Roadbed design method and device based on graph neural network, equipment and storage medium

Abstract

The application belongs to the technical field of traffic engineering, and discloses a subgrade design method, device and equipment based on a graph neural network and a storage medium; the method comprises the following steps: obtaining a subgrade detection parameter; predicting based on the subgrade detection parameter through a pre-established graph neural network to obtain a prediction weight sequence, wherein the graph neural network comprises a graph network and graph classification, and the graph network comprises a graph convolution layer and a preselected pooling layer constructed by a graph ontology; and determining a target subgrade design scheme from a preset design scheme according to the prediction weight sequence; the application combines the advantages of graph theory and neural networks, uses the deep reasoning capability of the neural network and the explicit logic of the graph ontology to provide an accurate and interpretable prediction algorithm for intelligent subgrade decision-making, realizes the topological constraint on the input characteristic parameters, improves the accuracy of the decision-making, enhances the interpretability of the decision-making, and avoids the problems of the non-transparent internal working mechanism of the deep learning neural network and the lack of interpretability of the decision-making process.

Description

Technical Field

[0001] This invention relates to the field of traffic engineering technology, and in particular to a roadbed design method, apparatus, equipment and storage medium based on graph neural networks. Background Technology

[0002] Decision tree methods make decisions based on a set of predefined rules and conditions, meaning their adaptability and flexibility are relatively poor. When faced with unseen or undefined new scenarios, the decision-making ability of decision trees may be limited because they cannot flexibly adapt to new situations. When decision rules become complex and diverse, traditional decision trees can become very large and difficult to manage. This can lead to the expansion of the tree structure, increasing the depth of the tree and the number of rules, thus increasing the complexity of maintenance and updates. It can also lead to overfitting, i.e., too many rules and branches that may not be suitable for generalizing to real-world data.

[0003] While deep learning neural networks can handle complex nonlinear relationships, their internal workings are often opaque, making the decision-making process lack interpretability. This means that although neural networks can provide highly accurate predictions and decisions, it can be difficult to understand why they made a particular decision in certain situations. This can be a challenge for some application areas, such as medical diagnosis and financial risk analysis, because transparency and interpretability are crucial to the rationality and credibility of decision-making. Summary of the Invention

[0004] The main objective of this invention is to provide a roadbed design method, apparatus, device, and storage medium based on graph neural networks, aiming to solve the technical problem that the internal working mechanism of deep learning neural networks in the prior art is not transparent and that the decision-making process using deep learning neural networks lacks interpretability.

[0005] To achieve the above objectives, this invention provides a roadbed design method based on graph neural networks, the method comprising the following steps:

[0006] Obtain roadbed detection parameters;

[0007] A prediction weight sequence is obtained by making predictions based on the roadbed detection parameters through a pre-established graph neural network. The graph neural network includes a graph network and a graph classification. The graph network includes a graph convolutional layer and a pre-selected pooling layer constructed from graph ontology.

[0008] The target roadbed design scheme is determined from the preset design schemes based on the predicted weight sequence.

[0009] Optionally, before obtaining the prediction weight sequence by making predictions based on the roadbed detection parameters using a pre-established graph neural network, the method further includes:

[0010] Obtain the influence factors and lower-level influence factors of the subgrade design scheme, and determine the expert relationship between the influence factors and the lower-level influence factors;

[0011] Construct a graph ontology for each impact factor based on the lower-level impact factors and the expert relationships;

[0012] An initial graph neural network is constructed based on the graph ontology, preselected pooling layer, preselected convolutional layer, and preselected fully connected layer;

[0013] The initial graph neural network is trained to obtain a graph neural network.

[0014] Optionally, constructing the initial graph neural network based on the graph ontology, pre-selected pooling layer, pre-selected convolutional layer, and pre-selected fully connected layer includes:

[0015] The adjacency matrix between each influencing factor is obtained based on the adjacency relationship between each influencing factor in the graph ontology.

[0016] The adjacency matrix is ​​embedded into a pre-selected convolutional layer to obtain a graph convolutional layer;

[0017] A graph network is constructed based on the graph convolutional layers and pre-selected pooling layers, and a graph classification is constructed based on the fully connected layers;

[0018] An initial graph neural network is obtained from the graph network and the graph classification.

[0019] Optionally, training the initial graph neural network to obtain a graph neural network includes:

[0020] Sample roadbed data and sample design schemes are obtained from a preset database, wherein the preset data includes multiple successfully verified roadbed design schemes and their corresponding roadbed data;

[0021] Based on the initial graph neural network, the sample roadbed data is predicted to obtain the predicted sample weights;

[0022] A prediction design scheme is obtained based on the predicted sample weights;

[0023] A loss function is constructed based on the sample design scheme and the prediction design scheme. The initial graph neural network is then optimized based on the loss function to obtain the graph neural network.

[0024] Optionally, the step of predicting based on the roadbed detection parameters using a pre-established graph neural network to obtain a prediction weight sequence includes:

[0025] The roadbed detection parameters are preprocessed to obtain the parameters to be predicted;

[0026] The parameters to be predicted are extracted using graph convolutional layers to obtain parameter-related features;

[0027] The parameter association features are reduced in dimensionality by a pre-selected pooling layer to obtain the dimensionality-reduced parameter association features.

[0028] The reduced parameter association features are classified using a fully connected layer for graph classification to obtain a predicted weight sequence.

[0029] Optionally, the preprocessing of the roadbed detection parameters to obtain the parameters to be predicted includes:

[0030] The roadbed detection parameters are classified to obtain continuous data and discrete data;

[0031] One-hot encoding is performed on the discrete data to obtain discrete data parameter encoding;

[0032] The continuous data is initially encoded to obtain integer data, and the integer data is then one-hot encoded to obtain the continuous data parameter encoding.

[0033] The parameters to be predicted are obtained based on the discrete data parameter encoding and the continuous data parameter encoding.

