A deep learning highway maintenance cost index dynamic measurement and intelligent auditing method and system

Abstract

The application provides a kind of deep learning's highway maintenance cost index dynamic measurement and intelligent auditing method and system, related to highway maintenance cost management technical field, the method includes: obtaining highway maintenance cost data, according to highway maintenance cost data, constructs multidimensional data cube;Extract the cost sample features of multidimensional data cube and carry out dimension reduction processing to the extracted features, obtain dimension reduction feature data;According to dimension reduction feature data, clustering analysis is carried out, and the clustering result is obtained;According to the clustering result, the abnormal sample is eliminated, and the target sample data is obtained.The application realizes the dynamic measurement and intelligent auditing of highway maintenance cost index, improves the accuracy, reliability and auditing efficiency of index interval.

Description

Technical Field

[0001] This invention relates to the field of highway maintenance cost management technology, and in particular to a method and system for dynamic calculation and intelligent review of highway maintenance cost indicators using deep learning. Background Technology

[0002] Highway maintenance projects involve different maintenance objects such as roadbed, pavement, bridges and culverts, and ancillary facilities. Project costs are affected by factors such as maintenance type, region, highway technical grade, construction time, and project scale. Current highway maintenance cost management typically collects historical project bill of quantities, material unit prices, and project cost data, categorizing them into limited project types and forming cost indicators through methods such as averages and fixed fluctuation ranges. During project review, reviewers look for similar historical indicators based on project attributes and combine their personal experience to judge whether the bill of quantities items and material unit prices are reasonable.

[0003] Because cost samples corresponding to different project attributes differ in terms of cost composition, data distribution, and fluctuation range, mixing different types of cost samples for calculation can easily cause the resulting index range to deviate from the actual cost level of the corresponding project. At the same time, there are abnormal samples in historical cost data, and some cost samples do not meet the normal distribution. It is difficult to form an index range that matches the sample distribution using a uniform statistical method. As new project data is continuously added to the database and material unit prices change, cost indicators calculated according to a fixed period may also be updated with lag, resulting in the index range retrieved by the auditors not matching the project to be audited. Summary of the Invention

[0004] This invention provides a method and system for dynamic calculation and intelligent review of highway maintenance cost indicators based on deep learning. It realizes dynamic calculation and intelligent review of highway maintenance cost indicators, and improves the accuracy, reliability and review efficiency of indicator ranges.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: The first aspect is a deep learning-based method for dynamic calculation and intelligent verification of highway maintenance cost indicators, the method comprising: Obtain highway maintenance cost data and construct a multidimensional data cube based on the highway maintenance cost data; Extract cost sample features from the multidimensional data cube and perform dimensionality reduction on the extracted features to obtain dimensionality-reduced feature data; perform cluster analysis based on the dimensionality-reduced feature data to obtain clustering results; remove outlier samples based on the clustering results to obtain target sample data; Perform a normality test on the target sample data, and determine the reasonable range of cost indicators based on the test results; A least squares support vector machine regression model is trained based on the target sample data and its corresponding project attributes to obtain a cost prediction model; the cost prediction result is obtained based on the cost prediction model. Acquire new highway maintenance cost data, incorporate the new highway maintenance cost data into a multidimensional data cube, re-perform cluster analysis and determine indicator intervals to obtain updated indicators, and store the updated indicators in the indicator database. Obtain the cost documents to be reviewed and extract the data to be reviewed from them. Based on the project attributes of the data to be reviewed, retrieve the reasonable range of the corresponding cost indicators from the indicator database as the matching indicator range. Calculate the deviation based on the matching indicator range and the data to be reviewed. If the deviation exceeds the preset deviation threshold, generate an early warning and difference analysis report to obtain review auxiliary decision-making information.

[0006] Secondly, a deep learning-based system for dynamic calculation and intelligent verification of highway maintenance cost indicators includes: The acquisition module is used to acquire highway maintenance cost data and construct a multi-dimensional data cube based on the highway maintenance cost data. The extraction module is used to extract cost sample features from the multidimensional data cube and perform dimensionality reduction processing on the extracted features to obtain dimensionality-reduced feature data; perform cluster analysis based on the dimensionality-reduced feature data to obtain clustering results; and remove outlier samples based on the clustering results to obtain target sample data. The testing module is used to perform normality tests on the target sample data and determine the reasonable range of cost indicators based on the test results. The prediction module is used to train a least squares support vector machine regression model based on the target sample data and its corresponding project attributes to obtain a cost prediction model; and to obtain the cost prediction result based on the cost prediction model. The analysis module is used to acquire newly added highway maintenance cost data, classify the newly added highway maintenance cost data into a multidimensional data cube, re-perform cluster analysis and determine indicator intervals to obtain updated indicators, and store the updated indicators in the indicator database. The matching module is used to obtain the cost documents to be reviewed and extract the data to be reviewed from them. Based on the project attributes of the data to be reviewed, it retrieves the reasonable range of the corresponding cost indicators from the indicator database as the matching indicator range. It calculates the deviation based on the matching indicator range and the data to be reviewed. If the deviation exceeds the preset deviation threshold, it generates an early warning and difference analysis report to obtain review auxiliary decision-making information.

[0007] The above-described solution of the present invention has at least the following beneficial effects: The cost management platform classifies highway maintenance cost data according to maintenance type, region, highway technical grade, construction time and project scale, constructs a multi-dimensional data cube, and performs dimensionality reduction and cluster analysis on cost sample characteristics. This allows samples with similar project attributes and cost characteristics to be grouped into the corresponding cluster results, reducing the impact of mixing different types of cost samples on indicator calculation.

[0008] The cost management platform removes outlier samples based on clustering results and performs a normality test on the target sample data. Based on the test results, it determines a reasonable range for cost indicators, ensuring that the indicator range matches the actual distribution characteristics of the target sample data and reducing the impact of outlier samples and distribution differences on the stability of the indicator range. The cost management platform trains a least squares support vector machine regression model based on the target sample data and its corresponding project attributes. This enables the cost prediction model to establish a data relationship between project attributes and cost results, and to obtain cost prediction results based on this data relationship. This provides a basis for predicting highway maintenance costs corresponding to different project attributes.

[0009] The cost management platform incorporates newly added highway maintenance cost data into a multi-dimensional data cube, performs re-cluster analysis and determines indicator ranges, and stores the updated indicators in the indicator database. This allows cost indicators to be updated along with newly added highway maintenance cost data, reducing the deviation in indicator ranges caused by the lag in updating fixed indicators.

[0010] The cost management platform retrieves matching indicator ranges from the indicator database based on the project attributes of the data to be audited, and calculates the deviation based on the matching indicator range and the data to be audited. When the deviation exceeds the preset deviation threshold, it generates an early warning and a difference analysis report, enabling abnormal cost data in the cost documents to be audited to be located and form corresponding audit basis, thereby improving the consistency of highway maintenance cost audit standards and the traceability of audit auxiliary decision-making information. Attached Figure Description

[0011] Figure 1 This is a flowchart illustrating a method for dynamic calculation and intelligent review of highway maintenance cost indicators based on deep learning, provided by an embodiment of the present invention.

[0012] Figure 2 This is a schematic diagram of a deep learning-based dynamic calculation and intelligent review system for highway maintenance cost indicators, provided by an embodiment of the present invention.

[0013] Figure 3 This is a flowchart illustrating the process of calculating the deviation based on the matching index range and the data to be reviewed, and generating an early warning and difference analysis report if the deviation exceeds a preset deviation threshold, provided by an embodiment of the present invention. Detailed Implementation

[0014] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0015] The embodiments of the present invention propose a method and system for dynamic calculation and intelligent review of highway maintenance cost indicators using deep learning.

[0016] The cost management platform in this embodiment refers to a computing platform used to receive, store, and process highway maintenance cost data, and to perform indicator calculation, cost prediction, indicator updating, and cost review. The cost management platform can be deployed on a server, server cluster, or cloud computing environment, and includes at least a processor, memory, data interface, and indicator database. The processor reads program instructions from the memory and executes the method of this embodiment. The data interface receives highway maintenance cost data, update trigger information, and cost documents to be reviewed. The indicator database stores reasonable ranges for cost indicators, updated indicators, and indicator version management data associated with the updated indicators.

[0017] Highway maintenance cost data is recorded using projects as the basic unit. Each project record includes project attributes, list items, and material unit prices. Project attributes include maintenance type, region, highway technical grade, construction time, and project scale. List items include list code, list name, unit of measurement, quantity, comprehensive unit price, and total list price. Material unit prices include material code, material name, specifications, unit of measurement, price corresponding time, and unit price. Unit cost represents the ratio of project cost to the quantity of the corresponding measurement object. The same measurement object and unit of measurement are used within the same classification unit.

