Industrial large model automatic optimization method and system based on artificial intelligence

Abstract

The invention relates to the technical field of model optimization, and particularly discloses an industry large model automatic optimization method and system based on artificial intelligence, and the method comprises the steps: obtaining multi-source heterogeneous data of a target industry, determining a data source entity set and an industry knowledge graph, and determining a training data set and a test data set of the target industry; determining an industry task demand set of the target industry based on the industry large model and an initial large model; determining an industry large model of the target industry, determining an evaluation value of each industry sub-model in the industry large model, and determining model evaluation data of the industry large model; and performing iterative optimization on the industry large model based on the industry task demand set of the industry large model and the model evaluation data. The industrial core demand can be quickly identified, the industrial task demand set adaptive to the complex industry is generated, the task collaboration is improved, the accuracy and generalization of the industrial large model are improved, the development cost is reduced, and the industrial intelligence is promoted.

Description

Technical Field

[0001] The present invention relates to the field of data retrieval technology, and in particular to an artificial intelligence-based automatic optimization method and system for large industry models. Background Art

[0002] The development of large industry models stems from the deepening application of artificial intelligence in vertical fields. In the early days, industry models mostly relied on rule systems or simple machine learning (such as SVM, decision trees), which were limited to a single task and difficult to process multi-source heterogeneous data. With the development of artificial intelligence technology, breakthroughs in deep learning have promoted the development of multi-task models, but still require a lot of manual feature engineering, and lack of generalization. At present, traditional methods have bottlenecks in task mining, model integration and automated optimization, and are difficult to adapt to dynamic industry needs. At the same time, the widespread existence of multi-source heterogeneous data and the increasing demand for large models in specific industries have prompted researchers to explore how to use these data to build and optimize industry large models to better meet the needs of industry-specific tasks.

[0003] Therefore, the present invention proposes an automatic optimization method and system for industry large models based on artificial intelligence. Summary of the Invention

[0004] The present invention provides an artificial intelligence-based industry big model automatic optimization method and system, which obtains and analyzes multi-source heterogeneous data, determines the data source entity set and the industry knowledge graph, determines the training data set and test data set of the target industry, determines the industry task requirement set of the target industry based on the industry big model and constructs the initial big model, trains the initial big model, determines and evaluates the industry big model, and iteratively optimizes the industry big model according to the industry task requirement set and model evaluation data. It can quickly identify the core needs of the industry, generate an industry task requirement set that is suitable for complex industries, improve task collaboration, improve the accuracy and generalization of the industry big model, reduce development costs, and promote industry intelligence.

[0005] The present invention provides an artificial intelligence-based automatic optimization method for large industry models, comprising:

[0006] S1: Obtain multi-source heterogeneous data for the target industry, determine the data source entity set and industry knowledge graph based on the multi-source heterogeneous data, and determine the training dataset and test dataset for the target industry;

[0007] S2: Determine the target industry's task requirement set based on the industry macro model based on multi-source heterogeneous data and data source entity sets, and determine the initial macro model based on the industry task requirement set and the industry knowledge graph;

[0008] S3: Train the initial large model based on the training data set to determine the industry large model of the target industry. Based on the test training set and the industry large model, determine the evaluation value of each industry sub-model in the industry large model, and determine the model evaluation data of the industry large model.

[0009] S4: Iteratively optimize the industry big model based on its industry task requirement set and model evaluation data.

[0010] Preferably, an artificial intelligence-based automatic optimization method for an industry large model obtains multi-source heterogeneous data of a target industry, including:

[0011] Obtain source data from each data source in the target industry and perform data preprocessing on the source data obtained from each data source. The source data includes structured data, semi-structured data, and unstructured data.

[0012] Determine multi-source heterogeneous data based on the source data of all data sources after data preprocessing.

[0013] Preferably, an artificial intelligence-based industry large model automatic optimization method determines a data source entity set and an industry knowledge graph based on multi-source heterogeneous data, and determines a training data set and a test data set for a target industry, including:

[0014] Performing entity extraction on the source data of each data source in the preprocessed multi-source heterogeneous data to determine a data source sub-entity set of each data source, wherein the data source sub-entity set includes multiple sub-entity objects and sub-entity attributes of each sub-entity object;

[0015] Perform entity disambiguation and entity deduplication on the data source entity set of all data sources to determine the data source entity set of all data sources, wherein the data source entity set includes multiple entity objects and entity attributes of each entity object;

[0016] Extract relationships from all source data based on the data source sub-entity set of each data source and the data source entity set of all data sources, determine the logical relationship between every two entity objects in the data source entity set of all data sources, and determine the data source relationship set of all data sources;

[0017] Build an industry knowledge graph based on the data source entity set and data source relationship set of all data sources;

[0018] Based on the industry knowledge graph, stratified sampling is performed on multi-source heterogeneous data to determine the training data set and test data set.

