A method for constructing railway engineering design knowledge base based on cloud platform

Abstract

The present invention proposes a method for constructing a railway engineering design knowledge base based on a cloud platform, which is applied to the cloud platform. The method includes: using the cloud platform to collect multimodal data, designing a distributed file system to store preprocessed data, and establishing a data indexing mechanism to retrieve multimodal data; using a deep learning model to intelligently analyze these multimodal data; identifying and extracting entities and relationships in railway engineering design from feature vectors of multimodal data to construct a knowledge graph, and then using a graph neural network to train the knowledge graph; automating design suggestions and optimization plans; constructing a cross-railway engineering design institution sharing model based on federated learning for knowledge sharing and privacy protection; designing a reinforcement learning model to train the simulation environment, and automatically generating a railway engineering design plan based on the trained reinforcement model. The present invention provides an efficient, intelligent and reliable knowledge base construction method for railway engineering design.

Description

Technical Field

[0001] The present invention belongs to the field of cloud platform data processing, and in particular relates to a method for constructing a railway engineering design knowledge base based on a cloud platform. Background Art

[0002] Railway engineering design is a complex process involving multiple disciplines and multiple data sources, covering a large number of design elements, technical specifications and engineering data. Traditional railway design methods rely on the experience of experts and manual operations, resulting in low design efficiency, prone to errors, and difficult to effectively manage and share design data. These data are usually scattered in different systems and departments, lacking unified management and efficient means of utilization. The existing railway design knowledge base system is mainly based on static document storage and simple keyword retrieval, which is difficult to cope with the complex needs of modern railway engineering design.

[0003] Some attempts at existing technologies, such as natural language processing, image recognition, and machine learning, have improved the railway design process to a certain extent, but there are still obvious shortcomings:

[0004] Data integration and management: Railway engineering design data includes text, drawings, videos and other forms. Existing systems are unable to effectively integrate and manage these multimodal data, resulting in low data utilization.

[0005] Intelligent analysis and information extraction: Although natural language processing and image recognition technologies are used, the effects of automated knowledge extraction and analysis are not ideal, and it is impossible to deeply understand complex design information.

[0006] Knowledge sharing and privacy protection: Cross-organizational knowledge sharing faces data privacy and security issues. Existing technologies make it difficult to find a balance between knowledge sharing and data privacy protection.

[0007] Design optimization and decision support: Existing design optimization methods mainly rely on expert experience and lack intelligent decision support systems, making it difficult to automatically generate and optimize design solutions.

[0008] To solve these problems, a new technical approach is needed that can effectively integrate and manage multimodal data, provide intelligent knowledge analysis and extraction functions, ensure cross-organizational knowledge sharing and data privacy, and achieve intelligent optimization of design solutions and decision support. Summary of the invention

[0009] The purpose of this invention is to propose a method for constructing a railway engineering design knowledge base based on a cloud platform. Through innovative points such as intelligent analysis and information extraction driven by deep learning, intelligent query and decision support based on knowledge graphs, and cross-organizational knowledge sharing and privacy protection driven by federated learning, the shortcomings of the existing technology are comprehensively solved, providing an efficient, intelligent and reliable knowledge base construction method for railway engineering design.

[0010] In order to achieve the above object, the present invention provides a method for constructing a railway engineering design knowledge base based on a cloud platform, which is applied to the cloud platform. The method includes:

[0011] S1. Using a cloud platform to collect multimodal data and preprocess the multimodal data, designing a distributed file system to store the preprocessed data, and establishing a data indexing mechanism to retrieve the multimodal data; wherein the multimodal data includes text data, image data, and video data;

[0012] S2. Use deep learning models to intelligently analyze these multimodal data, extract key information and technical parameters of railway engineering design, and then fuse the feature vectors of multimodal data to generate a comprehensive feature representation; use pre-trained language models to extract and analyze text data; use convolutional neural networks to extract and analyze image data; and use three-dimensional convolutional neural networks to extract and analyze video data;

[0013] S3, identify and extract entities and relationships in railway engineering design from the feature vectors of multimodal data to construct a knowledge graph, and then use a graph neural network to train the knowledge graph;

[0014] S4. Develop an intelligent query system based on the trained knowledge graph to provide automated design suggestions and optimization solutions, including:

[0015] S401, the user inputs the design requirements, and the intelligent query system uses the pre-trained language model to convert the natural language requirements into feature vector representation;

[0016] S402, querying the knowledge graph according to the feature vector representation, calculating the similarity between the user demand feature vector and the feature vector of each node in the knowledge graph, and matching the design elements and relationships;

[0017] S403, performing reasoning through a graph neural network based on the query results, using similarity to calculate the relationship weight between the candidate node and its neighbor nodes to obtain a final similarity, sorting the candidate nodes according to the final similarity value, and selecting the nodes with the highest similarity and their associated design elements to generate optimization suggestions;