[0034] Optionally, determining the target roadbed design scheme from the preset design schemes based on the predicted weight sequence includes:

[0035] The target prediction weight with the largest weight is obtained from the predicted weight sequence;

[0036] The reference design scheme in the preset design scheme is determined based on the target prediction weight;

[0037] The reference design scheme is verified, and if the verification is successful, the reference design scheme is adopted as the target roadbed design scheme.

[0038] After determining the target roadbed design scheme from the preset design schemes based on the predicted weight sequence, the method further includes:

[0039] The preset database is updated based on the roadbed detection parameters and the target roadbed design scheme.

[0040] The graph neural network is trained based on the preset database to obtain the training results;

[0041] The graph neural network is updated based on the training results.

[0042] Furthermore, to achieve the above objectives, the present invention also proposes a roadbed design device based on graph neural networks, the roadbed design device based on graph neural networks comprising:

[0043] The parameter acquisition module is used to acquire roadbed detection parameters;

[0044] The model prediction module is used to make predictions based on the roadbed detection parameters through a pre-established graph neural network to obtain a prediction weight sequence. The graph neural network includes a graph network and a graph classification. The graph network includes a graph convolutional layer and a pre-selected pooling layer constructed from a graph ontology.

[0045] The scheme generation module is used to determine the target roadbed design scheme from the preset design schemes based on the predicted weight sequence.

[0046] Furthermore, to achieve the above objectives, the present invention also proposes a roadbed design device based on graph neural networks. The roadbed design device based on graph neural networks includes: a memory, a processor, and a roadbed design program based on graph neural networks stored in the memory and executable on the processor. The roadbed design program based on graph neural networks is configured to implement the steps of the roadbed design method based on graph neural networks as described above.

[0047] Furthermore, to achieve the above objectives, the present invention also proposes a storage medium storing a roadbed design program based on a graph neural network, wherein when the roadbed design program based on the graph neural network is executed by a processor, it implements the steps of the roadbed design method based on the graph neural network described above.

[0048] This invention combines the advantages of deep learning technology with graph theory and neural networks. It utilizes the deep reasoning capabilities of neural networks and the explicit logic of graph ontologies to provide accurate and interpretable prediction algorithms for intelligent roadbed decision-making. It implements topological constraints on input feature parameters, thereby constraining the abstract information that neural networks can learn. This not only improves the accuracy of decision-making but also enhances its interpretability. It avoids the problem that although deep learning neural networks can handle complex nonlinear relationships, their internal working mechanisms are usually opaque and their decision-making processes lack interpretability. Attached Figure Description

[0049] Figure 1 This is a schematic diagram of the structure of a roadbed design device based on a graph neural network, which is part of the hardware operating environment of the embodiment of the present invention.

[0050] Figure 2 This is a flowchart illustrating the first embodiment of the roadbed design method based on graph neural networks of the present invention.

[0051] Figure 3 This is a schematic diagram of one-hot encoding format conversion in an embodiment of the roadbed design method based on graph neural networks of the present invention;

[0052] Figure 4This is a schematic diagram of graph neural network prediction in an embodiment of the roadbed design method based on graph neural networks of the present invention;

[0053] Figure 5 This is a flowchart of graph neural network convolution calculation in an embodiment of the roadbed design method based on graph neural networks of the present invention;

[0054] Figure 6 This is a flowchart illustrating the complete roadbed design scheme of an embodiment of the roadbed design method based on graph neural networks of the present invention.

[0055] Figure 7 This is a flowchart illustrating the second embodiment of the roadbed design method based on graph neural networks of the present invention.

[0056] Figure 8 This is a case study diagram of foundation reinforcement in an embodiment of the roadbed design method based on graph neural networks of the present invention.

[0057] Figure 9 This is a case study diagram of a retaining structure in an embodiment of the roadbed design method based on graph neural networks of the present invention.

[0058] Figure 10 This is a slope protection case diagram of an embodiment of the roadbed design method based on graph neural networks of the present invention;

[0059] Figure 11 This is a structural block diagram of the first embodiment of the roadbed design device based on graph neural networks of the present invention.

[0060] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0061] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0062] Reference Figure 1 , Figure 1 This is a schematic diagram of the roadbed design equipment structure based on graph neural networks, which is part of the hardware operating environment involved in the embodiments of the present invention.

[0063] like Figure 1As shown, the roadbed design device based on graph neural networks may include: a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include a display screen or an input unit such as a keyboard; optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wireless-Fidelity (Wi-Fi) interface). The memory 1005 may be a high-speed random access memory (RAM) or a stable non-volatile memory (NVM), such as a disk drive. Optionally, the memory 1005 may also be a storage device independent of the aforementioned processor 1001.

[0064] Those skilled in the art will understand that Figure 1 The structure shown does not constitute a limitation on the roadbed design device based on graph neural networks, and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0065] like Figure 1 As shown, the memory 1005, which serves as a storage medium, may include an operating system, a network communication module, a user interface module, and a roadbed design program based on a graph neural network.

[0066] exist Figure 1 In the roadbed design device based on graph neural networks shown, the network interface 1004 is mainly used for data communication with the network server; the user interface 1003 is mainly used for data interaction with the user; the processor 1001 and memory 1005 in the roadbed design device based on graph neural networks of the present invention can be set in the roadbed design device based on graph neural networks. The roadbed design device based on graph neural networks calls the roadbed design program based on graph neural networks stored in the memory 1005 through the processor 1001 and executes the roadbed design method based on graph neural networks provided in the embodiment of the present invention.

[0067] This invention provides a roadbed design method based on graph neural networks, referring to... Figure 2 , Figure 2 This is a flowchart illustrating the first embodiment of a roadbed design method based on graph neural networks according to the present invention.

[0068] In this embodiment, the roadbed design method based on graph neural networks includes the following steps:

[0069] Step S10: Obtain roadbed detection parameters.

