A Deep Learning-Based Method and System for Comparing Bridge Maintenance Schemes

Abstract

This invention discloses a method and system for comparing bridge maintenance schemes based on deep learning. The method includes: determining alternative maintenance schemes for the bridge to be maintained; collecting historical data of the bridge maintenance project and processing it to obtain training and testing sets for each maintenance feature; training a corresponding deep learning model for each maintenance feature to learn the relative importance of each maintenance feature under the combined influence of multiple factors; calculating the objective weights of each maintenance indicator using the trained model and the testing set; calculating the value index of each alternative maintenance scheme based on the weights, and determining the scheme with the highest value index as the optimal maintenance scheme; and generating a quantitative decision report based on the optimal maintenance scheme to support scheme approval and construction. This invention achieves objective quantification of maintenance indicator weights through a deep learning model, replacing traditional subjective or simple statistical weighting methods, and improving the objectivity, economy, and intelligence of bridge maintenance decision-making.

Description

Technical Field

[0001] This application belongs to the field of artificial intelligence-assisted engineering decision-making technology, specifically involving a method and system for comparing bridge maintenance schemes based on deep learning. Background Technology

[0002] With urban development, bridges, as major infrastructure, are under increasing traffic pressure, and bridge maintenance is becoming more frequent. Selecting the most valuable maintenance plan before maintenance has become one of the urgent tasks to be addressed.

[0003] Current methods for selecting bridge maintenance solutions often employ the "lowest bid method." However, the lowest bidder frequently results in cost overruns or substandard maintenance outcomes. The main reasons for this decision-making dilemma are: first, the selection process is price-driven, while maintenance effectiveness is influenced by numerous factors; second, while multiple maintenance indicators are listed for comparison, the importance of each indicator is often treated uniformly; third, even when weights are assigned to each indicator, these are often subjectively determined by expert experience or through simple statistical weighting; and fourth, historical data from previously maintained bridges are ignored. These factors make it difficult to obtain highly accurate and quantifiable weights for maintenance indicators, hindering the selection of high-value solutions that offer both good maintenance results and low costs. Consequently, current bridge maintenance decisions lack a scientific basis and demonstrate insufficient overall intelligence. Summary of the Invention

[0004] To address the aforementioned problems in the existing technology, this application provides a method and system for comparing and selecting bridge maintenance schemes based on deep learning. The technical problem to be solved by this application is achieved through the following technical solution: Firstly, this application provides a method for comparing and selecting bridge maintenance schemes based on deep learning, including: S100, determine the alternative maintenance plans for the bridge to be maintained; S200, for each maintenance index of each alternative maintenance scheme, calculate the weight of each maintenance index; wherein, the weight of each maintenance index is formed by collecting historical maintenance data of multiple completed bridge maintenance projects to form matrix data, obtaining training set and test set for each maintenance feature from the matrix data, using the training set for each maintenance feature to train a deep learning model for analyzing the relative importance of the maintenance feature under the combined influence of multiple factors, and then calculating the relative importance of each maintenance feature using the trained deep learning model and the test set; S300, based on the weight of each maintenance indicator, calculate the value index of each alternative maintenance plan, and determine the plan with the highest value index among the alternative maintenance plans as the best maintenance plan; S400, based on the optimal maintenance plan and maintenance index weights, generates a quantitative decision report for the bridge to be maintained.

[0005] Optionally, the process of collecting historical maintenance data from multiple completed bridge maintenance projects to form a matrix includes: S11, Collect historical maintenance data from multiple completed bridge maintenance projects and form sample data from the maintenance data; wherein, the maintenance data includes multiple maintenance indicators and corresponding maintenance indicator values; S12, summarize all sample data into a matrix form to obtain matrix data; wherein, the matrix data includes multiple maintenance features and corresponding maintenance feature values.

[0006] Optionally, obtaining the training set and test set for each maintenance feature from the matrix data includes: S21, the non-numerical maintenance feature values ​​in the matrix data are converted into numerical form through feature encoding, and the maintenance feature values ​​in all numerical form are cleaned to obtain preprocessed matrix data. The data cleaning process includes: missing value imputation, unified measurement unit and outlier handling. S22, Select the input dataset and its corresponding target value for each maintenance feature from the preprocessed matrix data; S23, standardize the input dataset and its corresponding target value for each maintenance feature to obtain the standardized input dataset and its corresponding target value; S24, the input dataset and its corresponding target value of each standardized maintenance feature are divided into a test set and a training set according to a predetermined ratio. After the division, a test set and a training set for each maintenance feature are obtained. The test set and the training set for each maintenance feature are composed of the input dataset and its corresponding target value in a corresponding ratio.

