A method and device for constructing a high-speed rail technology evolution map

Abstract

The application discloses a kind of construction methods and devices of high-speed rail technology evolution graph.The construction method includes: obtaining patent data set related to high-speed rail technology and classifying according to technical field, obtain multiple technical subsets;Based on PatentsBERTa model and cosine similarity calculation formula, obtain multiple similarity data sets;Obtain the similarity threshold of each technical subset, and based on the similarity threshold, obtain multiple candidate patent groups corresponding to each technical subset respectively;Based on the candidate patent group corresponding to each technical subset, build the time evolution graph and citation network graph corresponding to each technical subset, and convert into network data using graph projection, construct multiple evolution knowledge graphs;Obtain the target technology of pre-set, and based on the evolution knowledge graph corresponding to the technical subset of target technology, construct the evolution subgraph of target technology.The construction method provided by the application can identify the evolution path of each sub-technology in high-speed rail technology.

Description

Technical Field

[0001] This invention relates to the field of technical path implementation methods, specifically to a method and apparatus for constructing a high-speed rail technology evolution map. Background Technology

[0002] In recent years, high-speed rail technology has developed rapidly, with its complexity and diversity increasing year by year. However, this has also given rise to problems such as unclear technology types and difficulty in identifying technological development paths. Currently, there is a lack of a comprehensive and systematic analytical framework for the evolution of high-speed rail technology. Existing research mostly focuses on independent sub-fields, ignoring the cross-correlation between different technological dimensions. Traditional methods rely on expert subjective judgment and manual classification of patent information, which is labor-intensive and difficult to maintain label consistency and repeatability. Faced with massive amounts of patent data, the "manual + qualitative" model is inadequate and easily obscures potential technological features and innovation nodes. At the same time, traditional methods rely solely on patent citation networks to trace development, but the citation motivation may be influenced by patent layout strategies or legal considerations, making it difficult to truly reflect the intrinsic connections between technologies, and the deeper information in the patent description abstract is not effectively mined. Summary of the Invention

[0003] The purpose of this invention is to provide a method for identifying the evolution path of each sub-technology by effectively mining the deep information in the patent description abstract, which is a complex system of high-speed rail technology that includes many sub-technology, so as to solve the problems raised in the background art.

[0004] To achieve the above objectives, according to a first aspect of the present invention, the present invention provides a method for constructing a high-speed rail technology evolution map, comprising the following steps: Step S10: Obtaining a patent dataset related to high-speed rail technology, classifying the patents in the patent dataset according to the technical field to obtain multiple technology subsets, each technology subset including multiple patents; Step S20: Calculating the similarity value of a specific patent pair based on the PatentsBERTa model and the cosine similarity calculation formula, obtaining multiple similarity datasets corresponding one-to-one with the multiple technology subsets, each similarity dataset including multiple similarity values, the multiple similarity values ​​corresponding one-to-one with multiple specific patent pairs in the same technology subset, wherein the specific patent pair consists of two patents with a direct reference relationship; Step S30: Based on the similarity dataset corresponding to each technology subset, obtaining the similarity threshold of each technology subset, and based on each technology subset... Step S40: Based on the candidate patent groups corresponding to each technology subset, multiple candidate patent groups are obtained. Each candidate patent group consists of multiple specific patent pairs with similarity values ​​greater than the similarity threshold. Step S50: Based on the candidate patent groups corresponding to each technology subset, a temporal evolution graph and a citation network graph based on citation relationships are constructed for each technology subset using Neo4j. The graph triplet relationships are converted into network data using graph projection to construct multiple evolutionary knowledge graphs that correspond one-to-one with multiple technology subsets. Step S50: A preset target technology is obtained. Based on the evolutionary knowledge graph corresponding to the technology subset to which the target technology belongs, an evolutionary subgraph of the target technology is constructed. The evolutionary subgraph uses the patent with the earliest application date among the patents related to the target technology as the seed node and is formed by all subsequent patents that have direct and indirect citation relationships with the seed node, extracted along the evolutionary and temporal relationships.

[0005] In one specific implementation, the evolutionary subgraph of the target technology consists of multiple nodes and relationship lines connecting two nodes with direct reference relationships. The multiple nodes include a seed node and multiple non-seed nodes. The construction method further includes: step (a) calculating the betweenness centrality score of each non-seed node in the evolutionary subgraph of the target technology based on the betweenness centrality calculation formula of the nodes, and performing normalization processing to obtain multiple betweenness centrality scores; step (b) sorting all betweenness centrality scores in descending order, and selecting at least one bridge node group for stage division from the M target nodes with the largest betweenness centrality scores, each bridge node group including one or more target nodes, where M is a positive integer and 5≤M≤10; step (c) dividing the evolutionary subgraph of the target technology into stages based on the at least one bridge node group obtained in step (b), to obtain an evolutionary subgraph divided into multiple development stages.

[0006] In one specific implementation, at least one bridge node group for phase division is selected based on the patent application dates of M target nodes.

[0007] In one specific implementation, the plurality of said technology subsets include a train technology subset, an infrastructure technology subset, a train control technology subset, a digital and intelligent technology subset, a testing and experimentation technology subset, and a magnetic levitation technology subset.

[0008] In one specific implementation, step S10 includes: step S11, obtaining invention patents with patent application dates from January 1, 1986 to December 31, 2024 based on the Incopat patent database through keyword retrieval, thus obtaining a patent dataset related to high-speed rail technology; step S12, annotating some patents in the patent dataset using ChatGPT-4o and manual verification, determining the technical field of each annotated patent, obtaining an annotated dataset including multiple annotated samples, and collecting the specification abstracts of unannotated patents as unannotated samples to obtain an unannotated dataset, wherein the annotated... Each labeled sample in the labeled dataset includes a specification summary and a technical field determined based on the specification summary; Step S13: Train the RoBERTa_BiLSTM model using the labeled dataset to obtain a trained RoBERTa_BiLSTM model, and input the unlabeled dataset into the trained RoBERTa_BiLSTM model for classification to obtain the technical field of each unlabeled sample in the unlabeled dataset; Step S14: Based on the technical field corresponding to each patent in the patent dataset, collect all patents in the same technical field to obtain multiple technical subsets that correspond one-to-one with multiple technical fields.