[0070] Understandably, roadbed detection parameters can be data related to foundation reinforcement, such as line grade, track type, geological conditions, treatment purpose, and environmental conditions; or data related to retaining structures, such as slope grade, cut / fill state, site conditions, backfill, and work site type; or data related to slope protection, such as slope grade, cut / fill state, slope surface soil and rock properties, site conditions, and work site type.

[0071] It should be understood that the collected roadbed detection parameters correspond to the pre-established graph neural network. If the pre-established graph neural network is a model for predicting the foundation reinforcement design scheme, then the obtained roadbed detection parameters are the line grade, track type, stratum condition, treatment purpose, and environmental conditions.

[0072] It should be noted that predictions can be made for any one aspect of foundation reinforcement, retaining structure, or slope protection, or three different aspects can be built and predicted simultaneously.

[0073] It should be noted that the execution subject of this embodiment is a roadbed design device based on graph neural networks. This roadbed design device based on graph neural networks has functions such as data processing, data communication, and program execution. The roadbed design device based on graph neural networks can be an integrated controller, a control computer, or other devices with similar functions. This embodiment does not limit the scope of these devices.

[0074] Step S20: Based on the roadbed detection parameters, a prediction weight sequence is obtained by using a pre-established graph neural network. The graph neural network includes a graph network and a graph classification. The graph network includes a graph convolutional layer and a pre-selected pooling layer constructed from a graph ontology.

[0075] Understandably, the graph neural network includes a graph network and a graph classification. The graph network includes graph convolutional layers and pre-selected pooling layers constructed from graph ontology. The graph ontology is constructed from the indices corresponding to each roadbed detection parameter.

[0076] In practical implementation, if a graph neural network is used to predict the design scheme of foundation reinforcement, the graph ontology is constructed from indicators such as line grade, track type, stratum condition, treatment purpose and environmental conditions; similarly, if a graph neural network is used to predict the design scheme of retaining structure, the graph ontology is constructed from indicators such as slope grade, cut and fill state, site conditions, backfill of wall, and work point type.

[0077] It should be noted that by constructing corresponding graph ontologies based on the indicator data corresponding to different prediction needs, graph neural networks can be built. The graph ontologies can be replaced as a separate module, allowing for flexible construction of prediction models based on different needs.

[0078] It is worth noting that the dynamic nature of the graph ontology allows it to adjust in real time according to different decision-making scenarios, providing highly customized solutions. This means that the algorithm of this invention can flexibly adapt to different needs, whether it is for roadbed foundation treatment, retaining schemes, or slope treatment.

[0079] Furthermore, the design of the neural network modules is also modular; they can be trained independently and combined with different graph ontologies. This increases the flexibility and scalability of the entire system. Different neural network modules can be selected according to the requirements of specific tasks, and then combined with corresponding graph ontologies, making it easier to update and improve various parts of the system without making major changes to the entire system. Therefore, it can adapt more quickly to new decision-making scenarios and changes in knowledge, thereby continuously improving the system's performance and adaptability, ensuring that the current recommendation and prediction solutions remain up-to-date and most effective.

[0080] It should be noted that the graph neural network may also include a preprocessing layer, which can be used to perform preliminary processing on the data input to the graph neural network.

[0081] It should be understood that data feature analysis of the dataset input to the graph neural network reveals that the input factors all contain both discrete and continuous values. For example, cut and fill volume can be categorized into three types: fill, cut, and partial fill / partial cut. Fill height and depth, on the other hand, are continuous values ​​with decimal values, and these values ​​are not uniformly distributed. Therefore, preprocessing of the input dataset is necessary to convert it into a format that the graph neural network can recognize.

[0082] It should be noted that the preprocessing method for the input data can be one-hot encoding. The roadbed detection parameters are one-hot encoded using the preprocessing layer in the graph neural network to obtain the roadbed parameter encoding. For specific one-hot encoding format conversion, please refer to... Figure 3 The one-hot array in the diagram can be understood as the data after one-hot encoding. The diagram illustrates the one-hot encoding process and may not have actual encoding meaning; among them, Figure 3 The encoding criteria for the raw data can be referenced in Table 1 below:

[0083]

[0084] Table 1 shows the input datasets and output labels for various graph neural networks.

[0085] Furthermore, the roadbed parameter encoding is convolved based on the graph convolutional layer and preset activation function in the graph neural network to obtain roadbed features; the roadbed features are dimensionality-reduced and sampled based on the graph pooling layer in the graph neural network to obtain sampled features; and the sampled features and the prediction weights of each preset design scheme are predicted based on the graph classification layer in the graph neural network.

[0086] The complete prediction process of graph neural networks can be found by referring to... Figure 4 The graph includes a preprocessing layer, a graph network, and a graph classification layer; further, the graph network includes a graph convolutional layer and a graph pooling layer constructed from graph ontology.

[0087] In practice, the preprocessing layer preprocesses the input roadbed detection parameters (i.e., the influencing factors in the graph) using one-hot encoding. Based on the preprocessed data, a graph ontology is constructed, and word embedding vectors are obtained. These word embedding vectors are then embedded into the word embedding layer of the graph convolutional layer. Simultaneously, the preprocessed data undergoes convolution calculation using a deep learning network and the activation function of the graph convolutional layer to obtain roadbed features. Further, these roadbed features are dimensionality-reduced and sampled using the activation function of the graph pooling layer and an attention mechanism to obtain sampled features. Finally, a deep learning neural network based on the graph classification layer classifies the sampled features to obtain prediction weights.

[0088] It should be noted that, in summary, this embodiment preprocesses data, covering the discretization of continuous values ​​and the encoding transformation of discrete values; then, it employs a graph neural network, where nodes represent geological parameters and edges represent the relationships between parameters, coupled with activation functions such as ReLU and Sigmoid; furthermore, it classifies and incorporates external factors such as geographical and geological influence factors into the roadbed design scheme, combines historical cases and expert experience, integrates expert schemes with the predictions of the graph neural network, and utilizes the structure of the graph ontology and the training of the graph neural network to ensure that the decision-making is both in-depth and clear.