[0018] S404: Collect user feedback on design suggestions, record feedback information, adjust the parameters of the graph neural network based on the feedback information, introduce feedback adjustment items, and update the loss function of the graph neural network, which is expressed as follows:

[0019] ;

[0020] in, is the loss function after introducing the feedback adjustment term, is the original loss function, For user feedback, is the current node feature vector, is the target node feature vector, is the feedback adjustment coefficient;

[0021] S5. Construct a cross-railway engineering design agency sharing model based on federated learning to share knowledge and protect privacy among multiple railway design agencies;

[0022] S6. Build a simulation environment for railway engineering design, then design a reinforcement learning model to train the simulation environment, automatically generate railway engineering design plans based on the trained reinforcement model, and perform evaluation and optimization.

[0023] Preferably, the preprocessing includes: performing word segmentation, stop word removal and stem extraction on text data, performing grayscale, denoising and edge detection on image data, and performing frame segmentation on video data; the data indexing mechanism includes: indexing text data using inverted indexing technology, extracting feature vectors from image data and indexing using kd tree or LSH, and performing feature extraction and indexing on key frames of video data.

[0024] Preferably, the constructing of the knowledge graph includes:

[0025] Extract entities and relationships from text features and build a preliminary graph structure , where V represents the node set and E represents the edge set. The image features and video features are aligned with the text entities, the graph structure is updated, and the graph embedding method is used to represent the features of the nodes and edges.

[0026] Preferably, the use of a graph neural network to train the knowledge graph includes:

[0027] Initialize the feature representation of nodes and edges, where nodes The initial eigenvector of It is the result of multimodal feature fusion;

[0028] During each layer of propagation, the feature representation of a node is updated according to the features of its neighboring nodes, as shown below:

[0029] ;

[0030] in, Representation Node In the The feature representation of the layer, and denote the weight matrix and bias vector respectively, Representation Node The set of neighbor nodes of and Respectively represent nodes and The degree, represents the activation function;

[0031] A regularization term based on edge weight is introduced to ensure the optimization of edge features, which is expressed as follows:

[0032] ;

[0033] in, represents the edge weight regularization term, Represents edge The weight of Representation Node and characteristic differences.

[0034] Preferably, in S401, the text requirement input by the user is tokenized to obtain a token sequence , use the pre-trained language model to process the Token sequence and obtain the implicit representation of each Token , will implicitly represent After the average pooling operation, the feature vector of user demand is obtained ;

[0035] In S402, the user demand feature vector is calculated And each node feature vector in the knowledge graph The similarity is expressed as follows:

[0036] ;

[0037] in, represents the user demand feature vector, Represents the feature vector of each node in the knowledge graph, is the similarity between the demand vector and the node vector, represents the vector inner product, Represents the vector norm.

[0038] Preferably, generating optimization suggestions based on the nodes with the highest similarity and their associated design elements includes the following steps:

[0039] The graph neural network is used for reasoning and the relationship weights between candidate nodes and their neighbor nodes are calculated, which can be expressed as follows:

[0040] ;

[0041] in, Representation Node In the The feature representation of the layer, and denote the weight matrix and bias vector respectively, Representation Node The set of neighbor nodes of and Respectively represent nodes and The degree, represents the activation function;

[0042] Calculate the node feature vector after inference User demand feature vector The final similarity is expressed as follows:

[0043] ;

[0044] in, is the final similarity, is the optimized node feature vector, represents the vector inner product, represents the vector norm;

[0045] The candidate nodes are sorted according to the final similarity values, and the nodes with the highest similarity and their associated design elements are selected to generate optimization suggestions.

[0046] Preferably, the S5 specifically includes:

[0047] S501. Extract features and labels from local data of each institution and build a training set , initialize local model parameters , using stochastic gradient descent locally using the training set Train the local model and update the parameters;

[0048] S502: Introduce a differential privacy mechanism to protect data privacy during local training and parameter sharing, and obtain the model parameters after adding the differential privacy mechanism. ;

[0049] S503. Each organization will use the model parameters trained locally Sent to the central server for aggregation to obtain global model parameters ;

[0050] S504, introducing a privacy-preserving regularization term into the loss function of the global model to limit the range of variation of the global model parameters;

[0051] S505: The central server aggregates the global model parameters Distribute to each organization, and each organization uses the updated global model to perform the next round of local training, and repeat S501 to S504.

