A scientific research business management and control system based on large model technology

Abstract

The application provides a scientific research business management and control system based on large model technology, comprising a data acquisition module for acquiring multi-source heterogeneous data in the whole process of power scientific research business; a knowledge extraction module for extracting knowledge from the multi-source heterogeneous data according to a power knowledge graph and a data extraction model to obtain a semantic matrix; a model reasoning module for reasoning knowledge from the semantic matrix according to an intelligent reasoning large model to obtain decision information; a bidirectional association interface for determining a reasoning direction according to a structured field constraint of the data extraction model and determining prior knowledge of the intelligent reasoning large model according to the power knowledge graph; and a business management and control module for managing and controlling the scientific research business according to the decision information. Through multi-module cooperation and deep integration of large models and domain knowledge, the application realizes full mining and accurate analysis of multi-source heterogeneous data in the whole process of power scientific research, makes the management and control basis more comprehensive, and guarantees the accuracy and scene adaptability of the reasoning direction.

Description

Technical Field

[0001] This invention relates to the field of power business management technology, and in particular to a scientific research business control system based on large model technology. Background Technology

[0002] As the power industry accelerates its transformation towards new energy and intelligent technologies, power research is exhibiting significant characteristics: diversified research directions, complex participating entities, and heterogeneous data formats. The entire power research process encompasses multiple stages, including project application, experiment execution, and results acceptance, involving multi-source heterogeneous data such as text, time-series, and image data, with the data volume growing exponentially. Traditional power research management relies heavily on manual review and experience-based judgment, resulting in significant problems such as low management efficiency, delayed risk prediction, and imbalanced resource allocation. Specifically, in the project application stage, manual review struggles to quickly and accurately identify technical feasibility defects and compliance issues in application materials; in the experiment execution stage, it's impossible to track experimental progress and data quality in real time, leading to potential deviations from experimental procedures and data distortion; and in the results acceptance stage, the lack of objective data to assess the innovation and practicality of research results results results in a high degree of subjectivity in management decisions, making it difficult to meet the refined management needs of modern power research.

[0003] While existing technologies for scientific research business management have attempted to improve the intelligence level of management by introducing technologies such as data mining and knowledge graphs, there are still significant limitations: On the one hand, traditional data processing technologies are unable to achieve deep integration and efficient analysis of multimodal heterogeneous data, and cannot fully mine the scientific research business-related information contained in the data, resulting in incomplete management basis; on the other hand, existing management models are mostly single-function models for specific links, lacking deep coupling with knowledge in the field of power research, and the models have insufficient adaptive capabilities, making it difficult to adapt to the dynamic changes in different research scenarios and unable to form a closed-loop management system covering the entire process and all dimensions. Summary of the Invention

[0004] This invention provides a scientific research business management and control system based on large model technology, which addresses the problem of insufficient intelligence in existing scientific research business management and control systems.

[0005] A first aspect of this invention provides a scientific research business management and control system based on large model technology, comprising: The data acquisition module is used to acquire multi-source heterogeneous data throughout the entire process of power research and development. The knowledge extraction module is used to extract knowledge from multi-source heterogeneous data based on the power knowledge graph and data extraction model to obtain a semantic matrix. The model reasoning module is used to perform knowledge reasoning on the semantic matrix based on the intelligent reasoning big model to obtain decision information. The intelligent reasoning big model is a four-level structure consisting of a base layer, an adaptive enhancement layer, a reasoning layer, and a bidirectional association interface. The intelligent reasoning big model establishes a deep association with the data extraction model and the power knowledge graph through the bidirectional association interface. The bidirectional association interface determines the reasoning direction based on the structured field constraints of the data extraction model and determines the prior knowledge of the intelligent reasoning big model based on the power knowledge graph. The business management module is used to manage scientific research operations based on decision-making information.

[0006] In one possible implementation, the model inference module is used for: Based on the bidirectional association interface, the structured field constraints of the data extraction model are extracted as the reasoning direction, and the knowledge association relationships of the power knowledge graph are extracted as prior knowledge. The semantic matrix is ​​input into the base layer to obtain the general semantic features fused from multiple modalities; By inputting general semantic features into the adaptive enhancement layer of the intelligent reasoning model, we obtain semantic features specific to power research. The semantic features specific to power research are input into the reasoning layer of the intelligent reasoning model to obtain decision information.

[0007] In one possible implementation, the model inference module is used for: Establish a data interaction link with the data extraction model; The data extraction model is analyzed by its built-in structured field configuration specific to power research, and field constraints that represent the core business objectives of power research are filtered out. Transform field constraints into inference direction instructions that the model can recognize; Establish knowledge interaction links with the power knowledge graph; Through knowledge query and extraction algorithms, we extract various types of knowledge relationships in the field of power research from the power knowledge graph. The knowledge relationships are transformed into prior knowledge in vector form through knowledge embedding technology.

[0008] In one possible implementation, the model inference module is used for: Modality dimension parsing is performed on the input semantic matrix to identify the submatrices corresponding to the semantic matrix; where each submatrix corresponds to a modality. For sub-matrices of different modes, the feature extraction unit of the corresponding mode is used to perform initial feature extraction to obtain the initial feature vector corresponding to each mode; The association weights between the initial feature vectors of different modalities are calculated, and weighted fusion is performed based on the association weights to obtain general semantic features.

[0009] In one possible implementation, the model inference module is used for: Based on prior knowledge and inference direction instructions, a two-layer enhanced benchmark is constructed; By decoupling the general semantic features, the core semantic vector is obtained; The core semantic vector is enhanced in a hierarchical manner using a multi-level enhancement network to obtain scene enhancement features. The semantic fit between the enhanced features of the calculation scenario and the power knowledge graph and the matching degree with the reasoning direction instructions are calculated. If the semantic fit does not reach the first preset threshold or the matching degree does not reach the second preset threshold, the parameters of the multi-level enhanced network are dynamically adjusted and the hierarchical enhancement process is re-executed. If the semantic fit reaches the first preset threshold and the matching degree reaches the second preset threshold, then the scene enhancement features are dimensionally integrated and normalized to output semantic features specific to power research.

[0010] In one possible implementation, the model inference module is used for: Establish a reasoning constraint framework based on the reasoning direction instructions; The semantic features specific to power research are input into the inference layer to generate preliminary inference results under the inference constraint framework; The preliminary reasoning results are verified to determine the decision-making information.

[0011] In one possible implementation, the knowledge extraction module is used for: The acquired heterogeneous data from multiple sources is preprocessed to obtain the data to be extracted. The similarity between the data to be extracted and the power knowledge graph is calculated by a cross-layer semantic matching algorithm to determine the core extraction task. Activate the scene sub-network and task output header in the data extraction model that correspond to the core extraction task; The data to be extracted is input into the data extraction model to obtain the semantic matrix.

[0012] In one possible implementation, the knowledge extraction module is used for: Construct a joint semantic vector space for the core entity layer and the rule logic layer in the power knowledge graph; The data to be extracted is mapped to the joint semantic vector space to obtain the data semantic vector; Calculate the semantic similarity between the data semantic vector and various scientific research entity nodes in the core entity layer, as well as with various business rule nodes in the rule logic layer; Based on preset entity thresholds and rule thresholds, entity nodes and rule nodes with high similarity matching are selected. Based on the category attributes of the selected entity nodes and the constraint types of the rule nodes, the scientific research business scenarios corresponding to the data to be extracted and the core extraction tasks to be performed are determined.

[0013] In one possible implementation, the knowledge extraction module is used for: Based on the determined scientific research business scenario and core extraction task, locate and activate the scenario sub-network corresponding to the scientific research business scenario from the multi-level adaptation network of the data extraction model, and load the professional feature dictionary and extraction template built into the scenario sub-network. Synchronously locate and activate the task output head corresponding to the core extraction task, and configure the feature mapping parameters and result output format of the output head.

[0014] In one possible implementation, the knowledge extraction module is used for: The preprocessed data to be extracted is input into the data extraction model to generate enhanced data with domain labels; Based on the extraction template built into the scenario sub-network, targeted knowledge extraction is performed on the enhanced data to obtain initial knowledge fragments containing scientific research entities, entity relationships, and business rule matching information. The initial knowledge fragments are fed back into the power knowledge graph to obtain a standardized knowledge set; The standardized knowledge set is converted into semantic vectors with consistent dimensions, and the semantic vectors are arranged in an ordered manner and integrated into a matrix to obtain a semantic matrix.

[0015] Compared to traditional technologies, this invention provides a research business management and control system based on a large-scale model technology. The system includes a data acquisition module for acquiring multi-source heterogeneous data throughout the entire power research business process; a knowledge extraction module for extracting knowledge from the multi-source heterogeneous data based on a power knowledge graph and a data extraction model to obtain a semantic matrix; and a model reasoning module for performing knowledge reasoning on the semantic matrix based on an intelligent reasoning large-scale model to obtain decision information. The intelligent reasoning large-scale model has a four-level structure consisting of a base layer, an adaptive enhancement layer, a reasoning layer, and a bidirectional association interface. The intelligent reasoning large-scale model establishes a deep association with the data extraction model and the power knowledge graph through the bidirectional association interface. The bidirectional association interface determines the reasoning direction based on the structured field constraints of the data extraction model and determines the prior knowledge of the intelligent reasoning large-scale model based on the power knowledge graph. Finally, a business management and control module is used to manage and control research business based on the decision information. This invention, through multi-module collaboration and deep integration of large models and domain knowledge, achieves, on the one hand, the full mining and accurate analysis of multi-source heterogeneous data throughout the entire process of power research, making the basis for control more comprehensive; on the other hand, it ensures the accuracy of reasoning direction and scenario adaptability, ultimately realizing closed-loop control of the entire process of scientific research, significantly improving control efficiency and accuracy. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the structure of the scientific research business management and control system based on large model technology provided in the embodiment of the present invention. Detailed Implementation

[0017] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0018] Figure 1 This is a schematic diagram of the structure of a scientific research business management and control system based on large model technology provided in an embodiment of the present invention. Figure 1 As shown, the system includes: Data acquisition module 11 is used to acquire multi-source heterogeneous data in the entire process of power scientific research business; The knowledge extraction module 12 is used to extract knowledge from multi-source heterogeneous data based on the power knowledge graph and data extraction model to obtain a semantic matrix. Model reasoning module 13 is used to perform knowledge reasoning on the semantic matrix based on the intelligent reasoning big model to obtain decision information. The intelligent reasoning big model is a four-level structure consisting of a base layer, an adaptive enhancement layer, a reasoning layer, and a bidirectional association interface. The intelligent reasoning big model establishes a deep association with the data extraction model and the power knowledge graph based on the bidirectional association interface. The bidirectional association interface determines the reasoning direction based on the structured field constraints of the data extraction model and determines the prior knowledge of the intelligent reasoning big model based on the power knowledge graph. The business management module 14 is used to manage scientific research operations based on decision-making information.

[0019] In this embodiment of the invention, the data acquisition module 11 serves as the core of the system's data input. Its functionality directly determines the accuracy of subsequent knowledge extraction and reasoning. It is mainly responsible for acquiring multi-source heterogeneous data throughout the entire process of power research and development. Here, "entire process" covers the entire lifecycle stages of research project application, project approval, experiment implementation, data monitoring, results acceptance, and results transformation. This ensures that the collected data fully reflects the status of research and development and avoids biased management. "Multi-source heterogeneous data" includes various sources of data such as text data uploaded by researchers, time-series data collected by laboratory monitoring equipment, and simulation data output by simulation platforms, as well as data in different formats such as text, time series, and images. To achieve efficient data acquisition, the data acquisition module 11 adopts a hybrid strategy of "multi-interface adaptation + real-time / timed acquisition". It configures corresponding adaptation interfaces for data from different sources, such as industrial communication interfaces with laboratory monitoring equipment and API interfaces with simulation platforms. The acquisition frequency can be dynamically adjusted according to business needs. At the same time, it has a built-in data preprocessing unit to perform preliminary noise reduction, format standardization and add metadata to the acquired data, forming a standardized raw data set that is transmitted to the knowledge extraction module 12.

