A large model-based TGCOA-MDAN personnel training scheme intelligent generation model

Abstract

The application relates to the technical field of personnel training, and discloses an intelligent generation method of a personnel training scheme based on a large model-TGCOA-MDAN, which comprises a data preprocessing and index selection module, which is used for structurally fusing multi-source heterogeneous data of personnel, and outputs comprehensive evaluation indexes containing personnel skill states, post standard requirements and business influence levels. A scheduling and cooperative control module analyzes the comprehensive evaluation indexes, generates algorithm configuration parameters, and issues control instructions and business data streams carrying corresponding parameters. An individualized scheme generation engine based on TGCOA responds to the control instructions, optimizes under constraints through a global optimization algorithm, generates and feeds back an initial training scheme. A cross-domain self-adaption and effect iteration module driven by MDAN completes feature cross-domain alignment and correction, and outputs a target training scheme. A result analysis, explanation and dynamic iteration optimization mechanism performs explainability analysis, generates optimization signals in combination with offline implementation effect feedback.

Description

Technical Field

[0001] This application relates to the field of support personnel training technology, and more specifically, to an intelligent generation model for support personnel training schemes based on the large model-TGCOA-MDAN. Background Technology

[0002] With the increasing complexity of equipment systems and support missions, the professional skills training and competency matching of support personnel have become crucial. Currently, the development of training programs for support personnel mainly relies on human experience, fixed templates, or simple rule-based recommendation systems. Although some systems have initially introduced traditional machine learning models in hopes of achieving automated matching, they still face many bottlenecks in practical applications.

[0003] First, existing information systems often struggle to effectively extract and integrate heterogeneous data from multiple sources, such as personnel, equipment logs, and business feedback. This results in inaccurate personnel capability profiles and often leads to standardized, one-size-fits-all training programs with low personalization, failing to dynamically match actual personnel capability changes with the urgency of business operations. Second, traditional predictive models are typically trained on a single data distribution, exhibiting extremely weak cross-domain adaptability. When these models are extended to new units with different equipment systems or mission modes, they often face severe domain shift issues, causing the generated solutions to be significantly out of sync with actual needs. Furthermore, existing solution optimization processes often involve fixed parameters, leading to low optimization efficiency when facing complex resource constraints and a high risk of getting trapped in local optima. Finally, existing system outputs are often "black boxes," lacking interpretability of the generation logic and failing to establish a closed-loop iterative mechanism based on actual training performance feedback, hindering long-term self-evolution and decision support. These shortcomings significantly impede the progress of personnel training management towards intelligence and precision. Summary of the Invention

[0004] The main purpose of this application is to provide an intelligent generation model based on the large model-TGCOA-MDAN training scheme for security personnel, so as to solve the technical problems in the background art.

[0005] To achieve the above objectives, the first aspect of this application proposes an intelligent generation model based on the large model-TGCOA-MDAN training scheme for support personnel, comprising: The data preprocessing and indicator selection module is used to acquire multi-source heterogeneous data related to support personnel, and to perform structured fusion of the multi-source heterogeneous data to output a comprehensive evaluation indicator that includes personnel skill status, job standard requirements, and business impact level. The scheduling and collaborative control module, connected to the data preprocessing and indicator selection module, is used to receive the comprehensive evaluation indicators, parse the comprehensive evaluation indicators to form algorithm configuration parameters, and generate control instructions with algorithm configuration parameters and corresponding business data streams. A personalized scheme generation engine based on TGCOA is connected to the scheduling and collaborative control module. It is used to respond to the control command, receive the business data stream, and within the preset training resource constraint boundary, apply a global optimization algorithm to optimize according to the algorithm configuration parameters, generate an initial training scheme, and feed it back to the scheduling and collaborative control module. The MDAN-driven cross-domain adaptation and effect iteration module is connected to the scheduling and collaborative control module. It is used to respond to the control command, receive the initial training scheme in the business data stream, introduce the target domain task environment data to be adapted, apply the domain adaptation network to perform feature cross-domain alignment and correction, and output the target training scheme. The result analysis, interpretation, and dynamic iterative optimization mechanism has an input end connected to the MDAN-driven cross-domain adaptive and effect iteration module, and an output end connected to the scheduling and collaborative control module. It is used to perform interpretability analysis on the target training scheme, and combine the obtained offline implementation effect feedback data of the target training scheme to form an optimization signal. The optimization signal is then fed back to the scheduling and collaborative control module to drive the scheduling and collaborative control module to dynamically adjust the control commands and the business data stream.

[0006] In some feasible implementations, the data preprocessing and indicator selection module is configured as follows: The system acquires multi-source heterogeneous business data streams containing text, numbers, and time-series data related to support personnel, and calls a pre-trained large language model to perform semantic parsing on the unstructured text in the multi-source heterogeneous business data streams. It identifies key events, capability performance, and problem focus information in the unstructured text and converts them into structured capability tags, wherein the capability tags represent the actual skill status of individual support personnel. The structured capability tags are integrated and associated with the digital and time-series data in the multi-source heterogeneous business data stream and linked to the individual support personnel. The pre-trained large language model is used to infer and complete the missing data features. At the same time, logical verification is used to filter out contradictory data between different business systems to obtain the standardized business dataset of the individual support personnel. Pre-existing job descriptions and competency model data are selected from a pre-defined external standard knowledge base, and the standardized business dataset is input into the pre-trained large language model. The selected job descriptions and competency model data are parsed, and a hierarchical and categorized competency indicator library is extracted. Combined with the pre-defined business logic, the pre-trained large language model is used to analyze the impact of structured competency label features in the standardized business dataset on the success rate of specific support tasks and assign corresponding weights. The business impact level corresponding to the individual support personnel is quantified, and a comprehensive evaluation index including personnel skill status, job standard requirements, and business impact level is obtained.

[0007] In some feasible implementation methods, the data preprocessing and indicator selection module, when obtaining a comprehensive evaluation indicator that includes personnel skill status, job standard requirements, and business impact level, is further configured as follows: The pre-trained large language model is driven to parse the preset business logic, analyze the potential impact of the structured capability label features in the standardized business dataset on the success rate of specific guarantee tasks, and assign corresponding guarantee weight values ​​according to the degree of impact to generate an initial business impact level that represents the relevance of the task. The system acquires the individual promotion plans, recent task priorities, and technology deployment plans of the support personnel, drives the pre-trained large language model to parse and transform these into future capability mapping features, core task capability features, and new equipment skill features for the corresponding positions, and matches them with the capability gaps determined by comparing the structured capability tag features with the hierarchical and classified capability index library. The system evaluates the learning urgency coefficient of the training needs corresponding to each capability gap, and uses the learning urgency coefficient and the support weight value to correct the initial business impact level, generating a corrected target business impact level. The target business impact level is integrated and mapped with the structured capability label features in the standardized business dataset and the hierarchical classification capability indicator library to output a comprehensive evaluation indicator that includes personnel skill status, job standard requirements, and business impact level.

[0008] In some possible implementation methods, the scheduling and coordination control module is configured as follows: Receive the comprehensive evaluation index, call the pre-trained large language model to parse the comprehensive evaluation index, and extract the scene features and data attributes corresponding to the current task; Based on the scene characteristics and the data attributes, determine the functional sub-modules that the current task needs to call, and combine the operation requirements of the functional sub-modules with the analysis results of the comprehensive evaluation index to generate algorithm configuration parameters adapted to the current task; Based on the algorithm configuration parameters, control instructions carrying corresponding parameters and matching business data streams are generated, and the control instructions and business data streams are respectively distributed to the TGCOA-based personalized solution generation engine and the MDAN-driven cross-domain adaptive and effect iteration module. The algorithm receives optimization signals from the result analysis, interpretation, and dynamic iterative optimization mechanism, adjusts the algorithm configuration parameters according to the optimization signals, updates the corresponding control commands and business data streams, and then outputs them to each functional module.

[0009] In some feasible implementations, the TGCOA-based personalized solution generation engine is configured as follows: In response to the control command, receive the corresponding business data stream, configure the algorithm running parameters according to the algorithm configuration parameters carried by the control command, extract personnel skill status, job standard requirements, and business impact level from the comprehensive evaluation index, and construct the candidate solution space of the training scheme and the fitness evaluation function of the fit in combination with the preset training resource constraint boundary; A preset number of candidate training schemes are randomly generated as stingless bee individuals in the candidate solution space of the training scheme to complete the bee colony initialization. The fitness value corresponding to each stingless bee individual is calculated based on the fitness evaluation function to obtain the initial bee colony set and the initial global optimal candidate training scheme. The initial bee colony set is divided into scout bees, foraging bees, and follower bees according to a preset division of labor strategy. Using the initial globally optimal candidate training scheme as the iterative benchmark, multi-role collaborative optimization is performed. The scout bees are scheduled to conduct a global random search within the candidate solution space of the training scheme and generate new candidate training schemes. The foraging bees are scheduled to conduct a local search around the currently better candidate training scheme and adjust the components of the candidate training scheme to optimize the fitness value. The follower bees are scheduled to follow the search in the direction of a high-quality candidate training scheme based on the population information sharing mechanism. After completing the fitness verification of all candidate training schemes based on the fitness evaluation function, the updated bee colony set and the globally optimal candidate training scheme for this iteration are obtained. After each round of optimization, all candidate training schemes in the updated bee colony set are screened based on the calculation results of the fitness evaluation function. New random candidate training schemes are added to the candidate solution space of the training scheme to maintain population diversity. It is then determined whether the preset termination condition has been met. If not, the multi-role collaborative optimization is returned. If it has been met, an initial training scheme is generated and fed back to the scheduling and collaborative control module.

[0010] In some implementable methods, the MDAN-driven cross-domain adaptation and effect iteration module is configured as follows: In response to the control command, the system receives the corresponding business data stream, obtains the initial training scheme from the business data stream, introduces the target domain task environment data to be adapted, and retrieves the preset source domain labeled training sample data to form cross-domain adaptation data preparation. The feature extraction network is trained based on the source domain labeled training sample data. During the training process, a hybrid regularization strategy is adopted to enhance the generalization of the feature space. Self-supervised training is performed by combining controlled reconstruction learning to obtain the source domain pre-trained feature extraction model. The source domain pre-trained feature extraction model is invoked to predict the target domain task environment data to be adapted, generating corresponding pseudo labels. The mixing ratio of the source domain labeled training sample data and the target domain task environment data to be adapted is adjusted using Wasserstein distance as a distribution difference metric. In the input layer and feature layer of the source domain pre-trained feature extraction model, an intermediate mixed domain formed by the proportional mixing of the two domain data is constructed respectively. Progressive feature alignment is performed to gradually reduce the distribution difference between the two domains to obtain the intermediate domain adaptation feature representation. Based on the intermediate domain adaptation feature representation, the pseudo-labels are used as supervision information to perform self-learning training on the target domain task environment data to be adapted. In the feature space corresponding to the target domain task environment data to be adapted, a hybrid regularization strategy is used to construct consistency constraints, filter low-confidence pseudo-labels, and obtain the target domain optimized feature extraction model. The initial training scheme is modified for cross-domain features and adapted to the scene by using the feature extraction model optimized for the target domain, thereby generating the target training scheme.

