Multi-agent cooperation-based multi-module local linkage reconstruction method and system

Abstract

The present application relates to artificial intelligence and education management information technology field, specifically, it relates to a kind of multi-module local linkage reconstruction method and system based on multi-agent cooperation.It includes the following steps: talent training scheme is decoupled into independent business module by text analysis engine and industrial reference data is extracted;Then execute cross-dimension feature matching, generate local reconstruction instruction;Based on instruction execution node feature reconstruction, and automatically address associated module execution includes credit optimization, timing drift correction and physical space avoidance cascading linkage correction;Finally, execute multi-dimensional boundary consistency state check and assemble output reconstruction document.The present application uses time series penalty operator and peak avoidance algorithm, effectively solves the teaching logic inversion and physical resource conflict problem in reconstruction process, ensures that the reconstruction scheme is self-consistent in logic, space-time dimension, significantly improves the engineering executability and landing accuracy of scheme.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and educational management information technology, specifically to a multi-module local linkage reconstruction method and system based on multi-agent collaboration. Background Technology

[0002] The vocational talent training program is a legally binding document for vocational schools to standardize the teaching process, organize teaching operations, and implement quality monitoring. Its revision process needs to dynamically respond to the shifts and reshaping of the core competency map of positions under the evolution of regional industries. With the increasing complexity of professional cluster construction, the revision of the program has changed from a global reorganization to a local dynamic reconstruction based on specific industry data.

[0003] However, existing talent training program development systems suffer from technical flaws when dealing with parameter mapping and logical linkage between multiple modules, including a crude underlying mapping mechanism and a lack of temporal constraints. On the one hand, when reconstructing a curriculum system based on competency gaps, existing technologies generally adopt a discrete matching and linear parameter superposition mode under static rules. Because the system fails to quantitatively assess the implicit knowledge graph overlap between the new candidate courses and the existing curriculum system, it cannot remove redundant competency parameters. This reconstruction method, which lacks a marginal incremental evaluation mechanism, will cause the parameter aggregation results of the course dimension to exceed the rigid threshold boundaries of total class hours and total credits set by the national teaching standards, making the reconstruction scheme lose its engineering feasibility. On the other hand, when triggering cascading updates between modules, existing systems lack spatiotemporal logical topology control. For course nodes that are locally updated or introduced, default isolated timestamps are assigned only based on the traversal order, failing to establish a cross-node prerequisite dependency matrix from a global perspective. This time-reversed reconstruction output not only disrupts the gradual learning and cognitive patterns but also directly leads to serious logical conflicts and scheduling failures in the physical scheduling phase of the academic affairs system for the derived teaching arrangement modules. Therefore, this paper proposes a multi-module local linkage reconstruction method and system based on multi-agent collaboration. Summary of the Invention

[0004] The purpose of this invention is to provide a multi-module local linkage reconstruction method and system based on multi-agent collaboration, so as to solve the problem of credit parameter overflow caused by the lack of marginal incremental evaluation when reconstructing talent training programs in the prior art, as well as the problem of teaching logic inversion and physical resource conflict caused by the lack of spatiotemporal topological constraints.

[0005] To achieve the above objectives, the present invention aims to provide a multi-module local linkage reconstruction method based on multi-agent collaboration, applied to a talent training program intelligent revision system. This intelligent revision system is configured with interconnected talent training program revision agents, regional industry analysis agents, and talent demand analysis agents. The multi-module local linkage reconstruction method includes the following steps: S1, obtaining the initial talent training program to be revised through the talent training program revision agent, and calling a preset text parsing engine to decouple the initial talent training program according to preset business logic boundaries. Each is an independent business module, among which... ; Simultaneously, establish data communication links with the regional industry analysis intelligence and talent demand analysis intelligence to extract external industry dynamic reference data; S2. Perform cross-dimensional feature matching between each of the decoupled independent business modules and the external dynamic reference data; when the matching degree of any independent business module is lower than the preset compliance feature threshold, mark it as the target business module and generate the corresponding local reconstruction instruction; S3. Based on the local reconstruction instruction, perform node feature reconstruction within the boundary of the target business module to generate updated feature nodes; Based on a preset module logic topology map, the system automatically addresses related business modules affected by the state change of the target business module; the updated feature node is transmitted to the related business module through the coupling bus, triggering the linkage correction of the underlying support parameters in the related business module. S4. Perform consistency state verification on each independent business module that has completed partial reconstruction and linkage correction; based on the verification confirmation signal of the verification terminal for the feature reconstruction node, perform structured assembly on each independent business module that is in closed-loop steady state, and generate and output reconstruction document.

[0006] As a further improvement to this technical solution, in S1, the preset text parsing engine is a multimodal document parsing model that includes a visual typography feature extraction network and a semantic sequence recognition network; the preset text parsing engine is called to decouple it according to the business logic boundary. Each independent business module includes the following steps: The physical coordinate layout features of the initial talent training program document are extracted through the visual layout feature extraction network, wherein the visual layout feature extraction network is constructed based on a convolutional neural network and is used to identify and extract text box coordinates, font size and paragraph spacing parameters. Simultaneously, semantic features of the text sequence are extracted through the semantic sequence recognition network, wherein the semantic sequence recognition network is constructed based on a bidirectional encoder representation model and is used to perform vectorized semantic encoding on the document text stream; Multimodal feature alignment and concatenation are performed on the physical coordinate layout features and the text sequence semantic features to construct a document logical structure tree; The node distribution of the document logical structure tree is compared with the preset standard training scheme topology map to identify and locate the first-level title feature anchor points corresponding to each business logic boundary; The document logical structure tree is subjected to continuous segmentation and isolation along the physical coordinate axes of the first-level heading feature anchor points, and each independent branch node obtained after segmentation is structurally encapsulated into a structure carrying a unique logical identifier. Each is an independent business module.

[0007] As a further improvement to this technical solution, in step S2, cross-dimensional feature matching is performed between the decoupled independent business modules and the external dynamic reference data to calculate the feature matching degree of any independent business module. The specific steps involved are as follows: Extract the semantic distribution feature vector of the target independent business module, and the industry capability node feature vector in the external dynamic reference data that is mapped and associated with the target independent business module; A semantic tolerance matrix is ​​introduced to calculate the effective spatial cosine similarity between the semantic distribution feature vector and the industry capability node feature vector; Extract the attribute slot parameters of the independent business module and compare them with the preset rigid constraint threshold boundary. When an out-of-bounds error occurs, generate an asymmetric compliance penalty operator. Extract the skill word frequency popularity weight and ability life cycle decay rate embedded in the external dynamic reference data; The effective spatial cosine similarity, the skill word frequency popularity weight, and the ability life cycle decay rate are weighted and summed, and the summation result is subjected to dimensionality reduction multiplication mapping with the asymmetric compliance penalty operator to output the feature matching degree. When the feature matching degree is lower than the preset compliance feature threshold, the feature difference that maps the feature matching degree shortfall is extracted, encapsulated as a capability gap indicator vector, the attribute slot parameter is extracted as a rigid constraint parameter, and the corresponding atomic operation operator is configured based on the feature difference. The logical identifier of the target independent business module, the capability gap indication vector, the rigid constraint parameter, and the atomic operation operator are structurally encapsulated into a local reconstruction instruction, and then transmitted in a directional manner through the system's underlying communication bus.

