Enterprise technology mining data analysis management system based on large language model

Abstract

The application relates to the technical field of enterprise R&D data analysis and artificial intelligence text processing, and discloses an enterprise technology mining data analysis management system based on a large language model, which comprises a text structured analysis module, a double-layer knowledge graph construction module, a time sequence conflict matrix calculation module, an extraction rule feedback control module and a local topological verification and output module.The text structured analysis module receives a multi-source technology text stream, calls a large language model to extract a structured tuple set, generates a three-state trend vector value corresponding to an engineering parameter, and takes the three-state trend vector value as a substitute state parameter when an absolute numerical boundary is missing.The double-layer knowledge graph construction module establishes a technical component structure layer and a parameter constraint layer, and assigns edge weights.The time sequence conflict matrix calculation module generates a time sequence conflict adjacency matrix and calculates a coupling interference degree value.The extraction rule feedback control module generates a directional constraint template.The local topological verification and output module extracts a K-Hop local neighborhood subgraph, calculates an approximate graph editing distance and generates a technology evolution analysis report.The application can maintain parameter coupling relationship calculation continuity and improve enterprise technology evolution analysis accuracy.

Description

Technical Field

[0001] This invention relates to the fields of enterprise R&D data analysis and artificial intelligence text processing technology, specifically to an enterprise technology mining data analysis and management system based on a large language model. Background Technology

[0002] Enterprise technology research and development typically relies on a large amount of historical documentation. R&D personnel need to analyze past patent documents, internal R&D documents, and engineering test records to understand the patterns of technological evolution and to help formulate new product development strategies. Technical documents contain information such as component iteration paths, changes in engineering parameters, and parameter constraints. Systematically organizing this information helps identify product design bottlenecks and shortens the technology pre-research cycle.

[0003] Current data analysis systems widely incorporate natural language processing (NLP) techniques for text mining. Conventional methods can scan massive amounts of unstructured documents and extract explicitly disclosed entity names, parameter names, and experimental values ​​in batches using pre-defined matching rules. After categorizing and writing the extracted results into a graph database, the system can generate macro-level statistical graphs, enabling researchers to perform cross-format document retrieval and technical terminology analysis.

[0004] However, many engineering documents tend to use qualitative terms to describe performance changes. When specific numbers and units of measurement are lacking in the text, conventional systems struggle to perform structured analysis of the technical features, easily overlooking parameter information with directional change intentions, resulting in isolated entities in the graph and difficulty in establishing deep parameter coupling relationships. On the other hand, technical conflicts often recur over long periods, while existing analytical models often perform indiscriminate extraction of incoming text according to fixed templates, making it difficult to adjust the analytical scope based on the cumulative state of historical conflicts and to focus on monitoring key technical components. When performance indicators improve in subsequent texts, existing tools often rely on adding keywords or statistical frequency to infer technical trends, lacking quantitative means to assess the differences in the local technical graph structure between different stages. It is difficult to distinguish whether the improvement originates from system-level component structural reconstruction or from improvements in the original component's materials, processes, control algorithms, etc., making it difficult to directly convert the output results into component-level R&D reference information. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an enterprise technology mining data analysis and management system based on a large language model. This system solves the problems of existing technologies, such as difficulty in continuously parsing parameter relationships in engineering texts lacking explicit numerical boundaries, inability to adjust the extraction range based on historical conflict feedback, and difficulty in quantifying differences in component-level technology evolution.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an enterprise technology mining data analysis and management system based on a large language model, comprising:

[0007] The text structure parsing module is used to receive multi-source technical text streams, call a large language model to extract a set of structured tuples, perform directional semantic classification on the statements containing engineering parameters to generate three-state trend vector values, and use the three-state trend vector values ​​as alternative state parameters when the absolute numerical boundaries of the engineering parameters are missing.

[0008] A two-layer knowledge graph construction module is used to establish a technical component structure layer and a parameter constraint layer based on the structured tuple set, configure cross-layer mounting pointers, and allocate edge weights between engineering parameters in the parameter constraint layer according to the three-state trend vector values.

[0009] The temporal conflict matrix calculation module is used to convert the parameter constraint layer into a temporal conflict adjacency matrix, and calculate the coupling interference degree of the target parameter pair consisting of the starting engineering parameter and the target engineering parameter corresponding to the negative weight data in the edge weights.

[0010] The extraction rule feedback control module is used to determine the target entity node set based on the cross-layer mounting pointer when the coupling interference degree value exceeds the interference degree trigger threshold, and generate a directional constraint template. The directional constraint template is fed back to the text structure parsing module so that when the subsequent input text hits the entity corresponding to the target entity node set, the text parsing range is limited based on the directional constraint template.

[0011] The local topology verification and output module is used to extract the K-Hop local neighborhood subgraph associated with the target entity node set when the edge weight between the target parameter pairs in the parameter constraint layer changes from a historical negative value to a zero value or a positive value, calculate the approximate graph edit distance, and generate a technology evolution analysis report based on the approximate graph edit distance.

[0012] Preferably, the structured tuple set contains multiple tuple objects, and the tuple objects include at least technical component entities, engineering problems, engineering parameters, absolute numerical boundaries, three-state trend vector values, coupling influence parameters, and extraction confidence.

[0013] The text structure parsing module is specifically used to perform the following operations:

[0014] The combination of numeric characters and units of measurement in adjacent text segments of the engineering parameters is used as the absolute numerical boundary.

[0015] The large language model is instructed to perform directional semantic classification on the statement containing the engineering parameter, and output the three-state trend vector value with a value of +1, 0 or −1;

[0016] When the absolute numerical boundary is missing, a trend state mapping instruction is triggered, and the three-state trend vector value is used as a substitute state parameter for the absolute numerical boundary.

[0017] When the extraction confidence level is lower than the preset confidence threshold, the tri-state trend vector value in the corresponding tuple object is not used for the calculation of the negative weight data, or the corresponding tuple object is written into the manual review queue.

[0018] Preferably, the two-layer knowledge graph construction module is specifically used to perform the following operations:

[0019] The extracted technical component entities are constructed as entity nodes of the technical component structure layer;

[0020] The extracted engineering parameters are constructed as parameter nodes of the parameter constraint layer;

[0021] Write the globally unique identifier of the technical component entity into the parameter node attribute corresponding to the engineering parameter to generate the cross-layer mounting pointer;

[0022] When the coupling influence parameter exists in the tuple object, a directed edge is generated between the originating parameter node and the target parameter node, where both the originating parameter node and the target parameter node are parameter nodes in the parameter constraint layer.

[0023] Retrieve associated tuple objects in the structured tuple set that match the engineering parameters with the coupling influence parameters, and read the tri-state trend vector value in the associated tuple object as the tri-state trend vector value corresponding to the target parameter node;

[0024] Compare the tri-state trend vector values ​​corresponding to the initiating parameter node and the target parameter node with the expected evolution direction label;

[0025] When the originating parameter node is in a performance optimization state and the target parameter node enters a performance degradation state, the edge weight of the directed edge is assigned to the negative weight data of a preset constant.

[0026] Preferably, the temporal conflict matrix calculation module is specifically used to perform the following operations when converting the parameter constraint layer into a temporal conflict adjacency matrix:

[0027] Establish a global project parameter node pool;

[0028] Assign a fixed matrix index to each engineering parameter in the parameter constraint layer;

[0029] The capacity of the global engineering parameter node pool is dynamically expanded when a new engineering parameter is detected.

[0030] Add a new fixed matrix index to the newly added engineering parameters;

[0031] In the temporal conflict adjacency matrix of the historical time node, zero-padding is performed on the row and column positions corresponding to the new fixed matrix index.

[0032] Preferably, the time-series conflict matrix calculation module performs the following operations when performing time-series weighted accumulation on the negative weight data:

[0033] For the target parameter pairs within a specific observation period, obtain the absolute value scalar of the negative weight data at each time node;

[0034] Calculate the time span scalar between the current time point and the previous time point;

[0035] The time decay weight corresponding to each time node is calculated using an exponential decay method.

[0036] The absolute value scalar, the time span scalar, and the time decay weight are multiplied to obtain the product value;

[0037] The product values ​​are accumulated and the coupling interference degree value is output.

[0038] Preferably, the extraction rule feedback control module is specifically used to perform the following operations:

[0039] When the coupling interference value exceeds the interference trigger threshold, read the cross-layer mounting pointer of each engineering parameter in the target parameter pair;

[0040] A search is initiated at the technical component structure layer based on the globally unique identifier recorded in the cross-layer mounting pointer;

[0041] The shortest path algorithm is invoked to search for component connection paths in the technical component structure layer that connect the technical component entities corresponding to the two engineering parameters in the target parameter pair;

[0042] The set of components and endpoint entities involved in the component connection path are marked as the target entity node set;

[0043] When there is no connection path between the two technical component entities, the two technical component entities are added to the target entity node set respectively.

[0044] Preferably, the specific logic of the extraction rule feedback control module in generating the directional constraint template to constrain the subsequent text parsing range is as follows:

[0045] When the input text matches the entity name or alias corresponding to the target entity node set, the orientation constraint template is activated to perform text parsing, and a local text window is determined based on the position of the entity name or alias in the input text;

[0046] When the input text does not match the entity name or entity alias corresponding to the target entity node set, the initial constraint template is used to perform text parsing.

[0047] When the large language model is in interface call mode, only the text sequence within the local text window is input into the large language model;

[0048] When the large language model is deployed locally and the inference framework has the attention mask parameter enabled, the attention mask matrix elements corresponding to the text identifiers outside the local text window are set to negative infinity.

[0049] In the directional constraint template, the decision priority of the three-state trend vector value is set to be higher than that of the three-state trend vector value in the initial constraint template.