[0009] In one specific implementation, step S10 further includes a step of selecting the best classification model prior to step S13. The best classification model is the RoBERTa_BiLSTM model. The step of selecting the best classification model includes: step (a) splitting the labeled dataset into a training set and a validation set, and using the training set to train the BERT_uncased model, RoBERTa_base_cased model, and RoBERTa_BiLSTM model respectively, to obtain the trained BERT_uncased model, RoBERTa_base_cased model, and RoBERTa_BiLSTM model; step (b) using the validation set to evaluate the trained BERT_uncased model, RoBERTa_base_cased model, and RoBERTa_BiLSTM model, to obtain the evaluation metric for each model, and taking the model with the best evaluation metric as the best classification model. The evaluation metric includes accuracy, precision, recall, and F1 score.

[0010] In one specific implementation, step S20, which involves obtaining the similarity dataset corresponding to any subset of technologies based on the PatentsBERTa model and the cosine similarity calculation formula, includes: step (1) determining whether there is a direct reference relationship between any two patents in the subset of technologies, and forming a specific patent pair with the two patents having a direct reference relationship to obtain multiple specific patent pairs; step (2) calculating the similarity value of each specific patent pair one by one according to the order of the application dates of the referenced patents in each specific patent pair, based on the PatentsBERTa model and the cosine similarity calculation formula, to obtain the similarity dataset, wherein the similarity dataset includes multiple similarity values ​​corresponding one-to-one with multiple specific patent pairs.

[0011] In one specific implementation, the step of calculating the similarity value of a specific patent pair based on the PatentsBERTa model and the cosine similarity calculation formula includes: using the PatentsBERTa model to convert the specification summaries of the two patents in the specific patent pair into high-dimensional classification vectors, obtaining two high-dimensional classification vectors corresponding to the specific patent pair. Using the two high-dimensional classification vectors corresponding to the specific patent pair as variables, the similarity value corresponding to the specific patent pair is calculated using the cosine similarity calculation formula.

[0012] In one specific implementation, step S50 includes: Step A, obtaining a preset target technology, and determining the technology subset to which the target technology belongs based on the preset target technology; Step B, using Neo4j's query language to search for target patents that meet the query conditions in the evolutionary knowledge graph corresponding to the technology subset to which the target technology belongs, wherein the number of target patents is multiple; Step C, taking the earliest applied patent among the multiple target patents as a seed node, extracting all subsequent patents that have direct and indirect reference relationships with the seed node along the evolutionary and temporal relationships, forming an evolutionary subgraph of the target technology.

[0013] According to a second aspect of the present invention, the present invention provides an apparatus for constructing a high-speed rail technology evolution map, the apparatus comprising: a classification module, configured to acquire a patent dataset related to high-speed rail technology, classify the patents in the patent dataset according to the technical field to obtain multiple technology subsets, each technology subset including multiple patents; a calculation module, configured to calculate the similarity value of a specific patent pair based on the PatentsBERTa model and the cosine similarity calculation formula, acquire multiple similarity datasets corresponding one-to-one with the multiple technology subsets, each similarity dataset including multiple similarity values, the multiple similarity values ​​corresponding one-to-one with multiple specific patent pairs in the same technology subset, wherein the specific patent pair consists of two patents with a direct reference relationship; and an acquisition module, configured to acquire a similarity threshold for each technology subset based on the similarity dataset corresponding to each technology subset, and based on the similarity of each technology subset... The system employs a threshold-based approach to acquire multiple candidate patent groups corresponding to each technology subset. Each candidate patent group consists of multiple specific patent pairs with similarity values ​​exceeding a similarity threshold. A first construction module, based on the candidate patent groups corresponding to each technology subset, uses Neo4j to build a temporal evolution graph and a citation network graph based on citation relationships for each technology subset. Graph projection is then used to convert graph triplet relationships into network data, constructing multiple evolutionary knowledge graphs that correspond one-to-one with the multiple technology subsets. A second construction module acquires a preset target technology and, based on the evolutionary knowledge graph corresponding to the technology subset to which the target technology belongs, constructs an evolutionary subgraph of the target technology. This evolutionary subgraph uses the patent with the earliest application date among the patents related to the target technology as the seed node, and is formed by all subsequent patents extracted along evolutionary and temporal relationships that have direct and indirect citation relationships with the seed node.

[0014] Compared with the prior art, the present invention has the following beneficial effects:

[0015] I. This invention provides a method for constructing a high-speed rail technology evolution map. This method, which addresses the complexity and numerous sub-technologies of high-speed rail technology, can identify the evolution paths of each sub-technology, providing operational guidance for the innovation and development of high-speed rail technology, decision support for the coordinated progress of related industrial chains, and a practical basis for government departments to formulate technology policies and strategic plans.

[0016] Second, this invention first uses the PatentsBERTa model to convert the specification summaries of the two patents in a specific patent pair into high-dimensional classification vectors; then, it uses the cosine similarity calculation formula to calculate the similarity value corresponding to the specific patent pair; next, it uses the arithmetic mean of the similarity values ​​of multiple specific patent pairs with direct citation relationships in each technology subset as the similarity threshold of each technology subset, and uses the similarity threshold as a screening criterion to obtain a candidate patent group; then, based on the candidate patent group, it uses Neo4j to build a time-series evolution graph of high-speed rail technology and a citation network graph based on citation relationships, and combines the two to obtain an evolutionary knowledge graph; compared with the traditional citation network, the evolutionary knowledge graph based on fused semantic features has a moderate degree of modularity and higher technology dissemination efficiency.

