Railway line shape constraint perception type deep learning generation method based on B2C framework

Abstract

The invention provides a railway line shape constraint perception type deep learning generation method based on a B2C framework, and belongs to the technical field of railway engineering design and deep learning. 2) carrying out B2C constraint compiling; 3) preparing multi-source geographic data; according to the method, the multi-main-body game logic and the deep learning model are organically fused, a new path is provided for solving the multi-target conflict problem in linear infrastructure line selection, and the line selection efficiency and the scheme reasonability can be effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of railway engineering design and deep learning technology, and particularly relates to a railway line constraint perception type deep learning generation method based on a B2C framework. BACKGROUND

[0002] The route design of large-scale transportation infrastructure (such as cross-regional railways and highways) is a typical socio-technical system (Socio-technical Systems) challenge, and the core complexity comes from the deep interweaving of three dimensions: first, the uncertainty of the physical environment (such as geological conditions and terrain undulations), second, the hard constraints of the engineering level (such as curvature limits, slope standards and construction feasibility), and third, the strategic interaction of multiple interest subjects (investment parties pursuing cost-benefit balance, design parties focusing on technical specification compliance, and local governments considering social development and ecological protection). This interweaving makes the "optimal design" no longer a simple physical adaptation or single-objective optimization problem - the final landing of the project not only needs to meet the engineering technical requirements, but also needs to coordinate the conflicting goals of interest subjects in the information asymmetry scenario, and the current mainstream method is exactly difficult to break through this core contradiction.

[0003] Existing researches are mostly focused on single-dimensional optimization: pure physical driven methods (such as path search based on terrain elevation) only focus on the adaptability of the route and the geographical environment, completely ignoring the strategic uncertainty brought by "human factors"; single-objective engineering optimization (such as minimizing total cost or construction period) often leads to social cost spillover (such as ecological damage caused by excessive ground disturbance and land expropriation compensation disputes); even a few studies considering multi-objective optimization also simplify the interest conflicts by using weight allocation, and fail to depict the real interaction logic of "decision-making mutual influence and incomplete information" among subjects. These limitations make it difficult for traditional methods to meet the complex decision-making needs of large-scale projects, and there is an urgent need for a theoretical framework that can unify physical constraints, engineering requirements and interest balance from the source. SUMMARY

[0004] The application provides a railway line shape constraint perception type deep learning generation method based on a B2C framework, and a Bayes-to-Constraints (B2C) framework is used for intelligent route selection of linear infrastructure. The framework first constructs a three-party game model of an investor, a design institute and a local government, quantizes multi-agent interest demands into a set of interpretable design constraint parameters (including an upper limit of a tunnel proportion, a safety margin of earth covering, a longitudinal slope threshold value and a curvature / structure penalty factor) through Bayes Nash equilibrium deduction, then takes multi-source geographic data (digital elevation terrain data, DWG engineering data and the like) as input, trains a route-attribute double-task probability model based on deep learning, generates a line shape space probability distribution graph, and finally integrates the design parameters obtained through game deduction into a probability graph decoding process through a constraint perception decoding mechanism, and automatically outputs an infrastructure center line scheme meeting multi-agent interest equilibrium and satisfying engineering construction requirements. The method realizes organic fusion of multi-agent game logic and a deep learning model, provides a new path for solving multi-objective conflict problems in linear infrastructure route selection, and can effectively improve route selection efficiency and scheme rationality.

[0005] To achieve the above object, the application adopts the following technical scheme: The railway line shape constraint perception type deep learning generation method based on the B2C framework comprises the following steps: 1) Multi-agent Bayes game modeling: determining the participants as an investor, a design institute and a local government, calculating expected utilities of the participants, solving a Bayes Nash equilibrium by using an iterative optimal response algorithm, and obtaining optimal strategies of the participants; 2) B2C constraint compilation: constructing a compilation function, mapping the optimal strategies into an engineering constraint parameter set, and the engineering constraint parameter set comprising an upper limit of a tunnel proportion, a safety margin of earth covering, a ground cost weight, a curvature penalty coefficient, a longitudinal slope penalty coefficient and a structure penalty coefficient; 3) Multi-source geographic data preparation: obtaining multi-source geographic data, and performing standardization processing to obtain a training set, a verification set and a test set; 4) Deep learning model training: adopting an improved UNet as a training model, taking the training set as input, outputting a line shape space probability distribution graph by a route prediction head of the model, outputting attribute probability graphs of a roadbed, a bridge and a tunnel by an attribute prediction head, adopting a positive sample priority sampling strategy and a sample enhancement strategy in a training process, using a composite loss function to optimize the model, and obtaining a deep learning model; 5) Reasoning and constraint decoding: using the deep learning model, sliding window reasoning is adopted on the validation set or test set, the prediction results of the overlapping areas of the sliding windows are fused by Hanning weight to obtain the linear space probability distribution graph and the attribute probability graph; combining the engineering constraint parameter set, the linear space probability distribution graph and the attribute probability graph are decoded: based on the upper limit of the tunnel proportion, the candidate linear that exceeds the limit is filtered, combined with the safety margin of the overburden, the tunnel segment prediction that does not meet the requirements is removed, and the attribute probability is weighted according to the ground cost weight to preferentially select the linear attribute with low land cost; 6) Output railway linear center line scheme: the decoded result is output as a constructible railway center line scheme in the same coordinate system as the multi-source geographic data, including mileage pile number, geographic coordinates and roadbed / bridge / tunnel attributes.

[0006] In this specification, the step 1) of solving the Bayesian Nash equilibrium by using the iterative optimal response algorithm specifically includes the following steps: 1.1) Initialization: set the initial tunnel proportion of the investor, the initial risk threshold of the design institute for different geological types, and the initial compensation standard of the local government for different subsidy types, and set the iteration counter to 0; 1.2) Iterative update: repeat the following operations until convergence: 1.2.1) Investor updates strategy: according to the current risk threshold of the design institute and the current compensation standard of the local government, solve the tunnel proportion that maximizes the expected utility of the investor; 1.2.2) Design institute updates strategy: for each geological type, according to the current tunnel proportion of the investor, solve the risk threshold that maximizes the expected utility of the design institute; 1.2.3) Local government updates strategy: for each subsidy type, according to the current tunnel proportion of the investor, solve the compensation standard that maximizes the expected utility of the local government; 1.3) Convergence test: calculate the relative change of the tunnel proportion before and after updating, the relative change of the risk threshold of each geological type, and the relative change of the compensation standard of each subsidy type; if the maximum relative change is less than the tolerance of one hundred thousandth, the algorithm terminates and the optimal strategy of the Bayesian Nash equilibrium is obtained; otherwise, increase the iteration counter by 1 and return to step 1.2).

[0007] In this specification, the step 2) of mapping the optimal strategy to the engineering constraint parameter set by using the compiling function specifically includes the following steps: 2.1) Calculate the upper limit of the tunnel proportion: the upper limit of the tunnel proportion is equal to the first aggregation weight multiplied by the minimum-maximum normalized optimal tunnel proportion of the investor, plus the second aggregation weight multiplied by the minimum-maximum normalized average of the optimal compensation standard of the local government, wherein the sum of the first aggregation weight and the second aggregation weight is 1; 2.2) Calculate the earthwork safety margin: the earthwork safety margin is equal to the margin scaling coefficient multiplied by the minimum-maximum normalized design institute optimal risk threshold average value; 2.3) Calculate the ground cost weight: the ground cost weight is equal to the unit conversion coefficient multiplied by the local government optimal compensation standard average value, and the unit conversion coefficient is determined by the regression of the corresponding relationship between the land comprehensive cost in the historical railway route selection project and the monetary unit in the game model; 2.4) Calculate the curvature penalty coefficient: the curvature penalty coefficient is equal to the third aggregation weight multiplied by (1 minus the minimum-maximum normalized design institute optimal risk threshold average value), plus the fourth aggregation weight multiplied by (1 minus the minimum-maximum normalized investor optimal tunnel proportion), wherein the sum of the third aggregation weight and the fourth aggregation weight is 1.