[0052] Preferably, the stochastic gradient descent method is used locally to train the Training the local model is as follows:

[0053] ;

[0054] in, represents the model parameters of the kth iteration in the tth round of training, represents the learning rate, represents the loss function, and Represent the input features and labels respectively;

[0055] The S504 specifically includes:

[0056] Add regularization term during local training , limiting the range of changes in model parameters shared across railway engineering design organizations, as shown below:

[0057] ;

[0058] in, represents the total loss function of local training, represents the regularization coefficient, represents the regularization term, and denote input features and labels respectively, represents the model parameters of the kth iteration in the tth round of training;

[0059] Defining privacy-preserving regularization terms , so that it is penalized when the global model parameters change too much, as shown below:

[0060] ;

[0061] Among them, w represents the current model parameters, are the initial model parameters, is the L2 norm.

[0062] Preferably, the S6 specifically includes:

[0063] S601. Construct a simulation environment based on the actual railway engineering design process, including: defining the state space , including all design states; defining the action space , which contains all design decisions; defines the reward function , used to evaluate the quality of each design decision;

[0064] S602. Build a deep reinforcement learning model to train the design simulation environment and obtain a trained strategy ;

[0065] S603. According to the trained strategy , generate the optimal design solution ,Then the generated design scheme is simulated and evaluated multiple times, and the design scheme is optimized based on the evaluation results.

[0066] Preferably, the S602 includes:

[0067] Initialize the parameters of the Q network and the parameters of the target Q network , and set the experience replay pool D, for each time step t, in state Next, select Action , perform the action , observe the reward and the next state , storage experience into D, and then randomly sample a mini-batch from D , calculate the target Q value, expressed as follows:

[0068] ;

[0069] in, represents the discount factor, represents the Q value of the target Q network;

[0070] Minimize the loss function using gradient descent , update the parameters of the Q network , which is expressed as follows:

[0071] ;

[0072] In the periodic Q network parameters Parameters copied to the target Q network ;

[0073] Then define the multi-objective reward function, taking into account multiple design goals, as follows:

[0074] ;

[0075] in, represents the reward function of the ith target, represents the corresponding weight coefficient;

[0076] In the loss function The constraint processing item is introduced to ensure that the optimization process meets the design constraints, which is expressed as follows:

[0077] ;

[0078] in, represents the loss function with constraint processing, represents the kth constraint, represents the constraint penalty coefficient;

[0079] Then dynamically adjust the reward function based on real-time feedback The weight of the Q network is then updated using an adaptive optimization algorithm, and a reward adjustment mechanism based on uncommon regularization terms is introduced to make the reinforcement learning model calculation design scheme as follows:

[0080] ;

[0081] in, Representation strategy In Status The entropy under represents the regularization coefficient.

[0082] The beneficial technical effects of the present invention are at least as follows:

[0083] (1) The present invention uses advanced deep learning models to intelligently analyze railway design documents and extract key design information and technical parameters. At the same time, convolutional neural network technology is used to analyze design drawings and construction videos, extract image features and video content, and achieve in-depth understanding of multimodal data. This solves the shortcomings of existing technologies in data integration and intelligent analysis.

[0084] (2) The present invention uses graph neural network technology to construct a knowledge graph for railway engineering design, representing design elements and their relationships as a graph structure. The intelligent query system based on the knowledge graph can automatically provide design suggestions and optimization solutions, improve design quality and decision-making efficiency. This solves the shortcomings of existing technologies in intelligent decision support.

[0085] (3) The present invention uses federated learning technology among multiple railway design institutions to achieve knowledge sharing while protecting their respective data privacy. Each institution trains the model locally, and the cloud platform updates the joint model parameters to ensure a balance between data privacy and sharing. This solves the contradiction between knowledge sharing and privacy protection in the existing technology.

[0086] (4) The present invention comprehensively solves the deficiencies in the existing technology through innovative features such as intelligent analysis and information extraction driven by deep learning, intelligent query and decision support based on knowledge graph, and cross-organizational knowledge sharing and privacy protection driven by federated learning, and provides an efficient, intelligent and reliable knowledge base construction method for railway engineering design. BRIEF DESCRIPTION OF THE DRAWINGS

[0087] The present invention is further described using the accompanying drawings, but the embodiments in the accompanying drawings do not constitute any limitation to the present invention. A person skilled in the art can obtain other drawings based on the following drawings without creative work.

[0088] Figure 1 This is a flow chart of a method for constructing a railway engineering design knowledge base based on a cloud platform according to the present invention. DETAILED DESCRIPTION

[0089] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0090] like Figure 1 As shown, a method for constructing a railway engineering design knowledge base based on a cloud platform is provided in an embodiment of the present invention, and is applied to a cloud platform. The method includes the following steps S1-S6:

[0091] S1. Utilize a cloud platform to collect multimodal data and preprocess the multimodal data, design a distributed file system to store the preprocessed data, and establish a data indexing mechanism to retrieve the multimodal data; wherein the multimodal data includes text data, image data, and video data.