[0008] As a further improvement to this technical solution, the specific steps involved in generating and updating feature nodes in step S3 are as follows: The partial reconstruction instructions are parsed, and logical identifiers are extracted to lock the physical memory boundaries of the target business module; Based on the atomic operation operators in the local reconstruction instruction, the capability gap indicator vector is incrementally injected into the feature distribution matrix of the target business module, and weight reallocation is performed under the constraint of rigid constraint parameters. The cryptographic hash function is invoked to concatenate the underlying logic execution operator with the original industry data index pointer representing the basis for modification, generating a unique traceability hash label; The feature distribution matrix after weight redistribution is instantiated to generate an updated feature node containing the reconstructed semantic feature distribution, capability attribute labels, and the source hash labels.

[0009] As a further improvement to this technical solution, the specific steps involved in triggering the linkage correction of the underlying support parameters in the associated business module are as follows: Based on a preset module logical topology map, the logical dependency path starting from the target business module is analyzed, and the affected related business modules are located along the logical dependency path. The related business modules include at least: a course setting module, a teaching execution module, and a teaching facility module. The updated feature node is transmitted to the associated service module via the coupling bus in a cascading manner to trigger a linkage correction of the underlying support parameters based on the updated feature node. The underlying support parameter linkage correction specifically refers to the cascading mapping correction performed along the logical dependency path, which includes: Incremental optimization is performed on the credit capacity boundary of the aforementioned course setting module; Perform sequence drift correction on the temporal topology of the teaching execution module; Peak shifting and avoidance are implemented for the physical space load of the teaching facility module.

[0010] As a further improvement to this technical solution, the credit capacity boundary of the course setting module is incrementally optimized, involving the following specific steps: Extract the capability attribute labels and capability gap indication vectors embedded in the updated feature nodes of the pass-through, and combine them with the rigid constraint parameters defined by the local reconstruction instruction to generate a candidate course set; The underlying teaching elements of the candidate course set are analyzed, including course content, course objectives, and course knowledge points; the capability gap indicator vector is mapped to industry capability requirements, and the industry capability requirements are matched with the course content, course objectives, and course knowledge points of each candidate course using multi-dimensional feature matching. Based on the above feature matching results, retrieve the initial capability coverage of each candidate course relative to the missing capability node mapped by the capability gap indicator vector. By comparing the candidate course set with the existing course set in the course setting module, the overlap features of knowledge points between the two are extracted, and a capability overlap attenuation coefficient is constructed to characterize the scale of knowledge redundancy. The initial capability coverage is nonlinearly attenuated using the capability overlap attenuation coefficient, and the initial capability coverage is corrected to an effective marginal capability increment. Using the maximization of the effective marginal capability increment as the objective function and the class hour allocation boundary as the constraint, a knapsack optimization calculation is performed to select target courses and generate the optimal class hour allocation scheme. Based on the preset class hour and credit conversion rules, the credit allocation vector of the course setting module is refreshed.

[0011] As a further improvement to this technical solution, the temporal topology relationship of the teaching execution module is corrected for order drift, and the specific steps involved are as follows: The underlying subject progression relationships formed after the introduction of the target course are analyzed, and a prerequisite time dependency matrix reflecting the graph structure relationship between courses is constructed. Extract the default scheduling timestamp assigned to the target course in the initial reconstruction state; The prerequisite time-series dependency matrix is ​​invoked to perform a topological sorting constraint check on the default scheduling timestamps; When the verification result does not meet the preset minimum teaching cycle span, a time series penalty operator is introduced to perform a sequence drift correction on the default scheduling timestamp until the one-way teaching logic constraint is met, and the teaching process time sequence vector is output.

[0012] As a further improvement to this technical solution, the physical space load of the teaching facility module is subject to peak shifting and avoidance, and the specific steps involved are as follows: Extract the training site requirements parameters for each time slice in the reconstructed teaching process time-series vector; By comparing the real-time occupancy feature matrix of the physical resources preset in the teaching facility module, the reusability conflict rate of the physical space in each target time period is calculated. Determine whether the conflict rate of the physical space reuse exceeds a preset physical load threshold; If the conflict rate exceeds the threshold, an avoidance step size parameter is introduced for the course nodes that cause conflict in the time sequence vector of the teaching process, and a second peak shifting correction iteration is performed until the conflict rate of the theoretical and practical space reuse in all time slices converges to within the preset physical load threshold.

[0013] As a further improvement to this technical solution, in step S4, a consistency state check is performed on each independent business module that has completed partial reconstruction and linkage correction. The specific steps involved include: Extract the final feature data from the underlying cache of each independent business module. The final feature data includes at least: the updated credit allocation vector, the teaching process time sequence vector, and the theory-practice space reuse conflict rate within each time slice. The preset multi-dimensional boundary verification engine is invoked to perform cross-module closed-loop logic verification based on the final feature data. The closed-loop logic verification includes at least: overflow verification for data logic capacity boundaries, deadlock prevention verification for the underlying time-series scheduling network, and security load verification for physical resources. If the closed-loop logic verification passes, the underlying state flag of each independent business module is set to the closed-loop steady state. At the same time, the final feature data and the traceability hash label are packaged to generate a feature reconstruction node, which is then pushed to the verification terminal.

[0014] On the other hand, the present invention provides a multi-module local linkage reconstruction system based on multi-agent collaboration, including a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor executes the computer program to implement the multi-module local linkage reconstruction method based on multi-agent collaboration described above.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. In this multi-module local linkage reconstruction method and system based on multi-agent collaboration, by introducing a capability overlap attenuation coefficient and a nonlinear attenuation model, the knowledge redundancy between candidate courses and the existing system can be quantitatively evaluated; by using the corrected effective marginal capability increment to perform knapsack optimization calculation, it is ensured that while filling capability gaps, redundant weights are automatically stripped away, so that the reconstructed credit vector is strictly limited to the rigid threshold of the national standard.

[0016] 2. In this multi-module local linkage reconstruction method and system based on multi-agent collaboration, the abstract subject progression relationship is transformed into a computer-recognizable graph structure constraint by constructing a prerequisite time-series dependency matrix; and by using a time-series penalty operator to perform unidirectional drift correction, it can automatically identify and repair the time axis inversion problem caused by node insertion, ensuring that the teaching process fully conforms to the educational psychology logic of "pre-requisite and post-requisite".

[0017] 3. In this multi-module local linkage reconstruction method and system based on multi-agent collaboration, the overload risk of the physical site can be predicted in advance by calculating the conflict rate of the physical space reuse and comparing it with the physical load threshold in real time; by using the avoidance step size parameter to perform a second peak shift correction iteration, the overload demand is slidably released to the low time slice, which completely solves the physical collision problem that may occur in the actual scheduling stage after the scheme is reconstructed. Attached Figure Description

[0018] Figure 1 This is a flowchart of the overall method of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Example 1: Please refer to Figure 1 As shown, this embodiment provides a multi-module local linkage reconstruction method based on multi-agent collaboration, applied to a talent training program intelligent revision system. This system is configured with interconnected talent training program revision agents, regional industry analysis agents, and talent demand analysis agents. The multi-module local linkage reconstruction method includes the following steps: S1. Obtain the initial talent development plan to be revised through the talent development plan revision agent, and call the preset text parsing engine to decouple the initial talent development plan according to the preset business logic boundaries. Each is an independent business module, among which... ; Simultaneously, establish data communication links with the regional industry analysis intelligence and talent demand analysis intelligence to extract external industry dynamic reference data; The intelligent revision system for talent training programs of the present invention is based on a multi-agent collaborative architecture, specifically including a talent training program revision agent, a regional industry analysis agent, and a talent demand analysis agent; Specifically, the intelligent agent for revising talent training programs extracts job requirements, work tasks, and professional competency elements by connecting with recruitment platforms, industry associations, and enterprise human resource databases. In this invention, the intelligent agent is responsible for constructing a three-level topological map of "job-task-competency" to provide micro-level technical skill gap criteria for program reconstruction.

