An industry knowledge graph-based large model enhancement training method

Abstract

The application provides an industry knowledge graph-based large model enhancement training method, which comprises the following steps: constructing an industry task structure graph, structurally encoding task nodes to obtain a task structure vector set and a task structure hop number matrix; dividing a training task into stages based on the graph to form a stage task set and a corresponding stage sample set; training a large model stage by stage, and introducing a structural consistency constraint in each stage to align the model internal task representation with the task structure vector and make the relationship between tasks comply with the hop number matrix; and finally, fusing the model parameters and task intermediate representations of each stage, optimizing in a unified structure space, making the task representation of the final model close to the original graph structure, retaining the stage training memory and maintaining the hop number distance relationship between tasks. The method effectively improves the structural understanding and generalization ability of the large model in the industry task.

Description

Technical Field

[0001] This invention belongs to the field of large model technology, and in particular relates to a large model enhancement training method based on industry knowledge graph. Background Technology

[0002] With breakthroughs achieved by large-scale AI models in fields such as general language understanding, image recognition, and decision control, their application in complex industry-level systems has become a crucial direction for technological evolution. Especially in key industries such as education, transportation, energy, government, and the military, large-scale models are highly anticipated to enhance the intelligence level of task processing and decision-making efficiency. However, unlike general-purpose domains, these industries often have highly structured business processes, stringent regulations, and dynamically changing task chains, placing higher demands on the model's understanding capabilities, controllability, and task adaptability. Existing technologies generally employ fine-tuning or hint-based optimization methods based on general large models. These methods adapt the model to industry tasks by inputting domain data or constructing domain-specific question-answering templates. However, in practical applications, three significant problems are exposed: First, industry tasks have clear process dependencies and temporal transition logic, while general large models lack the ability to model such structured knowledge, easily leading to task order confusion or contextual comprehension breaks, making it impossible to accurately execute multi-step tasks in complex scenarios. Second, in actual deployment, the output of large models often lacks verifiability and dynamic feedback mechanisms for policy effectiveness or prediction accuracy, preventing the training and optimization process from forming a closed loop and severely impacting the model's stability and reliability. Third, during multi-task training, the lack of effective task structure control and stage management mechanisms often results in task interference and sample shifts, leading to uneven model performance across multiple sub-tasks and making it difficult to guarantee consistent and reliable performance in critical scenarios. To address the aforementioned issues, although some studies have attempted to introduce knowledge graphs or multi-task training frameworks, most of these are only used to enhance input information or assist in the reasoning stage. They have not yet proposed systematic improvement strategies for the model training methods themselves, nor have they fully considered the impact of industry-specific task flow structures and business feedback mechanisms on the training process. Summary of the Invention

[0003] The purpose of this invention is to design a large model augmentation training method based on industry knowledge graphs, which can deeply integrate industry knowledge structure with real feedback mechanism, improve the learning efficiency of large models of industry knowledge, and enhance their reliable execution capability in key scenarios.

[0004] To achieve the above objectives, this invention provides a method for large-scale model enhancement training based on industry knowledge graphs, the method comprising:

[0005] Construct an industry task structure graph and encode the task nodes in the graph to obtain a set of task structure vectors and a task structure hop count matrix;

[0006] Based on the task structure vector set and the task structure hop count matrix, the training task is divided into stages to form a stage task set and a corresponding stage sample set.

[0007] The model is trained stage by stage based on the set of stage tasks and the set of stage samples. In each stage of training, a structural consistency constraint is introduced to align the task representation within the model with the set of task structure vectors and to keep the structural relationship between tasks consistent with the task structure hop count matrix.

[0008] The model parameters obtained from each training stage and the corresponding intermediate task representations are fused into the final model. During the fusion process, a unified structure space optimization is used to ensure that the task representation of the final model closely matches the original task structure map, retains the training memory from each stage, and maintains the hop count distance between tasks.

[0009] Furthermore, the step of constructing the industry task structure graph includes: extracting task execution sequences from the historical task processing logs of the industry information system; constructing a directed graph based on the co-occurrence relationship of tasks in the sequence, where nodes represent tasks and directed edges represent dependencies between tasks; and calculating the structure vector of each task node and the shortest path hop count between any two task nodes based on the directed graph, thereby forming the task structure vector set and the task structure hop count matrix.

