A Road Material Cost Calculation Method Based on Generative Adversarial Networks and Conditional Regression

Abstract

This invention relates to a method for calculating road material costs based on generative adversarial networks and conditional regression, belonging to the field of data processing and analysis for road material cost data calculation. It includes the following steps: collecting road material cost data to construct a road surface cost dataset; preprocessing the road surface cost dataset to obtain engineering feature data, environmental feature data, and actual road costs; constructing a road material cost calculation model; inputting the preprocessed data into the model to train it; calculating the error between the actual road costs and the predicted data, using a loss function to measure model performance, and using an optimizer to optimize the parameters of each module to obtain a trained model; inputting the preprocessed road data to be predicted into the trained model, and outputting the predicted material costs for the road project. This invention can ensure high-accuracy prediction in changing engineering scenarios and enhance the model's generalization ability and environmental adaptability.

Description

Technical Field

[0001] This invention belongs to the field of road material cost data calculation and data processing analysis, specifically involving a road material cost calculation method based on generative adversarial networks and conditional regression. Background Technology

[0002] In road construction projects, material costs typically account for a major proportion of the total project cost. Therefore, accurate calculation of material costs is crucial for project budgeting, cost control, and investment decisions. Traditional methods for calculating material costs largely rely on manual calculations, historical data, and expert experience. While effective in simple cases, these methods have become increasingly limited as engineering projects become more complex, especially when faced with varying design parameters, geographical environments, and construction conditions, where the accuracy and reliability of predictions are restricted. Currently common methods for calculating road material costs, such as experience-based estimation, statistical regression analysis, and rule-based parametric models, while providing reference in certain situations, suffer from low accuracy and poor adaptability: experience-based methods struggle to handle complex requirements; statistical regression relies on large amounts of historical data and is not adept at handling nonlinear relationships; and rule-based models perform poorly under different design and environmental conditions.

[0003] In recent years, Generative Adversarial Networks (GANs) and Conditional Regression Models have been successfully applied in various fields, especially in generating data and making conditional predictions. GANs, through adversarial training of the generator and discriminator, can generate high-quality sample data in unsupervised learning; Conditional Regression Models can make accurate predictions based on input conditional information. In material cost estimation, GANs can generate diverse cost samples, while Conditional Regression Models can accurately predict actual costs based on different design and environmental conditions. However, existing GANs and regression models are mostly applied independently and lack a comprehensive solution for the specific task of road material cost estimation. Existing techniques cannot effectively improve prediction accuracy when data is insufficient or unevenly distributed, and they struggle to fully capture the nonlinear relationships between complex designs and environmental conditions. Therefore, a road material cost estimation method based on GANs and Conditional Regression can overcome these shortcomings and provide more accurate and reliable prediction results. By simulating cost fluctuations under different conditions using Conditional Generative Adversarial Networks (cGANs) and combining this with accurate predictions using Conditional Regression Models, the stability and predictive ability of the model can be significantly improved. Summary of the Invention

[0004] To achieve the above objectives, the present invention employs the following technical solution:

[0005] This invention provides a method for calculating road material costs based on generative adversarial networks and conditional regression, comprising the following steps:

[0006] Collect road material cost data to construct a pavement cost dataset; preprocess the pavement cost dataset to obtain engineering feature data. Environmental characteristic data and the actual road construction cost In the road construction cost dataset, the original engineering feature data includes road type, road grade, road width, project scale, pavement structure layers and thickness, design load, and material proportion; the original environmental feature data includes average annual precipitation, temperature range, and humidity; the original actual road construction cost includes concrete cost, asphalt cost, steel reinforcement cost, and aggregate cost.

[0007] A road material cost calculation model is constructed, including a condition-driven data builder, a condition-guided generation engine, an adversarial control joint optimizer, a multimodal adversarial fusion engine, an environmental feature optimization module, and a cost-aware regression network; the engineering feature data is then used to calculate the cost. Environmental characteristic data and the actual road construction cost Input the data into the road material cost calculation model to train the model;

[0008] By calculating the actual road construction cost The error between the model and the predicted data O is used to measure the model performance using a loss function, and the Adam optimizer is used to backpropagate and iteratively optimize the parameters of each module until the model converges, thus obtaining the trained model.

[0009] After preprocessing, the road data to be predicted is input into the trained model, which outputs the predicted material costs of the road project.