[0007] Optionally, S22 includes: S221, For each maintenance feature, select the feature values ​​of other maintenance features besides the maintenance feature from the preprocessed matrix data to form the input dataset for the maintenance feature; S222, for each maintenance feature, select the feature value of the maintenance feature from the preprocessed matrix data as the target value corresponding to the input dataset of the maintenance feature.

[0008] Optionally, training a deep learning model to analyze the relative importance of a maintenance feature under the combined influence of multiple factors, using the training set for each maintenance feature, includes: S31, For each maintenance feature, input the input dataset in the training set of that maintenance feature into the deep learning model that analyzes the importance of that maintenance feature, and obtain the training prediction value of that maintenance feature. S32, For each maintenance feature, calculate and analyze the loss function relative to the importance of the maintenance feature based on the target value in the training set of the maintenance feature and the training prediction value of the maintenance feature; S33, for each maintenance feature, the parameters of the deep learning model that analyzes the importance of the maintenance feature are optimized by using the loss function that analyzes the importance of the maintenance feature. After I rounds of iteration, the optimal parameters of the deep learning model that analyzes the importance of the maintenance feature are obtained. S34. For each maintenance feature, the deep learning model with optimal parameters that is used to analyze the relative importance of the maintenance feature under the combined influence of multiple factors will be determined as the trained deep learning model corresponding to the maintenance feature.

[0009] Optionally, the S200 includes: S201, For each maintenance feature, input the input dataset in the test set of the maintenance feature into the trained deep learning model corresponding to the maintenance feature to obtain the predicted value of the maintenance feature; S202, For each maintenance feature, calculate the gradient of the predicted value of the maintenance feature with respect to each maintenance feature in the input dataset of the test set, and calculate the average of the absolute values ​​of the gradients of each maintenance feature over the entire test set to obtain the importance value of each maintenance feature in the input dataset relative to the maintenance feature under the combined influence of other maintenance features. S203, for each maintenance feature, sum the importance values ​​of the maintenance feature relative to each other maintenance feature under the combined influence of other maintenance features to obtain the relative importance sum of the maintenance feature; S204. Calculate the weight of the maintenance index corresponding to each maintenance feature by using the relative importance of each maintenance feature.

[0010] Optional, the S300 includes: S310: Collect scores from various experts for each maintenance indicator in each alternative maintenance plan and the maintenance cost of each alternative maintenance plan. S320: For each alternative maintenance plan, the value index of the alternative maintenance plan is calculated by using the scores of each maintenance indicator in the alternative maintenance plan by each expert, the weight of each maintenance indicator, and the maintenance cost of the alternative maintenance plan. S330 identifies the maintenance option with the highest value index as the optimal maintenance option.

[0011] Optionally, the S320 includes: S321, for each alternative maintenance plan, the expert scores of each maintenance indicator of the alternative maintenance plan are weighted and summed to obtain the technical score of the alternative maintenance plan, and the technical index of the alternative maintenance plan is calculated based on the technical score. S322, For each alternative maintenance option, calculate the cost index of that alternative maintenance option based on its maintenance cost; S323, for each alternative maintenance plan, the value index of the alternative maintenance plan is obtained by calculating the ratio of the technical index to the cost index of the alternative maintenance plan.

[0012] Optional, the S400 includes: S410, Based on the alternative maintenance solutions and their technical index, cost index, and value index, and combined with the quantitative weights of the maintenance indicators, generate a quantitative decision report that includes solution comparison, weight basis, final selection reasons, and the best maintenance solution; S420, the quantitative decision report is output to the bridge maintenance department to guide the final approval and specific construction of the bridge maintenance plan.

[0013] Secondly, this application provides a bridge maintenance scheme comparison system based on deep learning, comprising: one or more processors; and a memory having instructions executable by the one or more processors, the instructions causing the bridge maintenance scheme comparison system based on deep learning to perform the method described in the first aspect.

[0014] Beneficial effects: 1. This application performs feature encoding, data cleaning, and standardization on maintenance index values ​​of completed bridge maintenance projects. It constructs an input dataset and its corresponding target value for each maintenance feature, then divides it into training and test sets, and initializes the corresponding deep learning model. By inputting the input dataset of the training set into the model, training prediction values ​​are obtained. The loss function of the model is calculated by combining the target value corresponding to the input dataset, and the model is iteratively optimized to finally obtain a well-trained deep learning model for each maintenance feature. The model can accurately fit the relative importance of each maintenance feature under the combined influence of multiple factors.

