Technical evolution prediction method and system based on double-source evidence fusion, and terminal

Abstract

The present application belongs to the technical field of scientific and technological information analysis, and relates to a technology evolution prediction method and system based on double-source evidence fusion and a terminal. The present application respectively constructs a patent supply side evidence pool and a project demand side evidence pool, and performs double-source collection, shunting BERTopic topic modeling, cross-source topic alignment, construction of an evidence graph, calculation of weak signal scores and demand verification scores, and technology evolution prediction on patent supply side data and project demand side data, and outputs technology evolution path classification, graphed evidence chains and mechanism level prediction. The present application can simultaneously utilize technology reserve information reflected by patent data and task demand information reflected by project data, effectively solves the problems of supply and demand evidence fragmentation, mixed topic modeling collapse, lack of demand verification of weak signals and untraceability of prediction results in heterogeneous scientific and technological information, and makes the prediction results interpretable and verifiable.

Description

Technical Field

[0001] This invention relates to the field of science and technology intelligence analysis technology, and in particular to a method, system and terminal for predicting technology evolution based on dual-source evidence fusion. Background Technology

[0002] Currently, the monitoring and analysis of foreign (especially US) science and technology intelligence typically relies on a single data source. For example, this might involve analyzing only patent texts, collecting publicly available patent data from the US Patent and Trademark Office (USPTO), and using methods such as keyword frequency statistics, citation network analysis, and topic modeling (e.g., LDA) to identify technological hotspots and development trends. Alternatively, it might focus on analyzing only academic papers, project announcements, or procurement requirements. For instance, it might collect publicly available information from the US federal government, such as project announcements, procurement requirements, and technical reports, from sources like SAM.gov (Federal Procurement Data System), SBIR / STTR (Small Business Innovation Research / Technology Transfer Program), and DTIC (Defense Technology Information Center), to analyze government funding directions and capacity gaps. While single-source analysis can identify hotspots within a specific data source, it suffers from the following problems: First, there is a disconnect between supply-side and demand-side evidence. Patent data can reflect technical solutions, claims, specifications, and feasible paths, but it may not reflect recent task requirements. Project demand data can reflect the capability gaps that government departments, research funders, or procurement parties are focusing on, but it may not indicate whether the corresponding technical solutions have already been accumulated. If only patent data is used, it is easy to misjudge long-term mature technologies as recent evolution signals; if only project demand data is used, it is easy to misjudge short-term procurement hotspots as directions for technological breakthroughs.

[0003] Second, directly mixing heterogeneous texts in modeling can easily lead to topic collapse. Patent texts contain numerous templated expressions such as "method," "apparatus," "system," "wherein," and "claim," while project texts contain numerous procurement and project management expressions such as "proposal," "contract," "award," "solicitation," "phase," and "government." If patent and project texts are directly mixed into the same topic model, the model may prioritize learning the differences in text genres rather than learning the true technical semantics, resulting in the patent or project topic being swallowed up by a few large topics.

[0004] Third, current trend predictions often rely on keyword frequency, citation count, or topic popularity, making it difficult to identify weak signals that "have a low historical baseline but have recently risen, and are cross-validated by demand." Weak signals are usually not the most frequent topics, but rather early signals that accelerate in the recent window, semantically bridge with project needs, and can be drilled down to specific mechanisms or technical parameters.

[0005] Fourth, existing knowledge graph methods are mostly used for entity retrieval, supply and demand recommendation, or relationship display, lacking closed-loop calculations from original documents, topics, institutions, years, and technical characteristics to weak signal prediction indicators. The analysis results often fail to answer "which patents, which projects, which annual trends, and which semantic bridging edges support this weak signal."

[0006] Fifth, existing technology predictions often remain at the level of topic names or trend curves, making it difficult to further map them to mechanism-level predictions, simulation modules, parameters, and indicators. For example, in the context of space-based tactical data links, simply determining that "Link-16 is a hotspot" is insufficient; it is more important to determine whether its evolution mechanism is gateway adaptation, dynamic anti-interference, low-latency TDMA control, PNT degradation self-synchronization, or LPI / LPD reconfigurable waveforms.

[0007] In view of the above problems, there is an urgent need to propose a technology evolution prediction scheme that is oriented towards heterogeneous scientific and technological intelligence and outputs interpretable, traceable, and predictable weak signals of technology evolution. Summary of the Invention

[0008] The purpose of this invention is to propose a technology evolution prediction method, system, and terminal based on dual-source evidence fusion. Through dual-source acquisition, diversion modeling, cross-source topic alignment, evidence graph construction, weak signal scoring, and technology evolution prediction of patent supply-side data and project demand-side data, the invention outputs a graph-based evidence chain and mechanism-level prediction, including corresponding patents, corresponding projects, bridging similarity, evolution path, simulation modules, parameters, and evaluation indicators. This addresses the problems of fragmented supply and demand evidence, collapsed mixed-topic modeling, lack of demand verification for weak signals, and untraceable prediction results in heterogeneous scientific and technological intelligence, making the prediction results interpretable and verifiable.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for predicting technological evolution based on dual-source evidence fusion, comprising the following steps: S1. Obtain patent supply-side data and project demand-side data, and construct a patent supply evidence pool and a project demand evidence pool; S2. Extract document identifiers from the patent supply evidence pool and the project demand evidence pool, normalize all document identifiers, and obtain the standard field set of the documents; S3. Use a text embedding model to convert the standard field set into semantic vectors; S4. Construct a patent document set and a project document set respectively. The patent document set includes documents and their semantic vectors from the patent supply evidence pool after normalization. The project document set includes documents and their semantic vectors from the project demand evidence pool after normalization. S5. Perform BERTopic topic modeling on the patent document set and the project document set respectively to obtain the patent topic set and the project topic set; S6. Calculate normalized topic centroids based on the semantic vectors of documents in the patent topic set and project topic set, and calculate the cosine similarity between each patent topic centroid and each project topic centroid; configure the similarity threshold and the number of candidates K, and select the top K cross-source topic candidate pairs with cosine similarity greater than or equal to the similarity threshold from the patent topic set and project topic set respectively, merge and deduplicate the selection results, and generate cross-source topic alignment edges; S7. Construct an evidence graph based on a standard field set, semantic vectors, and cross-source topic alignment edges. The evidence graph includes document nodes, topic nodes, organization nodes, year nodes, and technical feature nodes. S8. Preset the recent window and baseline window, and calculate the weak signal score and demand verification score for each technical feature based on the evidence spectrum, the recent window and the baseline window; S9. Weigh and fuse weak signal scores and demand verification scores to obtain a technology evolution prediction index. Based on the technology evolution prediction index, output a technology evolution path classification, a mechanism-level prediction list, and an evidence package.

[0010] As one possible implementation, the patent supply evidence pool is constructed using the following method: Patent data is obtained from the patent database, and field completion, key field extraction, and patent relevance scoring are performed to generate a patent supply-side evidence pool. The patent relevance score includes at least the following indicators, which are sorted by weight from highest to lowest as follows: Co-occurrence of spatial and tactical data links, key organizational hits, core keyword hits, government interest statements, classification number hits, tactical data link keyword hits, resilience mechanism term hits, spatial term hits, citation counts; The patent relevance score also includes an exclusion word penalty item. When the patent text contains an exclusion word, a penalty score is deducted from the score. The following method was used to construct the project requirements evidence pool: Collect project demand-side data from multiple sources, perform multi-source entity deduplication and merging on the project demand-side data, and score project relevance to generate a project demand-side evidence pool. The project relevance score includes at least the following indicators, which are sorted by weight from highest to lowest as follows: Tactical data chain keyword hit, union of demand keywords and domain keywords hit, key organizational structure hit, resilience mechanism keyword hit, spatial keyword hit, and response deadline exists.

[0011] As one possible implementation, the evidence graph is constructed using the following method: Create evidence graph nodes, which include: document nodes, topic nodes, organization nodes, year nodes, and technical feature nodes; The edges for creating the evidence graph include: Calculate the first relevance score, the second relevance score, and the third relevance score respectively; wherein, the first relevance score is the relevance score between any two patent documents in the patent document set; the second relevance score is the relevance score between any two project documents in the project document set; and the third relevance score is the relevance score between any patent document in the patent document set and any project document in the project document set. Configure the first threshold, the second threshold, and the third threshold respectively; Connect two patent documents with a first correlation score greater than the first threshold, connect two project documents with a second correlation score greater than the second threshold, and connect patent documents and project documents with a third correlation score greater than the third threshold to form the edges of the evidence graph.