[0089] Understandably, the output of a graph neural network can be a sequence of predicted weights, which can be the weights of the input roadbed detection parameters relative to each roadbed design scheme.

[0090] In practical implementation, for example, when performing classification, the input roadbed detection parameter S can be calculated, and the output can be the weight of S using roadbed design scheme 1, the weight of S using roadbed design scheme 2, ..., the weight of S using roadbed design scheme 10, which is a weight sequence.

[0091] It should be further explained that the step of predicting the roadbed detection parameters based on the pre-established graph neural network to obtain the prediction weight sequence includes: preprocessing the roadbed detection parameters to obtain parameters to be predicted; extracting features from the parameters to be predicted through a graph convolutional layer to obtain parameter association features; performing dimensionality reduction processing on the parameter association features through a pre-selected pooling layer to obtain dimensionality-reduced parameter association features; and classifying the dimensionality-reduced parameter association features through a fully connected graph classification layer to obtain the prediction weight sequence.

[0092] It should be noted that by using graph neural networks to train deep learning on datasets of roadbed foundation reinforcement, retaining structures, and slope protection, predictions of design schemes can be made. The network output stage is the final output target of the entire decision-making process. In this stage, fully connected layers fuse the features processed by graph convolutional layers to form the final decision prediction. This prediction is usually a vector representing the probability of different design schemes. For example, in the foundation reinforcement case, if there are five different reinforcement schemes, the network output might be a vector containing five elements, such as [0.1, 0.3, 0.4, 0.15, 0.05]. Each element in this vector represents the probability of the corresponding scheme being selected. In this example, the third scheme (with a value of 0.4) is the most likely choice. In addition, the network can also provide a probability score for each predicted classification, which is not just a classification decision but provides the level of confidence in the decision, offering decision-makers richer information.

[0093] In this context, it's understandable that two graph convolutional layers are used to learn the relationships between attributes. Graph convolution operations can extract features from the graph structure and capture the mutual influence between nodes. Taking a foundation as an example, the dataset contains the total number of strata and their properties, where the total number of strata is an integer, and the properties can be silt and silty soil, cohesive soil, sand, etc. Through graph convolutional layers, the correlation between the total number of strata and their properties can be learned and used for decision prediction. The graph neural network convolution calculation process can be found in [reference needed]. Figure 5 .

[0094] In this architecture, after the graph convolutional layer, two fully connected layers connect the output of the graph convolutional layer to the final output layer. The fully connected layers can further process features and perform decision predictions.

[0095] In practical implementation, the treatment objectives include controlling settlement, improving foundation bearing capacity, enhancing stability, and improving liquefaction resistance. Performance evaluation and iterative optimization are crucial steps in network training to ensure the accuracy and reliability of the network output. First, to evaluate model performance, the dataset needs to be divided into training and test sets. In each training epoch, the model learns on the training set and is then evaluated on the test set. In the training process of this invention, the Adam optimizer is used to update parameters. The Adam optimizer combines the advantages of momentum and RMSprop algorithms, automatically adjusting the learning rate based on historical gradient information of the parameters. This method adjusts the learning rate of each parameter by saving estimates of the first moment (mean) and second moment (unbiased variance) of past gradients. Therefore, the Adam optimizer not only improves the model's accuracy but also accelerates convergence. After each training and testing cycle, the current loss value and accuracy are calculated and recorded to monitor model performance and make necessary adjustments. Through this iterative process, the model can learn and optimize its performance more effectively. Based on these evaluation results, the parameters of the graph neural network can be adjusted to improve its performance and interpretability in practical applications. Through this iterative process, the model can continuously optimize and adjust itself to adapt to different datasets and application scenarios.

[0096] The process of preprocessing roadbed detection parameters to obtain parameters to be predicted includes: classifying the roadbed detection parameters to obtain continuous data and discrete data; performing one-hot encoding on the discrete data to obtain discrete data parameter encoding; performing preliminary encoding on the continuous data to obtain integer data; performing one-hot encoding on the integer data to obtain continuous data parameter encoding; and obtaining the parameters to be predicted based on the discrete data parameter encoding and the continuous data parameter encoding.

[0097] Understandably, data preprocessing encompasses data extraction from the case library, discretization of continuous data, and encoding transformation of discrete values. Subsequently, a graph ontology is constructed within the graph neural network. This construction process connects input data nodes via the graph ontology based on expert experience, thereby integrating expert knowledge with the neural network's ability to extract abstract features for prediction, enhancing the accuracy and interpretability of the recommended solutions. Next, the specific structures of the neural network, such as word embedding layers, convolutional layers, and pooling layers, are constructed. Then, the output of the neural network is calculated, and the internal weights are updated based on the difference between the output and the true values, undergoing iterative training. Finally, the trained graph neural network is used to calculate the input geological parameters and output the weight vector of the predicted solution.

[0098] Step S30: Determine the target roadbed design scheme from the preset design schemes based on the predicted weight sequence.

[0099] Understandably, the preset design scheme can be the classification scheme given when training the graph neural network. Simply put, if there are 10 design schemes, then there are 10 categories corresponding to the classification of the sampled features in the graph neural network. Classifying the sampled features can obtain the degree of matching between the sampled features and each category, and the degree of matching can be understood as the prediction weight.

[0100] Understandably, based on the weight series, the matching degree between the roadbed detection parameters and each roadbed design scheme can be known. The higher the weight, the higher the matching degree. The roadbed design scheme corresponding to the highest weight is taken as the target roadbed design scheme.

[0101] In specific implementation, determining the target roadbed design scheme from the preset design schemes based on the predicted weight sequence can be as follows: obtaining the target predicted weight with the largest weight from the predicted weight sequence; determining the corresponding reference design scheme in the preset design schemes based on the target predicted weight; and using the reference design scheme as the target roadbed design scheme.

[0102] It should be noted that the step of determining the target roadbed design scheme from the preset design schemes based on the predicted weight sequence includes: obtaining the target predicted weight with the largest weight from the predicted weight sequence; determining the corresponding reference design scheme in the preset design schemes based on the target predicted weight; verifying the reference design scheme, and when the verification is successful, using the reference design scheme as the target roadbed design scheme.