[0010] Furthermore, the data preprocessing operation involves encoding all non-numerical categorical variables in the dataset to obtain engineering feature data. Environmental characteristic data and the actual road construction cost Encoding methods include: tag encoding and one-hot encoding.

[0011] Furthermore, the condition-driven data builder includes a feature embedding layer, a first fully connected layer, a second fully connected layer, a third fully connected layer, and a fourth fully connected layer;

[0012] The engineering feature data Features are categorized into categorical features and continuous features. Categorical features include road type, road grade, and design load. Continuous features include road width, project scale, pavement structure thickness, and material proportion.

[0013] Each categorical feature is represented by an embedding function, mapping each category index to an embedding vector; the continuous features are uniformly projected through a linear layer to obtain a projection vector. In the feature embedding layer, the embedding vectors corresponding to each category are concatenated to obtain the category vector. The projection vector The activation vector is obtained after processing with the ReLU activation function. ; class vector and activation vector By concatenating the vectors, we obtain the intermediate vector. The intermediate vector After processing through the first, second, third, and fourth fully connected layers, the engineering-driven features are obtained. .

[0014] Furthermore, a condition-guided generation engine and an adversarial regulation joint optimizer are employed to effectively capture complex nonlinear feature interactions and optimize feature representations; the condition-guided generation engine includes a channel attention extraction module, a first bottleneck mapping layer, a second bottleneck mapping layer, and a residual enhancement output module;

[0015] The channel attention extraction module is used to perform channel-level feature weighting on the input engineering-driven features, highlighting key dimensions and suppressing invalid features. The channel attention extraction module includes a first compression layer and a second expansion layer; the engineering-driven features... The extended feature is obtained by sequentially passing through the first compression layer and the second expansion layer. The extended features With engineering-driven features Perform element-wise multiplication to obtain the weighted features. ;

[0016] Weighted features The input is fed into the first bottleneck mapping layer to obtain the first bottleneck mapping features. Map the first bottleneck features The input is fed into the second bottleneck mapping layer to obtain the second bottleneck mapping features. The second bottleneck mapping feature and engineering-driven features After weighted residual fusion by the residual enhancement output module, the engineering guidance features are obtained. .

[0017] Furthermore, the adversarial regulation joint optimizer includes a dual-stream feature alignment module, a feature interaction fusion unit, and a difference enhancement module;

[0018] The dual-stream feature alignment module is used to structurally reorganize two sets of features—engineering-driven features and engineering-guided features—to make them fusionable. The dual-stream feature alignment module includes a first branch path and a second branch path; the engineering-driven features... The first branch feature is obtained through the first branch path. Engineering-driven features Features of the first branch Adding them together yields the first merged feature. The second branch path includes a LayerNorm layer and a fifth fully connected layer; the engineering guidance feature After normalization using the LayerNorm layer, normalized features are obtained. The normalized features The first activation feature is obtained after passing through the fifth fully connected layer. ;

[0019] The first merging feature and the first activation feature After being concatenated by the feature interaction fusion unit, joint features are obtained. The joint features After passing through the sixth fully connected layer, the second activation feature is obtained. ;

[0020] The difference enhancement module includes a ReLU activation function and a Sigmoid activation function; the second activation feature The third activation feature is obtained by preserving the feature representation after applying the ReLU activation function. The second activation feature The fourth activation feature is obtained by generating gated weights using the Sigmoid activation function. The third activation feature and the fourth activation feature Element-wise multiplication is performed to obtain the engineering optimization features. .

[0021] Furthermore, the engineering guidance features Engineering optimization features and environmental characteristic data The input is fed into a multimodal adversarial fusion processor to obtain fused features. A multimodal adversarial fusion engine is introduced to optimize the fusion between different feature sources through adversarial learning, thereby improving the stability and accuracy of the model. The data processing procedure of the multimodal adversarial fusion engine is as follows:

[0022] ,

[0023] in, Indicates fusion characteristics; Represents the L2 norm; · indicates a constant; · indicates element-wise multiplication; Indicates engineering guidance features The transpose of [the transpose of the ...

[0024] Furthermore, the fusion features and environmental characteristic data The input is fed into the environmental feature optimization module to obtain the environmental optimization features. The environmental feature optimization module enables the model to make more accurate predictions under complex environmental changes by optimizing environmental features; this module ensures that the model can adaptively handle the impact of different environments on road material costs by adjusting the weights of environmental features.