[0008] In this specification, the multi-source geographic data preparation in step 3) specifically includes the following steps: 4.1) Digital elevation terrain data processing: Obtain digital elevation terrain data in the railway route selection area, fill the NoData value in the data with neighborhood interpolation method to ensure data continuity; 4.2) DWG engineering data processing: Obtain DWG engineering data in the railway route selection area, extract the horizontal and vertical section data, bridge and tunnel pile number data and ground object data therefrom by using CAD data analysis tool, and rasterize the ground object data according to types, wherein the non-ground object area is the first type, the general cultivated land and other general ground object area is the second type, and the building and river and other important ground object area is the third type; 4.3) Standardization and cropping: unify the processed digital elevation terrain data and DWG engineering data to the same coordinate system by using coordinate conversion tool, and crop them into 10 kilometer section case packages of 1000x1000 pixels by using slicing tool, each case package is associated with corresponding regional location information; 4.4) Quality inspection and data division: generate a quality assurance report containing coordinate system offset, NoData value distribution and data integrity information based on multi-source geographic data, filter qualified case packages according to the quality assurance report, and remove case packages with coordinate system offset exceeding tolerance and NoData value ratio exceeding 5%; divide the training set, validation set and test set in the ratio of 7:2:1, and ensure that each data set is non-overlapping to avoid sliding window leakage.

[0009] In this specification, the architecture of the improved UNet in step 4) specifically includes the following parts: 5.1) Encoder side: Set up 4 convolutional modules, each of which contains 3×3 convolution operation, BatchNorm normalization operation and ReLU activation function in sequence; downsampling is achieved by 3×3 convolution with stride of 2, and the number of feature map channels doubles with each downsampling, and multi-scale features of digital elevation terrain data, restricted area data and cost area data are extracted step by step; 5.2) Decoder side: Each layer is upsampled through 2×2 transposed convolution, which halves the number of channels in the feature map after upsampling; the upsampled feature map is then merged with the feature map of the corresponding layer in the encoder through skipconnection for channel concatenation and feature alignment, finally restoring the same pixel-level prediction resolution as the input data; 5.3) Dual-head output layer: The route prediction head is a 1-channel output, which outputs a linear spatial probability distribution map of the 0-1 interval through the Sigmoid activation function, used to characterize the probability that each pixel is the railway centerline; the attribute prediction head is a 3-channel output, which outputs the category probability map of roadbed, bridge and tunnel through the Softmax activation function, and invalid prediction results in non-linear areas are masked by setting ignore_index=255.

[0010] In this specification, the positive sample priority sampling strategy and sample enhancement strategy in step 4) specifically include the following: 6.1) Positive Sample Priority Sampling: The proportion of positive samples is controlled by parameters. The positive samples are the labeled railway line area data. In the initial training stage, the positive sample clipping ratio is set to 1.0, and all positive samples are sampled to ensure that the model can quickly learn the line features. In the later training stage, difficult negative samples with a ratio of 0.1-0.3 are introduced. The difficult negative samples are non-target area data with a distance of less than 5 pixels from the line area, which improves the model's ability to distinguish between target and non-target areas. 6.2) Sample augmentation: A center-jitter sample augmentation method is adopted, which randomly shifts the center position of each sample horizontally and vertically within a 32-pixel range, while keeping the sample size unchanged. This avoids the model overfitting to a fixed linear position and enhances the model's generalization ability in different location scenarios.

[0011] In this specification, the composite loss function in step 4) specifically includes the following: 7.1) Route prediction loss: The BCEWithLogits loss plus 0.5 times the SoftDice loss is adopted. The BCEWithLogits loss addresses the problem of imbalanced linear sample classes by setting positive and negative sample weights, while the SoftDice loss optimizes the prediction accuracy of sparse linear regions by calculating the intersection-union ratio of the predicted linear and the real linear, thus alleviating the gradient vanishing problem caused by the scarcity of positive samples. 7.2) Attribute prediction loss: Masked cross-entropy loss is used, and the loss is calculated only within the route neighborhood. The route neighborhood is obtained by performing 3×3 pixel dilation on the true linear label or linear region with a prediction probability greater than 0.2. At the same time, the ground cost weight obtained in step 2) is incorporated. The prediction loss of bridge attribute is multiplied by (1 plus ground cost weight), the prediction loss of tunnel attribute is multiplied by (1 plus 2 times ground cost weight), and the prediction loss of roadbed attribute is multiplied by 1, so that the attribute prediction result conforms to the cost constraint in multi-agent game.

[0012] In this instruction manual, step 4) of deep learning model training is specifically a phased training process, including the following steps: 8.1) First stage training: Set the attribute loss weight coefficient to 0 and train only the improved UNet route prediction head; use the training set data prepared in step 3) as input, adopt the Adam optimizer, the initial learning rate is 1 / 100,000, the positive and negative sample weights are set to 12, and optimize the route prediction loss; validate once every 2 training cycles until the linear intersection-union ratio of the validation set is not less than 0.6 to ensure that the linear position prediction is stable; 8.2) Second stage training: Set the attribute loss weight coefficient to 0.5-0.7, and jointly train the route prediction head and the attribute prediction head; keep the Adam optimizer and the initial learning rate unchanged, set the positive and negative sample weights to 10, and simultaneously optimize the route prediction loss and the attribute prediction loss; enable automatic mixed precision training to reduce memory usage and accelerate convergence, and combine gradient clipping to prevent gradient explosion during training. 8.3) Resume training from breakpoint: The model parameter file is saved every 5 training cycles during the training process; when training is interrupted, the most recently saved model parameter file is loaded and training continues from the interrupted training cycle, improving training efficiency.

[0013] In this specification, step 5) of reasoning and constraint decoding specifically includes the following steps: 9.1) Sliding window inference: For the standardized multi-source geographic data of the validation or test set in step 3), a sliding window of size 512×512 pixels is used to traverse the entire data area with a step size of 384 pixels; the prediction results of the overlapping areas of adjacent sliding windows are weighted and fused using the Hanning weight function to eliminate the prediction bias at the edge of the sliding window; finally, a linear spatial probability distribution map and an attribute probability map are output. 9.2) Constraint Decoding: Load the set of engineering constraint parameters obtained in step 2) and perform decoding in the following order: 9.2.1) Tunnel Proportion Filtering: Calculate the proportion of tunnel segment length for each candidate alignment and remove candidate alignments whose proportion exceeds the upper limit of tunnel proportion; 9.2.2) Soil cover safety inspection: Calculate the tunnel section burial depth based on the digital elevation topographic data processed in step 3), remove the predicted tunnel section burial depth that is less than the soil cover safety margin, and supplement the roadbed or bridge attributes. 9.2.3) Attribute Cost Weighting: For each pixel in the attribute probability map, the probabilities of roadbed, bridge, and tunnel are multiplied by 1, (1 plus ground cost weight), and (1 plus ground cost weight plus curvature penalty coefficient), respectively. The attribute with the highest weighted probability is selected as the final attribute of the pixel.

[0014] In this specification, step 6) of outputting the railway alignment centerline scheme specifically includes the following steps: The alignment spatial probability distribution map and attribute probability map after constraint decoding in step 5) are used to obtain the pixel coordinates of the railway centerline through probability threshold filtering and skeleton extraction algorithms; the pixel coordinates are converted into geographic coordinates in the same coordinate system as the digital elevation data in step 3); mileage stations are assigned to the geographic coordinates according to mileage order, and the corresponding roadbed / bridge / tunnel attributes are associated; the above information is output as a constructable railway centerline scheme.

[0015] In summary, the present invention has at least the following beneficial effects: Achieving a balance of interests among multiple stakeholders and resolving the traditional problem of conflicting interests in route selection. This method constructs a Bayesian game model involving investors, design institutes, and local governments, solves for the Bayesian Nash Equilibrium (BNE), and compiles it into actionable engineering constraint parameters, overcoming the limitations of traditional methods that rely on "single-objective optimization" or "weighted simplification of conflicts." On one hand, the cost control demands of investors (such as tunnel construction costs and land compensation costs) are addressed through parameters such as "tunnel ratio upper limit" and "surface cost weight." On the other hand, the engineering safety needs of design institutes (such as geological risks and construction feasibility) are transformed into hard constraints such as "cover safety margin" and "curvature penalty coefficient." The social costs that local governments are concerned about (such as land acquisition disputes and ecological impacts) are incorporated into attribute selection through "surface cost weight mapped by compensation standards." The final linear solution can simultaneously satisfy economic efficiency, safety, and social feasibility, avoiding difficulties in implementation caused by the demands of a single entity.

