selected feature data makes the model training result more accurate.

[0206] Specifically, this disclosure improves the accuracy, training speed, and generalization ability of bridge technical condition rating prediction. This disclosure establishes a bridge technical condition rating prediction model using the PSO-LightGBM algorithm. LightGBM, based on tree-based parallel learning and feature histogram binning, offers fast training speed, low memory consumption, and is more suitable for handling large volumes of data. Furthermore, PSO optimization of model parameters further avoids overfitting and enhances generalization ability. In predicting bridge technical condition ratings, the optimization of data processing, feature extraction, and model training methods improves the accuracy of bridge technical condition rating predictions.

[0207] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0208] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems and methods according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0209] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this disclosure, and not to limit them; under the concept of this disclosure, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of this disclosure as described above, which are not provided in detail for the sake of brevity; although this disclosure has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this disclosure.

Claims

1. A method for predicting the technical condition of bridges by integrating maintenance history information, characterized in that, Includes the following steps: Step 1: Obtain multiple bridge maintenance information records. Based on the profile coefficient method and K-means algorithm, construct a bridge maintenance history information dataset that includes the classification categories of bridge maintenance information records. The classification categories include maintenance cost clustering, reconstruction, and no maintenance. Step 2: Construct a bridge scheduled inspection information database that includes a dataset of bridge maintenance history information, basic bridge information, bridge design information, bridge deck layout information, and traffic statistics information. Step 3: Classify the data in the bridge inspection information database into numerical data and categorical data, and perform data cleaning; Step 4: Discretize the cleaned numerical data and convert it into categorical data; calculate Cramer's V coefficient between all categorical data and bridge technical condition rating information; the data in the bridge inspection information database obtained in Step 3 corresponding to the categorical data with Cramer's V coefficient greater than the coefficient threshold are selected as feature data. Step 5: Convert the selected feature data with categorical data attributes into one-hot encoded data and use principal component analysis to reduce dimensionality to obtain dimensionality-reduced categorical data; normalize the bridge inspection information dataset, which includes the selected feature data with numerical data attributes and the dimensionality-reduced categorical data. Step 6: The normalized bridge inspection information dataset is subjected to equalization processing using the SMOTE oversampling method and the NearMiss undersampling method to obtain a normalized and balanced bridge inspection information dataset. ; Step 7, using the PSO-LightGBM algorithm and Training a bridge technical condition rating prediction model; Step 8: Use the bridge technical condition rating prediction model to predict the bridge technical condition rating level.

2. The bridge technical condition rating and prediction method integrating maintenance history information according to claim 1, characterized in that, Some of the bridge maintenance information records include bridge technical condition rating information but exclude maintenance cost information and reconstruction records; some of the bridge maintenance information records include bridge technical condition rating information and maintenance cost information but exclude reconstruction records; some of the bridge maintenance information records include bridge technical condition rating information and reconstruction records. Step 1 includes: The number of clusters for maintenance cost was determined based on the silhouette coefficient method and the K-means algorithm. ; Bridge maintenance information records with reconstruction records are marked as reconstruction type, denoted as No. The following categories are defined: Bridge maintenance information records without reconstruction or maintenance cost information are marked as "No Maintenance" (Category 0); Bridge maintenance information records that simultaneously meet the following criteria—bridge technical condition rating not increasing compared to the previous bridge maintenance information record, maintenance cost greater than or equal to the previous bridge maintenance information record, and no reconstruction record—are also marked as "No Maintenance" (Category 0); Bridge maintenance information records without reconstruction records and with maintenance costs less than the previous bridge maintenance information record are further categorized by maintenance cost, and these are designated as "Category 1". One of the classes.

3. The bridge technical condition rating and prediction method integrating maintenance history information according to claim 1, characterized in that, The basic information of the bridge includes at least one of the following: bridge age, longitude, latitude, structural classification type, and material classification type. The bridge design information includes at least one of the following: design load, total structural length, maximum span, number of main bridge spans, number of approach bridge spans, skew angle, minimum lateral clearance, minimum vertical clearance, and total horizontal clearance. The bridge deck layout information includes at least one of the following: bridge deck pavement structure type, bridge deck pavement material type, total bridge deck width, roadway width, number of upper lanes, and number of lower lanes. The traffic statistics information includes at least one of the following: average daily traffic volume, average daily truck traffic volume percentage, and traffic volume statistics year.

4. The bridge technical condition rating and prediction method integrating maintenance history information according to claim 1, characterized in that, In step 3, linear interpolation is used to impute missing values ​​in numerical data, Z-score is used to remove outliers in numerical data, and mode imputation is used to impute missing values ​​in categorical data.