[0010] Furthermore, the step of dividing the training task into stages includes: using the Euclidean distance between task structure vectors, the hop distance in the task structure hop matrix, and whether there are dependent edges between tasks as constraints to cluster and group task nodes, so that tasks in the same stage are tightly aggregated in terms of structural semantics and have a short hop distance in the process path.

[0011] Furthermore, the stage sample set is obtained by filtering samples from historical task processing logs whose task numbers belong to the same stage task set, and each sample contains task input features and task output labels.

[0012] Furthermore, the structural consistency constraint includes: a supervision task loss term, a model task representation and task structure vector alignment loss term, a semantic similarity and structural relationship consistency loss term between task pairs, and a distance penalty term based on the task structure hop count matrix.

[0013] Furthermore, the distance penalty term based on the task structure hop count matrix is ​​used to encourage tasks with similar hop counts to have high similarity in the model representation space, while suppressing tasks with different hop counts from having excessively high similarity.

[0014] Furthermore, the intermediate representation of the task is extracted from fixed network layers during the training process of the large model and is used to characterize the model's structural understanding of the task.

[0015] Furthermore, the unified structure space optimization includes: setting the final structure vector of each task in the unified structure space, so that it simultaneously satisfies the following conditions: minimum distance to the set of task structure vectors, maximum consistency with the intermediate representation of the corresponding stage task, and Euclidean distance to other tasks reflecting the hop number relationship in the task structure hop number matrix.

[0016] Furthermore, the fusion of model parameters adopts a structure-aware weighting strategy, and the fusion weight of model parameters at each stage is determined according to the density of structural vectors in the unified structural space of the task at that stage.

[0017] Furthermore, the density of the structure vector is quantified by the mean norm or structure density of all task structure vectors in this stage, and then normalized and used as the fusion weight for the corresponding stage.

[0018] The beneficial technical effects of the present invention are at least as follows:

[0019] To address the aforementioned problems, this invention provides a large-scale model enhancement training method based on industry knowledge graphs. It establishes a structured knowledge representation model based on the logical relationships and execution paths between industry tasks to guide the stage division and task objective configuration of the training process, fundamentally solving the problems of multi-task interference and chaotic training paths. Simultaneously, it introduces a simulation system or digital twin system corresponding to the industry scenario to construct a feedback channel between model output and actual results, enabling dynamic adjustment and optimization of training behavior, thereby improving training closure and the verifiability of results. Furthermore, this invention's method achieves dynamic scheduling of model training strategies through a structural constraint mechanism, making the training process more closely aligned with the operational rules of actual tasks and ensuring the interpretability, stability, and adaptability of the model in complex industry environments. This invention not only improves the learning efficiency of large models in acquiring industry knowledge but also significantly enhances their reliable execution capabilities in key scenarios, filling a critical gap in existing large-scale model training technologies for intelligent industry applications. Attached Figure Description

[0020] The present invention will be further described with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present invention. For those skilled in the art, other drawings can be obtained based on the following drawings without creative effort.

[0021] Figure 1 This is a flowchart of a large model enhancement training method based on industry knowledge graphs according to the present invention. Detailed Implementation

[0022] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0023] In one or more embodiments, such as Figure 1 As shown, a method for augmenting training large models based on industry knowledge graphs is disclosed, the method comprising the following:

[0024] S1: Construct an industry task structure graph and perform structural encoding on the task nodes in the graph to obtain a set of task structure vectors and a task structure hop count matrix;

[0025] Specifically, this step aims to construct a directed graph structure that can express the sequence of tasks and control dependencies for typical industry scenarios with process structure characteristics (such as traffic control, energy scheduling, and education approval), and transform it into a vector representation that can be directly used in subsequent model training, thereby providing structural priors for the task process logic.

[0026] The input data for the graph construction comes from historical task processing logs accumulated in industry information systems. These logs record task instances generated by various business systems during actual operation, including fields such as task number, task type, department, task level, trigger time, end time, processing status, and associated task identifiers. These fields are stored in structured forms and are typically managed uniformly by the scheduling system or work order system, and can be obtained through data export interfaces or database query interfaces. Taking an urban traffic scheduling system as an example, the task types recorded in the log data include: event recognition, task assignment, signal control adjustment, video linkage, and execution confirmation.

[0027] A large number of task execution sequences were extracted from historical logs. Each sequence represents a complete task flow process, and this set is denoted as the task sequence set.