[0025] In practical engineering, environmental factors (such as geographical location and climate) have a significant impact on material costs, but traditional models often ignore changes in these factors. The environmental feature optimization module addresses this issue by dynamically adjusting the weights of environmental features, reducing the uncertainty of environmental factors, and improving the model's prediction accuracy under different environmental conditions. The data processing procedure of the environmental feature optimization module is as follows:

[0026] ,

[0027] in, exp represents the environmental optimization characteristics; exp represents the exponential function.

[0028] Furthermore, the environmental optimization features Engineering optimization features and engineering feature data The data is input into a cost-aware regression network to obtain predicted road construction cost data O. The cost-aware regression network includes a feature fusion input layer, a deep regression modeling layer, and a multi-output mapping layer. The deep regression modeling layer includes a three-layer perception network. The multi-output mapping layer includes a material single-item prediction head and a total cost prediction head.

[0029] The environmental optimization features and engineering optimization features After element-wise addition at the feature fusion input layer, the first fused feature is obtained. The first fusion feature Engineering feature data After concatenation, the features are obtained by applying the SiLU activation function. The splicing features After processing by a deep regression modeling layer, perceptual features are obtained. In the multi-output mapping layer, the perceptual features The material cost prediction result is obtained through the material item prediction head; the perceived features The total cost forecast is obtained by using the total cost forecast head; the material cost forecast results and the total cost forecast results are combined to obtain the road construction cost forecast data O.

[0030] The advantages of this invention are:

[0031] This invention significantly improves the accuracy and stability of road material cost prediction through the synergistic effect of its various modules. Using a multi-output regression network, the model can not only accurately predict the costs of various materials but also simultaneously estimate the total cost, providing a more detailed cost analysis. Generative adversarial optimization and adversarial learning mechanisms enhance the model's ability to capture complex nonlinear relationships, improving its robustness in situations with insufficient or imbalanced data. The environmental feature optimization module enables the model to adapt to different environmental conditions, ensuring high-accuracy predictions even in dynamic real-world engineering scenarios. This effectively solves problems such as data incompatibility and overfitting in traditional methods, enhancing the model's generalization ability and environmental adaptability. Attached Figure Description

[0032] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0033] Figure 1 This is a flowchart of the steps of the method of the present invention;

[0034] Figure 2 The contribution of different models to the SHAP value of input features. Detailed Implementation

[0035] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0036] Example 1

[0037] In this embodiment, as Figure 1As shown, this invention provides a method for calculating road material costs based on generative adversarial networks and conditional regression, the specific steps of which include:

[0038] Step 1: Collect road material cost data and construct the road cost dataset Rocost; preprocess the road cost dataset to obtain engineering feature data. Environmental characteristic data and the actual road construction cost ;

[0039] Specifically, the first step is to define the engineering characteristic data, environmental characteristic data, and actual road construction cost in detail, and collect relevant data. The original engineering characteristic data mainly includes data reflecting the structure, scale, design, and construction requirements of the road construction project. Specific parameters include: road type: expressway, urban road, rural road, secondary road, bridge; road grade: national, provincial, county road; road width (unit: meters); project scale (unit: kilometers); pavement structure layers and thickness: base course, surface course, subbase (unit: meters); design load: heavy, extra-heavy load; material percentage: asphalt, concrete, steel reinforcement, aggregate (unit: %). The original environmental characteristic data mainly describes environmental and climate-related factors that affect material costs, construction progress, and road lifespan. Specific parameters include: average annual precipitation: low, medium, high; temperature range: low, medium, high; humidity: low, medium, high. The original actual road construction cost reflects the actual expenditure data of the project, and specific parameters include: concrete cost: cost per cubic meter (unit: yuan / cubic meter), asphalt cost: cost per ton (unit: yuan / ton), steel reinforcement cost: cost per ton (unit: yuan / ton), and sand and gravel cost: cost per ton (unit: yuan / ton).

[0040] All data is compiled into a single spreadsheet, with each data point categorized by type. The units for each data point are determined and standardized. Missing values ​​are imputed and outliers are handled to ensure data quality and prevent inaccuracies in the prediction model due to missing or outlier data. Z-score normalization is used to ensure that values ​​of different features can be compared on the same scale. The data is divided into training and testing sets, typically in an 80%-20% ratio, with 80% used for training and 20% for testing, resulting in the Rocost road cost dataset.