[0012] As one possible implementation, the weak signal score is calculated using the following method: For a given technical feature, obtain the number of first patent documents mentioned in the recent window, the number of second patent documents mentioned in the baseline window, the number of third patent documents mentioned in the last year of the recent window, and the number of fourth patent documents mentioned in the year before the recent window; preset the minimum baseline upper limit and the minimum scale threshold; The low baseline score is calculated based on the minimum baseline upper limit and the number of second patent documents; the growth score is calculated based on the number of first patent documents and the number of second patent documents; the acceleration score is calculated based on the number of third patent documents and the number of fourth patent documents; and the recent size score is calculated based on the number of first patent documents and the minimum size threshold. The weak signal score of this technical feature is obtained by weighting and summing the low baseline score, growth score, acceleration score and recent size score.

[0013] As one possible implementation, the requirement verification score is calculated using the following method: For a certain technical feature, obtain the number of first project documents appearing in the recent window, the number of second project documents from priority requests in the recent window, the number of semantic bridge edges between patent documents and project documents related to the technical feature, the maximum semantic similarity between patent documents and project documents related to the technical feature, the number of active project topics in the recent window related to the technical feature, and the number of sources of projects in the recent window related to the technical feature. The bridging density score is calculated based on the number of semantic bridging edges; the semantic verification score is calculated based on the bridging density score and the maximum semantic similarity; the demand scale score is calculated based on the number of documents in the first project; the priority source score is calculated based on the number of documents in the second project; the project topic activity score is calculated based on the number of active project topics; and the source diversity score is calculated based on the number of sources. The bridging density score, semantic verification score, demand scale score, priority source score, project topic activity score, and source diversity score are weighted and summed to obtain the demand verification score for this technical feature.

[0014] As one possible implementation, when weighting and fusing weak signal scores and demand verification scores, the weight of demand verification scores is greater than that of weak signal scores.

[0015] As one possible implementation approach, the technology evolution path classification includes: demand-driven, technology reserve-driven, supply and demand synchronized, and single-source observation-driven; the mechanism-level prediction list includes: mechanism identifier, candidate parent technology feature set, mechanism name, technology claim, operational requirements, predicted evolution direction, alternative term set, simulation module, parameters and indicators; the evidence package includes: a list of patent document identifiers supporting the prediction results, a list of project document identifiers, a list of cross-source topic alignment edges, a list of semantic bridging edges between patents and projects, and the confidence level of the matching between technology features and mechanisms.

[0016] Secondly, the present invention provides a technology evolution prediction system based on dual-source evidence fusion, comprising: The dual-source evidence pool construction unit acquires patent supply-side data and project demand-side data to construct a patent supply evidence pool and a project demand evidence pool. The standard field set generation unit extracts document identifiers from the patent supply evidence pool and the project demand evidence pool, normalizes all document identifiers, and obtains the standard field set of the document. The document collection construction unit uses a text embedding model to convert the standard field set into semantic vectors, and constructs a patent document collection and a project document collection respectively. The patent document collection includes documents and their semantic vectors in the patent supply evidence pool after normalization, and the project document collection includes documents and their semantic vectors in the project demand evidence pool after normalization. The topic set acquisition unit performs BERTopic topic modeling on the patent document set and the project document set respectively to obtain the patent topic set and the project topic set; The cross-source topic alignment edge generation unit calculates normalized topic centroids based on the semantic vectors of documents in the patent topic set and the project topic set, and calculates the cosine similarity between each patent topic centroid and each project topic centroid; it configures a similarity threshold and a candidate number K, and selects the top K cross-source topic candidate pairs with a cosine similarity greater than or equal to the similarity threshold from the patent topic set and the project topic set respectively, merges and removes duplicates from the selection results, and generates cross-source topic alignment edges; The evidence graph construction unit constructs an evidence graph based on joint analysis documents, semantic vectors, and cross-source topic alignment edges. The evidence graph includes document nodes, topic nodes, organization nodes, year nodes, and technical feature nodes. The scoring unit has a preset recent window and a baseline window. Based on the evidence spectrum, the recent window, and the baseline window, it calculates the weak signal score and demand verification score for each technical feature. The output unit calculates and fuses the weak signal score and the demand verification score to obtain the technology evolution prediction index. Based on the technology evolution prediction index, it outputs a technology evolution path classification, a mechanism-level prediction list, and an evidence package.

[0017] Thirdly, the present invention provides a terminal including a processor and a communication interface coupled to the processor, the processor being used to run computer programs or instructions to implement the technology evolution prediction method based on dual-source evidence fusion provided in the first aspect.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. The technology evolution prediction method, system and terminal based on dual-source evidence fusion proposed in this invention, by constructing a patent supply-side evidence pool and a project demand-side evidence pool respectively, and performing normalization processing and joint analysis on the two, can simultaneously utilize the technology reserve information reflected by patent data and the task demand information reflected by project data, overcome the one-sidedness problem caused by single data source analysis, and realize the organic integration of supply and demand evidence.

[0019] 2. This invention addresses the differences in style, expression, and terminology between patent texts and project texts by employing a dual-channel BERTopic topic modeling strategy. It independently models the patent document set and the project document set, then achieves cross-source semantic bridging through topic centroid alignment. This "fidelity-first, fusion-later" technical architecture effectively avoids misinterpreting stylistic differences between patent template terms and project procurement terms as technical topic differences, significantly improving the semantic accuracy and interpretability of topic modeling.

[0020] 3. This invention maps text into dense semantic vectors and calculates the cosine similarity between patent topics and project topics based on the normalized centroids of document vectors within a topic. A bidirectional Top-K alignment strategy is then used to generate cross-source topic alignment edges. Even if the patent text and project text use different keywords and expressions, this method can accurately identify the semantic proximity between them, achieving automatic alignment of supply-side and demand-side topics.

[0021] 4. This invention constructs an evidence graph containing document nodes, topic nodes, organizational structure nodes, year nodes, and technical feature nodes, and establishes equilateral relationships such as document-attributed topics, cross-source topic alignment, semantic bridging of patent projects, and document-mentioned technical features. Each technology evolution prediction result can be traced back along the graph to the specific patent document, project document, topic alignment relationship, and semantic bridging edge supporting the prediction, giving the prediction results complete evidence chain support and significantly improving the interpretability and credibility of technical intelligence analysis.

[0022] 5. This invention proposes a three-tiered quantitative index system: Patent Weak Signal Score (WS), Demand Verification Score (DV), and Technology Evolution Prediction Index (EPI). The WS score emphasizes the weak signal characteristics of "low baseline, recent growth, and accelerated emergence," while the DV score emphasizes the cross-validation mechanism of "project demand-driven and semantically bridging verification." The EPI index weights and merges the two, giving higher weight to demand verification. This quantitative system can effectively identify early technological trends with low historical baselines but recent accelerated growth, and cross-validation by demand, unlike traditional heat analysis methods that rely solely on keyword frequency or citation count.

[0023] 6. This invention can start from high-value weak signals, match RSTP semantic bridging evidence with a predefined mechanism catalog, and output mechanism-level prediction results with specific technical claims, operational requirements, simulation modules, parameters, and indicators. This capability allows technology prediction to move beyond the level of macro-topic names (such as "Link-16 is a hot topic"), and instead answer verifiable and simulable technical questions such as "Link-16 specifically evolves towards interference-resistant dynamic reconstruction, PWSA tactical edge gateway adaptation, or low-latency TDMA control," significantly enhancing the practical value of technical intelligence.