5. The bridge technical condition rating and prediction method integrating maintenance history information according to claim 1, characterized in that, The process of discretizing the cleaned numerical data and transforming it into categorical data specifically includes: The cleaned numerical data is arranged in ascending order. dimensional numerical dataset ; based on The method of determining the optimal split point based on the Gini index to divide subsets is used. Based on the criterion that the difference between the Gini index of a subset and the Gini index of its optimal split point is less than a preset threshold, several final subsets with no union are determined. The set of all such final subsets is the... .

6. The bridge technical condition rating and prediction method integrating maintenance history information according to claim 1, characterized in that, The process of converting selected feature data with categorical attributes into one-hot encoded data and using principal component analysis to reduce dimensionality to obtain dimensionality-reduced categorical data specifically includes: The selected feature data with categorical attributes are transformed into one-hot encoded matrices. Through matrix The decentralized matrix is ​​obtained by subtracting the mean of the column containing each data point. ,matrix covariance matrix ; For covariance matrix Perform eigenvalue decomposition. ,in, This represents a matrix composed of eigenvectors. Represents a diagonal matrix composed of eigenvalues, according to Selecting features from largest to smallest The eigenvectors corresponding to the eigenvalues ​​form a matrix. ; according to Calculate matrix The matrix This is a matrix composed of dimensionality-reduced classification data.

7. The bridge technical condition rating and prediction method integrating maintenance history information according to claim 1, characterized in that, Step 6 includes: The normalized bridge inspection information dataset is divided into a minority class sample data set. and the majority class sample data set ; calculate Each sample dataset and Euclidean distance to all other sample datasets in the dataset; For each sample dataset The nearest neighbor is selected based on Euclidean distance. Each sample dataset is considered as the set of its nearest neighbor samples. ; From the set of each nearest neighbor sample A sample dataset is randomly selected from all datasets. ,according to Generate new sample dataset , This represents the new dataset generated through interpolation. It is a random number within the interval (0, 1). and all By merging, we obtain the aggregate of oversampled minority class sample data. ; calculate Each sample dataset To the minority class sample data collection The Euclidean distance; For each sample dataset The nearest neighbor is selected based on its Euclidean distance. Using Euclidean distances as the nearest distances, calculate the nearest neighbors. The average of Euclidean distances ; All sample datasets according to Sort by size in ascending order, before sorting Sample dataset Constructing the majority class sample data set after undersampling ; Combine the oversampled minority class sample data The collection of majority class sample data after undersampling The bridge inspection information dataset is obtained by merging the data. .

8. The bridge technical condition rating and prediction method integrating maintenance history information according to claim 1, characterized in that, The formula for the bridge technical condition rating prediction model is as follows: in, This is a predicted value for the bridge's technical condition rating. This is the set of feature vectors in the normalized and balanced bridge inspection information dataset. For the first Decision tree function, This represents the decision tree data space.

9. The bridge technical condition rating and prediction method integrating maintenance history information according to claim 1, characterized in that, The inputs to the bridge technical condition rating prediction model include basic bridge information, bridge design information, bridge deck layout information, and traffic statistics information of the bridge to be predicted. The inputs also include the classification information of the bridge maintenance information records of the bridge to be predicted. The output of the model is the bridge technical condition rating level.

10. A bridge technical condition rating and prediction system integrating maintenance history information, characterized in that, include: The dataset construction module is used to obtain multiple bridge maintenance information records. Based on the profile coefficient method and K-means algorithm, a bridge maintenance history information dataset is constructed, which includes the classification category information of bridge maintenance information records. The classification categories include maintenance cost clustering, reconstruction, and no maintenance. The database construction module is used to build a bridge scheduled inspection information database that includes a dataset of bridge maintenance history information, basic bridge information, bridge design information, bridge deck layout information, and traffic statistics information. The data cleaning module is used to classify the data in the bridge inspection information database into numerical data and categorical data, and to perform data cleaning. The feature data selection module is used to discretize the cleaned numerical data and transform it into categorical data; calculate Cramer's V coefficient between all categorical data and bridge technical condition rating information, and select the data in the cleaned bridge inspection information database corresponding to the categorical data with Cramer's V coefficient greater than the coefficient threshold as the selected feature data; The data dimensionality reduction and normalization module is used to transform data with categorical attributes in the selected feature data into one-hot encoding and use principal component analysis to reduce dimensionality to obtain dimensionality-reduced categorical data; and to normalize the bridge inspection information dataset, which includes data with numerical attributes in the selected feature data and the dimensionality-reduced categorical data. The data equalization module is used to equalize the normalized bridge inspection information dataset using the SMOTE oversampling method and the NearMiss undersampling method, resulting in a normalized equalized bridge inspection information dataset. ; The model training module is used to utilize the PSO-LightGBM algorithm and Training a bridge technical condition rating prediction model; The prediction module is used to predict the bridge technical condition rating level using a bridge technical condition rating prediction model.

Images