[0050] Preferably, the local topology verification and output module is specifically used to perform the following operations when extracting the K-Hop local neighborhood subgraph associated with the target entity node set:

[0051] Send a write lock command to the system main control unit for the target parameter pair and its associated map region;

[0052] Using the target entity node set as the seed node, a restricted breadth-first search algorithm is executed in the technical component structure layer. The execution logic of the restricted breadth-first search algorithm is to expand the search for adjacent nodes layer by layer according to a set number of hops.

[0053] Extract the topology from the graph cross section corresponding to the old time node and the graph cross section corresponding to the new time node, respectively;

[0054] The extracted topology is used to generate old K-Hop local neighborhood subgraphs and new K-Hop local neighborhood subgraphs respectively;

[0055] The new time node is the time node at which the edge weight between the target parameter pairs in the parameter constraint layer changes from a negative value to a zero value or a positive value, and triggers the extraction of the K-Hop local neighborhood subgraph; the old time node is determined by the alarm record that is closest to the new time node before the new time node, and the alarm record is the alarm record formed when the edge weight between the target parameter pairs in the parameter constraint layer is negative and the coupling interference degree of the target parameter pairs exceeds the interference degree trigger threshold.

[0056] Preferably, the local topology verification and output module is specifically used to perform the following operations when calculating the approximate map editing distance:

[0057] A heuristic algorithm based on bipartite graph matching is invoked to compare the differences between the old K-Hop local neighborhood subgraph and the new K-Hop local neighborhood subgraph;

[0058] Calculate the edit distance of the node name strings between nodes in the old subgraph and nodes in the new subgraph;

[0059] Extract the set of locally connected edges associated with each node in the old subgraph and each node in the new subgraph, and calculate the difference in the number of edges and the difference in the in-degree direction;

[0060] The node replacement cost is generated by normalizing and summing the edit distance of the node name string, the difference in the number of edges, and the difference in the in-degree direction;

[0061] Unmatched detached nodes in the old subgraph are included in the cost of deleting nodes, and unmatched detached nodes in the new subgraph are included in the cost of inserting nodes;

[0062] Find the minimum total cost in the bipartite graph mapping scheme space, and use the minimum total cost as the approximate graph edit distance.

[0063] Preferably, the local topology verification and output module is specifically used to perform the following operations when generating the technology evolution analysis report:

[0064] Extract the cardinality of the first number of nodes contained in the old K-Hop local neighborhood subgraph and the cardinality of the second number of nodes contained in the new K-Hop local neighborhood subgraph respectively;

[0065] The maximum value between the first and second node number bases is multiplied by a preset topology tolerance coefficient to obtain the topology change threshold.

[0066] When the approximate graph editing distance is greater than the topology change threshold, the current technology evolution path is classified as a structural reconstruction type evolution, and the unmatched old subgraph free nodes and the unmatched new subgraph free nodes are marked as structural change nodes;

[0067] When the approximate graph editing distance is not greater than the topology change threshold, the current technology evolution path is classified as an original component attribute improvement candidate path, and material properties, manufacturing process, control algorithm, component model, operating condition, interface protocol or environmental constraint fields are extracted a second time from one or more input texts corresponding to the new time node that are used to determine whether the edge weights between the target parameter pairs change from negative to zero or positive as R&D focus items.

[0068] The target parameter pair, each engineering parameter in the target parameter pair, and the corresponding technical component entity, evolution path type, approximate graph editing distance, topology change threshold, structural change node or R&D focus item, along with the absolute numerical boundary or the three-state trend vector value, are encapsulated into the technology evolution analysis report.

[0069] This invention provides an enterprise technology mining data analysis and management system based on a large language model. It has the following beneficial effects:

[0070] 1. This invention uses a large language model to generate tri-state trend vector values ​​and combines this with two-layer knowledge graph rules to assign positive and negative weights to the connected edges of parameter nodes. Even when text lacks absolute numerical boundaries, it can still establish coupling relationships between parameters based on the direction of parameter changes and allow these relationships to participate in subsequent calculations. Compared to existing solutions that rely on explicit numerical extraction to construct node connections, this invention reduces the risk of data processing link interruptions caused by fuzzy engineering descriptions and minimizes the omission of implicit influence relationships.

[0071] 2. This invention employs a temporal conflict adjacency matrix accumulation to calculate coupling interference degree and generates a directional constraint template after triggering a threshold. This limits the local parsing window of a large language model, enabling the identification of frequently occurring candidate coupling conflict objects and increasing the parsing priority of text related to the target entity. Compared to schemes that continuously perform global indiscriminate extraction of streaming text, this invention reduces the consumption of parsing resources by irrelevant text and adjusts the extraction range based on analysis feedback.

[0072] 3. This invention employs a negative edge weight state change monitoring mechanism. When the edge weight of a target parameter pair changes from negative to zero or positive, it extracts the new and old K-Hop local neighborhood subgraphs and calculates the approximate graph edit distance. Combined with a dynamic topology change threshold to classify the evolution path type, it can characterize the degree of structural change of corresponding technical components during the suspected mitigation or resolution of candidate coupling conflict objects, and output candidate paths for structural reconstruction or improvement of original component attributes. Compared to solutions that rely solely on the frequency of text keywords at different times to deduce technical trends, this invention provides component-level topology change evidence and R&D focus areas. Attached Figure Description

[0073] Figure 1 This is a schematic diagram of the system structure according to an embodiment of the present invention;

[0074] Figure 2 This is a schematic diagram of the method flow according to an embodiment of the present invention;

[0075] Figure 3 This is a flowchart illustrating the text structure parsing and tuple extraction process in an embodiment of the present invention.

[0076] Figure 4 This is a flowchart illustrating the principle of two-layer knowledge graph construction and association mapping in an embodiment of the present invention.

[0077] Figure 5 This is a flowchart illustrating the principle of temporal conflict adjacency matrix generation and temporal weighted accumulation calculation in an embodiment of the present invention.

[0078] Figure 6 This is a flowchart illustrating the principle of directional constraint template generation and cross-layer entity query in an embodiment of the present invention.

[0079] Figure 7 This is a flowchart illustrating the principle of local subgraph truncation extraction and approximate graph matching verification in an embodiment of the present invention.

[0080] Figure 8 This is a flowchart of the asynchronous topology change verification and technology evolution analysis report output of an embodiment of the present invention;

[0081] Figure 9 This is a time-series change diagram of data analysis under the scenario of the evolution of drive motor technology in an embodiment of the present invention, where (a) is the step change curve of the target parameter in the parameter constraint layer with respect to the original edge weight state, and (b) is the change curve of the coupling interference degree value with time weighted accumulation. Detailed Implementation

[0082] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0083] Reference Figure 1 This invention provides an enterprise technology mining data analysis and management system based on a large language model, including: a text structure parsing module, a two-layer knowledge graph construction module, a temporal conflict matrix calculation module, an extraction rule feedback control module, and a local topology verification and output module.

[0084] In this embodiment, the enterprise technology mining data analysis and management system based on a large language model is deployed on a server device. The server device is equipped with a central processing unit (CPU), a graphics processing unit (GPU), memory, and a network communication interface. The memory contains a graph database software environment for persistent storage and retrieval of technical component nodes, engineering parameter nodes, node-related edges, and edge weight data. The CPU performs text scheduling, data verification, graph database read / write operations, and inter-module task control. The CPU runs the system main control program and forms the system main control unit, which performs module scheduling, graph database write control, locking of related graph regions, cache management, and write lock release. The GPU performs forward inference of the large language model and matrix time-series weighted accumulation calculations.

[0085] The text structuring parsing module receives multi-source technical text stream data, including patent publications, internal R&D documents, technical specifications, and R&D records. It then uses a large language model to identify information in the unstructured text according to a pre-defined initial constraint template. The text structuring parsing module extracts technical elements from the text and outputs a set of structured tuples. Based on the text semantics, the module classifies the direction of change of engineering parameters into discrete three-state trend vector values. For text content where absolute numerical boundaries are not explicitly disclosed, the three-state trend vector values ​​are used as substitute state parameters.

[0086] The two-layer knowledge graph construction module acquires a set of structured tuple data and instantiates a technical component structure layer and a parameter constraint layer in a graph database. The module maps technical component entities to nodes in the technical component structure layer and engineering parameters to nodes in the parameter constraint layer. Based on the combination determination rules of the three-state trend vector values, the module assigns positive or negative weight values ​​to the connected edges between nodes in the parameter constraint layer.

[0087] The temporal conflict matrix calculation module transforms the network topology of the parameter constraint layer into a temporal conflict adjacency matrix based on the timestamp sequence of the text stream. This module performs a temporal weighted accumulation operation on the persistent negative weight data in the temporal conflict adjacency matrix and calculates the coupling interference degree of the target parameter pairs. This coupling interference degree characterizes the sustained strength of the negative coupling relationship between the target parameter pairs in the historical text sequence.

[0088] The rule extraction feedback control module monitors changes in the coupling interference degree. When the coupling interference degree exceeds a set interference trigger threshold, the rule extraction feedback control module locates the target entity node corresponding to the target parameter pair in the technical component structure layer. The module then uses the information from the target entity node to update the input constraint template of the large language model and generates a directional constraint template.

[0089] The local topology verification and output module monitors the state of the original edge weights between target parameter pairs in the parameter constraint layer. When it determines that an original edge weight has changed from a historical negative value to zero or a positive value, the local topology verification and output module initiates an asynchronous local verification task. The module extracts the K-Hop local neighborhood subgraphs before and after the time change and calculates the approximate graph edit distance. Based on the approximate graph edit distance, the module generates a technology evolution analysis report containing the target range of engineering parameters.

[0090] Reference Figure 2 This invention provides a method for enterprise technology mining data analysis and management based on a large language model, comprising the following steps:

[0091] To facilitate the explanation of the data transmission relationships between the modules in the system of this invention, the method flow of this embodiment is described below in the order of multi-source technical text access, structured parsing, graph construction, matrix calculation, feedback control, and topology verification output. The data processing order between each step can be executed sequentially according to the text stream timestamps, or it can be scheduled and executed in parallel by the server device if data dependencies are satisfied.