[0016] Improve the interpretability and operability of engineering constraints and lower the threshold for technology implementation. Traditional multi-objective optimization often relies on abstract weight allocation, making it difficult for engineers to intuitively understand the decision-making logic behind the weights. This method, however, uses a B2C framework to transform the game equilibrium solution into specific, quantifiable engineering parameters (such as the upper limit of tunnel proportions and a safety margin of 0.5–1.5 meters of soil cover). Each parameter corresponds to a clear benefit and engineering significance (e.g., the "curvature penalty coefficient" is directly related to the design institute's safety requirements and the investor's cost control). Engineers do not need to understand complex game theory to directly apply the parameters to linear selection and optimization, solving the problem of "theoretical models being disconnected from engineering practice" and significantly lowering the technical application threshold.

[0017] Improve the accuracy of linear prediction and attribute classification to ensure the engineering adaptability of the solution. This method addresses the challenges of sparse and imbalanced railway alignment samples. It employs an improved UNet dual-task prediction architecture (simultaneously outputting alignment location and roadbed / bridge / tunnel attributes), combining positive sample priority sampling (initially sampling all positive samples to ensure alignment visibility, and later introducing difficult negative samples to improve discriminative power), a composite loss function (BCEWithLogits loss to address class imbalance, and SoftDice loss to optimize accuracy in sparse regions), and sliding window fusion inference (Hanning weights to eliminate edge prediction bias). This significantly improves the accuracy of alignment location and attribute classification. In actual training and validation, the alignment intersection-over-union ratio (route_iou) consistently reaches above 0.6, and attribute classification accurately matches different terrains and cost constraints, avoiding the engineering rework risks caused by "alignment misclassification" and "attribute misjudgment" in traditional models.

[0018] Improve route selection and design efficiency and shorten the early stage of the project. This method constructs an end-to-end automated workflow encompassing "data preparation - model training - inference decoding - scheme output": multi-source geographic data (DEM, DWG) can be automatically standardized, cropped, and quality-checked (generating QA reports); the model supports phased training and breakpoint continuation training (reducing repetitive training time); during the inference phase, sliding window fusion and constraint decoding automatically output constructable centerline schemes with mileage markers, geographic coordinates, and attributes. Compared to traditional manual route selection (requiring repeated coordination of multiple parties' demands and manual adjustment of alignment), this method significantly reduces manual intervention, shortening the route selection scheme generation cycle by more than 30%, while avoiding subjective biases in human decision-making, improving design efficiency and scheme consistency.

[0019] Enhances scenario adaptability and generalization capabilities to adapt to complex routing environments. This method integrates multi-source data (compatible with 5-30 meter resolution DEM terrain data and various DWG feature data), enhances samples (center jitter within 32 pixels to avoid overfitting), and uses segment-by-segment independent data partitioning (to avoid sliding window leakage), enabling the model to adapt to different terrains (mountains, plains) and different feature distributions (densely built areas, farmland, rivers) for route selection scenarios. Simultaneously, engineering constraint parameters (such as tunnel ratio upper limit and soil cover margin) can be dynamically adjusted through a game theory model, automatically optimizing constraint thresholds for special scenarios such as "high geological risk areas" and "high subsidy areas." This solves the problem of insufficient generalization in traditional models that rely on "single parameters to adapt to all scenarios," and can stably output qualified solutions in cross-regional railway and complex terrain railway route selection.

[0020] Ensure the reliability of the solution and the traceability of the project, and connect with subsequent design processes. This method uses QA reports to screen qualified data during the data preparation phase (eliminating cases with excessive coordinate system offset and high NoData ratios). During model training, objective indicators such as AUPRC and route_diou are used to monitor accuracy. The output centerline scheme maintains the same coordinate system as the DEM data and includes complete mileage markers, geographic coordinates, and roadbed / bridge / tunnel attributes. It can be directly exported to GeoTIFF (visualization) or CSV (engineering calculation) formats, seamlessly connecting to subsequent railway detailed design, engineering quantity calculation, and construction planning stages. Furthermore, the constraint parameters of the scheme can be traced back to specific game equilibrium solutions, facilitating backtesting of the scheme's rationality by multiple parties and improving the reliability of engineering decisions. Attached Figure Description

[0021] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a schematic diagram of the railway alignment constraint-aware deep learning generation method based on the B2C framework involved in this invention.

[0023] Figure 2 This is a schematic diagram of the training and validation loss curves involved in this invention.

[0024] Figure 3 This is a schematic diagram comparing the reference line shape and the model-generated line shape involved in this invention. Detailed Implementation

[0025] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the embodiments of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0026] The following disclosure provides many different implementations or examples for carrying out different structures of the embodiments of the present invention. To simplify the disclosure of the embodiments of the present invention, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the embodiments of the present invention. Furthermore, reference numerals and / or reference letters may be repeated in different examples of the embodiments of the present invention; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various implementations and / or arrangements discussed.

[0027] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0028] like Figure 1 As shown, this embodiment provides a railway alignment constraint-aware deep learning generation method based on a B2C framework, including the following steps: 1) Multi-agent Bayesian game modeling: The participants are identified as investors, design institutes, and local governments. The type space of the participants includes complex and simple geological types for design institutes, and subsidized and unsubsidized types for local governments. The prior probabilities of the type space are obtained by statistically analyzing the frequency of different geological types of design institutes and the implementation frequency of different subsidy policies of local governments in historical railway route selection projects. The strategies of investors are defined as tunnel proportions, the strategies of design institutes are risk thresholds corresponding to different geological types, and the strategies of local governments are compensation standards corresponding to different subsidy types. The expected utility of each participant is calculated, where the expected utility of investors is determined based on basic costs, tunnel construction costs, bridge construction costs, and land compensation costs; the expected utility of design institutes is determined based on geological risk costs and engineering risk costs; and the expected utility of local governments is determined based on social impact costs, social benefits, and land compensation expenditures. The iterative optimal response algorithm is used to solve the Bayesian Nash equilibrium to obtain the optimal strategies of each participant. 2) B2C Constraint Compilation: Construct a compilation function to map the optimal Bayesian Nash equilibrium strategy obtained in step 1) to a set of engineering constraint parameters. The set of engineering constraint parameters includes the upper limit of tunnel proportion, the safety margin of soil cover, the weight of ground cost, the curvature penalty coefficient, the longitudinal slope penalty coefficient, and the structural penalty coefficient. The compilation function calculates each engineering constraint parameter by performing minimum-maximum normalization on the optimal strategy and combining the aggregation weight, the unit conversion coefficient, and the margin scaling coefficient. 3) Multi-source geographic data preparation: Acquire multi-source geographic data, including DEM topographic data with a resolution of 5-30 meters within the railway alignment area, and DWG engineering data containing horizontal and vertical profile data, bridge and tunnel stationing data, and feature data; Standardize the multi-source geographic data: Unify all data to the same coordinate system, rasterize the feature data into three categories (corresponding to non-feature areas, general feature areas, and important feature areas), and crop them into 1000×1000 pixel case packages with 10-kilometer segments; Generate a QA report, which includes coordinate system offset, NoData value distribution, and data integrity information; Divide the training set, validation set, and test set according to the "segment-by-segment independence" principle to avoid data leakage during sliding window inference; 4) Deep Learning Model Training: An improved UNet is used as the training model. The improved UNet is optimized based on the standard UNet to adapt to the requirements of multi-source geographic data input and route-attribute dual-task prediction. The training set data prepared in step 3) is used as the model input. The input data is a 3-channel feature map (including normalized DEM data, rasterized no-passage area data, and rasterized cost area data). The route prediction head of the model outputs a linear spatial probability distribution map of the 0-1 interval, and the attribute prediction head outputs the category probability maps of roadbed, bridge, and tunnel. During the training process, a positive sample priority sampling strategy and a sample augmentation strategy are adopted, and a composite loss function is used to optimize the model. 5) Inference and Constraint Decoding: Sliding window inference is applied to the standardized multi-source geographic data of the validation or test set in step 3). The sliding window size is 512×512 pixels and the step size is 384 pixels. The prediction results of the overlapping area of ​​the sliding window are fused by Hanning weight to obtain the linear spatial probability distribution map and attribute probability map. The set of engineering constraint parameters obtained in step 2) is loaded, and constraint decoding is performed on the probability map: Candidate linear shapes that exceed the limit are filtered based on the upper limit of the tunnel ratio, and tunnel segments that do not meet the requirements are eliminated by combining the soil cover safety margin. The attribute probability is weighted according to the ground cost weight to prioritize the linear attributes with low land cost. 6) Output railway alignment centerline scheme: Output the decoded result from step 5) as a constructable railway centerline scheme in the same coordinate system as the DEM data from step 3); the centerline scheme includes mileage station, geographic coordinates, and roadbed / bridge / tunnel attributes; this scheme is used in subsequent railway engineering design processes.