[0103] Understandably, the verification of the reference design can be evaluated through an expert system. After the graph neural network outputs the weight matrix of the recommended solution, the algorithm's recommendation results need to be reviewed by experts to eliminate obviously unsuitable solutions and increase the stability of the entire process.

[0104] After passing expert review, this design scheme will be pushed to downstream professionals for subsequent practical design and application in the project. After the project is completed, important design parameters and details will be automatically entered into the data lake to increase the number of samples in the case library, and the algorithm will be iteratively trained to increase the correctness and stability of the entire process. The entire update and iteration process is as follows:

[0105] In practice, upon completion of each engineering project, key design parameters and actual performance data are automatically collected and uploaded to a central data lake. This data covers everything from engineering geological data to the types of materials used, construction methods, costs and efficiency, and ultimately, stability and performance. This newly collected data will be merged with the existing case library, and subsequent automated and manual data cleaning and preprocessing steps will ensure data quality and adapt it to the model's input requirements, thereby enriching and updating the case library.

[0106] Understandably, the graph neural network in this embodiment can perform self-updating. This mechanism involves periodically retraining the model using the latest example library to ensure that the model can capture the latest trends and patterns. Furthermore, the real-time performance monitoring system continuously tracks the model's performance in real-world applications, such as prediction accuracy and decision time. If the monitoring system detects a performance decline or new trends and patterns, it automatically triggers additional training cycles to ensure the model can quickly adapt to these changes.

[0107] It should be noted that the roadbed design system based on graph neural networks described in this embodiment also possesses self-diagnostic capabilities, enabling it to periodically check for potential errors or data inconsistencies and automatically repair them, or notify the maintenance team. Simultaneously, a user feedback mechanism is established, allowing engineers and domain experts to provide feedback on the model's prediction results, thereby further optimizing the model. Furthermore, the system will automatically perform software updates, including algorithm updates, security patches, and performance improvements, ensuring the system is always up-to-date; the system regularly generates reports containing key performance indicators such as prediction accuracy and user satisfaction. These reports not only provide decision support for management but also guide the technical team in continuous optimization work. Long-term data analysis can provide valuable insights for algorithm optimization, such as adjusting specific parameters or introducing new data processing techniques.

[0108] It should be further explained that after determining the target roadbed design scheme from the preset design schemes based on the predicted weight sequence, the process further includes: updating the preset database based on the roadbed detection parameters and the target roadbed design scheme; training the graph neural network based on the preset database to obtain training results; and updating the graph neural network based on the training results. The complete roadbed design scheme process can be found by referring to [reference needed]. Figure 6 .

[0109] It is understood that the roadbed design process in this embodiment is the overall implementation process of the intelligent decision-making scheme, including the processing of survey data, forward propagation calculation of graph neural network, scheme weight recommendation, expert review and final scheme push, downstream design, and finally, the historical case library will be expanded, so as to iteratively train the graph neural network, update the weight of the neural network, and thus enhance the capabilities of the entire intelligent decision-making algorithm.

[0110] In practical implementation, using graph neural networks for intelligent decision-making in the field of transportation construction is a novel algorithm for prediction and recommendation. Leveraging deep learning technology, it combines the advantages of graph theory and neural networks to provide accurate and interpretable prediction algorithms for roadbed intelligent decision-making. By employing graph neural networks, topological constraints can be imposed on the input feature parameters, thereby constraining the abstract information that the neural network can learn, increasing the interpretability and stability of the neural network.

[0111] In designing graph networks, the construction of graph ontologies is a crucial step, involving the integration of expert knowledge into the network's structure. This invention employs a unique method to fuse prior expert knowledge with the topological structure of graph ontologies, thereby creating predefined graph connections within the feature data input to the neural network, forming a graph ontology-reinforced neural network. For example, feature data from multiple strata can be connected to a shared stratum ontology node, which can then be used for further graph neural network feature extraction. This approach not only enhances the interpretability of decisions, making it easier for users to understand the logic behind them, but also strengthens the system's fault tolerance, ensuring appropriate decisions can still be made in specific contexts.

[0112] It is important to emphasize that, in this embodiment, the data used to train the graph neural network comes from a pre-built roadbed engineering scheme design case library. This establishes a real-time self-learning mechanism for the graph neural network, enabling the decision-making algorithm to learn itself and its decision-making functions to maintain themselves. By constructing a roadbed engineering scheme design case library, which gathers various roadbed design examples and related parameters, these cases cover different geological conditions, environmental factors, and engineering requirements, providing rich learning materials for the graph neural network. Through learning from these cases, the graph neural network can identify and extract key design patterns and decision-making logic.

[0113] It should be noted that the graph neural network in this invention is equipped with a real-time self-learning mechanism. This means that the network can not only learn from the initial case library, but also continuously update and adjust its decision model based on new design cases and data collected in practical applications. This self-learning capability enables the decision-making algorithm to adapt to constantly changing design conditions and technological advancements, maintaining its predictive accuracy and relevance. By combining the roadbed engineering scheme design case library with the graph neural network's self-learning mechanism, this invention achieves continuous self-optimization and maintenance of the decision-making function. This not only improves the practicality and flexibility of the decision-making algorithm, but also ensures the continuous improvement of decision quality over the long term, meeting the needs of the rapidly developing field of transportation engineering.

[0114] It should be further explained that the intelligent decision-making and implementation process of the design scheme is as follows: This process comprehensively considers the processing of survey data, forward propagation calculation of graph neural networks, scheme weight recommendation, expert review and final scheme push, downstream design and other links to ensure the accuracy and feasibility of the final scheme; the calculation and training process of graph neural networks.