[0020] The knowledge extraction module 12 receives the output from the data acquisition module 11. Its core function is to convert unstructured / semi-structured raw data into a structured knowledge representation that computers can understand, namely a semantic matrix, providing high-quality input for subsequent intelligent reasoning. This module performs basic entity and relation extraction through a data extraction model, and then performs semantic verification and correlation completion on the extraction results based on the power knowledge graph. Finally, it integrates standardized knowledge fragments through vector encoding and matrixization to generate a semantic matrix, serving as a key bridge connecting raw data and intelligent reasoning. The model reasoning module 13, as the core reasoning unit of the system, relies on a four-level intelligent reasoning model to perform knowledge reasoning on the semantic matrix to obtain decision information. This four-level structure includes a base layer, an adaptive enhancement layer, a reasoning layer, and a bidirectional association interface. Through the bidirectional association interface, structured field constraints can be extracted from the data extraction model to clarify the reasoning direction, and knowledge relationships can be extracted from the power knowledge graph as prior knowledge to ensure the professionalism and accuracy of the reasoning.

[0021] The bidirectional association interface is a software interface with built-in data integrity verification and exception retransmission mechanisms, enabling synchronous interaction with the data extraction model and the power knowledge graph. Its bidirectional data flow logic is as follows: the structured field constraints of the data extraction model and the knowledge relationships of the power knowledge graph are synchronously transmitted to the interface. After the interface completes the format adaptation of the constraints and knowledge through a semantic conversion engine, it is pushed to the intelligent inference model in real time, while simultaneously receiving status feedback information from the model, forming a closed-loop interaction of 'extraction-adaptation-transmission-feedback'. The interface establishes logical connections with other modules through memory sharing, ensuring low latency and high reliability of data flow, while also supporting dynamic adjustment of the interaction frequency based on business complexity.

[0022] Interface core architecture and components Two-way interactive channel: It includes two independent data transmission links, namely the "constraint instruction transmission link" and the "knowledge support transmission link". The two links operate in parallel and verify each other to ensure the independence and integrity of data transmission.

[0023] Semantic conversion engine: Built-in semantic mapping dictionary and format adaptation rule library specifically for the field of power research, supporting the conversion of unstructured field constraints and knowledge associations into vector instructions and prior knowledge vectors that the model can recognize.

[0024] Data verification unit: It adopts a triple verification mechanism of "format verification + logic verification + integrity verification". Format verification ensures that the transmitted data conforms to the interface preset specifications. Logic verification verifies the business rationality of the constraint instructions and knowledge association (such as avoiding the mismatch transmission of "experiment parameter constraints" and "result acceptance knowledge"). Integrity verification ensures that the data is not lost or tampered with through hash value comparison.

[0025] Dynamic Feedback Module: It collects the inference status data of the intelligent inference large model in real time (such as semantic fit degree, matching degree verification result, parameter adjustment record), and generates a feedback report to provide a basis for the field constraint optimization of the data extraction model and the knowledge update of the power knowledge graph.

[0026] Cache and Scheduling Unit: It has a built-in distributed cache pool to cache the structured field constraints and core knowledge association relationships with high-frequency access. The cache validity period can be dynamically adjusted according to the business update frequency (default 24 hours); the scheduling unit adopts a load balancing algorithm. When there are concurrent requests in multiple business scenarios, resources are preferentially allocated to high-priority scenarios (such as the experimental risk warning scenario).

[0027] Business Management and Control Module 14 is the core of the system's management and control output. Its core function is to convert abstract decision-making information into specific implementable management and control actions, and implement precise management and control over the entire process of scientific research business, which is the ultimate embodiment of the system's value. Its management and control scope comprehensively covers all life cycle links of power scientific research from project application and approval to achievement transformation. The management and control method is not single and fixed, but is designed differentially based on the business characteristics and risk points of each link: In the application stage, for the compliance review conclusion in the decision-making information, hierarchical management and control are implemented for the applied projects. For projects that meet the application conditions and have feasible technical solutions, the project approval review process is automatically triggered. For projects with problems such as missing materials and unreasonable plans, a specific rectification list is accurately output (such as supplementing certain types of technical verification materials, optimizing the design of key technical parameters), and the subsequent process is suspended until the rectification is completed and passed the review; In the experimental implementation stage, the response mechanism is triggered hierarchically based on the risk warning information. If it is a general parameter deviation warning, a warning notice and parameter adjustment suggestions are automatically pushed to the scientific research personnel. If it is a major risk warning such as equipment failure and abnormal experimental environment, the laboratory control system is immediately linked to suspend the relevant experimental process, and the equipment maintenance personnel and project leaders are notified to handle it simultaneously. At the same time, combined with the analysis results of the progress deviation, a resource coordination plan is generated for the lagging experimental links (such as allocating additional experimental equipment, adjusting personnel scheduling) to ensure the progress of the project; In the achievement acceptance stage, according to the innovation, practicality, and compliance scores in the achievement evaluation report, hierarchical identification of the achievements is implemented. For qualified achievements, the acceptance approval procedures are handled and achievement transformation suggestions are pushed. For unqualified achievements, specific improvement directions and supplementary verification requirements are clarified; In addition, this module also has the ability of cross-link collaborative management and control, which can empower the scheme review in the application stage with the risk handling experience in the experimental stage and optimize the review index system.

[0028] To ensure the accurate and efficient implementation of control actions, the business control module 14 adopts a three-layer architecture of "rule engine + visual interaction + linkage execution," with each layer working in tandem: The rule engine is the core of the control logic, with a built-in standardized control rule library in the field of power research, covering project review rules, experimental safety specifications, and results acceptance standards. It can automatically match the corresponding control rules based on the type and content of the input decision information, generate standardized control instructions, and support managers to dynamically update the rule library according to business changes; The visual interaction layer provides managers with a panoramic control interface, which can not only display the status of the entire control process, details of decision information, and the execution progress of control instructions in real time, but also support managers to review the results of automatic control, implement manual intervention in special scenarios, and issue custom control instructions, balancing the efficiency of automated control with the flexibility of manual control; The linkage execution layer is the key to the implementation of control instructions. Through standardized interfaces, it connects with various business systems such as project management system, laboratory control system, results management system, and financial system, transforming control instructions into specific system operations, such as process start / stop, notification push, parameter adjustment, and permission configuration, to achieve automated execution of control actions. At the collaborative closed-loop level, the business control module 14 does not output control actions in a one-way manner, but forms a complete feedback link: on the one hand, it receives the decision information output by the model inference module 13 as control input, and on the other hand, it feeds back the execution results of control instructions (such as the completion status of rectification, the effect of early warning and disposal, and the conclusion of result acceptance) to the model inference module 13 and the data acquisition module 11 in real time. The data fed back to the model inference module 13 is used to optimize the inference parameters (such as adjusting the risk warning threshold and improving the weight of the evaluation indicators), and the data fed back to the data acquisition module 11 is used to clarify the key directions of subsequent data collection (such as increasing the collection frequency of materials after rectification and strengthening the monitoring of data related to experimental risk points).

[0029] In some embodiments, the model reasoning module is configured to: extract structured field constraints of the data extraction model as reasoning direction based on the bidirectional association interface, and extract knowledge association relationships of the power knowledge graph as prior knowledge; input the semantic matrix into the base layer to obtain multimodal fusion general semantic features; input the general semantic features into the adaptive enhancement layer of the intelligent reasoning large model to obtain power research-specific semantic features; and input the power research-specific semantic features into the reasoning layer of the intelligent reasoning large model to obtain decision information.

[0030] In this embodiment of the invention, the model reasoning module serves as the core unit for intelligent decision-making in the system. Its workflow follows a core framework of "knowledge association guidance - hierarchical feature processing - precise reasoning output." It establishes deep collaboration with the data extraction model and the power knowledge graph through a bidirectional association interface, ensuring that the reasoning process aligns with the needs of power research and development and is supported by professional knowledge. Specifically, this module first conducts bidirectional knowledge interaction and constraint extraction through the bidirectional association interface: on the one hand, it establishes a standardized data interaction link with the data extraction model, parsing the power research-specific structured field configurations built into the model. These configurations cover fields strongly related to core business, such as research project number, experimental indicator type, equipment parameter range, and results acceptance dimensions. The module then selects field constraints representing the core business objectives of power research and development, and transforms them into reasoning direction instructions recognizable by the intelligent reasoning model through a semantic mapping algorithm. For example, it transforms the field constraints related to "experimental parameter compliance verification" into "reasoning for the risk of parameter deviation in the experimental process," clarifying the core objectives and scope of the reasoning and preventing the reasoning process from deviating from actual business management needs; on the other hand... On the one hand, an efficient knowledge interaction link is established with the power knowledge graph. Through knowledge query and extraction algorithms (such as knowledge walking algorithms based on graph neural networks), various types of knowledge relationships in the field of power research are extracted from the power knowledge graph. These include the relationship between research entities (such as power equipment and experimental indicators), the temporal relationship between business processes (such as application-review-project approval), and the constraint relationship between technical standards (such as experimental parameter thresholds and compliance judgment). Then, through knowledge embedding technology (such as TransE, DistMult, etc.), these knowledge relationships are transformed into prior knowledge in vector form, which can be directly integrated into the reasoning and calculation process of large models. This provides professional domain knowledge support for reasoning and ensures that the reasoning results comply with the technical specifications and business logic of the power industry. After acquiring the reasoning direction and prior knowledge, the model reasoning module processes the semantic matrix output by the knowledge extraction module step by step according to the hierarchical logic of the base layer, adaptive enhancement layer, and reasoning layer, realizing the transformation from general features to specific features, and then to decision information. The core role of the base layer is to achieve unified feature fusion of the multimodal semantic matrix. Since the input semantic matrix is ​​a structured representation of multi-source heterogeneous data after knowledge extraction, it contains sub-matrices corresponding to different modalities such as text, time series, and images. The base layer first performs modal dimension parsing on the semantic matrix, accurately locates the data source type corresponding to each sub-matrix through modality recognition algorithms, and then matches specific feature extraction units for different modal sub-matrices. For example, a BERT-like pre-trained model is used to extract semantic features from text sub-matrices, a temporal convolutional network (TCN) is used to extract trend features from temporal sub-matrices, and a convolutional neural network (CNN) is used to extract visual features from image sub-matrices, thus obtaining initial feature vectors for each modality. Subsequently, the correlation strength between the initial feature vectors of different modalities is quantified based on the mutual information calculation method to generate dynamic correlation weights. Then, the feature vectors of each modality are weighted and fused according to the weights to obtain a general semantic feature of multimodal fusion, realizing a unified semantic representation of data of different forms, and laying the foundation for subsequent domain feature enhancement. After the general semantic features are generated, the model inference module inputs them into the adaptive enhancement layer. The core objective of this layer is to strengthen features specific to the power research field and weaken general domain-irrelevant features, thereby achieving scenario-based feature adaptation. Specifically, the adaptive enhancement layer first constructs a two-layer enhancement benchmark based on prior knowledge obtained from the bidirectional association interface and inference direction instructions: a "domain knowledge benchmark" and a "business target benchmark." The domain knowledge benchmark is constructed based on prior knowledge from the power knowledge graph and is used to measure the domain relevance of features, while the business target benchmark is constructed based on inference direction instructions and is used to measure the target adaptability of features. Subsequently, the general semantic features are decoupled through an attention mechanism, separating the core semantic vectors that are strongly correlated with the two-layer enhancement benchmark and irrelevant redundant vectors. Redundant vectors are removed to improve feature purity. Next, a multi-level enhancement network is used to perform hierarchical enhancement processing on the core semantic vectors. Each layer of this network introduces prior knowledge vectors for feature calibration, gradually strengthening the features. The system incorporates features specific to power research (such as parameter features of specific power equipment and key indicator features of new energy experiments). During the enhancement process, the module calculates in real time the semantic fit between the enhanced features and the power knowledge graph (using a cosine similarity algorithm) and the matching degree with the inference direction instructions (using a semantic matching score). If the semantic fit does not reach the first preset threshold or the matching degree does not reach the second preset threshold, the inter-layer weights and activation function parameters of the multi-level enhancement network are dynamically adjusted using a gradient descent algorithm, and the layered enhancement process is re-executed. Once both indicators meet the criteria, the enhanced features are dimensionally integrated and L2 normalized to ensure uniform feature dimensions and standardized numerical ranges, ultimately outputting semantic features specific to power research that accurately match the needs of power research scenarios. Finally, the model inference module inputs the semantic features specific to power research into the inference layer, completing the final transformation from features to decision information. The inference layer first establishes a targeted inference constraint framework based on the inference direction instructions determined by the bidirectional association interface. This framework contains inference rules and logical paradigms corresponding to the inference direction. For example, for the "experimental risk warning" inference direction, the framework contains "parameter deviation threshold - risk level correspondence rules" and "multi-parameter collaborative deviation - risk superposition rules," etc. Subsequently, the semantic features specific to power research are input into this constraint framework, and deep semantic inference is performed through a Transformer-based inference decoder to generate preliminary inference results, such as "the voltage parameter of a certain experimental link deviates from the threshold by 15%, indicating a medium risk" or "a project application material is missing 3 essential technical certifications, failing to meet compliance standards," etc. To ensure the accuracy of the inference results, the module also uses prior knowledge to perform a secondary verification of the preliminary inference results, including verifying the consistency of the inference results with the technical standards in the power knowledge graph and the coherence with the business process logic, eliminating deviation results with semantic conflicts and logical contradictions. If there are deviations in the preliminary inference results, the module will trace back to the adaptive enhancement layer, adjust the feature enhancement parameters, and re-process and infer the features until the final decision information that meets the requirements is obtained. This decision-making information is output in a structured format, including core content such as control conclusions, risk levels, rectification suggestions, and supporting clauses. It can directly provide clear and accurate control basis for the business control module, supporting the implementation of precise control actions in subsequent scientific research operations.