[0018] Figure 1 and Figure 3 This illustration shows a flowchart of a method for dynamic calculation and intelligent review of highway maintenance cost indicators based on multi-dimensional clustering analysis and deep learning, as provided in an embodiment of this application. (See also...) Figure 1 and Figure 3 The method is executed by the cost management platform and includes steps S101 to S109.

[0019] S101, Obtain highway maintenance cost data and construct a multi-dimensional data cube, including: The cost management platform receives historical highway maintenance project data that has been reviewed and approved through a data interface, maps fields such as project identifier, road section identifier, project attributes, list items and material unit prices to form project records.

[0020] During the data verification phase, the platform first checks whether the project identifier already exists in the system. If duplicate project identifiers exist, the platform compares the review status field and data version number field of each record: the review status field records the review progress of the record, with values ​​including pending review, under review, reviewed, rejected, etc.; the data version number field records the number of times the record has been updated. The platform prioritizes retaining records with the review status of "reviewed". If multiple records are in the "reviewed" status, the platform further compares the data version numbers and retains the record with the largest data version number. The duplicate records to be discarded are moved to the historical archive. The original data is retained in the archive and an archive time and archive reason mark are attached. The archived records no longer participate in any indicator calculation process, but can be used for subsequent data auditing and traceability queries.

[0021] For project attribute fields, the platform checks each of the five dimensions—maintenance type, region, highway technical grade, construction time, and project scale—to ensure that all are not empty. If any dimension is missing, the attribute is considered incomplete. The platform writes the record into the dataset to be corrected and generates a correction prompt in the management interface. The prompt includes the record identifier, the name of the missing field, and the original position of the missing field in the data source. The record in the dataset to be corrected is frozen and does not participate in subsequent feature extraction, cluster analysis, and indicator range determination. After the data management personnel complete the missing fields, the record is removed from the dataset to be corrected and re-enters the verification process.

[0022] For the verification of the list items, the platform adopts a cross-validation method: it checks whether the unit of measurement of each item in the list is consistent with the measurement caliber of the platform's preset classification unit, and at the same time verifies whether the calculation result of the quantity multiplied by the comprehensive unit price matches the total price of the list item. The platform allows rounding errors due to rounding, but sets the error tolerance to one-thousandth, that is, the absolute value of the difference between the total price and the quantity multiplied by the comprehensive unit price divided by the total price does not exceed one-thousandth; for records with an error tolerance exceeding one-thousandth, the platform generates a manual confirmation prompt, and the auditor verifies the data source and confirms whether to accept it.

[0023] For verifying material unit prices, the platform checks whether the material code exists in the material standard code library, which pre-stores the standard material code system issued by the transportation and housing construction authorities. It also checks whether the unit of measurement for the material unit price is convertible with the unit of measurement of the materials referenced in the list item. When the units of measurement are different but have a definite conversion relationship (e.g., a 1000x conversion between tons and kilograms, or a 1000x conversion between cubic meters and liters), the platform automatically converts the material unit price and material consumption to the unified unit of measurement used in the classification unit according to the conversion factor. The converted values ​​are synchronously updated in the project record, and the values ​​before and after the conversion and the conversion factor are recorded in the change log of the project record, retaining a complete backup of the original values ​​for traceability.

[0024] Maintenance types can be categorized according to maintenance engineering costs, routine maintenance costs, and maintenance inspection costs. Maintenance engineering costs can be further divided into roadbed engineering, pavement engineering, bridge and culvert engineering, tunnel engineering, traffic safety facilities engineering, landscaping and environmental protection engineering, and management and maintenance facilities engineering. The geographical area can be divided into several zones based on the target area's geographical conditions and construction environment. Each zone covers a geographical area with similar geographical conditions, economic development levels, and construction environments. The division comprehensively considers factors such as topographic features, climate conditions, transportation conditions, material supply convenience, and local economic development level. Highway technical grades can include expressways and Class I highways. For roads, secondary highways, and tertiary and lower-level highways, construction time can be categorized into annual, quarterly, and monthly timeframes. The platform allows users to switch between these three timeframes based on their actual analysis needs. Higher timeframes result in higher sample aggregation and larger sample sizes, but lower time resolution. Project scale can be divided into multiple continuous and non-overlapping scale ranges based on project cost. As one implementation method, project scale can be categorized into small projects, medium projects, large projects, and extra-large projects, with corresponding project cost ranges of less than 1 million yuan, more than 1 million yuan but less than 5 million yuan, more than 5 million yuan but less than 20 million yuan, and more than 20 million yuan, respectively.

[0025] The cost management platform sequentially reads the maintenance type, region, highway technical grade, construction time, and project scale from each project record, and generates a category key by concatenating these five attribute values ​​in a fixed order. The generation rule for the category key is as follows: the five attribute values ​​are concatenated sequentially, separated by double underscores to form a unique identifier string. The construction time dimension supports hierarchical aggregation: when generating the category key, the platform extracts the corresponding portion of the construction time according to the user-specified time granularity. When aggregating by year, only the year information is retained, and projects within the same year are grouped into the same category unit; when aggregating by quarter, the year is extracted along with the quarter number; and when aggregating by month, both year and month information are extracted.

[0026] The platform writes project records with the same category key into the same category unit and constructs a multidimensional data cube according to the dimensional relationships between the category units. The multidimensional data cube is constructed with five dimensions as axes: maintenance type, region, highway technical grade, construction time, and project scale. The set of values ​​on each dimension axis is determined by the attribute values ​​appearing in the actual data. Each cell in the data cube corresponds to a category unit. Each category unit stores the following information: a list of project record indexes, a sample count, a description of the measurement caliber, a data time range, and the current data version number. When the sample count of a category unit is less than a preset lower limit, the platform marks the category unit as having insufficient samples. This category unit does not perform indicator calculations independently but is merged with adjacent category units or directly references the indicators of the higher-level aggregated category unit. Subsequent feature extraction, cluster analysis, and indicator interval determination are all performed within category units with consistent measurement calibers and sufficient samples.

[0027] S102, extracting cost sample features and performing dimensionality reduction processing, including: The cost management platform extracts cost sample features from the project records of each category. These cost sample features include four main categories: unit cost features, cost composition features, list item features, and material consumption features. The unit cost characteristic is obtained by dividing the total project cost by the corresponding quantity of measurement objects. The selection of measurement objects depends on the maintenance type: for pavement engineering, the pavement area is used as the measurement object, and a comprehensive unit price per square meter is extracted; for subgrade engineering, the route length is used as the measurement object, and a comprehensive cost per kilometer is extracted; for bridge and culvert engineering, the bridge deck area or culvert length is used as the measurement object; for tunnel engineering, the tunnel length is used as the measurement object; for traffic safety facility engineering, the route length is used as the measurement object; for greening and environmental protection engineering, the greening area is used as the measurement object. For maintenance types not listed above, the platform uses the method of dividing the total project cost by the main engineering quantity index corresponding to the maintenance type as the unit cost.

[0028] The cost composition characteristics include the ratio of labor costs to the total project cost, the ratio of material costs to the total project cost, the ratio of machinery costs to the total project cost, and the ratio of management fees to the total project cost. These four costs are directly extracted from the project cost summary table, and each cost is divided by the total project cost to obtain the ratio. The platform verifies the sum of the four ratios: if the sum of the four ratios deviates from 1 by more than 0.5%, the record is marked, indicating that there may be a data error or cost omission. This mark is carried along with the record to subsequent processes for analysis and reference.

[0029] The features of the bill of quantities items include quantity, unit price, and total price. During extraction, for bill of quantities items with the same code within the same category unit, the platform extracts their quantity, unit price, and total price as feature values. If the bill of quantities items contained in different items within the same category unit are not completely consistent, the platform uses the set of bill of quantities items with the highest frequency in that category unit as the baseline bill of quantities template. The method for determining the baseline bill of quantities template is as follows: count the frequency of each bill of quantities code in all items within that category unit, select bill of quantities items whose frequency exceeds a preset threshold (e.g., exceeding 50%) and include them in the template. For template bill of quantities items that are missing in some items, their feature values ​​are padded with zeros; for bill of quantities items that exist in some items but are not included in the template, they are not included in the feature vector.

[0030] Material consumption characteristics are obtained by dividing the consumption of major materials by the quantity of the corresponding measurement objects. The platform selects the top few major materials with the highest usage percentage within this category unit, defaulting to the top five, and arranges them in descending order of usage. For each material, its consumption is divided by the quantity of the corresponding measurement objects to obtain the unit consumption. For example, the unit consumption of asphalt concrete in pavement engineering is expressed in tons per square meter, and the unit consumption of fill materials in subgrade engineering is expressed in cubic meters per kilometer.