[0115] This embodiment combines the advantages of deep learning technology with graph theory and neural networks. It utilizes the deep reasoning capabilities of neural networks and the explicit logic of graph ontologies to provide an accurate and interpretable prediction algorithm for intelligent roadbed decision-making. It can implement topological constraints on input feature parameters, thereby constraining the abstract information that the neural network can learn. This not only improves the accuracy of decision-making but also enhances its interpretability. It avoids the problem that although deep learning neural networks can handle complex nonlinear relationships, their internal working mechanisms are usually opaque, which makes the decision-making process lack interpretability.

[0116] refer to Figure 7 , Figure 7 This is a flowchart illustrating a second embodiment of a roadbed design method based on graph neural networks according to the present invention.

[0117] Based on the first embodiment described above, the roadbed design method based on graph neural networks in this embodiment further includes, before step S20:

[0118] Step S201: Obtain the influence factors and lower-level influence factors of the subgrade design scheme, and determine the expert relationship between the influence factors and the lower-level influence factors.

[0119] Understandably, as shown in Table 1, different roadbed design schemes have different influencing factors. For example, roadbed design schemes include foundation reinforcement, retaining structures, and slope protection; each design scheme has many influencing factors and complex relationships.

[0120] It should be understood that the influencing factors can be factors related to the subgrade design, such as the influencing factors of foundation reinforcement. The foundation reinforcement influencing factors include multiple sub-influencing factors, such as the line grade, cut and fill conditions, foundation conditions, treatment purpose, environmental conditions, etc.

[0121] It should be noted that the expert relationship between the impact factor and the lower-level impact factors can be a characteristic that is pre-assigned to each impact factor based on experience and actual circumstances.

[0122] Step S202: Construct a graph ontology for each impact factor based on the lower-level impact factors and the expert relationships.

[0123] Understandably, the ontology of a roadbed design case study mainly consists of its influencing factors and design schemes; different combinations of influencing factors determine the selection of the design scheme. Foundation reinforcement, retaining structures, and slope protection schemes have numerous influencing factors with complex relationships. Taking foundation reinforcement as an example, it mainly includes aspects such as route grade, cut and fill conditions, foundation conditions, treatment objectives, and environmental conditions. Each aspect can be further subdivided; for example, foundation conditions include factors such as stratum name, stratum properties, special soils and rocks, and adverse geological conditions. Based on expert knowledge, different attributes are interconnected through the ontology. Compared to simple fully connected layers, this sparse connection method can better limit the extraction of more practically meaningful information by the graph neural network, improving the accuracy and interpretability of the graph neural network model.

[0124] It should be noted that foundation reinforcement, retaining structures, and slope protection schemes are influenced by numerous and complex factors. Taking foundation reinforcement as an example, it mainly includes aspects such as the railway line grade, cut and fill conditions, foundation conditions, treatment objectives, and environmental conditions. Each aspect can be further subdivided; for example, foundation conditions include factors such as stratum name, stratum properties, special soils and rocks, and adverse geological conditions. For details, please refer to [link to relevant documentation]. Figure 8 , Figure 9 as well as Figure 10 In the picture, Figure 8 This is the main body of the foundation reinforcement case study diagram. Figure 9 This is the main body of the support structure case diagram. Figure 10 This is the main body of the slope protection case diagram.

[0125] Furthermore, by connecting different attributes through ontology based on expert knowledge, compared to simple fully connected networks, this sparse connection method can better limit the graph neural network to extract more practically meaningful information, thereby improving the accuracy and interpretability of the graph neural network model.

[0126] Step S203: Construct an initial graph neural network based on the graph ontology, pre-selected pooling layer, pre-selected convolutional layer, and pre-selected fully connected layer.

[0127] It should be noted that the construction of the initial graph neural network based on the graph ontology, pre-selected pooling layer, pre-selected convolutional layer, and pre-selected fully connected layer includes: obtaining an adjacency matrix between each influencing factor based on the adjacency relationship between each influencing factor in the graph ontology; embedding the adjacency matrix into the pre-selected convolutional layer to obtain a graph convolutional layer; constructing a graph network based on the graph convolutional layer and the pre-selected pooling layer; constructing a graph classification based on the fully connected layer; and obtaining the initial graph neural network from the graph network and the graph classification.

[0128] Among them, reference Figure 4Understandably, embedding the adjacency matrix into a preset convolutional layer to obtain a graph convolutional layer can be achieved by connecting each node in the graph ontology based on the adjacency matrix. After connection, a connection matrix between each node can be obtained. Word embedding vectors are obtained based on the connection matrix. The word embedding vectors and the graph ontology are then embedded into the convolutional layer to obtain, as shown below. Figure 3 The graph convolutional layer in the image.

[0129] It's important to emphasize that graph neural networks (GNNs) differ from traditional neural networks. In addition to the input factor values ​​for each sample, GNNs require the connections between these input factors, i.e., an adjacency matrix. In graph theory, a graph consists of nodes (vertices) and edges connecting them. The adjacency matrix represents these connections and the adjacency relationships between nodes. For a graph with n nodes, the adjacency matrix is ​​an n×n matrix where each element represents a connection between nodes. The size of the adjacency matrix depends on the number of nodes in the graph. For a graph with n nodes, the size of the adjacency matrix is ​​n×n.

[0130] Step S204: Train the initial graph neural network to obtain a graph neural network.

[0131] It should be noted that training the initial graph neural network to obtain the graph neural network includes: obtaining sample roadbed data and sample design schemes from a preset database, wherein the preset data includes multiple successfully verified roadbed design schemes and corresponding roadbed data; predicting the sample roadbed data based on the initial graph neural network to obtain predicted sample weights; obtaining predicted design schemes based on the predicted sample weights; constructing a loss function based on the sample design schemes and the predicted design schemes; and optimizing the initial graph neural network based on the loss function to obtain the graph neural network.

[0132] This embodiment constructs a graph ontology by connecting the influencing factors of the roadbed design scheme with the quality inspection experts of each factor. It integrates the prior knowledge of the experts with the topological structure of the graph ontology, thereby creating predefined graph connections in the feature data input to the neural network. This allows feature data from multiple strata to be connected to a shared stratum ontology node. Then, this stratum ontology node is used for further graph neural network feature extraction, which improves the interpretability of the decision, makes it easier for users to understand the logic behind the decision, and enhances the fault tolerance of the system, ensuring that appropriate decisions can still be made in specific situations.