[0015] 2. For each maintenance feature, this application inputs the test set input dataset of the maintenance feature into the corresponding trained deep learning model to obtain the predicted value of the maintenance feature. By calculating the gradient of the predicted value of the maintenance feature with respect to each maintenance feature in the input dataset, and calculating the average of the absolute values ​​of the gradients of each maintenance feature over the entire test set, the importance of each maintenance feature relative to the maintenance feature under the combined influence of other maintenance features is obtained. This value is used to quantify the relative importance of each maintenance feature under the combined influence of multiple factors, laying the foundation for calculating the weights of bridge maintenance indicators.

[0016] 3. This application calculates and sums the relative importance values ​​of each maintenance feature relative to other maintenance features under the combined influence of multiple factors for each feature, thus obtaining the relative importance sum of that maintenance feature. Based on the relative importance sum of each maintenance feature, the weight of the maintenance index corresponding to each maintenance feature is further calculated, achieving objective quantification of the maintenance index weights. This overcomes the subjective arbitrariness of manually setting weights and the irrationality of simple statistical weighting, making the weight allocation more reasonable.

[0017] 4. The method described in this application is applicable to the comparative selection of bridge maintenance schemes. Each alternative maintenance scheme receives a technical score based on a weighted average of key maintenance indicators, which more accurately reflects the maintenance effectiveness achievable by the scheme and avoids abnormal score deviations. Furthermore, a value index is calculated by examining the technical and cost indices of each maintenance scheme. The maintenance scheme with the highest value index is then determined as the optimal maintenance scheme, and a decision report is generated to guide scheme approval and specific construction. This method significantly improves the objectivity, economy, and intelligence of bridge maintenance decision-making.

[0018] The present application will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating a bridge maintenance scheme comparison method based on deep learning provided in this application. Detailed Implementation

[0020] The present application will be described in further detail below with reference to specific embodiments, but the implementation of the present application is not limited thereto.

[0021] like Figure 1 As shown, this application provides a method for comparing and selecting bridge maintenance schemes based on deep learning, including: S100, determine the alternative maintenance plans for the bridge to be maintained; Before an actual bridge maintenance project can commence, there are usually several feasible maintenance plans. The best plan needs to be selected and a decision report needs to be prepared to guide the approval of the plan and the specific construction. S200, for each maintenance index of each alternative maintenance scheme, calculate the weight of each maintenance index; wherein, the weight of each maintenance index is formed by collecting historical maintenance data of multiple completed bridge maintenance projects to form matrix data, obtaining training set and test set for each maintenance feature from the matrix data, using the training set for each maintenance feature to train a deep learning model for analyzing the relative importance of the maintenance feature under the combined influence of multiple factors, and then calculating the relative importance of each maintenance feature using the trained deep learning model and the test set; S300, based on the weight of each maintenance indicator, calculate the value index of each alternative maintenance plan, and determine the plan with the highest value index among the alternative maintenance plans as the best maintenance plan; S400, based on the optimal maintenance plan and maintenance index weights, generates a quantitative decision report for the bridge to be maintained.

[0022] In one specific embodiment of this application, the step of collecting historical maintenance data from multiple completed bridge maintenance projects to form matrix data includes: S11, Collect historical maintenance data from multiple completed bridge maintenance projects and form sample data from the maintenance data; wherein, the maintenance data includes multiple maintenance indicators and corresponding maintenance indicator values; The maintenance indicators in this application can be one or more of the following nine types: building and installation ownership or leasing type, average maintenance interval time, labor type, bridge service life, maintenance structure system type, maintenance material type, maintenance durability duration, formwork assembly and disassembly type, and traffic control duration. In practical applications, maintenance indicators such as average maintenance interval time, bridge service life, and maintenance durability duration can be selected from the nine, or some indicators such as building and installation ownership or leasing type and labor type can be selected, or all nine maintenance indicators can be selected. Specifically, the sample data in this application comes from the historical maintenance data of completed bridge maintenance projects. The maintenance index value of one completed bridge maintenance project constitutes one sample data, and the maintenance index values ​​of all completed bridge maintenance projects constitute the complete sample data.

[0023] S12, All sample data are summarized into a matrix form to obtain matrix data, wherein the matrix data includes multiple maintenance features and corresponding maintenance feature values; Specifically, this application summarizes all sample data into a matrix form, where each row of the matrix represents a sample and each column represents a maintenance feature (i.e., a corresponding maintenance index). If the total number of samples is... Select all 9 maintenance features to form a The matrix is ​​composed of the eigenvalues ​​of all maintenance features for each sample, forming the row vector of the matrix, and the values ​​of each maintenance feature across all samples forming the column vector of the matrix. The column vector corresponding to each maintenance feature is: ,in, Indicates the first Such maintenance features Indicates the first Such maintenance features in the first The values ​​in each sample , .