[0031] In some embodiments, the model reasoning module is used to: establish a data interaction link with the data extraction model; parse the power research-specific structured field configuration built into the data extraction model and filter out field constraints that represent the core business objectives of power research; convert the field constraints into reasoning direction instructions that the model can recognize; establish a knowledge interaction link with the power knowledge graph; extract multiple types of knowledge relationships in the power research field from the power knowledge graph through knowledge query and extraction algorithms; and convert the knowledge relationships into prior knowledge in vector form through knowledge embedding technology.

[0032] In this embodiment of the invention, the model reasoning module serves as the core unit for the system to achieve intelligent decision-making. Its workflow revolves around "knowledge association guidance - hierarchical feature processing - accurate reasoning output." Through a bidirectional association interface, it establishes deep collaboration with the data extraction model and the power knowledge graph. This collaboration is not simply information transmission, but rather forms a bidirectional driving mechanism of "constraint guidance - knowledge empowerment." This ensures that the reasoning process closely aligns with the actual management and control needs of power scientific research and business, while also possessing solid professional knowledge support, thereby fundamentally improving the accuracy and practicality of decision-making information.Specifically, this module first conducts bidirectional knowledge interaction and constraint extraction through a bidirectional association interface. This process is a prerequisite for achieving targeted and professional reasoning: On the one hand, a data interaction link based on a standardized protocol is established between the module and the data extraction model to ensure the stability and compatibility of data transmission. Subsequently, the structured parsing engine parses the power research-specific structured field configuration built into the data extraction model. This configuration is based on the pre-set business requirements of the entire power research process and covers field information strongly related to core business, such as research project number, experimental indicator type, equipment parameter range, result acceptance dimension, and funding usage category. Each field also includes metadata such as business meaning description, data type constraints, and value range definition. The module uses a core target identification algorithm to filter out field constraints that represent the core business objectives of power research, eliminates redundant auxiliary field information, and then uses a semantic mapping algorithm (such as a semantic conversion algorithm based on a pre-trained language model) to transform the filtered field constraints into structured reasoning direction instructions that can be recognized by the intelligent reasoning model. The instructions clearly contain key information such as reasoning topic, core dimension, and judgment criteria. For example, the field constraints related to "experimental parameter compliance verification" are transformed into "for experimental parameters deviating from the wind direction verification." The core dimensions of the reasoning for risk assessment are voltage, current, and temperature, which clearly define the core objectives and scope of the reasoning, fundamentally preventing the reasoning process from deviating from actual business management needs. On the other hand, an efficient knowledge interaction link based on graph query protocols is established between the module and the power knowledge graph. Through knowledge query and extraction algorithms (such as knowledge walking algorithms based on graph neural networks and rule-based knowledge extraction algorithms), various types of knowledge relationships in the field of power research are accurately extracted from the power knowledge graph. Specifically, this includes the relationships between research entities (such as power equipment-experimental indicators, materials-performance parameters) and business processes (such as application-review-...). The module establishes a series of relationships, including the temporal relationships of project initiation, implementation, and acceptance; the constraint relationships of technical standards (such as experimental parameter thresholds and compliance judgments, and achievement indicators and grade classifications); and the citation relationships of scientific research results (such as papers, patents, and experimental data). To enable these knowledge relationships to be directly integrated into the reasoning and calculation process of the large model, the module uses knowledge embedding technology to transform them into prior knowledge in the form of fixed-dimensional vectors. This prior knowledge in vector form can be efficiently fused and calculated with subsequent feature vectors, providing solid professional domain knowledge support for the reasoning process and ensuring that the reasoning results strictly comply with the technical specifications and business logic of the power industry.

[0033] After acquiring the reasoning direction and prior knowledge, the model reasoning module strictly follows the hierarchical progression logic of the base layer, adaptive enhancement layer, and reasoning layer. It gradually refines the semantic matrix output by the knowledge extraction module, realizing the transformation from general features to domain-specific features, and then to precise decision-making information. Each level of processing provides a high-quality data foundation for subsequent stages. The core function of the base layer is to achieve unified feature fusion and representation of multimodal semantic matrices. Since the input semantic matrix is ​​a structured result of knowledge extraction from multi-source heterogeneous data (text, time series, images, etc.), it contains sub-matrices corresponding to different modalities, and the feature dimensions and representation forms of each sub-matrix are significantly different. Direct subsequent processing would lead to feature conflicts or information loss. Therefore, the base layer first performs a comprehensive analysis of the semantic matrix through a modality dimension parsing unit, and uses modality recognition algorithms (such as feature-based modality classification algorithms) to accurately locate the data source type (text, time series, images, etc.) corresponding to each sub-matrix, and adds a modality identifier to each sub-matrix. Subsequently, based on the modality identifier, it matches a dedicated feature extraction unit for different modality sub-matrices to ensure the targeting and effectiveness of feature extraction. For example, for text-based sub-matrices (such as semantic sub-matrices corresponding to application materials and technical documents), a BERT-like pre-trained model is used for semantic feature extraction. A multi-head attention mechanism is used to capture the contextual semantic relationships within the text, generating high-dimensional text semantic feature vectors. For time-series sub-matrices (such as semantic sub-matrices corresponding to experimental monitoring data and simulation curves), a temporal convolutional network (TCN) combined with a long short-term memory network (LSTM) is used to extract trend features. This captures both the local variation features of the time-series data and preserves long-term dependencies, generating time-series feature vectors. For image-based sub-matrices (such as semantic sub-matrices corresponding to equipment status images and experimental scene images), a convolutional neural network (CNN) (such as ResNet and MobileNet) is used to extract visual features. Multi-layer convolution and pooling operations are used to capture key texture, shape, and other features in the images, generating visual feature vectors. After obtaining the initial feature vectors corresponding to each modality, the module quantifies the correlation strength between the initial feature vectors of different modalities using a mutual information calculation method (such as normalized mutual information NMI), generating dynamic correlation weights. Modal features with higher correlation strength have greater corresponding weights, thus highlighting the role of key modal data. Finally, the feature vectors of each modality are weighted and fused according to the dynamic correlation weights. During the fusion process, a feature alignment algorithm is used to ensure the uniformity of the dimensions and semantic alignment of the feature vectors of different modalities. Finally, a general semantic feature of multimodal fusion is obtained. This feature realizes a unified semantic representation of data of different forms, which not only retains the core information of each modality data, but also eliminates the interference caused by modal differences, laying a solid foundation for subsequent domain feature enhancement.

[0034] After the general semantic features are generated, the model inference module inputs them into the adaptive enhancement layer. This layer is crucial for achieving scenario-based feature adaptation. Its core objective is to generate exclusive semantic features that highly match the power research scenario through precise feature enhancement and redundancy removal. This strengthens information strongly related to power research business within the general semantic features, weakens irrelevant features from the general domain, and improves the efficiency and accuracy of subsequent inference. Specifically, the adaptive enhancement layer first uses prior knowledge obtained from the bidirectional association interface and inference direction instructions as its core basis to construct a two-layer enhancement benchmark: "domain knowledge benchmark - business target benchmark," forming a dual judgment standard for feature selection and enhancement. The domain knowledge benchmark is constructed based on prior knowledge from the power knowledge graph. It generates domain knowledge feature clusters by clustering prior knowledge vectors, which measure the relevance of input features to the power research field; higher relevance equates to greater feature value. The business objective benchmark is constructed based on reasoning direction instructions, transforming these instructions into target feature templates to measure the suitability of input features to specific business management objectives; higher suitability equates to greater feature contribution to reasoning. Subsequently, the module decouples general semantic features using a multi-head attention mechanism. Utilizing a dual-layer enhanced benchmark as the basis for calculating attention weights, it scores the importance of each feature dimension within the general semantic features, thus separating core semantic vectors strongly correlated with the dual-layer enhanced benchmark and redundant vectors weakly correlated or irrelevant to the benchmark. Vector filtering removes redundant vectors, effectively improving the purity of the core semantic vectors and reducing interference from irrelevant information in subsequent reasoning. Next, a multi-level enhanced network is used to perform hierarchical enhancement processing on the core semantic vectors. This network employs a progressive structure design, with each level introducing a feature where the corresponding prior knowledge vector is multiplied element-wise with the core semantic vector. The calibration process gradually strengthens the characteristics specific to power research (such as the rated parameter characteristics of specific power equipment, the key indicator characteristics of new energy grid connection experiments, and the constraint characteristics of power system safe operation), while suppressing non-domain characteristics. During the enhancement process, the module performs dual validity checks on the generated scene enhancement features in real time: on the one hand, it calculates the semantic fit between the scene enhancement features and the domain knowledge feature clusters in the power knowledge graph using the cosine similarity algorithm to determine whether the domain relevance of the features meets the standard; on the other hand, it measures the matching degree between the scene enhancement features and the business target benchmark by calculating the semantic matching score (such as the matching degree calculation based on cross-entropy) to determine whether the target adaptability of the features meets the standard. If the semantic fit does not reach the first preset threshold (this threshold can be dynamically adjusted according to the business needs of the power research field, such as setting it to 0.8) or the matching degree does not reach the second preset threshold (such as setting it to 0.75), the inter-layer weights and activation function parameters (such as ReLU, GELU, etc.) of the multi-level enhancement network are dynamically adjusted using the gradient descent algorithm, and the hierarchical enhancement process is re-executed. When both indicators meet the standard, the scene enhancement features are subjected to dimensional integration and L2 normalization.Dimensional integration ensures that the dimension of the feature vector is consistent with the input requirements of the subsequent inference layer. L2 normalization standardizes the numerical range of the feature vector to the [0,1] interval to avoid the inference result being affected by excessive numerical differences. Finally, the output is a semantic feature specific to power research that accurately matches the needs of power research scenarios. This feature is the key link connecting general features and decision information.