[0092] S100 receives multi-source technical text streams with discrete timestamp features, calls a large language model to parse the unstructured text, and outputs a set of structured tuples containing three-state trend vector values.

[0093] S200: Obtain the set of structured tuples and establish a technical component structure layer and parameter constraint layer in the graph database. Assign positive or negative weight values ​​to the connected edges between nodes in the parameter constraint layer.

[0094] S300 transforms the network topology of the parameter constraint layer into a temporal conflict adjacency matrix, performs temporal weighted accumulation on the negative weight data in the temporal conflict adjacency matrix, and calculates the coupling interference degree.

[0095] S400 monitors the coupling interference degree value and, when the coupling interference degree value exceeds the interference degree trigger threshold, determines the target entity node set and generates a directional constraint template. When the input text hits the entity name or entity alias corresponding to the target entity node set, it performs data parsing within a limited range based on the directional constraint template.

[0096] S500 monitors the state of the original edge weights between target parameter pairs in the parameter constraint layer. When the original edge weights change from historical negative values ​​to zero or positive values, it extracts the K-Hop local neighborhood subgraphs before and after the time change and calculates the approximate graph editing distance.

[0097] S600 compares the approximate map editing distance with the preset topology change threshold and classifies the technology evolution path type, generating a technology evolution analysis report.

[0098] The following section explains the specific implementation of each step in the above process, taking into account data fields, graph structure, matrix calculations, and output rules.

[0099] Reference Figure 3 The method step S100 provided by the present invention specifically includes the following sub-steps:

[0100] In this embodiment, step S100 is used to convert multi-source technical text with timestamps into structured data that can be read for subsequent graph construction and matrix calculation. This step includes at least timestamp extraction, initial constraint template loading, structured field extraction, output result verification, and trend state mapping when numerical values ​​are missing.

[0101] S101, Multi-source text preprocessing and initial constraint template loading. In this embodiment, the text structure parsing module receives multi-source technical text streams and extracts timestamp sequences based on text metadata. .

[0102] In the timestamp extraction rules, the application date or publication date is extracted as the time node mark for patent publications, and the system entry date or version finalization date is extracted as the time node mark for internal R&D specification documents. For text records with multiple time fields, the text structure parsing module selects the time node according to a preset priority; within the same type of text, the priority remains consistent to ensure the stability of the time series sorting results.

[0103] As a preferred implementation, the large language model employs a pre-trained language model based on the Transformer architecture. To enable the large language model to perform domain-specific extraction tasks, the text structuring parsing module loads pre-defined initial constraint templates into its internal memory.

[0104] The initial constraint template includes a JSON data framework specification for defining the output data structure, task role definition prompts, and few-shot annotation examples for standardizing the output format. The JSON data framework specification at least defines the following fields: technical component entity field, engineering problem field, engineering parameter field, absolute numerical boundary field, three-state trend vector value field, coupling influence parameter field, and extraction confidence field. The extraction confidence field characterizes the reliability of the large language model's judgment on the direction of parameters in the current tuple.

[0105] The text structure parsing module will be at the current time point. The text data and the initial constraint template are concatenated and packaged together, and then transformed into a sequence of identifiers (Tokens) that are adapted to the input format of the large language model, and then input into the large language model.

[0106] S102, Structured Feature Extraction and Result Verification. The large language model performs semantic recognition on unstructured natural language text based on the extraction criteria set in the initial constraint template, and extracts specified technical features within the text context window.

[0107] In terms of benchmark extraction, large language models identify the names of components in text that have independent physical forms or perform specific control logic as technical component entities. Identify system defect states indicated in the textual background description as engineering problems. Identify the names of physical quantities that measure the operational status of the technical component as engineering parameters. Identify other parameters whose state changes with changes in engineering parameters as coupling influence parameters. .

[0108] When the text simultaneously discloses coupling effect parameters When the direction of change is determined, the text structure parsing module couples the influencing parameters. As engineering parameters Generate an associative tuple object and record the coupling effect parameters in the associative tuple object. The corresponding three-state trend vector value; when the current text does not disclose the coupling effect parameter. When the direction of change is determined, the system retrieves engineering parameters from the set of structured tuples at the same or adjacent time nodes. Coupling effect parameters Matching associated tuple objects.

[0109] The above fields are used to characterize the correspondence between technical components, engineering problems, parameter changes, and the influence relationships between parameters in the technical text.

[0110] If the engineering problem is not explicitly stated in the specific text Or coupling affects parameters The text structuring parsing module records the corresponding fields as missing values ​​to maintain the continuity of the data processing flow. Specifically, in engineering problems... When missing, the system still retains the tuple and uses it as a parameter evolution record for subsequent graph updates; when coupling affects parameters When missing, the system only establishes the corresponding engineering parameter node and its cross-layer mounting relationship with the technical component entity, and does not generate directed coupling edges in the parameter constraint layer.

[0111] The text structuring parsing module combines and packages the extracted elements from each iteration, and performs format validation on the output of the large language model according to the JSON data framework specification in the initial constraint template. This validation is applied when output field types do not match, except for engineering problem fields that are allowed to be missing. Coupling affects parameter fields If any required fields other than those specified are missing, or if the three-state trend vector value exceeds the preset value set, the text structure parsing module triggers a second re-query process. If a qualified output still cannot be obtained after the second re-query, the text record is written to the exception cache table and is not included in the matrix calculation at the current time node.

[0112] The text structure parsing module will aggregate the tuple objects that have undergone format validation to generate the current time node. The following is a set of structured tuples Any one of the extracted tuple objects is defined as:

[0113] ;

[0114] in, Represents the physical entity of the technical component. Representative engineering issues, Represents the name of the engineering parameter. This represents the absolute numerical boundary of the engineering parameter. Represents the three-state trend vector value. This represents the parameters affected by coupling. The confidence field is extracted and stored as an additional attribute of this tuple object.

[0115] S103, Trend state mapping when values ​​are missing. The text structure parsing module executes the logic for determining missing values ​​when processing the engineering parameters of tuple objects.

[0116] The large language model prioritizes matching specific numeric characters and units of measurement within adjacent text segments associated with engineering parameter names. When a match is successful, the numeric characters and unit of measurement are recorded as absolute numerical boundaries. When the same engineering parameter corresponds to multiple numerical expressions, the text structure parsing module prioritizes selecting the numerical expression that is closest to the engineering parameter name and matches the unit of measurement; if a range value exists, then that range value is used as the absolute numerical boundary. .

[0117] When a relevant text paragraph uses vague quantifiers and does not contain specific numeric characters, the text structure parsing module determines that absolute numerical boundaries are missing and assigns them a null value. .

[0118] Regardless of whether absolute numerical boundaries are missing, the text structure parsing module instructs the large language model to perform directional semantic classification on the sentences containing the engineering parameters and outputs three-state trend vector values. When absolute numerical boundaries exist, the system also saves the absolute numerical boundaries. and three-state trend vector values When absolute numerical boundaries are missing, the text structure parsing module triggers a trend state mapping instruction and maps the three-state trend vector values. As the absolute numerical boundary The alternative state parameters are used in subsequent calculations. The large language model outputs discrete three-state trend vector values. Use it as the absolute numerical boundary When a state parameter is missing, it is used as an alternative input for subsequent calculations.

[0119] Three-state trend vector value The value space is limited to the set {+1, 0, −1}. When a statement contains verbs and modifiers indicating an increase, rise, or performance enhancement of engineering parameters, the text structuring parsing module will convert the three-state trend vector values... The value is determined to be +1; when the statement contains verbs and modifiers indicating a decrease, reduction, or weakening of engineering parameters, the text structure parsing module will determine the three-state trend vector value. The value is determined to be -1; when the statement maintains its current state, there is no clear parameter change indicating a direction, or there is a conflict in the context that makes it impossible to establish a direction, the text structure parsing module will determine the three-state trend vector value. The value of is determined to be 0.

[0120] The text structuring parsing module synchronously records the extraction confidence level of the three-state trend vector values. As one possible implementation, the text structuring parsing module performs multiple independent extractions on the same text segment and uses the ratio of the frequency of the most frequent three-state trend vector value to the total number of extractions as the extraction confidence level. The extraction confidence level ranges from 0 to 1. For example, if three independent extractions are performed on the same text segment, with two outputs of +1 and one output of 0, the extraction confidence level corresponding to +1 is 2 / 3. When the extraction confidence level is lower than a preset confidence threshold, the text structuring parsing module either does not use the three-state trend vector values ​​in the tuple for negative weight calculations or writes the tuple to the manual review queue. The preset confidence threshold can be set based on the manual verification results of historical extracted samples or a system default threshold can be used.

[0121] Through the above processing, natural language expressions that lack definite numerical boundaries are converted into discrete state parameters that can participate in matrix calculations, thereby reducing the probability of calculation process interruption due to the lack of explicit disclosure of core parameter values ​​in technical texts.

[0122] Reference Figure 4 The method step S200 provided by the present invention specifically includes the following sub-steps:

[0123] In this embodiment, step S200 is used to map the structured tuple set into a two-layer knowledge graph. The two-layer knowledge graph includes a technical component structure layer for representing technical components and their connections, and a parameter constraint layer for representing engineering parameters and the influence relationships between parameters. The technical component structure layer and the parameter constraint layer establish a correspondence through cross-layer pointers.

[0124] S201, Construction of the technical component structure layer. The two-layer knowledge graph construction module receives the structured tuple set data and instantiates and establishes the technical component structure layer in the graph database. In this embodiment, the graph database adopts a non-relational database that supports the attribute graph model.

[0125] The two-layer knowledge graph construction module traverses the structured tuple set at different time points, extracts the technical component entities, and constructs the extracted technical component entities as nodes of the technical component structure layer. Each technical component entity node includes at least the entity name, standardized name, entity source text identifier, timestamp, and globally unique node identifier.