[0021] The regional industry analysis intelligent agent is used to monitor and collect data on industrial policies, economic structure changes, and emerging industry orientation within the target region in real time. In this invention, the intelligent agent is responsible for transforming unstructured macro-industry information into calculable industry background parameters, providing a macro-level compliance and forward-looking basis for revisions.

[0022] The talent demand analysis intelligent agent integrates document parsing, feature comparison, linkage reconstruction, and compliance detection algorithms. In this invention, the intelligent agent is responsible for coordinating the output data of the above two intelligent agents (talent training program revision intelligent agent and regional industry analysis intelligent agent) to perform atomic decoupling, local reconstruction, and final solution assembly and output of the initial document.

[0023] In this embodiment, a multi-source authoritative knowledge base is also configured as the prior benchmark data base for system operation. The multi-source authoritative knowledge base non-volatilely stores the national vocational education teaching standards database, the childcare industry policy and regulation database, and the guiding documents issued by the education administrative departments of various provinces and cities. The pre-defined standard training program topology graph is a structured logical skeleton extracted from standardized text in a multi-source authoritative knowledge base after entity extraction and relation alignment. It uses a directed acyclic graph (DAG) as the data structure for storage and computation. The specific parameter features of this topology graph include node semantic feature parameters, hierarchical topological constraint logic, and attribute slot definitions. Among them, the semantic feature parameters of the nodes are: each semantic node (such as the "Course Setting" node) is embedded with a semantic distribution feature vector composed of a set of core keywords and their preset weights, forming the "semantic fingerprint" of the module. Hierarchical topological constraint logic: defined by directed edges The unidirectional dependencies and depth offset relationships at each core level (e.g., the "Training Objectives" node must be the predecessor node of the "Curriculum Setting" node) are used to correct logical overlaps in the parsing process. Attribute slot definition: Configure a preset attribute slot (such as "Total Learning Hours: Numeric") for each node to enforce the integrity of the decoupled content.

[0024] Furthermore, in order to achieve accurate decoupling of the initial talent training program document containing complex tables, multi-level headings, and mixed layouts, the talent training program revision agent performs feature extraction by calling a preset text parsing engine; wherein, the preset text parsing engine is a multimodal document parsing model that includes a visual layout feature extraction network and a semantic sequence recognition network. The visual typography feature extraction network performs pixel-level scanning of each page of the initial talent training program document (converted into image tensors). In specific implementation, the visual typography feature extraction network preferably uses a deep residual network (ResNet-50) combined with a feature pyramid network (FPN) to construct the backbone, uses a region proposal network (RPN) to generate candidate text boxes, and uses a bounding box regression algorithm to extract physical coordinate typography features including text box coordinates (absolute pixel coordinates of the top left and bottom right corners), font size (estimated values ​​based on text box height and pixel density), and paragraph spacing parameters (difference in Y-axis coordinates between adjacent text boxes). Simultaneously, the original character stream data is extracted through a semantic sequence recognition network. The semantic sequence recognition network preferably adopts a bidirectional encoder representation model based on the Transformer architecture (such as the BERT or LayoutLM text branch based on the Transformer architecture). It uses its self-attention mechanism to perform high-dimensional vectorized semantic encoding on the character stream and extracts text sequence semantic features containing contextual dependencies. After extracting the above bimodal features (physical coordinate layout features and text sequence semantic features), the parsing engine performs multimodal feature alignment and splicing on the physical coordinate layout features and the text sequence semantic features in the tensor dimension, so that each text paragraph not only has a high-dimensional semantic vector, but is also bound to a unique page physical coordinate; based on this, a document logical structure tree that maps the hierarchical dependency relationship of the entire text paragraphs is constructed.

[0025] Subsequently, a graph matching algorithm (preferably a graph edit distance (GED) based topological similarity algorithm) is invoked to extract the hierarchical depth and parent-child relationship vectors between nodes in the document logical structure tree, and to calculate the minimum node insertion, deletion, and feature substitution cost function required to map the graph topological structure of the document logical structure tree into the topological graph of the standard training scheme. The node distribution of the document logical structure tree is then compared with the aforementioned topological graph of the standard training scheme to identify graph structure features. Specifically, when a node in the document's logical structure tree is identified, and the joint probability distribution of its physical layout features (such as "bold, centered, size 2") and semantic features (such as high-confidence mapping of benchmark words like "course setting") highly matches a certain level anchor point in the standard training program's topology map, the physical absolute coordinates of that node are determined and located as a first-level heading feature anchor point; the horizontal section of the page where the first-level heading feature anchor point is located (i.e., the Y-axis coordinate section) is mapped and established as the business logic boundary.

[0026] After establishing the boundaries, the document logical structure tree is subjected to continuous data flow segmentation and isolation along the physical coordinate axes of each of the first-level heading feature anchor points.

[0027] To ensure that the subsequent "partial linkage reconstruction" has an absolutely atomic operational basis, the various text and table data blocks obtained from the segmentation are forcibly mapped and encapsulated as... (of which) This is an independent business module. Specifically, it maps to: career orientation, training objectives, training specifications, curriculum design, teaching arrangements, faculty, teaching facilities, teaching resources, teaching plans, teaching evaluation, quality management, graduation requirements, and final program presentation.

[0028] Each encapsulated independent business module is automatically assigned a globally unique logical identifier (UUID) and a microsecond-level version control timestamp by the system, and serialized into an atomic format (such as a lightweight JSON node) and stored in a distributed cache. This physical encapsulation mechanism reconstructs the original highly coupled long document of dozens of pages into a discrete business container that supports independent memory addressing, completely eliminating the risk of memory overwriting during subsequent multi-agent concurrent reconstruction.

[0029] Furthermore, while completing the decoupling of static documents, the talent training program revision agent establishes a data communication link with the regional industry analysis agent and the talent demand analysis agent through the configured API data access mechanism, and extracts external industry dynamic reference data (specifically, using a structured data stream push or on-demand query mechanism, directly retrieving formatted feature vector sequences from the knowledge graph library of the upstream agent through a preset data interface protocol (such as RESTful API or gRPC-based message mechanism). The extracted external dynamic reference data includes, but is not limited to: three-level topology map parameters of job-task-ability (including node identification features and associated weight parameters), regional skill popularity vector (including skill word frequency popularity value), ability life cycle decay features (including technology iteration rate parameters), and data timeliness verification tags; Furthermore, to prevent the introduction of outdated industry data from causing revision distortion, this embodiment introduces a timeliness verification mechanism at the underlying data extraction layer; that is, when receiving external dynamic reference data, the timeliness verification tag bound to its metadata layer (including the accurate timestamp of data generation and the confidence matrix of data source) is forcibly extracted. During the link handshake, the timeliness verification tag is compared with the current server timestamp in real time. When the calculated time difference exceeds the limit, i.e., it exceeds the system's preset freshness threshold (such as 30 calendar days), the system automatically intercepts the batch of expired data and triggers a cross-end API incremental update instruction to the upstream intelligent agent to force refresh the underlying industry data base. By introducing the timeliness parameter, dynamic interception and correction of static and lagging data are realized.