[0011] In some feasible implementation methods, the result analysis, interpretation, and dynamic iterative optimization mechanism is configured as follows: Receive the target training scheme, obtain the offline implementation effect feedback data corresponding to the target training scheme, call the pre-trained large language model to perform semantic parsing and structural transformation on the unstructured content in the offline implementation effect feedback data, filter out abnormal data, and match and align the processed offline implementation effect feedback data with the personnel skill status, job standard requirements and business impact level corresponding to the target training scheme one by one to obtain standardized effect feedback data and corresponding analysis benchmarks; Based on the aforementioned analytical benchmark, the pre-trained large language model is invoked to verify and interpret the target training scheme from four dimensions: capability coverage, resource matching status, expected business impact, and cross-domain adaptation effect, generating a structured scheme interpretability report. Using the expected indicators of each dimension in the structured solution interpretability report as a comparison benchmark, the standardized effect feedback data is compared with the corresponding expected effect, effect attribution analysis is performed, the solution deviation points and performance change causes are located, effect evaluation results are generated, and optimization signals are converted into optimization signals and fed back to the scheduling and collaborative control module. The optimization signals include algorithm configuration parameter adjustment instructions and solution generation logic correction directions.

[0012] Secondly, this application provides an intelligent generation method for support personnel training schemes based on the large model-TGCOA-MDAN, applied to the aforementioned intelligent generation model for support personnel training schemes based on the large model-TGCOA-MDAN, including: Acquire multi-source heterogeneous data related to support personnel, and perform structured fusion of the multi-source heterogeneous data to output a comprehensive evaluation index that includes personnel skill status, job standard requirements, and business impact level; The system receives the comprehensive evaluation index, parses the comprehensive evaluation index to form algorithm configuration parameters, and generates control instructions with algorithm configuration parameters and corresponding business data streams. In response to the control command, the system receives the service data stream and, within the preset training resource constraint boundary, applies a global optimization algorithm to perform optimization based on the algorithm configuration parameters, generates an initial training scheme, and feeds it back to the scheduling and collaborative control module. In response to the control command, the system receives the initial training scheme from the business data stream, introduces the target domain task environment data to be adapted, applies the domain adaptation network to perform feature cross-domain alignment and correction, and outputs the target training scheme. The target training scheme is subjected to interpretability analysis, and combined with the offline implementation effect feedback data of the target training scheme, an optimization signal is formed. The optimization signal is then fed back to the scheduling and collaborative control module to drive the scheduling and collaborative control module to dynamically adjust the control commands and the business data stream.

[0013] Thirdly, this application provides a computer storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the aforementioned intelligent generation method for support personnel training schemes based on large model-TGCOA-MDAN.

[0014] Fourthly, this application provides a computer program that, when executed by a processor, implements the steps of the aforementioned intelligent generation method for support personnel training schemes based on large model-TGCOA-MDAN.

[0015] The technical solutions provided by the embodiments of this application may include the following beneficial effects: This application presents an intelligent generation model for support personnel training schemes based on the large model-TGCOA-MDAN framework. It constructs an integrated intelligent generation model architecture, achieving structured fusion of multi-source heterogeneous data through data preprocessing and indicator selection modules. The model outputs comprehensive evaluation indicators covering personnel skill status, job standard requirements, and business impact levels, accurately characterizing personnel capabilities and business needs. This effectively addresses the pain points of existing systems, such as insufficient data fusion capabilities, one-size-fits-all training schemes, and low personalization matching. Furthermore, by using a TGCOA-based personalized scheme generation engine to perform global optimization within preset training resource constraints, it improves upon traditional optimization methods that suffer from parameter rigidity and susceptibility to traps. By overcoming the limitations of local optima, the efficiency and quality of solution optimization under complex constraints are improved. The MDAN-driven cross-domain adaptation and effect iteration modules achieve cross-domain feature alignment and solution correction, mitigating domain offset issues caused by differences in equipment systems and mission modes among different support units, and ensuring the model's solution adaptation accuracy in new application scenarios. Through result analysis, interpretation, and dynamic iterative optimization mechanisms, solution interpretability analysis is conducted, and combined with feedback from offline implementation effects, optimization signals are generated to drive dynamic adjustments across the entire chain. This breaks the black box state of model output, establishes a sustainable evolutionary closed loop, and comprehensively promotes the intelligent and precise development of support personnel training and management. Attached Figure Description

[0016] The accompanying drawings, which form part of this application, are used to provide a further understanding of the application and to make other features, objects, and advantages of the application more apparent. The illustrative embodiments and descriptions of this application are used to explain the application and do not constitute an undue limitation of the application. In the drawings: Figure 1 The logical diagram of the intelligent generation model for the support personnel training scheme based on the large model-TGCOA-MDAN provided in this application. Detailed Implementation

[0017] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0018] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of this application described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0019] In this application, the terms "upper," "lower," "left," "right," "front," "rear," "top," "bottom," "inner," "outer," "middle," "vertical," "horizontal," "lateral," and "longitudinal" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. These terms are primarily for the purpose of better describing this application and its embodiments, and are not intended to limit the indicated device, element, or component to having a specific orientation, or to be constructed and operated in a specific orientation.

[0020] Furthermore, in addition to indicating location or positional relationship, some of the aforementioned terms may also have other meanings. For example, the term "above" may also be used in some cases to indicate a certain dependency or connection relationship. Those skilled in the art can understand the specific meaning of these terms in this application based on the specific circumstances.

[0021] Furthermore, the terms "installation," "setup," "equipped with," "connection," "linked," and "socketing" should be interpreted broadly. For example, "connection" can be a fixed connection, a detachable connection, or an integral structure; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, or an internal connection between two devices, components, or parts. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0022] The following explanations of some terms used in this application are provided to aid in understanding the technical solution of this application: Large Language Models (LLMs) refer to pre-trained language models with a large number of parameters (usually billions to trillions), such as ChatGLM, GPT-4, and ERNIE, which have powerful capabilities in semantic understanding, text generation, logical reasoning, and multimodal data processing.

[0023] TGCOA (TetraGonula Carbonaria Optimization Algorithm) is a metaheuristic optimization algorithm based on the behavior of stingless bee colonies. It simulates the division of labor and cooperation mechanism of bee colonies in foraging, nest building, and information exchange, and achieves a balance between global exploration and local development through the division of labor among scout bees, foraging bees, and follower bees.

[0024] MDAN (Mixed Domain Adaptation Network) is a hybrid learning method for cross-domain adaptation. It achieves smooth transfer and adaptation of the model between different data distributions (domains) through a three-stage strategy of "source domain training - intermediate hybrid domain construction - target domain self-learning".

[0025] Domain adaptation is a transfer learning method that aims to adapt a model trained on one data distribution (source domain) to another different but related data distribution (target domain) to mitigate the performance degradation (domain shift) caused by differences in data distribution.

[0026] The source domain refers to a training environment (such as a pilot unit) with a large amount of labeled data, in which the model performs supervised learning and learns the basic rules and feature representations.

[0027] The target domain refers to a new application environment (such as a new security unit) where data is scarce or completely unlabeled. The model needs to maintain or improve its performance in this environment through domain adaptation techniques.

[0028] Mixup regularization is a data augmentation and regularization technique that enhances the generalization ability and consistency of a model by generating new samples through linear interpolation of two samples and their labels.

[0029] Pseudo-labels are temporary labels generated from unlabeled data using existing models. They are used for self-supervised or semi-supervised learning and usually require noise suppression strategies to improve reliability.

[0030] Wasserstein distance is a metric used to measure the difference between two probability distributions. It is often used in domain adaptation to assess the distributional differences between the source and target domains and to guide the adjustment of the mixing ratio.

[0031] Bidirectional Long Short-Term Memory (Bi-LSTM) is a recurrent neural network structure that can capture contextual information simultaneously and is suitable for time-series data analysis, such as operation logs and fault sequences.

[0032] Attention mechanism is a computational mechanism that simulates human attention allocation. By assigning different weights to input data, it highlights key information and improves the accuracy of the model in tasks such as matching and filtering.

[0033] The fitness function is a function used in optimization algorithms (such as TGCOA) to evaluate the merits of candidate solutions (such as training schemes), taking into account multiple objective factors such as capability coverage, resource matching degree, and business priority.

[0034] Closed-Loop Iterative Optimization refers to a system that, after outputting a solution, collects actual execution performance data and feeds it back into the model for parameter and logic adjustments, forming a continuous self-evolving optimization loop.

[0035] Interpretability refers to the ability of a model's output and decision-making process to be understood and traced by humans. It is usually achieved through natural language interpretation, decision logs, and visualization analysis.

[0036] The Progressive Mixup Strategy, used in MDAN, is a method for adjusting the mixing ratio between the source and target domains. It achieves smooth domain alignment by dynamically calculating the distribution distance and gradually adjusting the mixing weights.

[0037] Firstly, this application provides an intelligent generation model for support personnel training schemes based on large model-TGCOA-MDAN, including a data preprocessing and indicator selection module, a scheduling and collaborative control module, a personalized scheme generation engine based on TGCOA, an MDAN-driven cross-domain adaptive and effect iteration module, and a result analysis, interpretation, and dynamic iterative optimization mechanism.

[0038] The data preprocessing and indicator selection module is used to acquire multi-source heterogeneous data related to support personnel, and to perform structured fusion of the multi-source heterogeneous data to output a comprehensive evaluation indicator that includes personnel skill status, job standard requirements, and business impact level.

[0039] Specifically, the data preprocessing and indicator selection module is configured as follows: S101: Obtain multi-source heterogeneous business data streams containing text, numbers, and time-series data related to support personnel, and call a pre-trained large language model to perform semantic parsing on the unstructured text in the multi-source heterogeneous business data streams, identify key events, capability performance, and problem focus information in the unstructured text, and convert them into structured capability tags.

[0040] The capability tags represent the actual skill status of individual support personnel.

[0041] Specifically, the system first acquires multi-source heterogeneous business data streams related to support personnel, including text, numerical, and time-series data. The system extensively integrates data from multiple independent business systems, including but not limited to: personnel files and job information from the human resources system; historical scores and skill assessment results from the training and assessment system; operation logs and fault records from the equipment management system; task reports and performance evaluations from the business feedback system; and new equipment technical manuals and typical cases from external knowledge bases. This data integrates structured numerical and time-series data as well as unstructured text.

[0042] After acquiring the data, a pre-trained large language model (such as ChatGLM, ERNIE, etc.) is invoked to perform semantic parsing on the unstructured text in the multi-source heterogeneous business data stream. This identifies key events, capability performance, and problem focus information within the unstructured text and transforms them into structured capability tags. For example, leveraging the unstructured text understanding capabilities of the large model, key information is automatically extracted from task reports, fault descriptions, or personal summaries. Natural language descriptions such as "In XX task, I independently eliminated ZZ-type faults in YY-type equipment" are transformed into structured tags such as "Fault Troubleshooting Capability - YY Equipment - ZZ Fault Type." These capability tags represent the actual skill level of individual support personnel.

[0043] Furthermore, during the structured fusion of the multi-source heterogeneous business data streams, the data preprocessing and indicator selection module automatically adds "domain labels" to the structured capability labels and the standardized business datasets based on the scenario environment of the data source (e.g., daily simulation training, exercise assessment, real combat training, etc.). These domain labels are used to distinguish the distribution characteristics of the data, thereby providing underlying domain identification data support for downstream cross-domain capability adaptation and training scheme migration based on MDAN (Multi-Domain Adversarial Adaptive Network).

[0044] S102, the structured capability tags are integrated and associated with the digital and time-series data in the multi-source heterogeneous business data stream to the individual support personnel, and the pre-trained large language model is used to infer and complete the missing data features. At the same time, the contradictory data between different business systems is filtered out through logical verification to obtain the standardized business dataset of the individual support personnel.