[0092] Specifically, railway engineering design data includes text, drawings, videos and other forms, which need to be stored and managed in a unified manner. First, multimodal data is collected from different sources, including design documents (text), engineering drawings (images), and construction videos (videos). These data sources can be design software, document management systems, and on-site monitoring equipment.

[0093] For text data, word segmentation, stop word removal, and stem extraction are performed. For image data, grayscale, denoising, and edge detection are performed. For example, the image data denoising processing formula is:

[0094] ;

[0095] in, represents the denoised image; Indicates that the original image is at position The pixel value at ; Represents the weight matrix used for filtering; k represents half of the filter window size.

[0096] Furthermore, a distributed storage system is designed that can support the storage and fast retrieval of large-scale data. A distributed file system (such as HDFS) is used to store large-scale multimodal data. The system should have high reliability and high availability, and be able to automatically back up and restore data. A metadata management system (such as HBase) is used to store and manage the metadata of multimodal data, including file name, storage path, data type, timestamp, and other information.

[0097] Furthermore, in order to achieve fast data retrieval, an efficient data indexing mechanism is established. For text data, inverted indexing technology is used to index and support fast keyword retrieval.

[0098] Extract feature vectors (such as SIFT, SURF) from image data and use kd tree or LSH (Locality-Sensitive Hashing) for indexing to support fast image retrieval. For example, the formula for extracting image feature vectors is as follows:

[0099] ;

[0100] Among them, f represents the feature vector of the image; P represents the set of all key points in the image; h(p) represents the feature histogram of key point p; d(p) represents the feature descriptor of key point p.

[0101] Feature extraction and indexing of key frames of video data support fast content-based video retrieval.

[0102] S2. Use deep learning models to intelligently analyze these multimodal data, extract key information and technical parameters of railway engineering design, and then fuse the feature vectors of multimodal data to generate a comprehensive feature representation; use pre-trained language models to extract and analyze text data; use convolutional neural networks to extract and analyze image data; and use three-dimensional convolutional neural networks to extract and analyze video data.

[0103] Specifically, use the pre-trained language model to parse the text data and extract key information and technical parameters. The specific steps are as follows:

[0104] Tokenize the preprocessed text data to obtain a Token sequence ;

[0105] Input the Token sequence into the pre-trained language model to obtain the implicit representation of each Token :

[0106] ;

[0107] Use linear layer and Softmax layer to classify the implicit representation and extract key information K:

[0108] ;

[0109] Among them, W and b are the weight matrix and bias vector respectively.

[0110] In order to ensure that the extracted information is more representative and accurate, a sparse regularization term is introduced to penalize redundant information:

[0111] ;

[0112] in, is the true label, is the predicted probability, is the regularization coefficient, is the L1 norm of the implicit representation.

[0113] In addition, convolutional neural network (CNN) is used to extract and analyze image data. The specific steps are as follows:

[0114] Perform convolution operation on the preprocessed image data to obtain the feature map F:

[0115] ;

[0116] in, is the input image.

[0117] Perform a fully connected layer operation on the feature map to obtain the image feature vector :

[0118] ;

[0119] In order to enhance the distinguishability of features, the center loss (CenterLoss) is introduced to minimize the distance of features of the same type:

[0120] ;

[0121] in, is the image feature vector of the i-th sample, is the corresponding category center, and m is the number of samples.

[0122] In addition, a three-dimensional convolutional neural network (3D-CNN) is used to extract and analyze video data. The specific steps are as follows:

[0123] Perform a three-dimensional convolution operation on the preprocessed video data to obtain the feature map :

[0124] ;

[0125] Among them, V is the input video.

[0126] Perform a fully connected layer operation on the feature map to obtain the video feature vector :

[0127] ;

[0128] In order to capture the information of key frames in the video, the inter-frame similarity regularization term is introduced to ensure that the features of adjacent frames are similar:

[0129] ;

[0130] in, is the feature vector of the tth frame, and T is the total number of frames.

[0131] Furthermore, the feature vectors of text, image and video data are fused to generate a comprehensive feature representation The specific steps are as follows:

[0132] Weighted fusion of feature vectors of text, image and video:

[0133] ;

[0134] in, is the text feature vector, is the image feature vector, is the video feature vector, is the fusion weight coefficient.

[0135] The vector space model (VSM) is then used to represent the multimodal data, and the data association and retrieval are achieved by calculating the similarity between the multimodal data.

[0136] S3. Identify and extract entities and relationships in railway engineering design from the feature vectors of multimodal data to construct a knowledge graph, and then use a graph neural network to train the knowledge graph.

[0137] Among them, from the multimodal feature data (text, image, video) extracted in step S2, the entities and relationships in railway design are identified and extracted. These entities and relationships are represented as a graph structure, where nodes represent entities and edges represent relationships. The specific steps are as follows:

[0138] Extract entities (such as "bridge" and "tunnel") and relationships (such as "connection" and "inclusion") from the text and build a preliminary graph structure , where V represents the node set and E represents the edge set.