[0031] The extracted cost sample features are standardized. For each original feature dimension, the platform first calculates the arithmetic mean and standard deviation of all samples within the current classification unit. The standard deviation is calculated by dividing by... The method (i.e., Bezier correction), in which Given the number of samples, subtract the mean from the original feature value of each sample, and divide the difference by the standard deviation of that sample to obtain the standardized feature value. After standardization, the mean of each feature dimension is zero and the standard deviation of the sample is 1. If all samples in the current classification unit have the same value for a certain feature, that is, the standard deviation of the sample is zero, it means that the feature is a constant and does not have the ability to distinguish and identify samples. The platform removes the feature from the feature set participating in dimensionality reduction and clustering, but retains its original value and records it in the feature metadata, recording the name of the feature, the constant original value, and the reason for removal, for reference when interpreting the results and explaining the meaning of the indicators.

[0032] The cost management platform performs principal component analysis on the standardized feature data to achieve dimensionality reduction. The platform first organizes the standardized feature data into a matrix, where rows correspond to samples and columns correspond to feature dimensions. Let the number of samples be... The feature dimension is The data matrix is ​​then OK The platform calculates the covariance matrix of this matrix, which is: OK A square matrix of columns, where the first column is... Line number The element of the column is the first The first feature and the second The covariance among the features; since the data has been standardized, the covariance matrix is ​​equivalent to the correlation coefficient matrix. Eigenvalue decomposition is performed on the covariance matrix to obtain... eigenvalues and corresponding Unit eigenvectors Each eigenvalue corresponds to a principal component. The magnitude of the eigenvalue represents the amount of variance explained by that principal component; the larger the eigenvalue, the more data variation information the principal component contains. The principal components are arranged in descending order of eigenvalue, and the variance contribution rate of each principal component is calculated, which is the eigenvalue divided by the total variance. The sum of all eigenvalues; starting from the first principal component, the variance contribution rate is accumulated sequentially to calculate the cumulative variance contribution rate; the platform retains the top components with a cumulative variance contribution rate of 95%. The principal component, that is, when accumulated to the th principal component. If the cumulative variance contribution rate of the principal components first reaches or exceeds 95%, then the previous principal components are retained. Principal components, The value of is less than or equal to For each cost sample, the platform will generate a standardized feature vector of that sample. (dimensional) and the unit eigenvectors of each retained principal component (each is ) Perform a dot product operation on the (dimensional) to obtain the score of the sample on each retained principal component. The scores constitute the dimensionality-reduced feature vector of the sample. The dimensionality reduction (D) is achieved by combining the reduced feature vectors of all samples into dimensionality-reduced feature data. The data's dimensionality is reduced from the original number of features. Down to Dimension; the platform records in the dimensionality reduction metadata The values ​​of , the actual value of the cumulative variance contribution rate, the variance contribution rate of each retained principal component, and the loading matrix between each retained principal component and the original features are used to interpret the business meaning of each principal component after dimensionality reduction.

[0033] S103, Determine the target number of clusters based on the dimensionality reduction feature data, including: The cost management platform is based on the number of valid samples. Set multiple candidate clustering numbers When the number of valid samples When the number is less than 4, the platform does not perform cluster analysis, but directly treats all samples within the classification unit as a whole, grouping all samples into one cluster group. In this case, the target cluster number is directly set to 1. When the number of candidate clusters is greater than or equal to 4, The value starts from 2 and increases in increments of 1, with a maximum value not exceeding 2. The square root of a square root rounded down; for example When the value is 100, the square root is 10, and the number of candidate clusters is... Take the numbers 2, 3, 4, 5, 6, 7, 8, 9, 10 in sequence; when When the value is 4, the square root is 2, and the number of candidate clusters is... If there is only one possible value, 2, then the platform will directly use... To determine the target number of clusters, the selection process for the comprehensive scoring formula is skipped.

[0034] In each candidate cluster number Next, the platform performs pre-clustering on the dimensionality-reduced feature data. The pre-clustering uses the K-means algorithm, which starts by randomly selecting initial cluster centers and obtains the number of candidate clusters after iterative convergence. The clustering results include the cluster label of each sample and the centroid coordinates of each cluster. The maximum number of iterations for pre-clustering is set to 300. If the maximum number of iterations is reached but the clustering still fails to converge, the result of the last iteration is used as the clustering result for that candidate clustering number.

[0035] For each candidate clustering result, the cost management platform calculates the profile coefficient. Calinski-Harabasz index and Davies-Bouldin index Three clustering effectiveness evaluation indicators.

[0036] Profile coefficient This value is used to evaluate the compactness and separation of clustering results. A higher value indicates a better clustering effect. During the calculation, each sample is traversed, and the average distance between that sample and all other samples in the same cluster is calculated as the intra-cluster cohesion. The inter-cluster separation is defined as the average distance from the sample to all samples in the nearest neighbor cluster. The nearest neighbor cluster is defined as: the cluster with the smallest centroid distance from the sample to all other clusters, excluding the cluster to which the sample belongs, using the inter-cluster separation. Subtract intra-cluster cohesion Difference divided by the larger of the two The contour value of the sample is obtained, and the range of the contour value of a single sample is as follows: A value close to 1 indicates that the sample clustering is reasonable, while a value close to -1 indicates that the sample may have been misclassified. The arithmetic mean of all sample profile values ​​is the number of candidate clusters. Profile coefficients below .

[0037] Calinski-Harabasz index The ratio of inter-cluster dispersion to intra-cluster compactness is used to evaluate the clustering effect. A larger value indicates a better clustering effect. The calculation first calculates the global centroid, which is the mean vector of the dimensionality-reduced feature vectors of all samples. Then, the inter-cluster dispersion matrix is ​​calculated by multiplying the square of the difference between each cluster's centroid and the global centroid by the number of samples in that cluster, and summing the results over all clusters. Finally, the intra-cluster dispersion matrix is ​​calculated by summing the squares of the differences between each sample and the centroid of its cluster, and summing the results over all samples. It equals the inter-cluster dispersion divided by the intra-cluster dispersion, then multiplied by... The normalization factor, where This represents the number of samples.

[0038] Davies-Bouldin Index Used to evaluate inter-class similarity, a smaller value indicates better clustering results. During calculation, for each cluster... First, calculate the average distance from all samples within the cluster to the cluster centroid. Then for each other cluster , computation cluster with cluster similarity ,in For clusters Center of mass and cluster The Euclidean distance between the centroids, for a cluster ,Pick cluster Similarity to the most similar cluster, for all clusters The arithmetic mean is obtained. The cost management platform determines the optimal number of clusters according to the following comprehensive scoring formula: ; in, Indicates the number of candidate clusters; Indicates the target number of clusters; Represents the number of candidate clusters The corresponding contour coefficient; Represents the number of candidate clusters The corresponding Calinski-Harabasz index; Represents the total number of candidate clusters. The maximum value in; Represents the number of candidate clusters The corresponding Davies-Bouldin index; Represents the total number of candidate clusters. The maximum value in; , , These are the weighting coefficients.

[0039] First Divide by the total number of candidate clusters The maximum value is normalized so that Normalization to Interval; will Divide by the total number of candidate clusters The maximum value is normalized, which also makes Normalization to Interval; due to The smaller the better; its negative value is used in the overall score. Normalization eliminates the impact of numerical scale differences among the three indicators on the overall score, ensuring that the weighting coefficients truly reflect the relative importance of each indicator. , , The initial configuration is Take 0.5, Take 0.3, The platform also supports adjusting the weights based on the actual application scenario: when the sample data quality within the classification unit is high and there are few outliers, the silhouette coefficient can be appropriately increased. The weight of the silhouette coefficient is adjusted because it is more adaptable to the shape of the clusters; when the sample distribution is relatively uniform and the size of each cluster varies greatly, the weight can be appropriately increased. index The weight can be appropriately increased when strict control of inter-cluster separation is required. index The platform calculates the weights of each candidate cluster; The candidate cluster with the highest comprehensive score is determined as the target cluster number. .

[0040] S104, perform clustering based on the target cluster number and remove outlier samples, including: Based on the target cluster number Cluster initialization is performed multiple times. During each initialization, the platform uses a K-means++ initialization strategy to select initial cluster centers. The specific steps are as follows: A dimensionality-reduced feature vector is uniformly and randomly selected from the dimensionality-reduced feature data as the first cluster center. For each dimensionality-reduced feature vector not yet selected as a cluster center... Calculate the shortest Euclidean distance from the vector to all selected cluster centers, denoted as . The next cluster center is selected according to the probability distribution for each sample. The probability of being selected as the next cluster center is Divide by all unselected samples The sum of the probabilities is proportional to the square of the shortest distance. The farther a sample is from the selected cluster center, the greater its probability of being selected. This ensures that the initial cluster centers are as dispersed as possible in the data space, avoiding the local optimum problem caused by the initial cluster centers being too concentrated, until the selected cluster centers are found. The initial cluster centers.