[0030] S2. Perform cross-dimensional feature matching between each of the decoupled independent business modules and the external dynamic reference data; when the matching degree of any independent business module is lower than the preset compliance feature threshold, mark it as the target business module and generate the corresponding local reconstruction instruction; In this embodiment, the talent training program revision agent first extracts the semantic distribution feature vector of the underlying atomic independent business module, denoted as... Simultaneously extract external dynamic reference data and lock the mapping associated with the current target independent business module. Each industry capability node feature vector is denoted as... (in ); Due to the "false alarms of gaps" that may be caused by differences in literal expression, this embodiment introduces a semantic tolerance matrix to calculate the effective spatial cosine similarity between the semantic distribution feature vector and the industry capability node feature vector; Furthermore, extract the attribute slot parameters of the independent business module (denoted as...). (e.g., the total credit value of the current module), and compare it with the national teaching standard feature database preset in a multi-source authoritative knowledge base (wherein, the national teaching standard feature database has a preset rigid constraint threshold boundary). When an out-of-bounds error occurs, an asymmetric compliance penalty operator is generated to represent the veto state: when ; when ; In the formula, This represents the system's preset minimum penalty constant, and its value range is... ; This represents the absolute lower limit of the feature library. This represents the absolute upper limit of the feature library's preset values; This represents the real-time statistical values ​​of the attribute slots of the target independent business module in the current refactoring state. Specifically, when the parameters of the business module meet the national teaching standards (i.e., are within the rigid constraint threshold boundaries), the operator is assigned a constant value of 1, which does not have any gain effect on the subsequent matching results; however, once the parameters exceed the limits (such as the total credits exceeding the absolute upper limit), the operator is assigned a value of 1. Or, core competency indicators are below the absolute lower limit The operator will instantaneously change to a minimum value. .

[0031] Based on this, the asymmetric compliance penalty operator As a dimensionality reduction multiplier, it performs a product mapping with the effective spatial cosine similarity calculated based on the semantic tolerance matrix, outputting the final feature matching degree. : ; In the formula, This represents the frequency and popularity weight of skill terms, and its value range is... The value is determined by the normalized frequency of occurrence of the target skill node in the regional recruitment data in the external dynamic reference data, which is used to quantify the real-time urgent need for specific abilities. This represents the capability lifecycle degradation rate, and its value range is... The value is determined by a time-negative decay coefficient based on the technology iteration cycle, which is used to perform dimensionality reduction on technological features that are on the verge of being phased out. This represents the semantic tolerance matrix operation term.

[0032] Specifically, when hour: ; when hour; ; In the formula, This represents the preset allowable deviation subspace envelope threshold, whose value is determined based on the maximum spatial discrete radius of the synonym vector cluster in the natural language processing model (the preferred value range is...). ); The initial cosine similarity scalar is calculated between the semantic distribution feature vector and the industry capability node feature vector in standard high-dimensional Euclidean space. This represents the transpose matrix operation; This represents the preset semantic tolerance matrix. Specifically, let the semantic distribution feature vector be... 3D column vectors, then the semantic tolerance matrix The physical data structure is explicitly defined as having dimensions of It is a symmetric positive definite matrix, and is configured with the bias weights of the synonym vector space; The semantic tolerance matrix represents... The weighted spatial deviation distance calculated based on the metric benchmark; Specifically, by introducing a semantic tolerance matrix This makes the constraint terms This mechanism constructs a Mahalanobis distance metric for two semantic vectors within a specific industry context. This not only ensures that the calculated spatial deviation distance remains a non-negative real number, but also effectively counteracts the scaling and relevance interference of synonyms in the multidimensional feature space through the cross-weighting of off-diagonal elements, thus building a rigorous allowable deviation subspace. Notably, by calculating the high-dimensional cosine similarity between the semantic distribution feature vector and the industry capability node feature vector, and inputting it into a predefined semantic tolerance matrix for envelope comparison, the system determines that the vectors are synonymous heterogeneous mappings when their corners fall within the allowable deviation subspace defined by this matrix, retaining their valid similarity values. This mechanism leverages the continuity of mathematical space to accurately absorb differences in the natural language expressions of the same skill node from different institutions or industries (e.g., recognizing "infant psychological assessment" as "children's mental observation"), thereby correcting and filtering false feature gaps caused by literal inconsistencies.

[0033] Furthermore, the feature matching degree is calculated. Then, it is compared with the preset compliance feature threshold. Numerical comparisons are performed, where the preset compliance feature thresholds are dynamically distributed from the national teaching standard feature database linked to the system's underlying layer. Specifically, these thresholds are mapped to the minimum competency coverage rate required to meet national professional teaching standards, and their preferred value range is [missing information]. (In this embodiment, a scalar value of 0.80 is preferred.) When the system recognizes At that time, the talent training program revision agent marks the independent business module as the target business module and triggers the generation of the corresponding local reconstruction instruction; to ensure that the local linkage reconstruction in the subsequent step S3 can be implemented with absolutely atomic operation, the local reconstruction instruction is distributed through the system's internal event bus and is structured and encapsulated into a low-level data packet containing the following four core dimensions: Logical Identifier: Directly encapsulates the unique logical identifier (UUID) of the target business module in the distributed cache, enabling absolute addressing of the reconstructed target and blocking memory overwriting of other compliant modules; Capability Gap Indicator Vector: The system backtracks the aforementioned semantic tolerance matrix calculation steps to extract the capability gap indicator vector. Feature vectors of uncovered industry capability nodes (denoted as capability gap subsets) This indicates that after semantic tolerance matrix envelope filtering and multimodal representation difference elimination, the current target independent business module still has a set of core industry skill nodes with substantial technical gaps at its underlying level, and performs vector subtraction operation with the semantic distribution feature vector of the original module (i.e., The obtained mathematical residuals It is encapsulated as a capability gap indicator vector, providing precise target completion for downstream applications; Rigid constraint parameters: Synchronously extract the attribute slot capacity of the target business module (e.g., based on...) Subtract the used credit hours to calculate the remaining total available credits and write the instruction to prevent global logical parameters from overflowing after local node reconstruction; Atomic operators: based on The mathematical feature vector configuration corresponds to machine-readable opcodes, including the INSERT operator representing incremental injection of gap features, or the UPDATE operator representing dimensionality reduction coverage of old features.

[0034] S3. Based on the local reconstruction instruction, perform node feature reconstruction within the boundary of the target business module to generate updated feature nodes. The specific steps involved are as follows: The partial reconstruction instructions are parsed, and logical identifiers are extracted to lock the physical memory boundaries of the target business module; Based on the atomic operation operators in the local reconstruction instruction, the capability gap indicator vector is incrementally injected into the feature distribution matrix of the target business module, and weight reallocation is performed under the constraint of rigid constraint parameters. The cryptographic hash function is invoked to concatenate the underlying logic execution operator with the original industry data index pointer representing the basis for modification, generating a unique traceability hash label; Instantiate the feature distribution matrix after weight redistribution to generate an updated feature node containing the reconstructed semantic feature distribution, capability attribute labels, and the source hash labels; Based on a preset module logic topology map, the system automatically addresses related business modules affected by the state change of the target business module; the updated feature node is transmitted to the related business module through the coupling bus, triggering the linkage correction of the underlying support parameters in the related business module.