[0045] Specifically, this step uses the unique identity of the support personnel as the primary key to integrate and associate all the structured capability tags generated in S101 with the structured digital data and time-series record data in the original multi-source heterogeneous business data stream, so that all data items are collected under the corresponding individual support personnel name to form an initial fusion dataset.

[0046] To address the feature gaps in the initial fused dataset, a pre-trained large language model was invoked to perform inference and completion based on entity recognition and relational reasoning capabilities. The pre-trained large language model, combined with the overall ability distribution patterns of personnel in the same position and with similar qualifications, infers features such as missing skill records and assessment data. For example, for a personnel missing records related to complex electromagnetic environment operations, the model references the general performance of similar support personnel, combines this personnel's existing skill tags, predicts their corresponding skill weaknesses, and completes the corresponding feature values.

[0047] To address data inconsistencies between different business systems, a pre-trained large-scale language model is used to perform noise filtering and consistency verification, cross-checking the logical consistency of data from multiple sources. For example, when there is a significant discrepancy between a personnel's training assessment results and the actual operation records retained by the equipment system, the pre-trained large-scale language model, based on the cross-validation results of multi-source data, retains the conclusions corresponding to the more credible actual operation records and eliminates contradictory and abnormal data, ensuring the reliability of the dataset. After completing missing feature completion and contradictory data filtering, a standardized business dataset for that individual support personnel is generated.

[0048] When using the pre-trained large language model for inference completion and consistency verification, the system pre-constructs a domain prompt template containing "task instructions, role settings, contextual data, and output format." Specifically, the system fills the template with the capability distribution feature vector of the same job group as contextual prior knowledge, and inputs it into the large language model in combination with the existing skill entity extraction results of the support personnel. The large language model captures the dependencies between contextual entities based on a multi-layer self-attention mechanism and outputs the completion feature with the highest probability distribution. At the same time, by comparing the cosine similarity of feature entities in the semantic space of texts in different business systems, if the similarity is lower than a preset threshold, it is judged as contradictory data and filtered out.

[0049] S103, Select pre-existing job descriptions and competency model data from a preset external standard knowledge base, and input the standardized business dataset into the pre-trained large language model. Analyze the selected job descriptions and competency model data, extract a hierarchical and categorized competency indicator library, and combine it with preset business logic. Use the pre-trained large language model to analyze the impact of structured competency label features in the standardized business dataset on the success rate of specific support tasks and assign corresponding weights. Quantify the business impact level of the individual support personnel to obtain a comprehensive evaluation index that includes personnel skill status, job standard requirements, and business impact level.

[0050] Specifically, this step retrieves the job description and competency model data for the corresponding position from the preset external standard knowledge base, and inputs them together with the standardized business dataset generated in S102 into the pre-trained large language model to complete the extraction of competency indicators, generation of competency profiles, quantification of business impact, and processing of resource characteristics in sequence.

[0051] Competency model data is standardized data pre-built into an external standard knowledge base, used to define all the competency elements, level standards, and behavioral guidelines that personnel in corresponding support positions should possess. Competency model data serves as the official benchmark for measuring the actual competency level of support personnel and establishing job standard requirements. Competency model data, together with job descriptions, constitutes the basic input material for extracting competency indicators from a pre-trained large language model.

[0052] In the context of assurance scenarios, competency model data can specifically include the following four core categories: Capability Dimension Framework Data: Clearly define all major capability categories covered by the job, such as primary capability dimensions like equipment operation capability, troubleshooting capability, emergency repair capability, task coordination capability, and safety standard execution capability, as well as the hierarchical classification rules for equipment subcategories and detailed skills under each dimension. This forms the structural basis for generating a three-tiered capability indicator library.

[0053] Tiered skill standard data: For each sub-skill, define the mastery requirements corresponding to different ability levels. For example, divide the three-level skills into three levels: beginner, proficient and expert. Clearly define the quantitative or qualitative standards that personnel should achieve at the corresponding level, such as operational precision, processing speed and independent completion rate, so as to benchmark the actual ability of personnel and determine the ability gap.

[0054] Behavioral anchoring description data: Typical behavioral performance and task scenario examples corresponding to each ability level. For example, "proficient level fault diagnosis ability" corresponds to "being able to independently handle sudden faults such as radar signal interference in complex electromagnetic environments". This is used in large models to accurately match personnel's unstructured behavioral records, task performance and ability levels.

[0055] Job weighting configuration data: Clearly define the essential attributes and weight percentages of various capabilities in positions of different job levels and task directions. For example, the core fault diagnosis capability of the main combat equipment support position is a mandatory requirement with the highest weight, while the similar capability of the logistics support position is a low-level requirement with a lower weight. This is used for the weighting calculation of the impact level of subsequent business operations.

[0056] During the execution of step S103, the pre-trained large language model reads the complete job description and competency model data, performs semantic parsing and structural decomposition on the two types of data, and extracts a hierarchical and classified three-level competency indicator library to clarify the job standard requirements for the corresponding positions. At the same time, the pre-trained large language model uses the level standards in the competency model data as a reference to benchmark the structured competency labels of support personnel, determine the competency gaps of support personnel, and support the subsequent quantitative calculation of the business impact level.

[0057] First, a hierarchical and categorized competency indicator library is extracted. A pre-trained large language model parses and decomposes the job descriptions and competency model data, generating a three-tiered competency indicator system: Level 1 indicators correspond to broad competency categories such as equipment operation capabilities, troubleshooting capabilities, and emergency support capabilities; Level 2 indicators correspond to specific equipment dimensions such as a certain type of radar or a certain type of communication equipment; and Level 3 indicators correspond to detailed skill items such as power-on self-test, target acquisition, and interference troubleshooting. This hierarchical and categorized competency indicator library clearly defines the job standard requirements for the corresponding positions.

[0058] Second, generate personalized capability profiles. Based on standardized business datasets and capability indicator libraries, generate dynamic, multi-dimensional capability profiles for each support personnel. These profiles typically include two parts: first, quantitative scores corresponding to each of the three levels of capability indicators, calculated from historical assessment data and practical operation records; second, qualitative capability descriptions generated by a large model, such as "This comrade is proficient in operating conventional equipment, but lacks experience in handling sudden malfunctions in complex electromagnetic environments." The quantitative scores and qualitative descriptions together constitute a complete assessment of the personnel's skill level.

[0059] Third, quantify the level of business impact. Based on pre-defined equipment support business logic, a pre-trained large language model analyzes the potential impact of various capability states on the success rate of specific support tasks such as daily maintenance and wartime emergency repairs, assigning corresponding weights to each capability to obtain the basic business impact level. For example, the lack of "high-frequency communication equipment calibration capability" has a high impact weight on emergency communication support tasks. Building on this, by comprehensively considering three factors—personnel career development paths, unit's recent task priorities, and new technology equipment deployment plans—the urgency of various training needs is further assessed. This assessment, combined with the basic business impact level, quantifies the impact level for each individual support personnel.

[0060] Fourth, the feature processing of training resources is completed simultaneously. The data preprocessing and indicator selection module drives the pre-trained large language model to simultaneously extract features from all training resources in the preset course library. This generates detailed labels for each course and each instructor, covering knowledge points, skill points, difficulty level, teaching style, and target audience. These detailed labels are then transformed into resource feature vectors that can be matched by the algorithm. The data preprocessing and indicator selection module stores the extracted training resource feature vectors in a preset resource database for use by downstream modules.

[0061] Ultimately, by integrating three core information categories—personnel skill status, job standard requirements, and business impact level—standardized comprehensive evaluation indicators are output as input data for subsequent processing stages.

[0062] It should be noted that the data preprocessing and indicator selection module, when obtaining the comprehensive evaluation indicators that include personnel skill status, job standard requirements, and business impact level, is further configured as follows: S1031, drive the pre-trained large language model to parse the preset business logic, analyze the potential impact of the structured capability label features in the standardized business dataset on the success rate of specific guarantee tasks, assign corresponding guarantee weight values ​​according to the degree of impact, and generate an initial business impact level that represents the task relevance.

[0063] Specifically, the data preprocessing and indicator selection module drives the pre-trained large language model to parse the preset equipment support business logic rules. These rules can include success criteria and capability influence weighting rules for different types of support tasks, such as routine maintenance, wartime repair, and emergency support. The pre-trained large language model combines all structured capability label features in the standardized business dataset to analyze the missing or insufficient capabilities corresponding to each structured capability label feature, assessing their potential impact on the success rate of various specific support tasks. Based on the degree of influence, each structured capability label feature is assigned a corresponding support weight value, ranging from 0 to 1, with higher values ​​indicating a greater impact of the corresponding capability on the success rate of the support task.

[0064] Taking radar support personnel number BZ007 as an example, the pre-trained large language model analysis of the structured capability tag feature "fault troubleshooting capability - XX model radar - ZZ type signal interference fault handling" shows that the lack of this capability will directly lead to the failure of the target acquisition stage in emergency radar support missions, with a very high impact on the success rate of emergency support missions and a moderate impact on the success rate of routine maintenance missions. Considering the proportion and impact of the two types of missions, the pre-trained large language model assigns a support weight value of 0.72 to this structured capability tag feature.

[0065] The pre-trained large language model uses the guarantee weight value as the core basis to divide the initial business impact level into grade ranges: a guarantee weight value in the range of 0 to 0.3 corresponds to a low initial business impact level, in the range of 0.3 to 0.6 corresponds to a medium initial business impact level, and in the range of 0.6 to 1 corresponds to a high initial business impact level. Based on this, the initial business impact level corresponding to "fault troubleshooting capability - XX model radar - ZZ type signal interference fault handling" is determined to be a high initial business impact level. After all structured capability label features have completed impact analysis and weight assignment, a set of initial business impact levels representing the task relevance for each individual guarantee personnel is generated.

[0066] S1032, obtain the individual promotion plan, recent task priorities and technology deployment plan of the support personnel, drive the pre-trained large language model to parse and transform them into future capability mapping features, core task capability features and new equipment skill features for the corresponding positions, and match them with the capability gaps determined by comparing the structured capability label features with the hierarchical classification capability index library, evaluate the learning urgency coefficient of the training needs corresponding to each capability gap, and use the learning urgency coefficient and the support weight value to correct the initial business impact level, and generate the corrected target business impact level.

[0067] Specifically, the data preprocessing and indicator selection module first acquires three types of external related data: individual career advancement plans of support personnel, recent task priorities documents of the support personnel's units, and new technology equipment deployment plans of the support personnel's units. The data preprocessing and indicator selection module then drives a pre-trained large language model to perform semantic parsing and feature transformation on each of the three types of data. Analyze the career advancement plans of individual support personnel, transform them into future capability mapping characteristics for corresponding positions, and clarify the new skills and corresponding capability level requirements that support personnel need to master to advance to the next job level; Analyze the recent key tasks documents of the units to which the support personnel belong, transform them into core capability characteristics of the tasks, and clarify the core skills and capability requirements on which the units rely for the key support tasks to be performed in the near future; Analyze the new technology equipment deployment plans of the units to which the support personnel belong, transform them into the skill characteristics of the new equipment, and clarify the new skill requirements for operation, maintenance, and troubleshooting of the newly deployed equipment.