[0017] In addition to the objectives, features, and advantages described above, the present invention has other objectives, features, and advantages. The invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate embodiments of the present invention and form part of the specification. They are used together with the following detailed description to explain the embodiments of the present invention, but do not constitute a limitation thereof. In the drawings:

[0019] Figure 1 This is a flowchart illustrating the steps of a method for constructing a high-speed rail technology evolution map according to an embodiment of the present invention;

[0020] Figure 2 This is a flowchart illustrating sub-step S10 of a method for constructing a high-speed rail technology evolution map according to an embodiment of the present invention.

[0021] Figure 3 A diagram illustrating the triplet relationships;

[0022] Figure 4 Evolutionary knowledge graphs corresponding to the six technology subsets constructed for this invention;

[0023] Figure 5 A module diagram of a device for constructing a high-speed rail technology evolution map, provided in another embodiment of the present invention;

[0024] Figure 6This is a sub-diagram of the evolution of brake disc technology constructed in Embodiment 1 of the present invention. Detailed Implementation

[0025] The present invention will now be described in detail with reference to the embodiments shown in the accompanying drawings.

[0026] Please see Figure 1 and Figure 2 In a first aspect, the present invention provides a method for constructing a high-speed rail technology evolution map, the method comprising the following steps:

[0027] Step S10: Obtain a patent dataset related to high-speed rail technology, and classify the patents in the patent dataset according to the technical field to obtain multiple technology subsets, each of which includes multiple patents.

[0028] In this invention, the plurality of said technology subsets include a train technology subset, an infrastructure technology subset, a train control technology subset, a digital and intelligent technology subset, a testing and experimentation technology subset, and a magnetic levitation technology subset.

[0029] The patent dataset will be divided into six categories based on technical fields: train technology, infrastructure technology, train control technology, digitalization and intelligent technology, testing and experimental technology, and maglev technology.

[0030] In this invention, all patents in the patent dataset are invention patents, and the patent application dates range from January 1, 1986 to December 31, 2024. Specifically, invention patents related to my country's high-speed rail from 1986 to 2024 were collected from the Incopat patent database, and the application numbers were merged.

[0031] In this invention, the patent dataset includes a total of 9271 invention patents, of which the train technology subset includes 2173 invention patents, the infrastructure technology subset includes 3709 invention patents, the train control technology subset includes 1285 invention patents, the digital and intelligent technology subset includes 773 invention patents, the testing and experimental technology subset includes 1175 invention patents, and the magnetic levitation technology subset includes 121 invention patents.

[0032] In one optional implementation, step S10 includes:

[0033] Step S11: Based on the Incopat patent database, obtain invention patents with patent application dates from January 1, 1986 to December 31, 2024 through keyword search, and obtain a patent dataset related to high-speed rail technology.

[0034] In this invention, the search keywords include high-speed railway, high-speed train, high-speed railway station, high-speed passenger train, passenger transport, etc.

[0035] Step S12: Combine ChatGPT-4o and manual verification to annotate some patents in the patent dataset, determine the technical field of each annotated patent, obtain an annotated dataset including multiple annotated samples, and collect the specification abstracts of unannotated patents as unannotated samples to obtain an unannotated dataset. Each annotated sample in the annotated dataset includes a specification abstract and a technical field determined based on the specification abstract.

[0036] In this invention, the tagging of a portion of the patents in the patent dataset by combining ChatGPT-4o and manual verification specifically involves: based on the patent specification abstract, ChatGPT-4o is used to annotate the technical fields of a portion of the patents in the patent dataset; after the annotation is completed, manual verification is performed to determine the technical field of each annotated patent.

[0037] In this invention, each labeled sample in the labeled dataset includes a specification summary and a technical field determined based on the specification summary, while each unlabeled sample in the unlabeled dataset includes only a specification summary.

[0038] In one alternative implementation, the ratio of the labeled dataset to the unlabeled dataset is 7:3.

[0039] That is, assuming there are 10,000 patents in the patent dataset, the technical fields of 7,000 of these patents are labeled, while the technical fields of the other 3,000 patents are not labeled.

[0040] Step S13: Train the RoBERTa_BiLSTM model using the labeled dataset to obtain the trained RoBERTa_BiLSTM model, and input the unlabeled dataset into the trained RoBERTa_BiLSTM model for classification to obtain the technical field of each unlabeled sample in the unlabeled dataset.

[0041] BiLSTM is a bidirectional LSTM network that can encode text sequences in both forward and backward directions, thereby capturing more contextual information. This invention uses the contextual implicit semantic mining capabilities of the BERT model and the sequence feature modeling capabilities of bidirectional long short-term memory networks for classification, achieving the advantage of high classification accuracy.

[0042] Step S14: Based on the technical field corresponding to each patent in the patent dataset, collect all patents in the same technical field to obtain multiple technology subsets that correspond one-to-one with multiple technical fields.

[0043] In this invention, the technical fields of some patents are determined by combining ChatGPT-4o and manual verification, while the technical fields of other patents are predicted by a trained RoBERTa_BiLSTM model. In this way, the technical field of each patent is known. By aggregating patents in the same technical field, multiple technical subsets corresponding to multiple technical fields can be obtained.

[0044] In an optional implementation, step S10 further includes a step of selecting the best classification model prior to step S13, wherein the best classification model is a RoBERTa_BiLSTM model, and the step of selecting the best classification model includes:

[0045] Step (a): Split the labeled dataset into a training set and a validation set. Use the training set to train the BERT_uncased model, RoBERTa_base_cased model, and RoBERTA_BiLSTM model respectively to obtain the trained BERT_uncased model, RoBERTa_base_cased model, and RoBERTA_BiLSTM model.