[0035] Finally, the model inference module inputs the semantic features specific to power research into the inference layer. This layer completes the transformation from domain-specific features to final decision information and is the core link in realizing intelligent decision-making. Its design focuses on ensuring the logicality and compliance of the inference process, as well as the accuracy and interpretability of the inference results. The inference layer first establishes a targeted inference constraint framework through the rule parsing engine based on the inference direction instructions determined by the bidirectional association interface. This framework is not fixed but dynamically adjusted according to the inference direction, and contains inference rules, logical paradigms, and domain constraints that strictly correspond to the inference direction.For example, regarding the reasoning direction of "experimental risk warning," the framework incorporates rules such as "parameter deviation threshold - risk level correspondence rules" (e.g., deviation within 5% of the threshold is low risk, 5%-15% is medium risk, and above 15% is high risk), "multi-parameter collaborative deviation - risk superposition rules" (e.g., when voltage and current parameters deviate simultaneously, the risk level is increased by one level), and "experimental stage - risk threshold adaptation rules" (e.g., the risk threshold for critical experimental stages is stricter than that for ordinary stages). Subsequently, the semantic features specific to power research are input into this reasoning constraint framework, and deep semantic reasoning is performed through a Transformer-based reasoning decoder. The decoder utilizes a self-attention mechanism to capture key information and relationships within specific semantic features. Combined with rules and paradigms within the reasoning constraint framework, it generates structured preliminary reasoning results. These results not only contain core conclusions but also include crucial supporting information, such as "a voltage parameter in a certain experimental stage deviates from the threshold by 15%, indicating a medium risk" or "a project application lacks three essential technical certifications (material performance testing report, equipment calibration certificate, and experimental feasibility analysis report), failing to meet compliance standards." To further ensure the accuracy and reliability of the reasoning results, the module also utilizes prior knowledge obtained from the power knowledge graph. The initial inference results undergo secondary verification, employing a two-dimensional verification logic: first, compliance verification, checking the consistency between the initial inference results and the technical standards, industry norms, and business processes included in the power knowledge graph, eliminating deviations that violate regulations; second, logical verification, verifying the internal logical coherence of the initial inference results and their logical consistency with historical inference results and actual business data, eliminating results with logical contradictions. If the initial inference results contain deviations (such as compliance issues or logical contradictions), the module will initiate a reverse tracing mechanism, tracing back to the adaptive enhancement layer, and adjusting feature enhancement parameters according to the deviation type (such as adjusting the annotation). (Including intention weights, inter-layer parameters of multi-level augmented networks, etc.), the feature enhancement and inference process is re-executed until a compliant inference result is generated. The final decision information is output in a standardized structured format, including control conclusions, risk levels (if risk assessment is involved), rectification suggestions (if problem rectification is involved), legal basis, related business links, processing priorities, and other core content. This information can be directly parsed and utilized by the business control module, providing clear, accurate, and implementable control basis for the business control module, effectively supporting the implementation of precise control actions in each link of subsequent scientific research business, and ensuring the pertinence and effectiveness of control measures.

[0036] In some embodiments, the model inference module is used to: perform modality dimension parsing on the input semantic matrix and identify the sub-matrices corresponding to the semantic matrix; wherein each sub-matrix corresponds to a modality; for the sub-matrices of different modalities, use the feature extraction unit of the corresponding modality to perform initial feature extraction to obtain the initial feature vector corresponding to each modality; calculate the correlation weight between the initial feature vectors of different modalities, and perform weighted fusion based on the correlation weight to obtain general semantic features.

[0037] In this embodiment of the invention, constraint guidance ensures that the reasoning direction is anchored to the actual management and control needs of power research, while knowledge empowerment injects professional domain knowledge into the reasoning process. Together, these two aspects fundamentally improve the accuracy and practicality of decision-making information. Specifically, the module first conducts bidirectional knowledge interaction and constraint extraction through a bidirectional association interface. This process is a crucial prerequisite for achieving targeted and professional reasoning: on the one hand, a data interaction link based on a standardized protocol is established between the module and the data extraction model. This link has a built-in data verification and fault tolerance mechanism, which can perform integrity verification and abnormal retransmission of transmitted data, ensuring the stability, compatibility, and reliability of data transmission. Subsequently, the structured parsing engine is activated to deeply analyze the power research-specific structured field configurations built into the data extraction model. These configurations are based on a standardized configuration system preset according to the business needs of the entire power research process, covering research project numbers, experimental indicator types, and equipment. The module includes fields strongly related to core business, such as parameter range, results acceptance dimensions, and funding usage categories. Each field is accompanied by detailed explanations of its business meaning, data type constraints (e.g., numeric, text, enumeration), value range definitions, and validation rules, providing a clear basis for subsequent constraint extraction. The module uses a core objective identification algorithm (a machine learning-based target field classification model) to filter the parsed field configurations, accurately identifying and extracting field constraints representing the core business objectives of power research, and removing redundant auxiliary field information such as project name remarks and data entry time. Then, a semantic mapping algorithm (e.g., a semantic conversion algorithm based on a pre-trained language model) is used to transform the filtered fields... Segment constraints are transformed into structured reasoning direction instructions that can be recognized by the intelligent reasoning model. These instructions explicitly include key information such as the reasoning topic, core dimensions, judgment criteria, and output format. For example, constraints related to "experimental parameter compliance verification" are transformed into precise instructions such as "reasoning about the risk of parameter deviation in the experimental process, with voltage, current, and temperature as the core dimensions, and outputting a structured result including risk level, deviation value, and rectification suggestions." This clarifies the core objectives and scope of reasoning from the source, completely preventing the reasoning process from deviating from actual business control needs. On the other hand, an efficient knowledge interaction link based on a graph query protocol is established between the module and the power knowledge graph. This link employs batch query and caching optimization mechanisms. This system enhances the efficiency of knowledge acquisition. Through knowledge query and extraction algorithms (such as knowledge walk algorithms based on graph neural networks and rule-based knowledge extraction algorithms), it accurately extracts various types of knowledge relationships in the field of power research from the power knowledge graph. Specifically, this includes the relationships between research entities (such as power equipment-experimental indicators, materials-performance parameters, and researchers-project roles), the temporal relationships of business processes (such as application-review-project approval-implementation-acceptance) and the entry conditions for each stage, the constraint relationships of technical standards (such as experimental parameter thresholds-compliance judgment and achievement indicators-level classification), the citation relationships of research results (such as papers-patents-experimental data), and the requirements for achievement transformation.To enable these knowledge relationships to be directly integrated into the reasoning and computation process of large models, the module uses knowledge embedding techniques (such as classic knowledge embedding algorithms like TransE, DistMult, and RotatE) to transform them into fixed-dimensional (e.g., 768-dimensional or 1024-dimensional) vector-based prior knowledge. This vector-based prior knowledge can be efficiently fused with subsequent feature vectors through element-wise fusion, attention weighting, and other computational operations, providing solid domain-specific knowledge support for the reasoning process and ensuring that the reasoning results strictly comply with the technical specifications and business logic of the power industry.

[0038] After acquiring the reasoning direction and prior knowledge, the model reasoning module strictly follows the hierarchical progression logic of the base layer, adaptive enhancement layer, and reasoning layer. It progressively refines the semantic matrix output by the knowledge extraction module, achieving a layered transformation from general features to domain-specific features, and then to precise decision-making information. Each layer of processing generates standardized data output, providing a high-quality data foundation for subsequent stages. The core function of the base layer is to achieve unified feature fusion and representation of the multimodal semantic matrix. The core execution logic of the model reasoning module at the base layer is as follows: First, it performs modality dimension parsing on the input semantic matrix, identifying the corresponding sub-matrices. Since the input semantic matrix is ​​a structured result of knowledge extraction from multi-source heterogeneous data (text, time series, images, etc.), its essence is an aggregated representation of multimodal data, containing multiple sub-matrices corresponding to different data source types, and each sub-matrix uniquely corresponds to a modality (e.g., text modality, time series modality, image modality). To ensure the accuracy of modality recognition, the module executes a comprehensive parsing process through the modality dimension parsing unit, first performing dimensional splitting and... Feature statistics are performed to extract key statistical attributes (such as the number of dimensions, element sparsity, numerical distribution features, timestamp identifiers, etc.) from each sub-matrix. Then, modality recognition algorithms (such as modality classification models based on feature statistics) are used to match the extracted statistical attributes with preset modality feature templates (such as text modality sub-matrices are usually high-dimensional sparse vectors, time-series modality sub-matrices contain continuous timestamp association features, and image modality sub-matrices are two-dimensional pixel association vectors). This accurately identifies the specific modality corresponding to each sub-matrix and adds a unique modality identifier (such as "text-001", "time-002", "image-003") to each sub-matrix, providing accurate modality differentiation criteria for subsequent modality feature extraction.

[0039] After the general semantic features are generated, the model inference module inputs them into the adaptive enhancement layer. This layer is the key to achieving feature-based scenario adaptation. Its core objective is to generate exclusive semantic features that highly match the power research scenario through precise feature enhancement and redundancy removal. This strengthens information strongly related to power research business in the general semantic features (such as power equipment parameters and experimental compliance indicators) and weakens irrelevant features in the general domain (such as semantic noise in general text), thereby improving the efficiency and accuracy of subsequent inference from the feature level. Specifically, the adaptive enhancement layer first uses prior knowledge obtained from the bidirectional association interface and inference direction instructions as the core basis to construct a two-layer enhancement benchmark: "domain knowledge benchmark - business target benchmark." This forms a dual judgment standard for feature selection and enhancement, ensuring the accuracy of feature processing.The domain knowledge benchmark is constructed based on prior knowledge from the power knowledge graph. K-means clustering is used to cluster prior knowledge vectors to generate domain knowledge feature clusters. These feature clusters cover core knowledge dimensions in the power research field and are used to quantitatively measure the relevance of input features to the power research field; higher relevance indicates greater feature value. The business objective benchmark is constructed based on inference direction instructions. Inference direction instructions are transformed into fixed-dimensional target feature templates (e.g., the "experimental risk warning" target feature template includes key dimensions such as risk indicators and threshold ranges) to measure the adaptability of input features to specific business management objectives; higher adaptability indicates greater contribution of features to inference. Subsequently, the module... A multi-head attention mechanism is used to decouple general semantic features. A two-layer augmentation benchmark is used as the basis for calculating attention weights. The importance of each feature dimension in the general semantic features is scored (range 0-1). Feature dimensions with scores above a preset threshold (e.g., 0.6) are classified as core semantic vectors strongly correlated with the two-layer augmentation benchmark, while those below the threshold are classified as weakly correlated or irrelevant redundant vectors. Vector filtering (retaining core semantic vectors and removing redundant vectors) effectively improves the purity of core semantic vectors and reduces the interference of irrelevant information on subsequent reasoning. Next, a multi-level augmentation network is used to perform hierarchical augmentation processing on the core semantic vectors. This network adopts a progressive residual structure design, with each layer... Each stage introduces corresponding prior knowledge vectors and core semantic vectors for element-wise feature calibration. The domain attributes of prior knowledge are used to strengthen the power research-specific features in the core semantic vectors (such as rated parameter features of specific power equipment, key indicator features of new energy grid connection experiments, and constraint features for safe operation of the power system). Simultaneously, non-domain features are suppressed using activation functions. During the enhancement process, the module performs dual validity checks on the generated scene enhancement features in real time to ensure the reliability of the enhanced features: on the one hand, the semantic fit between the scene enhancement features and the domain knowledge feature clusters in the power knowledge graph is calculated using a cosine similarity algorithm to determine whether the domain relevance of the features meets the standard; on the other hand, semantic... Matching score calculation (such as matching degree calculation based on cross-entropy, cosine distance matching) measures the matching degree between scene enhancement features and business target benchmark, and judges whether the target adaptability of the features meets the standard. If the semantic fit does not reach the first preset threshold (this threshold can be dynamically adjusted according to the business needs of the power research field, such as setting it to 0.8) or the matching degree does not reach the second preset threshold (such as setting it to 0.75), the inter-layer weights and activation function parameters (such as ReLU, GELU, Swish, etc.) of the multi-level enhancement network are dynamically adjusted through gradient descent algorithm (such as Adam optimizer), and the hierarchical enhancement processing is re-executed. When both indicators meet the standard, the scene enhancement features are dimensionally integrated and L2 normalized.Dimension integration maps feature vectors to dimensions consistent with the input requirements of subsequent inference layers (e.g., 1024 dimensions) through a fully connected layer. L2 normalization standardizes the numerical range of feature vectors to the [0,1] interval, avoiding instability in inference model training or deviation in inference results due to excessive numerical differences. Finally, it outputs semantic features specific to power research that accurately match the needs of power research scenarios. These features are the key link between general features and decision information, possessing both domain-specific attributes and precise adaptation to business objectives.