[0068] Meanwhile, the pre-trained large language model compares the structured capability label features in the standardized business dataset with a hierarchical and categorized capability indicator library item by item to determine a complete list of capability gaps for individual support personnel. This list includes all skill items that have not met the job requirements and their corresponding gaps. The pre-trained large language model then matches the aforementioned three types of transformed features with each capability gap in the capability gap list, assessing the learning urgency coefficient of the training need corresponding to each capability gap. The learning urgency coefficient has a base value of 1.0, and when a matching urgency factor is found, the coefficient is added according to rules. A higher final coefficient value indicates a more urgent training need for that capability gap. Taking the "fault troubleshooting capability - XX model radar - ZZ type signal interference fault handling" gap of personnel number BZ007 as an example: Personnel number BZ007 has a radar support position promotion plan, and the future capability mapping characteristics of the corresponding position match this capability gap, with an added learning urgency coefficient of 1.2 times; Personnel number BZ007's unit is focusing on radar support tasks in electromagnetic countermeasures environment this quarter, and the corresponding core capability characteristics of the task match this capability gap, with an added learning urgency coefficient of 1.3 times; Personnel number BZ007's unit has no new equipment deployment plan of the same type this year, and the corresponding new equipment skill characteristics do not have an added effect, with a base coefficient of 1.0 times. After considering the three factors, the final learning urgency coefficient for this capability gap is calculated as follows: Learning Urgency Coefficient = 1.2 × 1.3 × 1.0 = 1.56. The pre-trained large language model multiplies the learning urgency coefficient with the assurance weight value generated in step S1031 to correct the assurance weight value corresponding to the initial business impact level, thus calculating the target business impact weight value. The calculation formula is: Target Business Impact Weight Value = Assurance Weight Value × Learning Urgency Coefficient. Substituting the corresponding value for "Fault Troubleshooting Capability - XX Model Radar - ZZ Type Signal Interference Fault Handling" into the formula, we get: Target Business Impact Weight Value = 0.72 × 1.56 = 1.1232. The pre-trained large language model updates the impact level classification standard based on the corrected target business impact weight value. In addition to the original low, medium, and high levels, an extremely high target business impact level is added, corresponding to a target business impact weight value greater than 1.0. Therefore, the target business impact level corresponding to this capability gap is determined to be an extremely high target business impact level. After all capability gaps have been corrected and calculated, a set of corrected target business impact levels for each individual support personnel is generated.

[0069] S1033, integrate and map the target business impact level with the structured capability label features in the standardized business dataset and the hierarchical classification capability indicator library, and output the comprehensive evaluation index that includes personnel skill status, job standard requirements and business impact level.

[0070] Specifically, the pre-trained large language model maps the target business impact level set to the structured capability label features and hierarchical classification capability indicator library in the standardized business dataset in a one-to-one correspondence. This ensures that each capability indicator in the hierarchical classification capability indicator library matches the actual skill status represented by the corresponding structured capability label feature and the corresponding target business impact level, forming a one-to-one mapping relationship.

[0071] After the integration and mapping are completed, the data preprocessing and indicator selection module outputs standardized comprehensive evaluation indicators, which include three core components: Personnel skill status: all structured competency tags, quantitative scores and qualitative competency descriptions for each sub-skill, and a complete list of competency gaps; Job standards requirements: A complete hierarchical and categorized competency indicator library, and corresponding job level standards for each skill item; Business Impact Level: The guarantee weight value, learning urgency coefficient, target business impact weight value, and final target business impact level for each capability item.

[0072] The comprehensive evaluation indicators are transmitted to the scheduling and collaborative control module through a standardized data interface, serving as input data for the scheduling and collaborative control module.

[0073] The scheduling and collaborative control module, connected to the data preprocessing and indicator selection module, is used to receive the comprehensive evaluation indicators, parse the comprehensive evaluation indicators to form algorithm configuration parameters, and generate control instructions with algorithm configuration parameters and corresponding business data streams.

[0074] Specifically, the scheduling and coordination control module is configured as follows: S201, Receive the comprehensive evaluation index, call the pre-trained large language model to parse the comprehensive evaluation index, and extract the scene features and data attributes corresponding to the current task.

[0075] Specifically, the scheduling and collaborative control module receives the comprehensive evaluation index output by the data preprocessing and index selection module. The scheduling and collaborative control module calls the pre-trained large language model to perform full semantic parsing on the comprehensive evaluation index and extract the scene features and data attributes corresponding to the task generated by the current training scheme.

[0076] The extraction dimensions of scenario features may include: the scale of support personnel covered by this task, the complexity of training resource constraints, the time sensitivity requirements for solution generation, and whether there are cross-domain adaptability requirements across support units; the extraction dimensions of data attributes may include: the degree of structuring of personnel capability data, the proportion of time-series operation data to the total data volume, the proportion of unstructured text data to the total data volume, and the overall quality level of the full data.

[0077] For example, taking the task of generating a cross-unit adaptation training plan for 32 support personnel of a radar support brigade as an example, the scenario characteristics obtained by the pre-trained large language model are: the scale of support personnel is 32 people, the training resource constraints include three types of hard constraints: the number of instructors, the duration of the training site occupation, and the available equipment time period, the requirement to complete the generation of all plans within 4 hours, and the existence of cross-domain adaptability requirements across support units; the data attributes obtained by the analysis are: 60% structured digital data, 25% time-series operation data, 15% unstructured text data, and the overall quality level of the full data is good.

[0078] S202, based on the scene characteristics and the data attributes, determine the functional sub-modules to be called for the current task, and generate algorithm configuration parameters adapted to the current task by combining the operation requirements of the functional sub-modules and the analysis results of the comprehensive evaluation index.

[0079] Specifically, the pre-trained large language model determines all the functional sub-modules to be called for the current task based on the extracted scene features and data attributes, combined with the applicable scenarios and computational requirements of each downstream functional sub-module, and generates algorithm configuration parameters adapted to the current task.

[0080] Functional submodule determination process: The pre-trained large language model makes the following judgments based on scene features and data attributes: This task requires global optimization of the training scheme, which necessitates calling a personalized scheme generation engine based on TGCOA. This task has cross-domain adaptability requirements across support units, which requires calling the MDAN-driven cross-domain adaptive and effect iteration module; In this mission, timing operation data accounted for 25%, involving the feature extraction requirements of equipment operation timing and fault handling timing, which required calling the bidirectional long short-term memory network submodule as the feature processing unit. This task requires prioritizing the matching of courses with skill gaps, necessitating the use of the attention mechanism network submodule as the weight allocation unit.

[0081] Among them, the bidirectional long short-term memory network submodule and the attention mechanism network submodule are both pre-set modules.

[0082] The pre-trained large language model simultaneously determines the calling logic of each functional sub-module: the bidirectional long short-term memory network sub-module and the attention mechanism network sub-module are the front-end feature processing units, which are serially connected to the input end of the personalized scheme generation engine based on TGCOA; the MDAN-driven cross-domain adaptation and effect iteration module is the back-end adaptation unit, which receives the output results of the personalized scheme generation engine based on TGCOA.

[0083] The pre-trained large language model, combined with the analysis results of comprehensive evaluation indicators, generates corresponding running parameters for each functional sub-module. All parameters together constitute the complete algorithm configuration parameters.

[0084] Configuration parameters for the TGCOA-based personalized solution generation engine include: swarm size, scout bee ratio, local search step size, information exchange intensity, and maximum iteration count. For example, the pre-trained large language model, based on a user base of 32 and the complexity of the three types of hard constraints, sets the initial swarm size to 120, the scout bee ratio to 15%, the local search step size to 0.08, the information exchange intensity to 0.7, and the maximum iteration count to 200.

[0085] Configuration parameters for the MDAN-driven cross-domain adaptation and performance iteration module include: domain offset detection threshold, initial blending ratio, pseudo-label confidence threshold, and progressive blending ratio adjustment strategy. The progressive blending ratio adjustment strategy is formulated by the pre-trained large language model based on Wasserstein distance, and the corresponding calculation formula is as follows: The first step is to calculate the weighting factors corresponding to the distribution differences between the source domain and the mixed domain, and between the target domain and the mixed domain. The formula is as follows: ; Where q is the source domain similarity weighting factor, used to characterize the similarity between the mixed domain and the source domain; The Wasserstein distance between the source domain data distribution and the mixed domain data distribution; The Wasserstein distance between the target domain data distribution and the mixed domain data distribution; This is a temperature constant, which can be fixed at 0.05, used to control the sensitivity of the weighting factor.

[0086] The second step is to calculate the baseline value of the mixing ratio in the nth iteration, using the following formula: ; in, This is the baseline value for the mixing ratio in the nth iteration; This is the index of the current iteration round; This represents the total number of iterations. This is the baseline value for the mixing ratio in the (n-1)th iteration; This is a weighting factor for source domain similarity.

[0087] The third step is to add a random perturbation to the mixing ratio baseline value and apply range clamping, using the following formula: ; in, This is the final mixing ratio value used in the nth iteration; The sampling function is a uniform distribution. This represents the local perturbation range, with a fixed value of 0.2, used to improve the stability of the training process. This is a numerical clamping function used to constrain the mixing ratio within a legal range of 0 to 1.

[0088] In this way, the pre-trained large language model uses the above-mentioned mixing ratio adjustment rules, initial mixing ratio (e.g., 0.3), and pseudo-label confidence threshold (e.g., 0.75) as configuration parameters for the MDAN-driven cross-domain adaptation and effect iteration module.

[0089] Configuration parameters for the bidirectional long short-term memory network submodule and the attention mechanism network submodule include the number of network layers, the number of hidden layer nodes, and weight allocation coefficients. The pre-trained large language model sets corresponding initial parameters based on the length of the temporal data and the number of matching dimensions.

[0090] S203, generate control instructions carrying corresponding parameters and matching business data streams according to the algorithm configuration parameters, and distribute the control instructions and business data streams to the TGCOA-based personalized solution generation engine and the MDAN-driven cross-domain adaptive and effect iteration module, respectively.

[0091] Based on the generated algorithm configuration parameters, the pre-trained large language model generates control instructions carrying complete parameter information, as well as business data streams matching the control instructions. The control instructions are used to trigger the start and operation of the corresponding functional sub-modules, specifying the sub-module's operating rules, parameter requirements, and execution order; the business data streams are used to provide all the input data required for the computation of each functional sub-module.

[0092] The scheduling and collaborative control module distributes control commands and corresponding business data streams to the TGCOA-based personalized solution generation engine and the MDAN-driven cross-domain adaptive and effect iteration module, respectively, according to the preset module connection relationship.

[0093] The control instructions distributed to the TGCOA-based personalized solution generation engine include: engine startup instructions, all TGCOA algorithm operation parameters, and calling rules and parameters for the bidirectional long short-term memory network submodule and the attention mechanism network submodule; the corresponding business data stream includes: personnel skill status data, job standard requirement data, business impact level data, and training resource constraint boundary data in the comprehensive evaluation indicators.

[0094] The control instructions distributed to the MDAN-driven cross-domain adaptation and effect iteration module include: module start instructions, mixing ratio adjustment strategy parameters, and pseudo-label confidence threshold parameters; the corresponding business data stream includes: preset source domain labeled training sample data, target domain task environment data to be adapted, and output data docking rules based on the TGCOA-based personalized solution generation engine.

[0095] S204, receive the optimization signal fed back by the result analysis, interpretation and dynamic iterative optimization mechanism, adjust the algorithm configuration parameters according to the optimization signal, and update the corresponding control instructions and business data streams before outputting them to each functional module.

[0096] Specifically, the scheduling and collaborative control module establishes a data connection with the result analysis, interpretation, and dynamic iterative optimization mechanism, and receives optimization signals fed back from the mechanism. The optimization signals include four types of information: scheme quality assessment results, cross-domain adaptation effects, module operating efficiency, and abnormal operating states.