[0035] In this embodiment, for the target business module, the received partial reconstruction instruction is parsed. First, the logical identifier (UUID) in the address packet header of the instruction is extracted. The physical memory boundary of the target business module is accurately located and locked in the distributed cache. (Specifically, the physical memory boundary refers to the protected address space or memory partition allocated by the system for a specific business module. The "partial reconstruction instruction" generated in step S2 contains the logical identifier (UUID), which is mapped to the corresponding starting memory address and offset in the distributed cache (such as Redis or in-memory database). The purpose of setting the "boundary" is to ensure that the write operation is limited to the address segment of the target module through the memory isolation mechanism when performing reconstruction, thereby preventing "adjacent overflow" or "illegal memory tampering" and ensuring the absolute atomicity of the reconstruction operation.)

[0036] Based on the atomic operation operators carried in the local reconstruction instruction, the capability gap indication vector obtained in step S2 is... The feature distribution matrix at the bottom layer of the target business module is written using incremental injection. During this process, to prevent the total module weight (mapped to total credits or total class hours) from exceeding the absolute upper limit of the rigid constraint parameter due to incremental injection... Under the constraints of rigid parameters defined by the local reconstruction instruction (such as the legal credit limit stipulated in the teaching syllabus), a redistribution algorithm (such as the normalized weight redistribution algorithm) is performed on the local weights in the matrix to obtain the new weights after redistribution. This is used to accurately compensate for the capacity gap while automatically reducing the weight redundancy of non-core features through negative feedback adjustment, ensuring that the reconstructed feature distribution matrix is ​​always constrained by the boundary conditions of the national teaching standards in terms of logical summation, and avoiding logical collapse caused by parameter overflow. Simultaneously, to establish a tamper-proof data lineage record, the system invokes a cryptographic hash function (such as the SHA-256 algorithm) to generate a traceability marker; specifically, it executes the underlying logic operator that performs the modification action. The original industry data index pointer on which the characterization is modified. Perform low-level byte stream concatenation to generate a unique traceability hash label. : ; In the formula, The underlying logic execution operator records exactly "what actions" the agent took in response to the talent development program; It serves as an index pointer for original industry data, recording where the "objective basis (evidence)" for the intelligent agent to perform the aforementioned actions lies; The feature distribution matrix after weight redistribution is instantiated to generate an updated feature node containing the reconstructed semantic feature distribution, the newly added capability attribute label, and the source hash label. The updated feature node is structured and encapsulated into an information carrier containing the following three dimensions: the reconstructed semantic feature distribution (representing the high-dimensional feature vector matrix after capability gap injection), the newly added capability attribute label (representing the class hours, credit value, and theory-practice ratio parameter after being updated due to rigid constraint parameters), and the source hash label (a security authentication packet used for data lineage verification).

[0037] Furthermore, after generating the updated feature nodes, a preset module logical topology graph is loaded; specifically, the preset module logical topology graph is represented as a directed graph model in the underlying data structure. And persistently store it as an adjacency list or adjacency matrix; where the vertex set The edge set represents the discrete, independent business module identifiers in the talent development program. This represents the strong coupling and dependency relationships between modules formed by the teaching logic (for example, changes in "curriculum design" will inevitably lead to changes in "teaching execution"). Based on the logical topology graph of the module, a graph traversal algorithm (preferably a depth-first search (DFS) algorithm or a topology sorting algorithm) is used to recursively explore along the out-degree direction of each vertex, starting with the identifier of the target business module as the root node, and to resolve one or more logical dependency paths.

[0038] Subsequently, the associated business modules affected downstream are located along the logical dependency path. Specifically, when the graph traversal algorithm detects a specific affected vertex on the path, the module identifier (UUID) encapsulated in that vertex is extracted. The addressing engine inputs the module identifier into the microservice registry or addressing routing table at the system's underlying layer for hash mapping to obtain the physical memory base address or listening port of the associated business module bound to the identifier, thereby completing the absolute addressing and location at the computer execution level.

[0039] In this embodiment, the affected related business modules include, in order of hierarchy, at least: a course setting module, a teaching execution module, and a teaching facility module; Subsequently, the updated feature node is transmitted to the aforementioned associated service module via the internal coupling bus to trigger a linked correction of the underlying support parameters based on the updated feature node. The specific correction steps are as follows: Incremental optimization is performed on the credit capacity boundary of the aforementioned course setting module: After receiving the pass-through, the capability attribute labels embedded in the updated feature nodes and the capability gap indicator vector are extracted, and combined with the rigid constraint parameters (such as the remaining total credits) defined by the local reconstruction instruction, a candidate course set that meets the preliminary conditions is generated. .

[0040] The underlying teaching elements of the candidate course set are analyzed, and the underlying teaching elements are specifically decoupled into course content (representing the teaching carrier), course objectives (representing the teaching orientation), and course knowledge points (representing micro-skills); the extracted capability gap indicator vector is mapped to the actual industry capability requirements, and the industry capability requirements are matched with the course content, course objectives, and course knowledge points of the candidate courses through multi-dimensional feature mapping and matching. Based on the above multi-dimensional feature matching results, the initial capability coverage of each candidate course relative to the missing capability node mapped by the capability gap indicator vector is retrieved. ; Retrieve the initial capability coverage of each candidate course relative to the missing capability node mapped by the capability gap indicator vector. To address the knowledge redundancy caused by excessive course content, the candidate course sets were compared. With the existing course set within the course setting module We extract the set of overlapping knowledge points from the underlying graphs of both systems, and based on this, we construct a capability overlap attenuation coefficient to characterize the scale of knowledge redundancy. : ; In the formula, Represents a set of knowledge point nodes; The cardinality of the elements in a set is represented by the total number of unique physical knowledge point nodes within the set. This is the ability overlap attenuation coefficient, used to quantify the overlap rate of knowledge points between candidate courses and existing courses. A higher value indicates higher knowledge redundancy, and its range is [value range missing]. ; Indicates the candidate course set The knowledge point graph node set of the underlying mapping; Indicates existing course sets The knowledge point graph node set of the underlying mapping; Furthermore, utilizing the capacity overlap attenuation coefficient The initial capacity coverage is nonlinearly decayed, and the effective marginal capacity increment is obtained by correction. : ; In the formula, The redundancy penalty adjustment factor configured for the system is used to control the sensitivity of knowledge redundancy to the negative penalty generated by capability increments (i.e., the steepness of the exponential decay curve), and its value is preferably limited to... Within the range, when facing the scenario of restructuring a standard higher vocational education system, its optimal value range is: ; The effective marginal capacity increment is used to represent the net remaining effective capacity coverage of the candidate courses to the system after deducting duplicate teaching knowledge points. It represents the base (constant) of the natural logarithm, used to construct a smooth nonlinear exponential decay model and avoid rigid step abrupt changes at the penalty boundary; Ultimately, to maximize The system performs dynamic programming knapsack optimization calculation for the objective function. Since credits are essentially standardized indicators derived from class hour allocation by each school, this system introduces class hour allocation parameters as underlying constraints in the optimization calculation. Within the range of rigid constraint parameters, the optimal target course is selected, and the credit allocation vector of the course setting module is refreshed.

[0041] Specifically, the knapsack optimization calculation is performed with the objective function of maximizing the effective marginal capacity increment, which includes: For the candidate course set The system's underlying configuration corresponds to the Boolean decision variables for each optional course. (in The course was selected to represent the course. (This indicates that the course has been dropped).