[0041] Specifically, the non-numerical categorical variables in all data in the dataset are encoded to obtain engineering feature data. Environmental characteristic data and the actual road construction cost ;

[0042] Encoding methods include: label encoding and one-hot encoding; Label Encoding: suitable for category features with a natural order, such as road grades (high, medium, low); One-Hot Encoding: suitable for category features without an order relationship, such as road type.

[0043] Step 2: Construct a road material cost calculation model, including a condition-driven data builder, a condition-guided generation engine, an adversarial control joint optimizer, a multimodal adversarial fusion engine, an environmental feature optimization module, and a cost-aware regression network. The engineering feature data... Environmental characteristic data and the actual road construction cost The data is input into the road material cost calculation model to train the model.

[0044] Specifically, the engineering feature data After processing by the condition-driven data builder, engineering-driven features are obtained. ;

[0045] The conditional data builder includes a feature embedding layer, a first fully connected layer, a second fully connected layer, a third fully connected layer, and a fourth fully connected layer. The first fully connected layer has an output dimension of 64 and uses the ReLU activation function; the second fully connected layer has an output dimension of 128 and uses the ReLU activation function with Dropout set to 0.3; the third fully connected layer has an output dimension of 128 and uses the ReLU activation function; and the fourth fully connected layer has an output dimension of 64 and does not use an activation function. The design of the conditional data builder allows each layer to progressively enhance the model's feature extraction, fusion, and abstraction capabilities. From initial categorical data embedding to feature optimization through deep fully connected layers, the model can efficiently learn the complex relationships between different features. This progressive structure effectively improves the model's ability to handle complex data and ensures that it can fully capture the complex patterns hidden in the data when processing data under different engineering conditions, thereby improving the accuracy and stability of road material cost prediction.

[0046] The engineering feature data Features are categorized into categorical features and continuous features. Categorical features include road type, road grade, and design load. Continuous features include road width, project scale, pavement structure thickness, and material composition.

[0047] Each feature in the categorical features is represented by an embedding function, mapping each category index to a 16-dimensional embedding vector. The continuous features are uniformly projected through a linear layer to obtain a 64-dimensional projection vector. In the feature embedding layer, the embedding vector corresponding to each category is... The vectors are concatenated to obtain a 48-dimensional category vector. The projection vector The activation vector is obtained after processing with the ReLU activation function. ; class vector and activation vector The vectors are concatenated and merged into a single 112-dimensional intermediate vector. ;

[0048] The intermediate vector After processing through the first, second, third, and fourth fully connected layers, the engineering-driven features are obtained. .

[0049] Specifically, the engineering-driven features The input is fed into the conditional guidance generation engine to obtain engineering guidance features. ;

[0050] The condition-guided generation engine includes a channel attention extraction module, a first bottleneck mapping layer, a second bottleneck mapping layer, and a residual enhancement output module.

[0051] The channel attention extraction module is used to perform channel-level feature weighting on the input engineering-driven features, highlighting key dimensions and suppressing invalid features. The channel attention extraction module includes a first compression layer and a second expansion layer; the first compression layer is configured to compress 64-dimensional features into 16-dimensional features using a fully connected layer; the parameters are: output dimension 16, activation function ReLU activation function; the second expansion layer is configured to restore the features to 64-dimensional features using a fully connected layer; the parameters are: output dimension 64, activation function Sigmoid activation function, used to generate channel attention weights.

[0052] The engineering driving features The extended feature is obtained by sequentially passing through the first compression layer and the second expansion layer. The extended features With engineering-driven features Perform element-wise multiplication to obtain the weighted features. ;

[0053] The first bottleneck mapping layer is implemented using a fully connected layer with an input dimension of 64 and an output dimension of 128, and the activation function is ReLU; weighted features are then processed. The input is fed into the first bottleneck mapping layer to obtain the first bottleneck mapping features. ;

[0054] The second bottleneck mapping layer is implemented using a fully connected layer with an input dimension of 128 and an output dimension of 64, and does not use an activation function; it maps the features of the first bottleneck layer. The input is fed into the second bottleneck mapping layer to obtain the second bottleneck mapping features. ;

[0055] The second bottleneck mapping feature and engineering-driven features After weighted residual fusion by the residual enhancement output module, the engineering guidance features are obtained. The fusion method is as follows: Preferably, Set it to 0.6.