[0097] The scheduling and collaborative control module calls the pre-trained large language model to analyze the optimization signal, locate the functional sub-modules and parameter items corresponding to performance deviations, and make targeted adjustments to the original algorithm configuration parameters. For example, when the optimization signal shows that the cross-domain adaptation effect is not as expected and the target domain solution matching degree is low, the pre-trained large language model determines that the mixing ratio adjustment rhythm of the MDAN-driven cross-domain adaptation and effect iteration module is too fast and the pseudo-label filtering threshold is too low. It then adjusts the iteration step size of the gradual adjustment of the mixing ratio and raises the pseudo-label confidence threshold to 0.8. When the optimization signal shows that the solution optimization convergence speed is too slow and does not meet the time requirements, the pre-trained large language model determines that the global exploration intensity of the personalized solution generation engine based on TGCOA is insufficient. It then adjusts the bee colony size parameter and the reconnaissance bee ratio parameter to improve the global optimization efficiency.

[0098] After the algorithm configuration parameters are adjusted, the pre-trained large language model synchronously updates the corresponding control instructions and business data streams. The updated control instructions and business data streams are then redistributed to each functional sub-module, driving the entire chain to execute the dynamically adjusted computation process and achieving closed-loop dynamic optimization of scheduling and collaborative control.

[0099] A personalized scheme generation engine based on TGCOA is connected to the scheduling and collaborative control module. It is used to respond to the control commands, receive the business data stream, and within the preset training resource constraints, apply a global optimization algorithm to optimize according to the algorithm configuration parameters, generate an initial training scheme, and feed it back to the scheduling and collaborative control module.

[0100] Specifically, the personalized solution generation engine based on TGCOA is configured as follows: S301, responding to the control command, receiving the corresponding business data stream, configuring the algorithm running parameters according to the algorithm configuration parameters carried by the control command, extracting personnel skill status, job standard requirements, and business impact level from the comprehensive evaluation index, and constructing a candidate solution space for training schemes and an adaptation fitness evaluation function in combination with preset training resource constraint boundaries.

[0101] Specifically, the personalized solution generation engine based on TGCOA responds to control commands issued by the scheduling and collaborative control module and receives business data streams that match the control commands. The engine first configures all the running parameters of the stingless bee optimization algorithm according to the algorithm configuration parameters carried in the control commands. The core parameters include colony size, reconnaissance bee ratio, local search step size, information exchange intensity, and maximum number of iterations. All parameter values ​​are generated by the scheduling and collaborative control module based on the task scenario.

[0102] Subsequently, from the comprehensive evaluation indicators contained in the business data flow, three types of core data are extracted: personnel skill status, job standard requirements, and business impact level. Combined with the preset training resource constraints, including hard constraints such as the upper limit of the number of instructors, the available time of training venues, the time limit of equipment operation, and the total number of courses that can be offered, a candidate solution space for training schemes is constructed, and an appropriate fitness evaluation function is constructed.

[0103] The training scheme candidate solution space is defined as follows: each candidate solution in the solution space corresponds to a complete candidate training scheme. The scheme components include five categories of content: course combination matched for each support personnel, learning sequence arrangement, learning format setting, time schedule arrangement, and supporting resource configuration. All legal candidate schemes that satisfy the training resource constraint boundary together constitute the training scheme candidate solution space.

[0104] The fitness evaluation function is used to quantitatively assess the quality of each candidate training scheme. Taking into account five recorded evaluation dimensions, the formula for calculating the overall fitness value is: ; Among them, F is the overall adaptability value of the candidate training program, and the higher the value, the better the overall quality of the program; C is the capability coverage, which represents the degree to which the candidate program makes up for the personnel's capability shortcomings; M is the resource matching degree, which represents the degree to which the course content and difficulty level of the candidate program match the personnel's current skill level; E is efficiency and cost, which represents the rationality of the total training time and core resource usage of the candidate program; P is business priority, which represents the degree to which the candidate program prioritizes and meets the training needs with high business impact and high urgency; and D is personal development fit, which represents the degree to which the candidate program matches the career development aspirations of the support personnel. , , , , The weighting coefficients for the corresponding evaluation dimensions can be uniformly set by the preset equipment support business logic.

[0105] Furthermore, the calculation of the comprehensive fitness value adopts a preset scoring logic: the system pre-sets base scores and weights for each evaluation dimension. For example, regarding capability coverage, the system iterates through all courses in the candidate schemes, accumulating the corresponding base score for each capability gap covered; regarding business priority, the system verifies the time arrangement order of the candidate schemes, and if a skill deficiency with a high target business impact level is identified as being placed at the beginning of the scheme (e.g., within the first 30% of class time), a bonus reward rule is triggered. Finally, the scores of each dimension are summarized according to the preset business weights to obtain a quantified comprehensive score as the comprehensive fitness value.

[0106] S302, a preset number of candidate training schemes are randomly generated in the candidate solution space of the training scheme as stingless bee individuals to complete the bee colony initialization. The fitness value corresponding to each stingless bee individual is calculated based on the fitness evaluation function to obtain the initial bee colony set and the initial global optimal candidate training scheme.

[0107] Specifically, after the algorithm's runtime parameters are configured, the personalized solution generation engine based on TGCOA randomly generates candidate training schemes with the same number of individuals as the hive size N within the candidate solution space. Each candidate training scheme corresponds to one stingless bee individual, completing the hive initialization operation. Taking this radar support post training task with a scale of 32 people as an example, the hive size is configured as 120, that is, 120 independent candidate training schemes are generated as the initial set of stingless bee individuals.

[0108] After the bee colony initialization is completed, based on the constructed fitness evaluation function, the comprehensive fitness value of each stingless bee individual is calculated one by one, and the fitness calculation results of all individuals are recorded to obtain the initial bee colony set. At the same time, the candidate training scheme with the highest comprehensive fitness value is selected from the initial bee colony set and determined as the initial globally optimal candidate training scheme, which serves as the benchmark scheme for subsequent iterative optimization.

[0109] S303, the initial bee colony set is divided into scout bees, foraging bees, and follower bees according to a preset division of labor strategy. Using the initial globally optimal candidate training scheme as the iterative benchmark, multi-role collaborative optimization is performed. The scout bees are scheduled to perform a global random search within the candidate solution space of the training scheme and generate new candidate training schemes. The foraging bees are scheduled to perform a local search around the currently better candidate training scheme and adjust the components of the candidate training scheme to optimize the fitness value. The follower bees are scheduled to follow the search in the direction of a high-quality candidate training scheme based on the population information sharing mechanism. After completing the fitness verification of all candidate training schemes based on the fitness evaluation function, the updated bee colony set and the globally optimal candidate training scheme for this iteration are obtained.

[0110] Specifically, the personalized solution generation engine based on TGCOA divides stingless bees in the initial bee colony set into three categories: scout bees, foraging bees, and follower bees according to a preset division of labor strategy. It then uses the initial global optimal candidate training scheme as the iterative benchmark and executes a multi-role collaborative optimization process.

[0111] Recon Bee Global Random Search: Based on TGCOA's personalized solution generation engine, all recon bees are scheduled to conduct a global random search within the candidate solution space of training solutions. This attempts to introduce entirely new course combinations and disruptive learning path arrangements to explore unknown areas of the solution space and generate entirely new candidate training solutions, avoiding the optimization process from getting stuck in the local optima corresponding to traditional fixed course packages. For example, recon bees can break through the conventional course arrangement logic, prioritizing practical courses corresponding to high-priority weak skills to generate differentiated candidate solutions.

[0112] Local search optimization by the honeybee: Based on the personalized solution generation engine of TGCOA, all honeybees are scheduled to conduct local searches around candidate training schemes with the best overall fitness value. The fitness value is optimized by fine-tuning the components of the candidate training schemes. The adjustment operations include: the system randomly selects two course nodes in the course sequence of the candidate training scheme and directly swaps their time arrangement; or the system searches the resource library for alternative courses with the same knowledge point tags as a current long course, but with lower difficulty or time consumption, and performs node replacement operations in the candidate schemes. Through such specific list adjustment and replacement actions, the accuracy of the local optimum is improved without compromising the validity of the solution.

[0113] Follow-up bee search: Based on TGCOA's personalized solution generation engine, all follower bees are scheduled to acquire fitness information for all candidate solutions through a population information sharing mechanism. At this stage, the system sorts all current candidate training solutions in descending order of their overall fitness value and selects the top 20% (e.g., the top percentage) as a pool of high-quality candidate solutions. Subsequently, the follower bees randomly select a target solution from this pool and follow the direction of that high-quality solution in their search, thereby accelerating the propagation of high-quality solution patterns within the population without requiring complex probability calculations.

[0114] After completing the search operations for all three types of bees in this round, the personalized scheme generation engine based on TGCOA performs fitness verification on all newly generated candidate training schemes based on the fitness evaluation function, updates the corresponding schemes and fitness values ​​of all stingless bee individuals, obtains the bee colony set updated in this round, and selects the globally optimal candidate training scheme corresponding to this iteration.

[0115] S304. After each round of optimization, based on the calculation result of the fitness evaluation function, all candidate training schemes in the updated bee colony set are screened, and new random candidate training schemes are added to the candidate solution space of the training scheme to maintain population diversity. It is then determined whether the preset termination condition has been met. If not, the multi-role collaborative optimization is returned. If it has been met, an initial training scheme is generated and fed back to the scheduling and collaborative control module.

[0116] Specifically, after each round of optimization, the TGCOA-based personalized solution generation engine filters all candidate training schemes in the updated bee colony set based on the fitness evaluation function calculation results. It retains high-quality candidate schemes with high overall fitness values ​​and eliminates low-quality candidate schemes with low overall fitness values. Simultaneously, the TGCOA-based personalized solution generation engine supplements the candidate solution space with entirely new random candidate training schemes, maintaining a consistent number of additions and eliminations to preserve bee colony diversity and prevent premature convergence during the optimization process.

[0117] After the population update is completed, the TGCOA-based personalized solution generation engine determines whether the current optimization process has reached the preset termination conditions. The preset termination conditions include two types: the first is that the number of iterations reaches the preset maximum number of iterations; the second is that the overall fitness value of the globally optimal candidate training scheme has not shown a significant improvement for multiple consecutive generations, i.e., the convergence condition is met. If the preset termination conditions are not met, the multi-role collaborative optimization step is returned to start the next iteration; if the preset termination conditions are met, the current globally optimal candidate training scheme is determined as the final initial training scheme, and the TGCOA-based personalized solution generation engine feeds back the initial training scheme to the scheduling and collaborative control module, completing the entire execution process of the TGCOA-based personalized solution generation engine.

[0118] The MDAN-driven cross-domain adaptation and effect iteration module is connected to the scheduling and collaborative control module. It is used to respond to the control command, receive the initial training scheme in the business data stream, introduce the target domain task environment data to be adapted, apply the domain adaptation network to perform feature cross-domain alignment and correction, and output the target training scheme.

[0119] Specifically, the MDAN-driven cross-domain adaptive and effect iteration module is configured as follows: S401, in response to the control command, receive the corresponding business data stream, obtain the initial training scheme from the business data stream, introduce the target domain task environment data to be adapted, and retrieve the preset source domain labeled training sample data to form cross-domain adaptation data preparation.

[0120] Specifically, the MDAN-driven cross-domain adaptation and effect iteration module responds to control commands issued by the scheduling and collaborative control module and receives business data streams that match the control commands. The MDAN-driven cross-domain adaptation and effect iteration module obtains the initial training scheme output by the personalized scheme generation engine based on TGCOA from the business data stream, and at the same time introduces the target domain task environment data to be adapted, and retrieves the preset source domain labeled training sample data to complete all data preparation work for cross-domain adaptation.