[0029] In some embodiments, step 1) of "solving the Bayesian Nash equilibrium using an iterative optimal response algorithm" specifically includes the following steps: 1.1) Initialization: Set the initial tunnel ratio for investors, the initial risk thresholds for different geological types for design institutes, and the initial compensation standards for different subsidy types for local governments, and set the iteration counter to 0; 1.2) Iterative Update: Repeat the following operations until convergence: 1.2.1) Investor's Update Strategy: Based on the design institute's current risk threshold and the local government's current compensation standards, determine the tunnel proportion that maximizes the investor's expected utility; 1.2.2) Design Institute Update Strategy: For each geological type, based on the current tunnel ratio of the investor, solve for the risk threshold that maximizes the expected utility of the design institute; 1.2.3) Local government update strategy: For each type of subsidy, based on the current tunnel ratio of the investor, solve for the compensation standard that maximizes the expected utility of the local government; 1.3) Convergence test: Calculate the relative change in the tunnel ratio after the update compared to before the update, the relative change in the risk threshold of each geological type, and the relative change in the compensation standard of each subsidy type; if the maximum relative change is less than the tolerance of one ten-thousandth, the algorithm terminates and obtains the optimal strategy of Bayesian Nash equilibrium; otherwise, increment the iteration counter by 1 and return to step 1.2).

[0030] In some embodiments, step 2) of "mapping the Bayesian Nash equilibrium optimal policy into a set of engineering constraint parameters through a compilation function" specifically includes the following steps: 2.1) Calculate the upper limit of the tunnel ratio: The upper limit of the tunnel ratio is equal to the first aggregation weight multiplied by the minimum-maximum normalized optimal tunnel ratio of the investor, plus the second aggregation weight multiplied by the average of the minimum-maximum normalized optimal compensation standards of the local government, where the sum of the first aggregation weight and the second aggregation weight is 1; 2.2) Calculate the safety margin of the backfill: The safety margin of the backfill is equal to the margin scaling factor multiplied by the average value of the minimum-maximum normalized optimal risk threshold of the design institute, wherein the average value of the minimum-maximum normalized optimal risk threshold of the design institute is mapped to a range of 0.5 to 1.5 meters; 2.3) Calculate the ground cost weight: The ground cost weight is equal to the unit conversion factor multiplied by the average value of the local government's optimal compensation standard. The unit conversion factor is determined by regressing the correspondence between the comprehensive land cost in historical railway route selection projects and the monetary unit in the game model. 2.4) Calculate the curvature penalty coefficient: The curvature penalty coefficient is equal to the third aggregation weight multiplied by (1 minus the average value of the design institute's optimal risk threshold after minimum-maximum normalization), plus the fourth aggregation weight multiplied by (1 minus the investor's optimal tunnel ratio after minimum-maximum normalization), where the sum of the third aggregation weight and the fourth aggregation weight is 1.

[0031] In some embodiments, step 3) of "preparing multi-source geographic data" specifically includes the following steps: 4.1) DEM topographic data processing: Obtain DEM topographic data within the railway alignment area, and fill the NoData values ​​in the data using neighborhood interpolation to ensure data continuity; 4.2) DWG Engineering Data Processing: Obtain DWG engineering data within the railway alignment area, extract horizontal and vertical profile data, bridge and tunnel station data, and ground feature data using CAD data parsing tools, and rasterize the ground feature data according to type, with non-ground feature areas classified as Category 1, general ground feature areas such as ordinary farmland classified as Category 2, and important ground feature areas such as buildings and rivers classified as Category 3. 4.3) Standardization and cropping: The processed DEM terrain data and DWG engineering data are unified to the same coordinate system using coordinate transformation tools, and cropped into 1000×1000 pixel 10-kilometer segment case packages using tiling tools. Each case package is associated with the corresponding regional location information. 4.4) Quality Inspection and Data Division: Select qualified case packages based on QA reports, and remove case packages with coordinate system offset exceeding tolerance or NoData value exceeding 5%; divide the training set, validation set, and test set in a 7:2:1 ratio to ensure that the datasets do not overlap to avoid sliding window leakage.

[0032] In some embodiments, the architecture of the "improved UNet" in step 4) specifically includes the following parts: 5.1) Encoder side: Set up 4 convolutional modules, each of which contains 3×3 convolution operation, BatchNorm normalization operation and ReLU activation function in sequence; downsampling is achieved by 3×3 convolution with stride of 2, and the number of feature map channels doubles with each downsampling, and multi-scale features of DEM terrain data, restricted area data and cost area data are extracted step by step; 5.2) Decoder side: Each layer is upsampled through 2×2 transposed convolution, which halves the number of channels in the feature map after upsampling; the upsampled feature map is then concatenated with the feature map of the corresponding layer in the encoder through skip connection and fused with the channel alignment, finally restoring the same pixel-level prediction resolution as the input data; 5.3) Dual-head output layer: The route prediction head is a 1-channel output, which outputs a linear spatial probability distribution map of the 0-1 interval through the Sigmoid activation function, used to characterize the probability that each pixel is the railway centerline; the attribute prediction head is a 3-channel output, which outputs the category probability map of roadbed, bridge and tunnel through the Softmax activation function, and invalid prediction results in non-linear areas are masked by setting ignore_index=255.

[0033] In some embodiments, step 4) specifically includes the following: "positive sample priority sampling strategy and sample enhancement strategy": 6.1) Positive Sample Priority Sampling: The proportion of positive samples is controlled by parameters. The positive samples are the labeled railway line area data. In the initial training phase (the first 20 training cycles), the positive sample clipping ratio is set to 1.0, and all positive samples are sampled to ensure that the model can quickly learn the line features. In the later training phase (after 20 training cycles), difficult negative samples with a ratio of 0.1-0.3 are introduced. The difficult negative samples are non-target area data with a distance of less than 5 pixels from the line area, which improves the model's ability to distinguish between target and non-target areas. 6.2) Sample augmentation: A center-jitter sample augmentation method is adopted, which randomly shifts the center position of each sample horizontally and vertically within a 32-pixel range, while keeping the sample size unchanged. This avoids the model overfitting to a fixed linear position and enhances the model's generalization ability in different location scenarios.

[0034] In some embodiments, the "composite loss function" in step 4) specifically includes the following: 7.1) Route prediction loss: The BCEWithLogits loss plus 0.5 times the SoftDice loss is adopted. The BCEWithLogits loss addresses the problem of imbalanced linear sample classes by setting positive and negative sample weights, while the SoftDice loss optimizes the prediction accuracy of sparse linear regions by calculating the intersection-union ratio of the predicted linear and the real linear, thus alleviating the gradient vanishing problem caused by the scarcity of positive samples. 7.2) Attribute prediction loss: Masked cross-entropy loss is used, and the loss is calculated only within the "route neighborhood". The "route neighborhood" is obtained by performing 3×3 pixel dilation on the true linear label or linear area with a prediction probability greater than 0.2. At the same time, the ground cost weight obtained in step 2) is incorporated. The prediction loss of bridge attribute is multiplied by (1 plus ground cost weight), the prediction loss of tunnel attribute is multiplied by (1 plus 2 times ground cost weight), and the prediction loss of roadbed attribute is multiplied by 1, so that the attribute prediction result conforms to the cost constraint in multi-agent game.

[0035] In some embodiments, step 4) "deep learning model training" specifically refers to phased training, including the following steps: 8.1) First stage training: Set the attribute loss weight coefficient to 0 and train only the improved UNet route prediction head; use the training set data prepared in step 3) as input, adopt the Adam optimizer, the initial learning rate is 1 / 100,000, the positive and negative sample weights are set to 12, and optimize the route prediction loss; validate once every 2 training cycles until the linear intersection-union ratio of the validation set is not less than 0.6 to ensure that the linear position prediction is stable; 8.2) Second stage training: Set the attribute loss weight coefficient to 0.5-0.7, and jointly train the route prediction head and attribute prediction head; keep the Adam optimizer and initial learning rate unchanged, set the positive and negative sample weights to 10, and simultaneously optimize the route prediction loss and attribute prediction loss; enable AMP mixed precision training to reduce memory usage and accelerate convergence, and combine gradient clipping (limit the maximum gradient value to 5.0) to prevent gradient explosion during training; 8.3) Resume training from breakpoint: The model parameter file is saved every 5 training cycles during the training process (the file is named last.pt); when training is interrupted, the most recently saved model parameter file is loaded and training continues from the interrupted training cycle, thereby improving training efficiency.