[0056] Specifically, the engineering-driven features and the engineering guidance features The input is fed into the adversarial regulation joint optimizer to obtain the engineering optimization features. ;

[0057] The adversarial regulation joint optimizer includes: a dual-stream feature alignment module, a feature interaction fusion unit, and a difference enhancement module;

[0058] The dual-stream feature alignment module is used to structurally reorganize the two sets of features, namely engineering-driven features and engineering-guided features, so that they can be fused.

[0059] The dual-stream feature alignment module includes a first branch path and a second branch path; the first branch path includes two fully connected layers, the first fully connected layer having an input dimension of 64 and an output dimension of 64, and not using an activation function; the second fully connected layer also has an input dimension of 64 and an output dimension of 64, and does not use an activation function; the engineering-driven feature... The first branch feature is obtained through the first branch path. Engineering-driven features Features of the first branch Adding them together yields the first merged feature. ;

[0060] The second branch path includes a LayerNorm layer and a fifth fully connected layer; the fifth fully connected layer has an input dimension of 64 and an output dimension of 64, and uses the SiLU activation function; the engineering-guided features After normalization using the LayerNorm layer, normalized features are obtained. The normalized features The first activation feature is obtained after passing through the fifth fully connected layer. ;

[0061] The first merging feature and the first activation feature After being concatenated by the feature interaction fusion unit, 128-dimensional joint features are obtained. The joint features After passing through the sixth fully connected layer, the second activation feature is obtained. The sixth fully connected layer has an input dimension of 128 and an output dimension of 128, and uses the SiLU activation function.

[0062] The difference enhancement module includes a ReLU activation function and a Sigmoid activation function; the second activation feature After passing through the ReLU activation function, the feature representation is preserved, resulting in the third activation feature. The second activation feature The fourth activation feature is obtained by generating gated weights using the Sigmoid activation function. The third activation feature and the fourth activation feature Element-wise multiplication is performed to obtain the engineering optimization features. .

[0063] Specifically, the engineering guidance features Engineering optimization features and environmental characteristic data The input is fed into a multimodal adversarial fusion processor to obtain fused features. ;

[0064] The data processing procedure of the multimodal adversarial fusion engine is as follows:

[0065] ,

[0066] in, Indicates fusion characteristics; Represents the L2 norm; · indicates a constant; · indicates element-wise multiplication; Indicates engineering guidance features The transpose of the multimodal adversarial fusion engine enables the model to effectively learn and optimize the relationship between multiple sources of information by adversarially fusing information from different modalities; by introducing an adversarial learning mechanism, the fused features can reflect the nonlinear interaction between various feature sources.

[0067] Specifically, the fusion features and environmental characteristic data The input is fed into the environmental feature optimization module to obtain the environmental optimization features. ;

[0068] The environmental feature optimization module enables the model to make more accurate predictions under complex environmental changes by optimizing environmental features; this module also ensures that the model can adaptively handle the impact of different environments on road material costs by adjusting the weights of environmental features.

[0069] In practical engineering, environmental factors (such as geographical location and climate) have a significant impact on material costs, but traditional models often ignore changes in these factors. The environmental feature optimization module addresses this issue by dynamically adjusting the weights of environmental features, reducing the uncertainty of environmental factors, and improving the model's prediction accuracy under different environmental conditions. The data processing procedure of the environmental feature optimization module is as follows:

[0070] ,

[0071] in, exp represents the environmental optimization characteristics; exp represents the exponential function.

[0072] Specifically, the environmental optimization features Engineering optimization features and engineering feature data The data is input into a cost-aware regression network to obtain predicted road construction costs O.

[0073] The cost-aware regression network includes a feature fusion input layer, a deep regression modeling layer, and a multi-output mapping layer.