[0121] For example, the source domain labeled training sample data can be taken from the data-rich pilot radar support brigade, which can contain complete personnel capability data, training plan data, and training effect labeling data, and has complete supervision labels; the target domain mission environment data to be adapted is taken from the newly established radar support unit. The equipment system, mission mode, and personnel base of the newly established radar support unit are significantly different from those of the pilot radar support brigade; the newly established radar support unit only has personnel capability data and mission environment data, but does not have the labeled training effect data. Directly applying the model trained in the source domain will result in a domain offset problem where the generated plan is out of touch with the actual needs.

[0122] S402, a feature extraction network is trained based on the source domain labeled training sample data. During the training process, a hybrid regularization strategy is adopted to enhance the generalization of the feature space. Self-supervised training is performed by combining controlled reconstruction learning to obtain the source domain pre-trained feature extraction model.

[0123] Specifically, the MDAN-driven cross-domain adaptive and performance iteration module trains the feature extraction network based on source domain labeled training sample data. The feature extraction network adopts a five-layer bidirectional long short-term memory network structure. During the training process, supervised learning and self-supervised learning are performed simultaneously. Supervised learning uses a hybrid regularization strategy to enhance the generalization of the feature space, while self-supervised learning uses a controlled reconstruction learning strategy to improve the transferability of features. Finally, a source domain pre-trained feature extraction model is obtained.

[0124] The loss calculation rules for the training process are as follows: 1. Source domain original sample supervision loss Supervised learning is performed using the mean squared error loss function on the original labeled samples in the source domain. The calculation formula is as follows: ; in, This represents the supervised loss value for the original samples in the source domain. y represents the source domain data distribution; x represents the input source domain sample data; y represents the true label corresponding to the source domain sample. For parameters Feature extraction network; For parameters Prediction network; This is the function for calculating the mean squared error loss; The expectation operator is a mathematical operator that represents the calculation of the statistical expectation under a specified data distribution, that is, the average of the calculation results of all samples in the corresponding distribution. It is a standard mathematical symbol in the fields of probability statistics and machine learning.

[0125] 2. Source domain mixed sample regularization loss At the feature level, a mixing operation is performed on the source domain samples to generate mixed feature samples and corresponding mixed labels. Supervised learning of the mixed samples is used to achieve mixing regularization, enriching the intrinsic structure of the latent feature space. The calculation formula is as follows: ; ; ; in, and These are two randomly selected source domain samples; , These are the true labels corresponding to the two source domain samples; The mixing coefficients are randomly sampled from the beta distribution; This represents the hybrid feature representation; The mixed label; This represents the supervised loss value for mixed samples from the source domain.

[0126] 3. Controlled reconstruction of self-supervised loss Partial occlusion of source domain temporal samples is achieved through masking operations, followed by controlled reconstruction learning. This introduces a self-supervised signal to prevent supervised collapse and improve feature transferability. The controlled reconstruction loss consists of two parts: masking loss and unmasked loss, calculated using the following formula: ; ; ; in, For the binary mask at time step t, the input data at the mask position is set to zero; This represents the reconstruction loss value at the mask location; This represents the reconstruction loss value at the unmasked location; To balance the parameters, random sampling is performed from a uniform distribution; This represents the total loss value of controlled reconstruction.

[0127] 4. Overall training objective of the source domain The overall optimization objective of source domain training is to minimize the weighted sum of the three types of losses mentioned above, calculated using the following formula: ; in, This is a tradeoff constant for the hybrid regularization loss; For the trade-off constant of controlled reconstruction loss, The set of all trainable parameters for the prediction network; This is the complete set of trainable parameters for the feature extraction network.

[0128] After completing all iterations of training, the source domain pre-trained feature extraction model is obtained.

[0129] S403, the source domain pre-trained feature extraction model is invoked to predict the target domain task environment data to be adapted, generating corresponding pseudo-labels. The mixing ratio of the source domain labeled training sample data and the target domain task environment data to be adapted is adjusted using Wasserstein distance as a distribution difference metric. In the input layer and feature layer of the source domain pre-trained feature extraction model, intermediate mixed domains formed by proportional mixing of the two domains are constructed respectively. Progressive feature alignment is performed to gradually reduce the distribution difference between the two domains to obtain the intermediate domain adaptation feature representation.

[0130] Specifically, the MDAN-driven cross-domain adaptation and effect iteration module calls the source domain pre-trained feature extraction model to predict the target domain task environment data to be adapted, generating corresponding pseudo-labels. Using the Wasserstein distance as a distribution difference metric, the MDAN-driven cross-domain adaptation and effect iteration module dynamically adjusts the mixing ratio of the source domain labeled training sample data and the target domain task environment data to be adapted. It constructs intermediate mixed domains formed by proportionally mixing the data from the two domains in the input and feature layers of the source domain pre-trained feature extraction model, performs progressive feature alignment, gradually reducing the distribution difference between the two domains, and finally obtains the intermediate domain adapted feature representation.

[0131] The rules for constructing the intermediate blending domain and dynamically adjusting the blending ratio are as follows: 1. Construction of mixed samples from input layer and feature layer Linear interpolation is performed on the source domain samples and target domain samples at both the input and feature levels to generate mixed samples and corresponding mixed labels. The calculation formula is as follows: Input layer mixing: ; ; Feature-level blending: ; ; in, For source domain samples, The source domain's true label; For the target domain sample, These are the pseudo-labels corresponding to the target domain samples; The value represents the mixing ratio; the larger the value, the closer the mixed samples are to the source domain, and the smaller the value, the closer they are to the target domain.

[0132] 2. Gradual mixing ratio adjustment based on Wassstein distance A progressive blending strategy is adopted to dynamically adjust the blending ratio. In the initial stage, the intermediate blending domain is closer to the target domain, and it gradually moves closer to the source domain to achieve smooth domain alignment.

[0133] The first step is to calculate the source domain similarity weighting factor, using the following formula: ; The second step is to calculate the baseline value of the mixing ratio in the nth iteration, using the following formula: ; The third step is to add a random perturbation to the mixing ratio baseline value and apply range clamping, using the following formula: ; Steps one through three can be referred to in the description of step S202, and will not be repeated here; The training loss of the intermediate mixing domain is the sum of the supervision losses of the input layer mixed samples and the feature layer mixed samples, calculated as follows: ; The overall optimization objective for intermediate domain training is: ; in, This is a tradeoff constant for the training loss in the intermediate domain.

[0134] After completing the training of the intermediate domain, the intermediate domain adaptation feature representation is obtained.

[0135] S404, based on the intermediate domain adaptation feature representation, self-learning training is performed on the target domain task environment data to be adapted using the pseudo-labels as supervision information. In the feature space corresponding to the target domain task environment data to be adapted, a hybrid regularization strategy is used to construct consistency constraints, filter low-confidence pseudo-labels, and obtain the target domain optimized feature extraction model.

[0136] Specifically, the MDAN-driven cross-domain adaptation and performance iteration module performs self-learning training on the target domain task environment data to be adapted, based on the intermediate domain adaptation feature representation and using pseudo-labels as supervision information. The MDAN-driven cross-domain adaptation and performance iteration module employs a hybrid regularization strategy to construct consistency constraints in the feature space corresponding to the target domain task environment data to be adapted, filtering out low-confidence pseudo-labels with confidence levels below a preset threshold, suppressing the negative impact of pseudo-label noise on the training process, and finally obtaining the target domain optimized feature extraction model.

[0137] The loss calculation rules for target domain training are as follows: 1. Self-learning loss of original samples in the target domain Using pseudo-labels as supervisory information, self-learning is performed on the original samples of the target domain. The calculation formula is as follows: ; In the formula: The self-learning loss value is the value of the original samples in the target domain. Data distribution for the target domain; The pseudo-labels are the samples corresponding to the target domain. Only pseudo-labels with a confidence level higher than a preset threshold are retained for training.

[0138] 2. Consistency loss of mixed samples in the target domain A fusion operation is performed in the target domain feature space to constrain the consistency between the original samples and the fused samples output. Fusion regularization is used as a consistency regularization method to reduce the interference of false label noise. The calculation formula is as follows: ; In the formula, This represents the consistency loss value for mixed samples in the target domain; These are the mixed pseudo-labels corresponding to the mixed samples in the target domain.

[0139] 3. Overall training objective for the target domain The overall optimization objective of target domain training is to minimize the weighted sum of the two types of losses mentioned above, calculated using the following formula: ; In the formula: This is a tradeoff constant for the self-learning loss in the target domain; is a tradeoff constant for the target domain hybrid regularization loss.

[0140] After completing all iterative training in the target domain, the optimized feature extraction model for the target domain is obtained.

[0141] S405, the initial training scheme is modified for cross-domain features and adapted to the scene by using the feature extraction model optimized for the target domain, thereby generating the target training scheme.

[0142] Specifically, the MDAN-driven cross-domain adaptation and effect iteration module uses a feature extraction model optimized for the target domain to perform cross-domain feature correction and scenario adaptation adjustments on the initial training scheme. These adjustments include: adjusting the course knowledge points and practical content of the training scheme according to the equipment system of the support unit corresponding to the target domain task environment data; adjusting the course difficulty gradient and learning progress arrangement according to the personnel foundation of the support unit corresponding to the target domain task environment data; and adjusting the training focus and high-priority skill arrangement according to the task mode of the support unit corresponding to the target domain task environment data, ensuring the training scheme fully adapts to the actual support needs of the target domain. After completing all adaptation adjustments, the MDAN-driven cross-domain adaptation and effect iteration module generates the final target training scheme and outputs it to the downstream result analysis, interpretation, and dynamic iterative optimization mechanism.

[0143] The result analysis, interpretation, and dynamic iterative optimization mechanism has an input end connected to the MDAN-driven cross-domain adaptive and effect iteration module, and an output end connected to the scheduling and collaborative control module. It is used to perform interpretability analysis on the target training scheme, and combine the obtained offline implementation effect feedback data of the target training scheme to form an optimization signal. The optimization signal is then fed back to the scheduling and collaborative control module to drive the scheduling and collaborative control module to dynamically adjust the control commands and the business data stream.

[0144] Specifically, the result analysis, interpretation, and dynamic iterative optimization mechanism is configured as follows: S501, receive the target training scheme, obtain the offline implementation effect feedback data corresponding to the target training scheme, call the pre-trained large language model to perform semantic parsing and structural transformation on the unstructured content in the offline implementation effect feedback data, filter out abnormal data, and match and align the processed offline implementation effect feedback data with the personnel skill status, job standard requirements and business impact level corresponding to the target training scheme one by one to obtain standardized effect feedback data and corresponding analysis benchmarks.

[0145] Specifically, the results analysis, interpretation, and dynamic iterative optimization mechanism receives the target training scheme output by the MDAN-driven cross-domain adaptive and effect iteration module, and simultaneously acquires offline implementation effect feedback data generated after the target training scheme is implemented. The offline implementation effect feedback data includes structured assessment scores, skills evaluation results, and task execution success rates, as well as two types of unstructured content: student learning feedback texts, instructor teaching evaluation texts, and task execution summary texts.

[0146] The results analysis, interpretation, and dynamic iterative optimization mechanism invokes a pre-trained large language model to perform semantic parsing and structural transformation of unstructured content in the feedback data on offline implementation effects. Specifically, the large model uses entity relation extraction and sentiment analysis mechanisms to identify key entities in the text (such as specific courses, instructors, and equipment) and their corresponding sentiment tendencies (such as positive, negative, and neutral). For example, the unstructured feedback that the practical training time for the new radar was too short and not fully mastered is accurately transformed into a unified format of structured tags: [Entity: New Radar Practical Training Course], [Attribute: Training Duration], [Status: Insufficient], and [Satisfaction: Negative]. After the structural transformation is completed, the results analysis, interpretation, and dynamic iterative optimization mechanism performs anomaly filtering, eliminating invalid feedback data with severe data gaps, obvious logical contradictions, or non-compliant collection processes, ensuring the authenticity and reliability of the feedback data.