[0019] Preferably, an artificial intelligence-based automatic optimization method for an industry large model determines the task requirements of a target industry based on the industry large model based on multi-source heterogeneous data and a data source entity set, including:

[0020] Analyze multi-source heterogeneous data and the data source entity set of all data sources to determine the first task requirement set of all data sources in the target industry, where the first task requirement set includes multiple first subtask requirements and the priority of each first subtask requirement;

[0021] Performing requirement preprocessing on the task requirement set to determine a second task requirement set, wherein the requirement preprocessing includes at least requirement conflict resolution and requirement coordination and integration, and the second task requirement set includes a plurality of second subtask requirements and a priority of each second subtask requirement;

[0022] Analyze the second task requirement set and determine the consistency of every two second subtask requirements in the second task requirement set;

[0023] Based on the priorities of all second subtask requirements in the second task requirement set and the consistency of all second subtask requirements, the industry task requirement set of the target industry based on the industry big model is determined, wherein the industry task requirement set includes multiple industry subtask requirements and the priority of each industry subtask requirement.

[0024] Preferably, an artificial intelligence-based industry large model automatic optimization method determines the initial large model based on task requirements and industry knowledge graph, including:

[0025] Map each industry subtask requirement in the industry task requirement set to all entity objects in the industry knowledge graph and the logical relationships between all entity objects, and determine the subtask entity set and subtask relationship set of each industry subtask requirement in the industry task requirement set;

[0026] Determine the subtask entity set of each industry subtask requirement in the industry task requirement set as the model input of each initial sub-model, and determine each industry subtask requirement in the industry task requirement set as the model output of the corresponding initial sub-model;

[0027] Determine an initial sub-model for each industry sub-task requirement in the industry task requirement set based on the model input and model output of each initial sub-model;

[0028] Based on the initial sub-models of all industry sub-task requirements in the industry task requirement set, the initial large model of the target industry is determined.

[0029] Preferably, an automatic optimization method for an industry big model based on artificial intelligence is used to train an initial big model based on a training data set to determine the industry big model of a target industry, including:

[0030] Based on the model input and model output of the industry subtask requirements of each initial sub-model in the initial large model, the training data set is extracted to determine the sub-model training set of the initial sub-model required by each industry subtask;

[0031] Based on the sub-model training set required by each industry sub-task, the corresponding initial sub-model is trained to determine the industry sub-model of the target industry based on the industry sub-task requirements;

[0032] Based on the industry sub-models of all industry sub-task requirements in the industry task requirement set, determine the industry macro-model of the target industry.

[0033] Preferably, an automatic optimization method for an industry large model based on artificial intelligence determines the evaluation value of each industry sub-model in the industry large model based on a test training set and the industry large model, and determines the model evaluation data of the industry large model, including:

[0034] Based on the model input of the industry subtask requirements of each initial sub-model in the initial large model, the test data set is extracted to determine the sub-model test set of the industry sub-model required by each industry subtask;

[0035] Inputting the sub-model test set of each industry sub-task requirement in the industry task requirement set into the corresponding industry sub-model, and determining the prediction result of the industry sub-model of each industry sub-task requirement in the industry task requirement set based on the output result of the industry sub-model, wherein the prediction output includes at least one or more prediction sub-results;

[0036] Based on the model output of the industry subtask requirements of each initial sub-model in the initial large model, extract the test data set to determine the actual result of each industry subtask requirement, wherein the actual result includes at least one or more actual sub-results;

[0037] Determine the evaluation value of the industry sub-model of each industry sub-task requirement in the industry task requirement set based on the sub-task entity set, the sub-task relationship set, and the prediction results of the industry sub-model of each industry sub-task requirement in the industry task requirement set;

[0038] Based on the evaluation values of the industry sub-models of all industry sub-task requirements in the industry task requirement set, the model evaluation data of the industry large model is determined.

[0039] Preferably, an artificial intelligence-based automatic optimization method for an industry large model determines the evaluation value of the industry sub-model of each industry sub-task requirement in the industry task requirement set, including:

[0040]

[0041] Among them, A k represents the industry sub-model evaluation value of the kth industry sub-task requirement in the industry task requirement set, p kj represents the jth prediction sub-result in the prediction results of the industry sub-model of the kth industry sub-task requirement in the industry task requirement set, a kj represents the jth actual sub-result in the actual results of the kth industry sub-task requirement in the industry task requirement set, d(p kj ,a kj ) represents the j-th prediction result p kj and the jth actual sub-result a kj The difference measurement function between kj represents the weight of the jth actual sub-result in the actual results of the kth industry sub-task requirement in the industry task requirement set, kN4 represents the number of actual sub-results in the actual results of the kth industry sub-task requirement in the industry task requirement set, E k The impact value of the subtask entity set and subtask relationship set of the kth industry subtask requirement in the industry task requirement set on the industry sub-model evaluation, E1 k The first sub-influence value of the kth industry sub-task requirement in the industry task requirement set on the corresponding industry sub-model, E2 k Pr represents the second sub-influence value of the kth industry sub-task requirement in the industry task requirement set on the corresponding industry sub-model, k Indicates the priority of the kth industry subtask requirement in the industry task requirement set, En k The subtask entity set representing the kth industry subtask requirement in the industry task requirement set, En i Represents the subtask entity set of the i-th industry subtask requirement in the industry task requirement set, Re k Represents the subtask relationship set of the kth industry subtask requirement in the industry task requirement set, Re i represents the subtask relationship set of the i-th industry subtask requirement in the industry task requirement set, represents the number of logical relationships between the ath entity object in the subtask entity set of the kth industry subtask requirement in the industry task requirement set and all other entity objects in the subtask entity set, kN1 represents the number of entity objects in the subtask entity set of the kth industry subtask requirement in the industry task requirement set, DE represents the data source entity set, DR represents the data source relationship set, Nu1 DE Indicates the number of entity objects in the data source entity collection, Nu2 DR It represents the number of logical relationships in the data source relationship set, α represents the coordination parameter, λ represents the density parameter, and N3 represents the number of industry subtask requirements in the industry task requirement set.