[0024] In one specific embodiment of this application, obtaining the training set and test set for each maintenance feature from the matrix data includes: S21, the non-numerical maintenance feature values ​​in the matrix data are converted into numerical form through feature encoding, and the maintenance feature values ​​in all numerical form are cleaned to obtain preprocessed matrix data. The data cleaning process includes: missing value imputation, unified measurement unit and outlier handling. Specifically, the feature encoding method provided in this application, for any maintenance feature ending with "type", includes the following encoding steps: identifying all possible value categories of the maintenance feature; assigning a unique integer code to each value category; and replacing the original text category value of the maintenance feature with the corresponding integer code. For example, for the maintenance feature "labor type", its value categories include "own labor", "outsourced labor" and "sporadic labor", which can be encoded as 1, 2, and 3 respectively using the feature encoding method described above. Specifically, in this application, data cleaning refers to performing certain mathematical transformations on the feature values ​​of the maintenance features to make them conform to the usage requirements of deep learning model calculations. The specific steps include: missing value imputation, filling the missing values ​​of the maintenance features with the mean, median or mode; unifying the unit of measurement, uniformly converting the unit of measurement of the maintenance feature values; and outlier handling, defining the value range of each maintenance feature information, deleting outliers that exceed the range, and then performing missing value imputation.

[0025] S22, Select the input dataset and its corresponding target value for each maintenance feature from the preprocessed matrix data.

[0026] In one specific embodiment of this application, S22 includes: S221, For each maintenance feature, select the feature values ​​of other maintenance features besides the maintenance feature from the preprocessed matrix data to form the input dataset for the maintenance feature; Specifically, this application selects data from the matrix data after feature encoding and data cleaning, excluding the first... The eigenvalues ​​of maintenance features other than the first maintenance feature are composed of features for the first maintenance feature. The input dataset for the maintenance feature. For the... The input dataset for each maintenance feature is: , .

[0027] S222, for each maintenance feature, select the feature value of the maintenance feature from the preprocessed matrix data as the target value corresponding to the input dataset of the maintenance feature; Specifically, this application selects the first element from the matrix data after feature encoding and data cleaning. The eigenvalue of the maintenance feature is used as the eigenvalue of the maintenance feature. The target value corresponding to each maintenance feature in the input dataset. The target value corresponding to the input dataset for each maintenance feature is: , .

[0028] S23, standardize the input dataset and its corresponding target value for each maintenance feature to obtain the standardized input dataset and its corresponding target value; Specifically, this application addresses the first... The input datasets for each maintenance feature and their corresponding target values ​​are standardized. The standardization formula used is: , ,in, and They represent the standardized first digits. The input dataset for each maintenance feature and its corresponding target value. These represent the mean values ​​of each maintenance feature across all samples. These represent the standard deviations of each maintenance feature across all samples. .

[0029] S24, the input dataset and its corresponding target value of each standardized maintenance feature are divided into a test set and a training set according to a predetermined ratio. After the division, a test set and a training set for each maintenance feature are obtained. The test set and the training set for each maintenance feature are composed of the input dataset and its corresponding target value in a corresponding ratio. Specifically, this application starts from the standardized version of the first... From the input dataset of the maintenance feature, samples are randomly selected at a certain proportion 'a', and corresponding target values ​​are selected at the same proportion to form a sample for the maintenance feature. The test set for the maintenance feature; the remaining standardized test set for the first feature. The input dataset for the maintenance feature and its corresponding target value constitute the dataset for the first maintenance feature. The training set for the maintenance feature. For the first... The input dataset in the test set for each maintenance feature is: Regarding the first The target value corresponding to the input dataset in the test set of the maintenance features is: Regarding the first The input dataset in the training set for each maintenance feature is: Regarding the first The target value corresponding to the input dataset in the training set for each maintenance feature is: ,in, These represent the maintained features in the input dataset of the test set, respectively. The value of , Indicates the target value in the test set as the first Maintenance features The value of , These represent the maintained features in the input dataset of the training set, respectively. The value of , The target value in the training set is the th Maintenance features The value of , In practice, the ratio 'a' can be taken as 10%.