[0126] Due to differences in expression habits among different text writers, the two-layer knowledge graph construction module performs entity alignment and deduplication operations on extracted technical component entities with the same name or synonymous aliases. As a preferred approach, the system integrates an industry-standard terminology library, calculates the string edit distance between the extracted entity and the standard terms in the library, and determines that the entity is the same physical component and merges the nodes when the distance is less than a preset threshold. When merging nodes, the system retains the original entity descriptions from different texts and writes the original entity descriptions into the alias attribute of the node for subsequent text source tracing.

[0127] The technical component structure layer is defined as a directed graph. Directed graph middle, Represents all technical component entities within the technical component structure layer. The set of nodes; This represents the set of directed edges between nodes.

[0128] The two-layer knowledge graph construction module reads the source text corresponding to the structured tuples and extracts the component relationships from the source text to generate component relationship records. Each component relationship record includes at least the source technology component entity, the target technology component entity, the component relationship type, the relationship direction, the source text identifier, the timestamp, and the relationship extraction confidence level. Based on the component relationship records, the two-layer knowledge graph construction module maps the physical assembly, electrical connection, and data communication link relationships between technology component entities into a set of directed edges. Directed edges in the relation. For component relation records with a relation extraction confidence level lower than the preset relation confidence level threshold, the system does not generate corresponding directed edges, or writes the component relation record into the manual review queue.

[0129] For each directed edge, the system records the relationship type, source text identifier, and timestamp in the edge attributes; when the text only records that two technical components appear side by side but does not explicitly record the connection relationship, the system does not generate the corresponding connection edge.

[0130] S202, Parameter Constraint Layer Construction and Cross-Layer Association. The two-layer knowledge graph construction module synchronously instantiates and builds the parameter constraint layer in the graph database and constructs cross-layer association mapping relationships. The parameter constraint layer is defined as a weighted directed graph. Weighted directed graph middle, All engineering parameters within the representative parameter constraint layer The set of nodes; Represents the parameters affected by coupling The set of directed edges that are determined; This represents the edge weight matrix.

[0131] The parameter constraint layer is used to store engineering parameter nodes, directed influence relationships between parameters, and corresponding edge weights.

[0132] To establish a clear data traceability chain, the two-layer knowledge graph construction module configures a cross-layer mounting pointer for each engineering parameter node in the parameter constraint layer. This cross-layer mounting pointer is implemented by writing the globally unique identifier (UUID) of the target entity node into the parameter node attributes of the graph database. This subordinate mounting relationship ensures that any engineering parameter at the lower level can correspond to a specific technical component entity, avoiding system identification ambiguity caused by different physical components having parameters with the same name.

[0133] In the parameter constraint layer, the globally unique identifier of the engineering parameter node is generated by the UUID of the target entity node, the standard name of the engineering parameter, and the parameter unit, ensuring that parameters with the same name under different entities are stored as different parameter nodes. The engineering parameter node also includes the original name of the parameter, the standardized name, the parameter unit, the parameter source text identifier, the timestamp, and the cross-layer mounting pointer corresponding to the technical component entity node.

[0134] Within the parameter constraint layer, when a change in an engineering parameter recorded in a structured tuple triggers a change in a coupled-influence parameter, the two-layer knowledge graph construction module generates a directed edge between the two corresponding parameter nodes. This directed edge takes the actively changing engineering parameter as the starting node and the passively changing coupled-influence parameter as the target node, representing the direction of influence between the parameters.

[0135] When the coupling influence parameter field in the structured tuple is missing, the two-layer knowledge graph construction module only saves the engineering parameter node and its cross-layer mounting pointer, does not generate directed edges in the parameter constraint layer, and waits for the text record containing the coupling influence parameter in the subsequent time node to complete the edge relationship.

[0136] S203, Parameter Edge Weight Allocation. The two-layer knowledge graph construction module performs coupled weight allocation calculations based on the combination relationship of three-state trend vector values.

[0137] To quantify the collaborative or conflicting relationships between engineering parameters, the two-layer knowledge graph construction module uses the three-state trend vector values ​​corresponding to the originating node and the target node of an edge. Numerical combination, edge weight matrix The elements in the dictionary are assigned values ​​for calculation. In practical applications, the direction of parameter change alone is insufficient to determine the impact of the change on system performance. Therefore, the two-layer knowledge graph construction module maintains an engineering parameter attribute dictionary. This dictionary pre-labels the expected evolution direction of each engineering parameter; for example, the expected evolution direction for "heat generation" is decrease, and for "operational frequency" it is increase. This dictionary supports dynamic switching or custom configuration based on specific analysis scenarios, industry terminology libraries, or domain ontology libraries to adapt to differences in the expected evolution direction of engineering parameters across different technical fields. Each parameter record in the dictionary includes at least the parameter standard name, parameter synonyms, unit of measurement, technical field, expected evolution direction label, applicable operating conditions, and data source.

[0138] When there are coupling effects on parameters in the structured tuple object At that time, the two-layer knowledge graph construction module will use the three-state trend vector values ​​in the current tuple object. As the tri-state trend vector value corresponding to the initial parameter node, the engineering parameter field and coupling influence parameter are retrieved from the structured tuple set at the current time node. For a matching associated tuple object, the tri-state trend vector value in the associated tuple object is used as the tri-state trend vector value corresponding to the target parameter node. If no associated tuple object is found, the system does not assign a valid edge weight to the corresponding directed edge, writes the directed edge to the pending completion queue, and directed edges in the pending completion state do not participate in edge weight state change monitoring.

[0139] The system compares the actual three-state trend vector value of the parameter with the expected evolution direction label to determine whether the current parameter state is a performance optimization state or a performance degradation state. When an engineering parameter is not included in the engineering parameter attribute dictionary, the system marks the parameter as an undefined expected direction parameter; for undefined expected direction parameters, the system does not generate negative weights based on the parameter, or adds the parameter to the list of parameters to be configured.

[0140] After determining the node state, the two-layer knowledge graph construction module allocates edge weights according to the execution criteria. .

[0141] When the initial parameter node In performance optimization state and target parameter node When both are in a performance optimization state, it is determined that they exhibit cooperative gain characteristics, and the edge weight of the connected edge is set to a constant + 1.

[0142] When the initial parameter node In performance optimization state and target parameter node When the system enters a performance degradation state, it is determined that there is a parameter coupling conflict between the two, and the edge weight of the connected edge is set to a constant -1.

[0143] When the initial parameter node The performance optimization state leads to the target parameter node When the system enters a performance degradation state, it determines that there is a parameter coupling conflict between the two components and marks this relationship as a candidate parameter coupling conflict. At this point, the two-layer knowledge graph construction module assigns negative weight values ​​to its connected edges, i.e. It is set to a constant -1, and this assignment is recorded as a negative weight in subsequent calculations. .

[0144] When the three-state trend vector value of the originating parameter node or the target parameter node is 0, or when there is no clear cooperative or conflicting direction in the text, the edge weight of the connected edge is set to zero.

[0145] Through the above weight allocation rules, the parameter influence relationships in the structured tuples are converted into symbolic weight data that can participate in matrix calculations.

[0146] Reference Figure 5 The method step S300 provided by the present invention specifically includes the following sub-steps:

[0147] In this embodiment, step S300 is used to convert the time-varying negative edge weights in the parameter constraint layer into a computable matrix sequence, and calculate the coupling interference degree value of the target parameter pair under a unified matrix dimension. The coupling interference degree value is used to represent the strength of the negative coupling relationship that the target parameter pair continuously exhibits in the historical text sequence.

[0148] S301, Temporal conflict adjacency matrix generation. In this embodiment, the temporal conflict matrix calculation module follows the timestamp sequence of the text stream to transform the network topology of the parameter constraint layer into a temporal conflict adjacency matrix carrying a time dimension.

[0149] To eliminate the inconsistency in matrix dimensions caused by the addition or removal of nodes in the dynamic graph, the temporal conflict matrix calculation module establishes a global engineering parameter node pool. Considering that streaming text will continuously introduce new engineering parameters, the temporal conflict matrix calculation module dynamically expands the capacity of the node pool when a new node is detected, synchronously increases the number of rows and columns of the matrix, and performs zero-padding operation on the newly added dimension positions in the matrix of historical time nodes, thereby maintaining a globally consistent dimension for all matrix sequences.

[0150] The global engineering parameter node pool assigns a fixed matrix index to each engineering parameter node. This fixed matrix index is used to determine the row and column positions of the engineering parameter node in the corresponding matrix at each time point. After a new engineering parameter node is written to the global engineering parameter node pool, the system appends a new matrix index to it and pads the existing historical matrix with zeros at the row and column positions corresponding to the new index.

[0151] At each independent point in time The temporal conflict matrix calculation module performs graph section extraction on the parameter constraint layer. The graph database configures version identifier fields, valid start time fields, and valid end time fields for nodes and edges. The temporal conflict matrix calculation module calculates the time section based on the current time node. The above fields are filtered to recover the parameter constraint layer map section at that time point, and a two-dimensional discrete adjacency matrix of the corresponding section is generated. .

[0152] The original edge weights in the parameter constraint layer represent the actual relationship between engineering parameters, and their values ​​include -1, 0, and +1. Among them, -1 represents a negative coupling relationship, 0 represents no explicit coupling relationship, and +1 represents a positive cooperative relationship.

[0153] The temporal conflict matrix calculation module extracts the negative weights present in the edge weight matrix and maps them to the discrete adjacency matrix. At specific coordinate locations, all other coordinates in the matrix that are not negatively connected are simultaneously filled with zeros. When a target parameter pair has no edge relationship at a certain time node, or only a positive weight relationship exists, the matrix element corresponding to that target parameter pair at that time node is set to zero. Therefore, the temporal conflict adjacency matrix is ​​only used to record and calculate the negative coupling state, and is not used to store the positive edge weights in the parameter constraint layer; the edge weight state change monitoring of the target parameter pair reads the original edge weights from the graph section of the corresponding time node of the parameter constraint layer.