[0042] The present invention provides an artificial intelligence-based industry large model automatic optimization system, which is used to execute any one of the artificial intelligence-based industry large model automatic optimization methods in Examples 1 to 8, including:

[0043] Acquisition module: Acquires multi-source heterogeneous data of the target industry, determines the data source entity set and industry knowledge graph based on the multi-source heterogeneous data, and determines the training data set and test data set of the target industry;

[0044] Determination module: Determines the target industry's industry task requirement set based on the industry big model based on multi-source heterogeneous data and data source entity sets, and determines the initial big model based on the industry task requirement set and the industry knowledge graph;

[0045] Evaluation module: Trains the initial large model based on the training data set, determines the industry large model of the target industry, determines the evaluation value of each industry sub-model in the industry large model based on the test training set and the industry large model, and determines the model evaluation data of the industry large model;

[0046] Optimization module: Iteratively optimizes the industry big model based on the industry task requirement set and model evaluation data of the industry big model.

[0047] Compared with the existing technology, the beneficial effects of the present invention are as follows: by acquiring and analyzing multi-source heterogeneous data, determining the data source entity set and the industry knowledge graph, determining the training data set and the test data set of the target industry, determining the industry task requirement set of the target industry based on the industry big model and building the initial big model, training the initial big model, determining and evaluating the industry big model, and iteratively optimizing the industry big model according to the industry task requirement set and model evaluation data, the core requirements of the industry can be quickly identified, and an industry task requirement set suitable for complex industries can be generated, thereby improving task collaboration, improving the accuracy and generalization of the industry big model, reducing development costs, and promoting industry intelligence.

[0048] Other features and advantages of the present invention will be described in the following description, and in part will become apparent from the description, or will be understood by practicing the present invention. The purpose and other advantages of the present invention can be achieved and obtained through the structures specifically pointed out in this application document.

[0049] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0051] Figure 1 This is an automatic optimization method for industry large models based on artificial intelligence in an embodiment of the present invention;

[0052] Figure 2 This is an artificial intelligence-based industry large model automatic optimization system in an embodiment of the present invention. DETAILED DESCRIPTION

[0053] The preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0054] Example 1:

[0055] The present invention provides an automatic optimization method for industry large models based on artificial intelligence, referring to Figure 1 ,include:

[0056] S1: Obtain multi-source heterogeneous data for the target industry, determine the data source entity set and industry knowledge graph based on the multi-source heterogeneous data, and determine the training dataset and test dataset for the target industry;

[0057] S2: Determine the target industry's task requirement set based on the industry macro model based on multi-source heterogeneous data and data source entity sets, and determine the initial macro model based on the industry task requirement set and the industry knowledge graph;

[0058] S3: Train the initial large model based on the training data set to determine the industry large model of the target industry. Based on the test training set and the industry large model, determine the evaluation value of each industry sub-model in the industry large model, and determine the model evaluation data of the industry large model.

[0059] S4: Iteratively optimize the industry big model based on its industry task requirement set and model evaluation data.

[0060] In this embodiment, by acquiring and processing multi-source heterogeneous data of the target industry, extracting data source entities, constructing an industry knowledge graph, and dividing the training data set and test data set, a data foundation is laid for the construction of a large industry model.

[0061] In this embodiment, by analyzing multi-source heterogeneous data and data source entities, industry task requirements are mined, and the initial large model framework is designed in combination with the knowledge graph.

[0062] In this embodiment, each subtask requirement is mapped to the entities and relationships of the knowledge graph, the subtask entities and subtask relationships are determined, and multiple initial sub-models are determined and integrated into an initial large model.

[0063] In this embodiment, the initial large model is optimized into an industry large model adapted to the industry through training and evaluation, and its performance is quantified.

[0064] In this embodiment, the industry large model is iteratively optimized based on industry task requirements (reflecting the priorities and goals of industry sub-task requirements) and model evaluation data (revealing the performance of industry sub-models).

[0065] In this embodiment, iterative optimization can be achieved through performance analysis: identifying weaknesses in evaluation data; task adjustment: prioritizing key tasks based on task priority; sub-model optimization: adjusting the architecture (such as increasing the number of layers), hyperparameters (such as learning rate) or data (such as enhanced samples); large model optimization: optimizing shared layers (such as Transformer encoders) or task weights to improve synergy; iterative training: retraining the optimized model, cyclically evaluating and improving.