[0028] ;

[0029] Each of them Let be a sequence of task numbers arranged chronologically. Based on this sequence, we can statistically analyze the co-occurrence relationships between tasks and construct a directed graph structure, denoted as .

[0030] ;

[0031] in For a set of task nodes, Let be the set of directed edges between tasks. If tasks It frequently appears in most historical sequences in the task Previously, and in accordance with the process definition in industry standards, directed edges were established. , indicating task Depends on The task is completed. The strength of the dependency between tasks is represented by the edge weights, which are obtained by normalizing the historical co-occurrence frequency.

[0032] After the graph structure is constructed, a structure representation vector needs to be generated for each task node in the graph for subsequent structure-aware training of the model. Let the total number of tasks be... For each task Construct its structure vector This structure vector is derived from the task. The metadata of all upstream tasks is aggregated together with the edge weights, and the specific calculation method is as follows:

[0033] ;

[0034] in, Represents all directions in the diagram The upstream task set, as shown in the graph The structure is directly determined; For the task The static metadata vector contains fields such as task type, level, and role code, and is obtained by linear transformation after one-hot encoding of the structured fields. For the task and The feature representation of the edges between them is derived from the statistical results of task co-occurrence frequency, and normalized according to the maximum frequency to represent the dependency strength between the two tasks. and This is a linear mapping matrix that can be optimized during training; This represents a nonlinear mapping function, using the modified linear unit (ReLU) to enhance the model's expressive power.

[0035] Furthermore, to assist in subsequent task phase division operations, we further calculate the shortest path hop count matrix between tasks, denoted as...

[0036] ;

[0037] Among them, the first The element represents the task. To the mission The shortest path length between tasks, expressed in hops. This matrix is ​​calculated using... The Floyd-Warshall algorithm is executed to generate the structural distance between tasks, which is used to quantify the structural distance between tasks and serve as the structural constraint for task clustering in subsequent stages.

[0038] Finally, this step outputs two results: (1) a set of task structure vectors. (2) Task structure hop count matrix This is used to describe the structural proximity between any two tasks.

[0039] S2: Based on the task structure vector set and the task structure hop count matrix, the training task is divided into stages to form a stage task set and a corresponding stage sample set.

[0040] Specifically, the input for this step is the two variables output from the first step: (1) the task structure vector set. Each of them For the first (1) The structure-aware vector of each task has been encoded with its upstream task relationship; (2) Task hop count matrix , of which elements Indicates from the task arrive In the task map The shortest path length on the graph. All variables are derived from the graph structure construction and coding process, data is extracted from industry system logs, and the symbols are consistent with those in the first step. No new variables or data sources were introduced.

[0041] To make the stage division more closely resemble real-world industry task flows, this step introduces a task stage aggregation method that combines structural similarity with process constraints. Specifically, by aggregating task points in the task structure space and using a hop count matrix as a graph constraint, task nodes are grouped, ensuring that tasks within each group are tightly clustered in the structure space and have short hop distances in the process structure, thus forming training stage units. This process not only maintains the consistency of task semantic representation but also ensures the continuity of tasks in the business process.

[0042] Let the set of task stages after partitioning be... Each of them For the stage This includes a subset of task IDs. This partitioning is achieved through the following optimization objective function:

[0043] ;

[0044] The objective function consists of three parts: the first term is the squared Euclidean distance of the task structure representation, which encourages tasks in the same stage to cluster tightly in the structure space; the second term is the squared task hop count, which inhibits tasks with excessively large process spans from being grouped into the same stage; and the third term is a hop count deviation regularization term, where... As an indicator function, when the task and Having logically dependent edges (i.e.) The value is 1 if the condition is met, and 0 otherwise. This is used to explicitly penalize situations where adjacent tasks are incorrectly classified. is the regularization coefficient. By introducing this structural regularization term, we can more effectively address the problem of "disordered" tasks in industry flowcharts. This problem is particularly serious in processes such as "traffic perception - route planning - traffic light adjustment" in traffic scheduling systems, where traditional clustering cannot model such weakly coupled but strongly sequential relationships.