[0041] In a single clustering iteration, the platform calculates the Euclidean distance from each reduced-dimensional feature vector to each cluster center. The Euclidean distance is calculated as the square root of the sum of the squares of the differences between the two vectors in each dimension. The platform assigns the reduced-dimensional feature vectors to the cluster group corresponding to the cluster center with the smallest distance. After all samples are assigned, the platform calculates the arithmetic mean of the reduced-dimensional feature vectors within each cluster group for each feature dimension, obtaining the updated cluster center vectors. The platform repeats the sample assignment and cluster center update steps. After each iteration, the sample assignment is compared with the previous iteration. If no sample's cluster assignment changes, the clustering converges and the iteration stops. If there is a change in assignment, the next iteration begins. The platform sets a maximum of 300 iterations. If convergence is not achieved after reaching the maximum number of iterations, the clustering result of the last iteration is used as the output, and a non-convergence message is recorded.

[0042] For a complete clustering run, the platform uses the following formula for evaluating clustering results: Intra-cluster sum of squares (WCSS): ; in, Indicates the number of candidate clusters; Indicates the first A sample set of each cluster; Indicates belonging to the first Samples from each cluster; Indicates the first The centroid of the cluster; Indicates sample With the corresponding cluster centroid WCSS represents the sum of the squared Euclidean distances between all samples within a cluster and the centroid of the corresponding cluster. The smaller the value, the more compact the samples within the cluster and the better the clustering effect.

[0043] Because the K-means algorithm is sensitive to initial cluster centers, different initializations may lead to different clustering results. Therefore, the cost management platform executes multiple clustering runs with different initial cluster centers. The number of runs is dynamically determined based on the sample size: when the sample size... Run 20 times when the value is less than 50; Run 15 times when the number is between 50 and 200; when When the WCSS value is greater than 200, the process is run 10 times. Each run uses a different random seed to generate initial cluster centers, resulting in multiple candidate clustering results. The platform calculates the WCSS value for each candidate clustering result and determines the candidate clustering result with the smallest WCSS as the final clustering result. If multiple candidate clustering results have the same WCSS, the silhouette coefficient is selected from them. The largest one is taken as the final clustering result.

[0044] For each cluster in the clustering results, the cost management platform calculates a statistic of the unit cost within that cluster, including the mean. The standard deviation is obtained by calculating the arithmetic mean of the unit costs of all samples within the group. The first quartile is obtained by taking the square root of the sum of the squares of the differences between the unit cost and the mean of each sample within the group, divided by the sample size minus one (i.e., using the sample standard deviation). The platform then sorts the unit cost data within the group in ascending order to determine the first quartile. and the third and fourth quartiles First quartile The third quartile is the value at the 25th percentile after sorting. If the 25th percentile lies between two data points, it is obtained through linear interpolation. For the values ​​at the 75th percentile after sorting, linear interpolation is also used to obtain the interquartile range by subtracting the first quartile from the third quartile. ,Right now .

[0045] The cost management platform uses a combination of standard deviation anomaly criteria and interquartile range anomaly criteria to identify anomalous samples. The standard deviation anomaly criteria are as follows: ; in, Indicates the unit cost of the sample; This represents the average cost per unit of the cluster to which the sample belongs; This represents the standard deviation of the unit cost of the cluster to which the sample belongs. This condition determines whether the sample's unit cost deviates from the mean of its cluster by more than 3 times the standard deviation. Under the assumption of normality, the probability of a sample deviating from the mean by more than 3 times the standard deviation is approximately 0.3%, which is a low-probability event. The interquartile range outlier condition is: ; in, and These represent the first and third quartiles of the cluster group, respectively. This represents the interquartile range, and this condition determines whether the unit cost of the sample falls within the lower margin. With the top edge Beyond the scope of the data distribution, this method does not rely on assumptions about the data distribution and is equally applicable to skewed distributions.

[0046] For each sample within a group, two outlier conditions are checked one by one. Only cost samples that simultaneously meet both the standard deviation outlier condition and the interquartile range outlier condition are marked as outliers and removed, thus obtaining the target sample data. Samples that meet only one condition but not the other are not marked as outliers and are retained. This double-insurance mechanism effectively avoids erroneous removal: for example, in a skewed distribution, a sample located in the long tail direction may only meet the interquartile range condition but not the standard deviation condition; in this case, the sample is retained. In a concentrated distribution, a sample located on the margin may only meet the standard deviation condition but not the interquartile range condition; in this case, the sample is also retained. Only when both conditions are triggered simultaneously is the sample considered to have truly deviated from the normal range. Records marked as outliers are removed from the target sample data but retained in the outlier sample log. The log records the outlier sample's project identifier, original unit cost, cluster group number, specific conditions triggering the outlier, removal time, and the person who performed the removal operation. This log is used by data management personnel for manual review to confirm the cause of the outlier and decide whether to permanently remove or restore the sample.

[0047] S105, Perform a normality test on the target sample data and determine the reasonable range of cost indicators, including: The cost management platform performs a Shapiro-Wilke normality test on the target sample data for each cluster. The platform sorts the target sample data within each cluster by numerical value and calculates the Shapiro-Wilke test statistic. , The statistic is calculated based on the correlation between the sorted data and the expected value of the standard normal distribution. The closer the value is to 1, the closer the data is to a normal distribution. The closer the value is to 0, the more severely the data deviates from a normal distribution. The corresponding significance probability is also calculated. , The meaning is: under the null hypothesis that the data actually comes from a normal distribution, the current... The probability of statistical measures or more extreme cases; the platform will Compare with the preset significance level of 0.05: If If the value is not less than 0.05, then there is insufficient evidence to reject the normal distribution hypothesis, and the data is determined to conform to a normal distribution; if... If the value is less than 0.05, the normality hypothesis is rejected, and the data is determined to be non-normally distributed. The Shapiro-Wilke test is applicable to datasets with sample sizes between 3 and 5000, and it has good applicability to the sample size range of highway maintenance cost data in this embodiment.

[0048] For data that does not conform to a normal distribution, the cost management platform uses the Box-Cox transformation to make the transformed data as close as possible to a normal distribution. The transformation formula is as follows: ; in, Represent the original unit cost data, requiring ; Indicates the transformation parameters; Represents the natural logarithm; when When the value is not equal to zero, use the original data. After subtracting one from the power and dividing by ;when When the value is zero, take the natural logarithm of the original data and transform the parameters. Automatically determined through maximum likelihood estimation: the platform operates within a preset timeframe. Find the optimal search range Value, search range set to The search step size is set to 0.1, meaning it will try each step sequentially. Take 101 candidate values: -5.0, -4.9, -4.8...4.9, 5.0; for each candidate... The platform calculates the likelihood function value of the transformed data. A larger likelihood function value indicates that the transformed data is closer to a normal distribution. The value with the largest likelihood function is selected. As the optimal transformation parameters, after the transformation, the Shapiro-Wilke normality test is performed again on the transformed data; if the transformed data... If the value is not less than 0.05, it is treated as a normal distribution, and subsequent interval estimation is performed on the transformed scale. After calculating the interval, it is then restored to the original scale through an inverse transformation. If the transformed value still does not satisfy a normal distribution, i.e. If it is still less than 0.05, then the nonparametric method of kernel density estimation is used for distribution estimation, no longer relying on the normal distribution assumption.

[0049] For datasets that do not satisfy the normality assumption, the cost management platform uses the following Gaussian kernel density estimation method: ; in, Indicates in numerical value The probability density estimate at point ; Indicates the number of samples; Indicates bandwidth parameter; This represents the standard normal kernel function, which is the probability density function of the standard normal distribution. ; Indicates the first The basic idea of ​​kernel density estimation for each sample is to place a Gaussian kernel function, i.e., the probability density curve of a standard normal distribution, centered on each sample point, and then apply this kernel function to all samples. The sum of kernel functions divided by With bandwidth The product of these factors yields the overall probability density estimation curve; the bandwidth parameter... Automatically determined using Silverman's rule of thumb: ; in, To express the sample standard deviation, use the sample standard deviation (divided by) ); Indicates sample size and bandwidth parameter Controlling the width of the kernel function: The larger the kernel size, the flatter each kernel function becomes, resulting in a smoother estimated curve after stacking, but this may cause the loss of detailed features of the data distribution. The smaller the kernel size, the sharper each kernel function becomes, and the better the superimposed curve reflects the local characteristics of the data. However, this may lead to overfitting of the sample data and the generation of spurious fluctuations. The Silverman rule of thumb takes into account both the standard deviation and the sample size to achieve a balance between bias (smoothing degree) and variance (fit degree), providing a bandwidth selection suitable for most scenarios. The cost management platform calculates the reasonable range of cost indicators based on the data distribution estimation results.