[0066] The beneficial effects of the above technologies are: by acquiring and analyzing multi-source heterogeneous data, determining the data source entity set and industry knowledge graph, determining the training data set and test data set of the target industry, determining the industry task requirement set of the target industry based on the industry big model and building the initial big model, training the initial big model, determining and evaluating the industry big model, and iteratively optimizing the industry big model according to the industry task requirement set and model evaluation data, it is possible to quickly identify the core needs of the industry, generate an industry task requirement set that is suitable for complex industries, improve task collaboration, improve the accuracy and generalization of the industry big model, reduce development costs, and promote industry intelligence.

[0067] Example 2:

[0068] Based on Example 1, an artificial intelligence-based automatic optimization method for a large industry model obtains multi-source heterogeneous data of a target industry, including:

[0069] Obtain source data from each data source in the target industry and perform data preprocessing on the source data obtained from each data source. The source data includes structured data, semi-structured data, and unstructured data.

[0070] Determine multi-source heterogeneous data based on the source data of all data sources after data preprocessing.

[0071] In this embodiment, source data is collected from various data sources in target industries (such as electricity, medical care, finance, etc.) Data sources may include but are not limited to sensors, databases, log files, text reports, image records, etc.

[0072] In this embodiment, the structured data format is standardized and easy to query, such as a table in a database (a load record table in the power industry).

[0073] In this embodiment, the semi-structured data has a certain organization but a flexible format, such as JSON and XML files (equipment configuration files in the power industry).

[0074] In this embodiment, the unstructured data has no fixed format, such as text, images, and videos (inspection reports and equipment photos in the power industry).

[0075] In this embodiment, the source data usually has problems such as noise, inconsistency, and missing data, and data preprocessing is required to improve the data quality to ensure the reliability of subsequent analysis and modeling. Different preprocessing technologies are used for different data types (structured, semi-structured, and unstructured).

[0076] The beneficial effects of the above technologies are: obtaining multi-source heterogeneous data of the target industry can ensure data quality and consistency, and provide a data basis for determining the data source entity set and industry knowledge graph.

[0077] Example 3:

[0078] Based on Example 2, an artificial intelligence-based industry large model automatic optimization method determines a data source entity set and an industry knowledge graph based on multi-source heterogeneous data, and determines a training data set and a test data set for a target industry, including:

[0079] Performing entity extraction on the source data of each data source in the preprocessed multi-source heterogeneous data to determine a data source sub-entity set of each data source, wherein the data source sub-entity set includes multiple sub-entity objects and sub-entity attributes of each sub-entity object;

[0080] Perform entity disambiguation and entity deduplication on the data source entity set of all data sources to determine the data source entity set of all data sources, wherein the data source entity set includes multiple entity objects and entity attributes of each entity object;

[0081] Extract relationships from all source data based on the data source sub-entity set of each data source and the data source entity set of all data sources, determine the logical relationship between every two entity objects in the data source entity set of all data sources, and determine the data source relationship set of all data sources;

[0082] Build an industry knowledge graph based on the data source entity set and data source relationship set of all data sources;

[0083] Based on the industry knowledge graph, stratified sampling is performed on multi-source heterogeneous data to determine the training data set and test data set.

[0084] In this embodiment, entity extraction is performed on the source data of each data source after preprocessing to determine the data source entity set of each data source. This can be based on rules: using industry dictionaries and regular expressions to match entities (such as IPC classification numbers in the patent field); or it can be based on models: using named entity recognition (NER) models (such as BERT, Flair), fine-tuning for industry data to ensure that the entity recognition accuracy is ≥95%.

[0085] In this embodiment, entity disambiguation means resolving the ambiguity of entities with the same name (e.g., "Huawei" may refer to a company or a product), and achieving entity linking through contextual semantic analysis or knowledge graph embedding (e.g., the TransE algorithm).

[0086] In this embodiment, entity deduplication means calculating text similarity by using Jaccard similarity and setting a threshold (such as 0.8) to filter out duplicate entities or relationships.

[0087] In this embodiment, the relationship extraction of the corresponding source data is performed based on the data source entity set of each data source, and the logical relationship between each entity object in the data source entity set of each data source is determined. Remote supervision can be performed: using an existing knowledge base (such as Wikipedia) to automatically annotate data and train a relationship extraction model. Supervised learning can also be performed: using a sequence annotation model (such as BiLSTM+CRF) or a graph neural network (GNN) to extract the relationship between entities, ensuring that the relationship extraction F1 value is ≥0.85 through cross-validation, and manually verifying high-frequency relationships through industry experts, or using logical rules (such as ) Eliminate contradictory relationships.

[0088] In this embodiment, stratified sampling is performed on multi-source heterogeneous data based on the industry knowledge graph to determine the training data set and the test data set. Stratification can be performed by entity type: for example, if the data contains three types of entities, namely "enterprise", "product" and "policy", sampling ensures that the proportion of each type of entity in the training / test set is consistent with the global ratio; stratification is performed by relationship complexity: high-frequency relationships (such as "product-model") and low-frequency relationships (such as "policy-impact") can also be sampled separately to avoid the loss of low-frequency relationship samples.