[0036] In some embodiments, step 5) "sliding window reasoning and constraint decoding" specifically includes the following steps: 9.1) Sliding Window Inference: For the standardized multi-source geographic data of the validation or test set in step 3), a sliding window of size 512×512 pixels is used to traverse the entire data area with a step size of 384 pixels; the prediction results of the overlapping area (overlapping part is 128 pixels) of adjacent sliding windows are weighted and fused using the Hanning weight function (the weight of the window edge is 0.1 and the weight of the window center is 1.0) to eliminate the prediction bias of the sliding window edge; finally, a linear spatial probability distribution map (saved as route_prob.npy file) and an attribute probability map (saved as attr_prob.npy file) are output. 9.2) Constraint Decoding: Load the set of engineering constraint parameters obtained in step 2) and perform decoding in the following order: 9.2.1) Tunnel Proportion Filtering: Calculate the proportion of tunnel segment length for each candidate alignment and remove candidate alignments whose proportion exceeds the upper limit of tunnel proportion; 9.2.2) Soil cover safety inspection: Calculate the tunnel section burial depth based on the DEM topographic data processed in step 3), remove the predicted tunnel section burial depth that is less than the soil cover safety margin, and supplement the roadbed or bridge attributes. 9.2.3) Attribute Cost Weighting: For each pixel in the attribute probability map, the probabilities of roadbed, bridge, and tunnel are multiplied by 1, (1 plus ground cost weight), and (1 plus ground cost weight plus curvature penalty coefficient), respectively. The attribute with the highest weighted probability is selected as the final attribute of the pixel.

[0037] In some embodiments, step 6) of “outputting the railway alignment centerline scheme” specifically includes the following steps: The alignment spatial probability distribution map and attribute probability map after constraint decoding in step 5) are used to obtain the pixel coordinates of the railway centerline through probability threshold filtering (selecting pixels with a probability greater than 0.7) and skeleton extraction algorithms; the pixel coordinates are converted into geographic coordinates in the same coordinate system as the DEM data in step 3); mileage stations are assigned to the geographic coordinates in mileage order, and the corresponding roadbed / bridge / tunnel attributes are associated; the above information is output as a constructable railway centerline scheme in GeoTIFF format (for visualization) or CSV format (for engineering calculation); the scheme is directly used in the subsequent detailed design, quantity calculation, and construction planning stages of railway engineering.

[0038] The technical concept of this invention is as follows: I. Overview of the Plan This invention proposes a Bayes-to-Constraints (B2C) framework, which for the first time systematically "compiles" the multi-agent Bayesian Nash Equilibrium (BNE) into a set of interpretable and operable engineering control parameters (Λ), thereby achieving synergistic optimization of project economic, social and environmental goals at the source of decision-making.

[0039] The end-to-end process of this invention is organized in the order of "game theory → constraint → learning → fitting → evaluation / case study": Stage A—(Game Theory Modeling): Define the type space and prior (p,q) of the three main parties (investors / design institutes / local governments), give the cost / risk / subsidy benefit functions, solve BNE, and obtain (α*,β*(·),γ*(·)).

[0040] Stage B—B2C (Game Theory → Engineering Constraints): The equilibrium strategy is compiled into a set of engineering knobs Λ={r_max,minDepth,λ_land,λ_κ}, where r_max and minDepth are used as hard constraints to participate in the fitting, and {λ} is used as a soft weight to adjust the preference.

[0041] Stage C—(Data and Case Preparation): Standardize the DEM (5–30m), full-length alignment (horizontal / vertical profile + bridge / tunnel stationing), and DWG features (rasterized to {0 / 1 / 2}) into 10km segment case packages; unify CRS, automatically crop to approximately 1000×1000, and output QA reports (CRS offset, NoData, etc.); divide the train / val / test by segment to avoid sliding window leakage.

[0042] Stage D—(Neural Network Learning): Generate Gaussian corridor labels using artificial linear shapes, and train pixel-level center lines Pr(x,y) on the DEM+ ground feature raster.

[0043] II. Game Theory Modeling 2.1 Representation and Equilibrium Definition of Bayesian Games Participants: Investor P, Design Institute D, Local Government L. Type of Space (Complex / Simple quality) (With / without subsidy), prior is .Strategy: (Tunnel ratio) (Risk threshold) (Compensation Standard). The expected utility is used to characterize the benefits to all three parties, and BNE is defined accordingly.

[0044] Expected utilities (expected returns, which can generally be quantified): ; ; ; BNE (Equilibrium Definition): ; 2.2 Calculation of Expected Utility The core of solving the BNE lies in calculating the expected utility of each agent, a process that integrates the uncertainties of other agent types.

[0045] For the investor (P), the type of designer needs to be determined. With government type Simultaneously, calculate the expected value: ; For the designer (D), only their own type is required. Seeking expectations—because its benefits depend on government action. However, it does not directly depend on the type of government. (Unless a correlation exists): ; For local governments (L), it is necessary to determine their own type. Expectation: ; These expected utility calculations form the basis for each agent to make optimal decisions.

[0046] 2.3 Iterative Algorithm for Solving BNE Considering the complexity of the policy space, an iterative optimal response algorithm is used to numerically solve the BNE. The algorithm steps are as follows: 1. Initialization: Set initial strategies for all subjects and types (e.g., ... , , ), and set the iteration counter. .

[0047] 2. Iterative Update: Repeat the following steps until convergence—for each agent, update its own policy based on the current policies of other agents to achieve the optimal response: Investor Update: ; Designer updates (for each type) ): ; Government updates (for each type) ): ; 3. Convergence Check: Calculates the relative change in the strategy, i.e. and to and Perform the same calculation. If the maximum change is below the tolerance level... (like If the condition is met, the algorithm terminates; otherwise, set... Then return to step 2.

[0048] 2.4 Convergence and Stability Under the standard assumptions, the existence of BNE has been theoretically proven. Empirical results show that the iterative algorithm converges stably to the solution in all simulations.

[0049] The detailed derivation of the Bayesian Nash equilibrium is as follows: 1. Expansion and Variable Definition of the Expected Return Function 1.1 Participant Types and Belief Distribution Design Institute Type: Geological Conditions ,in Indicates complex geology (probability of landslide) =0.01), Simply put, it represents the probability of collapse. =0.05). The investor's belief in the design institute type is... , (Common knowledge).

[0050] Local government type: policy subsidies ,in This indicates that there are subsidies (reduced compensation costs). ), This indicates no subsidies are offered. The investor's belief in the local government type is... , (Common knowledge).

[0051] Decision variable: Proportion of tunnels chosen by investors The design institute selects the risk threshold. Local governments choose compensation standards .

[0052] 1.2 Mathematical Expansion of the Expected Return Function (1) Investors' expected returns The investor's returns depend on the type of design institute. and local government type To find the joint distribution, we need to calculate the expectation for all combinations of types: ; in, Substituting, we get: ; Simplified expected terms: (Expected risk threshold of the design institute); (Expectations regarding local government compensation standards).

[0053] (2) The revenue function of the design institute and the local government Design Institute (Type Dependency): Where (k=0.5) is the geological risk amplification factor ( When the value is 1, the risk loss increases by 50%.

[0054] Local government (type dependency): ( When S=1, the subsidy offsets 30% of the cost. =0.3).

[0055] 2. Inverse solution of Bayesian Nash equilibrium 2.1 The optimal strategy for design institutes

[0056] The design institute's goal is to minimize risk losses; its benefit function is... Follow Monotonically decreasing, therefore for type Choose the least feasible option : ; Proof: For Differentiation Therefore The smaller, The larger it is, the proof is complete.

[0057] 2.2 Optimal Strategy of Local Government

[0058] Local governments aim to maximize compensation benefits, and their strategies depend on the type of subsidy (S): type (With subsidies): Compensation costs are reduced, and the highest compensation standard can be selected. ; type (No subsidy): The cost of compensation is high, so choose the lowest compensation standard. .

[0059] Mathematical verification: Differentiation Therefore The larger, The larger it is, the proof is complete.