[0139] Extract and parse features from image and video data, identify design elements, align them with text entities, and update graph structures.

[0140] Use graph embedding methods to represent nodes and edges. Assume that node The characteristic is expressed as ,side The characteristic is expressed as .

[0141] Furthermore, in order to optimize the representation of the knowledge graph, a graph neural network (GNN) is used to train the knowledge graph. The specific steps are as follows:

[0142] Initialize the feature representation of nodes and edges. The initial eigenvector of The multimodal feature fusion result from step S2.

[0143] During each layer of propagation, the feature representation of a node is updated based on the features of its neighboring nodes:

[0144] ;

[0145] in, For Node In the The feature representation of the layer, and are the weight matrix and bias vector respectively, Representation Node The set of neighbor nodes of and Node and The degree, is the activation function;

[0146] Introduce edge weight-based regularization to ensure the optimization of edge features:

[0147] ;

[0148] in, is the edge weight regularization term, For edge The weight of For Node and characteristic differences.

[0149] Furthermore, based on the trained knowledge graph, an intelligent query system is developed to realize automated design suggestions and optimization solutions. The specific steps are as follows:

[0150] The user inputs the design requirements, and the system queries through the knowledge graph to match relevant design elements and relationships.

[0151] Utilize the reasoning capabilities of graph neural networks to calculate the similarity between design requirements and the nodes and edges in the graph and generate optimization suggestions.

[0152] In order to improve the accuracy of query results, context similarity calculation is introduced:

[0153] ;

[0154] in, For Node The similarity with query q, is the feature representation of the query, represents the vector inner product, Represents the vector norm.

[0155] Through the above steps, the knowledge representation and association in railway engineering design are realized.

[0156] S4. Develop an intelligent query system based on the trained knowledge graph to provide automated design suggestions and optimization solutions, including the following steps:

[0157] S401, the user inputs the design requirements, and the system needs to convert the natural language requirements into feature vector representation. The specific steps are as follows:

[0158] Tokenize the text input by the user to get a token sequence ;

[0159] Use a pre-trained language model (such as BERT) to process the token sequence and obtain the implicit representation of each token :

[0160] ;

[0161] in, An implicit representation of the requirement, Represents the token sequence entered by the user.

[0162] The implicit representation is averaged and pooled to obtain the feature vector of the user's needs. :

[0163] ;

[0164] S402. Feature vector based on user needs , query in the knowledge graph and match relevant design elements and relationships. The specific steps are as follows:

[0165] Calculate user demand feature vector And each node feature vector in the knowledge graph Similarity:

[0166] ;

[0167] in, is the similarity between the demand vector and the node vector, represents the vector inner product, Represents the vector norm.

[0168] S403, sort the nodes according to the similarity values, and select the top k nodes with the highest similarity as candidate design elements. Generate design suggestions and optimization solutions based on the query results. The specific steps are as follows:

[0169] The graph neural network is used for reasoning and the relationship weights between candidate nodes and their neighbor nodes are calculated:

[0170] ;

[0171] in, For Node In the The feature representation of the layer, and are the weight matrix and bias vector respectively, Representation Node The set of neighbor nodes of and Respectively represent nodes and The degree, is the activation function;

[0172] Calculate the node feature vector after inference User demand feature vector The final similarity is:

[0173] ;

[0174] in, is the final similarity, is the optimized node feature vector, represents the vector inner product, represents the vector norm;

[0175] The candidate nodes are sorted according to the final similarity values, and the nodes with the highest similarity and their associated design elements are selected to generate optimization suggestions.

[0176] S404: In order to improve the intelligence of the system and user satisfaction, the system needs to make adaptive adjustments based on user feedback. The specific steps are as follows:

[0177] Users provide feedback on design suggestions and the system records the feedback information.

[0178] Based on the feedback information, the parameters of the graph neural network are adjusted so that the system can better meet user needs.

[0179] Introduce feedback adjustment terms and update the loss function of the graph neural network:

[0180] ;

[0181] in, is the loss function after introducing the feedback adjustment term, is the original loss function, For user feedback, is the current node feature vector, is the target node feature vector, is the feedback adjustment coefficient.

[0182] S5. Build a cross-railway engineering design agency sharing model based on federated learning to carry out knowledge sharing and privacy protection among multiple railway design agencies.

[0183] Specifically, each participating organization (such as different railway design agencies) uses its data to train the model locally. The specific steps are as follows:

[0184] Each institution extracts features and labels from its local data to build a training set ;

[0185] Initialize local model parameters ;

[0186] Use stochastic gradient descent (SGD) to train the model locally and update the parameters:

[0187] ;

[0188] in, represents the model parameters of the kth iteration in the tth round of training, is the learning rate, is the loss function, and are the input features and labels respectively.