[0089] In this embodiment, the industry knowledge graph is used to guide the partitioning of multi-source heterogeneous data. Training and test data sets are generated through stratified sampling to ensure that the data distribution reflects industry characteristics. Stratified sampling is based on the entities and relationships in the graph to partition the data set, meeting the requirements of representativeness: covering the main entity types in the industry (such as substations and users in the power industry); balance: maintaining the relationship distribution (such as the ratio of faults to normal states); and independence: no overlap between training and test data to avoid information leakage. The partition ratio (such as 80:20) is determined according to the task requirements. The training data set is used for model optimization, and the test data set is used for performance evaluation.

[0090] The beneficial effects of the above technologies are: determining the data source entity set and industry knowledge graph based on multi-source heterogeneous data, and determining the training data set and test data set of the target industry, which can ensure the high representativeness and consistency of the training data set and the test data set, provide data basis for building industry big models, and improve the adaptability and generalization ability of industry big models.

[0091] Example 4:

[0092] Based on Example 1, an artificial intelligence-based automatic optimization method for an industry big model determines the task requirements of a target industry based on the industry big model based on multi-source heterogeneous data and a data source entity set, including:

[0093] Analyze multi-source heterogeneous data and the data source entity set of all data sources to determine the first task requirement set of all data sources in the target industry, where the first task requirement set includes multiple first subtask requirements and the priority of each first subtask requirement;

[0094] Performing requirement preprocessing on the task requirement set to determine a second task requirement set, wherein the requirement preprocessing includes at least requirement conflict resolution and requirement coordination and integration, and the second task requirement set includes a plurality of second subtask requirements and a priority of each second subtask requirement;

[0095] Analyze the second task requirement set and determine the consistency of every two second subtask requirements in the second task requirement set;

[0096] Based on the priorities of all second subtask requirements in the second task requirement set and the consistency of all second subtask requirements, the industry task requirement set of the target industry based on the industry big model is determined, wherein the industry task requirement set includes multiple industry subtask requirements and the priority of each industry subtask requirement.

[0097] In this embodiment, the source data of each data source in the multi-source heterogeneous data is analyzed to determine the subtask requirements of each data source. A data source dependency graph can be drawn through a data lineage tool (such as Apache Atlas) to identify key data sources (such as "order data" affecting the two tasks of "inventory forecasting" and "user profiling").

[0098] In this embodiment, cross-data source requirement conflicts of shared entities (such as "user ID") are marked (for example, system A requires "user ID associated with consumption level", and system B requires "user ID associated with regional distribution").

[0099] In this embodiment, the first task requirements may be redundant, conflicting, or inconsistent, and need to be optimized through requirement preprocessing to generate clearer second task requirements.

[0100] In this embodiment, demand coordination and integration, for example, "integrate order data (structured) and user reviews (text) to generate personalized product recommendations."

[0101] In this embodiment, the requirement conflict resolution is, for example, "giving up full access to log data due to the real-time requirement of sensor data."

[0102] In this embodiment, the priorities of the first subtask requirements and the second subtask requirements can be evaluated through the Kano model. For example, basic requirements (such as data security) are set to the highest priority; expected requirements (such as reasoning speed) are set to dynamically adjustable priorities; and exciting requirements (such as multimodal output) are set to extended optimization items.

[0103] In this embodiment, the consistency of each second subtask requirement within the second task requirement is evaluated, that is, the compatibility of the semantics, goals, or resource requirements between the tasks. For example, in the power industry, "load forecasting" and "fault detection" may share a data source (sensor data) and have high consistency; whereas "real-time monitoring" and "offline analysis" have different time requirements and low consistency. Consistency analysis ensures inter-task collaboration.

[0104] In this embodiment, the priority and consistency of the second subtask requirements are integrated to optimize and generate the final industry task requirements. As the task of the industry big model, the industry task requirements include multiple industry subtask requirements (such as "comprehensive load forecasting" and "fault detection and scheduling") and their priorities, reflecting the importance of the tasks and the order of execution.

[0105] The beneficial effects of the above technologies are: based on multi-source heterogeneous data and data source entity sets, the task requirements of the target industry based on the industry big model can be determined, a task set suitable for complex industries can be generated, the core needs of the industry can be quickly identified, task collaboration can be improved, development costs can be reduced, model adaptability and business value can be enhanced, and industry intelligence can be promoted.

[0106] Example 5:

[0107] Based on Example 4, an artificial intelligence-based industry large model automatic optimization method determines an initial large model based on task requirements and industry knowledge graph, including:

[0108] Map each industry subtask requirement in the industry task requirement set to all entity objects in the industry knowledge graph and the logical relationships between all entity objects, and determine the subtask entity set and subtask relationship set of each industry subtask requirement in the industry task requirement set;

[0109] Determine the subtask entity set of each industry subtask requirement in the industry task requirement set as the model input of each initial sub-model, and determine each industry subtask requirement in the industry task requirement set as the model output of the corresponding initial sub-model;

[0110] Determine an initial sub-model for each industry sub-task requirement in the industry task requirement set based on the model input and model output of each initial sub-model;

[0111] Based on the initial sub-models of all industry sub-task requirements in the industry task requirement set, the initial large model of the target industry is determined.