[0060] 2.3 Optimal Strategy for Investors

[0061] Investors maximize expected returns choose ,right about Differentiate: ; Setting the derivative to zero, we can solve for the optimal tunnel ratio. Critical condition: ; Equilibrium strategy: like ,but (Entire tunnel); like ,but (All ground); If the equation holds true, then The strategy is a hybrid approach (partial tunnel + partial surface).

[0062] 3. Proof of the existence of Bayesian Nash equilibrium Definition: Strategy Combination A Bayesian Nash equilibrium is formed if, for all participants, given the strategies and type beliefs of other participants, their strategy is the optimal response.

[0063] prove: 1. Design Institute: (This likely refers to a specific institution or organization) , make Maximum (monotonicity verified); 2. Local government: Regarding , make Maximum (monotonicity verified); 3. Investors: satisfy This is the condition for maximizing expected returns.

[0064] In summary, the three strategies satisfy the optimal response of each other, and the equilibrium exists and is unique.

[0065] III. Game Theory → Engineering Constraints Bayesian Nash Equilibrium (BNE) Solution The optimization strategy encompasses all key components. However, this solution cannot be directly used as engineering parameters. Therefore, a compilation function is defined. This function can demap the BNE into a set of interpretable engineering adjustment parameters. The purpose of this function is to transform the economic and strategic preferences reflected in the equilibrium into specific design constraints.

[0066] For example, the upper limit of tunnel ratio The determination of this factor stems both from the investor's optimal cost-benefit analysis and from the local government's assessment of environmental and social costs. ; in This represents min-max normalization. This is to aggregate weights. This design ensures that the final engineering parameters reflect the equilibrium preferences of key stakeholders.

[0067] Similarly, other engineering adjustment parameters are defined as follows: Ground cost weighting: (Economic implications); Safety margin: (Risk implications); Curvature penalty coefficient: (The dual meaning of safety and cost); in: Unit conversion factor. It maps the "currency unit" in the game to the "cost unit" in the path optimization model. This factor can be determined by regressing historical data, so that the ground cost in the path cost model matches the comprehensive land cost in the actual project. Margin scaling factor. It maps the normalized risk preference to a specific engineering margin (e.g., mapping [0,1] to a cover margin of [0.5m,1.5m]). , Aggregate weights, satisfying .

[0068] This compilation process is the core of the B2C framework, bridging the gap between game theory equilibrium and practically executable engineering design.

[0069] IV. Artificial Linear Machine Learning with Embedded Game Constraints 4.1 Machine Learning Model Architecture Design An improved UNet is adopted as the core model for artificial linear machine learning, which is suitable for multi-source geographic data input and linear / attribute dual-task prediction requirements. The specific architecture is as follows: Backbone network: The encoder has 4 convolutional modules (including BatchNorm and ReLU activation), which use stride 2 convolutions to achieve downsampling and extract multi-scale features of DEM terrain, no-pass regions, and cost regions; the decoder uses transposed convolutions to upsample and align and fuse with the corresponding layer features (skip connection) of the encoder to restore pixel-level prediction resolution.

[0070] Dual-head output layer: Route prediction head (route_head): 1-channel output, activated by Sigmoid to output the centerline probability map of the [0,1] interval, capturing the linear spatial position; Attribute prediction head (attr_head): 3-channel output, activated by Softmax to output the class probabilities of roadbed (0), bridge (1) and tunnel (2), combined with ignore_index=255 to mask invalid areas, to achieve linear attribute classification.

[0071] 4.2 Training Strategy and Loss Function Design To address the issues of linear sample sparsity and class imbalance, a robust training strategy and a composite loss function are designed to ensure model convergence and prediction accuracy. Sample processing strategy: Positive sample priority sampling: The pos_crop_frac parameter controls the proportion of positive samples (linear regions) to be cropped (recommended 0.6-1.0). Initially, 1.0 is used to sample all positive samples to ensure that the model "sees the line". Later, hard_neg_frac=0.1-0.3 are introduced as difficult negative samples (near-linear non-target regions) to improve the model's discrimination ability. Sample augmentation: Center jitter (within 32 pixels) is used to prevent the model from overfitting to a fixed linear position and to enhance generalization.

[0072] Composite loss function: Route prediction loss: BCEWithLogits loss + 0.5×SoftDice loss is adopted. BCEWithLogits addresses class imbalance, while SoftDice loss focuses on optimizing the intersection-union ratio in sparse linear regions to alleviate gradient vanishing caused by the scarcity of positive samples. Attribute prediction loss: Masked cross-entropy (CE) loss is adopted, and the loss is calculated only in the "route neighborhood" (GT or 3×3 inflated area with prediction probability > 0.2) to shield the interference of nonlinear areas. At the same time, the land cost weight λ_land compiled by B2C is incorporated to perform cost weighting on bridge / tunnel attribute prediction, which conforms to game constraints.

[0073] 4.3 Model Training and Inference Process 4.3.1 Training Process and Optimized Configuration Based on standardized samples (*.npz format) generated from multi-source geographic data, model training is conducted in stages to ensure a gradual improvement in the accuracy of linear and attribute predictions. Training data input: Sample input x is a 3-channel feature map ([DEM_norm,forbidden(0 / 1),costly(0 / 1)]), label y_route is a Gaussian smoothed linear probability map, y_attr is the attribute category label, and the sample size is adjusted to the UNet adaptation size by pad_multiple=32; Optimize parameter configuration: Use the Adam optimizer with an initial learning rate of lr=1e-4, and set pos_weight=8-16 for sparse positive samples; enable AMP mixed precision training to accelerate convergence, and combine gradient clipping (grad_clip=5.0) to prevent gradient explosion; support resume function, and continue training based on last.pt breakpoint to improve training efficiency. Phased training: Phase 1 -- w_attr=0, only train the route prediction head to ensure stable linear position prediction (route_iou≥0.6); Phase 2 -- w_attr=0.5-0.7, jointly train the dual-output head to simultaneously optimize attribute classification accuracy.

[0074] 4.3.2 Reasoning and Constraint Decoding Process In the model inference phase, a sliding window fusion strategy and game-constraint decoding are introduced to transform the probability output into a constructible centerline. The process is as follows: Sliding window inference: A 512×512 sliding window (step size 384) is applied to the input DEM feature map. The prediction results of overlapping regions are fused through Hanning weights, and route_prob.npy (linear probability map) and attr_prob.npy (attribute probability map) are output to avoid edge prediction bias. Game Theory Constraint Decoding: Load the constraint set Λ output by the B2C compiler and decode the probability graph: Candidate linear shapes exceeding the tunnel ratio upper limit are filtered based on r_max; By combining the minDepth constraint on the tunnel section burial depth, predictions that do not meet the risk threshold are eliminated; Based on the weighted attribute probability of λ_land, linear attributes with lower land costs are prioritized (e.g., roadbeds are prioritized over bridges). Output format: After decoding, the output is the constructable centerline (including station, x, y coordinates and attr attribute) in the same coordinate system as the DEM. It can be exported as GeoTIFF or CSV format to connect with subsequent engineering design processes.

[0075] 4.4 Model Evaluation and Optimization Design an evaluation index system adapted to linear prediction tasks, dynamically monitor the model training process, and iteratively optimize it: Core evaluation indicators: Linear position accuracy: The predicted line shape is spatially overlapped with the actual line shape by using route_iou (line shape intersection-over-union ratio) and route_diou (dilated intersection-over-union ratio, which better matches the characteristics of thin lines). Class imbalance robustness: AUPRC (Area Under the Precision-Recall Curve) is adopted, which is more suitable for sparse positive sample scenarios than ACC, avoiding the distortion of indicators caused by an excessive proportion of negative samples. Visual monitoring: Plot the Training Loss / Validation Loss curves in real time using plot_training.py, and support dynamic refresh in the --watch mode to promptly detect overfitting (training loss decreases but validation metrics stagnate) and adjust hyperparameters (such as reducing the learning rate or increasing the proportion of hard negative samples).

[0076] Training and validation loss curves are as follows Figure 2 As shown; a comparison chart of the reference line shape and the model-generated line shape is shown below. Figure 3 As shown, Figure 3 In this context, gt represents the reference line shape, and pred_opt represents the line shape generated by the model.

[0077] Variable meaning explanation table:

[0078] The embodiments described above are for illustrative purposes only and are not intended to limit the invention. Therefore, any changes in numerical values ​​or substitutions of equivalent elements should still fall within the scope of this invention.