[0154] The above processing converts the graph structure in the parameter constraint layer into a two-dimensional adjacency matrix sequence, providing a unified data format for subsequent parallel matrix computation.

[0155] S302, Long-period time-weighted accumulation calculation of coupling interference. The time-series conflict matrix calculation module performs time-series weighted accumulation calculation on the continuously existing negative weight data in the time-series conflict adjacency matrix.

[0156] In engineering parameter analysis, if a pair of engineering parameters shows a negative coupling relationship multiple times over a long period of time, it usually indicates that there may be a long-term high-frequency coupling conflict between the parameter pair, which can be used as a candidate for subsequent technical gap mining and targeted analysis.

[0157] Therefore, the time-series conflict matrix calculation module targets specific objective parameters. During the observation period consisting of the start time and the current time point The coupling interference degree is calculated internally. The specific time-weighted accumulation formula for this value is set as follows:

[0158] ;

[0159] in, This represents the coupling interference degree of the target parameter during the observation period; Represents the time node Discrete adjacency matrix Medium parameter node and The actual negative weight values ​​between them are set to a constant -1 according to the calculation criteria of the previous steps, and after absolute value conversion, they are dimensionless mathematical scalars. Represents a time span scalar used to quantify the duration of this negative weight state; This represents the time decay weight, used to reduce the impact of early negative coupling relationships on the current observation period.

[0160] As a preferred implementation, time span scalar The value is the current time node. Compared with the previous time point The absolute natural day difference between the two values ​​is normalized by dividing the difference by a preset base duration.

[0161] To avoid starting node The lack of a calculation anchor point caused a program error. The system set the time when the target parameter was first extracted and entered into the map as the initial anchor point. The baseline duration is fixed at 365 days, thus unifying the dimension of the time span scalar to the calendar year.

[0162] To reflect the varying impact of technical texts on current analytical results over time, time decay weights are used. Calculated using the exponential decay method, set as The time variable This represents the latest time point at the end of the observation period. Indicates by time node Until the end of the observation period The annualized time difference obtained by normalizing the natural day difference between the two is to ensure that the weight of the early negative coupling relationship decreases exponentially over time.

[0163] Attenuation coefficient in the model The attenuation coefficient λ is pre-configured based on the empirical value of the average technology iteration cycle in the target technology field, with a value range between 0.01 and 0.05. When the system is first applied to a new technology field and historical iteration cycle data is lacking, the attenuation coefficient λ adopts the system default value; after accumulating historical samples in this technology field, the system reconfigures the attenuation coefficient based on the historical samples. .

[0164] This long-cycle time-series weighted cumulative calculation converts the negative coupling semantics extracted from historical text into coupling interference values, which are used to quantify the persistence of candidate coupling conflict relationships in the historical text sequence and provide numerical triggering basis for subsequent feedback control processes.

[0165] Reference Figure 6 The method step S400 provided by the present invention specifically includes the following sub-steps:

[0166] In this embodiment, step S400 is used to adjust the subsequent text parsing range based on the calculation result of the coupling interference degree value. This step includes comparing the coupling interference degree threshold, determining the target entity node set, and generating a directional constraint template. The directional constraint template is used to prioritize the processing of text fragments related to the target entity node set in subsequent text parsing.

[0167] S401, Coupling Interference Threshold Monitoring and Cross-Layer Entity Query. In this embodiment, the extraction rule feedback control module monitors the coupling interference value output by the time-series conflict matrix calculation module on the time axis.

[0168] At the end of a specific observation period The extraction rule feedback control module extracts target parameters. Calculation results of coupling interference And compare it with the system's preset interference trigger threshold. Perform numerical comparison.

[0169] Regarding the interference trigger threshold As a preferred implementation, the system extracts a set of coupling interference degree samples of all parameter pairs from the historical atlas library for statistical analysis, and uses the 80th percentile of the sample distribution as the threshold benchmark. When the number of samples in the historical atlas library is lower than a preset sample number threshold, the system uses a preset default threshold as the interference degree trigger threshold; after the number of historical samples reaches the preset sample number threshold, the interference degree trigger threshold is re-determined based on the sample distribution.

[0170] When the calculation meets the condition At that time, the system will set parameters and The relationships between them are marked as candidate coupling conflict objects that appear frequently in the long term, and these are used as the key monitoring objects for subsequent targeted analysis and local topology verification.

[0171] To map candidate coupling conflict objects in the parameter layer to specific technical components, the rule extraction feedback control module reads the parameter nodes. and Cross-layer mounting pointers are generated during the graph construction phase. The extraction rule feedback control module utilizes the primary key index retrieval function of the graph database, and based on the globally unique identifier recorded in the mounting pointer, directs the data to the technical component structure layer. Initiate a query.

[0172] When parameter and When they point to different entity nodes, for example, node and The rule extraction feedback control module utilizes the shortest path algorithm in the directed graph of the technical component structure layer. Search Connectivity and The system defines the component connection path. The system uniformly marks the set of components involved in this connection path and the endpoint entities as the target entity node set. If node and If there is no connecting path between them, the system will... With nodes Add to the target entity node set respectively They are then matched as independent target entities in subsequent parsing.

[0173] When parameter node With parameter nodes When a cross-layer mounting pointer points to the same technical component entity node, the system treats the same technical component entity node as a component connection path with zero length and directly adds the same technical component entity node to the target entity node set. .

[0174] The above query process is used to map the target parameter pairs in the parameter constraint layer to the corresponding technical component entities, thereby determining the entity range for subsequent targeted text parsing.

[0175] S402, directional constraint template generation and parsing range control. The extraction rule feedback control module utilizes the target entity node set. The information is used to control the parsing range of the large language model input in the text structure parsing module.

[0176] Subsequent timestamps Before introducing a new batch of technical text streams, the extraction rule feedback control module retains the system's default initial constraint template and uses an assembly script to generate a directional constraint template. When the input text hits the target entity node set... When the input text matches the corresponding entity name or entity alias, the system uses a targeted constraint template for parsing; when the input text does not match the target entity node set... When the corresponding entity name or entity alias is used, the system still uses the initial constraint template to perform regular parsing.

[0177] The directional constraint template includes at least the target entity node set, target parameter pairs, local text window length, engineering parameter fields to be prioritized, three-state trend vector value fields, and output format constraint fields.

[0178] The directional constraint template shrinks the effective context window of a large language model through input text pre-truncation or attention masking. Specifically, the extraction rule feedback control module writes entity position constraint instructions within the cue word framework of the directional constraint template. When the extraction rule feedback control module matches the target entity node set in the input unstructured text sequence... When the corresponding entity name or entity alias is used, the entity position constraint instruction control system truncates the local text sequence centered on the entity name or entity alias, generating a fixed-length local truncation window, such as 512 identifiers before and after.

[0179] When the large language model is deployed locally and the inference framework has the attention mask parameter enabled, for the remaining text sequence identifiers outside the local truncation window, the system modifies the attention mask matrix elements corresponding to the input of the large language model's self-attention mechanism to negative infinity.

[0180] When the large language model operates on an interface call model and the underlying attention mask parameters cannot be modified, the system first performs local truncation on the input text based on the position of the target entity name or entity alias in the input text, and only inputs the truncated local text sequence into the large language model to achieve an equivalent context window shrinkage.

[0181] Based on the scaling dot product attention calculation principle of the Transformer architecture The mask value of negative infinity will be mapped to zero after activation by the Softmax function. Therefore, text identifiers outside the local truncation window do not participate in the attention calculation of the current local window.

[0182] Suppose the original input text sequence contains N text identifiers, and the local text window contains L text identifiers, where L is less than N. In the interface call mode, the number of text identifiers input to the large language model is reduced from N to L, thus reducing the theoretical complexity of standard full attention computation from O(N^2) to L. 2 ) decreased to O(L 2 This reduces input cache usage and GPU memory usage during inference. In local deployment mode, the attention mask resets the attention weights of text identifiers outside the local text window to zero, reducing interference from irrelevant contexts on the target parameter extraction results. When the inference framework supports sparse attention computation or skips the computation of mask positions, it can further reduce the actual attention computation and GPU memory usage.

[0183] Based on shrinking the effective context window, the extraction rule feedback control module further sets the parameter extraction priority in the directional constraint template. In the normal parsing mode, the large language model prioritizes finding the absolute numerical boundary. After the directional constraint template is enabled, the system increases the three-state trend vector value for the statement containing the target parameter pair. The priority of judgment.

[0184] In this mode, even if absolute numerical boundaries are missing in the current literature context, the orientation constraint template requires the large language model to prioritize the identification of technical expressions related to directional changes in the target parameter pair. These technical expressions include directional terms indicating increase, decrease, maintenance, improvement, weakening, suppression, or enhancement, as well as descriptive fields such as material properties, process conditions, control strategies, or component models that co-occur with the target parameter pair.

[0185] Through the above processing, the system forms a feedback update process from parameter conflict objects to the target entity node set and then to the directional constraint template based on the threshold comparison results of the coupling interference degree values.

[0186] Reference Figure 7 The method step S500 provided by the present invention specifically includes the following sub-steps:

[0187] In this embodiment, step S500 is used to perform local verification of the technical component structure corresponding to the change when the edge weight of the target parameter pair changes from a negative value to a zero value or a positive value. This step includes edge weight state monitoring, target subgraph writing and locking, K-Hop local neighborhood subgraph extraction, and approximate graph editing distance calculation.

[0188] S501, Edge weight state change monitoring and target subgraph writing locking. In this embodiment, the local topology verification and output module monitors the original edge weight state between target parameter pairs in the parameter constraint layer.