[0112] In this embodiment, the initial sub-model for each industry sub-task requirement in the industry task requirement set is constructed. The corresponding initial sub-model, i.e., an untrained model framework, can be designed based on the input (sub-task entity) and output (sub-task requirement target) of each sub-task. If it is a classification task, convolutional neural network (CNN), recurrent neural network (RNN) and its variants (such as LSTM, GRU), Transformer architecture, etc. can be selected; if it is a regression task, linear regression, logistic regression, random forest regression and other models can be selected; if it is a generation task, generative adversarial network (GAN), variational autoencoder (VAE) and other models can be selected.

[0113] In this embodiment, each industry subtask requirement in the industry task requirement (such as "load forecasting" and "fault detection" in the power industry) is mapped with the entity objects (such as "substation" and "generator") in the industry knowledge graph and their logical relationships (such as "power supply" and "fault association") to identify the core objects (subtask entities) involved in each subtask and their associations (subtask relationships).

[0114] In this embodiment, the industry subtask requirements are decomposed, their goals and required information are clarified, and task-related entities and relationships are searched in the industry knowledge graph.

[0115] In this embodiment, a corresponding initial sub-model is designed for each industry sub-task requirement, and the input and output of the model are clarified: Model input: the sub-task entity (such as "substation" and "load data") and its attributes (such as voltage and time series) are used as the input data of the initial sub-model; Model output: the goal of the industry sub-task requirement is defined as the model output, such as the output of the "load forecast" task is "load value for the next 24 hours."

[0116] In this embodiment, all initial sub-models are integrated into a unified initial large model as the basic model framework for the target industry. The initial large model is a comprehensive architecture that can handle multiple sub-tasks. It may be implemented in the following ways: Model collection: each sub-model is run as an independent module, loosely coupled, and shares some data or features; Joint model: Design a multi-task learning (MTL) architecture with a shared underlying feature extraction layer (such as a shared Transformer encoder) and a top-level task-specific output layer; Graph-driven model: Combined with the industry knowledge graph, a graph neural network (GNN) is designed, which takes graph entities and relationships as input and adapts to multiple tasks.

[0117] The beneficial effects of the above technology are: determining the initial large model based on task requirements and industry knowledge graphs can achieve accurate mapping of task requirements and industry knowledge graphs, quickly adapt to dynamic tasks, and reduce model development costs.

[0118] Example 6:

[0119] Based on Example 5, an artificial intelligence-based industry big model automatic optimization method is provided, which trains an initial big model based on a training data set to determine an industry big model for a target industry, including:

[0120] Based on the model input and model output of the industry subtask requirements of each initial sub-model in the initial large model, the training data set is extracted to determine the sub-model training set of the initial sub-model required by each industry subtask;

[0121] Based on the sub-model training set required by each industry sub-task, the corresponding initial sub-model is trained to determine the industry sub-model of the target industry based on the industry sub-task requirements;

[0122] Based on the industry sub-models of all industry sub-task requirements in the industry task requirement set, determine the industry macro-model of the target industry.

[0123] In this example, a sub-model training set corresponding to each initial sub-model is extracted from the training dataset (generated through stratified sampling, encompassing a representative sample of heterogeneous data from multiple sources). Each initial sub-model is designed for a specific industry sub-task (e.g., "load forecasting" or "fault detection"), with well-defined model inputs and outputs. This extraction process ensures that the training data aligns with the sub-model's input and output requirements.

[0124] In this embodiment, each sub-model training set is used to train the corresponding initial sub-model, optimize its parameters, and generate an industry sub-model adapted to the needs of a specific industry sub-task. The initial sub-model is an untrained framework (such as LSTM or Transformer) that defines the input and output structure. The training process adjusts the model weights and other factors to achieve the expected performance (such as high prediction accuracy) on the sub-task.

[0125] In this embodiment, all industry sub-models are integrated into a unified industry model to form a comprehensive model of the target industry, which has the ability to handle multiple sub-tasks. The integration method takes into account the synergy between tasks and the model structure, and may include: modular combination: each industry sub-model runs independently, which is suitable for scenarios with weak dependencies between tasks; multi-task learning (MTL): sharing the underlying feature extraction layer (such as sharing the Transformer encoder), and the top layer is a task-specific output layer to improve synergy; graph neural network (GNN): with the industry knowledge graph as the backbone, it integrates multiple related industry sub-models and is suitable for strong relationship tasks; the industry model is jointly optimized (such as fine-tuning the shared layer) to ensure the overall performance and balance between tasks.

[0126] The beneficial effects of the above technologies are: training the initial large model based on the training data set, determining the industry large model of the target industry, efficiently utilizing multi-source heterogeneous data, reducing training costs, building an industry large model that is highly adaptable to the industry, and improving the accuracy and generalization ability of the model in complex industry scenarios.