[0133] Furthermore, this embodiment of the invention also proposes a storage medium storing a roadbed design program based on a graph neural network. When the roadbed design program based on the graph neural network is executed by a processor, it implements the steps of the roadbed design method based on the graph neural network described above.

[0134] Reference Figure 11 , Figure 11 This is a structural block diagram of the first embodiment of the roadbed design device based on graph neural networks of the present invention.

[0135] like Figure 11 As shown, the roadbed design device based on graph neural networks proposed in this embodiment of the invention includes:

[0136] Parameter acquisition module 10 is used to acquire roadbed detection parameters;

[0137] The model prediction module 20 is used to make predictions based on the roadbed detection parameters through a pre-established graph neural network to obtain a prediction weight sequence. The graph neural network includes a graph convolutional layer, a graph pooling layer, and a graph classification layer constructed from graph ontology.

[0138] The scheme generation module 30 is used to determine the target roadbed design scheme from the preset design schemes based on the predicted weight sequence.

[0139] This embodiment combines the advantages of deep learning technology with graph theory and neural networks. It utilizes the deep reasoning capabilities of neural networks and the explicit logic of graph ontologies to provide an accurate and interpretable prediction algorithm for intelligent roadbed decision-making. It can implement topological constraints on input feature parameters, thereby constraining the abstract information that the neural network can learn. This not only improves the accuracy of decision-making but also enhances its interpretability. It avoids the problem that although deep learning neural networks can handle complex nonlinear relationships, their internal working mechanisms are usually opaque, which makes the decision-making process lack interpretability.

[0140] In one embodiment, the model prediction module 20 is further configured to obtain the influence factors and lower-level influence factors of the roadbed design scheme, and determine the expert relationship between the influence factors and the lower-level influence factors;

[0141] Construct a graph ontology for each impact factor based on the lower-level impact factors and the expert relationships;

[0142] An initial graph neural network is constructed based on the graph ontology, preselected pooling layer, preselected convolutional layer, and preselected fully connected layer;

[0143] The initial graph neural network is trained to obtain a graph neural network.

[0144] In one embodiment, the model prediction module 20 is further configured to obtain an adjacency matrix between each influencing factor based on the adjacency relationship between each influencing factor in the graph ontology;

[0145] The adjacency matrix is ​​embedded into a pre-selected convolutional layer to obtain a graph convolutional layer;

[0146] A graph network is constructed based on the graph convolutional layers and pre-selected pooling layers, and a graph classification is constructed based on the fully connected layers;

[0147] An initial graph neural network is obtained from the graph network and the graph classification.

[0148] In one embodiment, the model prediction module 20 is further configured to obtain sample roadbed data and sample design schemes from a preset database, wherein the preset data includes multiple successfully verified roadbed design schemes and corresponding roadbed data;

[0149] Based on the initial graph neural network, the sample roadbed data is predicted to obtain the predicted sample weights;

[0150] A prediction design scheme is obtained based on the predicted sample weights;

[0151] A loss function is constructed based on the sample design scheme and the prediction design scheme. The initial graph neural network is then optimized based on the loss function to obtain the graph neural network.

[0152] In one embodiment, the model prediction module 20 is further configured to preprocess the roadbed detection parameters to obtain the parameters to be predicted;

[0153] The parameters to be predicted are extracted using graph convolutional layers to obtain parameter-related features;

[0154] The parameter association features are reduced in dimensionality by a pre-selected pooling layer to obtain the dimensionality-reduced parameter association features.

[0155] The reduced parameter association features are classified using a fully connected layer for graph classification to obtain a predicted weight sequence.

[0156] In one embodiment, the model prediction module 20 is further configured to classify the roadbed detection parameters to obtain continuous data and discrete data;

[0157] One-hot encoding is performed on the discrete data to obtain discrete data parameter encoding;

[0158] The continuous data is initially encoded to obtain integer data, and the integer data is then one-hot encoded to obtain the continuous data parameter encoding.

[0159] The parameters to be predicted are obtained based on the discrete data parameter encoding and the continuous data parameter encoding.

[0160] In one embodiment, the scheme generation module 30 is further configured to obtain the target prediction weight with the largest weight from the prediction weight sequence;

[0161] The reference design scheme in the preset design scheme is determined based on the target prediction weight;

[0162] The reference design scheme is verified, and if the verification is successful, the reference design scheme is adopted as the target roadbed design scheme.

[0163] After determining the target roadbed design scheme from the preset design schemes based on the predicted weight sequence, the method further includes:

[0164] The preset database is updated based on the roadbed detection parameters and the target roadbed design scheme.

[0165] The graph neural network is trained based on the preset database to obtain the training results;

[0166] The graph neural network is updated based on the training results.

[0167] It should be understood that the above are merely illustrative examples and do not constitute any limitation on the technical solutions of the present invention. In specific applications, those skilled in the art can make settings as needed, and the present invention does not impose any restrictions on this.

[0168] It should be noted that the workflow described above is merely illustrative and does not limit the scope of protection of this invention. In practical applications, those skilled in the art can select some or all of the workflow to achieve the purpose of this embodiment according to actual needs, and no restrictions are imposed here.

[0169] Furthermore, it should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0170] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0171] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as read-only memory (ROM) / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0172] It should be understood that although the steps in the flowcharts of this application's embodiments are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some of the steps in the figures may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be performed alternately or in turn with other steps or at least a portion of the sub-steps or stages of other steps.