[0046] In this invention, the ratio of the training set to the validation set is 7:3.

[0047] Step (b): Evaluate the trained BERT_uncased model, RoBERTA_base_cased model, and RoBERTA_BiLSTM model using the validation set, obtain the evaluation metric for each model, and take the model with the best evaluation metric as the best classification model. The evaluation metric includes accuracy, precision, recall, and F1 score.

[0048] The formulas for calculating accuracy, precision, recall, and F1 score are as follows:

[0049]

[0050] Wherein, TN represents negative samples predicted as negative by the model, TP represents positive samples predicted as positive by the model, FN represents negative samples predicted as positive by the model, and FP represents positive samples predicted as negative by the model.

[0051] Please refer to Table 1, which contains the evaluation metrics for the three classification models.

[0052] Table 1 Evaluation metrics for the three classification models

[0053]

[0054] As shown in Table 1, RoBERTa_base_cased outperforms BERT_uncased in precision, recall, and F1 score, but its precision is slightly lower. This is likely because RoBERTa_base_cased is more efficient at capturing various features when dealing with imbalanced data, resulting in better performance in precision and recall. After adding the BiLSTM layer, RoBERTa_BiLSTM outperforms RoBERTa_base_cased in all four evaluation metrics, with improvements of 0.93% in precision and 0.65% in F1 score. This indicates that the BiLSTM layer allows the model to more effectively capture contextual information in sequences and enhances its ability to recognize complex patterns. Therefore, it significantly improves both the accuracy and overall performance of text classification, especially in precision and F1 score, demonstrating the model's advantage in recognizing fine-grained features. The improved accuracy also indicates that after adding the BiLSTM layer, the model is more consistent in its overall predictions and can better classify complex high-speed rail technology patent data.

[0055] Step S20: Calculate the similarity value of a specific patent pair based on the PatentsBERTa model and the cosine similarity calculation formula, and obtain multiple similarity datasets that correspond one-to-one with multiple technology subsets. Each similarity dataset includes multiple similarity values, and the multiple similarity values ​​correspond one-to-one with multiple specific patent pairs in the same technology subset. The specific patent pair consists of two patents that have a direct reference relationship.

[0056] In one alternative implementation, the steps of obtaining the similarity dataset corresponding to any subset of technologies based on the PatentsBERTa model and the cosine similarity calculation formula include:

[0057] Step (1): Determine whether there is a direct reference relationship between any two patents in the technology subset, and form a specific patent pair with the two patents that have a direct reference relationship to obtain multiple specific patent pairs.

[0058] In this invention, the existence of a direct reference relationship between two patents is determined manually. This reference relationship specifically refers to a later reference relationship, that is, only the case where the later-filed patent references the earlier-filed patent is considered.

[0059] Step (2): According to the application time of the cited patents in each specific patent pair, calculate the similarity value of each specific patent pair one by one based on the PatentsBERTa model and the cosine similarity calculation formula to obtain the similarity dataset. The similarity dataset includes multiple similarity values ​​corresponding to multiple specific patent pairs.

[0060] In one optional implementation, the step of calculating the similarity value of a specific patent pair based on the PatentsBERTa model and the cosine similarity calculation formula includes:

[0061] The PatentsBERTa model is used to convert the specification summaries of the two patents in a specific patent pair into high-dimensional classification vectors, thereby obtaining two high-dimensional classification vectors corresponding to the specific patent pair.

[0062] By using two high-dimensional classification vectors corresponding to a specific patent pair as variables, the similarity value corresponding to the specific patent pair is calculated using the cosine similarity calculation formula.

[0063] The formula for calculating cosine similarity is:

[0064]

[0065] Where CosSimilarity(A, B) is the similarity value, and A and B are high-dimensional classification vectors.

[0066] In this invention, patents related to high-speed rail technology are divided into six technology subsets based on the technical field. Step S20 will obtain six similarity datasets, and the six similarity datasets correspond one-to-one with the six technology subsets. For ease of understanding, let's take an example. Suppose that the first similarity dataset in the six similarity datasets corresponds to train technology. Train technology includes 1,000 patents, and there are 100 specific patent pairs with citation relationships. Then the first similarity dataset includes 100 similarity values, and the 100 similarity values ​​correspond one-to-one with the 100 specific patent pairs.

[0067] Step S30: Based on the similarity dataset corresponding to each technology subset, obtain the similarity threshold of each technology subset, and obtain multiple candidate patent groups corresponding to each technology subset based on the similarity threshold of each technology subset. The multiple patents in each candidate patent group are formed by a set of multiple specific patent pairs with similarity values ​​greater than the similarity threshold.

[0068] In this invention, the similarity threshold for each technology subset is the arithmetic mean of multiple similarity values ​​in the similarity dataset.

[0069] Understandably, the similarity threshold for each technology subset is calculated separately, and is determined based on multiple similarity values ​​in the similarity dataset of the technology subset.

[0070] For ease of understanding, assume that the similarity dataset corresponding to the train technology subset includes 100 similarity values, from A1 to A100. Then, the similarity threshold for the train technology subset is (A1 + A2 + A3 + ... + A100) / 100 = B. If 60 of the 100 similarity values ​​are greater than the similarity threshold B, then the candidate patent group corresponding to the train technology subset consists of 60 specific patent pairs with similarity values ​​greater than the similarity threshold B. It is understandable that if there are duplicate patents among the 120 patents forming these 60 specific patent pairs, a deletion operation is needed to retain only one.