[0189] Furthermore, in order to further protect data privacy, a differential privacy mechanism is introduced in the local training and parameter sharing process. The specific steps are as follows:

[0190] Before sending model parameters, each organization adds noise to achieve differential privacy protection:

[0191] ;

[0192] in, are the model parameters after adding noise, The mean is 0 and the variance is Gaussian noise.

[0193] Furthermore, each organization sends the locally trained model parameters to the central server for aggregation. The specific steps are as follows:

[0194] The central server receives local model parameters sent by each organization ;

[0195] The central server performs weighted averaging on the received parameters to obtain the global model parameters :

[0196] ;

[0197] in, is the global model parameter after aggregation, M is the number of participating organizations, is the number of samples of the jth organization, is the total sample size.

[0198] Furthermore, in order to further ensure the privacy protection effect of model parameters, the present invention introduces a privacy protection regularization term into the loss function of federated learning:

[0199] Add regularization term during local training , to limit the range of variation of model parameters:

[0200] ;

[0201] in, is the total loss function for local training, is the regularization coefficient, is the regularization term.

[0202] Defining privacy-preserving regularization terms , so that it penalizes when the global model parameters change too much:

[0203] ;

[0204] in, are the initial model parameters, is the L2 norm.

[0205] Furthermore, the central server distributes the aggregated global model parameters to each organization, and each organization uses the updated global model for the next round of training. The specific steps are as follows:

[0206] The central server sends the global model parameters Sent to all participating organizations.

[0207] After receiving the global model parameters, each organization uses the parameters to perform the next round of local training and repeats the above steps.

[0208] S6. Build a simulation environment for railway engineering design, then design a reinforcement learning model to train the simulation environment, automatically generate railway engineering design plans based on the trained reinforcement model, and perform evaluation and optimization.

[0209] Specifically, the steps include:

[0210] First, a simulation environment is built, which can simulate the actual railway engineering design process. The specific steps are as follows:

[0211] Defining the state space , contains all possible design states. For example, the state Indicates the design status at time t, including design parameters, progress and other information;

[0212] Defining the action space , which contains all possible design decisions. For example, the action represents the design decision chosen at time t, such as adjusting design parameters, modifying the construction plan, etc.;

[0213] Defining the reward function , which is used to evaluate the quality of each design decision. The reward function can take into account factors such as design quality, cost, and time.

[0214] Furthermore, a deep reinforcement learning algorithm is used to train the design simulation environment. The specific steps are as follows:

[0215] Initialize the parameters of the Q network and the parameters of the target Q network , and set up experience replay pool D.

[0216] For each time step t, in state Next, select Action (use Greedy strategy), execute action , observe the reward and the next state , storage experience To D.

[0217] Randomly sample a mini-batch from D , calculate the target Q value:

[0218] ;

[0219] in, is the discount factor, is the Q value of the target Q network.

[0220] Then use the gradient descent method to minimize the loss function and update the parameters of the Q network :

[0221] ;

[0222] Periodically change the parameters of the Q network Parameters copied to the target Q network .

[0223] Furthermore, in the process of reinforcement learning, design optimization often needs to meet multiple objectives and constraints. To achieve this, the present invention introduces a multi-objective optimization and constraint processing mechanism. The specific steps are as follows:

[0224] Define a multi-objective reward function that takes into account multiple design goals (such as quality, cost, time, etc.):

[0225] ;

[0226] in, is the reward function of the ith target, is the corresponding weight coefficient.

[0227] Introduce constraint processing terms into the loss function to ensure that the optimization process meets the design constraints:

[0228] ;

[0229] in, is the loss function with constraint processing, is the kth constraint, Represents the constraint penalty coefficient.

[0230] Furthermore, in order to further optimize the design solution, a dynamic adjustment and adaptive mechanism is introduced. The specific steps are as follows:

[0231] During training, the reward function is dynamically adjusted based on real-time feedback The weights of , ensure that the reward function accurately reflects the current design goals and constraints.

[0232] Use an adaptive optimization algorithm (such as adaptive time-varying target optimization) to update the target value of the Q network to improve the convergence speed and optimization effect of the model.

[0233] Introducing reward adjustment mechanisms based on uncommon regularization terms, such as entropy-based rewards, encourages the model to explore more possible design solutions:

[0234] ;

[0235] in, For strategy In Status The entropy under is the regularization coefficient.

[0236] Furthermore, for design scheme generation and evaluation, the specific steps are as follows: based on the trained reinforcement learning model, the design scheme is automatically generated, evaluated and optimized.

[0237] Execute the trained policy in simulation , generate the optimal design solution .

[0238] The generated design scheme is simulated and evaluated multiple times, and its design quality, cost, time and other indicators are recorded.