[0147] Subsequently, the results analysis, interpretation, and dynamic iterative optimization mechanism matches the processed offline implementation effect feedback data with the personnel skill status, job standard requirements, and business impact level corresponding to the target training plan. Using the unique identifier of each support personnel as the primary key, it ensures that each piece of feedback data corresponds to a clearly defined individual support personnel, capability item, and training plan content, ultimately yielding standardized effect feedback data and corresponding analysis benchmarks. These benchmarks include the target training plan's preset capability improvement goals, resource allocation standards, expected business impact, and cross-domain adaptation requirements, serving as a reference for subsequent effect verification.

[0148] S502, based on the aforementioned analysis benchmark, the pre-trained large language model is invoked to verify and interpret the target training scheme from four dimensions: capability coverage, resource matching status, expected business impact, and cross-domain adaptation effect, generating a structured scheme interpretability report.

[0149] Specifically, the results analysis, interpretation, and dynamic iterative optimization mechanism, based on the analysis benchmark, calls the pre-trained large language model to comprehensively verify and interpret the target training scheme from four dimensions: capability coverage, resource matching status, expected business impact, and cross-domain adaptation effect, generating a structured scheme interpretability report.

[0150] First, an interpretation of the capability coverage dimension. The pre-trained large language model is compared with the hierarchical and categorized capability indicator library and the personnel capability gap list to verify the coverage and depth of the course content and practical training projects of the target training plan for all capability gaps. It is clearly stated which core skills are covered by the target training plan, which skills are only covered at a basic level, and which skills are not covered. At the same time, the training time ratio and expected capability improvement level of each skill are marked.

[0151] Second, the resource matching status dimension is interpreted. The pre-trained large language model, combined with the feature vector of training resources, verifies the degree of matching between the course difficulty, teaching style, practical training equipment and the current skill level and learning acceptance ability of the support personnel in the target training plan. It explains the rationality of resource allocation and the existing adaptation deviations, such as the insufficient suitability of high-difficulty practical courses for support personnel with weak foundations.

[0152] Third, interpretation of expected business impact. The pre-trained large language model, combined with business impact level data, analyzes the expected improvement in the success rate of various support tasks after the implementation of the target training plan. It clarifies the expected improvement effects for high-priority, high-business-impact skills and quantifies the value of the target training plan in enhancing the overall support capabilities of the organization.

[0153] Fourth, an interpretation of cross-domain adaptation effectiveness. The pre-trained large language model, considering the environmental differences between the source and target domains, analyzes the adaptation adjustments made by the target training scheme for the target domain's equipment system, task mode, and personnel base. This explains the rationality of the adaptation adjustments and the expected degree of adaptation, verifying whether the cross-domain adjustments effectively solved the domain offset problem.

[0154] The interpretations from all four dimensions are organized into standardized, structured items, which together constitute a structured interpretability report. The report includes both quantitative indicators and qualitative descriptions.

[0155] S503, using the expected indicators of each dimension in the structured scheme interpretability report as a comparison benchmark, compare the standardized effect feedback data with the corresponding expected effect, perform effect attribution analysis, locate the scheme deviation points and the causes of performance changes, generate effect evaluation results, and convert them into optimization signals to be fed back to the scheduling and collaborative control module. The optimization signals include algorithm configuration parameter adjustment instructions and scheme generation logic correction directions.

[0156] Specifically, the results analysis, interpretation, and dynamic iterative optimization mechanism uses the expected indicators of each dimension in the structured scheme interpretability report as the comparison benchmark. It compares the actual implementation effect in the standardized effect feedback data with the corresponding expected effect item by item, and calculates the deviation between the actual value and the expected value of each indicator.

[0157] Based on this, the results are analyzed, interpreted, and the dynamic iterative optimization mechanism is used for attribution analysis. The underlying logic of the attribution analysis is as follows: the pre-trained large language model extracts the target indicators that show negative deviations and traces back strongly correlated factors in the end-to-end data logs. For example, when the actual success rate of a specific task in the target domain is much lower than the expected value, the large model searches the resource configuration records and finds that the practical courses corresponding to this task account for a very low percentage of the total class hours, or finds that the environmental parameters of the target domain are too different from those of the source domain in the cross-domain configuration records. Thus, the cause of the deviation is logically deduced, and the specific reasons for the deviation and performance changes are located.

[0158] After completing the attribution analysis, the results analysis, interpretation, and dynamic iterative optimization mechanism generate complete performance evaluation results. Simultaneously, these results are transformed into structured key-value pair (Key-Value) format optimization signals, which are then fed back to the scheduling and collaborative control module. The optimization signals contain two core categories: the first is algorithm configuration parameter adjustment instructions (such as generating machine-readable instructions for {"Action":"Increase","Parameter":"TGCOA_Practical_Weight","Value":"+15%"})... The second category is the direction for correcting the solution generation logic, providing optimization directions for deviations at the underlying logic level, such as supplementing feature labels for special task scenarios in the target domain or adjusting the priority ranking rules for high-urgency skills.

[0159] After completing the attribution analysis, the results analysis, interpretation, and dynamic iterative optimization mechanism generate complete performance evaluation results. Simultaneously, these results are transformed into optimization signals and fed back to the scheduling and collaborative control module. The optimization signals contain two core categories: the first is algorithm configuration parameter adjustment instructions, providing specific parameter adjustment suggestions based on the causes of deviations, such as increasing the practical resource weight of the TGCOA-based personalized solution generation engine, adjusting the mixed-ratio iteration step size of the MDAN-driven cross-domain adaptive and performance iteration modules, and increasing the pseudo-label confidence threshold; the second is the direction of solution generation logic correction, providing optimization directions for deviations at the underlying logic level, such as supplementing feature labels for special task scenarios in the target domain and adjusting the priority ranking rules for high-urgency skills.

[0160] Once the scheduling and collaborative control module receives the optimization signal, it can adjust the algorithm configuration parameters, update the corresponding control commands and business data streams, and drive the entire link to complete the dynamic iterative optimization of the training scheme.

[0161] Example like Figure 1 As shown, a radar support brigade generated a precise training plan for 32 support personnel, strictly following the logic diagram link throughout the entire process: The process begins with data collection and preprocessing (data preprocessing and indicator selection module). This involves collecting multi-dimensional data on personnel, job positions, and training resources, including personnel files, job competency standards, course libraries, and instructor resources from the human resources system. Next, large-scale model data cleaning, fusion, structuring, and feature extraction are performed. A pre-trained large language model is used to parse unstructured task reports and fault records, converting them into structured capability labels, filling in missing features, and removing contradictory data. Finally, a dynamic indicator system is built to create capability profiles and quantify business impact, generating comprehensive evaluation indicators covering personnel skill status, job standard requirements, and business impact levels, thus completing the personnel capability profile and business impact measurement.

[0162] The processed data is output to the core intelligent generation process, where the large model intelligently schedules and dynamically calls TGCOA, MDAN, or neural networks (scheduling and collaborative control module) to analyze comprehensive evaluation indicators, determine whether the task has cross-unit adaptation requirements, generate corresponding algorithm configuration parameters, issue control commands and business data streams, and schedule subsequent algorithm modules to run collaboratively.

[0163] First, the TGOOA optimization scheme generation (a personalized scheme generation engine based on TGOOA) step is executed. Within the training resource constraint boundary, the global optimization is performed through the stingless bee colony intelligent algorithm to generate an initial training scheme that takes into account capability coverage, resource matching, and business priority.

[0164] Then, the process proceeds to the MDAN cross-domain adaptive source domain → hybrid domain → target domain (MDAN-driven cross-domain adaptive and effect iteration module). The pilot brigade's labeled data is used as the source domain, and the environmental data of the support brigade is used as the target domain. By constructing an intermediate hybrid domain, progressive feature alignment is performed to complete the cross-scene adaptation correction of the training scheme.

[0165] After the solution is generated, it enters the intelligent analysis and decision interpretation (result analysis, interpretation, and dynamic iterative optimization mechanism) stage: First, the large model results analysis and interpretation work is carried out to verify the rationality of the solution from four dimensions: capability coverage, resource matching, business impact, and cross-domain adaptation; then, a structured decision report (structured solution interpretability report) is generated to clarify the solution matching logic and expected effects; at the same time, human-computer interaction and adjustment support are provided to support managers to put forward adjustment suggestions based on actual needs.

[0166] Once the validated plan is completed, it enters the plan output and effect iteration chain. First, the personalized training plan output (course, time, resources, or assessment) (target training plan) is completed. The output includes a complete training plan containing course combinations, time arrangements, resource configurations, and assessment standards.

[0167] After the plan is implemented, training and effect tracking will be carried out, and all implementation data such as assessment results, task performance, and personnel feedback will be collected.

[0168] After tracking the data feedback, the model and parameters are optimized through a closed-loop feedback data process. The implementation effect is transformed into an optimization signal and fed back to the scheduling and collaborative control module. The algorithm parameters and scheme generation logic are dynamically adjusted to achieve continuous iterative evolution of the intelligent generation model for support personnel training schemes based on the large model-TGCOA-MDAN.

[0169] Secondly, this application provides an intelligent generation method for support personnel training schemes based on the large model-TGCOA-MDAN, applied to the aforementioned intelligent generation model for support personnel training schemes based on the large model-TGCOA-MDAN, including: S100: Acquire multi-source heterogeneous data related to support personnel, perform structured fusion of the multi-source heterogeneous data, and output a comprehensive evaluation index that includes personnel skill status, job standard requirements, and business impact level. S200: Receive the comprehensive evaluation index, parse the comprehensive evaluation index to form algorithm configuration parameters, and generate control instructions with algorithm configuration parameters and corresponding business data streams; S300, in response to the control command, receive the service data stream and within the preset training resource constraint boundary, apply a global optimization algorithm to perform optimization according to the algorithm configuration parameters, generate an initial training scheme and feed it back to the scheduling and collaborative control module; S400, in response to the control command, receive the initial training scheme in the service data stream, introduce the target domain task environment data to be adapted, apply the domain adaptation network to perform feature cross-domain alignment and correction, and output the target training scheme. S500, the interpretability analysis of the target training scheme is performed, and combined with the offline implementation effect feedback data of the target training scheme, an optimization signal is formed, and the optimization signal is fed back to the scheduling and collaborative control module to drive the scheduling and collaborative control module to dynamically adjust the control command and the business data stream.

[0170] Thirdly, this application provides a computer storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the aforementioned intelligent generation method for support personnel training schemes based on large model-TGCOA-MDAN.

[0171] Fourthly, this application provides a computer program that, when executed by a processor, implements the steps of the aforementioned intelligent generation method for support personnel training schemes based on large model-TGCOA-MDAN.

[0172] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0173] Obviously, those skilled in the art should understand that the various units or steps of this application described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by a computing device, or fabricating them separately as individual integrated circuit modules, or fabricating multiple modules or steps into a single integrated circuit module. Thus, this application is not limited to any particular combination of hardware and software.