[0024] 7. Although this invention has been validated in embodiments of space-based tactical data links and space-assisted Link-16, its core methods—including dual-source evidence pool construction, diversion topic modeling, cross-source topic alignment, evidence map construction, weak signal scoring, and mechanism-level prediction—possess good domain portability. As long as collectable patent supply-side data and project demand-side data exist, this invention can be applied to predicting technological evolution in multiple technical fields such as satellite communication, electronic warfare, artificial intelligence, semiconductor manufacturing, new materials, and biomedicine. Attached Figure Description

[0025] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a flowchart of the technology evolution prediction method based on dual-source evidence fusion proposed in this embodiment of the invention; Figure 2 This is a schematic diagram illustrating the process of constructing a cross-source topic similarity matrix in an embodiment of the present invention; Figure 3 This is a flowchart illustrating the construction of a dual-source evidence map in an embodiment of the present invention; Figure 4 This is a schematic diagram illustrating the process of calculating the weak signal score, demand verification score, and technology evolution prediction index in an embodiment of the present invention. Detailed Implementation

[0026] To facilitate a clear description of the technical solutions in the embodiments of the present invention, the terms "first" and "second" are used to distinguish identical or similar items with essentially the same function and effect. For example, the first threshold and the second threshold are merely used to distinguish different thresholds and do not limit their order. Those skilled in the art will understand that the terms "first" and "second" do not limit the quantity or execution order, and that the terms "first" and "second" are not necessarily different.

[0027] It should be noted that in this invention, the terms "exemplary" or "for example" are used to indicate examples, illustrations, or descriptions. Any embodiment or design described as "exemplary" or "for example" in this invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0028] In this invention, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one" or similar expressions refer to any combination of these items, including any combination of singular or plural items. For example, "at least one of a, b, or c" can represent: a, b, c, a combination of a and b, a combination of a and c, a combination of b and c, or a, b, and c, where a, b, and c can be single or multiple.

[0029] This invention aims to provide a method, system, and terminal for predicting technological evolution based on dual-source evidence fusion. By performing dual-source acquisition, diversion modeling, cross-source topic alignment, evidence graph construction, weak signal scoring, and technological evolution prediction on patent supply-side data and project demand-side data, it outputs a graph-based evidence chain and mechanism-level prediction, including corresponding patents, corresponding projects, bridging similarity, evolution paths, simulation modules, parameters, and evaluation indicators, making the prediction results interpretable and verifiable.

[0030] In this technical solution, "patent supply-side data" refers to technical reserve evidence embodied in patent texts, claims, specifications, classification numbers, government interests, applicants, and citation relationships. "Project demand-side data" refers to task-driven evidence embodied in government projects, procurement announcements, SBIR / STTR projects, BAA / RFI, technical reports, etc. "Dual-source evidence fusion" refers to the fusion of patent supply-side evidence and project demand-side evidence within a unified corpus, topic, graph, and scoring framework. "Topic centroid" refers to the semantic representation of a topic obtained by normalizing the average vector of all normalized document vectors within that topic. "Weak signals" refer to technical signals with low historical baselines but recent growth that can be verified by demand-side evidence. "Recent window" is a year window used to identify recent growth, such as the two most recent full years. "Baseline window" is a year window used to measure historical foundations, such as the three full years prior to the recent window. "Mechanism-level prediction" refers to specific technical mechanisms, operational requirements, parameters, and indicators obtained by drilling down further from the parent weak signals.

[0031] In a first aspect, embodiments of the present invention provide a method for predicting technological evolution based on dual-source evidence fusion, see [link to previous section]. Figure 1 It includes the following steps: S1. Obtain patent supply-side data and project demand-side data, and construct a patent supply evidence pool and a project demand evidence pool; As one possible implementation, the patent supply evidence pool is constructed using the following method: Patent data is obtained from the patent database, and field completion, key field extraction, and patent relevance scoring are performed to generate a patent supply-side evidence pool. The patent relevance score includes at least the following indicators, which are sorted by weight from highest to lowest as follows: Co-occurrence of spatial and tactical data links, key organizational hits, core keyword hits, government interest statements, classification number hits, tactical data link keyword hits, resilience mechanism term hits, spatial term hits, citation counts; As an example of constructing a patent supply evidence pool, a batch table of published patents, full-text patent data, a patent citation table, and a patent classification table are obtained from a patent database. For example, the batch table includes at least a basic patent table, an abstract table, an application table, an assignee table, an inventor table, a CPC classification table, an IPC classification table, a government interest declaration table, a government interest agency table, a patent citation table, and a detailed specification table.

[0032] First, read the patent baseline table and perform the following initial screening on each patent record: Determine if the patent year falls within the target year range, determine if the patent type is the target type, exclude withdrawn or invalid records, perform candidate text judgment in the title, perform candidate text judgment in the abstract, perform candidate text judgment in the government interest statement, and perform supplementary inclusion in the context of constellations, satellites, space links, etc.

[0033] It should be noted that candidate text evaluation is not based on a single keyword, but rather on rules based on a set of multiple signal types. Let the set of domain keywords be... The spatial word set is The set of tactical data chain keywords is The set of constellation and interstellar link terms is The set of words related to resilience and anti-interference is as follows Then for the text Define the hit function: in, It can be Any set in the range.

[0034] The candidate patent judgment function can be expressed as: The first item indicates that the domain keyword is hit; the second item indicates that spatial words or constellation and inter-satellite link words appear together with tactical data link keywords; and the third item indicates that resilience and anti-interference words appear together with tactical data link keywords.

[0035] After the initial screening, a set of patent numbers meeting the requirements can be obtained. Next, fields are completed from the patent batch table, including: completing the title, grant date, patent year, and patent type from the patent base table; completing the abstract from the abstract table; completing the earliest application date from the application table; completing fields such as organization, country, state, and city from the assignee table; completing inventor information from the inventor table; completing the classification number from the CPC and IPC classification tables; completing the government interest statement from the government interest table; completing the federal agency level from the government interest agency table; calculating the number of US patents cited by the patent and the number of times the patent has been cited by other candidate patents from the patent citation table; and completing the long text of the specification from the specification table.

[0036] The above fields are uniformly written into the patent supply evidence pool. Each patent in the evidence pool shall retain at least the following fields: patent number, title, abstract, authorization date, year, type, applicant, inventor, classification number, government interest statement, government interest agency, citation statistics, claims or specification text, candidate reason, and original source.

[0037] Next, a patent relevance score is calculated for each patent. The meaning of each parameter in the scoring formula is shown in Table 1: Table 1. Comparison of Patent Relevance Scoring Parameters and Their Meanings The formula for scoring patent relevance is as follows: in, , , , , , , , , These represent the weights for core keyword hits, spatial term hits, tactical data chain keyword hits, resilience mechanism term hits, key organization hits, classification number hits, government interest statements, spatial and tactical data chain co-occurrence, and citation counts, respectively. These weights determine the relative importance of different indicators; the higher the weight, the greater the contribution of that indicator to the final score.

[0038] For example, the values ​​of the above weights are as follows: , , , , , , , , .

[0039] Patent relevance scoring also includes exclusion penalty items. This indicates the penalty score weight for the excluded word penalty item. When the patent text matches an excluded word, a penalty score is deducted from the score. For example, .

[0040] It should be noted that the weight values ​​in the above examples are preferred embodiments of the present invention in the field of space-based tactical data links, and are not fixed limitations. Those skilled in the art can adjust the weights according to the characteristics of the specific technical field, data type distribution, or expert experience. For example, when analyzing the biomedical field, the weight of "Government Interest Statement" can be lowered, while the weight of "Core Keyword Hit" can be increased. The specific selection of weight values ​​is a conventional design choice for those skilled in the art and does not constitute a limitation on the scope of protection of the present invention.

[0041] As one possible implementation, the project requirements evidence pool can be constructed using the following method: Collect project demand-side data from multiple sources, perform multi-source entity deduplication and merging on the project demand-side data, and score project relevance to generate a project demand-side evidence pool. The project relevance score includes at least the following indicators, which are sorted by weight from highest to lowest as follows: Tactical data chain keyword hit, union of demand keywords and domain keywords hit, key organizational structure hit, resilience mechanism keyword hit, spatial keyword hit, and response deadline exists.

[0042] As an example of constructing a project requirements evidence pool, project requirements data is collected from multiple heterogeneous sources, including DTIC technical reports, SAM.gov government opportunities, DARPA announcements, SBIR / STTR projects, and SBIR / STTR awards. The data structures of these different sources vary significantly. For example, SAM.gov records emphasize announcement number, procurement type, response deadline, and office information; SBIR / STTR records emphasize project number, stage, abstract, technical field, and award information; and DARPA or DTIC records may emphasize project name, technical report number, project description, or main text. These different sources are uniformly mapped to project record objects, with fields including at least: source system, source record identifier, project type, title, abstract or overview, main text description, requirements or technical requirements text, publication date, update date, and response deadline, status, organization, sub-organization, and office, source link, solicitation number, announcement number, project number, award number, or technical report number, relevant organization, keywords, external identifiers, and original record.