[0189] When the system time advances to the time node At that time, the local topology verification and output module performs parameter constraint layer checks at the specified time node. Extract target parameter pairs from corresponding spectral sections The original edge weights between them. At time node... The corresponding parameter constraint layer map section does not contain target parameter pairs. When there is a directed edge between them, the system will set the target parameter to the time node. The edge weight state is marked as unobserved. An unobserved state is not equivalent to zero and does not participate in the determination of edge weight changes from negative to zero or positive. Local topology verification is triggered only when a valid directed edge is formed between the target parameter pairs at the current time node, and the edge weight of the valid directed edge is determined to be zero or positive. Target parameter pairs These are parameter pairs that have been previously identified as candidate coupling conflict objects that have appeared frequently over a long period of time.

[0190] If the system detects that the original edge weights between target parameter pairs in the parameter constraint layer change from historical negative values ​​to zero or positive values, i.e., the negative coupling relationship of the target parameter pair disappears or transforms into a cooperative relationship, then the system determines that the candidate coupling conflict object has a suspected mitigation or resolution event. The specific determination criteria are as follows:

[0191] and ;

[0192] in, This represents the edge weights of the target parameters at the old time nodes. This represents the edge weights of the target parameters at the latest time node.

[0193] Old time nodes Given that the original edge weights between the target parameter pairs in the parameter constraint layer are negative, the coupling interference degree of the target parameter pair most recently exceeded the interference degree trigger threshold. The generated alarm records are determined; if multiple alarm records meet the above conditions, the system selects the one with the specified distance. The most recent alarm time is used as the old time node. New Time Node For the old time node Then, the time point at which the original edge weight first changes from a negative value to a zero value or a positive value is detected.

[0194] To ensure the consistency of the graph data during subsequent verification, the local topology verification and output module sends a write lock command to the system main control unit for the target parameter pair and its associated graph region, temporarily locking the graph write operation for the target parameter pair and its associated graph region. Data to be written related to the target parameter pair at the current timestamp is written to the cache queue; regular text parsing tasks unrelated to the target parameter pair continue to execute, or enter a queue state according to the system scheduling rules.

[0195] Local topology verification tasks are started independently and asynchronously. After verification is complete, the system releases the write lock on the associated graph region and merges the data in the cache queue into the graph database.

[0196] S502, K-Hop local neighborhood subgraph extraction. Accurate graph edit distance calculation is typically an NP-hard problem. To control the computational scale of comparing the old and new graph structures, the local topology verification and output module uses the target entity node set... As seed nodes, in the directed graph of the technical component structure layer The restricted breadth-first search algorithm is executed.

[0197] The system uses hop count As a boundary for search depth, the search for adjacent nodes is expanded layer by layer according to the number of hops. As a preferred method, the number of hops... The value is configured within the integer range of 2 to 3. This range is used to cover the main assembly relationships, communication links, or electrical connections around the target component, and to reduce the number of irrelevant peripheral nodes entering the subsequent comparison process.

[0198] When the target entity node set When there are multiple entity nodes, the local topology verification and output module extracts the K-Hop neighborhood from each entity node, and merges and deduplicates the nodes and edges in the multiple neighborhoods to obtain the K-Hop local neighborhood subgraph at the corresponding time node.

[0199] Based on the above search boundaries, the system performs searches at the old time nodes. and new time nodes Local topology is extracted from the cross-section of the corresponding technical component structural layer map to generate the old K-Hop local neighborhood subgraph. With the new K-Hop local neighborhood subgraph .

[0200] Old time nodes and new time nodes The corresponding graph section is recovered based on the version identifier field, valid start time field, and valid end time field of the nodes and edges in the graph database.

[0201] The aforementioned local neighborhood extraction method, while preserving the connectivity relationships around the target entity, limits the number of nodes and edges that need to be processed for subsequent graph structure comparison.

[0202] S503, Approximate graph edit distance calculation. To quantify the degree of structural change of corresponding technical components of candidate coupling conflict objects during the suspected mitigation or resolution process, the local topology verification and output module calls a heuristic algorithm based on bipartite graph matching to compare the differences between the old and new K-Hop local neighborhood subgraphs.

[0203] The specific formula for calculating the approximate image editing distance is set as follows:

[0204] ;

[0205] in, The approximate graph edit distance scalar represents the distance between the old and new local subgraphs. This represents the space of all feasible bipartite graph mapping schemes between two sets of nodes in a local subgraph. The representation represents a specific set of node pairs in the scheme space, containing nodes from the old subgraph. With new subgraph nodes The matching and pairing relationships between them; The representative will use the old node Replace with new node algebraic cost; and They represent the current mapping set respectively The set of detached nodes in the old subgraph that were not successfully matched and the set of detached nodes in the new subgraph; This represents deleting a node from the old topology. The cost; This represents inserting a node into the new topology. The cost.

[0206] To enable the heuristic calculation process to be implemented directly, the system sets the following numerical allocation rules for each cost parameter.

[0207] The local topology verification and output module will remove node costs. Cost of inserting nodes The unit is uniformly set to a fixed constant of 1.

[0208] Regarding node replacement costs This cost incorporates differences in entity attributes and local connectivity. Specifically, the system first uses the Levenshtein Distance algorithm to calculate node... With nodes The edit distance of the component name string is carried; subsequently, the system extracts the node. With nodes Given the respective sets of locally connected edges, calculate the difference between the two sets in terms of edge count and in-degree. Node replacement cost. Ultimately, it is assigned the normalized sum of the string difference value and the difference of the connected edges.

[0209] String difference value The ratio of the edit distance between the node name strings to the maximum character length of the two node names; the difference in the number of edges. The ratio of the absolute value of the difference in the number of edges between two nodes to the maximum value of the number of edges between the two nodes; the difference in in-degree direction. This is the ratio of the sum of the absolute values ​​of the in-degree and out-degree differences between the two nodes to the maximum total degree of the two nodes. When the corresponding denominator is zero, the denominator is set to 1. The node replacement cost is calculated using the following formula:

[0210] ;

[0211] in, Each weight is pre-configured by the system.

[0212] This approach, which incorporates edge structure changes into node replacement costs, allows the bipartite graph matching process to represent node name differences and local connectivity differences with a single cost value.

[0213] If node With nodes For components belonging to different engineering component categories, the replacement cost is assigned a preset penalty constant, such as 9999, to avoid generating unreasonable node matching results. For the calculation to find the minimum total cost, those skilled in the art can use the Kuhn-Munkres (KM) algorithm or the standard Hungarian algorithm.

[0214] Through the above calculations, the system obtains an approximate graph edit distance value to characterize the degree of difference between the old and new K-Hop local neighborhood subgraphs, and uses it as input for subsequent evolution path classification.

[0215] Reference Figure 8 The method step S600 provided by the present invention specifically includes the following sub-steps:

[0216] In this embodiment, step S600 is used to classify the suspected mitigation or resolution paths of candidate coupling conflict objects based on the approximate graph edit distance, and generate a technology evolution analysis report by combining the classification results, related components, parameter boundaries, or trend status information. This step includes determining the topology change threshold, classifying evolution path types, and generating and outputting the report.

[0217] S601, Topology change threshold comparison and evolution path type classification. In this embodiment, after calculating the approximate graph edit distance of the old and new K-Hop local neighborhood subgraphs, the local topology verification and output module obtains the approximate graph edit distance. And compare it with the topology change threshold set by the system. Perform numerical comparison.

[0218] As a preferred approach, the topology change threshold It is not a fixed absolute constant, but rather dynamically calculated by the system based on the actual size of the truncated subgraph. The formula for calculating the dynamic threshold is set as follows:

[0219] ;

[0220] in, This represents the preset topology tolerance coefficient, which is based on empirical statistics of routine engineering changes, and its value range is set between 0.15 and 0.25. and These represent the cardinality of the number of nodes contained in the old K-Hop local neighborhood subgraph and the new K-Hop local neighborhood subgraph, respectively.

[0221] When the system is first applied to a new technical field and there are no regular engineering change samples, β adopts the system default value; after accumulating comparison samples of new and old subgraphs in this technical field, the system reconfigures the value of β based on the historical comparison results.

[0222] This dynamic threshold evaluation method is used to balance the impact of subgraph size on approximate graph edit distance and reduce classification bias caused by differences in the number of nodes in local subgraphs.

[0223] Based on the above comparison results, the local topology verification and output module categorizes the suspected mitigation or resolution paths of candidate coupling conflict objects into two preset evolution path types:

[0224] The first type is structural reconstruction-type evolution. When the system detects that the calculation result satisfies… At this time, the current technological evolution path is classified as a structural reconfiguration-type evolution. This classification result indicates that the suspected mitigation or resolution process of candidate coupling conflict objects is highly correlated with the addition, deletion, or replacement of external technical components or the redirection of network topology links. For this classification, the local topology verification and output module extracts unmatched free nodes recorded during the approximate graph editing distance calculation process and marks them as structural change nodes; structural change nodes include deleted old component nodes, inserted new component nodes, or replacement candidate nodes that failed to complete the matching.

[0225] The second category is candidate paths for improving the original component properties. When the system detects that the calculation result satisfies... At that time, the current technological evolution path is classified as a candidate path for improving the original component attributes. Under this classification, the structural difference between the old and new K-Hop local neighborhood subgraphs is lower than the topology change threshold, but the original edge weights between the target parameter pairs in the parameter constraint layer have changed from negative to zero or positive.

[0226] This indicates that changes in the target parameter pairs may depend less on changes in macroscopic connectivity within the technical component structure layer. The local topology verification and output module further extracts attribute fields from the latest text, such as material properties, manufacturing processes, control algorithms, component models, operating conditions, interface protocols, or environmental constraints, to assist in identifying attribute factors related to changes in the edge weights of the target parameter pairs.

[0227] To identify the target of interest in the report, the system uses cross-layer mounting pointers to backtrack and trigger the target parameters of the current verification to mark the source entity node directly attached to it as the target source component. The material properties, manufacturing process, control algorithm, component model or operating condition fields obtained in the secondary extraction are used as auxiliary evidence for the candidate path.