[0040] In some embodiments, the model inference module is configured to: construct a two-layer enhanced baseline based on prior knowledge and inference direction instructions; decouple general semantic features to obtain core semantic vectors; perform hierarchical enhancement processing on the core semantic vectors according to a multi-level enhancement network to obtain scene enhanced features; calculate the semantic fit between the scene enhanced features and the power knowledge graph and the matching degree with the inference direction instructions; if the semantic fit does not reach a first preset threshold or the matching degree does not reach a second preset threshold, dynamically adjust the parameters of the multi-level enhancement network and re-execute the hierarchical enhancement processing; if the semantic fit reaches the first preset threshold and the matching degree reaches the second preset threshold, perform dimensional integration and normalization processing on the scene enhanced features to output power research-specific semantic features.

[0041] In this embodiment of the invention, the core value of the dual-layer enhanced benchmark lies in providing clear and quantifiable judgment criteria for subsequent feature processing, ensuring that feature enhancement always aligns with the characteristics of the power research field and specific business management objectives. Its construction process combines domain expertise with target-specificity. The construction of the domain knowledge benchmark is supported by the multi-type knowledge relationships within the power knowledge graph. First, the knowledge relationships are extracted from the power knowledge graph, encompassing multiple dimensions such as research entity relationships (e.g., power equipment-experimental indicators, materials-performance parameters), business rule constraints (e.g., experimental parameter threshold ranges, achievement acceptance standards), and technical standards and specifications (e.g., new energy grid connection technical requirements). Then, knowledge embedding techniques (e.g., TransE, DistMult algorithms) are used to transform these unstructured knowledge relationships into fixed-dimensional vector-based prior knowledge, ensuring efficient computer processing. Finally, the K-means clustering algorithm is employed to cluster the vector-based prior knowledge, forming multiple domain-specific knowledge feature clusters. Each feature cluster corresponds to a core knowledge dimension in the power research field (e.g., equipment operation monitoring dimension, project compliance review dimension), constituting a core benchmark for measuring the "domain relevance" of features. Only features highly consistent with these feature clusters possess unique value specific to power research.

[0042] The construction of business target benchmarks is closely centered around specific management and control needs. First, the structured field constraints obtained from the data extraction model are analyzed. These constraints are concrete manifestations of the core business objectives of power research. For example, the field constraints corresponding to "compliance management of experimental parameters" include the threshold ranges of core parameters such as voltage and current, and the allowable range of parameter fluctuations. Then, a semantic mapping algorithm (based on the semantic transformation logic of a pre-trained language model) is used to transform these field constraints into standardized target feature templates. These templates explicitly include key information such as core indicator dimensions, judgment logic, and output formats. For example, the target feature template for "experimental risk warning" will clearly specify the risk indicator type, the threshold division for different risk levels, and the presentation format of the warning results, forming a precise basis for measuring the "target suitability" of features. Together, these two elements constitute a two-layer enhanced benchmark, forming a dual judgment dimension of "relevance to the domain" and "target suitability," fundamentally preventing feature processing from deviating from the actual business scenarios of power research.

[0043] Feature decoupling essentially involves "purifying" the general semantic features derived from multimodal fusion. By accurately identifying and separating core semantic vectors from redundant vectors, the relevance and effectiveness of features are improved, eliminating interference for subsequent enhancement processing. This process uses a dual-layer enhancement benchmark as the core judgment criterion and relies on a multi-head attention mechanism to achieve refined feature selection. First, the general semantic features are semantically aligned with the dual-layer enhancement benchmark to construct an attention calculation matrix. In this process, each dimension of the general semantic features undergoes semantic similarity calculation with the feature clusters of the domain knowledge benchmark and the feature templates of the business target benchmark. Then, based on the calculated semantic similarity, dynamic attention weights are assigned to each dimension of the general semantic features, with weight values ​​ranging from 0 to 1. Dimensions with high compatibility with the domain knowledge feature clusters and matching key indicators of the business target feature templates receive higher weights (e.g., above 0.6), while those with lower compatibility receive lower weights. Finally, a weight threshold is set (which can be dynamically adjusted according to the business scenario, such as 0.6), and feature dimensions with weights above the threshold are aggregated into core semantic vectors. These vectors collectively carry key information about power research and development, such as experimental equipment operating parameters, compliance indicators of project application materials, and key points for evaluating the innovativeness of achievements. Simultaneously, redundant vectors with weights below the threshold are filtered out. These redundant vectors are mostly irrelevant information from the general domain (e.g., modal particles in text data, meaningless fluctuations in time-series data). This filtering process significantly improves the purity of the core semantic vectors, reduces the interference of irrelevant information on subsequent reasoning, and lays the foundation for accurate feature enhancement.

[0044] Layered enhancement is a process that uses a multi-tiered enhancement network to gradually transform core semantic vectors into features specific to the particular scenarios of power research. Its core logic is to use prior knowledge for layer-by-layer calibration and enhancement to achieve "scenario-based adaptation" of features. The multi-tiered enhancement network adopts a progressive residual structure design, containing 3-5 enhancement layers (which can be dynamically adjusted according to business complexity). Each layer undertakes a specific enhancement function and introduces corresponding prior knowledge vectors for feature calibration. The first layer is the basic domain feature enhancement layer, which mainly performs element-wise multiplication operations between the core semantic vector and the prior knowledge of the core entities related to power research, strengthening basic association features such as equipment-indicators and project-processes. For example, it highlights the association features between rated parameters and actual operating parameters of power equipment, and the matching features between project application processes and compliance requirements. The middle layer is the subdivided scenario feature enhancement layer, which introduces corresponding business rule constraints and technical standard specifications for specific research scenarios (such as experimental implementation and results acceptance) to perform targeted feature enhancement. For example, in the experimental implementation scenario, it focuses on strengthening the comparison features between experimental parameters and threshold ranges, and the conformity features between equipment operating status and safety specifications. In the results acceptance scenario, it focuses on strengthening the matching features between results indicators and acceptance standards, and the differences between the innovativeness of the results and the current state of the industry. The last layer is the feature integration enhancement layer, which comprehensively integrates the features enhanced in the previous layers, eliminates feature conflicts between layers, and further amplifies the weight of core scenario features.

[0045] In the hierarchical enhancement process, each level outputs intermediate enhanced features, and the residual connection mechanism preserves the effective information of the preceding levels, avoiding feature gradient vanishing and ensuring the stability and effectiveness of the enhancement process. Through this progressive and targeted enhancement approach, the core semantic vector is gradually endowed with attributes specific to the power research field and adaptability to specific scenarios, forming preliminary scenario-enhanced features that provide a foundation for subsequent verification.

[0046] Dual validation and standardization are crucial for ensuring the effectiveness and usability of semantic features specific to power industry research. Through a closed-loop logic of "validity validation - parameter adjustment - standardization processing," the output features are ensured to conform to domain knowledge logic, adapt to specific business objectives, and possess a unified and standardized format. Dual validity validation employs a two-dimensional evaluation system of "semantic fit + matching degree": semantic fit is calculated using a cosine similarity algorithm, performing a dimension-by-dimensional semantic similarity calculation between the scene-enhanced features and the domain knowledge feature clusters in the domain knowledge benchmark to obtain an overall semantic fit score. This score reflects the degree of fit between the features and the knowledge in the power industry research domain; a higher score indicates stronger domain relevance. Matching degree is calculated using a cross-entropy-based semantic matching algorithm, precisely matching the scene-enhanced features with the target feature templates in the business objective benchmark to quantify the adaptability of the features to specific business objectives; a higher matching degree score indicates stronger target targeting.

[0047] The system presets a first threshold (semantic fit threshold, e.g., 0.8) and a second threshold (matching threshold, e.g., 0.75). If the semantic fit of the scene enhancement feature does not reach the first threshold, it indicates insufficient domain relevance of the feature, potentially indicating interference features from non-electrical research fields. If the matching degree does not reach the second threshold, it indicates that the feature is not accurately adapted to the specific business objective and cannot support subsequent inference decisions. In this case, the system activates a dynamic parameter adjustment mechanism, using the Adam optimizer to adaptively adjust the inter-layer weights and activation function parameters (e.g., ReLU, GELU, Swish) of the multi-tiered enhancement network. The adjustment direction is determined based on the verification deviation results. For example, if the semantic fit is insufficient, the weight ratio of the domain knowledge prior vector will be increased; if the matching degree is insufficient, the feature dimension weights corresponding to the core indicators of the business objective template will be strengthened. After adjustment, the hierarchical enhancement process is re-executed until both indicators of the scene enhancement feature meet the standards.

[0048] If the double verification passes, the standardization process begins: First, dimensional integration is performed by mapping the scene enhancement features to a unified dimension (e.g., 1024 dimensions) through a fully connected layer, ensuring that the feature dimensions fully match the input requirements of the subsequent inference layer and avoiding inference errors caused by inconsistent dimensions; then, L2 normalization is performed to standardize the numerical range of the feature vectors to the [0,1] interval, eliminating interference caused by excessive numerical differences between different feature dimensions and ensuring the stability of the inference process; finally, the output semantic features specific to power research not only have distinct attributes of the power research field and adaptability to specific scenarios, but also meet the requirements of the standardized format, and can be directly used as input to the model inference layer, providing high-quality feature support for subsequent accurate decision-making.

[0049] In some embodiments, the model reasoning module is used to: establish a reasoning constraint framework according to the reasoning direction instruction; input the electrical research-specific semantic features into the reasoning layer and generate preliminary reasoning results under the reasoning constraint framework; and verify the preliminary reasoning results to determine decision information.

[0050] In this embodiment of the invention, the reasoning direction instruction is the core basis for constructing the reasoning constraint framework. The business objectives, core dimensions, and judgment criteria it carries are directly transformed into the underlying logical rules of the framework. During the framework construction process, the reasoning direction instruction is first structurally decomposed using a rule parsing engine to extract constraint elements strongly related to power research business. For example, for the reasoning direction of "project application compliance review," core constraints such as material completeness requirements, technical solution feasibility indicators, and applicant qualification conditions are extracted. Subsequently, combining the domain specifications and business process logic in the power knowledge graph, specific judgment rules and logical paradigms are configured for each constraint item, forming a hierarchical constraint system. Taking the "Experimental Risk Warning" reasoning direction as an example, the framework clearly defines the "parameter deviation threshold - risk level corresponding rules" (e.g., voltage parameters deviating from the rated value within 5% is low risk, 5%-15% is medium risk, and above 15% is high risk), "multi-parameter collaborative deviation - risk superposition rules" (e.g., when voltage and current parameters deviate simultaneously, the risk level is automatically increased by one level), and "experimental stage adaptation rules" (e.g., the risk judgment threshold for critical experimental stages is stricter than that for ordinary stages). These rules and paradigms together constitute the boundary constraints of reasoning, ensuring that the reasoning process does not deviate from the actual business needs, while also ensuring the professionalism and compliance of the reasoning results.