[0043] Next, the following processing is performed on each source: select the collector according to the source parameters, filter records according to the year window, exclude records that lack source record identifiers, perform candidate project judgment on title, abstract, body text, requirement text, organization and keywords, group candidate records according to normalization key, merge records from multiple sources under the same normalization key, calculate the requirement-side score for the merged project entity, and write the project entity, source record and score result into the project requirement evidence pool.

[0044] Set project text It is composed of the project title, abstract, body text, requirement text, organization field, organizational structure, and keywords. Let the set of requirement-driving keywords be... The key institutional term set is The project candidate judgment function can be expressed as: Project requirement data often shows the same project appearing repeatedly from different sources, at different announcement stages, or at different reward stages. To avoid duplicate counting, a normalized key is constructed for each project record. Normalized key generation rules are arranged from highest to lowest reliability. From the given record set, select the record with the highest text quality as the master record: In the formula, Indicates the master record. , , Representing records respectively The main text, requirements text, and summary text, Indicates the length of the string. The sorting priority is to compare the length of the main text first, then the length of the required text, and then the length of the abstract.

[0045] Then, the abstract, main text, requirement text, organizational structure, keywords, external identifiers, and date fields are deduplicated and merged. For example, text fields are deduplicated and concatenated: In the formula, This represents the merged, deduplicated text. This indicates a function that removes duplicates and concatenates data in the original source order. Indicates the first The corresponding text field of the source record, This indicates the number of source records involved in the merger under the same project entity.

[0046] For the date field, the earliest publication date, latest update time, and latest response deadline can be used. Through the above normalization and deduplication, the same real project entity is counted as only one project in the evidence pool, but its evidence from multiple sources is still retained.

[0047] set up For project text, For project organization text, To ensure the existence of a response deadline, the project score is expressed as follows: in, This indicates that domain keywords and demand-driven terms are combined and deduplicated before counting to avoid awarding points for the same semantic meaning. The response deadline item indicates that the project has a relatively clear time constraint. This score is used to screen highly relevant project demands and serves as the strength of demand-side evidence in subsequent technology evolution predictions.

[0048] S2. Extract document identifiers from the patent supply evidence pool and the project demand evidence pool, normalize all document identifiers, and obtain the standard field set of the documents; As an example, extract document identifiers for each patent: Patent document identifiers must include at least: , Patent number, date, year, type, score, applicant, government interest agency, organization, IPC / CPC classification number, citation statistics. Extract document identifiers for each project: Project document identifiers should include at least: , Project identifier, project type, source system, title, abstract, body text, requirements text, release date, response deadline, event year, organization, sub-organization, office, organizational structure, keywords, and rating.

[0049] Normalization ensures that patent document identifiers and project document identifiers have aligned field types and use a unified field structure, facilitating subsequent vectorization, topic modeling, graph writing, and time statistics. Furthermore, within the standard field set formed by both, [the following is used]... and Isolating namespaces avoids cross-origin primary key conflicts and preserves... This enables subsequent topic-based modeling and cross-source alignment to be performed within the same processing framework. Furthermore, incorporating long texts such as patent claims, specifications, and project requirements into the analysis avoids semantic loss caused by relying solely on title summaries; and by marking incomplete years, it avoids misclassifying incompletely collected years as declining years in annual trend analysis.

[0050] S3. Use a text embedding model to convert the standard field set into semantic vectors; As an example, we use the text embedding model to convert a standard set of fields into semantic vectors. Let the text encoder be... ,in Indicates model parameters, This indicates the input text. If we represent the output vector, then: in, Represents the vector dimension. Represents the set of real numbers. express 3D real vector space, vector Treat as a document The semantic representation of.

[0051] Input text as follows: In the formula, Text used for topic modeling This indicates text used for other analyses.

[0052] To avoid duplicate requests and ensure traceability, a vector cache key is established: In the formula, Indicates the first The vector cache key for each document. Indicates the first Document identifier, This indicates the embedded input text. The calculated text hash, Indicates the name of the embedded model. This indicates a document index.

[0053] If a vector with the same document, text hash, and model name already exists in the cache, the cached vector is reused directly; otherwise, the text embedding model is called to generate a new vector.

[0054] S4. Construct a patent document set and a project document set respectively. The patent document set includes documents and their semantic vectors from the patent supply evidence pool after normalization. The project document set includes documents and their semantic vectors from the project demand evidence pool after normalization. S5. Perform BERTopic topic modeling on the patent document set and the project document set respectively to obtain the patent topic set and the project topic set; As an example, BERTopic topic modeling is performed using the following method: The Uniform Manifold Approximation and Projection (UMAP) algorithm is employed to perform nonlinear dimensionality reduction on the semantic vectors in the document set to be modeled, mapping the high-dimensional semantic vectors to a low-dimensional space. Specifically, UMAP first uses distance in the high-dimensional space... Build for each point Nearest neighbor set For example, The distance metric is cosine distance: In the formula, Representing vectors and cosine distance, Represents the midpoint of space Time The distance between, This indicates the cosine similarity between the two.

[0055] The UMAP algorithm processes each point Learning local scale parameters and and the midpoint of the high-dimensional space Time The fuzzy membership strength is represented as: in, Point Local connectivity distance to nearest neighbor Used to adjust the density of local neighborhoods.

[0056] UMAP synthesizes locally directed adjacency strengths into an undirected fuzzy graph. For points... With point The edge weights in a high-dimensional graph can be represented as: This formula is equivalent to the union operation of fuzzy sets, and is used to represent the strength of two points belonging to the same local neighborhood structure in a high-dimensional space.

[0057] UMAP searches for points in low-dimensional space. To make low-dimensional fuzzy graphs As close as possible to a high-dimensional fuzzy image Low-dimensional edge weights can typically be represented as: in, These are curve parameters determined by parameters such as min_dist. The optimization objective of UMAP can be written as the cross-entropy between the high-dimensional and low-dimensional graphs: Optimizing this objective can keep similar points in high-dimensional space similar in low-dimensional space, while maintaining a certain degree of separation between irrelevant points.

[0058] This solution introduces the UMAP algorithm to compress high-dimensional text vectors into a low-dimensional space, alleviating the sparsity of high-dimensional structures and preserving local neighborhood structures. This makes it easier for patents or projects with similar technical semantics to form density clusters. Simultaneously, it provides HDBSCAN with inputs more suitable for density clustering and improves the reproducibility of topic modeling results by using a fixed random seed.

[0059] Then, in the reduced-dimensional space, the HDBSCAN algorithm is used for density clustering to divide the documents into multiple document clusters, and documents that cannot be stably assigned to any cluster are marked as outliers. The HDBSCAN algorithm is a density-based hierarchical clustering method. It does not require a pre-specified number of topics, can identify clusters of arbitrary shapes, and marks low-density samples as outliers. This characteristic is suitable for patent and project texts, as scientific texts often contain long-tailed topics, noisy text, and cross-disciplinary mixed text.

[0060] For example, suppose the set of points after dimensionality reduction by UMAP is: in, This represents the set of low-dimensional points after UMAP dimensionality reduction. Indicates the first The coordinates of a document in a low-dimensional space. , This represents the number of low-dimensional points.

[0061] For any point Define its Distance to nearest core: In the formula, Point To its first The distance to the nearest neighbor, This represents the distance function in a low-dimensional space.

[0062] For example, Optimal choice 10. HDBSCAN uses mutual reachability distance instead of ordinary Euclidean distance, which has the following advantages: if a point is located in a sparse region, the core distance is larger, and the distance between points is amplified; if a point is located in a dense region, the core distance is smaller, and the distance between points is closer to the original distance; this processing makes clustering focus more on high-density regions and reduces the influence of noise points on the cluster boundaries. For two points... Its mutual reachability distance is: In the formula, Represent two points in a low-dimensional space and The distance between them.

[0063] HDBSCAN constructs a minimum spanning tree based on mutual reachability distance and progressively segments it into a hierarchical clustering tree. Each cluster is born, expands, or dies at different density thresholds. Let the density hierarchy parameters be: For a certain cluster Its stability can be summarized as the cumulative duration of points within a cluster at the density level: In the formula, Represents sample points Density level at the time of leaving the cluster or when the cluster disappears. This indicates the density level at which a cluster is formed.