[0074] The environmental optimization features and engineering optimization features After element-wise addition at the feature fusion input layer, a first fused feature with dimension 64 is obtained. The first fusion feature Engineering feature data After concatenation, the features are obtained by applying the SiLU activation function. ;

[0075] The deep regression modeling layer is the core regression modeling structure, employing a three-layer perceptual network for layer-by-layer feature abstraction. The deep regression modeling layer comprises three perceptual networks: the first layer has an input dimension of 88 and an output dimension of 128, uses the ReLU activation function, and has Dropout set to 0.3; the second layer has an input dimension of 128 and an output dimension of 128, uses the ReLU activation function, and employs Batch Normalization to stabilize the feature distribution; the third layer has an input dimension of 128 and an output dimension of 64, and uses the ReLU activation function; the concatenated features... The perceptual features are obtained by sequentially processing the data through a deep regression modeling layer. ;

[0076] The multi-output mapping layer includes a material single-item prediction header and a total cost prediction header;

[0077] In the multi-output mapping layer, the perceptual features The material cost prediction result is obtained through the material-specific prediction head. The material-specific prediction head uses a fully connected layer with an output dimension of 4, indicating that the cost of each material is predicted individually, and each output value corresponds to the cost of one material (asphalt, concrete, steel reinforcement, aggregate). The perceived features... The total cost prediction head is used to obtain the total cost prediction result; the total cost prediction head uses a fully connected layer and has an output dimension of 1; the material cost prediction result and the total cost prediction result are merged to obtain the road construction cost prediction data O.

[0078] Step 3: Calculate the actual road construction cost The error between the model and the predicted data O is used to measure the model performance using a loss function. The Adam optimizer is then used to backpropagate and iteratively optimize the parameters of each module until the model converges, resulting in a trained model.

[0079] Step 4: After preprocessing the road data to be predicted, input it into the trained model and output the predicted material cost of the road project.

[0080] Example 2

[0081] In this embodiment, data from 1000 actual road construction projects over the past 5 years were selected from a provincial transportation engineering database. The data was processed and divided into a training set (800 sets) and a test set (200 sets). Using this data, the predictive performance of the method of this invention was compared with that of a traditional regression model. The predictive performance comparison of different models is shown in Table 1.

[0082] Table 1. Comparison of the predictive performance of the method of the present invention and the traditional regression model.

[0083]

[0084] The table above shows the performance of five common cost prediction models on the same test dataset: Linear Regression, Support Vector Regression (SVR), Random Forest Regression (RF), Conditional Regression, and the model proposed in this invention. The comparison metrics used are Mean Absolute Percentage Error (MAPE) and Root Mean Square Error (RMSE), which measure the prediction accuracy and error fluctuation level of the models, respectively. It is evident that the method proposed in this invention performs best in both metrics: MAPE is 6.2%, significantly lower than the traditional linear model's 17.3%, SVR's 14.6%, RF's 12.1%, and RMSE's 10.8%, demonstrating extremely strong prediction stability and adaptability; RMSE is only 48.3 yuan / m², approximately 36% lower than Conditional Regression (75.6 yuan / m²) and RF (82.5 yuan / m²), and more than 50% lower than the linear model, proving that this method has a significant advantage in error control.

[0085] This invention utilizes conditional generative adversarial networks to simulate the distribution of cost samples under different engineering and environmental conditions, addressing the issues of insufficient and imbalanced data and enhancing the model's generalization ability to diverse scenarios. Furthermore, by fusing with conditional regression networks, it significantly improves prediction accuracy when input features are known. Compared to traditional methods that rely solely on historical data regression, this method can more deeply capture the complex nonlinear relationship between design parameters and material costs, maintaining stable prediction performance even when road type, material mix, and environmental climate change significantly.

[0086] The distribution of prediction bias of the model under complex environments is shown in Table 2:

[0087] Table 2. Prediction bias of the model in complex environments.

[0088]

[0089] The table above shows the distribution of prediction biases for road material costs under complex environmental conditions (such as high precipitation, high humidity, and extreme temperature differences). The horizontal axis represents different model types, and the vertical axis represents the deviation between the predicted and actual values ​​(unit: yuan / m²). This invention selected 200 road projects under complex environments as test samples and statistically analyzed the mean, standard deviation, and extreme error values ​​of each model. It can be observed that the traditional conditional regression model exhibits large fluctuations in prediction bias under complex environments, with a mean bias of +22.5 yuan, a standard deviation of ±42.3 yuan, and a maximum error exceeding 121 yuan / m², reflecting its instability in handling environmental diversity and nonlinear effects. The model proposed in this invention, under the same scenario, has a mean bias of only +5.3 yuan, a standard deviation of ±18.7 yuan, and a maximum error controlled within 49 yuan / m², indicating that this method significantly improves the adaptability to extreme and complex conditions.