[0042] The objective function constructed is: ; The corresponding rigid boundary constraint equations are: ; In the formula, Indicates the first The effective marginal capacity increment of each alternative course after attenuation calculation; Indicates the first The default number of class hours allocated to each of the optional courses; This indicates the current remaining total available class hours for the target business module. Indicates the first The pre-set standard credit values ​​for each elective course; This represents the rigid constraint parameter, i.e., the current remaining available total credit limit boundary of the target business module; This represents the total number of candidate target course nodes contained in the candidate course set, and its underlying data type is a positive integer; This represents the traversal index number in the candidate course set, used to locate a specific candidate course; Represented as the first Discrete Boolean decision variables for the configuration of optional courses.

[0043] The system calls an optimization algorithm to solve the objective function that satisfies the above constraint equations, and outputs the result that... The optimal decision vector that reaches the global maximum and the corresponding The system identifies the candidate courses as the target courses, generates an optimal class hour allocation scheme, and then calls the preset class hour to credit conversion rules (e.g., 16 standard class hours correspond to 1 credit) to convert the class hour allocation scheme into credit values. Finally, it refreshes the credit allocation vector in the course setting module and sets the corresponding credit values. The alternative courses are identified as target courses, and the credit allocation vector of the course setting module is refreshed.

[0044] Timing topology order drift correction for the instruction execution module: To address the potential inversion of prerequisite and subsequent requirements after incorporating the target course, this paper analyzes the underlying disciplinary progression relationships formed by introducing the target course and constructs a prerequisite time dependency matrix that reflects the graph structure relationships between courses. Its dimension is an asymmetric binary matrix, This indicates the total number of course nodes covered in the talent development program after partial restructuring; Let the row index and column index of the matrix be the course index. and courses (in ); if the underlying requirements are courses Prior to the course Teaching, then matrix elements If the course and courses If there are no strong prerequisite requirements between courses, or if parallel teaching is allowed, then the corresponding matrix element should be set to a low level. ; It is worth noting that this pre-prepared time dependency matrix The reflexivity and asymmetry constraints must be satisfied, meaning that for any node, there must be a diagonal element. , and if The system will then forcibly lock the reverse parity bit. ; Extract the default scheduling timestamp assigned to the target course in the initial reconstruction state. And call the prerequisite time-series dependency matrix to set the default scheduling timestamps. Perform topological sorting constraint verification; during this verification process, a preset minimum teaching cycle span is introduced. The preset minimum teaching cycle span The minimum atomic time interval used to characterize the time interval required to ensure that two highly dependent courses are not scheduled in parallel (its value is determined according to the academic administration standards, preferably one standard semester or one teaching month).

[0045] By verifying the formula Perform logical judgment; if the verification result does not meet the preset minimum teaching cycle span... When a time series inversion occurs (i.e., a time series penalty operator is introduced), the time series penalty operator is introduced. Perform a sequence drift correction on the default scheduling timestamps; specifically, the time series penalty operator is a forced advancement function based on the difference: ; In the formula, This indicates the courses that are prerequisites in the current state. The timestamp of the course schedule (such as the absolute teaching week sequence number or semester sequence number); This indicates the target course that will be used as a subsequent revision node in the current reconstruction state. The assigned default time stamp for the course scheduling; This represents the preset minimum teaching cycle span, which is the minimum atomic distance between two strongly dependent courses on the timeline allowed by the system, serving as a rigid lower bound for timeline verification. This indicates that after the unidirectional drift simulation, the target course... The final updated valid class schedule; This represents the anti-collision relaxation variable, and its preferred value range is... When the system has sufficient computing resources and the scheduling density is low, the preferred method is... To ensure a compact course schedule; when a statutory long holiday or concurrent training tasks are detected within the target time period, heuristic rules are used to assign... A step size compensation value greater than 0; This is a time series penalty operator, which is a one-way advancement function of the underlying time axis. During each topology sort constraint check, the underlying scheduler calls this operator to apply the passed-in default scheduling timestamps. The specific algorithm expression for the evaluation is as follows: ; Among them, when the computer determines When this occurs, it indicates that the current default scheduling timestamp not only satisfies the prerequisite constraint but also has a safety margin. At this time, the operator is in a silent state. Original value output of function No write operation memory overhead is generated at the underlying level; When the computer determines When (i.e., when a time reversal occurs or the safety distance is compressed to below the limit), the operator is activated instantaneously; Functions capture absolutely safe lower bounds and change the original default timestamp The value is forcibly overwritten to this lower bound. Through the unidirectional push operation of this arithmetic logic unit (ALU), the system completely eliminates the risk of deadlock loops in the topology network.

[0046] Furthermore, this time series penalty operator is called iteratively. All nodes that have been inverted are shifted backward along the timeline until the global network satisfies the unidirectional teaching logic constraint (i.e., a directed acyclic graph is formed), and the teaching process time sequence vector is output accordingly. .

[0047] Peak shifting and load avoidance are implemented for the physical space load of the aforementioned teaching facility modules: Extract each time slice from the reconstructed time-series vector of the teaching process. Training site requirements parameters This is then compared with the pre-set real-time physical resource occupancy feature matrix in the teaching facility module (which includes the maximum environmental capacity limit of the site). Compare the results; calculate the conflict rate of theoretical and practical space reuse within each target time period. : ; In the formula, The discrete time slice in the time-series vector representing the teaching process is matched with the minimum scheduling cycle of the scheduling system (such as a specific teaching week or teaching day) at the system level, and serves as a dynamic index scalar on the time axis. Indicates a slice at a specific time. The parameter for the required training space within the system. In the underlying calculation, it is specifically mapped to the total number of course nodes that simultaneously request to call the target training space (such as a specific type of virtual simulation training room), or the total number of physical workstations required, representing the real-time concurrent request volume of the system; The preset maximum environmental capacity limit represents the maximum number of concurrent effective physical workstations that the physical training space can provide; Indicates the conflict rate of reusing the theoretical and practical space; Physical load threshold At the computer's underlying level, it manifests as a rigid, static scalar boundary used to trigger exception interrupts and translation algorithms. Its physical essence is the maximum concurrent load capacity that the physical training space can provide under the dual constraints of meeting national fire safety mandatory red lines and basic teaching quality control; in the underlying data constraints, The absolute physical value space is strictly defined as .

[0048] Furthermore, the preset physical load threshold The specific value is dynamically configured based on the hazard level and physical buffer requirements of the training site: For conventional theoretical classrooms or pure computer lab environments, in order to maximize site utilization, the system preferentially configures its full-load scalar value to the absolute upper limit (i.e., For high-risk, high-consumption training spaces involving hazardous chemicals and heavy machinery operations, to reserve a safety buffer (such as a backup workstation for temporary equipment failure), the system preferentially lowers this threshold to a safe level range (e.g., ); when the real-time conflict rate When the configuration threshold is exceeded, the computer's underlying system throws a space overload exception and forcibly blocks the scheduling write operation of the current time slice.

[0049] The calculated reuse conflict rate is monitored, and it is determined whether it exceeds the preset physical load threshold. ; If it is determined that it exceeds (i.e.) This indicates a scheduling collision occurring in the physical reality space; in this case, an avoidance step size parameter is introduced for the course nodes that cause the conflict in the teaching process time sequence vector. Avoidance step size parameter It represents the smallest sliding unit in the time-slice resource scheduling sequence (e.g., extending by 1 teaching week). Specifically, its value is based on a fixed-length offset constant of the standard timestamp (to prevent time fragmentation caused by consecutive decimals).