[0239] Based on the evaluation results, the design scheme is further optimized to ensure that it meets actual application requirements and multiple constraints.

[0240] It should be noted that the workflow described above is merely illustrative and does not limit the scope of protection of the present invention. In practical applications, technicians in this field can select part or all of them according to actual needs to achieve the purpose of the present embodiment, and no limitation is made here.

[0241] In addition, for technical details that are not described in detail in this embodiment, reference can be made to the parameter operation method provided in any embodiment of the present invention, and will not be repeated here.

[0242] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or system including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or system. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or system including the element.

[0243] The serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments.

[0244] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an 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, which is stored in a storage medium (such as a read-only memory / random access memory, a magnetic disk, or an optical disk), and includes a number of instructions for enabling a terminal device (which can be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) to execute the methods described in each embodiment of the present invention.

[0245] The above are only preferred embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A method for constructing a railway engineering design knowledge base based on a cloud platform, applied to a cloud platform, characterized in that: The method comprises: S1. Using a cloud platform to collect multimodal data and preprocess the multimodal data, designing a distributed file system to store the preprocessed data, and establishing a data indexing mechanism to retrieve the multimodal data; wherein the multimodal data includes text data, image data, and video data; S2. Use deep learning models to intelligently analyze these multimodal data, extract key information and technical parameters of railway engineering design, and then fuse the feature vectors of multimodal data to generate a comprehensive feature representation; use pre-trained language models to extract and analyze text data; use convolutional neural networks to extract and analyze image data; and use three-dimensional convolutional neural networks to extract and analyze video data; S3, identify and extract entities and relationships in railway engineering design from the feature vectors of multimodal data to construct a knowledge graph, and then use a graph neural network to train the knowledge graph; S4. Develop an intelligent query system based on the trained knowledge graph to provide automated design suggestions and optimization solutions, including: S401, the user inputs the design requirements, and the intelligent query system uses the pre-trained language model to convert the natural language requirements into feature vector representation; S402, querying the knowledge graph according to the feature vector representation, calculating the similarity between the user demand feature vector and the feature vector of each node in the knowledge graph, and matching the design elements and relationships; S403, performing reasoning through a graph neural network based on the query results, using similarity to calculate the relationship weight between the candidate node and its neighbor nodes to obtain a final similarity, sorting the candidate nodes according to the final similarity value, and selecting the nodes with the highest similarity and their associated design elements to generate optimization suggestions; S404: Collect user feedback on design suggestions, record feedback information, adjust the parameters of the graph neural network based on the feedback information, introduce feedback adjustment items, and update the loss function of the graph neural network, which is expressed as follows: ; in, is the loss function after introducing the feedback adjustment term, is the original loss function, For user feedback, is the current node feature vector, is the target node feature vector, is the feedback adjustment coefficient; S5. Construct a cross-railway engineering design agency sharing model based on federated learning to share knowledge and protect privacy among multiple railway design agencies; S6. Build a simulation environment for railway engineering design, then design a reinforcement learning model to train the simulation environment, automatically generate railway engineering design plans based on the trained reinforcement model, and perform evaluation and optimization.

2. A method for constructing a railway engineering design knowledge base based on a cloud platform according to claim 1, characterized in that: The preprocessing includes: performing word segmentation, stop word removal and stem extraction on text data, performing grayscale, denoising and edge detection on image data, and performing frame segmentation on video data; the data indexing mechanism includes: indexing text data using inverted indexing technology, extracting feature vectors from image data and indexing using kd tree or LSH, and performing feature extraction and indexing on key frames of video data.

3. The method for constructing a railway engineering design knowledge base based on a cloud platform according to claim 1, characterized in that: The construction of the knowledge graph includes: Extract entities and relationships from text features and build a preliminary graph structure , where V represents the node set and E represents the edge set. The image features and video features are aligned with the text entities, the graph structure is updated, and the graph embedding method is used to represent the features of the nodes and edges.

4. The method for constructing a railway engineering design knowledge base based on a cloud platform according to claim 3 is characterized in that: The method of using a graph neural network to train a knowledge graph includes: Initialize the feature representation of nodes and edges, where nodes The initial eigenvector of It is the result of multimodal feature fusion; During each layer of propagation, the feature representation of a node is updated according to the features of its neighboring nodes, as shown below: ; in, Representation Node In the The feature representation of the layer, and denote the weight matrix and bias vector respectively, Representation Node The set of neighbor nodes of and Respectively represent nodes and The degree, represents the activation function; A regularization term based on edge weight is introduced to ensure the optimization of edge features, which is expressed as follows: ; in, represents the edge weight regularization term, Represents edge The weight of Representation Node and characteristic differences.