[0090] The multimodal adversarial fusion engine introduced in this method deeply integrates engineering-guided features, optimization features, and environmental variables, effectively capturing the cross-influence between environmental variables and material costs during the modeling process. Simultaneously, the adversarial training mechanism helps the model learn a broader data distribution, making the generated predictions more "realistic." Figure 2 As shown, the analysis illustrates the response strength of each model to the input variables, employing the SHAP (SHapley Additive exPlanations) method to reveal the contribution of different features to the model's prediction results. The figure lists the 10 most critical features and their average SHAP values ​​to the model output (material cost prediction). The Y-axis represents the average SHAP value (feature contribution), and the X-axis represents the input variables of different models. Blue bars represent traditional conditional regression models; orange bars represent the model of this invention.

[0091] As shown in the figure, traditional conditional regression models are limited by structural factors such as road width and project scale, while their response to environmental features (such as temperature and humidity) is weak, indicating that they fail to fully capture environmental impacts during modeling. In the SHAP graph presented by the method of this invention, environmental features (annual precipitation, humidity level, and temperature range) and engineering features together constitute the core basis for model judgment, with their average contribution on par with road structure parameters. In particular, the input variable "material proportion" has a significantly increased SHAP weight in this method, reflecting the model's ability to deeply understand the impact of different material combinations on cost structure, rather than simply performing linear weight fitting. This figure further verifies that the method of this invention, through the environmental feature optimization module and cost-aware regression network, not only improves prediction accuracy but also enhances model interpretability and engineering controllability. Users can use the SHAP analysis results to clearly determine which design variables are sensitive to cost, thereby assisting in optimizing engineering design and resource allocation.

[0092] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for calculating road material costs based on generative adversarial networks and conditional regression, characterized in that, Includes the following steps: Collect road material cost data to construct a pavement cost dataset; preprocess the pavement cost dataset to obtain engineering feature data. Environmental characteristic data and the actual road construction cost ; A road material cost calculation model is constructed, including a condition-driven data builder, a condition-guided generation engine, an adversarial control joint optimizer, a multimodal adversarial fusion engine, an environmental feature optimization module, and a cost-aware regression network; the engineering feature data is then used to calculate the cost. Environmental characteristic data and the actual road construction cost The data is input into the road material cost calculation model to train the model; specifically, The condition-driven data builder includes a feature embedding layer, a first fully connected layer, a second fully connected layer, a third fully connected layer, and a fourth fully connected layer; it stores the engineering feature data. After processing by the condition-driven data builder, engineering-driven features are obtained. ; The condition-guided generation engine includes a channel attention extraction module, a first bottleneck mapping layer, a second bottleneck mapping layer, and a residual enhancement output module; the channel attention extraction module includes a first compression layer and a second expansion layer; and it processes the engineering-driven features. The input is fed into the conditional guidance generation engine to obtain engineering guidance features. ; The adversarial regulation joint optimizer includes a two-stream feature alignment module, a feature interaction fusion unit, and a difference enhancement module; the two-stream feature alignment module includes a first branch path and a second branch path; the second branch path includes a layer normalization layer (LayerNorm) and a fifth fully connected layer; the difference enhancement module includes a ReLU activation function and a Sigmoid activation function; and the engineering-driven features are... and the engineering guidance features The input is fed into the adversarial regulation joint optimizer to obtain the engineering optimization features. ; Engineering guidance features Engineering optimization features and environmental characteristic data The input is fed into a multimodal adversarial fusion processor to obtain fused features. The data processing procedure of the multimodal adversarial fusion engine is as follows: ， in, Indicates fusion characteristics; Indicates engineering guidance features; Indicates engineering optimization characteristics; Represents environmental characteristic data; Represents the L2 norm; · indicates a constant; · indicates element-wise multiplication; Indicates engineering guidance features Transpose of; Fusion features and environmental characteristic data The input is fed into the environmental feature optimization module to obtain the environmental optimization features. The data processing procedure of the environmental feature optimization module is as follows: ， in, Represents environmental optimization characteristics; exp represents an exponential function; The cost-aware regression network includes a feature fusion input layer, a deep regression modeling layer, and a multi-output mapping layer; the environmental optimization features Engineering optimization features and engineering feature data The data is input into a cost-aware regression network to obtain predicted road construction costs O. By calculating the actual road construction cost The error between the model and the predicted data O is used to measure the model performance using a loss function, and the Adam optimizer is used to backpropagate and iteratively optimize the parameters of each module until the model converges, thus obtaining the trained model. After preprocessing, the road data to be predicted is input into the trained model, which outputs the predicted material costs of the road project.