[0064] HDBSCAN uses the Excess of Mass (EOM) method to select the most stable clusters from the agglomeration tree as the final document clusters. Documents that cannot stably belong to any cluster are marked as outliers and assigned a topic number of -1.

[0065] The advantages of using HDBSCAN in this solution are: it does not require a preset number of topics, making it suitable for exploring unknown technical fields; it can mark noisy text and unstable topics as outliers, reducing false alarms in weak signal identification; it can identify topic clusters of different densities, adapting to situations where patent topics and project topics are unevenly distributed; and when combined with UMAP, it can form interpretable technical topics in the semantic space.

[0066] After obtaining a document cluster, BERTopic concatenates all documents under it into a class document, uses the c-TF-IDF algorithm to extract the keyword with the highest c-TF-IDF weight in the class document, and uses it as the topic keyword of the document cluster. Each topic keyword is assigned a different number; and selects several documents with high topic probability for each topic as representative documents of the topic cluster.

[0067] As an example, suppose the set of document clusters is... Then a certain cluster The following document collection is Then the class document is: Then, bag-of-words analysis was performed on the class documents, using unary, bigram, and trigram phrases, and low-frequency words were filtered out. In addition, domain-specific stop words were used to reduce patent template terms and project procurement terms. For example, patent-side stop words included: claim, claims, method, apparatus, wherein, embodiment, etc., while project-side stop words included: proposal, contract, award, solidification, phase, government, etc.

[0068] For terms and class documents Define intra-class term frequencies: in, For terms In the class document The number of times it appears in the text.

[0069] set up This represents the average document length for each topic category. For terms The inverse document frequency of c-TF-IDF can be expressed as the frequency of occurrence or the number of occurrences of all classes in all documents: Then the term On the topic The class TF-IDF weights are: Standard TF-IDF treats each document as a statistical unit, suitable for document retrieval; c-TF-IDF treats each topic cluster as a statistical unit, suitable for topic naming. This highlights keywords that are more distinctive to a particular topic than to other topics, rather than just high-frequency words in a single document. The technical benefits of using c-TF-IDF in this solution include: highlighting technical solution terms in patent topics, such as frequency hopping, time slot allocation, uplink control, and inter-satellite relay; and highlighting task requirement terms in project topics, such as tactical edge, PWSA, anti-jam, network operations, MIDS, and Link-16. By training patent topics and project topics separately, and then extracting keywords separately, the system can obtain clearer supply-side and demand-side topics.

[0070] For each document BERTopic outputs local topic numbers. and topic probability Convert local numbering to global numbering: The patent topic offset can be 100,000, and the project topic offset can be 200,000. Each topic is assigned a record that retains the document identifier, document type, source system, event year, topic number, local topic number, topic probability, original topic name, keywords, and representative document tags.

[0071] For each non-outlier topic Select representative documents based on topic probability and event year: Preferably, five representative documents are selected for each topic. These representative documents are used to audit the quality of the topic, generate summaries, and support the embodiments in the technical disclosure document.

[0072] Calculate the number of documents, patents, projects, and annual percentage for each topic and year. Let... Theme In the year The number of documents, For the year If the total number of documents is: In the formula, Indicates the topic In the year The percentage of documents.

[0073] The following calculations can be performed separately for the patent side and the project side: S6. Calculate normalized topic centroids based on the semantic vectors of documents in the patent topic set and project topic set, and calculate the cosine similarity between each patent topic centroid and each project topic centroid; configure the similarity threshold and the number of candidates K, and select the top K cross-source topic candidate pairs with cosine similarity greater than or equal to the similarity threshold from the patent topic set and project topic set respectively, merge and deduplicate the selection results, and generate cross-source topic alignment edges; Since patent topics and project topics are trained separately, it is necessary to establish cross-source topic alignment relationships. This invention uses the normalized topic centroids of document vectors within a topic as the semantic representation of the topic.

[0074] Set a theme The document collection is The document vector is First, normalize the document vectors: Calculate the initial centroid of the topic: Then normalize the centroid of the theme: in, That is, the topic The semantic centroid.

[0075] Let the set of patent topics be The project theme collection is The following method is used to construct the centroid matrix of patent topics and the centroid matrix of project topics: See Figure 2 The cross-source topic similarity matrix is ​​as follows: Since the centroid has been normalized, the dot product is the cosine similarity.

[0076] Select alignment edges from both the patent topic perspective and the project topic perspective, and apply them to each patent topic. Select the one with the highest similarity that exceeds the threshold. The former Project Themes: For each project theme Select the one with the highest similarity that exceeds the threshold. The former Patent topics: The final set of aligned edges is the union of the two candidate sets: Preferably, , , .

[0077] Each alignment edge should include at least the source topic, target topic, source topic range, target topic range, number of source topic documents, number of target topic documents, similarity, source side ranking, target side ranking, and alignment method.

[0078] By aligning topics, we can avoid topic collapse caused by the direct mixing of patent text and project text, thus preserving the respective topic structures of the supply and demand sides. We establish cross-source semantic bridges through topic centroid similarity and prevent large topics from unilaterally swallowing smaller topics, providing topic-level evidence for edge creation and technology evolution prediction in subsequent evidence graphs.

[0079] S7. Construct an evidence graph based on a standard field set, semantic vectors, and cross-source topic alignment edges. The evidence graph includes document nodes, topic nodes, organization nodes, year nodes, and technical feature nodes. As one possible implementation, the evidence graph is constructed using the following method: Create evidence graph nodes, which include: document nodes, topic nodes, organization nodes, year nodes, and technical feature nodes; The edges for creating the evidence graph include: Calculate the first relevance score, the second relevance score, and the third relevance score respectively; wherein, the first relevance score is the relevance score between any two patent documents in the patent document set; the second relevance score is the relevance score between any two project documents in the project document set; and the third relevance score is the relevance score between any patent document in the patent document set and any project document in the project document set. Configure the first threshold, the second threshold, and the third threshold respectively; Connect two patent documents with a first correlation score greater than the first threshold, connect two project documents with a second correlation score greater than the second threshold, and connect patent documents and project documents with a third correlation score greater than the third threshold to form the edges of the evidence graph.

[0080] As an example, the evidence map includes at least the following relationships, as shown in Table 2: Table 2 Evidence Graph Relationship Table The dual-source evidence graph constructed based on the above nodes and relationships is as follows: Figure 3 As shown.

[0081] Technical features are extracted from document keywords, subject keywords, and the text dictionary, and then normalized. For example: "link 16" -> "link-16" "link16"->"link-16" "inter satellite" ->"inter-satellite" "lpi" ->"low probability of intercept" "lpd" ->"low probability of detection" "time synchronization" ->"network synchronization" "timing synchronization"->"network synchronization" Let the original term be The normalization function is The technical features are: If the word is empty after normalization, belongs to a stop word category, is purely numerical, is too short, or is not in the allowed list, it is discarded. Simultaneously, alias dictionary matching features are used from the text; feature edges obtained from text matching can be assigned lower weights, for example: in, This indicates technical features derived from the document keyword field. This indicates a technical feature derived from matching the alias dictionary in the main text. The evidence weights of the technical feature edges are represented.

[0082] For each technical feature and year ,statistics , , And calculate the annual percentage: in, Representation of features In the year The percentage of all documents Indicates the proportion of patents. Indicates the proportion of the project side. , , These represent the corresponding counts for all documents, patents, and projects, respectively. , , Representing the year Total number of documents, total number of patents, and total number of projects.

[0083] The above statistics provide a basis for setting the recent window and baseline window in step S8.

[0084] Next, a Semantic Bridging Pair (RSTP) algorithm is used to create edges in the evidence graph. RSTP represents the strength of evidence bridging that integrates semantic similarity, citation coupling, and co-occurrence of technical features. It can be understood as a score of relationship strength across documents, topics, and evidence sources, used to measure whether an evidence bridge that can be used to predict technological evolution is formed between two documents. RSTP integrates signals such as semantic similarity, citation coupling, direct citation relationships, and co-occurrence of technical features.

[0085] First, calculate the first relevance score, which is the relevance score between any two patent documents in the patent document set. For the patent document set... and project document collection First, construct the normalized vector matrix: in: The semantic similarity matrix of the patent projects is as follows: For each patent document, select the Top-K project documents, and for each project document, select the Top-K patent documents, retaining edges whose similarity exceeds a first threshold. For example, the first threshold is 0.62, and Top-K = 12 to 20.