[0030] In one specific embodiment of this application, training a deep learning model for analyzing the relative importance of a maintenance feature under the combined influence of multiple factors using the training set for each maintenance feature includes: S31, For each maintenance feature, input the input dataset in the training set of that maintenance feature into the deep learning model that analyzes the importance of that maintenance feature, and obtain the training prediction value of that maintenance feature. Specifically, this application will first analyze the relative first The deep learning model is initialized to maintain the importance of each feature, and the initialized analysis is obtained relative to the first feature. A deep learning model for maintaining the importance of features; secondly, a deep learning model for the first feature will be developed. The input dataset from the training set containing the maintained features is used as the input to the initialized deep learning model to obtain the th... The training prediction values ​​of each maintenance feature. The formula for calculating the training prediction value of each maintenance feature is: ,in, Indicates the first Training prediction values ​​for each maintenance feature, This indicates that the analysis after initialization is relative to the first... A deep learning model that maintains the importance of features. express The bias parameters, express learning rate, express The weight parameters, express activation function, Indicates the first The input dataset in the training set of each maintenance feature. In practice, the ReLU function is used as the activation function.

[0031] S32, For each maintenance feature, calculate and analyze the loss function relative to the importance of the maintenance feature based on the target value in the training set of the maintenance feature and the training prediction value of the maintenance feature; Specifically, this application is based on Article [number missing]. The target value in the training set of the maintenance feature is compared with the training predicted value of that maintenance feature, and the relative values ​​are calculated and analyzed. The loss function determines the importance of each maintenance feature. Analysis relative to the first... The loss function for maintaining the importance of each feature is: ,in, Indicates the first Training prediction values ​​for each maintenance feature, Indicates the first The target value in the training set of each maintenance feature. .

[0032] S33, for each maintenance feature, the parameters of the deep learning model that analyzes the importance of the maintenance feature are optimized by using the loss function that analyzes the importance of the maintenance feature. After I rounds of iteration, the optimal parameters of the deep learning model that analyzes the importance of the maintenance feature are obtained. Specifically, this application will analyze the relative third The loss function that determines the importance of each maintenance feature returns the analysis relative to the first feature. A deep learning model that maintains the importance of features is used for parameter optimization. After one round of iterations, the optimal parameters are obtained. In specific implementation, the Adam algorithm is used to optimize the parameters, and the first round of iterations consists of 500 iterations.

[0033] S34. For each maintenance feature, the deep learning model with optimal parameters that is used to analyze the relative importance of the maintenance feature under the combined influence of multiple factors will be determined as the trained deep learning model corresponding to the maintenance feature.

[0034] In one specific embodiment of this application, S200 includes: S201, For each maintenance feature, input the input dataset in the test set of the maintenance feature into the trained deep learning model corresponding to the maintenance feature to obtain the predicted value of the maintenance feature; Specifically, this application will [details about the first part]. The input dataset in the test set of the maintenance feature is used as the first... The input to the trained deep learning model corresponding to the maintenance feature is used to obtain the first feature. The predicted value of each maintenance feature. The formula for calculating the predicted value of each maintenance feature is as follows: ,in, Indicates the first Predicted values ​​of each maintenance feature Indicates the first Each maintenance feature corresponds to a trained deep learning model. express The optimal bias parameters. express The optimal learning rate, express The optimal weight parameters, express activation function, Indicates the first The input dataset in the test set of each maintenance feature. In practice, the ReLU function is used as the activation function.

[0035] S202, For each maintenance feature, calculate the gradient of the predicted value of the maintenance feature with respect to each maintenance feature in the input dataset of the test set, and calculate the average of the absolute values ​​of the gradients of each maintenance feature over the entire test set to obtain the importance value of each maintenance feature in the input dataset relative to the maintenance feature under the combined influence of other maintenance features. Specifically, this application calculates the first... The predicted value of the first maintenance feature is calculated relative to the gradient of each maintenance feature in the input dataset of the test set. The average of the absolute values ​​of the gradients of each maintenance feature is then calculated over the entire test set to obtain the relative value of each maintenance feature in the input dataset to the first maintenance feature under the combined influence of other maintenance features. The importance value of each maintenance feature. For the first... The input dataset in the test set for each maintenance feature is: The maintained features in the input dataset are: The relative maintenance features in the input dataset, under the combined influence of other maintenance features, are related to the first maintenance feature. The formula for calculating the importance value of each maintenance feature is as follows: ,in, These represent the relative positions of each maintenance feature in the input dataset with respect to the first maintenance feature under the combined influence of other maintenance features. The importance value of each maintenance feature, Indicates the first The gradient of the predicted value of each maintenance feature with respect to each maintenance feature in the input dataset. This indicates taking the absolute value. Indicates the number of samples in the test set. .

[0036] As shown in Table 1, the relative importance values ​​among the maintenance features obtained through the trained deep learning model under the combined influence of multiple factors are recorded in this table.