[0071] Step S40: Based on the candidate patent group corresponding to each technology subset, construct the temporal evolution graph and the citation network graph based on the citation relationship for each technology subset using Neo4j, and use graph projection to convert the graph triple relationship into network data to construct multiple evolutionary knowledge graphs that correspond one-to-one with multiple technology subsets.

[0072] In this invention, the evolutionary knowledge graph is a large, structured database that stores several "entities" (i.e., patents) and the "relationships" between them. Each patent is a node, and node attributes include patent number, name, application date, applicant, technical field, etc. Patent nodes are connected by lines, which are defined as "evolutionary relationships" and "temporal relationships".

[0073] Neo4j is a graph-based database management system that can handle graph-structured data well. In Neo4j, the basic representation of data is a triple, which is (node ​​1, relation, node 2). Such triple relations can well describe the interaction or connection between one entity and another entity. In the network of high-speed rail patent technology evolution, a patent node may be connected to another patent node through technology evolution relations.

[0074] It should be noted that the time-series evolution map can be understood as a time-series evolution map based on the application date.

[0075] Please see Figure 3 , Figure 3 To illustrate the triplet relationship, from Figure 3 It can be seen that the later-filed patent "A Multi-level High-speed Train Collision Buffer" is derived from the earlier patent "A Chamber-type..." The "expansion and stretching deformation energy absorption high-speed train anti-collision buffer" evolved from this, which also indicates that there is a reference relationship between the two patents.

[0076] Please see Figure 4 , Figure 4 (A) in the graph represents the evolutionary knowledge graph corresponding to the subset of train technology. Figure 4 (B) in the graph represents the evolutionary knowledge graph corresponding to the subset of train control technology. Figure 4 (C) in the graph represents the evolutionary knowledge graph corresponding to the subset of infrastructure technologies. Figure 4 In the graph, (D) represents the evolutionary knowledge graph corresponding to the subset of experimental and testing techniques. Figure 4 In the graph, (E) represents the evolutionary knowledge graph corresponding to a subset of digital and intelligent technologies. Figure 4 In the diagram, (F) represents the evolutionary knowledge graph corresponding to the subset of magnetic levitation technology.

[0077] The modularity of each evolutionary knowledge graph was determined using the Louvain community modularity test. Specifically, the modularity of the train technology subset was 0.821, the train control technology subset was 0.645, the infrastructure technology subset was 0.806, the digitalization and intelligent technology subset was 0.711, the experiment and testing technology subset was 0.757, and the magnetic levitation technology subset was 0.663. These data show that the evolutionary knowledge graph constructed using the method provided in this invention exhibits a moderate degree of modularity and higher technology dissemination efficiency.

[0078] Step S50: Obtain the preset target technology, and construct the evolutionary subgraph of the target technology based on the evolutionary knowledge graph corresponding to the subset of technologies to which the target technology belongs. The evolutionary subgraph takes the patent with the earliest application date among the patents related to the target technology as the seed node, and is formed by all subsequent patents that have direct and indirect reference relationships with the seed node, extracted along the evolutionary and temporal relationships.

[0079] Step S50 includes:

[0080] Step A: Obtain a preset target technology, and determine the technology subset to which the target technology belongs based on the preset target technology.

[0081] In this invention, each technical field includes multiple sub-technologies, and the preset target technology is a sub-technology within the corresponding technical field.

[0082] For ease of understanding and illustration, the sub-technologies within the technical field of train technology include brake disc technology, bogie technology, etc., and the preset target technologies can be brake disc technology, bogie technology, etc.

[0083] Step B: Using Neo4j's query language, search for target patents that meet the query conditions in the evolutionary knowledge graph corresponding to the technology subset to which the target technology belongs. The number of target patents is multiple.

[0084] In this invention, the target patent is located using Cypher statements.

[0085] In this invention, the query conditions may include keywords related to the target technology, as well as date-limiting terms that restrict the search scope.

[0086] Step C: Take the earliest patent filed among multiple target patents as the seed node, and extract all subsequent patents that have direct and indirect reference relationships with the seed node along the evolutionary and temporal relationships to form an evolutionary subgraph of the target technology.

[0087] To facilitate understanding, we will take the evolutionary subgraph of brake disc technology after 2001 as an example. When the preset target technology is brake disc technology, we first determine that the technical field to which brake disc technology belongs is train technology. Then, in the evolutionary knowledge graph corresponding to train technology, we use Cypher statements to locate the patent nodes related to brake disc technology with the application date from January 1, 2002 to the present. Next, we take the earliest application date node among the patent nodes related to brake disc technology as the seed node, expand along the evolutionary relationship and temporal relationship, and extract all subsequent patents that the technology corresponding to the seed node inherits, thus forming a complete subgraph of brake disc technology evolution.

[0088] In one optional implementation, the evolutionary subgraph of the target technology consists of multiple nodes and relationship lines connecting two nodes with direct reference relationships. The multiple nodes include a seed node and multiple non-seed nodes. The construction method further includes:

[0089] Step (a): Based on the betweenness centrality calculation formula of the node, calculate the betweenness centrality of each non-seed node in the evolution subgraph of the target technology, and perform normalization processing to obtain multiple betweenness centrality scores.

[0090] The formula for calculating the betweenness centrality of the node is:

[0091]

[0092] Among them, C B(v): Betweenness centrality of node v, representing the frequency and importance of node v as a "shortest path intermediary" for other node pairs in the network; (s,t): Two distinct nodes in the network, where s ≠ t ≠ v (i.e., excluding "pseudo-node pairs" of node v itself), traversing all such node pairs (s,t); σ(s,t): The total number of shortest paths from node s to node t (i.e., how many shortest paths exist between s and t); σ(s,t|v): The number of shortest paths from node s to node t that pass through node v. In this invention, the calculation process of this formula is as follows: for each pair of nodes s and t, calculate the proportion of the number of shortest paths passing through node v to the total number of shortest paths, and sum over all node pairs.