[0086] Patents and patents ,set up For patent The collection of cited patents, For patent The set of cited patents. The citation coupling strength between the two is determined using the Ochiai coefficient: like direct quote or direct quote ,but: The inter-patent citation strength is: For RSTP edges between patents, consider both semantic similarity and citation strength: in, Indicates semantic similarity. Indicating reference strength, for example, , .

[0087] Regarding the relationship between patents and projects, and between projects, since there is no patent citation relationship between projects, it can be stated that: in, and "Any" means any two patents, any two projects, or any one patent and one project.

[0088] Each RSTP edge can also record document type, topic number, whether they are on the same topic, difference in event year, number of shared technical features, and bidirectional ranking.

[0089] The RSTP method proposed in this invention can not only discover documents with similar text, but also incorporate patent citations, technical features, and thematic relationships into bridging evidence. For technology evolution prediction, patent item RSTP edges are particularly important because they can indicate which existing or recent patented technical solutions have semantic proximity relationships to a recent project requirement.

[0090] S8. Preset the recent window and baseline window, and calculate the weak signal score and demand verification score for each technical feature based on the evidence spectrum, the recent window and the baseline window; As an example, a recent window and a baseline window are constructed based on the complete year of analysis. Let the complete year series be: Let the recent window length be The baseline window length is Preferably, , ,but: Taking space-based tactical data links as an example, the near-term data window is 2024 and 2025, and the baseline window is 2021, 2022, and 2023. The system excludes the incomplete year of 2026 to avoid misjudgments of trends caused by partial data collection.

[0091] See Figure 4 As one possible implementation, the weak signal score is calculated using the following method: For a given technical feature, obtain the number of first patent documents mentioned in the recent window, the number of second patent documents mentioned in the baseline window, the number of third patent documents mentioned in the last year of the recent window, and the number of fourth patent documents mentioned in the year before the recent window; preset the minimum baseline upper limit and the minimum scale threshold; The low baseline score is calculated based on the minimum baseline upper limit and the number of second patent documents; the growth score is calculated based on the number of first patent documents and the number of second patent documents; the acceleration score is calculated based on the number of third patent documents and the number of fourth patent documents; and the recent size score is calculated based on the number of first patent documents and the minimum size threshold. The weak signal score of this technical feature is obtained by weighting and summing the low baseline score, growth score, acceleration score and recent size score.

[0092] As an example, let the number of the first patent documents be... The number of second patent documents is The number of third patent documents is The number of fourth patent documents is The minimum baseline upper limit is The minimum size threshold is Then calculate the low baseline score. as follows: Calculate growth score as follows: in, Indicates the growth rate: Calculate acceleration score as follows: Calculate recent size score as follows: The weak signal score is obtained by weighting and summing the low baseline score, growth score, acceleration score, and recent size score. : For example, the weights for the low baseline score, growth score, acceleration score, and recent size score are: , , , This scoring system emphasizes the combination of "low baseline, recent growth, accelerated emergence, and reaching a minimum scale." It differs from typical popularity scoring because for high-frequency, mature technologies, a low baseline score will decrease if the baseline period is already high.

[0093] See Figure 4 As one possible implementation, the requirement verification score is calculated using the following method: For a certain technical feature, obtain the number of first project documents appearing in the recent window, the number of second project documents from priority requests in the recent window, the number of semantic bridge edges between patent documents and project documents related to the technical feature, the maximum semantic similarity between patent documents and project documents related to the technical feature, the number of active project topics in the recent window related to the technical feature, and the number of sources of projects in the recent window related to the technical feature. The bridging density score is calculated based on the number of semantic bridging edges; the semantic verification score is calculated based on the bridging density score and the maximum semantic similarity; the demand scale score is calculated based on the number of documents in the first project; the priority source score is calculated based on the number of documents in the second project; the project topic activity score is calculated based on the number of active project topics; and the source diversity score is calculated based on the number of sources. The bridging density score, semantic verification score, demand scale score, priority source score, project topic activity score, and source diversity score are weighted and summed to obtain the demand verification score for this technical feature.

[0094] As an example, let the number of documents in the first project be... The number of documents in the second project is The number of semantic bridge edges is The maximum semantic similarity is The number of active project topics is The number of project sources is Then calculate the bridging density score. as follows: Calculate semantic verification score as follows: Calculate the demand size score as follows: Calculate the priority source score as follows: Calculate the activity score of the project topic as follows: Calculate the source diversity score as follows: The requirement verification score for this specific technical feature is obtained by weighting and summing the bridging density score, semantic verification score, requirement size score, priority source score, project topic activity score, and source diversity score. : For example, the weights for bridging density score, semantic validation score, demand size score, priority source score, project topic activity score, and source diversity score are as follows: , , , , The technical effect of this scoring is that a higher demand verification score will only be obtained when the weak signal from the patent side can be cross-validated by recent project demands, priority demand sources, semantic bridging edges, and project theme activities.

[0095] S9. Weigh and fuse weak signal scores and demand verification scores to obtain a technology evolution prediction index. Based on the technology evolution prediction index, output a technology evolution path classification, a mechanism-level prediction list, and an evidence package.

[0096] See Figure 4 As one possible implementation, when weighting and fusing weak signal scores and demand verification scores, the weight of demand verification scores is greater than that of weak signal scores.

[0097] As an example, the technology evolution prediction index , This represents the weight of the weak signal score, with a value of 0.4. This indicates the weight of the requirement verification score, with a value of 0.6.

[0098] The reason for assigning a higher weight to demand verification than to weak patent signals is that this invention focuses on predicting technological evolution, rather than simply patent trends. Recent project demands can provide task-driven evidence, and therefore are given a higher weight in predicting technological evolution.

[0099] Furthermore, the criteria for determining high-value weak signals are defined as follows: in, This indicates the threshold for identifying high-value weak signals. Indicates technical features Whether it is judged as a high-value weak signal, for example, , , If a patent meets the criteria of low baseline and rising, but its EPI has not reached the high-value threshold and there is evidence of demand, it can be added to the cross-validation watchlist; if it meets the criteria of low baseline and rising but there is currently no evidence of demand, it can be added to the patent-side emerging watchlist.

[0100] As one possible implementation approach, the technology evolution path classification includes: demand-driven, technology reserve-driven, supply and demand synchronized, and single-source observation-driven; the mechanism-level prediction list includes: mechanism identifier, candidate parent technology feature set, mechanism name, technology claim, operational requirements, predicted evolution direction, alternative term set, simulation module, parameters and indicators; the evidence package includes: a list of patent document identifiers supporting the prediction results, a list of project document identifiers, a list of cross-source topic alignment edges, a list of semantic bridging edges between patents and projects, and the confidence level of the matching between technology features and mechanisms.

[0101] As an example, regarding technical features ,set up Indicates the year of the patent in which this technical feature first appeared. Indicates the year in which this technical feature first appeared in a project. This indicates the number of semantic bridging edges in a patent project. Technology evolution paths are classified according to the following rules: When the first year of the project for this technical feature is empty, it is determined to be a patent-side observation path; When the first year of the patent for this technical feature is blank, it is determined to be a demand-driven path; When the year of the first project for this technical feature is earlier than the year of the first patent, it is determined to be a demand-driven path; When the patent for this technical feature was first filed in a year earlier than the project's initial year, it is determined to be a technology reserve-type path; When both the first year of the patent and the first year of the project for the technical feature are not empty, and there is a semantic bridging edge between the patent document and the project document, it is determined to be a supply and demand semantic bridging synchronous path. Otherwise, it is determined to be a supply and demand synchronized path.

[0102] Simply outputting weak signal names is insufficient to constitute high-value technology predictions. This invention further matches weak signals with a mechanism catalog, bridging evidence, and an alias dictionary to output mechanism-level predictions. Let the mechanism catalog be: Each mechanism It should include at least: mechanism identifier, set of candidate parent technical features, mechanism name, technical claims, operational requirements, predicted evolution direction, set of alternative terms, simulation module, parameters, and indicators.

[0103] For candidate weak signals ,like Belongs to mechanism If the set of candidate parent features, or the shared features in the bridging evidence intersect with the candidate features of the mechanism, then the mechanism is considered applicable to the weak signal.