[0228] S602, Technology Evolution Analysis Report Generation and Output. To generate a research and development reference report, the local topology verification and output module initiates a data processing procedure, performing cross-layer correlation between the aforementioned evolution path types and the indicator data extracted from the large language model in the early stages. The local topology verification and output module reads the target parameter pairs from the graph database at the latest time node. The corresponding structured tuple object.

[0229] The system determines the absolute numerical boundaries of the engineering parameters in the tuple. The data status.

[0230] If there is a clear absolute numerical boundary in the structured tuple, the system extracts the numerical index and establishes a mapping relationship between it and the evolution path type and the corresponding technical component entity to generate design reference information containing specific numerical boundaries or numerical ranges.

[0231] If the target text does not explicitly disclose specific numerical values, resulting in a lack of absolute numerical boundaries, i.e. The local topology verification and output module enables the missing value processing procedure. At this time, the system calls the tri-state trend vector value obtained by the large language model based on the directional constraint template in the previous steps. .

[0232] The system has a pre-set text generation template that combines characters based on the specific discrete values ​​of the three-state trend vector and the parameter name field in the tuple. Specifically, when the three-state trend vector value is +1, the system extracts the prefix "increase / raise" and concatenates it with the target parameter name; when the three-state trend vector value is -1, the system extracts the prefix "decrease / reduce" and concatenates it with the target parameter name; when the three-state trend vector value is 0, the system generates the prefix "maintain / pay attention" and concatenates it with the target parameter name.

[0233] Through the above processing, even when the technical text does not explicitly disclose the absolute numerical boundaries, the system can still output qualitative information to indicate the direction of parameter adjustment.

[0234] Finally, the local topology verification and output module formats and organizes the evolution path type, structural change nodes, absolute numerical boundaries, or three-state trend vector values. When it is necessary to generate report content in natural language, it generates corresponding parameter trend status information based on the three-state trend vector values. Specifically, for results classified as candidate paths for improving the original component attributes, the local topology verification and output module simultaneously writes the material properties, manufacturing process, control algorithm, component model, operating conditions, or interface protocol fields obtained from the secondary extraction to form R&D focus items corresponding to the candidate path.

[0235] The technology evolution analysis report should include at least the target parameter pairs, corresponding technical component entities, evolution path type, approximate graph edit distance, topology change threshold, and absolute numerical boundary or three-state trend vector value; for structural reconstruction type evolution, the technology evolution analysis report should also include structural change nodes; for candidate paths for improvement of original component attributes, the technology evolution analysis report should also include attribute fields obtained by secondary extraction and R&D focus items.

[0236] The system encapsulates the formatted dataset into a standardized technology evolution analysis report and distributes it to the graphical interface of the enterprise's R&D terminal equipment via a network communication interface for display.

[0237] To further illustrate the data processing procedure of the present invention in enterprise R&D data analysis scenarios, the following explanation uses the field of new energy vehicle drive motor technology as an example. This example is used to explain the data flow relationship between various modules and is not intended to limit the application area of ​​the present invention.

[0238] In this embodiment, the system is applied to the research and development data analysis scenario in the field of new energy vehicle drive motor technology.

[0239] In this embodiment, the target analysis object is the target parameter pair consisting of "drive motor output power" and "stator system temperature rise". Among them, "drive motor output power" is the engineering parameter that needs to be improved, and "stator system temperature rise" is the coupled influence parameter affected by it.

[0240] Reference Figure 9 When the system accesses historical patent documents and internal R&D document streams in this field, the text structure parsing module processes the text records one by one according to the timestamp sequence. During the observation period from time node t1 to t5, the system continuously received multiple text records describing drive motor power enhancement schemes.

[0241] When processing the above text records, the text structure parsing module extracts the three-state trend vector value of the engineering parameter "output power" as +1. Since the expected evolution direction of "output power" is increasing, this parameter state is determined to be a performance optimization state. Simultaneously, the text structure parsing module extracts the three-state trend vector value of the coupling effect parameter "stator system temperature rise" as +1. Since the expected evolution direction of "temperature rise" recorded in the engineering parameter attribute dictionary is decreasing, the increase in "stator system temperature rise" is determined to be a performance degradation state.

[0242] Based on the above extraction results, the two-layer knowledge graph construction module establishes directed edges for the parameter nodes corresponding to "output power" and "stator system temperature rise" in the parameter constraint layer of the graph database, and assigns negative weight values. .

[0243] The temporal conflict matrix calculation module follows the timestamp sequence, converting the negative edge weights in the parameter constraint layer into a temporal conflict adjacency matrix, and then performs a temporal weighted accumulation operation. Combined with... Figure 9 As shown in part (a), during the t1 to t6 stages, the original edge weights between the target parameter pairs in the parameter constraint layer remain negative, indicating that the parameter pairs have a continuous negative coupling relationship across multiple time points.

[0244] Because this negative coupling relationship appears multiple times in the time series, the coupling interference value calculated by the time series conflict matrix calculation module continues to rise. Combined with... Figure 9 As shown in section (b), the coupling interference degree of this target parameter pair is as follows when time progresses to node t5. Exceeding the system's preset interference trigger threshold .

[0245] exist Once the conditions are met, the extraction rule feedback control module initiates the feedback update process. The extraction rule feedback control module queries the technical component structure layer through the cross-layer mounting pointer in the parameter node to determine the technical component entities corresponding to "output power" and "stator system temperature rise". The query determines that the target entity node set includes technical component entities such as "stator core" and "winding assembly".

[0246] At time node t6, when the system subsequently enters the system, the rule feedback control module generates a directional constraint template. When the input text matches the entity name or alias of the target entity node such as "stator core" or "winding assembly," the system performs partial truncation on the input text based on the directional constraint template and increases the priority of the three-state trend vector value judgment related to the direction change of the target parameters. For text that does not match the entity name or alias, the system still uses the initial constraint template to perform conventional parsing.

[0247] As time progresses to time node t7, the system receives technical texts containing new cooling structures or cooling control schemes. The text structure parsing module extracts statements from this batch of texts indicating an increase in "output power" while "stator system temperature rise" decreases or remains stable. According to the parameter attribute dictionary, the increase in "output power" is determined to be a performance optimization state, and the decrease or stabilization of "stator system temperature rise" no longer constitutes a performance degradation state.

[0248] The two-layer knowledge graph construction module updates the edge weights of the target parameter pairs accordingly. Figure 9 As shown in section (a), the original edge weights between target parameter pairs in the parameter constraint layer change from a historical negative value of -1 to a zero or positive value at node t7. After detecting this edge weight state change, the local topology verification and output module marks this event as a suspected mitigation or resolution event for candidate coupling conflict objects.

[0249] To ensure the consistency of the graph data during the verification process, the local topology verification and output module temporarily locks the write operations of the target parameter pairs and their associated subgraphs, and writes the data to be written related to the target parameter pairs to a cache queue. Regular text parsing tasks unrelated to the target parameter pairs continue to execute, or enter a queued state according to the system scheduling rules.

[0250] The local topology verification and output module traces back to the target parameter pair to determine the most recent instance where the coupling interference value exceeded the interference trigger threshold. The generated alarm logs will identify t5 as the old time node. And set the current time node t7 as the new time node. Subsequently, the system uses the target entity node set As a seed node, K-Hop local neighborhood subgraph extraction is performed in the technical component structure layer. As a preferred approach, this embodiment searches for the depth hop count. Set to 2.

[0251] At the old time point In the corresponding K-Hop local neighborhood subgraph, the system extracts nodes and connecting edges related to the "stator core," "round wire winding," and external cooling structure. At the new time node... In the corresponding K-Hop local neighborhood subgraph, the system extracts nodes such as "flat wire winding," "oil-cooled spray pipe," and "flow guide ring," along with their corresponding connecting edges. The differences in nodes and edges between the old and new local neighborhood subgraphs are used for subsequent approximate graph editing distance calculations.

[0252] The local topology verification and output module calls a heuristic algorithm based on bipartite graph matching to calculate the approximate graph edit distance between the old and new K-Hop local neighborhood subgraphs. During the cost calculation process, the system calculates the node replacement cost based on the string differences in the node name, the component category to which the node belongs, and the differences in the in-degree and out-degree of locally connected edges, and includes unmatched nodes in the insertion cost or deletion cost.

[0253] After comparing with the topology change threshold If the system determines... If so, the current technological evolution path is classified as a structural reconfiguration type evolution. In this embodiment, the system marks the unmatched new node "oil-cooled spray pipe" as a structural change node and writes it into the structural change node field of the technological evolution analysis report.

[0254] During the report generation phase, the local topology verification and output module reads the target parameter pair at the latest time node. The corresponding structured tuple object. If the structured tuple object contains explicit flow rate, oil pressure, or temperature rise boundary values, the system will write the corresponding values ​​as design reference information into the technology evolution analysis report. If the relevant text does not explicitly disclose specific numerical boundaries, i.e. The system calls the three-state trend vector values ​​recorded in the preceding steps. It generates qualitative parameter adjustment information based on the text template, such as "increase the flow rate of cooling medium at the stator end" or "reduce the temperature rise of the stator system".

[0255] Finally, the local topology verification and output module formats and organizes fields such as target parameter pairs, corresponding technical component entities, evolution path types, approximate graph editing distance, topology change threshold, structural change nodes, absolute numerical boundaries or three-state trend vector values ​​to generate a technology evolution analysis report, which is then distributed to the graphical interface of the enterprise's R&D terminal equipment for display via the network communication interface.

[0256] While specific embodiments of the present invention have been described above, those skilled in the art should understand that these specific embodiments are merely illustrative. Those skilled in the art can omit, substitute, and modify the details of the above methods and systems in various ways without departing from the principles and essence of the present invention. For example, combining the above method steps to perform substantially the same function and achieve substantially the same result using substantially the same method falls within the scope of the present invention.