[0127] Example 7:

[0128] Based on Example 6, an artificial intelligence-based automatic optimization method for an industry large model determines the evaluation value of each industry sub-model in the industry large model based on a test training set and the industry large model, and determines the model evaluation data of the industry large model, including:

[0129] Based on the model input of the industry subtask requirements of each initial sub-model in the initial large model, the test data set is extracted to determine the sub-model test set of the industry sub-model required by each industry subtask;

[0130] Inputting the sub-model test set of each industry sub-task requirement in the industry task requirement set into the corresponding industry sub-model, and determining the prediction result of the industry sub-model of each industry sub-task requirement in the industry task requirement set based on the output result of the industry sub-model, wherein the prediction output includes at least one or more prediction sub-results;

[0131] Based on the model output of the industry subtask requirements of each initial sub-model in the initial large model, extract the test data set to determine the actual result of each industry subtask requirement, wherein the actual result includes at least one or more actual sub-results;

[0132] Determine the evaluation value of the industry sub-model of each industry sub-task requirement in the industry task requirement set based on the sub-task entity set, the sub-task relationship set, and the prediction results of the industry sub-model of each industry sub-task requirement in the industry task requirement set;

[0133] Based on the evaluation values of the industry sub-models of all industry sub-task requirements in the industry task requirement set, the model evaluation data of the industry large model is determined.

[0134] In this example, a sub-model test set corresponding to each industry sub-model is extracted from a test dataset (generated through stratified sampling, independent of the training dataset, and reflecting the industry data distribution). Each industry sub-model is designed for a specific industry sub-task requirement and is trained based on an initial sub-model with well-defined model inputs. The extraction process ensures that the test data aligns with the sub-model's input requirements.

[0135] In this embodiment, each sub-model test set is input into the corresponding industry sub-model to generate a prediction result. The prediction result is the output of the sub-model based on the test data, reflecting the target required by the sub-task. The prediction output contains one or more prediction sub-results, corresponding to the multi-dimensional target or multi-category result of the task.

[0136] In this embodiment, the actual results required by each industry subtask are extracted from the test dataset, that is, the true labels or target values of the test data, for comparison with the predicted results. The actual results are extracted based on the model output of the subtask and include one or more actual sub-results.

[0137] In this embodiment, the evaluation values of all industry sub-models are integrated to generate model evaluation data of the industry large model, which comprehensively reflects the performance of the large model on the target industry task.

[0138] The beneficial effects of the above technologies are: based on the test training set and the industry big model, the evaluation value of each industry sub-model in the industry big model is determined, and the model evaluation data of the industry big model is determined, which can quantify the performance of the industry big model, ensure that the evaluation results are highly consistent with the industry scenario, improve cross-task synergy, reduce evaluation bias, and improve the adaptability and reliability of the industry big model.

[0139] Example 8:

[0140] Based on Example 7, an artificial intelligence-based industry large model automatic optimization method determines the evaluation value of the industry sub-model of each industry sub-task requirement in the industry task requirement set, including:

[0141]

[0142]

[0143] Among them, A k represents the industry sub-model evaluation value of the kth industry sub-task requirement in the industry task requirement set, p kj represents the jth prediction sub-result in the prediction results of the industry sub-model of the kth industry sub-task requirement in the industry task requirement set, a kj represents the jth actual sub-result in the actual results of the kth industry sub-task requirement in the industry task requirement set, d(p kj ,a kj ) represents the j-th prediction result p kj and the jth actual sub-result a kj The difference measurement function between kj represents the weight of the jth actual sub-result in the actual results of the kth industry sub-task requirement in the industry task requirement set, kN4 represents the number of actual sub-results in the actual results of the kth industry sub-task requirement in the industry task requirement set, E k The impact value of the subtask entity set and subtask relationship set of the kth industry subtask requirement in the industry task requirement set on the industry sub-model evaluation, E1 k The first sub-influence value of the kth industry sub-task requirement in the industry task requirement set on the corresponding industry sub-model, E2 k Pr represents the second sub-influence value of the kth industry sub-task requirement in the industry task requirement set on the corresponding industry sub-model, k Indicates the priority of the kth industry subtask requirement in the industry task requirement set, En k The subtask entity set representing the kth industry subtask requirement in the industry task requirement set, En iRepresents the subtask entity set of the i-th industry subtask requirement in the industry task requirement set, k Represents the subtask relationship set of the kth industry subtask requirement in the industry task requirement set, Re i represents the subtask relationship set of the i-th industry subtask requirement in the industry task requirement set, represents the number of logical relationships between the ath entity object in the subtask entity set of the kth industry subtask requirement in the industry task requirement set and all other entity objects in the subtask entity set, kN1 represents the number of entity objects in the subtask entity set of the kth industry subtask requirement in the industry task requirement set, DE represents the data source entity set, DR represents the data source relationship set, Nu1 DE Indicates the number of entity objects in the data source entity collection, Nu2 DR It represents the number of logical relationships in the data source relationship set, α represents the coordination parameter, λ represents the density parameter, and N3 represents the number of industry subtask requirements in the industry task requirement set.