[0104] Bridging evidence for each patent project By concatenating the patent title, patent fragment, project title, project fragment, and shared technical features, the mechanism matching text is obtained: If the mechanism alias set The terms appear in In this case, it is denoted as the mechanism matching signal: By summarizing the matching signals from each bridging piece of evidence, mechanism prediction evidence is formed.

[0105] Let the number of pieces of evidence predicted by a certain mechanism be . The average semantic similarity is The number of different matching signals is The parent weak signal EPI is .definition: in, Representation mechanism Evidence density score, Representation mechanism The specificity score of the matched signal, Indicates a weak signal from the parent level. The normalized score of EPI.

[0106] The confidence level of the mechanism prediction is: in, Indicates semantic similarity. Indicates the specificity of the mechanism signal. Indicates the strength of the weak signal from the parent level. , , , These represent the weights of the number of pieces of evidence, semantic similarity, specificity of the mechanism signal, and strength of the parent weak signal, respectively. For example, , , , This confidence level considers the quantity of evidence, semantic similarity, the specificity of the mechanism signal, and the strength of the parent weak signal.

[0107] Each mechanism-level prediction output should include at least: parent weak signal, mechanism name, prediction claim, operational requirements, prediction evolution direction, confidence level, number of evidence items, matching signal, representative patent bridging evidence, and simulable modules, parameters, and metrics.

[0108] The technical solution of the present invention will be further described below with reference to specific embodiments.

[0109] In one embodiment for space-based tactical data links and space-assisted Link-16, the collected and processed data are shown in Table 3 below: Table 3. Number of data acquisition metrics for space-based tactical data links and space-assisted Link-16 After modeling the diversionary themes, we obtain the patent themes and project themes, as shown in Table 4 below: Table 4 Summary Table of Patent Topics and Project Topics These themes illustrate that the patent side and the project side have respectively formed the themes of technology reserves and demand-driven themes.

[0110] Alignment edges are generated based on the cosine similarity of the topic centroid. Table 5 shows some of the highly similar aligned patent topics and project topics: Table 5. Patent and Project Topics with High Similarity Alignment The results indicate that multiple patent supply-side themes are highly aligned with the project demand-side theme of "space tactical data link network," suggesting a clear semantic connection between patent reserves and recent project needs.

[0111] Recent window and baseline window analyses were performed on 29 patent-side technical features, identifying two high-value weak signals, as shown in Table 6 below: Table 6. Two high-value weak signals identified For Link-16, a low recent patent baseline was observed, but a rise is expected in 2024-2025. Simultaneously, there are 10 recent demand records and 38 semantic bridging edges for patent projects on the demand side, with demand primarily originating from SBIR and SAM. This weak signal not only indicates that Link-16 remains a focus of attention but also suggests its evolution from a traditional tactical data link towards space transport layer, tactical edge gateway, dynamic anti-jamming, and low-latency TDMA control.

[0112] Regarding Network synchronization, the baseline number of patents was 0, the recent number of patents was 2, the recent number of projects was 2, and the number of semantic bridging edges was 27. Evidence suggests that this term should not be understood merely as ordinary network synchronization, but rather as the self-synchronizing tactical network capability under conditions of GNSS interference, base station unavailability, or infrastructure deficiency.

[0113] For Link-16, this scheme outputs the following mechanism-level prediction: 1. Interference Resilience Dynamic Reconfiguration: Evidence bridging shows that the requirements of Advanced Interference Mitigation, DYNAMO, adaptive power control, dynamic relay allocation, etc. are semantically connected with the patented technology, predicting that Link-16 neighboring systems will add a software control loop to select relay, power, routing or waveform configuration according to interference and congestion status.

[0114] 2. PWSA Tactical Edge Gateway Adaptation: Evidence bridging shows that the requirements of PWSA Tactical Edge, SEARUS, MAVLink, STANAG 4586, CEC, Link 11 / 22, etc. are similar to the Link-16 patented technology, predicting that Link-16 will evolve from a single-link capability to a gateway adaptation layer between the tactical edge and the space transport layer.

[0115] 3. Low-latency, long-distance TDMA control: Evidence bridging reveals mechanisms such as low-latency subnets, long-distance time bases, fastterminal entry, and concurrent multi-netting. Predicting that the spatially assisted Link-16 will focus on solving time slot order, priority, and low-latency control under long propagation paths.

[0116] For network synchronization, this solution outputs the following mechanism-level prediction: 1. PNT Degradation and Self-Synchronization Tactical Networks: Evidence bridging shows the need for infrastructureless 5G, Alternative PNT, and Beyond 5G resilient tactical networks, predicting that synchronization mechanisms will shift to resynchronization recovery under GNSS degradation, base station unavailability, and self-organizing network conditions.

[0117] 2. LPI / LPD reconfigurable waveform control: Evidence bridging demonstrates FPGA-based LPx Waveform, low probability of detection, low probability of interception, and frequency hopping mechanisms. Predicting concealed anti-jamming waveforms will become a companion mechanism for network synchronization and tactical link resilience.

[0118] Secondly, the present invention provides a technology evolution prediction system based on dual-source evidence fusion, comprising: The dual-source evidence pool construction unit acquires patent supply-side data and project demand-side data to construct a patent supply evidence pool and a project demand evidence pool. The standard field set generation unit extracts document identifiers from the patent supply evidence pool and the project demand evidence pool, normalizes all document identifiers, and obtains the standard field set of the document. The document collection construction unit uses a text embedding model to convert the standard field set into semantic vectors, and constructs a patent document collection and a project document collection respectively. The patent document collection includes documents and their semantic vectors in the patent supply evidence pool after normalization, and the project document collection includes documents and their semantic vectors in the project demand evidence pool after normalization. The topic set acquisition unit performs BERTopic topic modeling on the patent document set and the project document set respectively to obtain the patent topic set and the project topic set; The cross-source topic alignment edge generation unit calculates normalized topic centroids based on the semantic vectors of documents in the patent topic set and the project topic set, and calculates the cosine similarity between each patent topic centroid and each project topic centroid. It configures a similarity threshold, retains topic pairs with a cosine similarity greater than the threshold and both sides ranking in the top K positions, and generates cross-source topic alignment edges. It then calculates normalized topic centroids based on the semantic vectors of documents in the patent topic set and the project topic set, and calculates the cosine similarity between each patent topic centroid and each project topic centroid. Finally, it configures a similarity threshold and a candidate number K, and selects the top K cross-source topic candidate pairs with a cosine similarity greater than or equal to the similarity threshold from both the patent topic side and the project topic side, and merges and deduplicates the candidate results from both sides to generate cross-source topic alignment edges.

[0119] The evidence graph construction unit constructs an evidence graph based on joint analysis documents, semantic vectors, and cross-source topic alignment edges. The evidence graph includes document nodes, topic nodes, organization nodes, year nodes, and technical feature nodes. The scoring unit has a preset recent window and a baseline window. Based on the evidence spectrum, the recent window, and the baseline window, it calculates the weak signal score and demand verification score for each technical feature. The output unit calculates and fuses the weak signal score and the demand verification score to obtain the technology evolution prediction index. Based on the technology evolution prediction index, it outputs a technology evolution path classification, a mechanism-level prediction list, and an evidence package.

[0120] Thirdly, the present invention provides a terminal including a processor and a communication interface coupled to the processor, the processor being used to run computer programs or instructions to implement the technology evolution prediction method based on dual-source evidence fusion provided in the first aspect.

[0121] Although the invention has been described herein in conjunction with various embodiments, those skilled in the art will understand and implement other variations of the disclosed embodiments by reviewing the accompanying drawings, the disclosure, and the description of the drawings, in carrying out the claimed invention. In this specification, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude multiple components. A single processor or other unit can implement several of the functions listed in the specification. While certain measures are described in different embodiments, this does not mean that these measures cannot be combined to produce good results.

[0122] Although the invention has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made therein without departing from the spirit and scope of the invention. Accordingly, this specification and drawings are merely illustrative of the invention and are considered to cover any and all modifications, variations, combinations, or equivalents within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if such modifications and modifications fall within the scope of the invention and its equivalents, the invention is also intended to include such modifications and modifications.