[0173] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A roadbed design method based on graph neural networks, characterized in that, The roadbed design method based on graph neural networks includes: Obtain roadbed detection parameters; A prediction weight sequence is obtained by making predictions based on the roadbed detection parameters through a pre-established graph neural network. The graph neural network includes a graph network and a graph classification. The graph network includes a graph convolutional layer and a pre-selected pooling layer constructed from graph ontology. The target roadbed design scheme is determined from the preset design scheme based on the predicted weight sequence; The step of predicting based on the roadbed detection parameters using a pre-established graph neural network to obtain a prediction weight sequence includes: The roadbed detection parameters are preprocessed to obtain the parameters to be predicted; The parameters to be predicted are extracted using graph convolutional layers to obtain parameter-related features; The parameter association features are reduced in dimensionality by a pre-selected pooling layer to obtain the dimensionality-reduced parameter association features. The reduced parameter association features are classified using a fully connected layer for graph classification to obtain a predicted weight sequence. The preprocessing of the roadbed detection parameters to obtain the parameters to be predicted includes: The roadbed detection parameters are classified to obtain continuous data and discrete data; One-hot encoding is performed on the discrete data to obtain discrete data parameter encoding; The continuous data is initially encoded to obtain integer data, and the integer data is then one-hot encoded to obtain the continuous data parameter encoding. The parameters to be predicted are obtained based on the discrete data parameter encoding and the continuous data parameter encoding. The step of determining the target roadbed design scheme from the preset design schemes based on the predicted weight sequence includes: The target prediction weight with the largest weight is obtained from the predicted weight sequence; The reference design scheme in the preset design scheme is determined based on the target prediction weight; The reference design scheme is verified, and if the verification is successful, the reference design scheme is adopted as the target roadbed design scheme. After determining the target roadbed design scheme from the preset design schemes based on the predicted weight sequence, the method further includes: The preset database is updated based on the roadbed detection parameters and the target roadbed design scheme. The graph neural network is trained based on the preset database to obtain the training results; The graph neural network is updated based on the training results.

2. The roadbed design method based on graph neural networks as described in claim 1, characterized in that, Before obtaining the prediction weight sequence by making predictions based on the roadbed detection parameters using a pre-established graph neural network, the method further includes: Obtain the influence factors and lower-level influence factors of the subgrade design scheme, and determine the expert relationship between the influence factors and the lower-level influence factors; Construct a graph ontology for each impact factor based on the lower-level impact factors and the expert relationships; An initial graph neural network is constructed based on the graph ontology, preselected pooling layer, preselected convolutional layer, and preselected fully connected layer; The initial graph neural network is trained to obtain a graph neural network.

3. The roadbed design method based on graph neural networks as described in claim 2, characterized in that, The construction of the initial graph neural network based on the graph ontology, pre-selected pooling layers, pre-selected convolutional layers, and pre-selected fully connected layers includes: The adjacency matrix between each influencing factor is obtained based on the adjacency relationship between each influencing factor in the graph ontology. The adjacency matrix is ​​embedded into a pre-selected convolutional layer to obtain a graph convolutional layer; A graph network is constructed based on the graph convolutional layers and pre-selected pooling layers, and a graph classification is constructed based on the fully connected layers; An initial graph neural network is obtained from the graph network and the graph classification.

4. The roadbed design method based on graph neural networks as described in claim 2, characterized in that, The step of training the initial graph neural network to obtain a graph neural network includes: Sample roadbed data and sample design schemes are obtained from a preset database, wherein the preset data includes multiple successfully verified roadbed design schemes and their corresponding roadbed data; Based on the initial graph neural network, the sample roadbed data is predicted to obtain the predicted sample weights; A prediction design scheme is obtained based on the predicted sample weights; A loss function is constructed based on the sample design scheme and the prediction design scheme. The initial graph neural network is then optimized based on the loss function to obtain the graph neural network.

5. A roadbed design device based on graph neural networks, characterized in that, The roadbed design device based on graph neural networks includes: The parameter acquisition module is used to acquire roadbed detection parameters; The model prediction module is used to make predictions based on the roadbed detection parameters through a pre-established graph neural network to obtain a prediction weight sequence. The graph neural network includes a graph network and a graph classification. The graph network includes a graph convolutional layer and a pre-selected pooling layer constructed from a graph ontology. The scheme generation module is used to determine the target roadbed design scheme from the preset design schemes based on the predicted weight sequence; The model prediction module is further configured to preprocess the roadbed detection parameters to obtain parameters to be predicted; extract features from the parameters to be predicted through a graph convolutional layer to obtain parameter association features; reduce the dimensionality of the parameter association features through a pre-selected pooling layer to obtain dimensionality-reduced parameter association features; and classify the dimensionality-reduced parameter association features through a fully connected graph classification layer to obtain a prediction weight sequence. The model prediction module is further configured to classify the roadbed detection parameters to obtain continuous data and discrete data; perform one-hot encoding on the discrete data to obtain discrete data parameter encoding; perform preliminary encoding on the continuous data to obtain integer data; perform one-hot encoding on the integer data to obtain continuous data parameter encoding; and obtain the parameters to be predicted based on the discrete data parameter encoding and the continuous data parameter encoding. The scheme generation module is further configured to obtain the target prediction weight with the largest weight from the prediction weight sequence; and determine the corresponding reference design scheme in the preset design scheme based on the target prediction weight. The reference design scheme is verified, and if the verification is successful, the reference design scheme is used as the target roadbed design scheme. After determining the target roadbed design scheme from the preset design schemes according to the predicted weight sequence, the method further includes: updating the preset database according to the roadbed detection parameters and the target roadbed design scheme; training the graph neural network based on the preset database to obtain the training result; and updating the graph neural network according to the training result.

6. A roadbed design device based on graph neural networks, characterized in that, The device includes: a memory, a processor, and a graph neural network-based roadbed design program stored in the memory and executable on the processor, the graph neural network-based roadbed design program being configured to implement the graph neural network-based roadbed design method as described in any one of claims 1 to 4.

7. A storage medium, characterized in that, The storage medium stores a roadbed design program based on a graph neural network, which, when executed by a processor, implements the roadbed design method based on a graph neural network as described in any one of claims 1 to 4.

Images