[0050] For data that follows a normal distribution, including both the original data that follows a normal distribution and the data that follows a normal distribution after a Box-Cox transformation, the normal distribution parameter method is used to determine the lower and upper limits of the index interval: ; ; in, and These represent the lower and upper limits of the reasonable range for cost indicators, respectively. This represents the sample mean; Indicates the sample standard deviation; This represents the two-tailed quantile of the standard normal distribution. The significance level; for example, a 95% confidence level corresponds to From the standard normal distribution table, we can find... For data that has undergone Box-Cox transformation, after calculating the upper and lower limits of the interval at the transformed scale, the upper and lower limits of the interval are restored to the original unit cost scale through inverse Box-Cox transformation.

[0051] For data that does not follow a normal distribution, the percentile method of kernel density estimation is used to determine the index interval: ; ; in, This represents the inverse function of the cumulative distribution function based on kernel density estimation; To indicate the significance level, the platform constructs the cumulative distribution function by numerically integrating the kernel density estimation curve. Divide the data range into several equal intervals, perform a trapezoidal integral on the density estimation function in each interval, and sum the results to obtain the range from negative infinity to... The cumulative probability is obtained by solving the inverse function. quantiles and The quantile corresponds to the value: find the cumulative probability on the cumulative distribution function. and Time corresponding The value is obtained by linear interpolation if an exact match does not exist, also at a 95% confidence level. The lower limit of the interval is set at the 2.5th percentile, and the upper limit is set at the 97.5th percentile, ensuring that approximately 95% of the data falls within this interval.

[0052] In one possible implementation, the cost management platform also verifies the linkage between the comprehensive unit cost and the list of items to ensure the multi-level internal consistency of the indicator system. In road maintenance engineering, the comprehensive index of road engineering (yuan / m²) should meet the following relationship: Multiply the comprehensive unit price of the first item in the bill of quantities by the unit quantity of that item to obtain the total price of the first item; multiply the comprehensive unit price of the second item in the bill of quantities by the unit quantity of that item to obtain the total price of the second item; and so on, repeating the above operation for each item in the bill of quantities until all items in the bill of quantities have been processed; sum the total prices corresponding to all items in the bill of quantities to obtain the total price; divide the total price by the total area of ​​the road surface works to obtain the comprehensive index of the road surface works; after determining the unit price index range for each item in the bill of quantities, it is also necessary to verify the logical consistency between the index range of each item in the bill of quantities and the comprehensive index range. The specific verification method is as follows: multiply the median value of the unit price index range of each item in the bill of quantities by the corresponding unit quantity of the work, sum the results, and then divide by the total area to obtain the estimated comprehensive unit price value. The calculated value is compared with the median value of the comprehensive indicator range. If the absolute value of the deviation exceeds the preset tolerance (e.g., 5%), it indicates an inconsistency between the comprehensive indicator and the list item indicator. The following strategy is used for fine-tuning: Prioritize adjusting the range of the list item with large fluctuations in the quantity of work (i.e., high coefficient of variation of the quantity of work), scaling the upper and lower limits of the range by the same ratio, which is the ratio of the target deviation to the current deviation; after adjustment, re-verify until the mathematical relationship between the comprehensive indicator and each list item indicator is established within the tolerance range. This multi-level linkage constraint ensures the inherent consistency of the indicator system, avoids logical contradictions between the comprehensive indicator and each list item indicator, and improves the reliability and usability of the indicator system.

[0053] S106, Training a least squares support vector machine regression model and predicting construction costs, including: Input features and prediction targets are extracted from the target sample data to construct a least squares support vector machine (LSSVM) regression prediction model. Input features include maintenance type, region, highway technical grade, project scale, and construction time. For categorical variables such as maintenance type, region, and highway technical grade, the platform uses one-hot encoding to convert them into numerical features: each category corresponds to a binary feature bit, with the bit set to 1 when the sample belongs to that category and 0 otherwise. Each category generates Each binary feature bit; for ordered variables such as project size, the platform uses sequential integer encoding, such as 1 for small projects, 2 for medium-sized projects, 3 for large projects, and 4 for extra-large projects; time information is extracted by year and quarter: the year is directly used as a numerical feature, and the quarter is processed through sine and cosine transforms to retain its periodicity, i.e., generating and Two characteristics; the prediction target is the unit cost index.

[0054] Standard SVM regression requires solving a quadratic programming problem with inequality constraints, which has high computational complexity and significantly increases training time when the sample size is large. LSSVM, on the other hand, replaces inequality constraints with equality constraints and... - Replacing the insensitive loss function with the squared error loss function transforms the quadratic programming problem into a linear equation system problem, reducing computational complexity while maintaining good nonlinear fitting ability and generalization performance. This makes it more suitable for regression prediction scenarios with medium sample size and multiple feature dimensions, such as highway maintenance cost data. The training objective function of the least squares support vector machine regression model is: ; The constraints are: ; in, Indicates the training target value; Represents the model weight vector; Represents the model bias term; Indicates the first The regression error of each training sample; This represents the regularization parameter, used to balance the relationship between model complexity and training error. The larger the value, the heavier the penalty for training error. The model tends to fit the training data more accurately, but this may lead to overfitting. The smaller the value, the smoother the model, but it may underfit. Indicates the number of training samples; Indicates the first The input feature vector of each training sample; Indicates the first The target for predicting the unit cost of each training sample; This represents a nonlinear mapping function that maps input features from the original feature space to a higher-dimensional feature space. By introducing kernel function techniques, explicit computation is not required. We only need to calculate the kernel function values ​​of the two samples in the original space.

[0055] The cost management platform uses radial basis functions (RBF) as its kernel function. The RBF kernel function is defined as follows: ,in The RBF kernel parameter controls the width of the kernel function. The RBF kernel function can map the original input features to an infinite-dimensional feature space, handling various nonlinear regression problems, and exhibits good numerical stability and generalization ability. A grid search combined with 5-fold cross-validation method is used to optimize the hyperparameters of the LSSVM model, including the RBF kernel parameter. and regularization parameters The specific process of grid search is as follows: the platform presets RBF kernel parameters. The range of candidate values ​​is Regularization parameters The range of candidate values ​​is ,form A two-dimensional mesh with a total of 81 hyperparameter combinations. For each combination in the mesh... The platform performs 5-fold cross-validation: the target sample data is randomly divided into 5 equal parts, and 4 parts are used as the training set and 1 part as the validation set in turn. An LSSVM model is trained on the training set, and the root mean square error (RMSE) between the predicted and true values ​​is calculated on the validation set. The arithmetic mean of the RMSE from 5 validations is used as the performance evaluation metric for this hyperparameter combination. After traversing all 81 hyperparameter combinations in the grid, the platform selects the combination with the smallest cross-validation RMSE. As the optimal hyperparameter, if the cross-validation RMSE of multiple hyperparameter combinations is the same, then the optimal hyperparameter is selected. The group with the smaller value will be prioritized for models that are smoother and have stronger generalization ability.

[0056] After determining the optimal hyperparameters, the platform retrains the LSSVM model using all target sample data to obtain the final cost prediction model. After the model training is complete, the platform saves the following model parameters for subsequent predictions: optimal RBF kernel parameters. Optimal regularization parameters Model bias terms The input feature vectors of the support vectors (i.e., the samples that provide support in the training samples) and their corresponding Lagrange multipliers. During prediction, for a new input feature vector The predicted value is calculated using the following formula: The summation iterates through all support vectors. This is the RBF kernel function.

[0057] Using a trained cost prediction model, the platform can perform functions such as cost trend prediction, cost estimation for new projects, and early warning of abnormal deviations.

[0058] The cost trend prediction function is used to predict the price trends of various category indicators over the next 6 to 12 months. The platform uses the current time point as a benchmark and generates input feature vectors for future time points in monthly or quarterly steps. It keeps the maintenance type, region, highway technical grade, and project scale unchanged, and only updates the time dimension features. The prediction model is then input sequentially to obtain the unit cost prediction value for each future time point. These prediction values ​​are connected to form a price trend curve. At the same time, the prediction value is compared with the historical indicator range of the category unit, and it is marked whether the prediction value exceeds the historical range.