[0079] The above detailed description will enable those skilled in the art to understand that the present invention can indeed achieve the aforementioned objectives and has complied with the provisions of the Patent Law.

[0080] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the invention. The above descriptions are merely preferred embodiments of the invention and are not intended to limit the invention. It should be noted that any modifications, equivalent substitutions, and improvements made within the spirit and principles of the invention should be included within the scope of protection of the invention.

[0081] It should be noted that the above description of the process is for illustrative purposes only and does not limit the scope of this specification. Those skilled in the art can make various modifications and changes to the process under the guidance of this specification. However, these modifications and changes remain within the scope of this specification.

[0082] The basic concepts have been described above. Obviously, for those skilled in the art who have read this application, the above disclosure is merely illustrative and does not constitute a limitation of this application. Although not explicitly stated herein, those skilled in the art may make various modifications, improvements, and corrections to this application. Such modifications, improvements, and corrections are suggested in this application, and therefore, such modifications, improvements, and corrections still fall within the spirit and scope of the exemplary embodiments of this application.

[0083] Furthermore, this application uses specific terms to describe its embodiments. For example, "an embodiment," "one embodiment," and / or "some embodiments" refer to a particular feature, structure, or characteristic related to at least one embodiment of this application. Therefore, it should be emphasized and noted that "an embodiment," "one embodiment," or "an alternative embodiment" mentioned twice or more in different positions in this specification do not necessarily refer to the same embodiment. In addition, certain features, structures, or characteristics in one or more embodiments of this application can be appropriately combined.

[0084] Furthermore, those skilled in the art will understand that aspects of this application can be described and illustrated through several patentable types or situations, including any new and useful combination of processes, machines, products, or substances, or any new and useful improvements thereof. Therefore, aspects of this application can be implemented entirely in hardware, entirely in software (including firmware, resident software, microcode, etc.), or a combination of hardware and software. All of the above hardware or software can be referred to as a “unit,” “module,” or “system.” Furthermore, aspects of this application can take the form of a computer program product embodied in one or more computer-readable media, wherein computer-readable program code is contained therein.

[0085] The computer program code required for the operation of each part of this application can be written in any one or more programming languages, including object-oriented programming languages ​​such as Java, Scala, Smalltalk, Eiffel, JADE, Emerald, C++, C#, VB.NET, and Python; general programming languages ​​such as C; Visual Basic, Fortran2103, Perl, COBOL2102, PHP, and ABAP; dynamic programming languages ​​such as Python, Ruby, and Groovy; or other programming languages. This program code can run entirely on the user's computer, or as a standalone software package on the user's computer, or partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the latter case, the remote computer can be connected to the user's computer via any network, such as a local area network (LAN) or wide area network (WAN), or connected to an external computer (e.g., via the Internet), or in a cloud computing environment, or used as a service such as Software as a Service (SaaS).

[0086] Furthermore, unless expressly stated in the claims, the order of processing elements and sequences, the use of numbers and letters, or other names described in this application are not intended to limit the order of the processes and methods of this application. Although some currently considered useful embodiments of the invention have been discussed in the foregoing disclosure by way of various examples, it should be understood that such details are for illustrative purposes only, and the appended claims are not limited to the disclosed embodiments; rather, the claims are intended to cover all modifications and equivalent combinations that conform to the substance and scope of the embodiments of this application. For example, although the implementation of the various components described above can be embodied in a hardware device, it can also be implemented as a purely software solution, such as an installation on an existing server or mobile device.

[0087] Similarly, it should be noted that, in order to simplify the description of the present application and thus aid in the understanding of one or more embodiments of the invention, the foregoing description of the embodiments of the present application sometimes combines multiple features into a single embodiment, drawing, or description thereof. However, this approach of the present application should not be construed as reflecting an intention that the claimed subject matter requires more features than expressly recited in each claim. Rather, the subject of the invention should possess fewer features than in any single embodiment described above.

Claims

1. A deep learning-based method for generating railway alignment constraints based on a B2C framework, characterized in that, include: 1) Multi-agent Bayesian game modeling: The participants are identified as investors, design institutes, and local governments. The expected utility of each participant is calculated, and the Bayesian Nash equilibrium is solved using the iterative optimal response algorithm to obtain the optimal strategy of each participant. 2) B2C Constraint Compilation: Construct a compilation function to map the optimal strategy to a set of engineering constraint parameters, which includes the upper limit of tunnel proportion, the safety margin of soil cover, the weight of ground cost, the curvature penalty coefficient, the longitudinal slope penalty coefficient, and the structural penalty coefficient. 3) Multi-source geographic data preparation: Acquire multi-source geographic data and perform standardization processing to obtain training set, validation set, and test set; 4) Deep learning model training: An improved UNet is used as the training model. The training set is used as input. The model's route prediction head outputs a linear spatial probability distribution map, and the attribute prediction head outputs attribute probability maps of roadbed, bridge, and tunnel. During training, a positive sample priority sampling strategy and a sample augmentation strategy are adopted. A composite loss function is used to optimize the model to obtain the deep learning model. 5) Reasoning and Constraint Decoding: Using a deep learning model, sliding window reasoning is applied to the validation or test set. The prediction results of the overlapping area of ​​the sliding window are fused through Hanning weights to obtain the linear spatial probability distribution map and attribute probability map. Constraint decoding is performed on the linear spatial probability distribution map and attribute probability map in combination with the set of engineering constraint parameters: Candidate linear shapes that exceed the limit are filtered based on the upper limit of the tunnel ratio, and tunnel segments that do not meet the requirements are eliminated based on the soil cover safety margin. The attribute probability is weighted according to the ground cost weight to prioritize linear attributes with low land cost. 6) Output railway alignment centerline scheme: Output the decoded results as a constructable railway centerline scheme in the same coordinate system as the multi-source geographic data, including mileage station, geographic coordinates and roadbed / bridge / tunnel attributes.

2. The railway alignment constraint-aware deep learning generation method based on a B2C framework according to claim 1, characterized in that, Step 1) uses the iterative optimal response algorithm to solve for the Bayesian Nash equilibrium, specifically including the following steps: 1.1) Initialization: Set the initial tunnel ratio for investors, the initial risk thresholds for different geological types for design institutes, and the initial compensation standards for different subsidy types for local governments, and set the iteration counter to 0; 1.2) Iterative Update: Repeat the following operations until convergence: 1.2.1) Investor's Update Strategy: Based on the design institute's current risk threshold and the local government's current compensation standards, determine the tunnel proportion that maximizes the investor's expected utility; 1.2.2) Design Institute Update Strategy: For each geological type, based on the current tunnel ratio of the investor, solve for the risk threshold that maximizes the expected utility of the design institute; 1.2.3) Local government update strategy: For each type of subsidy, based on the current tunnel ratio of the investor, solve for the compensation standard that maximizes the expected utility of the local government; 1.3) Convergence test: Calculate the relative change in the tunnel ratio after the update compared to before the update, the relative change in the risk threshold of each geological type, and the relative change in the compensation standard of each subsidy type; if the maximum relative change is less than the tolerance of one ten-thousandth, the algorithm terminates and obtains the optimal strategy of Bayesian Nash equilibrium; otherwise, increment the iteration counter by 1 and return to step 1.2).

3. The railway alignment constraint-aware deep learning generation method based on a B2C framework according to claim 1, characterized in that, Step 2) maps the optimal strategy to a set of engineering constraint parameters through a compilation function, specifically including the following steps: 2.1) Calculate the upper limit of the tunnel ratio: The upper limit of the tunnel ratio is equal to the first aggregation weight multiplied by the minimum-maximum normalized optimal tunnel ratio of the investor, plus the second aggregation weight multiplied by the average of the minimum-maximum normalized optimal compensation standards of the local government, where the sum of the first aggregation weight and the second aggregation weight is 1; 2.2) Calculate the safety margin of the backfill: The safety margin of the backfill is equal to the margin scaling factor multiplied by the average value of the optimal risk threshold of the design institute after minimum-maximum normalization; 2.3) Calculate the ground cost weight: The ground cost weight is equal to the unit conversion factor multiplied by the average value of the local government's optimal compensation standard. The unit conversion factor is determined by regressing the correspondence between the comprehensive land cost in historical railway route selection projects and the monetary unit in the game model. 2.4) Calculate the curvature penalty coefficient: The curvature penalty coefficient is equal to the third aggregation weight multiplied by (1 minus the average value of the design institute's optimal risk threshold after minimum-maximum normalization), plus the fourth aggregation weight multiplied by (1 minus the investor's optimal tunnel ratio after minimum-maximum normalization), where the sum of the third aggregation weight and the fourth aggregation weight is 1.