[0037] Table 1

[0038] In Table 1, both row labels and column labels are: The values ​​represent the nine main maintenance features. The relative importance values ​​between maintenance features themselves are invalid and are replaced with *. For example, consider the importance value relative to the sixth maintenance feature. This represents the 6th maintenance feature. The input dataset for the test set of the 6th maintenance feature is: The maintained features in the input dataset are: The relative maintenance features of each maintenance feature in the input dataset under the combined influence of other maintenance features. The importance value can be seen from the tag. In this column, the values ​​of the valid cells are... ,in, Indicates maintenance features Other maintenance features Relative maintenance characteristics under the combined influence of The importance value, and the same applies to others. Indicates maintenance features Other maintenance features Relative maintenance characteristics under the combined influence of Importance value, Indicates maintenance features Other maintenance features Relative maintenance characteristics under the combined influence of The importance value, and the same applies to others. Indicates maintenance features Other maintenance features Relative maintenance characteristics under the combined influence of The importance value.

[0039] S203, for each maintenance feature, sum the importance values ​​of the maintenance feature relative to each other maintenance feature under the combined influence of other maintenance features to obtain the relative importance sum of the maintenance feature; Specifically, this application will [details about the first part]. The importance values ​​of the first maintenance feature relative to each other under the combined influence of other maintenance features are summed to obtain the value of the second maintenance feature. The relative importance of each maintenance feature. They represent the excluding the first Each maintenance feature other than the first maintenance feature, the first maintenance feature, The formula for calculating the relative importance of each maintenance feature is: ,in, Indicates the first The relative importance of each maintenance feature and They represent the first The importance value of each maintenance feature relative to other maintenance features under the combined influence of these features. .

[0040] As shown in Table 1, the relative importance of the first maintenance feature is taken as an example. This indicates the first maintenance feature. These represent all maintenance features except the first one; the importance of the first maintenance feature relative to the other maintenance features under their combined influence can be seen in the labels. In this row, the value of the valid cells is ,in, Indicates maintenance features Other maintenance features Relative maintenance characteristics under the combined influence of Importance value, Indicates maintenance features Other maintenance features Relative maintenance characteristics under the combined influence of The importance value, and the same applies to others. Indicates maintenance features Other maintenance features Relative maintenance characteristics under the combined influence of The importance value is calculated by summing the importance values ​​of the first maintenance feature relative to all other maintenance features under their combined influence. The formula for calculating the relative importance sum of the first maintenance feature is as follows: ,in, This indicates the relative importance of the first maintenance feature. The relative importance of the first maintenance feature is shown in the cell. middle.

[0041] S204, using the relative importance of each maintenance feature, calculate the weight of the maintenance index corresponding to each maintenance feature; Specifically, this application first sums the relative importance of each maintenance feature, then removes the sum of the relative importance of each maintenance feature to obtain the weight of the maintenance index corresponding to each maintenance feature. The formula for calculating the weight of each maintenance indicator is as follows: And satisfy ,in, Indicates the first The weight of each maintenance metric, Indicates the first The relative importance of each maintenance feature and This represents the sum of the relative importance of each maintenance feature. .

[0042] In one specific embodiment of this application, S300 includes: S310: Collect scores from various experts for each maintenance indicator in each alternative maintenance plan and the maintenance cost of each alternative maintenance plan. S320: For each alternative maintenance plan, the value index of the alternative maintenance plan is calculated by using the scores of each maintenance indicator in the alternative maintenance plan by each expert, the weight of each maintenance indicator, and the maintenance cost of the alternative maintenance plan. S330 identifies the maintenance option with the highest value index as the optimal maintenance option.

[0043] In one specific embodiment of this application, S320 includes: S321, for each alternative maintenance plan, the expert scores of each maintenance indicator of the alternative maintenance plan are weighted and summed to obtain the technical score of the alternative maintenance plan, and the technical index of the alternative maintenance plan is calculated based on the technical score. Specifically, this application first weights and sums the expert scores for the maintenance indicators of each alternative maintenance scheme to obtain the technical score of each alternative maintenance scheme; secondly, it sums the technical scores of each alternative maintenance scheme; finally, it removes the technical score of each alternative maintenance scheme to obtain the technical index of each alternative maintenance scheme. The formula for calculating the technical score of each alternative maintenance solution is as follows: ,in, Indicates the first Technical scores for each alternative maintenance solution Indicates the first The weight of each maintenance metric, Indicates the first The first of the alternative maintenance options Expert scoring for each maintenance metric. The formula for calculating the technical index of each alternative maintenance option is as follows: ,in, Indicates the first Technical indicators of the alternative maintenance solutions This means summing the technical scores of each alternative maintenance solution. .

[0044] S322, For each alternative maintenance option, calculate the cost index of that alternative maintenance option based on its maintenance cost; Specifically, in this application, the first... The formula for calculating the cost index of each alternative maintenance option is as follows: ,in, Indicates the first The cost index of each alternative maintenance option Indicates the first Maintenance costs of each alternative maintenance option This means summing the maintenance costs of all alternative maintenance options. .