[0045] The objective function described above is optimized using a graph-structure-aware task clustering approach. During execution, the structural representation is first reduced in dimensionality, and then partitioned using a graph-structure-based hierarchical clustering algorithm. The process is reproducible and the computational complexity is controllable. Each stage of the task set obtained after partitioning... Used to extract its corresponding training sample set The training sample set is derived from real task logs. Each training sample contains input features and output labels, with the task number serving as the label field. Therefore, the sample set can be extracted and a stage sample set constructed by filtering the stage to which the task number belongs. .

[0046] At this point, this step outputs two results: first, a set of stage tasks. First, clearly define the tasks included in each training phase; second, define the sample set for each phase. This provides input for subsequent phased structure-guided training.

[0047] S3: Train the model stage by stage based on the set of stage tasks and the set of stage samples, and introduce structural consistency constraints in each stage of training to align the task representations within the model with the set of task structure vectors, and keep the structural relationships between tasks consistent with the task structure hop count matrix.

[0048] Specifically, this step is a crucial step in explicitly injecting industry structural knowledge into the large model training process. Its purpose is to integrate the phased tasks defined in the previous step. and corresponding stage sample set As input, the task structure representation output from the first step is introduced into the large model training. The expressed process depends on a structure, thus realizing the transformation from "structural priors" to "model parameters". Unlike traditional multi-task training mechanisms that only perform data-level task mixing, this step introduces three structural losses: structural consistency regularization, task comparison constraint, and hop count penalty mechanism. This effectively solves key challenges commonly encountered in industry task training, such as structural memory loss, out-of-order learning of processes, and fuzzy task boundaries. It is especially suitable for scenarios with rigid task processes, such as transportation, power, and government affairs.

[0049] The input for this step consists of three parts, all derived from the outputs of the previous two steps: the first being the set of stage task partitioning. and stage sample set The first is used to control the partitioning of training batches and sample scheduling; the second is a set of task structure representations. Each vector Indicates task The structural context, derived from graph structure path aggregation calculation, has been encoded with its position and upstream dependencies in the industry process; the third is the task hop count matrix. ,in Indicates task To the mission The shortest structural hop count is used to construct the task hop count constraint.

[0050] During training, a staged training mechanism is used for the large model. (Model parameters are denoted as) Updated in stages, each stage The training samples are Each sample middle, Input features for the task (such as scheduling instructions, sensor status, text descriptions, etc.). The task outputs objectives (such as state values, task labels, and result categories). During training, task-level representations are extracted from intermediate layers of the model. This is used to construct structural consistency and process constraints. To ensure that structural injection not only preserves structural information at the individual task level but also maintains the structural relationships between tasks without distortion, a training objective function composed of the following three loss terms is designed:

[0051] ;

[0052] The first term is the loss for a regular supervised task, corresponding to task label learning; the second term is the alignment term between the model representation and the structural graph, ensuring that the model's internal encoding closely approximates the task graph's structural representation; the third term is used to maintain the consistency of the semantic relationship structure between tasks, expressing similarity through an inner product form, making the relationship between task pairs in the model's semantic space consistent with the graph structure; the fourth term is the "hop count-aware structure penalty term," based on the hop count matrix. A distance function is defined, and the sigmoid function controls the penalty for distant task pairs. This encourages similarity between tasks with similar hop counts, while penalizing tasks with different hop counts less. This mechanism simulates the real-world pattern of "high coupling between nearby tasks and weak coupling between distant tasks" in industry processes, and has practical engineering implications. For example, in power equipment scheduling, "detection → initial judgment → control" must be highly similar, while "data archiving" can be significantly different.

[0053] Of the aforementioned losses, , , These are weighting factors, set empirically or tuned offline based on task type and structural complexity. Model Structure The specific configuration is determined based on the actual task data format. For example, a three-layer convolutional network and a two-layer feedforward network can be used for processing structured time series tasks, while a multi-layer encoding structure with an attention mechanism is used for text tasks. All training is based on a set of stage tasks. and sample set Execution, model parameters at each stage Continuously update based on previous stages, maintaining parameter sharing and structural continuity across stages.

[0054] After training, two types of variables are output: the first is the model parameter version after each training stage. The first is used for further training or deployment fusion in the next stage; the second is the set of intermediate representations of the model. This is used to construct a unified structural space representation in the subsequent structural fusion stage.

[0055] S4: The model parameters and corresponding intermediate task representations obtained from each training stage are fused into the final model. During the fusion process, the unified structure space is optimized to make the task representation of the final model closely resemble the original task structure map, retain the training memory of each stage, and maintain the hop distance between tasks.