[0059] The new project cost estimation function allows users to input basic parameters such as maintenance type, region, technical level, project scale, and planned construction time for a new project. The platform automatically generates corresponding input feature vectors, which are then input into the prediction model to obtain point prediction values. Simultaneously, the platform calculates the prediction interval based on the residual distribution of the prediction model on the training set: it calculates the standard deviation of the residuals (actual value minus predicted value) on the training set. The predicted value is doubled or tripled for each point in the prediction interval. Assuming the residuals follow a normal distribution, this interval contains approximately 95% of the true values. The predicted interval and the reasonable range of the indicator can corroborate each other, providing a multi-dimensional reference for cost estimation of new projects.

[0060] The abnormal deviation warning function is used to trigger a manual review warning when the cost of a new project deviates from the predicted value by more than 30%. The platform automatically calculates the percentage deviation between the cost of the new project and the predicted value. The difference is obtained by subtracting the predicted value from the actual value. The absolute value of the difference is obtained as the absolute deviation. The absolute deviation is divided by the predicted value to obtain the deviation ratio. The deviation ratio is multiplied by 100 to obtain the deviation percentage. When the deviation percentage exceeds the preset 30% threshold, the platform automatically sends a warning notification to the reviewers. The warning information includes basic project information, predicted value, actual value, deviation percentage, and historical indicator range of the category unit, prompting the reviewers to pay close attention to the project and conduct manual verification to check for errors in cost preparation or abnormal market fluctuations.

[0061] S107, Update cost indicators based on update trigger information, including: The cost management platform continuously monitors three types of update trigger information: newly added highway maintenance cost data, information price updates, and regular maintenance, to ensure that cost indicators can reflect the latest market changes and data accumulation in a timely manner, and maintain the timeliness and accuracy of the indicator system.

[0062] The first type of update trigger: After the newly added highway maintenance cost data is reviewed and entered into the database, the platform verifies the project attributes, list items, and material unit prices of the new data. The verification rules are completely consistent with the initial data verification rules to ensure the consistency of data quality. After the verification is passed, the platform extracts the five attribute values ​​of the new data to generate a classification key and locates the corresponding classification unit in the multidimensional data cube. If the classification unit corresponding to the classification key already exists, the platform adds the project index of the new data to the project record index list of the classification unit, updates the sample count (increments by 1), and updates the data time range. If the classification unit corresponding to the classification key does not yet exist, the platform creates the classification unit in the data cube and initializes all attributes. The platform then determines the update strategy: if only new project data is added without changing the original feature definition and classification rules, then an incremental update of the indicator range under that classification unit is triggered, that is, only the complete process from S102 to S105 is re-executed for the affected classification unit; a change in feature definition refers to a change in the feature type, number of features, or feature calculation method selected during feature extraction; a change in classification rules refers to a change in the dimension combination of the classification key, the attribute value division standard, or the time aggregation granularity; if the addition of new data leads to a change in feature definition or classification rules, or the addition of attribute value combinations that have not appeared before, causing the classification unit to need to be split or merged, then a full update is performed on all classification units within the affected range.

[0063] The second type of update trigger: Information price update. When the transportation or housing construction authorities release new material information prices, the platform receives the information price update file through the data interface. The file contains the material code, material name, specifications, applicable area, applicable time, and latest information price. Based on the material code and applicable conditions (applicable area and applicable time), the platform searches the indicator database for all historical project records that reference the material code and meet the applicable conditions. For each record found, the platform adjusts it according to the following material price conversion and refresh algorithm: divide the latest information price of the material by the original information price to obtain the price adjustment coefficient; multiply the original unit price of the material by the price adjustment coefficient to obtain the adjusted unit price of the material. The original unit price represents the material unit price recorded in historical projects; the latest information price represents the information price recently released by the authorities; and the original information price represents the information price corresponding to the material at the time of project construction. All three prices must use the same material. To ensure comparability and accuracy of price conversions, the platform uses the same coding, material specifications, region, and unit of measurement. If the original information price is missing from historical project records, the platform employs the following alternative strategies: First, it prioritizes finding the most recently published information price for the material before or after the project's construction period as the original information price. If this is still not possible, it uses the material's historical price index for estimation, obtained from the material price index database published by the relevant authority. If none of the above methods are feasible, the record is marked as having a missing information price and will not participate in price conversion. The original material unit price is retained, and the reason for the missing price is recorded in the indicator update log. Based on the adjusted material unit price, the platform recalculates upwards from the quota item level: first, it updates the total price of the quota item, then summarizes and updates the total price of the list items, and finally summarizes and updates the total project cost. Since information price updates may affect multiple classification units simultaneously, the platform performs a full update for all classification units within the affected scope.

[0064] The third type of update trigger: Regular maintenance. The platform automatically initiates the regular maintenance process at the end of each quarter, with the start time set to the early morning of that day to minimize the impact on daily business operations. The platform reads all valid highway maintenance cost data from the multidimensional data cube, traverses each category unit, and sequentially executes the complete process from S102 to S105 for each category unit: re-extracting cost sample features, re-standardizing and reducing dimensionality, redetermining the target number of clusters, re-performing clustering and outlier removal, re-performing normality tests, and determining the reasonable range of indicators. Regular maintenance does not depend on whether there is new data. Even if some category units have no new data in the quarter, the indicators are recalculated to ensure that the indicators can reflect the overall data and the latest statistical characteristics. After the regular maintenance is completed, the platform generates a maintenance report, recording the total number of category units involved in the maintenance, the number of category units where indicators changed, and the statistics of the change range.

[0065] Based on the updated multidimensional data cube, feature extraction, dimensionality reduction, clustering, outlier removal, and indicator range determination are re-executed to obtain updated indicators. The update strategy is divided into two methods: incremental update and full update. Incremental update is only performed on the affected classification units, which is fast and consumes less computing resources. It is suitable for small-scale data change scenarios where only a small amount of project data is added and the features and classification rules are not changed. Full update performs the complete process on all classification units within the affected range. It is suitable for scenarios such as information price updates, feature definition changes, classification rule changes, or periodic maintenance, ensuring that the calculation logic and statistical caliber of all related indicators are consistent.

[0066] S108, Generate indicator version management data and store it in the indicator database, including: Each time an updated indicator is received, the cost management platform generates an indicator version snapshot. This version snapshot is a metadata document that fully records the entire update process, containing the following information: The update trigger information includes recording the trigger type as either new data entry, information price update, or periodic maintenance. If it is new data entry, the list and quantity of newly added project identifiers are also recorded. If it is an information price update, the updated material code and the latest information price source are also recorded. The update time includes the precise timestamp of the indicator update, accurate to the second, in the format YYYY-MM-DDHH:MM:SS, and the identifier of the person who performed the update operation is also recorded. The indicator value before the update includes the lower and upper limits of the reasonable range of cost indicators for this category unit before the update, as well as the corresponding median value of the indicator. If this category unit is created for the first time, the indicator value field before the update is marked as first creation. The indicator value after the update includes the lower and upper limits of the reasonable range of cost indicators for this category unit after the update, as well as the corresponding median value of the indicator, and the amount and percentage change of the indicator value before and after the update are also recorded. The data scope involved in the update includes the list of affected category units, the number of newly added project records, the number of changed project records, and the number of abnormal samples removed.

[0067] The classification key includes the concatenated result of the five attribute values ​​of the classification unit; the feature definition includes a list of features involved in feature extraction, including the name, type, calculation method and parameters, mean and standard deviation of each feature during standardization; the dimensionality reduction parameters include the number of principal components to be retained. The actual value of the cumulative variance contribution rate, the variance contribution rate and eigenvalue of each retained principal component; the target cluster number includes the target cluster number determined by the comprehensive scoring formula. The data includes the silhouette coefficient, CH index, DB index, and overall score for each candidate cluster number; the outlier identification includes a list of removed outliers, including the project identifier, original unit cost, cluster group number, specific triggering condition type, and removal time for each outlier; the normality test results include the Shapiro-Wilke test results. Statistics and Whether it has undergone Box-Cox transformation and the transformation parameters If the transformed value still does not satisfy a normal distribution, record the distribution estimation method used and the bandwidth parameter. The interval calculation method includes recording the final interval estimation method used, which is one of the following: normal distribution parameter method, Box-Cox transformation parameter method, or kernel density estimation percentile method. The significance level is also recorded. Model parameters include recording the RBF kernel parameters and regularization parameters of the LSSVM model. The model's performance metrics on the training set (RMSE and ), as well as the number of training samples and the number of support vectors.

[0068] The platform assigns version identifiers to updated metrics. These identifiers are generated using an incrementing integer version number combined with a timestamp, in the following format: Where YYYYMMDD represents the update date, and the serial number is the version number for that day, starting from 001 and incrementing, for example... This indicates the first version generated on March 1, 2025.