5. The method for constructing a railway engineering design knowledge base based on a cloud platform according to claim 1, characterized in that: In S401, the text input by the user is tokenized to obtain a token sequence. , use the pre-trained language model to process the Token sequence and obtain the implicit representation of each Token , will implicitly represent After the average pooling operation, the feature vector of user demand is obtained ; In S402, the user demand feature vector is calculated And each node feature vector in the knowledge graph The similarity is expressed as follows: ; in, represents the user demand feature vector, Represents the feature vector of each node in the knowledge graph, is the similarity between the demand vector and the node vector, represents the vector inner product, Represents the vector norm.

6. A method for constructing a railway engineering design knowledge base based on a cloud platform according to claim 5, characterized in that: Generate optimization suggestions based on the nodes with the highest similarity and their associated design elements, including the following steps: The graph neural network is used for reasoning and the relationship weights between candidate nodes and their neighbor nodes are calculated, which can be expressed as follows: ; in, Representation Node In the The feature representation of the layer, and denote the weight matrix and bias vector respectively, Representation Node The set of neighbor nodes of and Respectively represent nodes and The degree, represents the activation function; Calculate the node feature vector after inference User demand feature vector The final similarity is expressed as follows: ； in, is the final similarity, is the optimized node feature vector, represents the vector inner product, represents the vector norm; The candidate nodes are sorted according to the final similarity values, and the nodes with the highest similarity and their associated design elements are selected to generate optimization suggestions.

7. The method for constructing a railway engineering design knowledge base based on a cloud platform according to claim 1, characterized in that: The S5 specifically includes: S501. Extract features and labels from local data of each institution and build a training set , initialize local model parameters , using stochastic gradient descent locally using the training set Train the local model and update the parameters; S502: Introduce a differential privacy mechanism to protect data privacy during local training and parameter sharing, and obtain the model parameters after adding the differential privacy mechanism ; S503. Each organization will use the model parameters trained locally Sent to the central server for aggregation to obtain global model parameters ; S504, introducing a privacy-preserving regularization term into the loss function of the global model to limit the range of variation of the global model parameters; S505: The central server aggregates the global model parameters Distribute to each organization, and each organization uses the updated global model to perform the next round of local training, and repeat S501 to S504.

8. The method for constructing a railway engineering design knowledge base based on a cloud platform according to claim 7, characterized in that: The stochastic gradient descent method is used locally to train the Training the local model is as follows: ; in, represents the model parameters of the kth iteration in the tth round of training, represents the learning rate, represents the loss function, and Represent the input features and labels respectively; The S504 specifically includes: Add regularization term during local training , limiting the range of changes in model parameters shared across railway engineering design organizations, as shown below: ; in, represents the total loss function of local training, represents the regularization coefficient, represents the regularization term, and denote input features and labels respectively, represents the model parameters of the kth iteration in the tth round of training; Defining privacy-preserving regularization terms , so that it is penalized when the global model parameters change too much, as shown below: ; Among them, w represents the current model parameters, are the initial model parameters, is the L2 norm.

9. The method for constructing a railway engineering design knowledge base based on a cloud platform according to claim 1, characterized in that: The S6 specifically includes: S601. Construct a simulation environment based on the actual railway engineering design process, including: defining the state space , including all design states; defining the action space , which contains all design decisions; defines the reward function , used to evaluate the quality of each design decision; S602. Build a deep reinforcement learning model to train the design simulation environment and obtain a trained strategy ; S603, according to the trained strategy , generate the optimal design solution ,Then the generated design scheme is simulated and evaluated multiple times, and the design scheme is optimized based on the evaluation results.

10. A method for constructing a railway engineering design knowledge base based on a cloud platform according to claim 9, characterized in that: The S602 includes: Initialize the parameters of the Q network and the parameters of the target Q network , and set the experience replay pool D, for each time step t, in state Next, select Action , perform the action , observe the reward and the next state , storage experience into D, and then randomly sample a mini-batch from D , calculate the target Q value, expressed as follows: ; in, represents the discount factor, represents the Q value of the target Q network; Minimize the loss function using gradient descent , update the parameters of the Q network , which is expressed as follows: ; In the periodic Q network parameters Parameters copied to the target Q network ; Then define the multi-objective reward function, taking into account multiple design goals, as follows: ; in, represents the reward function of the ith target, represents the corresponding weight coefficient; In the loss function The constraint processing item is introduced to ensure that the optimization process meets the design constraints, which is expressed as follows: ; in, represents the loss function with constraint processing, represents the kth constraint, represents the constraint penalty coefficient; Then dynamically adjust the reward function based on real-time feedback The weight of the Q network is then updated using an adaptive optimization algorithm, and a reward adjustment mechanism based on uncommon regularization terms is introduced to make the reinforcement learning model calculation design scheme as follows: ; in, Representation strategy In Status The entropy under represents the regularization coefficient.

Images