[0051] The semantic features specific to power research, used as input data for the inference layer, possess distinct domain attributes and scenario adaptability. Their core semantic information (such as experimental parameter data, project material characteristics, and achievement evaluation indicators) provides a high-quality foundation for accurate inference. Guided by the inference constraint framework, the inference layer conducts deep semantic inference through a Transformer-based inference decoder. The decoder utilizes a self-attention mechanism to capture the inherent relationships and weight proportions of information in each dimension of the specific semantic features. For example, in the "achievement acceptance innovation evaluation" scenario, it automatically strengthens the weights of key dimensions such as the differences between the achievement's technical solution and existing technologies, and patent citation association features. Simultaneously, combined with the pre-set inference rules and logical paradigms within the framework, it performs compliance judgments, risk level assessments, and result quantification scoring on the feature information. During the inference process, the system generates structured preliminary inference results, including not only core conclusions (such as "project application materials are compliant" and "the experiment has a moderate safety risk"), but also detailed supporting information, such as the specific triggering rule clauses, the numerical range of key feature parameters, and comparison results with standard thresholds, ensuring the traceability of the preliminary inference results.

[0052] To ensure the accuracy and reliability of decision-making information, the preliminary reasoning results need to be verified in two dimensions. The first is compliance verification, which compares the preliminary reasoning results with the benchmarks based on the technical standards, industry norms, and business process rules included in the power knowledge graph. For example, it verifies whether the experimental risk level determination complies with the power industry's safe operation specifications and whether the achievement innovation score meets the acceptance standards. If there is a conflict, it is marked as a deviation result. The second is logical verification. On the one hand, it checks the internal logical coherence of the preliminary reasoning results to avoid self-contradictory conclusions such as "the materials are complete but the compliance is not up to standard." On the other hand, it cross-validates the results with historical reasoning data and actual business operation data. For example, it compares the audit results of similar projects and verifies the authenticity of experimental parameter data, eliminating results that are logically contradictory or inconsistent with reality. If the verification finds a deviation, the system will initiate a reverse tracing mechanism to locate the root cause of the deviation (such as insufficient feature enhancement, deviation in constraint rule configuration, etc.), and adjust the feature processing parameters of the adaptive enhancement layer or the rule details of the inference constraint framework, and re-execute the feature enhancement and inference process; if the verification passes, the preliminary inference results will be structured and integrated to form the final decision information containing core content such as control conclusions, risk levels, rectification suggestions, basis clauses, and processing priorities, providing the business control module with accurate and implementable control basis.

[0053] In some embodiments, the knowledge extraction module is configured to: preprocess the acquired multi-source heterogeneous data to obtain data to be extracted; calculate the similarity between the data to be extracted and the power knowledge graph through a cross-layer semantic matching algorithm to determine the core extraction task; activate the scene sub-network and task output head corresponding to the core extraction task in the data extraction model; and input the data to be extracted into the data extraction model to obtain a semantic matrix.

[0054] In the embodiments of the present invention, the multi-source heterogeneous data in the whole process of electric power scientific research business cover various forms such as text types (declaration materials, technical documents), time series types (equipment monitoring data, energy consumption statistics), image types (equipment status images, experimental scene pictures), etc. The data formats are messy, with noise and missing values. Different cleaning and standardization operations need to be implemented for different types of data during preprocessing. For text data, irrelevant information such as punctuation marks, format tags, and redundant spaces are first removed through regular expressions, and then semantic segmentation is performed using word segmentation algorithms (such as jieba segmentation, BERT segmentation). Combined with the stop word dictionary in the electric power field, meaningless words (such as "de", "le") are filtered. Finally, text normalization processing is performed to unify professional terms and abbreviations into standard expressions (such as "new energy grid connection" is uniformly replaced with "new energy grid connection"); for time series data, missing values are filled using linear interpolation or mean filling methods, and abnormal fluctuation data beyond a reasonable range are identified and removed through anomaly detection algorithms such as Isolation Forest and DBSCAN. Then, the data is sorted and segmented in chronological order to ensure coherent time series logic; for image data, algorithms such as Gaussian filtering and median filtering are used to remove noise interference, and the size and pixel range are unified through image scaling and normalization operations. At the same time, key visual features such as image edges and textures are extracted and converted into numerical forms. In addition, metadata tags are added to all data to record information such as data sources (such as "laboratory equipment A", "declaration system"), collection time, and the business processes they belong to (such as "project declaration", "experimental implementation"), and finally a set of data to be extracted with unified structure, complete information, and low noise is formed.

[0055] The core of cross-layer semantic matching is to establish a semantic association between the data to be extracted and the power knowledge graph. By quantifying similarity, the extraction target is clearly defined, avoiding blind extraction. First, a joint semantic vector space is constructed between the core entity layer and the rule logic layer of the power knowledge graph. The core entity layer contains key entities in the power research field (such as power equipment, technical methods, research results, and application entities), while the rule logic layer covers constraint information such as business process specifications, technical standards, and review rules. The joint space uses a unified vector encoding algorithm (such as Word2Vec or BERT) to map the semantic features of entities and rules to the same dimensional space, forming a standardized vector representation system. Then, the preprocessed data to be extracted is input into a pre-trained language model (such as RoBERTa or Electra) to generate data semantic vectors with the same dimension as the joint semantic vector space, achieving semantic alignment between the data and the knowledge graph. Finally, cosine similarity is used to calculate... The method calculates the semantic similarity between the data semantic vector and various scientific research entity nodes in the core entity layer and various business rule nodes in the rule logic layer, respectively, to obtain entity similarity scores and rule similarity scores. Finally, based on preset entity thresholds and rule thresholds (e.g., both set to 0.7), highly similar matching entity nodes and rule nodes are selected. According to the category attributes of these nodes (e.g., entity nodes are "experimental equipment" or "project application") and constraint types (e.g., rule nodes are "compliance review rules" or "parameter verification rules"), the scientific research business scenario (e.g., project application, experiment implementation, and result acceptance) corresponding to the data to be extracted is accurately determined, and the core extraction tasks are clarified (e.g., extracting project application entity qualification information, key experimental parameter data, and matching relationship between results and technical standards).

[0056] The data extraction model incorporates a multi-level adaptation network, including multiple scenario sub-networks for different scientific research business scenarios (such as project application scenario sub-network, experimental monitoring scenario sub-network, and results acceptance scenario sub-network) and task output heads corresponding to various core extraction tasks (such as entity extraction output head, relation extraction output head, and rule matching output head). The activation process requires precise adaptation between the scenario and the task. Based on the scientific research business scenario determined by cross-layer semantic matching, the model locates the corresponding scenario sub-network from the multi-level adaptation network. This sub-network contains a professional feature dictionary strongly related to the scenario (such as the dictionary for experimental monitoring scenarios containing parameter terms such as "voltage," "current," and "temperature") and extraction templates (such as the entity extraction template containing the format "[device name]-[parameter type]-[value]"). After activation, these dedicated resources are automatically loaded to support targeted extraction. At the same time, based on the type of core extraction task (such as entity extraction, relation extraction, and rule matching), the corresponding task output head is located and activated synchronously. The feature mapping parameters and result output format of the output head are dynamically configured according to the task type. For example, the entity extraction output header is configured with entity category label mapping parameters, and the output format is key-value pairs of "entity name-entity category"; the relation extraction output header is configured with relation type mapping parameters, and the output format is triples of "entity1-relation type-entity2"; the rule matching output header is configured with rule matching degree calculation parameters, and the output format is structured data of "data fragment-matching rule-matching score", ensuring that the extraction results meet the requirements of subsequent semantic matrix construction.

[0057] After the data to be extracted is adapted to the scene subnetwork and the task output head, the input data extraction model performs deep knowledge extraction and structured transformation. The model first strengthens the input data by adding domain labels to key information in the data using the scene subnetwork's professional feature dictionary (e.g., labeling "voltage 380V" as "[parameter]-voltage-[value]380V"), generating domain-labeled strengthened data. Subsequently, based on the extraction templates built into the scene subnetwork, targeted knowledge extraction is performed on the strengthened data. For example, entity extraction templates are used to extract research entities (e.g., equipment name, technical terms, application entity), relationship extraction templates are used to extract relationships between entities (e.g., "equipment A-monitoring-voltage parameter" "application entity B-application-project C"), and rule matching templates are used to extract matching information between data and business rules (e.g., "experimental data D-compliant-standard E" "application material F-missing-supporting material G"), forming initial knowledge containing research entities, entity relationships, and business rule matching information. Knowledge fragments are identified; to ensure the accuracy and standardization of knowledge fragments, they are fed back to the power knowledge graph for semantic verification and association completion. For example, the knowledge graph verifies the legality of entities (such as confirming whether the extracted equipment name is a standard equipment in the power field) and completes the implicit relationships between entities (such as completing the relationship "Project C - Field - New Energy" based on the knowledge graph), ultimately obtaining a standardized knowledge set. Finally, each knowledge fragment in the standardized knowledge set is converted into a semantic vector with consistent dimensions using vector encoding algorithms (such as FastText and GloVe). The semantic vectors are arranged in an ordered manner according to the logical order of "entity vector - relation vector - rule matching vector". Then, the vector sequence is transformed into a structured semantic matrix through matrix integration operations. This matrix can completely preserve the semantic association and hierarchical structure of knowledge, providing high-quality input for subsequent model inference.

[0058] In some embodiments, the knowledge extraction module is used to: construct a joint semantic vector space of the core entity layer and the rule logic layer in the power knowledge graph; map the data to be extracted to the joint semantic vector space to obtain data semantic vectors; calculate the semantic similarity between the data semantic vectors and various scientific research entity nodes in the core entity layer, and various business rule nodes in the rule logic layer; based on preset entity thresholds and rule thresholds, filter out entity nodes and rule nodes with high similarity matching; and determine the scientific research business scenario corresponding to the data to be extracted and the core extraction task to be executed according to the category attributes of the selected entity nodes and the constraint type of the rule nodes.