[0069] The platform associates and stores updated metrics and metric version management data into the metric database. When storing the data, the platform sets the effective end time of the previous version as the effective time of the new version, so that the version timeline of this category unit forms a continuous and non-overlapping time series. The effective start time of a version is the update time of that version, and the effective end time is the update time of the next version. In this way, any point in time can uniquely identify a valid version, enabling precise time tracking of metrics.

[0070] When querying historical versions, the cost management platform retrieves version identifiers that are valid at the target time based on the category key and target time. The platform then searches the indicator database for all version records under that category key, filters versions according to their valid time range, and returns the version identifier and the corresponding reasonable range of cost indicators. If the specified version does not exist, the platform returns a version unavailable flag and indicates the available time range. If the data range is incomplete, the platform returns a version unavailable flag and explains the missing data items. If the measurement caliber corresponding to the version is inconsistent with the data to be audited, the platform also returns a version unavailable flag and explains the specific reason for the measurement caliber mismatch, such as the measurement unit in the indicator database being yuan / m², the measurement unit in the data to be audited being yuan / km, and there is no definite conversion relationship.

[0071] S109, Match indicator ranges and generate audit support decision information, including: The platform retrieves the cost estimates documents to be reviewed, parses the project attributes, list items, and material unit prices within them, and generates data to be reviewed. The platform supports parsing cost estimates documents in various formats, including but not limited to spreadsheet formats and standardized cost data exchange formats. During the parsing process, the platform automatically identifies the file format and calls the corresponding parser to extract basic project information, list item details, and material unit price details from the file. For format anomalies or missing fields encountered during the parsing process, the platform records the anomaly location and anomaly type in the parsing log and attempts to continue parsing the remaining parts.

[0072] The platform verifies the data to be audited according to the list code, material code, and unit of measurement. It verifies whether each list code exists in the standard list code library; if not, it marks it as an unknown list code and prompts the auditor for confirmation. It also verifies whether each material code exists in the material standard code library; if not, it marks it as an unknown material code. Finally, it verifies whether the unit of measurement is valid and consistent with the units of measurement specified in the standard list code library or material standard code library, converting convertible units of measurement to the units of measurement used in the indicator database.

[0073] For the data to be audited that has passed the verification, the cost management platform generates a category key for auditing based on five dimensions: maintenance type, region, highway technical grade, construction time and project scale. The platform then retrieves a reasonable range of currently valid cost indicators with consistent measurement standards from the indicator database based on the category key for auditing, and uses this range as the matching indicator range.

[0074] During the search, the platform employs a strategy of prioritizing exact matching and using approximate matching as a fallback. It first attempts exact matching: searching the indicator database for currently valid versions whose category keys exactly match the category key to be reviewed. If an exact match is successful, it directly returns the reasonable range for the cost indicators of that version. If an exact match fails, the platform attempts approximate matching: successively relaxing the time dimension, project size dimension, and region dimension, searching again after relaxing one dimension at a time until a matching indicator range is found. The approximate matching results are marked with the matching method in the review results; for example, if the time dimension has been relaxed to a quarterly match or the project size dimension has been relaxed to an adjacent range, the reviewer is advised that the matching precision may have decreased. If all relaxation strategies fail to find a matching indicator range, the platform returns a message indicating that no matching indicator range exists and suggests that the reviewer make a manual judgment.

[0075] The cost management platform determines the auditable indicator values ​​based on the list items and material unit prices in the data to be audited, and determines the deviation according to the following deviation calculation method: The platform determines whether the value of the indicator to be reviewed falls within the matching indicator range. If the value of the indicator to be reviewed is greater than or equal to the lower limit of the range and less than or equal to the upper limit of the range, the deviation is zero, indicating that the value of the indicator to be reviewed is within the normal fluctuation range and does not require special attention.

[0076] If the value of the indicator to be audited falls outside the interval, the deviation is calculated. The specific calculation method is as follows: take the difference between the value of the indicator to be audited and the nearest interval boundary value, and take the absolute value of the difference to obtain the absolute deviation; take the difference between the upper limit of the interval and the lower limit of the interval to obtain the interval width; divide the absolute deviation by the interval width to obtain the deviation ratio; multiply the deviation ratio by 100% to obtain the deviation percentage. Among them, the nearest interval boundary value refers to the boundary value that is closer to the value of the indicator to be audited between the upper limit of the interval and the lower limit of the interval. If the value of the indicator to be audited is higher than the upper limit of the interval, the nearest interval boundary value is the upper limit of the interval; if the value of the indicator to be audited is lower than the lower limit of the interval, the nearest interval boundary value is the lower limit of the interval. The deviation is a dimensionless percentage and is always a non-negative value. The larger the deviation, the more serious the deviation of the value of the indicator to be audited from the normal range.

[0077] The platform handles two special cases: First, if the upper limit of the matching indicator range equals the lower limit, and the value of the indicator to be reviewed equals the range value, the numerator of the deviation calculation formula is zero. The platform directly sets the deviation to zero percent, indicating that the value of the indicator to be reviewed perfectly matches the range, requiring no further calculation. Second, if the upper limit of the matching indicator range equals the lower limit, and the value of the indicator to be reviewed does not equal the range value, the denominator of the deviation calculation formula is zero, making division impossible. The platform does not perform division calculations, directly assigns this item to the third warning status, and generates an anomaly explanation for a zero range width. The anomaly explanation includes: the sample size on which the indicator range is based, possible reasons for a zero range width, and a reminder to reviewers that the reference value of this indicator range is limited, suggesting manual judgment based on indicators from other sources or expert experience.

[0078] The cost management platform compares the deviation to a first deviation threshold and a second deviation threshold sequentially. The first deviation threshold is set at 20%, and the second deviation threshold is set at 50%. When the deviation is not greater than the first deviation threshold, it is determined to be in the first warning state, marked as green (normal), and the item is displayed in green on the review interface, indicating that the item has passed automatic review and requires no manual intervention. When the deviation is greater than the first deviation threshold but not greater than the second deviation threshold, it is determined to be in the second warning state, marked as yellow (warning), and the item is highlighted in yellow on the review interface with a suggested verification mark. When the deviation is greater than the second deviation threshold, it is determined to be in the third warning state, marked as red (warning), and the item is highlighted in red on the review interface with a key attention mark.

[0079] The cost management platform automatically generates a variance analysis report, which includes three levels of content: The first level is project-level information, which records: project ID to be reviewed, category key to be reviewed, matching indicator version ID, and matching method.

[0080] The second level is item-level details, which record each list item or material unit price that deviates from the green normal range, item by item: list item name or material name, list code or material code, unit of measurement, indicator value to be reviewed, lower and upper limits of the matching indicator range, range width, deviation calculation value, warning status, and triggering reason.

[0081] The third level is a summary of statistical information, recording: the total number of audited items, the number and percentage of green normal items, the number and percentage of yellow alert items, the number and percentage of red warning items, and the number and percentage of items in the third warning state.

[0082] The platform aggregates all project results within the same cost document awaiting review, packages the difference analysis report and review support decision information, and sends them to the review interface. The review support decision information includes: the overall conclusion of this review, a list of red warning items, a list of yellow reminder items, and a detailed deviation analysis of each warning item. Reviewers can view the details, confirm the review results, and fill in the review comments on the review interface. The review results and comments are simultaneously written back to the cost management platform, forming a complete review loop.

[0083] like Figure 2 As shown, embodiments of the present invention also provide a deep learning-based dynamic calculation and intelligent review system for highway maintenance cost indicators, comprising: The acquisition module is used to acquire highway maintenance cost data and construct a multi-dimensional data cube based on the highway maintenance cost data. The extraction module is used to extract cost sample features from the multidimensional data cube and perform dimensionality reduction processing on the extracted features to obtain dimensionality-reduced feature data; perform cluster analysis based on the dimensionality-reduced feature data to obtain clustering results; and remove outlier samples based on the clustering results to obtain target sample data. The testing module is used to perform normality tests on the target sample data and determine the reasonable range of cost indicators based on the test results. The prediction module is used to train a least squares support vector machine regression model based on the target sample data and its corresponding project attributes to obtain a cost prediction model; and to obtain the cost prediction result based on the cost prediction model. The analysis module is used to acquire newly added highway maintenance cost data, classify the newly added highway maintenance cost data into a multidimensional data cube, re-perform cluster analysis and determine indicator intervals to obtain updated indicators, and store the updated indicators in the indicator database. The matching module is used to obtain the cost documents to be reviewed and extract the data to be reviewed from them. Based on the project attributes of the data to be reviewed, it retrieves the reasonable range of the corresponding cost indicators from the indicator database as the matching indicator range. It calculates the deviation based on the matching indicator range and the data to be reviewed. If the deviation exceeds the preset deviation threshold, it generates an early warning and difference analysis report to obtain review auxiliary decision-making information.