[0174] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. An intelligent generation model for a support personnel training scheme based on the large model-TGCOA-MDAN, characterized in that, include: The data preprocessing and indicator selection module is used to acquire multi-source heterogeneous data related to support personnel, and to perform structured fusion of the multi-source heterogeneous data to output a comprehensive evaluation indicator that includes personnel skill status, job standard requirements, and business impact level. The scheduling and collaborative control module, connected to the data preprocessing and indicator selection module, is used to receive the comprehensive evaluation indicators, parse the comprehensive evaluation indicators to form algorithm configuration parameters, and generate control instructions with algorithm configuration parameters and corresponding business data streams. A personalized scheme generation engine based on TGCOA is connected to the scheduling and collaborative control module. It is used to respond to the control command, receive the business data stream, and within the preset training resource constraint boundary, apply a global optimization algorithm to optimize according to the algorithm configuration parameters, generate an initial training scheme, and feed it back to the scheduling and collaborative control module. The MDAN-driven cross-domain adaptation and effect iteration module is connected to the scheduling and collaborative control module. It is used to respond to the control command, receive the initial training scheme in the business data stream, introduce the target domain task environment data to be adapted, apply the domain adaptation network to perform feature cross-domain alignment and correction, and output the target training scheme. The result analysis, interpretation, and dynamic iterative optimization mechanism has an input end connected to the MDAN-driven cross-domain adaptive and effect iteration module, and an output end connected to the scheduling and collaborative control module. It is used to perform interpretability analysis on the target training scheme, and combine the obtained offline implementation effect feedback data of the target training scheme to form an optimization signal. The optimization signal is then fed back to the scheduling and collaborative control module to drive the scheduling and collaborative control module to dynamically adjust the control commands and the business data stream.

2. The intelligent generation model for the support personnel training scheme based on the large model-TGCOA-MDAN as described in claim 1, characterized in that, The data preprocessing and indicator selection module is configured as follows: The system acquires multi-source heterogeneous business data streams containing text, numbers, and time-series data related to support personnel, and calls a pre-trained large language model to perform semantic parsing on the unstructured text in the multi-source heterogeneous business data streams. It identifies key events, capability performance, and problem focus information in the unstructured text and converts them into structured capability tags, wherein the capability tags represent the actual skill status of individual support personnel. The structured capability tags are integrated and associated with the digital and time-series data in the multi-source heterogeneous business data stream and linked to the individual support personnel. The pre-trained large language model is used to infer and complete the missing data features. At the same time, logical verification is used to filter out contradictory data between different business systems to obtain the standardized business dataset of the individual support personnel. Pre-existing job descriptions and competency model data are selected from a pre-defined external standard knowledge base, and the standardized business dataset is input into the pre-trained large language model. The selected job descriptions and competency model data are parsed, and a hierarchical and categorized competency indicator library is extracted. Combined with the pre-defined business logic, the pre-trained large language model is used to analyze the impact of structured competency label features in the standardized business dataset on the success rate of specific support tasks and assign corresponding weights. The business impact level corresponding to the individual support personnel is quantified, and a comprehensive evaluation index including personnel skill status, job standard requirements, and business impact level is obtained.

3. The intelligent generation model for the support personnel training scheme based on the large model-TGCOA-MDAN as described in claim 2, characterized in that, When the data preprocessing and indicator selection module obtains a comprehensive evaluation indicator that includes personnel skill status, job standard requirements, and business impact level, it is further configured as follows: The pre-trained large language model is driven to parse the preset business logic, analyze the potential impact of the structured capability label features in the standardized business dataset on the success rate of specific guarantee tasks, and assign corresponding guarantee weight values ​​according to the degree of impact to generate an initial business impact level that represents the relevance of the task. The system acquires the individual promotion plans, recent task priorities, and technology deployment plans of the support personnel, drives the pre-trained large language model to parse and transform these into future capability mapping features, core task capability features, and new equipment skill features for the corresponding positions, and matches them with the capability gaps determined by comparing the structured capability tag features with the hierarchical and classified capability index library. The system evaluates the learning urgency coefficient of the training needs corresponding to each capability gap, and uses the learning urgency coefficient and the support weight value to correct the initial business impact level, generating a corrected target business impact level. The target business impact level is integrated and mapped with the structured capability label features in the standardized business dataset and the hierarchical classification capability indicator library to output a comprehensive evaluation indicator that includes personnel skill status, job standard requirements, and business impact level.

4. The intelligent generation model for the support personnel training scheme based on the large model-TGCOA-MDAN as described in claim 1, characterized in that, The scheduling and coordination control module is configured as follows: Receive the comprehensive evaluation index, call the pre-trained large language model to parse the comprehensive evaluation index, and extract the scene features and data attributes corresponding to the current task; Based on the scene characteristics and the data attributes, determine the functional sub-modules that the current task needs to call, and combine the operation requirements of the functional sub-modules with the analysis results of the comprehensive evaluation index to generate algorithm configuration parameters adapted to the current task; Based on the algorithm configuration parameters, control instructions carrying corresponding parameters and matching business data streams are generated, and the control instructions and business data streams are respectively distributed to the TGCOA-based personalized solution generation engine and the MDAN-driven cross-domain adaptive and effect iteration module. The algorithm receives optimization signals from the result analysis, interpretation, and dynamic iterative optimization mechanism, adjusts the algorithm configuration parameters according to the optimization signals, updates the corresponding control commands and business data streams, and then outputs them to each functional module.

5. The intelligent generation model for the support personnel training scheme based on the large model-TGCOA-MDAN as described in claim 1, characterized in that, The TGCOA-based personalized solution generation engine is configured as follows: In response to the control command, receive the corresponding business data stream, configure the algorithm running parameters according to the algorithm configuration parameters carried by the control command, extract personnel skill status, job standard requirements, and business impact level from the comprehensive evaluation index, and construct the candidate solution space of the training scheme and the fitness evaluation function of the fit in combination with the preset training resource constraint boundary; A preset number of candidate training schemes are randomly generated as stingless bee individuals in the candidate solution space of the training scheme to complete the bee colony initialization. The fitness value corresponding to each stingless bee individual is calculated based on the fitness evaluation function to obtain the initial bee colony set and the initial global optimal candidate training scheme. The initial bee colony set is divided into scout bees, foraging bees, and follower bees according to a preset division of labor strategy. Using the initial globally optimal candidate training scheme as the iterative benchmark, multi-role collaborative optimization is performed. The scout bees are scheduled to conduct a global random search within the candidate solution space of the training scheme and generate new candidate training schemes. The foraging bees are scheduled to conduct a local search around the currently better candidate training scheme and adjust the components of the candidate training scheme to optimize the fitness value. The follower bees are scheduled to follow the search in the direction of a high-quality candidate training scheme based on the population information sharing mechanism. After completing the fitness verification of all candidate training schemes based on the fitness evaluation function, the updated bee colony set and the globally optimal candidate training scheme for this iteration are obtained. After each round of optimization, all candidate training schemes in the updated bee colony set are screened based on the calculation results of the fitness evaluation function. New random candidate training schemes are added to the candidate solution space of the training scheme to maintain population diversity. It is then determined whether the preset termination condition has been met. If not, the multi-role collaborative optimization is returned. If it has been met, an initial training scheme is generated and fed back to the scheduling and collaborative control module.

6. The intelligent generation model for the support personnel training scheme based on the large model-TGCOA-MDAN as described in claim 1, characterized in that, The MDAN-driven cross-domain adaptive and effect iteration module is configured as follows: In response to the control command, the system receives the corresponding business data stream, obtains the initial training scheme from the business data stream, introduces the target domain task environment data to be adapted, and retrieves the preset source domain labeled training sample data to form cross-domain adaptation data preparation. The feature extraction network is trained based on the source domain labeled training sample data. During the training process, a hybrid regularization strategy is adopted to enhance the generalization of the feature space. Self-supervised training is performed by combining controlled reconstruction learning to obtain the source domain pre-trained feature extraction model. The source domain pre-trained feature extraction model is invoked to predict the target domain task environment data to be adapted, generating corresponding pseudo labels. The mixing ratio of the source domain labeled training sample data and the target domain task environment data to be adapted is adjusted using Wasserstein distance as a distribution difference metric. In the input layer and feature layer of the source domain pre-trained feature extraction model, an intermediate mixed domain formed by the proportional mixing of the two domain data is constructed respectively. Progressive feature alignment is performed to gradually reduce the distribution difference between the two domains to obtain the intermediate domain adaptation feature representation. Based on the intermediate domain adaptation feature representation, the pseudo-labels are used as supervision information to perform self-learning training on the target domain task environment data to be adapted. In the feature space corresponding to the target domain task environment data to be adapted, a hybrid regularization strategy is used to construct consistency constraints, filter low-confidence pseudo-labels, and obtain the target domain optimized feature extraction model. The initial training scheme is modified for cross-domain features and adapted to the scene by using the feature extraction model optimized for the target domain, thereby generating the target training scheme.

7. The intelligent generation model for the support personnel training scheme based on the large model-TGCOA-MDAN as described in claim 1, characterized in that, The result analysis, interpretation, and dynamic iterative optimization mechanism is configured as follows: Receive the target training scheme, obtain the offline implementation effect feedback data corresponding to the target training scheme, call the pre-trained large language model to perform semantic parsing and structural transformation on the unstructured content in the offline implementation effect feedback data, filter out abnormal data, and match and align the processed offline implementation effect feedback data with the personnel skill status, job standard requirements and business impact level corresponding to the target training scheme one by one to obtain standardized effect feedback data and corresponding analysis benchmarks; Based on the aforementioned analytical benchmark, the pre-trained large language model is invoked to verify and interpret the target training scheme from four dimensions: capability coverage, resource matching status, expected business impact, and cross-domain adaptation effect, generating a structured scheme interpretability report. Using the expected indicators of each dimension in the structured solution interpretability report as a comparison benchmark, the standardized effect feedback data is compared with the corresponding expected effect, effect attribution analysis is performed, the solution deviation points and performance change causes are located, effect evaluation results are generated, and optimization signals are converted into optimization signals and fed back to the scheduling and collaborative control module. The optimization signals include algorithm configuration parameter adjustment instructions and solution generation logic correction directions.

8. A method for intelligently generating training schemes for support personnel based on a large model-TGCOA-MDAN, applied to the intelligent generation model for training schemes for support personnel based on a large model-TGCOA-MDAN as described in any one of claims 1-7, comprising: Acquire multi-source heterogeneous data related to support personnel, and perform structured fusion of the multi-source heterogeneous data to output a comprehensive evaluation index that includes personnel skill status, job standard requirements, and business impact level; The system receives the comprehensive evaluation index, parses the comprehensive evaluation index to form algorithm configuration parameters, and generates control instructions with algorithm configuration parameters and corresponding business data streams. In response to the control command, the system receives the service data stream and, within the preset training resource constraint boundary, applies a global optimization algorithm to perform optimization based on the algorithm configuration parameters, generates an initial training scheme, and feeds it back to the scheduling and collaborative control module. In response to the control command, the system receives the initial training scheme from the business data stream, introduces the target domain task environment data to be adapted, applies the domain adaptation network to perform feature cross-domain alignment and correction, and outputs the target training scheme. The target training scheme is subjected to interpretability analysis, and combined with the offline implementation effect feedback data of the target training scheme, an optimization signal is formed. The optimization signal is then fed back to the scheduling and collaborative control module to drive the scheduling and collaborative control module to dynamically adjust the control commands and the business data stream.

9. A computer storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the intelligent generation method for the support personnel training scheme based on the large model-TGCOA-MDAN as described in claim 8.

10. A computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the intelligent generation method for the support personnel training scheme based on the large model-TGCOA-MDAN as described in claim 8.

Images