[0093] Step (b): Sort all betweenness centrality scores in descending order, and select at least one bridge node group for phase partitioning from the M target nodes with the largest betweenness centrality scores. Each bridge node group includes one or more target nodes, where M is a positive integer and 8≤M≤12.

[0094] The number of bridge node groups can be 1, 2, 3, 4, etc. When there is 1 bridge node group, the evolution subgraph of the target technology is divided into two stages. When there are 2 bridge node groups, the evolution subgraph of the target technology is divided into three stages. When there are 3 bridge node groups, the evolution subgraph of the target technology is divided into four stages. When there are 4 bridge node groups, the evolution subgraph of the target technology is divided into five stages.

[0095] In embodiments of the present invention, the number of target nodes can be 8, 9, 10, 11 and 12. In other embodiments, the number of target nodes can also be 6, 7, 13, etc. The number of bridge node groups is determined by human experience, and whether a target node can be a bridge node in a bridge node group is determined by a combination of human experience and manual determination.

[0096] In one alternative implementation, multiple bridge node groups for phase division are selected based on the application dates of the M target nodes.

[0097] In one optional implementation, when a bridge node group includes multiple target nodes, the interval between the earliest and latest application dates corresponding to the multiple target nodes in the same bridge node group is less than or equal to a preset duration; the maximum application date interval between target nodes in two adjacent bridge node groups is greater than the preset interval duration.

[0098] In this invention, the preset duration can be 6 months, 12 months, 24 months, 36 months, etc., and will not be listed exhaustively here. The specific duration is determined according to the application date range limited during the search. For ease of understanding, for example, if the patent dataset related to high-speed rail technology obtained in step S10 covers a 20-year period, the preset duration can be 24 months or 36 months, etc.; if the patent dataset related to high-speed rail technology obtained in step S10 covers a 10-year period, the preset duration can be 12 months or 18 months, etc.; if the patent dataset related to high-speed rail technology obtained in step S10 covers a 5-year period, the preset duration can be 4 months or 6 months, etc., and the specific duration is determined based on experience and actual conditions.

[0099] In this invention, the preset interval can be 3 years (36 months), 4 years (48 months), 5 years (60 months), etc., specifically determined according to the application date range limited during the search. When the application date range limited during the search is large, the preset interval can also be large; when the application date range limited during the search is small, the preset interval can also be small. It is understood that the preset interval is also related to the number of stages to be divided. A smaller number of stages to be divided results in a larger preset interval, and a larger number of stages to be divided results in a smaller preset interval.

[0100] To facilitate understanding, let's illustrate with an example. Assume there are 10 target nodes, a preset duration of 12 months, and a preset interval of 48 months. The method for selecting multiple groups of bridge nodes for phase division can be as follows: First, select the first non-seed node with the highest betweenness centrality score as one of the first group of bridge nodes. Then, compare the application dates of the other 9 target nodes with the application date of the first non-seed node. Nodes with application dates less than 12 months apart are also included in the first group of bridge nodes. Next, select the second non-seed node with the highest betweenness centrality score from the other target nodes not included in the first group. If the interval between the second non-seed node and the maximum application date of the first group of bridge nodes is greater than 48 months, it is included in the second group of bridge nodes. Then, refer to the selection criteria for the first group of bridge nodes to determine if other target nodes meet the requirements for inclusion in the second group. If the interval between the second non-seed node and the maximum application date of the first group of bridge nodes is less than 48 months, re-evaluate whether the third non-seed node with the second highest betweenness centrality score among the other target nodes not included in the first group meets the requirements… and so on.

[0101] Step (c) Based on the at least one bridge node group obtained in step (b), the evolution subgraph of the target technology is divided into stages to obtain an evolution subgraph divided into multiple development stages.

[0102] The construction method provided by this invention addresses the complexity and numerous sub-technologies of high-speed rail technology. It can identify the evolution paths of each sub-technology in each technical field, providing operational guidance for the innovation and development of high-speed rail technology, decision support for the coordinated progress of related industrial chains, and a practical basis for government departments to formulate technology policies and strategic plans.

[0103] Please see Figure 5 According to a second aspect of the present invention, the present invention also provides a device 100 for constructing a high-speed rail technology evolution map. The device includes: a classification module 101, used to acquire a patent dataset related to high-speed rail technology, classify the patents in the patent dataset according to the technical field, and obtain multiple technology subsets, each of which includes multiple patents; a calculation module 102, used to calculate the similarity value of a specific patent pair based on the PatentsBERTa model and the cosine similarity calculation formula, and acquire multiple similarity datasets corresponding one-to-one with the multiple technology subsets, each similarity dataset including multiple similarity values, the multiple similarity values ​​corresponding one-to-one with multiple specific patent pairs in the same technology subset, wherein the specific patent pair consists of two patents with a direct reference relationship; and an acquisition module 103, used to acquire the similarity threshold of each technology subset based on the similarity dataset corresponding to each technology subset, and based on the similarity of each technology subset... The first construction module 104 is used to construct a sub-technology temporal evolution graph and a citation network graph based on citation relationships for each technology subset based on the candidate patent groups corresponding to each technology subset, and to convert the graph triplet relationships into network data using graph projection, thereby constructing multiple evolutionary knowledge graphs that correspond one-to-one with multiple technology subsets. The second construction module 105 is used to obtain a preset target technology, and to construct an evolutionary subgraph of the target technology based on the evolutionary knowledge graph corresponding to the technology subset to which the target technology belongs. The evolutionary subgraph uses the patent with the earliest application date among the patents related to the target technology as the seed node, and is formed by all subsequent patents that have direct and indirect citation relationships with the seed node extracted along the evolutionary and temporal relationships.