4. The railway alignment constraint-aware deep learning generation method based on a B2C framework according to claim 1, characterized in that, Step 3) involves the preparation of multi-source geographic data, specifically including the following steps: 4.1) Digital Elevation Topographic Data Processing: Obtain digital elevation topographic data within the railway alignment area, and fill the NoData values ​​in the data using neighborhood interpolation to ensure data continuity; 4.2) DWG Engineering Data Processing: Obtain DWG engineering data within the railway alignment area, extract horizontal and vertical profile data, bridge and tunnel station data, and ground feature data using CAD data parsing tools, and rasterize the ground feature data according to type, with non-ground feature areas classified as Category 1, general ground feature areas such as ordinary farmland classified as Category 2, and important ground feature areas such as buildings and rivers classified as Category 3. 4.3) Standardization and cropping: The processed digital elevation topographic data and DWG engineering data are unified to the same coordinate system by using coordinate transformation tools. The data is then cropped into 1000×1000 pixel 10-kilometer segment case packages, and each case package is associated with the corresponding regional location information. 4.4) Quality Inspection and Data Division: Generate a quality assurance report based on multi-source geographic data, which includes coordinate system offset, NoData value distribution and data integrity information. Select qualified case packages based on the quality assurance report and remove case packages with out-of-tolerance coordinate system offset and NoData value ratio exceeding 5%. Divide the training set, validation set and test set into a 7:2:1 ratio to ensure that the datasets do not overlap to avoid sliding window leakage.

5. The railway alignment constraint-aware deep learning generation method based on a B2C framework according to claim 1, characterized in that, The improved UNet architecture in step 4) specifically includes the following parts: 5.1) Encoder side: Set up 4 convolutional modules, each of which contains 3×3 convolution operation, BatchNorm normalization operation and ReLU activation function in sequence; downsampling is achieved by 3×3 convolution with stride of 2, and the number of feature map channels doubles with each downsampling, and multi-scale features of digital elevation terrain data, restricted area data and cost area data are extracted step by step; 5.2) Decoder side: Each layer is upsampled through 2×2 transposed convolution, which halves the number of channels in the feature map after upsampling; the upsampled feature map is then merged with the feature map of the corresponding layer in the encoder through skipconnection for channel concatenation and feature alignment, finally restoring the same pixel-level prediction resolution as the input data; 5.3) Dual-head output layer: The route prediction head is a 1-channel output, which outputs a linear spatial probability distribution map of the 0-1 interval through the Sigmoid activation function, used to characterize the probability that each pixel is the railway centerline; the attribute prediction head is a 3-channel output, which outputs the category probability map of roadbed, bridge and tunnel through the Softmax activation function, and invalid prediction results in non-linear areas are masked by setting ignore_index=255.

6. The railway alignment constraint-aware deep learning generation method based on a B2C framework according to claim 1, characterized in that, Step 4) The positive sample priority sampling strategy and sample augmentation strategy specifically include the following: 6.1) Positive Sample Priority Sampling: The positive sample clipping ratio is controlled by parameters. The positive samples are the labeled railway alignment area data. In the initial training phase, the positive sample clipping ratio is set to 1.0, and all positive samples are sampled to ensure that the model can quickly learn the alignment features. In the later training phase, a number of difficult negative samples with a ratio of 0.1-0.3 are introduced. These difficult negative samples are non-target region data that are less than 5 pixels away from the linear region, which improves the model's ability to distinguish between target and non-target regions. 6.2) Sample augmentation: A center-jitter sample augmentation method is adopted, which randomly shifts the center position of each sample horizontally and vertically within a 32-pixel range, while keeping the sample size unchanged. This avoids the model overfitting to a fixed linear position and enhances the model's generalization ability in different location scenarios.

7. The railway alignment constraint-aware deep learning generation method based on a B2C framework according to claim 1, characterized in that, Step 4) The composite loss function specifically includes the following: 7.1) Route prediction loss: The BCEWithLogits loss plus 0.5 times the SoftDice loss is adopted. The BCEWithLogits loss addresses the problem of imbalanced linear sample classes by setting positive and negative sample weights, while the SoftDice loss optimizes the prediction accuracy of sparse linear regions by calculating the intersection-union ratio of the predicted linear and the real linear, thus alleviating the gradient vanishing problem caused by the scarcity of positive samples. 7.2) Attribute prediction loss: Masked cross-entropy loss is used, and the loss is calculated only within the route neighborhood. The route neighborhood is obtained by performing 3×3 pixel dilation on the true linear label or linear region with a prediction probability greater than 0.

2. At the same time, the ground cost weight obtained in step 2) is incorporated. The prediction loss of bridge attribute is multiplied by (1 plus ground cost weight), the prediction loss of tunnel attribute is multiplied by (1 plus 2 times ground cost weight), and the prediction loss of roadbed attribute is multiplied by 1, so that the attribute prediction result conforms to the cost constraint in multi-agent game.

8. The railway alignment constraint-aware deep learning generation method based on a B2C framework according to claim 1, characterized in that, Step 4) involves training the deep learning model in stages, including the following steps: 8.1) First stage training: Set the attribute loss weight coefficient to 0 and train only the improved UNet route prediction head; use the training set data prepared in step 3) as input, adopt the Adam optimizer, the initial learning rate is 1 / 100,000, the positive and negative sample weights are set to 12, and optimize the route prediction loss; validate once every 2 training cycles until the linear intersection-union ratio of the validation set is not less than 0.6 to ensure that the linear position prediction is stable; 8.2) Second stage training: Set the attribute loss weight coefficient to 0.5-0.7, and jointly train the route prediction head and the attribute prediction head; keep the Adam optimizer and the initial learning rate unchanged, set the positive and negative sample weights to 10, and simultaneously optimize the route prediction loss and the attribute prediction loss; enable automatic mixed precision training to reduce memory usage and accelerate convergence, and combine gradient clipping to prevent gradient explosion during training. 8.3) Resume training from breakpoint: The model parameter file is saved every 5 training cycles during the training process; when training is interrupted, the most recently saved model parameter file is loaded and training continues from the interrupted training cycle, improving training efficiency.

9. The railway alignment constraint-aware deep learning generation method based on a B2C framework according to claim 1, characterized in that, Step 5) specifically includes the following steps for reasoning and constraint decoding: 9.1) Sliding window inference: For the standardized multi-source geographic data of the validation or test set in step 3), a sliding window of size 512×512 pixels is used to traverse the entire data area with a step size of 384 pixels; the prediction results of the overlapping areas of adjacent sliding windows are weighted and fused using the Hanning weight function to eliminate the prediction bias at the edge of the sliding window; finally, a linear spatial probability distribution map and an attribute probability map are output. 9.2) Constraint Decoding: Load the set of engineering constraint parameters obtained in step 2) and perform decoding in the following order: 9.2.1) Tunnel Proportion Filtering: Calculate the proportion of tunnel segment length for each candidate alignment and remove candidate alignments whose proportion exceeds the upper limit of tunnel proportion; 9.2.2) Soil cover safety inspection: Calculate the tunnel section burial depth based on the digital elevation topographic data processed in step 3), remove the predicted tunnel section burial depth that is less than the soil cover safety margin, and supplement the roadbed or bridge attributes. 9.2.3) Attribute Cost Weighting: For each pixel in the attribute probability map, the probabilities of roadbed, bridge, and tunnel are multiplied by 1, (1 plus ground cost weight), and (1 plus ground cost weight plus curvature penalty coefficient), respectively. The attribute with the highest weighted probability is selected as the final attribute of the pixel.

10. The railway alignment constraint-aware deep learning generation method based on a B2C framework according to claim 1, characterized in that, Step 6) outputs the railway alignment centerline scheme, which specifically includes the following steps: The alignment spatial probability distribution map and attribute probability map obtained from the constraint decoding in Step 5) are processed using probability threshold filtering and skeleton extraction algorithms to obtain the pixel coordinates of the railway centerline; the pixel coordinates are converted to geographic coordinates in the same coordinate system as the digital elevation data in Step 3); mileage stations are assigned to the geographic coordinates in mileage order, and the corresponding roadbed / bridge / tunnel attributes are associated; the above information is output as a constructable railway centerline scheme.

Images