[0059] In this embodiment of the invention, constructing a joint semantic vector space of the core entity layer and the rule logic layer in the power knowledge graph is the foundation for achieving accurate scenario matching and task localization of the data to be extracted. Its core objective is to break down the semantic separation between the two layers and form a unified semantic representation dimension. Specifically, the core entity layer covers key entity nodes in the field of power research, including research equipment, experimental indicators, material samples, researchers, project stages, etc. Each entity node is accompanied by attribute information (such as equipment model, indicator threshold, project stage characteristics, etc.). The rule logic layer contains constraint rule nodes for the entire research business process, such as application and review rules, experimental operation specifications, result acceptance standards, parameter compliance constraints, etc. Each rule node is associated with semantic information such as triggering conditions, constraint objects, and execution requirements. During the construction process, a unified knowledge embedding algorithm (such as TransE, RotatE, etc.) is first used to vectorize the entity nodes of the core entity layer and the rule nodes of the rule logic layer, respectively, to generate initial entity vectors and initial rule vectors. Then, a cross-layer semantic association enhancement mechanism is introduced. By calculating the potential semantic associations between entity nodes and rule nodes (such as the association between the "experimental equipment" entity and the "equipment operation specification" rule, and the association between the "experimental index" entity and the "parameter compliance constraint" rule), the initial vectors are iteratively optimized to make the entity vectors and rule vectors have the same semantic dimension. Finally, through a vector space alignment algorithm, the optimized entity vector set and the rule vector set are merged to form a unified joint semantic vector space, ensuring that any vector in this space can be directly compared for semantic similarity. Mapping the data to be extracted to a joint semantic vector space to obtain the data semantic vector is a key step in establishing the semantic association between data and knowledge graphs. The core of this step is to achieve standardized semantic representation of unstructured / semi-structured data. The data to be extracted covers multi-source heterogeneous data from the entire process of power research and business, such as experimental record texts, equipment monitoring logs, project application materials, simulation reports, etc. First, preprocessing is required, including text cleaning (removing invalid characters and standardizing formats) and data structuring transformation (decomposing unstructured text into phrase-level semantic units and converting time-series data into feature sequences). Then, a semantic encoding model with the same origin as the knowledge graph node encoding is used to transform the preprocessed data to be extracted into initial data vectors, ensuring that the initial data vectors have the same dimension as the joint semantic vector space. To improve mapping accuracy, a domain adaptation enhancement strategy is introduced, using pre-trained corpus from the power research domain to fine-tune the encoding model, strengthening the model's semantic understanding of domain-specific terms and business expressions. Finally, through the vector mapping calibration module, the initial data vectors are projected onto the constructed joint semantic vector space to obtain the final data semantic vector. This vector accurately carries the core semantic information of the data to be extracted and can be directly semantically associated with entity nodes and rule node vectors within the space. Calculating the semantic similarity between the data semantic vector and various scientific research entity nodes in the core entity layer, as well as various business rule nodes in the rule logic layer, is a crucial step in quantifying the strength of the association between data and the core elements of the knowledge graph, providing a quantitative basis for subsequent node selection and matching. The similarity calculation employs the cosine similarity algorithm, which accurately characterizes the semantic fit of vectors by measuring the cosine of the angle between two vectors; the closer the cosine value is to 1, the higher the semantic similarity. During the calculation, for the core entity layer, the cosine similarity between the data semantic vector and all entity node vectors under that category is calculated in batches according to entity category (e.g., equipment, indicators, project stages), yielding the similarity distribution between the data and various entities. Similarly, for the rule logic layer, the cosine similarity between the data semantic vector and all rule node vectors under the corresponding rule type is calculated in batches according to rule type (e.g., review, operation, compliance), yielding the similarity distribution between the data and various rules. Meanwhile, to improve computational efficiency, a vector index optimization mechanism (such as KD tree, Ball Tree, etc.) is introduced to index the set of entity node vectors and the set of rule node vectors, enabling fast retrieval and calculation of similarity, and avoiding the inefficiency caused by full traversal. Based on preset entity and rule thresholds, selecting highly similar entity and rule nodes is a crucial step in achieving accurate semantic matching of data. The core principle lies in eliminating weakly related nodes and retaining core related information through threshold settings. Both entity and rule thresholds are preset empirical values ​​based on the business characteristics of the power research field and can be dynamically adjusted according to actual extraction accuracy requirements. The entity threshold is used to determine the effectiveness of the association between the data semantic vector and research entity nodes, while the rule threshold is used to determine the effectiveness of the association between the data and business rule nodes. During the selection process, firstly, the similarity results between the data and various entity nodes are compared with the entity thresholds, retaining entity nodes with similarity higher than the thresholds to form a high-similarity entity set. This set reflects the core research entities involved in the data to be extracted. Subsequently, the similarity results between the data and various rule nodes are compared with the rule thresholds, retaining rule nodes with similarity higher than the thresholds to form a high-similarity rule set. This set reflects the core business rules associated with the data to be extracted. Simultaneously, to address the possibility of multiple categories of entities and multiple types of rules coexisting after selection, the similarity ranking of various nodes is recorded to provide a priority basis for subsequent scenario and task determination. The ultimate goal of this process is to determine the scientific research business scenario corresponding to the data to be extracted and the core extraction tasks to be performed based on the category attributes of the selected entity nodes and the constraint types of the rule nodes. This achieves the transformation from semantic matching to business implementation. Specifically, the category attributes of entity nodes directly point to the business object dimension associated with the data. For example, entity nodes containing "experimental equipment" or "experimental indicators" indicate that the data is associated with the experimental implementation stage, while entities containing "project application materials" or "application review" point to the project application stage. The constraint types of rule nodes further refine the business scenario. For example, the "parameter compliance constraint" rule corresponds to the experimental parameter monitoring scenario, and the "application material completeness review" rule corresponds to the project application review scenario. By integrating the category attributes of a set of highly similar entities with the constraint types of a set of highly similar rules, the specific scientific research business scenario corresponding to the data to be extracted can be accurately located. Based on this, the core extraction tasks are determined by combining the core needs of the business scenario. For example, if the scenario is "experimental implementation parameter monitoring", the core extraction task is to extract information such as the experimental equipment model, real-time parameter values, and parameter deviations from the data. If the scenario is "project application review", the core extraction task is to extract key information such as the project name, applicant unit, core points of the technical solution, and list of required supporting materials from the application materials, ensuring that the extraction tasks are accurately matched with the business needs.

[0060] In some embodiments, the knowledge extraction module is used to: locate and activate the scene sub-network corresponding to the scientific research business scenario from the multi-level adaptation network of the data extraction model according to the determined scientific research business scenario and core extraction task, load the professional feature dictionary and extraction template built into the scene sub-network; synchronously locate and activate the task output head corresponding to the core extraction task, and configure the feature mapping parameters and result output format of the output head.

[0061] In this embodiment of the invention, completing the targeted adaptation configuration of the data extraction model based on the determined scientific research business scenario and core extraction task is a prerequisite for achieving accurate data extraction. The core logic lies in activating the target sub-module in the multi-level adaptation network of the model to achieve precise matching between the model's capabilities and business requirements. The multi-level adaptation network is one of the core architectures of the data extraction model. It adopts a hierarchical design of "scenario classification - task subdivision," modularly encapsulating the business scenarios and corresponding extraction tasks of the entire power research process. Each business scenario corresponds to an independent scenario sub-network, and each core extraction task corresponds to a dedicated task output head. This architectural design enables the model to quickly adapt to different scenarios and tasks. The first step in the adaptation configuration is to locate and activate the scenario sub-network in the multi-level adaptation network based on the determined scientific research business scenario. This process is fundamental to enabling the model to possess scenario-specific extraction capabilities. The localization process of the scene sub-network relies on the scene index mapping mechanism built into the multi-level adaptation network. This mechanism pre-establishes a mapping relationship library of "business scene tags - sub-network address indexes," where business scene tags accurately correspond to various sub-scenarios in power research (such as experimental parameter monitoring scenarios, project application review scenarios, and achievement acceptance and evaluation scenarios), and sub-network address indexes point to the physical storage address of the corresponding scene sub-network in the model storage architecture. During localization, the identified research business scenarios are first converted into standardized scene tags. A tag matching algorithm is then used to quickly compare the tags with the mapping relationship library to accurately locate the storage address of the corresponding scene sub-network. Subsequently, the sub-network activation mechanism is activated. The model controller sends an activation command to the target sub-network, waking up the core components of the sub-network, such as the feature extraction unit and semantic encoding unit, switching them from a dormant state to an active state, preparing them for subsequent professional feature extraction and data parsing. This process ensures that the model only activates modules related to the current scene, avoiding irrelevant modules from occupying computing resources and improving model operating efficiency. Simultaneously with the activation of the scene sub-network, the professional feature dictionary and extraction template built into this sub-network are loaded. This is the core step in enabling the model to possess scene-specific semantic understanding and extraction logic. The professional feature dictionary is a domain-specific resource built for the corresponding scientific research business scenario, covering core terms, entity names, attribute keywords, etc. in that scenario. For example, the dictionary for the experimental parameter monitoring scenario includes various power equipment models, experimental indicator terms, parameter units, etc., while the dictionary for the project application review scenario includes application material-related terms, approval keywords, etc. During the loading process, the feature dictionary is transformed into a vector representation that the model can recognize through the dictionary parser and integrated into the semantic encoding unit of the scene sub-network, improving the model's ability to recognize and encode scene-specific semantics. The extraction template is a structured parsing rule preset based on the core extraction requirements of the scenario. It defines the location features, contextual features, etc., of the extraction targets such as entities and relationships. After loading, it will be embedded as a constraint condition into the feature extraction unit, guiding the model to accurately locate the extraction targets according to the scene-specific logic. Based on the adaptation of the scene sub-network, the task output head corresponding to the core extraction task is simultaneously located and activated, achieving precise alignment between the model's extraction capabilities and specific task requirements. The task output head is the core module in the data extraction model responsible for result generation and format adaptation. Different core extraction tasks (such as entity extraction, relation extraction, attribute extraction, etc.) correspond to independent output heads, each with built-in feature mapping logic and result assembly rules matching the task. The localization process also relies on an index mapping mechanism. The model pre-establishes a "task type - output head index" association library, quickly matching the corresponding task output head by transforming the current core extraction task into a standardized task type. The activation process activates the feature mapping unit and result formatting unit of the output head through the model parameter controller, enabling it to receive scene sub-network output features and complete the extraction result transformation, while ensuring smooth data transmission between the output head and the scene sub-network, achieving efficient feature transfer. Finally, configure the feature mapping parameters and result output format of the task output head to ensure that the extraction results meet the usage requirements of business management. Configuring the feature mapping parameters is crucial for the output head to accurately parse the output features of the scene sub-network. This parameter defines the mapping relationship between feature vectors and extraction targets (such as entity categories and relationship types). During configuration, the model will call preset basic parameters from the parameter configuration library based on the needs of the core extraction task and dynamically fine-tune them in conjunction with the professional feature dictionary for the current scene to ensure that the mapping relationship accurately matches the extraction targets in the scene. Configuring the result output format is to improve the usability of the extraction results. According to the data format requirements of the business management module, a structured form for the output results (such as JSON or XML format) is set, specifying the fields included in the extraction results (such as entity name, category, confidence level, extraction location, etc.). A format converter assembles the raw results extracted by the model according to the set format, ultimately outputting standardized extraction results, providing high-quality data input for subsequent knowledge extraction and inference processes.

[0062] In some embodiments, the knowledge extraction module is configured to: input preprocessed data to be extracted into a data extraction model to generate enhanced data with domain labels; perform targeted knowledge extraction on the enhanced data based on the extraction template built into the scene sub-network to obtain initial knowledge fragments containing scientific research entities, entity relationships, and business rule matching information; feed the initial knowledge fragments back to the power knowledge graph to obtain a standardized knowledge set; convert the standardized knowledge set into semantic vectors with consistent dimensions, and perform ordered arrangement and matrix integration of the semantic vectors to obtain a semantic matrix.