Claims

1. A method for predicting technology evolution based on dual-source evidence fusion, characterized in that, Includes the following steps: S1. Obtain patent supply-side data and project demand-side data, and construct a patent supply evidence pool and a project demand evidence pool; S2. Extract document identifiers from the patent supply evidence pool and the project demand evidence pool, normalize all document identifiers, and obtain the standard field set of the documents; S3. Use a text embedding model to convert the standard field set into semantic vectors; S4. Construct a patent document set and a project document set respectively. The patent document set includes documents and their semantic vectors from the patent supply evidence pool after normalization. The project document set includes documents and their semantic vectors from the project demand evidence pool after normalization. S5. Perform BERTopic topic modeling on the patent document set and the project document set respectively to obtain the patent topic set and the project topic set; S6. Calculate normalized topic centroids based on the semantic vectors of documents in the patent topic set and project topic set, and calculate the cosine similarity between each patent topic centroid and each project topic centroid; configure the similarity threshold and the number of candidates K, and select the top K cross-source topic candidate pairs with cosine similarity greater than or equal to the similarity threshold from the patent topic set and project topic set respectively, merge and deduplicate the selection results, and generate cross-source topic alignment edges; S7. Construct an evidence graph based on a standard field set, semantic vectors, and cross-source topic alignment edges. The evidence graph includes document nodes, topic nodes, organization nodes, year nodes, and technical feature nodes. S8. Preset the recent window and baseline window, and calculate the weak signal score and demand verification score for each technical feature based on the evidence spectrum, the recent window and the baseline window; S9. Weigh and fuse weak signal scores and demand verification scores to obtain a technology evolution prediction index. Based on the technology evolution prediction index, output a technology evolution path classification, a mechanism-level prediction list, and an evidence package. 2.The technology evolution prediction method based on dual-source evidence fusion according to claim 1, characterized in that, The following method is used to construct the patent supply evidence pool: Patent data is obtained from a patent database, and field completion, key field extraction, and patent relevance scoring are performed to generate a patent supply-side evidence pool. The patent relevance score includes at least the following indicators, which are sorted by weight from high to low as follows: Co-occurrence of spatial and tactical data links, key organizational hits, core keyword hits, government interest statements, classification number hits, tactical data link keyword hits, resilience mechanism term hits, spatial term hits, citation counts; The patent relevance score also includes an exclusion word penalty item. When the patent text contains an exclusion word, a penalty score is deducted from the score. The following method was used to construct the project requirements evidence pool: Project demand-side data is collected from multiple sources. This data undergoes multi-source entity deduplication and merging, and project relevance scoring is performed to generate a project demand-side evidence pool. The project relevance score includes at least the following indicators, which are ranked from highest to lowest weight: Tactical data chain keyword hit, union of demand keywords and domain keywords hit, key organizational structure hit, resilience mechanism keyword hit, spatial keyword hit, and response deadline exists. 3.The technology evolution prediction method based on dual-source evidence fusion according to claim 1, characterized in that, The evidence map was constructed using the following method: Create evidence graph nodes, which include: document nodes, topic nodes, organization nodes, year nodes, and technical feature nodes; The edges for creating the evidence graph include: Calculate the first relevance score, the second relevance score, and the third relevance score respectively; wherein, the first relevance score is the relevance score between any two patent documents in the patent document set; the second relevance score is the relevance score between any two project documents in the project document set; and the third relevance score is the relevance score between any patent document in the patent document set and any project document in the project document set. Configure the first threshold, the second threshold, and the third threshold respectively; Connect two patent documents with a first correlation score greater than the first threshold, connect two project documents with a second correlation score greater than the second threshold, and connect patent documents and project documents with a third correlation score greater than the third threshold to form the edges of the evidence graph. 4.The technology evolution prediction method based on dual-source evidence fusion according to claim 1, wherein, The weak signal score is calculated using the following method: For a given technical feature, obtain the number of first patent documents mentioned in the recent window, the number of second patent documents mentioned in the baseline window, the number of third patent documents mentioned in the last year of the recent window, and the number of fourth patent documents mentioned in the year before the recent window; preset the minimum baseline upper limit and the minimum scale threshold; The low baseline score is calculated based on the minimum baseline upper limit and the number of second patent documents; the growth score is calculated based on the number of first patent documents and the number of second patent documents; the acceleration score is calculated based on the number of third patent documents and the number of fourth patent documents; and the recent size score is calculated based on the number of first patent documents and the minimum size threshold. The weak signal score for this technical feature is obtained by weighting and summing the low baseline score, growth score, acceleration score, and recent scale score. 5.The technology evolution prediction method based on dual-source evidence fusion according to claim 1, wherein, The requirements verification score is calculated using the following method: For a certain technical feature, obtain the number of first project documents appearing in the recent window, the number of second project documents from priority requests in the recent window, the number of semantic bridge edges between patent documents and project documents related to the technical feature, the maximum semantic similarity between patent documents and project documents related to the technical feature, the number of active project topics in the recent window related to the technical feature, and the number of sources of projects in the recent window related to the technical feature. The bridging density score is calculated based on the number of semantic bridging edges; the semantic verification score is calculated based on the bridging density score and the maximum semantic similarity; the demand scale score is calculated based on the number of documents in the first project; the priority source score is calculated based on the number of documents in the second project; the project topic activity score is calculated based on the number of active project topics; and the source diversity score is calculated based on the number of sources. The bridging density score, semantic verification score, demand scale score, priority source score, project topic activity score, and source diversity score are weighted and summed to obtain the demand verification score for this technical feature.

6. The technology evolution prediction method based on dual-source evidence fusion according to claim 1, characterized in that, When weighted and fused weak signal scores and demand verification scores, the weight of demand verification scores is greater than that of weak signal scores.

7. The technology evolution prediction method based on dual-source evidence fusion according to claim 1, characterized in that, The technology evolution path classification includes: demand-driven, technology reserve-driven, supply and demand synchronous, and single-source observation-driven; the mechanism-level prediction list includes: mechanism identifier, candidate parent technology feature set, mechanism name, technology claim, operational requirements, predicted evolution direction, alternative term set, simulation module, parameters and indicators; the evidence package includes: a list of patent document identifiers supporting the prediction results, a list of project document identifiers, a list of cross-source topic alignment edges, a list of semantic bridge edges between patents and projects, and the confidence level of the matching between technology features and mechanisms.

8. A technology evolution prediction system based on dual-source evidence fusion, characterized in that, include: The dual-source evidence pool construction unit acquires patent supply-side data and project demand-side data to construct a patent supply evidence pool and a project demand evidence pool. The standard field set generation unit extracts document identifiers from the patent supply evidence pool and the project demand evidence pool, normalizes all document identifiers, and obtains the standard field set of the document. The document collection construction unit uses a text embedding model to convert the standard field set into semantic vectors, and constructs a patent document collection and a project document collection respectively. The patent document collection includes documents and their semantic vectors in the patent supply evidence pool after normalization, and the project document collection includes documents and their semantic vectors in the project demand evidence pool after normalization. The topic set acquisition unit performs BERTopic topic modeling on the patent document set and the project document set respectively to obtain the patent topic set and the project topic set; The cross-source topic alignment edge generation unit calculates normalized topic centroids based on the semantic vectors of documents in the patent topic set and the project topic set, and calculates the cosine similarity between each patent topic centroid and each project topic centroid; it configures a similarity threshold and a candidate number K, and selects the top K cross-source topic candidate pairs with a cosine similarity greater than or equal to the similarity threshold from the patent topic set and the project topic set respectively, merges and removes duplicates from the selection results, and generates cross-source topic alignment edges; The evidence graph construction unit constructs an evidence graph based on joint analysis documents, semantic vectors, and cross-source topic alignment edges. The evidence graph includes document nodes, topic nodes, organization nodes, year nodes, and technical feature nodes. The scoring unit has a preset recent window and a baseline window. Based on the evidence spectrum, the recent window, and the baseline window, it calculates the weak signal score and demand verification score for each technical feature. The output unit calculates and fuses the weak signal score and the demand verification score to obtain the technology evolution prediction index. Based on the technology evolution prediction index, it outputs a technology evolution path classification, a mechanism-level prediction list, and an evidence package.

9. A terminal, characterized in that, It includes a processor and a communication interface coupled to the processor, the processor being used to run computer programs or instructions to implement the technology evolution prediction method based on dual-source evidence fusion as described in any one of claims 1 to 7.

Images