[0056] Specifically, the purpose of this step is to combine the multiple training model parameters output from the previous stage. and the set of intermediate representations of tasks corresponding to each stage. Unified and integrated into a single deployable final model. Meanwhile, the consistency of the task structure must be maintained during the fusion process. Since the models in each stage were trained for different industry sub-tasks, their internal representations are each attached to the local task structure. If structural calibration is not performed during the fusion process, it can easily lead to the destruction of the global process relationship, such as "process reverse modeling" or "blurred task boundaries". This is especially true in highly dependent industries such as traffic emergency dispatch, power management, and education affairs approval, which can directly cause business logic disorder.

[0057] The input includes: (1) the parameters after each stage of model training. It consists of a multi-layer network structure, including an encoding layer, an attention layer, a normalization layer, etc., which is derived from the model trained in stages in step three; (2) the intermediate representation set of each stage , is the task vector extracted during training through a fixed layer inside the model (such as the latent vector of the penultimate layer), used to represent the model's structural understanding of the task; (3) the structural graph embedding representation output in the first step Each vector is generated by the task graph structure path encoding and has industry process constraint meaning; (4) Task hop matrix , which represents the shortest path length between task pairs and is used for structural continuity modeling.

[0058] To achieve structurally consistent fusion operations, a unified structural space is first constructed. , to make each task There is a final structure vector in this space. This vector must not only be close to the structure graph representation. It also needs to be combined with intermediate training. In addition to maintaining consistency, it is also necessary to maintain the relative hop count relationships between structures. Therefore, the following optimization objective is set:

[0059] ;

[0060] The first term represents the final structure vector of the task. With structural spectrum vector The distance ensures that the fusion result does not deviate from the original task flow structure. The second term represents... What I learned in training To maintain consistency, the fusion result retains training memory. The third term is a hop count constraint, which forces... and The Euclidean distance is used to restore the number of task jumps in the flowchart as much as possible. For example, if ,express and If the tasks are sequential in process, then their structural distances must be very close. This objective function can be optimized using gradient descent. and This is a hyperparameter that controls the sensitivity to training representation and structure hop count.

[0061] In terms of specific implementation Extracted from a specified layer during the training phase, in the standard Transformer architecture, the second-to-last layer encoding vector is usually selected; Derived from a graph structure representation, it has been generated through task graph path aggregation; and Then initialized to and The weighted average is then iteratively optimized using the aforementioned loss function. This process can be performed offline before fusion, does not rely on an online learning process, and has clear engineering feasibility.

[0062] Optimization yields a unified structural space Then, in order to achieve parameter-level fusion of the models, it is necessary to... The information is incorporated into the merging operation of model parameters at each stage. Considering the differences in the number of tasks and structural complexity at different stages, a structure-aware weighting strategy is adopted to fuse the model parameters at each stage to prevent the structure-dominant stage from being diluted. Let the final fused model parameters be:

[0063] ;

[0064] in The stage weights are calculated as follows: First, based on... Tasks at each stage The corresponding structural vector density (such as mean norm or structural density) is used to calculate a scalar representing the structural dominance of that stage, which is then normalized and used as a weighting coefficient. Taking a traffic task as an example, if the stage... The corresponding task is "path planning, signal control, and scheduling," which has a higher structural concentration than other stages (such as monitoring and archiving tasks), and therefore should be given higher weight. In order to retain its dominant influence in structural control.

[0065] This invention also provides a large model augmentation training device based on an industry knowledge graph, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the steps described in the above embodiment of a large model augmentation training method based on an industry knowledge graph. Figure 1 The steps S1 to S4 described above; or, when the processor executes the computer program, it implements the functions of each module in the above system embodiments.

[0066] For example, the computer program may be divided into one or more modules, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules may be a series of computer program instruction segments capable of performing specific functions, which describe the execution process of the computer program in the large model augmentation training device based on industry knowledge graphs.

[0067] The aforementioned large-scale model augmentation training device based on industry knowledge graphs can be a computing device such as a desktop computer, laptop, handheld computer, or cloud server. This device may include, but is not limited to, processors and memory. Those skilled in the art will understand that the device may also include input / output devices, network access devices, buses, etc.

[0068] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASACs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor. This processor is the control center of the large-scale model augmentation training device based on industry knowledge graphs, connecting all parts of the device via various interfaces and lines.