[0063] In this embodiment of the invention, after completing the scenario-task-oriented adaptation configuration of the data extraction model, the preprocessed data to be extracted is input into the model to generate domain-labeled enhanced data. This is a fundamental step in realizing domain-specific knowledge extraction, and the core objective is to enhance the domain semantic representation of the data through the model's scenario adaptation capability. The preprocessed data to be extracted has undergone format standardization and noise removal, possessing a unified input form. After being input into the model, it first enters the semantic encoding unit of the activated scenario sub-network. This unit integrates scenario-specific professional feature dictionary vectors and performs deep semantic encoding on the input data through a domain-adapted encoding model. This not only captures the general semantic information of the data but also accurately identifies scenario-specific terms, business expressions, and other core information. Simultaneously, based on the domain labeling system built into the scenario sub-network, the model adds corresponding domain labels to the encoded semantic features, such as "experimental equipment," "parameter threshold," and "application materials," which are specific to power research, forming domain-labeled enhanced data. These domain labels provide semantic anchors for the subsequent extraction process, effectively improving the positioning accuracy of the extraction target and avoiding interference from general semantics. Based on the extraction template built into the scenario sub-network, targeted knowledge extraction is performed on the reinforced data to obtain initial knowledge fragments containing scientific research entities, entity relationships, and business rule matching information. This is the core extraction step in the process, realizing the transformation from reinforced data to structured knowledge fragments. The extraction template, as a scenario-specific structured parsing rule, predefines the contextual features, positional constraints, and label matching logic for different extraction targets. For example, the template for the experimental parameter monitoring scenario clearly defines the entity relationship extraction logic of "equipment name - parameter type - parameter value," while the template for the project application review scenario defines the extraction rule of "application material type - required elements - missing status." During the extraction process, the model's feature extraction unit, constrained by the extraction template, accurately matches the domain labels and semantic features of the reinforced data. It first locates scientific research entities that meet the template constraints (such as power equipment, experimental indicators, application materials, etc.), then mines the relationships between entities (such as "equipment - monitoring indicators," "materials - review elements," etc.) through relationship extraction algorithms, and simultaneously matches the corresponding business rule node information in the rule logic layer, ultimately integrating them to form the initial knowledge fragments. The initial knowledge fragments have a preliminary structured form, containing the core information of the extracted target and corresponding domain labels, laying the foundation for subsequent standardization processing. Feeding initial knowledge fragments into the power knowledge graph to obtain a standardized knowledge set is a crucial step in ensuring the uniformity and standardization of knowledge. The core logic relies on the authoritative semantic system of the knowledge graph to calibrate and complete the initial extraction results. During the feedback process, semantic connections are first established between the initial knowledge fragments and the core entity layer and rule logic layer of the power knowledge graph. Vector similarity matching aligns the research entities and business rules in the fragments with the standard nodes in the knowledge graph. For successfully matched knowledge fragments, missing attribute information is supplemented according to the standardized attribute system of the knowledge graph, unifying the expression standards for entity names and relationship types. For knowledge fragments with semantic ambiguity or partial deficiencies, the associative reasoning capabilities of the knowledge graph are used for completion; for example, the "standard parameter range" information is supplemented based on the association relationships of the "experimental equipment" entity node. Abnormal knowledge fragments that cannot be matched are marked as pending verification and removed to ensure the reliability of the standardized knowledge set. The final output standardized knowledge set possesses unified semantic specifications and structural form, eliminating redundancy and biases in the initial knowledge fragments. Converting standardized knowledge sets into semantic vectors with consistent dimensions is a prerequisite for achieving structured knowledge representation and subsequent reasoning computation. The core objective is to transform discrete standardized knowledge into quantifiable and computable vector forms. The conversion process employs a knowledge embedding algorithm that shares the same semantic vector space as the power knowledge graph, ensuring that the generated semantic vectors possess the same semantic dimension and can be directly correlated with knowledge graph node vectors for computation. First, each knowledge unit in the standardized knowledge set (a single research entity, a set of entity relationships, or a rule matching information) is independently vectorized to generate initial knowledge vectors. Then, through vector normalization, all initial knowledge vectors are uniformly mapped to a vector space of the same dimension (e.g., 768-dimensional or 1024-dimensional), resulting in semantic vectors with consistent dimensions. This step eliminates the differences in vector dimensions between different types of knowledge units, providing a unified data foundation for subsequent matrix integration. The final step in the process is to arrange and matrix-integrate semantic vectors to obtain a semantic matrix. This transforms scattered semantic vectors into a structured matrix, providing standardized input for the model inference module. The arrangement process follows pre-defined ordering rules, designed in conjunction with the business logic of power research and subsequent inference requirements. For example, these rules may prioritize knowledge types according to "research entity - entity relationship - business rule," or sort according to the chronological order of the business processes corresponding to knowledge units, ensuring that the order of vectors in the matrix has a clear business logic connection. After sorting, a matrix integration algorithm combines all semantic vectors with consistent dimensions according to their arrangement order to form a two-dimensional semantic matrix. The row vectors of the matrix correspond to the semantic vectors of individual knowledge units, while the column vectors correspond to different dimensions of semantic features. The final semantic matrix not only retains the core semantic information of each knowledge unit but also strengthens the logical connections between knowledge through structured arrangement. It can be directly parsed by the model inference module, providing high-quality structured input for subsequent multimodal feature fusion, domain feature enhancement, and other inference processes.

[0064] It should be noted that the bidirectional association interface adopts a software interface form, with built-in data integrity verification and exception retransmission mechanisms, enabling synchronous interaction with the data extraction model and the power knowledge graph. Its bidirectional data flow logic is as follows: the structured field constraints of the data extraction model and the knowledge associations of the power knowledge graph are synchronously transmitted to the interface. After the interface completes the format adaptation of the constraints and knowledge through a semantic conversion engine, it is pushed to the intelligent inference model in real time, while simultaneously receiving status feedback information from the large model, forming a closed-loop interaction of "extraction-adaptation-transmission-feedback". The interface establishes logical connections with other modules through memory sharing, ensuring low latency and high reliability of data flow, while also supporting dynamic adjustment of the interaction frequency according to business complexity.

[0065] In the model inference module, the base layer uses the mutual information entropy algorithm to calculate the association weights of the initial feature vectors of different modalities. The semantic association strength between vectors is quantified by a formula, and multimodal feature weighting fusion is completed according to the weight ratio. The multi-level enhancement network of the adaptive enhancement layer contains five progressive layers: feature decoupling layer, domain enhancement layer, scene adaptation layer, feature integration layer, and verification layer. Each layer is configured with a 3×3 convolutional kernel and a ReLU activation function. The parameters are dynamically adjusted, including the inter-layer learning rate (initial value 0.001) and the attention weight allocation ratio. The inference constraint framework of the inference layer adopts a "rule base + logic engine" architecture. The rule base stores the inference rules in the form of structured tables. The logic engine parses the rule associations through a forward inference algorithm. The initial inference result verification includes three core indicators: logical consistency, data integrity, and domain compliance. When the verification fails, a reverse tracing mechanism is triggered, and the parameters of the adaptive enhancement layer are adjusted by a gradient descent step size of 0.0005.

[0066] In the preprocessing stage of the knowledge extraction module, text data is cleaned using a stop word dictionary specific to the power industry. The anomaly detection threshold for time-series data is set to mean ± 3 times the standard deviation. Image data is normalized using the formula (pixel value / 255). The cross-layer semantic matching algorithm uses an improved cosine similarity algorithm based on graph neural networks. The input is the semantic vector of the data to be extracted and the knowledge graph node vector. The number of iterations is set to 100, and the output similarity score is retained to 4 decimal places. The multi-level adaptation network is divided into a scene layer, a task layer, and an execution layer. The scene layer and the task layer are associated one-to-one or one-to-many through a mapping relationship table. For example, the "experimental parameter monitoring scene" corresponds to the two types of task output heads: "entity extraction" and "relationship extraction". The feature mapping parameters of the output heads are dynamically adjusted according to the feature distribution of the power research field. The output format is uniformly JSON, which includes 8 core fields such as entity name, category, and confidence.

[0067] In the above embodiments, the descriptions of each embodiment have their own emphasis. Parts not detailed or described in a particular embodiment can be referred to in the relevant descriptions of other embodiments. Unless otherwise specified or in conflict with logic, the terminology and / or descriptions between different embodiments are consistent and can be referenced interchangeably. Technical features in different embodiments can be combined to form new embodiments based on their inherent logical relationships.

[0068] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A scientific research business management and control system based on large-scale model technology, characterized in that, include: The data acquisition module is used to acquire multi-source heterogeneous data throughout the entire process of power research and development. The knowledge extraction module is used to extract knowledge from the multi-source heterogeneous data based on the power knowledge graph and the data extraction model to obtain a semantic matrix. The model reasoning module is used to perform knowledge reasoning on the semantic matrix based on the intelligent reasoning big model to obtain decision information; wherein, the intelligent reasoning big model is a four-level structure consisting of a base layer, an adaptive enhancement layer, a reasoning layer, and a bidirectional association interface; the intelligent reasoning big model establishes a deep association with the data extraction model and the power knowledge graph based on the bidirectional association interface; the bidirectional association interface determines the reasoning direction based on the structured field constraints of the data extraction model and determines the prior knowledge of the intelligent reasoning big model based on the power knowledge graph; The business management module is used to manage the scientific research business based on the decision information.

2. The scientific research business management and control system based on large model technology according to claim 1, characterized in that, The model inference module is used for: Based on the bidirectional association interface, the structured field constraints of the data extraction model are extracted as the reasoning direction, and the knowledge associations of the power knowledge graph are extracted as prior knowledge. The semantic matrix is ​​input into the base layer to obtain the general semantic features fused from the multimodal model; The general semantic features are input into the adaptive enhancement layer of the intelligent reasoning model to obtain the semantic features specific to power research. The specific semantic features for power research are input into the reasoning layer of the intelligent reasoning model to obtain decision information.

3. The scientific research business management and control system based on large model technology according to claim 2, characterized in that, The model inference module is used for: Establish a data interaction link with the data extraction model; The data extraction model is analyzed by its built-in structured field configuration specific to power research, and field constraints that represent the core business objectives of power research are filtered out. The field constraints are converted into inference direction instructions that the model can recognize; Establish knowledge interaction links with the power knowledge graph; Through knowledge query and extraction algorithms, we extract various types of knowledge relationships in the field of power research from the power knowledge graph. The knowledge relationships are transformed into prior knowledge in vector form using knowledge embedding technology.

4. The scientific research business management and control system based on large model technology according to claim 2, characterized in that, The model inference module is used for: Modality dimension parsing is performed on the input semantic matrix to identify the submatrices corresponding to the semantic matrix; where each submatrix corresponds to a modality. For sub-matrices of different modes, the feature extraction unit of the corresponding mode is used to perform initial feature extraction to obtain the initial feature vector corresponding to each mode; The association weights between the initial feature vectors of different modalities are calculated, and weighted fusion is performed based on the association weights to obtain general semantic features.

5. The scientific research business management and control system based on large model technology according to claim 4, characterized in that, The model inference module is used for: Based on the prior knowledge and reasoning direction instructions, a two-layer enhanced benchmark is constructed; The general semantic features are decoupled to obtain the core semantic vector; The core semantic vector is subjected to hierarchical enhancement processing based on a multi-cascaded enhancement network to obtain scene enhancement features; Calculate the semantic fit between the scene enhancement features and the power knowledge graph, and the matching degree with the reasoning direction instructions. If the semantic fit does not reach the first preset threshold or the matching degree does not reach the second preset threshold, then dynamically adjust the parameters of the multi-cascade enhancement network and re-execute the hierarchical enhancement process. If the semantic fit reaches the first preset threshold and the matching degree reaches the second preset threshold, then the scene enhancement features are subjected to dimensional integration and normalization processing to output the power research-specific semantic features.

6. The scientific research business management and control system based on large model technology according to claim 2, characterized in that, The model inference module is used for: Establish a reasoning constraint framework based on the reasoning direction instructions; The semantic features specific to power research are input into the inference layer, and preliminary inference results are generated under the inference constraint framework. The preliminary reasoning results are verified to determine the decision information.

7. The scientific research business management and control system based on large model technology according to claim 1, characterized in that, The knowledge extraction module is used for: The acquired heterogeneous data from multiple sources is preprocessed to obtain the data to be extracted. The similarity between the data to be extracted and the power knowledge graph is calculated using a cross-layer semantic matching algorithm to determine the core extraction task. Activate the scene sub-network and task output head in the data extraction model that correspond to the core extraction task; The data to be extracted is input into the data extraction model to obtain the semantic matrix.

8. The scientific research business management and control system based on large model technology according to claim 7, characterized in that, The knowledge extraction module is used for: Construct a joint semantic vector space for the core entity layer and the rule logic layer in the power knowledge graph; The data to be extracted is mapped to the joint semantic vector space to obtain the data semantic vector; Calculate the semantic similarity between the data semantic vector and various scientific research entity nodes in the core entity layer, and between the data semantic vector and various business rule nodes in the rule logic layer. Based on preset entity thresholds and rule thresholds, entity nodes and rule nodes with high similarity matching are selected. Based on the category attributes of the selected entity nodes and the constraint types of the rule nodes, the scientific research business scenario corresponding to the data to be extracted and the core extraction tasks to be performed are determined.

9. The scientific research business management and control system based on large model technology according to claim 7, characterized in that, The knowledge extraction module is used for: Based on the determined scientific research business scenario and core extraction task, locate and activate the scenario sub-network corresponding to the scientific research business scenario from the multi-level adaptation network of the data extraction model, and load the professional feature dictionary and extraction template built into the scenario sub-network. Synchronously locate and activate the task output head corresponding to the core extraction task, and configure the feature mapping parameters and result output format of the output head.

10. The scientific research business management and control system based on large model technology according to claim 9, characterized in that, The knowledge extraction module is used for: The preprocessed data to be extracted is input into the data extraction model to generate enhanced data with domain labels; Based on the extraction template built into the scenario sub-network, targeted knowledge extraction is performed on the enhanced data to obtain initial knowledge fragments containing scientific research entities, entity relationships, and business rule matching information. The initial knowledge fragments are fed back into the power knowledge graph to obtain a standardized knowledge set; The standardized knowledge set is converted into semantic vectors with consistent dimensions, and the semantic vectors are arranged in an ordered manner and integrated into a matrix to obtain a semantic matrix.

Images