[0104] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the working process of the above-described construction device can be referred to the content of the foregoing method embodiments, and will not be repeated here.

[0105] Example 1

[0106] Obtaining the Evolutionary Subgraph of Brake Disc Technology Based on the Corresponding Evolutionary Knowledge Graph of Train Technology

[0107] Step 1: In the evolutionary knowledge graph of train technology, the patent nodes related to brake disc technology with the application date from January 1, 2002 to the present were located using Cypher statements; then, the earliest application date node among the patent nodes related to brake disc technology was taken as the seed node, and the evolutionary and temporal relationships were expanded to extract all subsequent patents corresponding to the seed node, thus forming a complete brake disc technology evolution subgraph.

[0108] Step 2: Based on the betweenness centrality calculation formula of the node, calculate the betweenness centrality score of each non-seed node in the evolution subgraph of the target technology, and perform normalization processing to obtain multiple betweenness centrality scores. Sort them in descending order to obtain the 10 target nodes with the largest betweenness centrality scores.

[0109] Step 3: Based on the patent application dates of 10 target nodes, two bridge node groups for stage division are selected, and the brake disc technology evolution sub-graph is divided into three stages based on the selected bridge node groups.

[0110] In this embodiment, two bridge node groups are selected. The first bridge node group includes 5 target patents, and the second bridge node group includes 1 target patent. The first and second bridge node groups divide the brake disc technology evolution sub-graph into stage I, stage II and stage III.

[0111] Please refer to the following: Figure 6 , Figure 6 This is the brake disc evolution sub-graph constructed for this embodiment.

[0112] This evolutionary path can be roughly divided into three stages:

[0113] (1) Material composite and modification stage (2002-2008)

[0114] Starting with the first patent for composite material brake discs in 2002, various patents iterated on aspects such as material modification and the proportion of wear-resistant additives, forming an early group of material innovation technologies.

[0115] (2) Friction performance and structural optimization stage (2009-2014)

[0116] Subsequently, a batch of patents applied for between 2009 and 2014 in the sub-figures focused on friction coefficient adjustment and heat dissipation structure, using methods such as micro-surface texturing and ventilation groove improvement to ensure the stable performance of the brake disc in high-temperature environments.

[0117] (3) Intelligent monitoring and diagnostic integration stage (2015 to present)

[0118] During this period, patent applications focused on online monitoring, fault diagnosis, and intelligent algorithm support for brake discs. This signifies a shift in brake disc technology from a "passive improvement" model to a "proactive protection and improvement" model. By analyzing the abstract text information between entities, the path of brake disc technology improvement—from single-performance enhancements to comprehensive enhancements of multiple performance aspects—can be traced. For example, from "improving wear resistance" to "balancing heat dissipation and shock absorption" and then to "modular intelligent manufacturing," forming a clear technological evolution chain. Patent network analysis aided by knowledge graphs can accurately identify the main development lines and branches of brake disc technology, providing visualization assistance and decision-making references for predicting future technological development trends and R&D roadmaps.

[0119] This invention uses Neo4j visualization to obtain a complete sub-graph of the evolution of brake disc technology and summarizes its development into three major stages through text information: material composite and modification stage, friction performance and structural optimization stage, and intelligent monitoring and diagnostic integration stage.

[0120] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for constructing a high-speed rail technology evolution map, characterized in that, Includes the following steps: Step S10: Obtain a patent dataset related to high-speed rail technology, and classify the patents in the patent dataset according to the technical field to obtain multiple technical subsets, each of which includes multiple patents. Step S20: Calculate the similarity value of a specific patent pair based on the PatentsBERTa model and the cosine similarity calculation formula, and obtain multiple similarity datasets that correspond one-to-one with multiple technology subsets. Each similarity dataset includes multiple similarity values, and the multiple similarity values ​​correspond one-to-one with multiple specific patent pairs in the same technology subset. The specific patent pair consists of two patents that have a direct reference relationship. Step S20, which involves obtaining the similarity dataset corresponding to any subset of techniques based on the PatentsBERTa model and the cosine similarity calculation formula, includes: Step (1): Determine whether there is a direct reference relationship between any two patents in the technology subset, and form a specific patent pair with two patents that have a direct reference relationship to obtain multiple specific patent pairs; Step (2): According to the chronological order of the application dates of the cited patents in each specific patent pair, calculate the similarity value of each specific patent pair one by one based on the PatentsBERTa model and the cosine similarity calculation formula to obtain the similarity dataset. The similarity dataset includes multiple similarity values ​​corresponding one-to-one with multiple specific patent pairs. The step of calculating the similarity value of a specific patent pair based on the PatentsBERTa model and the cosine similarity calculation formula includes: The PatentsBERTa model is used to convert the specification summaries of the two patents in a specific patent pair into high-dimensional classification vectors, thereby obtaining two high-dimensional classification vectors corresponding to the specific patent pair. Using two high-dimensional classification vectors corresponding to a specific patent pair as variables, the cosine similarity calculation formula is used to calculate the similarity value corresponding to the specific patent pair. Step S30: Based on the similarity dataset corresponding to each technology subset, obtain the similarity threshold of each technology subset, and obtain multiple candidate patent groups corresponding to each technology subset based on the similarity threshold of each technology subset. The multiple patents in each candidate patent group are formed by a set of multiple specific patent pairs with similarity values ​​greater than the similarity threshold. Step S40: Based on the candidate patent group corresponding to each technology subset, construct the temporal evolution graph and the citation network graph based on the citation relationship for each technology subset using Neo4j, and use graph projection to convert the graph triple relationship into network data to construct multiple evolutionary knowledge graphs that correspond one-to-one with multiple technology subsets. Step S50: Obtain the preset target technology, and construct the evolutionary subgraph of the target technology based on the evolutionary knowledge graph corresponding to the technology subset to which the target technology belongs. The evolutionary subgraph is formed by taking the patent with the earliest application date among the patents related to the target technology as the seed node, and by all subsequent patents that have direct and indirect reference relationships with the seed node extracted along the evolutionary and temporal relationships. The evolutionary subgraph of the target technology consists of multiple nodes and relation lines connecting two nodes with direct reference relationships. The multiple nodes include a seed node and multiple non-seed nodes. The construction method further includes: Step (a): Based on the betweenness centrality calculation formula of the node, calculate the betweenness centrality of each non-seed node in the evolution subgraph of the target technology, and perform normalization processing to obtain multiple betweenness centrality scores. Step (b): Sort all betweenness centrality scores in descending order, and select at least one bridge node group for stage partitioning from the M target nodes with the largest betweenness centrality scores. Each bridge node group includes one or more target nodes, where M is a positive integer and 8≤M≤12. Step (c): Based on the at least one bridge node group obtained in step (b), the evolution subgraph of the target technology is divided into stages to obtain an evolution subgraph divided into multiple development stages.