[0084] It should be noted that this system is a system corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0085] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for dynamic calculation and intelligent review of highway maintenance cost indicators based on deep learning, characterized in that, Executed by the cost management platform, including: Obtain highway maintenance cost data and construct a multi-dimensional data cube based on the highway maintenance cost data; Extract cost sample features from the multidimensional data cube and perform dimensionality reduction on the extracted features to obtain dimensionality-reduced feature data; perform cluster analysis based on the dimensionality-reduced feature data to obtain clustering results; remove outlier samples based on the clustering results to obtain target sample data; Perform a normality test on the target sample data, and determine the reasonable range of cost indicators based on the test results; A least squares support vector machine regression model is trained based on the target sample data and its corresponding project attributes to obtain a cost prediction model; the cost prediction result is obtained based on the cost prediction model. Acquire new highway maintenance cost data, incorporate the new highway maintenance cost data into a multidimensional data cube, re-perform cluster analysis and determine indicator intervals to obtain updated indicators, and store the updated indicators in the indicator database. Obtain the cost documents to be reviewed and extract the data to be reviewed from them. Based on the project attributes of the data to be reviewed, retrieve the reasonable range of the corresponding cost indicators from the indicator database as the matching indicator range. Calculate the deviation based on the matching indicator range and the data to be reviewed. If the deviation exceeds the preset deviation threshold, generate an early warning and difference analysis report to obtain review auxiliary decision-making information.

2. The method according to claim 1, characterized in that, Cluster analysis was performed on the dimensionality-reduced feature data to obtain the clustering results, including: Based on the number of samples in the dimensionality reduction feature data, multiple candidate cluster numbers are set; Cluster the dimensionality-reduced feature data under each candidate cluster number to obtain the candidate clustering results; The silhouette coefficient, inter-class dispersion-intra-class compactness ratio, and inter-class similarity index were calculated for each candidate clustering result. The silhouette coefficient and the inter-class dispersion and intra-class compactness ratio indices are positively normalized, while the inter-class similarity index is negatively normalized. The comprehensive score is calculated based on each normalization result and its corresponding preset weight, and the candidate cluster number with the highest comprehensive score is determined as the target cluster number. Cluster the dimensionality-reduced feature data according to the target number of clusters to obtain the clustering results.

3. The method according to claim 2, characterized in that, Clustering is performed on the dimensionality-reduced feature data based on the target cluster number to obtain the clustering results, including: Based on the target number of clusters, an improved mean clustering initialization method is used to determine multiple initial cluster centers from the dimensionality-reduced feature data. Clustering iterations were performed on the dimensionality-reduced feature data using the initial cluster centers of each group to obtain multiple candidate clustering results; Calculate the sum of squared distances between samples within each cluster and the corresponding cluster center of each candidate clustering result, and determine the candidate clustering result with the smallest sum of squared distances as the clustering result.

4. The method according to claim 3, characterized in that, Based on the clustering results, outlier samples are removed to obtain the target sample data, including: Based on the clustering results, the mean, standard deviation, first quartile, and third quartile of the cost samples within each cluster group were determined. Construct standard deviation abnormality conditions based on the mean and standard deviation; The interquartile range is determined based on the first and third quartiles, and interquartile range outlier conditions are constructed. Cost samples that simultaneously meet the standard deviation anomaly condition and the interquartile range anomaly condition are removed from the clustering results to obtain the target sample data.

5. The method according to claim 4, characterized in that, Perform a normality test on the target sample data, and determine the reasonable range of cost indicators based on the test results, including: The Shapiro-Wilke normality test was performed on the target sample data, and the first test result was obtained; For target sample data whose first test results satisfy a normal distribution, the normal distribution index interval is determined according to the normal distribution interval estimation method, and used as the reasonable range of cost index. For target sample data whose first test result does not meet the normal distribution, the Box-Cox transformation is performed on the target sample data, and the normality test is performed again on the transformed sample data to obtain the second test result; For transformed sample data whose second test results satisfy a normal distribution, the transformation index interval is determined according to the normal distribution interval estimation method, and the transformation index interval is inversely transformed to obtain the reasonable cost index interval. For transformed sample data whose second test results do not meet the normal distribution, Gaussian kernel density estimation is performed on the transformed sample data, and the nonparametric index interval is determined based on the estimated fractions, which serves as the reasonable range for the cost index.

6. The method according to claim 1, characterized in that, Based on the target sample data and its corresponding project attributes, a least squares support vector machine regression model is trained to obtain the cost prediction model, including: Extract project attributes from the target sample data as model input features, and extract unit cost as the prediction target; Based on the preset candidate kernel parameter range and candidate regularization parameter range, construct multiple sets of candidate model parameter combinations; Cross-validation of the target sample data was performed using the parameter combinations of each candidate model, and the cross-validation error of each candidate model was calculated. The candidate model parameter combination with the minimum cross-validation error is determined as the target kernel parameter and the target regularization parameter; Based on the target kernel parameters and target regularization parameters, with project attributes as input and unit cost as the prediction target, a least squares support vector machine regression model is trained to obtain the cost prediction model.

7. The method according to claim 5, characterized in that, Acquire newly added highway maintenance cost data, incorporate this data into a multidimensional data cube, and re-perform cluster analysis and indicator interval determination to obtain updated indicators, including: Get update trigger information; When the update triggers the addition of new highway maintenance cost data to the database, the new highway maintenance cost data will be included in the multidimensional data cube. When an update triggers an update of the information representation price, the material unit price in the multidimensional data cube is refreshed according to the updated material unit price, and the project cost of the corresponding list item is re-determined. When the update triggers the periodic maintenance of the information representation, all highway maintenance cost data in the multidimensional data cube are read. Based on the updated multidimensional data cube, feature extraction, dimensionality reduction, clustering, outlier removal, and indicator range determination are performed again to obtain updated indicators.

8. The method according to claim 7, characterized in that, The updated metrics will be stored in the metrics database, including: Generate a snapshot of the indicator version based on the updated indicators; Write the update trigger information, update time, indicator values ​​before and after the update, and the data range involved in the update into the indicator version snapshot to generate indicator version management data; Updated metrics and metric version management data are associated and stored in the metric database. The metric database supports retrieving historical versions of updated metrics based on metric version management data.

9. The method according to claim 8, characterized in that, The deviation is calculated based on the matching indicator range and the data to be reviewed. If the deviation exceeds the preset deviation threshold, an early warning and difference analysis report are generated, including: The values ​​of the indicators to be audited are determined based on the list items and material unit prices in the data to be audited. Determine the absolute value of the difference between the value of the indicator to be reviewed and the center value of the matching indicator interval; The difference between the upper and lower limits of the matching indicator range is determined as the range width; The deviation is determined by the ratio of the absolute value of the difference to the width of the interval; When the deviation is not greater than the first deviation threshold, the first warning state is determined and the corresponding level of warning information is generated. When the deviation is greater than the first deviation threshold but not greater than the second deviation threshold, the second warning state is determined and the corresponding level of warning information and difference analysis report is generated. When the deviation exceeds the second deviation threshold, a third warning state is determined and a corresponding level of warning information and difference analysis report is generated.

10. A deep learning-based dynamic calculation and intelligent review system for highway maintenance cost indicators, wherein the system implements the method as described in any one of claims 1 to 9, characterized in that, include: The acquisition module is used to acquire highway maintenance cost data and construct a multi-dimensional data cube based on the highway maintenance cost data. The extraction module is used to extract cost sample features from the multidimensional data cube and perform dimensionality reduction processing on the extracted features to obtain dimensionality-reduced feature data. Cluster analysis is performed on the dimensionality-reduced feature data to obtain the clustering results; Based on the clustering results, outlier samples are removed to obtain the target sample data; The testing module is used to perform normality tests on the target sample data and determine the reasonable range of cost indicators based on the test results. The prediction module is used to train a least squares support vector machine regression model based on the target sample data and its corresponding project attributes to obtain a cost prediction model; and to obtain the cost prediction result based on the cost prediction model. The analysis module is used to acquire newly added highway maintenance cost data, classify the newly added highway maintenance cost data into a multidimensional data cube, re-perform cluster analysis and determine indicator intervals to obtain updated indicators, and store the updated indicators in the indicator database. The matching module is used to obtain the cost documents to be reviewed and extract the data to be reviewed from them. Based on the project attributes of the data to be reviewed, it retrieves the reasonable range of the corresponding cost indicators from the indicator database as the matching indicator range. It calculates the deviation based on the matching indicator range and the data to be reviewed. If the deviation exceeds the preset deviation threshold, it generates an early warning and difference analysis report to obtain review auxiliary decision-making information.

Images