Claims

1. A data analysis and management system for enterprise technology mining based on a large language model, characterized in that: include: The text structure parsing module is used to receive multi-source technical text streams, call a large language model to extract a set of structured tuples, perform directional semantic classification on the statements containing engineering parameters to generate three-state trend vector values, and use the three-state trend vector values ​​as alternative state parameters when the absolute numerical boundaries of the engineering parameters are missing. A two-layer knowledge graph construction module is used to establish a technical component structure layer and a parameter constraint layer based on the structured tuple set, configure cross-layer mounting pointers, and allocate edge weights between engineering parameters in the parameter constraint layer according to the three-state trend vector values. The temporal conflict matrix calculation module is used to convert the parameter constraint layer into a temporal conflict adjacency matrix, and calculate the coupling interference degree of the target parameter pair consisting of the starting engineering parameter and the target engineering parameter corresponding to the negative weight data in the edge weights. The extraction rule feedback control module is used to determine the target entity node set based on the cross-layer mounting pointer when the coupling interference degree value exceeds the interference degree trigger threshold, and generate a directional constraint template. The directional constraint template is fed back to the text structure parsing module so that when the subsequent input text hits the entity corresponding to the target entity node set, the text parsing range is limited based on the directional constraint template. The local topology verification and output module is used to extract the K-Hop local neighborhood subgraph associated with the target entity node set when the edge weight between the target parameter pairs in the parameter constraint layer changes from a historical negative value to a zero value or a positive value, calculate the approximate graph edit distance, and generate a technology evolution analysis report based on the approximate graph edit distance.

2. The enterprise technology mining data analysis and management system based on a large language model according to claim 1, characterized in that, The structured tuple set contains multiple tuple objects, and the tuple objects include at least technical component entities, engineering problems, engineering parameters, absolute numerical boundaries, three-state trend vector values, coupling effect parameters, and extraction confidence. The text structure parsing module is specifically used to perform the following operations: The combination of numeric characters and units of measurement in adjacent text segments of the engineering parameters is used as the absolute numerical boundary. The large language model is instructed to perform directional semantic classification on the statement containing the engineering parameter, and output the three-state trend vector value with a value of +1, 0 or −1; When the absolute numerical boundary is missing, a trend state mapping instruction is triggered, and the three-state trend vector value is used as a substitute state parameter for the absolute numerical boundary. When the extraction confidence level is lower than the preset confidence threshold, the tri-state trend vector value in the corresponding tuple object is not used for the calculation of the negative weight data, or the corresponding tuple object is written into the manual review queue.

3. The enterprise technology mining data analysis and management system based on a large language model according to claim 2, characterized in that, The two-layer knowledge graph construction module is specifically used to perform the following operations: The extracted technical component entities are constructed as entity nodes of the technical component structure layer; The extracted engineering parameters are constructed as parameter nodes of the parameter constraint layer; Write the globally unique identifier of the technical component entity into the parameter node attribute corresponding to the engineering parameter to generate the cross-layer mounting pointer; When the coupling influence parameter exists in the tuple object, a directed edge is generated between the originating parameter node and the target parameter node, where both the originating parameter node and the target parameter node are parameter nodes in the parameter constraint layer. Retrieve associated tuple objects in the structured tuple set that match the engineering parameters with the coupling influence parameters, and read the tri-state trend vector value in the associated tuple object as the tri-state trend vector value corresponding to the target parameter node; Compare the tri-state trend vector values ​​corresponding to the initiating parameter node and the target parameter node with the expected evolution direction label; When the originating parameter node is in a performance optimization state and the target parameter node enters a performance degradation state, the edge weight of the directed edge is assigned to the negative weight data of a preset constant.

4. The enterprise technology mining data analysis and management system based on a large language model according to claim 1, characterized in that, The temporal conflict matrix calculation module performs the following operations when converting the parameter constraint layer into a temporal conflict adjacency matrix: Establish a global project parameter node pool; Assign a fixed matrix index to each engineering parameter in the parameter constraint layer; The capacity of the global engineering parameter node pool is dynamically expanded when a new engineering parameter is detected. Add a new fixed matrix index to the newly added engineering parameters; In the temporal conflict adjacency matrix of the historical time node, zero-padding is performed on the row and column positions corresponding to the new fixed matrix index.

5. The enterprise technology mining data analysis and management system based on a large language model according to claim 1, characterized in that, The temporal conflict matrix calculation module performs the following operations when performing temporal weighted accumulation on the negative weight data: For the target parameter pairs within a specific observation period, obtain the absolute value scalar of the negative weight data at each time node; Calculate the time span scalar between the current time point and the previous time point; The time decay weight corresponding to each time node is calculated using an exponential decay method. The absolute value scalar, the time span scalar, and the time decay weight are multiplied to obtain the product value; The product values ​​are accumulated and the coupling interference degree value is output.

6. The enterprise technology mining data analysis and management system based on a large language model according to claim 1, characterized in that, The extraction rule feedback control module is specifically used to perform the following operations: When the coupling interference value exceeds the interference trigger threshold, read the cross-layer mounting pointer of each engineering parameter in the target parameter pair; A search is initiated at the technical component structure layer based on the globally unique identifier recorded in the cross-layer mounting pointer; The shortest path algorithm is invoked to search for component connection paths in the technical component structure layer that connect the technical component entities corresponding to the two engineering parameters in the target parameter pair; The set of components and endpoint entities involved in the component connection path are marked as the target entity node set; When there is no connection path between the two technical component entities, the two technical component entities are added to the target entity node set respectively.

7. The enterprise technology mining data analysis and management system based on a large language model according to claim 6, characterized in that, The specific logic of the extraction rule feedback control module in generating a directional constraint template to constrain the subsequent text parsing range is as follows: When the input text matches the entity name or alias corresponding to the target entity node set, the orientation constraint template is activated to perform text parsing, and a local text window is determined based on the position of the entity name or alias in the input text; When the input text does not match the entity name or entity alias corresponding to the target entity node set, the initial constraint template is used to perform text parsing. When the large language model is in interface call mode, only the text sequence within the local text window is input into the large language model; When the large language model is deployed locally and the inference framework has the attention mask parameter enabled, the attention mask matrix elements corresponding to the text identifiers outside the local text window are set to negative infinity. In the directional constraint template, the decision priority of the three-state trend vector value is set to be higher than that of the three-state trend vector value in the initial constraint template.

8. The enterprise technology mining data analysis management system based on a large language model according to claim 1, characterized in that, The local topology verification and output module is specifically used to perform the following operations when extracting the K-Hop local neighborhood subgraph associated with the target entity node set: Send a write lock command to the system main control unit for the target parameter pair and its associated map region; Using the target entity node set as the seed node, a restricted breadth-first search algorithm is executed in the technical component structure layer. The execution logic of the restricted breadth-first search algorithm is to expand the search for adjacent nodes layer by layer according to a set number of hops. Extract the topology from the graph cross section corresponding to the old time node and the graph cross section corresponding to the new time node, respectively; The extracted topology is used to generate old K-Hop local neighborhood subgraphs and new K-Hop local neighborhood subgraphs, respectively. The new time node is the time node at which the edge weight between the target parameter pairs in the parameter constraint layer changes from a negative value to a zero value or a positive value, and triggers the extraction of the K-Hop local neighborhood subgraph; the old time node is determined by the alarm record that is closest to the new time node before the new time node, and the alarm record is the alarm record formed when the edge weight between the target parameter pairs in the parameter constraint layer is negative and the coupling interference degree of the target parameter pairs exceeds the interference degree trigger threshold.

9. The enterprise technology mining data analysis management system based on a large language model according to claim 8, characterized in that, The local topology verification and output module is specifically used to perform the following operations when calculating the approximate map editing distance: A heuristic algorithm based on bipartite graph matching is invoked to compare the differences between the old K-Hop local neighborhood subgraph and the new K-Hop local neighborhood subgraph. Calculate the edit distance of the node name strings between nodes in the old subgraph and nodes in the new subgraph; Extract the set of locally connected edges associated with each node in the old subgraph and each node in the new subgraph, and calculate the difference in the number of edges and the difference in the in-degree direction; The node replacement cost is generated by normalizing and summing the edit distance of the node name string, the difference in the number of edges, and the difference in the in-degree direction; Unmatched detached nodes in the old subgraph are included in the cost of deleting nodes, and unmatched detached nodes in the new subgraph are included in the cost of inserting nodes; Find the minimum total cost in the bipartite graph mapping scheme space, and use the minimum total cost as the approximate graph edit distance.

10. The enterprise technology mining data analysis and management system based on a large language model according to claim 9, characterized in that, The local topology verification and output module is specifically used to perform the following operations when generating the technology evolution analysis report: Extract the cardinality of the first number of nodes contained in the old K-Hop local neighborhood subgraph and the cardinality of the second number of nodes contained in the new K-Hop local neighborhood subgraph respectively; The maximum value between the first and second node number bases is multiplied by a preset topology tolerance coefficient to obtain the topology change threshold. When the approximate graph editing distance is greater than the topology change threshold, the current technology evolution path is classified as a structural reconstruction type evolution, and the unmatched old subgraph free nodes and the unmatched new subgraph free nodes are marked as structural change nodes; When the approximate graph editing distance is not greater than the topology change threshold, the current technology evolution path is classified as an original component attribute improvement candidate path, and material properties, manufacturing process, control algorithm, component model, operating condition, interface protocol or environmental constraint fields are extracted a second time from one or more input texts corresponding to the new time node that are used to determine whether the edge weights between the target parameter pairs change from negative to zero or positive as R&D focus items. The target parameter pair, each engineering parameter in the target parameter pair, and the corresponding technical component entity, evolution path type, approximate graph editing distance, topology change threshold, structural change node or R&D focus item, along with the absolute numerical boundary or the three-state trend vector value, are encapsulated into the technology evolution analysis report.

Images