[0069] The memory can be used to store the computer programs and / or modules. The processor implements various functions of the large model augmentation training device based on industry knowledge graph by running or executing the computer programs and / or modules stored in the memory and calling the data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function, etc.; the data storage area may store data created based on the operation of the air conditioner controller, etc. In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital card (SD card), flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage devices.

[0070] The module integrated into the large model augmentation training device based on industry knowledge graphs, if implemented as a software functional unit and sold or used as an independent product, can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc.

[0071] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0072] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for augmenting training large models based on industry knowledge graphs, characterized in that, The method includes: A task structure graph is constructed for the industry, and the task nodes in the graph are structurally encoded to obtain a set of task structure vectors and a task structure hop count matrix. Specifically, for industry scenarios such as traffic control, energy scheduling, and education approval, task execution sequences are extracted from the historical task processing logs of industry information systems. The logs record task instances generated by various business systems during actual operation, including task number, task type, department, task level, trigger time, end time, processing status, and associated task identifiers. A directed graph is constructed based on the co-occurrence relationship of tasks in the sequence, where nodes represent tasks and directed edges represent dependencies between tasks. Based on the directed graph and the static meta-information vectors of the tasks, the structure vector of each task node and the shortest path hop count between any two task nodes are calculated to form the set of task structure vectors and the task structure hop count matrix. Based on the task structure vector set and the task structure hop count matrix, the training tasks are divided into stages to form a stage task set and a corresponding stage sample set. The step of dividing the training tasks into stages includes: using the Euclidean distance between task structure vectors, the hop count distance in the task structure hop count matrix, and whether there are dependency edges between tasks as constraints, clustering and grouping task nodes so that tasks within the same stage are tightly aggregated in structural semantics and have a short hop count distance in the process path. The model is trained stage by stage based on the set of stage tasks and the set of stage samples. In each stage of training, a structural consistency constraint is introduced to align the task representation within the model with the set of task structure vectors and to keep the structural relationship between tasks consistent with the task structure hop count matrix. The model parameters obtained from each training stage and the corresponding intermediate task representations are fused into the final model. During the fusion process, a unified structure space optimization is used to ensure that the task representation of the final model closely matches the original task structure map, retains the training memory of each stage, and maintains the hop distance relationship between tasks.

2. The method for large-scale model augmentation training based on industry knowledge graphs according to claim 1, characterized in that, The stage sample set is obtained by filtering samples from historical task processing logs whose task numbers belong to the same stage task set. Each sample contains task input features and task output labels.

3. The method for large-scale model augmentation training based on industry knowledge graphs according to claim 1, characterized in that, The structural consistency constraints include: a supervision task loss term, a model task representation and task structure vector alignment loss term, a semantic similarity and structural relationship consistency loss term between task pairs, and a distance penalty term based on the task structure hop count matrix.

4. The method for large-scale model augmentation training based on industry knowledge graphs according to claim 3, characterized in that, The distance penalty term based on the task structure hop count matrix is ​​used to encourage tasks with similar hop counts to have high similarity in the model representation space, while suppressing tasks with different hop counts from having excessively high similarity.

5. The method for large-scale model augmentation training based on industry knowledge graphs according to claim 1, characterized in that, The intermediate representation of the task is extracted from fixed network layers during the training process of the large model and is used to characterize the model's structural understanding of the task.

6. The method for large-scale model augmentation training based on industry knowledge graphs according to claim 1, characterized in that, The unified structure space optimization includes: setting the final structure vector of each task in the unified structure space, so that it simultaneously satisfies the following conditions: minimum distance to the set of task structure vectors, maximum consistency with the intermediate representation of the corresponding stage task, and Euclidean distance to other tasks reflecting the hop number relationship in the task structure hop number matrix.

7. The method for large-scale model augmentation training based on industry knowledge graphs according to claim 1, characterized in that, The model parameters are fused using a structure-aware weighting strategy, and the fusion weights of the model parameters at each stage are determined based on the density of the structural vectors in the unified structural space for that stage.

8. The method for large-scale model augmentation training based on industry knowledge graphs according to claim 7, characterized in that, The density of the structure vectors is quantified by the mean norm or structure density of all task structure vectors in this stage, and then normalized and used as the fusion weight for the corresponding stage.

Images