[0050] pass Perform a second-order peak-shifting correction iteration, where, This is a new time slice after the second peak shifting correction iteration, representing the valid scheduling timestamp of the abnormal course node that caused the spatial collision, which was eventually reallocated and written into memory after passing through the computer's underlying sliding scheduling. The integer step size multiplier representing the peak shifting and avoidance iteration (i.e., the iteration number variable, system control) (Starting from 1, proceed with backward exploration by incrementing the levels). The overloaded demand is gradually released to adjacent low-load time slices until the conflict rate of the physical space reuse in all time slices converges to within the preset physical load threshold (i.e., This outputs the final revised data that completes the full-dimensional physical boundary mapping.

[0051] S4. Perform consistency state verification on each independent business module that has completed partial reconstruction and linkage correction; based on the verification confirmation signal of the verification terminal for the feature reconstruction node, structurally assemble each independent business module in the closed-loop steady state, generate and output the reconstruction document, and the specific steps involved include: Extract the final feature data from the underlying cache of each independent business module. The final feature data includes at least: the updated credit allocation vector, the teaching process time sequence vector, and the theory-practice space reuse conflict rate within each time slice. The pre-defined multi-dimensional boundary verification engine (a self-deployed, asynchronous event-driven verification microservice component on the underlying computer software architecture, consisting of a rule parsing and mapping unit, a concurrent sandbox execution unit, and a Boolean aggregation decision unit) is invoked to perform cross-module closed-loop logic verification based on the final feature data. The closed-loop logic verification includes at least: overflow verification for data logic capacity boundaries, deadlock prevention verification for the underlying time-series scheduling network, and security load verification for physical resources. If the closed-loop logic verification passes, the underlying state flag of each independent business module is set to the closed-loop steady state. At the same time, the final feature data and the traceability hash label are packaged to generate a feature reconstruction node, which is then pushed to the verification terminal.

[0052] In this embodiment, closed-loop logic verification is performed in the following three dimensions: Overflow verification for data logic capacity boundaries: Compare the sum of the total credits in the credit allocation vector with the rigid constraint parameter embedded in the local reconstruction instruction (i.e., the absolute upper limit of legal credits) to determine whether the underlying accumulator register has overflowed. If the sum exceeds the upper limit, the system determines that the current reconstruction logic violates the compliance boundary of the scheme; For deadlock prevention verification of the underlying time-series scheduling network: Based on the directed acyclic graph (DAG) model, the time-series vector of the teaching process is traversed, and the loop detection algorithm is called to determine whether there is a timestamp reversal caused by node insertion or drift correction in the time-series network, thereby identifying whether there is a deadlock loop that causes the scheduling logic to be unable to close. Security load verification for physical resources: Scan the global timeline and extract the spatial reuse conflict rate of the physical resources within each time slice. And compare them one by one with the preset physical load threshold. The comparison is performed, and the verification engine determines whether it meets the requirements. Consistency constraints are applied to ensure that the refactoring scheme does not have the risk of resource breakdown in the physical space dimension.

[0053] The consistency state check is considered successful only if all three boundary checks mentioned above return a high-level pass signal. At this time, the system triggers the underlying state machine to forcibly flip the system-level state flag of each independent business module from the "under revision" state and lock it to the "closed-loop steady state".

[0054] Simultaneously, the system packages the final feature data with the underlying traceability hash tag generated in step S3. The system performs memory serialization encapsulation to generate feature reconstruction nodes containing complete data lineage characteristics, and pushes them to the verification terminal via the system coupling bus for experts to perform visual verification.

[0055] It is worth noting that the parsing and mapping unit in the multi-dimensional boundary verification engine is used to extract rigid constraint parameters (such as the legal credit limit) and system-preset physical red lines (such as...) from the local reconstruction instructions after receiving the final feature data. ), and compile it into an assertion script that can be directly executed by a computer; To avoid blocking the main system thread during cross-module verification of massive amounts of data, the concurrent sandbox execution unit creates an independent sandbox runtime environment at the system level and invokes a multi-threaded concurrency mechanism. Specifically, the system simultaneously starts three isolated daemon threads, which inject the extracted final feature data into three parallel logical verifiers, respectively, and execute the aforementioned "logical capacity overflow verification," "time-series network deadlock prevention verification," and "physical security load verification" in strictly synchronized time slices. The Boolean aggregation decision unit is used to collect the asynchronous callback results of the three daemon threads mentioned above. At the underlying logic circuit simulation level, this unit is represented as a multi-input Boolean AND gate controller.

[0056] Example 2: This example provides a multi-module local linkage reconfiguration system based on multi-agent collaboration, including a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor executes the computer program to implement the multi-module local linkage reconfiguration method based on multi-agent collaboration described above.

[0057] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A multi-module local linkage reconstruction method based on multi-agent collaboration is applied to an intelligent revision system for talent training programs. This intelligent revision system is configured with interconnected intelligent agents for talent training program revision, regional industry analysis, and talent demand analysis. Its characteristics are as follows: The multi-module local linkage reconstruction method includes the following steps: S1. Obtain the initial talent development plan to be revised through the talent development plan revision agent, and call the preset text parsing engine to decouple the initial talent development plan according to the preset business logic boundaries. Each is an independent business module, among which... ; Simultaneously, establish data communication links with the regional industry analysis intelligence and talent demand analysis intelligence to extract external industry dynamic reference data; S2. Perform cross-dimensional feature matching between each of the decoupled independent business modules and the external dynamic reference data; when the matching degree of any independent business module is lower than the preset compliance feature threshold, mark it as the target business module and generate the corresponding local reconstruction instruction; S3. Based on the local reconstruction instruction, perform node feature reconstruction within the boundary of the target business module to generate updated feature nodes; Based on a preset module logic topology map, the system automatically addresses related business modules affected by the state change of the target business module; the updated feature node is transmitted to the related business module through the coupling bus, triggering the linkage correction of the underlying support parameters in the related business module. S4. Perform consistency state verification on each independent business module that has completed partial reconstruction and linkage correction; based on the verification confirmation signal of the verification terminal for the feature reconstruction node, perform structured assembly on each independent business module that is in closed-loop steady state, and generate and output reconstruction document.

2. The multi-module local linkage reconstruction method based on multi-agent collaboration according to claim 1, characterized in that, In step S1, the preset text parsing engine is a multimodal document parsing model that includes a visual typography feature extraction network and a semantic sequence recognition network; the preset text parsing engine is called to decouple it according to the business logic boundaries. Each independent business module includes the following steps: The physical coordinate layout features of the initial talent training program document are extracted through the visual layout feature extraction network, wherein the visual layout feature extraction network is constructed based on a convolutional neural network and is used to identify and extract text box coordinates, font size and paragraph spacing parameters. Simultaneously, semantic features of the text sequence are extracted through the semantic sequence recognition network, wherein the semantic sequence recognition network is constructed based on a bidirectional encoder representation model and is used to perform vectorized semantic encoding on the document text stream; Multimodal feature alignment and concatenation are performed on the physical coordinate layout features and the text sequence semantic features to construct a document logical structure tree; The node distribution of the document logical structure tree is compared with the preset standard training scheme topology map to identify and locate the first-level title feature anchor points corresponding to each business logic boundary; The document logical structure tree is subjected to continuous segmentation and isolation along the physical coordinate axes of the first-level heading feature anchor points, and each independent branch node obtained after segmentation is structurally encapsulated into a structure carrying a unique logical identifier. Each is an independent business module.