2. The construction method according to claim 1, characterized in that, Based on the patent application dates of M target nodes, at least one bridge node group is selected for phase division.

3. The construction method according to claim 1, characterized in that, The various technology subsets include train technology subsets, infrastructure technology subsets, train control technology subsets, digitalization and intelligent technology subsets, testing and experimental technology subsets, and magnetic levitation technology subsets.

4. The construction method according to claim 3, characterized in that, Step S10 includes: Step S11: Based on the Incopat patent database, obtain invention patents with patent application dates from January 1, 1986 to December 31, 2024 through keyword search, and obtain a patent dataset related to high-speed rail technology. Step S12: Combine ChatGPT-4o and manual verification to annotate some patents in the patent dataset, determine the technical field of each annotated patent, obtain an annotated dataset including multiple annotated samples, and collect the specification abstracts of unannotated patents as unannotated samples to obtain an unannotated dataset. Each annotated sample in the annotated dataset includes a specification abstract and a technical field determined based on the specification abstract. Step S13: Train the RoBERTa_BiLSTM model using the labeled dataset to obtain the trained RoBERTa_BiLSTM model, and input the unlabeled dataset into the trained RoBERTa_BiLSTM model for classification to obtain the technical field of each unlabeled sample in the unlabeled dataset. Step S14: Based on the technical field corresponding to each patent in the patent dataset, collect all patents in the same technical field to obtain multiple technology subsets that correspond one-to-one with multiple technical fields.

5. The construction method according to claim 4, characterized in that, Step S10 also includes a step of selecting the best classification model prior to step S13, wherein the best classification model is the RoBERTa_BiLSTM model, and the step of selecting the best classification model includes: Step (a): Split the labeled dataset into a training set and a validation set, and use the training set to train the BERT_uncased model, RoBERTa_base_cased model and RoBERTA_BiLSTM model respectively, to obtain the trained BERT_uncased model, RoBERTa_base_cased model and RoBERTA_BiLSTM model. Step (b): Evaluate the trained BERT_uncased model, RoBERTa_base_cased model, and RoBERTA_BiLSTM model using the validation set, obtain the evaluation metric for each model, and take the model with the best evaluation metric as the best classification model. The evaluation metric includes accuracy, precision, recall, and F1 score.

6. The construction method according to claim 1, characterized in that, Step S50 includes: Step A: Obtain a preset target technology, and determine the technology subset to which the target technology belongs based on the preset target technology; Step B: Using Neo4j's query language, search for target patents that meet the query conditions in the evolutionary knowledge graph corresponding to the technology subset to which the target technology belongs. The number of target patents is multiple. Step C: Take the earliest patent filed among multiple target patents as the seed node, and extract all subsequent patents that have direct and indirect reference relationships with the seed node along the evolutionary and temporal relationships to form an evolutionary subgraph of the target technology.

7. A device for constructing a high-speed rail technology evolution map, characterized in that, The construction device is used to implement the method for constructing the high-speed rail technology evolution map as described in claim 1, and the construction device includes: The classification module is used to acquire a patent dataset related to high-speed rail technology. Based on the technical field, the patents in the patent dataset are classified to obtain multiple technical subsets, and each technical subset includes multiple patents. The calculation module is used to calculate the similarity value of a specific patent pair based on the PatentsBERTa model and the cosine similarity calculation formula, and to obtain multiple similarity datasets that correspond one-to-one with multiple technology subsets. Each similarity dataset includes multiple similarity values, and the multiple similarity values ​​correspond one-to-one with multiple specific patent pairs in the same technology subset. The specific patent pair consists of two patents that have a direct reference relationship. The acquisition module is used to acquire the similarity threshold of each technology subset based on the similarity dataset corresponding to each technology subset, and to acquire multiple candidate patent groups corresponding to each technology subset based on the similarity threshold of each technology subset. The multiple patents in each candidate patent group are formed by a set of multiple specific patent pairs with similarity values ​​greater than the similarity threshold. The first construction module is used to build a temporal evolution graph and a citation network graph based on citation relationships for each technology subset using Neo4j, based on the candidate patent group corresponding to each technology subset. The graph triplet relationship is converted into network data using graph projection to construct multiple evolutionary knowledge graphs that correspond one-to-one with multiple technology subsets. The second construction module is used to obtain a preset target technology and construct an evolutionary subgraph of the target technology based on the evolutionary knowledge graph corresponding to the subset of technologies to which the target technology belongs. The evolutionary subgraph is formed by taking the patent with the earliest application date among the patents related to the target technology as the seed node and extracting all subsequent patents that have direct and indirect reference relationships with the seed node along the evolutionary and temporal relationships.

Images