[0045] S323, For each alternative maintenance plan, the value index of the alternative maintenance plan is obtained by calculating the ratio of the technical index to the cost index of the alternative maintenance plan; Specifically, the value index of each alternative maintenance solution in this application is the ratio of the technical index to the cost index of that alternative maintenance solution. The formula for calculating the value index of each alternative maintenance option is as follows: ,in, Indicates the first The value index of each alternative maintenance solution Indicates the first Technical indicators of the alternative maintenance solutions Indicates the first The cost index of each alternative maintenance option .

[0046] This application identifies the maintenance option with the highest value index as the optimal maintenance option.

[0047] In one specific embodiment of this application, S400 includes: S410, Based on the alternative maintenance solutions and their technical index, cost index, and value index, and combined with the quantitative weights of the maintenance indicators, generate a quantitative decision report that includes solution comparison, weight basis, final selection reasons, and the best maintenance solution; Specifically, the process of generating a quantitative decision report for the bridge to be maintained in this application includes: explaining the calculation basis and data source for the weights of each maintenance indicator; visually displaying and comparing the expert scores and weighted scores of each alternative maintenance scheme on each maintenance indicator, the maintenance cost and cost index of each alternative maintenance scheme, and the value index of each alternative maintenance scheme in tabular form; recommending the alternative maintenance scheme with the highest value as the best maintenance scheme; explaining, based on the maintenance indicators and quantitative weights, that the scheme achieves the best balance between maintenance effect and maintenance cost, and forming an electronic document.

[0048] S420, The quantitative decision report is output to the bridge maintenance department to guide the final approval and specific construction of the bridge maintenance plan; Specifically, this application outlines the best maintenance plan and decision-making basis to facilitate the approval of the plan by the management department. The best maintenance plan can be directly used to guide specific maintenance construction.

[0049] This application also provides a bridge maintenance scheme comparison system based on deep learning, comprising: one or more processors; and a memory having instructions executable by the one or more processors, the instructions causing the bridge maintenance scheme comparison system based on deep learning to execute a bridge maintenance scheme comparison method based on deep learning.

[0050] It is worth noting that the terms "first" and "second" in this application are used for descriptive purposes only and should not be construed as indicating or implying a relative degree of importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0051] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of this application and should not be construed as limiting the specific implementation of this application to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of this application, and all such modifications or substitutions should be considered within the scope of protection of this application.

Claims

1. A method for comparing and selecting bridge maintenance schemes based on deep learning, characterized in that, include: S100, determine the alternative maintenance plans for the bridge to be maintained; S200, for each maintenance index of each alternative maintenance scheme, calculate the weight of each maintenance index; wherein, the weight of each maintenance index is formed by collecting historical maintenance data of multiple completed bridge maintenance projects to form matrix data, obtaining training set and test set for each maintenance feature from the matrix data, using the training set for each maintenance feature to train a deep learning model for analyzing the relative importance of the maintenance feature under the combined influence of multiple factors, and then calculating the relative importance of each maintenance feature using the trained deep learning model and the test set; S300, based on the weight of each maintenance indicator, calculate the value index of each alternative maintenance plan, and determine the plan with the highest value index among the alternative maintenance plans as the best maintenance plan; S400, based on the optimal maintenance plan and maintenance index weights, generates a quantitative decision report for the bridge to be maintained.

2. The bridge maintenance scheme comparison method based on deep learning according to claim 1, characterized in that, The matrix data formed by collecting historical maintenance data from multiple completed bridge maintenance projects includes: S11, Collect historical maintenance data from multiple completed bridge maintenance projects and form sample data from the maintenance data; wherein, the maintenance data includes multiple maintenance indicators and corresponding maintenance indicator values; S12, summarize all sample data into a matrix form to obtain matrix data; wherein, the matrix data includes multiple maintenance features and corresponding maintenance feature values.

3. The bridge maintenance scheme comparison method based on deep learning according to claim 1, characterized in that, The process of obtaining the training set and test set for each maintenance feature from the matrix data includes: S21, the non-numerical maintenance feature values ​​in the matrix data are converted into numerical form through feature encoding, and the maintenance feature values ​​in all numerical form are cleaned to obtain preprocessed matrix data. The data cleaning process includes: missing value imputation, unified measurement unit and outlier handling. S22, Select the input dataset and its corresponding target value for each maintenance feature from the preprocessed matrix data; S23, standardize the input dataset and its corresponding target value for each maintenance feature to obtain the standardized input dataset and its corresponding target value; S24, the input dataset and its corresponding target value of each standardized maintenance feature are divided into a test set and a training set according to a predetermined ratio. After the division, a test set and a training set for each maintenance feature are obtained. The test set and the training set for each maintenance feature are composed of the input dataset and its corresponding target value in a corresponding ratio.