2. The method for calculating road material costs based on generative adversarial networks and conditional regression according to claim 1, characterized in that, The engineering feature data Features are categorized into categorical features and continuous features; Each categorical feature is represented by an embedding function, mapping each category index to an embedding vector; the continuous features are uniformly projected through a linear layer to obtain a projection vector. In the feature embedding layer, the embedding vectors corresponding to each category are concatenated to obtain the category vector. The projection vector The activation vector is obtained after processing with the ReLU activation function. ; class vector and activation vector By concatenating the vectors, we obtain the intermediate vector. The intermediate vector After processing through the first, second, third, and fourth fully connected layers, the engineering-driven features are obtained. .

3. The method for calculating road material costs based on generative adversarial networks and conditional regression according to claim 2, characterized in that, The engineering driving features The extended feature is obtained by sequentially passing through the first compression layer and the second expansion layer. The extended features With engineering-driven features Perform element-wise multiplication to obtain the weighted features. Weighted features The input is fed into the first bottleneck mapping layer to obtain the first bottleneck mapping features. Map the first bottleneck features The input is fed into the second bottleneck mapping layer to obtain the second bottleneck mapping features. The second bottleneck mapping feature and engineering-driven features After weighted residual fusion by the residual enhancement output module, the engineering guidance features are obtained. .

4. The method for calculating road material costs based on generative adversarial networks and conditional regression according to claim 3, characterized in that, The engineering driving features The first branch feature is obtained through the first branch path. Engineering-driven features Features of the first branch Adding them together yields the first merged feature. The engineering guidance features After normalization using the LayerNorm layer, normalized features are obtained. The normalized features The first activation feature is obtained after passing through the fifth fully connected layer. ; The first merging feature and the first activation feature After being concatenated by the feature interaction fusion unit, joint features are obtained. The joint features After passing through the sixth fully connected layer, the second activation feature is obtained. ; In the difference enhancement module, the second activation feature The third activation feature is obtained by preserving the feature representation after applying the ReLU activation function. The second activation feature The fourth activation feature is obtained by generating gated weights using the Sigmoid activation function. The third activation feature and the fourth activation feature Element-wise multiplication is performed to obtain the engineering optimization features. .

5. The method for calculating road material costs based on generative adversarial networks and conditional regression according to claim 4, characterized in that, The environmental optimization features and engineering optimization features After element-wise addition at the feature fusion input layer, the first fused feature is obtained. The first fusion feature Engineering feature data After concatenation, the features are obtained by applying the SiLU activation function. ; The deep regression modeling layer includes a three-layer perceptual network; the spliced ​​features After processing by a deep regression modeling layer, perceptual features are obtained. ; The multi-output mapping layer includes a material single-item prediction header and a total cost prediction header; In the multi-output mapping layer, the perceptual features The material cost prediction result is obtained through the material item prediction head; the perceived features The total cost forecast is obtained by using the total cost forecast head; the material cost forecast results and the total cost forecast results are combined to obtain the road construction cost forecast data O.

6. The method for calculating road material costs based on generative adversarial networks and conditional regression according to claim 5, characterized in that, The road construction cost dataset contains original engineering feature data including road type, road grade, road width, project scale, pavement structure layers and thickness, design load, and material proportion; original environmental feature data including average annual precipitation, temperature range, and humidity; and original actual road construction cost including concrete cost, asphalt cost, steel reinforcement cost, and aggregate cost.

7. The method for calculating road material costs based on generative adversarial networks and conditional regression according to claim 6, characterized in that, The data preprocessing operation is as follows: Encode the non-numerical categorical variables in all data in the dataset to obtain engineering feature data. Environmental characteristic data and the actual road construction cost ; Encoding methods include: tag encoding and one-hot encoding.

8. The method for calculating road material costs based on generative adversarial networks and conditional regression according to claim 7, characterized in that, The categorical features include road type, road grade, and design load; the continuous features include road width, project scale, pavement structure thickness, and material proportion.

Images