3. The multi-module local linkage reconstruction method based on multi-agent collaboration according to claim 1, characterized in that, In step S2, the decoupled independent business modules are matched with the external dynamic reference data across dimensions to calculate the feature matching degree of any independent business module. The specific steps involved are as follows: Extract the semantic distribution feature vector of the target independent business module, and the industry capability node feature vector in the external dynamic reference data that is mapped and associated with the target independent business module; A semantic tolerance matrix is ​​introduced to calculate the effective spatial cosine similarity between the semantic distribution feature vector and the industry capability node feature vector; Extract the attribute slot parameters of the independent business module and compare them with the preset rigid constraint threshold boundary. When an out-of-bounds error occurs, generate an asymmetric compliance penalty operator. Extract the skill word frequency popularity weight and ability life cycle decay rate embedded in the external dynamic reference data; The effective spatial cosine similarity, the skill word frequency popularity weight, and the ability life cycle decay rate are weighted and summed, and the summation result is subjected to dimensionality reduction multiplication mapping with the asymmetric compliance penalty operator to output the feature matching degree. When the feature matching degree is lower than the preset compliance feature threshold, the feature difference that maps the feature matching degree shortfall is extracted, encapsulated as a capability gap indicator vector, the attribute slot parameter is extracted as a rigid constraint parameter, and the corresponding atomic operation operator is configured based on the feature difference. The logical identifier of the target independent business module, the capability gap indication vector, the rigid constraint parameter, and the atomic operation operator are structurally encapsulated into a local reconstruction instruction, and then transmitted in a directional manner through the system's underlying communication bus.

4. The multi-module local linkage reconstruction method based on multi-agent collaboration according to claim 3, characterized in that, In step S3, the specific steps involved in generating and updating feature nodes are as follows: The partial reconstruction instructions are parsed, and logical identifiers are extracted to lock the physical memory boundaries of the target business module; Based on the atomic operation operators in the local reconstruction instruction, the capability gap indicator vector is incrementally injected into the feature distribution matrix of the target business module, and weight reallocation is performed under the constraint of rigid constraint parameters. The cryptographic hash function is invoked to concatenate the underlying logic execution operator with the original industry data index pointer representing the basis for modification, generating a unique traceability hash label; The feature distribution matrix after weight redistribution is instantiated to generate an updated feature node containing the reconstructed semantic feature distribution, capability attribute labels, and the source hash labels.

5. The multi-module local linkage reconstruction method based on multi-agent collaboration according to claim 4, characterized in that, The specific steps involved in triggering the linkage correction of the underlying support parameters in the associated business module are as follows: Based on a preset module logical topology map, the logical dependency path starting from the target business module is analyzed, and the affected related business modules are located along the logical dependency path. The related business modules include at least: a course setting module, a teaching execution module, and a teaching facility module. The updated feature node is transmitted to the associated service module via the coupling bus in a cascading manner to trigger a linkage correction of the underlying support parameters based on the updated feature node. The underlying support parameter linkage correction specifically refers to the cascading mapping correction performed along the logical dependency path, which includes: Incremental optimization is performed on the credit capacity boundary of the aforementioned course setting module; Perform sequence drift correction on the temporal topology of the teaching execution module; Peak shifting and avoidance are implemented for the physical space load of the teaching facility module.

6. The multi-module local linkage reconstruction method based on multi-agent collaboration according to claim 5, characterized in that, The credit capacity boundary of the curriculum design module is incrementally optimized, and the specific steps involved are as follows: Extract the capability attribute labels and capability gap indication vectors embedded in the updated feature nodes of the pass-through, and combine them with the rigid constraint parameters defined by the local reconstruction instruction to generate a candidate course set; The underlying teaching elements of the candidate course set are analyzed, including course content, course objectives, and course knowledge points; the capability gap indicator vector is mapped to industry capability requirements, and the industry capability requirements are matched with the course content, course objectives, and course knowledge points of each candidate course using multi-dimensional feature matching. Based on the above feature matching results, retrieve the initial capability coverage of each candidate course relative to the missing capability node mapped by the capability gap indicator vector. By comparing the candidate course set with the existing course set in the course setting module, the overlap features of knowledge points between the two are extracted, and a capability overlap attenuation coefficient is constructed to characterize the scale of knowledge redundancy. The initial capability coverage is nonlinearly attenuated using the capability overlap attenuation coefficient, and the initial capability coverage is corrected to an effective marginal capability increment. Using the maximization of the effective marginal capability increment as the objective function and the class hour allocation boundary as the constraint, a knapsack optimization calculation is performed to select target courses and generate the optimal class hour allocation scheme. Based on the preset class hour and credit conversion rules, the credit allocation vector of the course setting module is refreshed.

7. The multi-module local linkage reconstruction method based on multi-agent collaboration according to claim 6, characterized in that, The timing topology relationship of the teaching execution module is corrected for order drift, and the specific steps involved are as follows: The underlying subject progression relationships formed after the introduction of the target course are analyzed, and a prerequisite time dependency matrix reflecting the graph structure relationship between courses is constructed. Extract the default scheduling timestamp assigned to the target course in the initial reconstruction state; The prerequisite time-series dependency matrix is ​​invoked to perform a topological sorting constraint check on the default scheduling timestamps; When the verification result does not meet the preset minimum teaching cycle span, a time series penalty operator is introduced to perform a sequence drift correction on the default scheduling timestamp until the one-way teaching logic constraint is met, and the teaching process time sequence vector is output.

8. The multi-module local linkage reconstruction method based on multi-agent collaboration according to claim 7, characterized in that, The physical space load of the teaching facility module is shifted to avoid peak loads, and the specific steps involved are as follows: Extract the training site requirements parameters for each time slice in the reconstructed teaching process time-series vector; By comparing the real-time occupancy feature matrix of the physical resources preset in the teaching facility module, the reusability conflict rate of the physical space in each target time period is calculated. Determine whether the conflict rate of the physical space reuse exceeds a preset physical load threshold; If the conflict rate exceeds the threshold, an avoidance step size parameter is introduced for the course nodes that cause conflict in the time sequence vector of the teaching process, and a second peak shifting correction iteration is performed until the conflict rate of the theoretical and practical space reuse in all time slices converges to within the preset physical load threshold.

9. The multi-module local linkage reconstruction method based on multi-agent collaboration according to claim 1, characterized in that, In step S4, a consistency state check is performed on each independent business module that has completed partial reconstruction and linkage correction. The specific steps involved include: Extract the final feature data from the underlying cache of each independent business module. The final feature data includes at least: the updated credit allocation vector, the teaching process time sequence vector, and the theory-practice space reuse conflict rate within each time slice. The preset multi-dimensional boundary verification engine is invoked to perform cross-module closed-loop logic verification based on the final feature data. The closed-loop logic verification includes at least: overflow verification for data logic capacity boundaries, deadlock prevention verification for the underlying time-series scheduling network, and security load verification for physical resources. If the closed-loop logic verification passes, the underlying state flag of each independent business module is set to the closed-loop steady state. At the same time, the final feature data and the traceability hash label are packaged to generate a feature reconstruction node, which is then pushed to the verification terminal.

10. A multi-module local linkage reconfiguration system based on multi-agent collaboration, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, The processor executes a computer program to implement the multi-module local linkage reconstruction method based on multi-agent collaboration as described in any one of claims 1-9.

Images