4. The bridge maintenance scheme comparison method based on deep learning according to claim 3, characterized in that, S22 includes: S221, For each maintenance feature, select the feature values ​​of other maintenance features besides the maintenance feature from the preprocessed matrix data to form the input dataset for the maintenance feature; S222, for each maintenance feature, select the feature value of the maintenance feature from the preprocessed matrix data as the target value corresponding to the input dataset of the maintenance feature.

5. The method for comparing and selecting bridge maintenance schemes based on deep learning according to claim 1, characterized in that, The step of training a deep learning model using the training set for each maintenance feature to analyze the relative importance of that maintenance feature under the combined influence of multiple factors includes: S31, For each maintenance feature, input the input dataset in the training set of that maintenance feature into the deep learning model that analyzes the importance of that maintenance feature, and obtain the training prediction value of that maintenance feature. S32, For each maintenance feature, calculate and analyze the loss function relative to the importance of the maintenance feature based on the target value in the training set of the maintenance feature and the training prediction value of the maintenance feature; S33, for each maintenance feature, the parameters of the deep learning model that analyzes the importance of the maintenance feature are optimized by using the loss function that analyzes the importance of the maintenance feature. After I rounds of iteration, the optimal parameters of the deep learning model that analyzes the importance of the maintenance feature are obtained. S34. For each maintenance feature, the deep learning model with optimal parameters that is used to analyze the relative importance of the maintenance feature under the combined influence of multiple factors will be determined as the trained deep learning model corresponding to the maintenance feature.

6. The bridge maintenance scheme comparison method based on deep learning according to claim 5, characterized in that, S200 includes: S201, For each maintenance feature, input the input dataset in the test set of the maintenance feature into the trained deep learning model corresponding to the maintenance feature to obtain the predicted value of the maintenance feature; S202, For each maintenance feature, calculate the gradient of the predicted value of the maintenance feature with respect to each maintenance feature in the input dataset of the test set, and calculate the average of the absolute values ​​of the gradients of each maintenance feature over the entire test set to obtain the importance value of each maintenance feature in the input dataset relative to the maintenance feature under the combined influence of other maintenance features. S203, for each maintenance feature, sum the importance values ​​of the maintenance feature relative to each other maintenance feature under the combined influence of other maintenance features to obtain the relative importance sum of the maintenance feature; S204. Calculate the weight of the maintenance index corresponding to each maintenance feature by using the relative importance of each maintenance feature.

7. The method for comparing and selecting bridge maintenance schemes based on deep learning according to claim 5, characterized in that, The S300 includes: S310: Collect scores from various experts for each maintenance indicator in each alternative maintenance plan and the maintenance cost of each alternative maintenance plan. S320: For each alternative maintenance plan, the value index of the alternative maintenance plan is calculated by using the scores of each maintenance indicator in the alternative maintenance plan by each expert, the weight of each maintenance indicator, and the maintenance cost of the alternative maintenance plan. S330 identifies the maintenance option with the highest value index as the optimal maintenance option.

8. The bridge maintenance scheme comparison method based on deep learning according to claim 7, characterized in that, S320 includes: S321, for each alternative maintenance plan, the expert scores of each maintenance indicator of the alternative maintenance plan are weighted and summed to obtain the technical score of the alternative maintenance plan, and the technical index of the alternative maintenance plan is calculated based on the technical score. S322, For each alternative maintenance option, calculate the cost index of that alternative maintenance option based on its maintenance cost; S323, for each alternative maintenance plan, the value index of the alternative maintenance plan is obtained by calculating the ratio of the technical index to the cost index of the alternative maintenance plan.

9. The method for comparing and selecting bridge maintenance schemes based on deep learning according to claim 1, characterized in that, The S400 includes: S410, Based on the alternative maintenance solutions and their technical index, cost index, and value index, and combined with the quantitative weights of the maintenance indicators, generate a quantitative decision report that includes solution comparison, weight basis, final selection reasons, and the best maintenance solution. S420, the quantitative decision report is output to the bridge maintenance department to guide the final approval and specific construction of the bridge maintenance plan.

10. A bridge maintenance scheme comparison system based on deep learning, characterized in that, include: One or more processors; The system also includes a memory having instructions executable by the one or more processors, which cause the deep learning-based bridge maintenance scheme comparison system to perform the method according to any one of claims 1 to 9.

Images