[0144] In this embodiment, Nu2 DR / Nu1 DE Represents the average logical relationship value of entity objects in the industry knowledge graph.

[0145] In this embodiment, The entity objects and logical relationships representing the kth industry subtask requirement in the industry task requirement set are based on the coverage value of the industry knowledge graph.

[0146] In this embodiment, Represents the entity object and logical relationship of the kth industry subtask requirement in the industry task requirement set, and the collaborative value of the remaining industry subtask requirements in the industry task requirement set.

[0147] In this embodiment, The logical relationship representing the kth industry subtask requirement in the industry task requirement set is based on the relationship density value of the industry knowledge graph, that is, the relationship richness of the industry subtask requirement based on the industry knowledge graph.

[0148] In this embodiment, the second sub-influence value E2 k It represents the density impact value of the relationship density value of the kth industry sub-task requirement in the industry task requirement set on the evaluation value of the corresponding industry sub-model.

[0149] In this embodiment, the first sub-influence value E1 k It represents the collaborative impact value of the collaborative value of the kth industry sub-task requirement in the industry task requirement set on the evaluation value of the corresponding industry sub-model.

[0150] In this embodiment, It represents the priority of the kth industry subtask requirement in the industry task requirement set and the priority-coverage impact value of the coverage value on the evaluation value of the corresponding industry sub-model.

[0151] In this embodiment, the difference metric function d(p kj ,a kj ) can be selected according to the specific task, for example: the actual sub-result is numerical: d(p kj ,a kj )=|p kj -a kj |or d(p kj ,a kj )=(p kj -a kj ) 2 ; Non-numeric type, determines whether they are completely consistent, or calculates similarity.

[0152] In this embodiment, max(d(p kj ,a kj )) is a normalization factor to ensure that the difference values are within a reasonable range.

[0153] The beneficial effects of the above technology are: based on the subtask entity set, subtask relationship set and prediction results of the industry sub-model of each industry subtask requirement in the industry task requirement set, the evaluation value of the industry sub-model of each industry subtask requirement in the industry task requirement set is determined, the performance of all industry sub-models can be quantified, and high-quality data support can be provided for determining the model evaluation data to achieve iterative optimization of the industry large model.

[0154] Example 9:

[0155] The present invention provides an industry large model automatic optimization system based on artificial intelligence, which is used to execute any one of the industry large model automatic optimization methods based on artificial intelligence in embodiments 1 to 8, with reference to Figure 2 ,include:

[0156] Acquisition module: Acquires multi-source heterogeneous data of the target industry, determines the data source entity set and industry knowledge graph based on the multi-source heterogeneous data, and determines the training data set and test data set of the target industry;

[0157] Determination module: Determines the target industry's industry task requirement set based on the industry big model based on multi-source heterogeneous data and data source entity sets, and determines the initial big model based on the industry task requirement set and the industry knowledge graph;

[0158] Evaluation module: Trains the initial large model based on the training data set, determines the industry large model of the target industry, determines the evaluation value of each industry sub-model in the industry large model based on the test training set and the industry large model, and determines the model evaluation data of the industry large model;

[0159] Optimization module: Iteratively optimizes the industry big model based on the industry task requirement set and model evaluation data of the industry big model.

[0160] The beneficial effects of the above technologies are: by acquiring and analyzing multi-source heterogeneous data, determining the data source entity set and industry knowledge graph, determining the training data set and test data set of the target industry, determining the industry task requirement set of the target industry based on the industry big model and building the initial big model, training the initial big model, determining and evaluating the industry big model, and iteratively optimizing the industry big model according to the industry task requirement set and model evaluation data, it is possible to quickly identify the core needs of the industry, generate an industry task requirement set that is suitable for complex industries, improve task collaboration, improve the accuracy and generalization of the industry big model, reduce development costs, and promote industry intelligence.

[0161] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if such changes and modifications fall within the scope of the claims and their equivalents, the present invention is intended to include such changes and modifications.

Claims

1. An automatic optimization method for industry large models based on artificial intelligence, characterized by: include: S1: Obtain multi-source heterogeneous data for the target industry, determine the data source entity set and industry knowledge graph based on the multi-source heterogeneous data, and determine the training dataset and test dataset for the target industry; S2: Determine the target industry's task requirement set based on the industry macro model based on multi-source heterogeneous data and data source entity sets, and determine the initial macro model based on the industry task requirement set and the industry knowledge graph; S3: Train the initial large model based on the training data set to determine the industry large model of the target industry. Based on the test training set and the industry large model, determine the evaluation value of each industry sub-model in the industry large model, and determine the model evaluation data of the industry large model. S4: Iteratively optimize the industry big model based on its industry task requirement set and model evaluation data.

2. The method for automatic optimization of industry large models based on artificial intelligence according to claim 1 is characterized in that: Acquire multi-source heterogeneous data for the target industry, including: Obtain source data from each data source in the target industry and perform data preprocessing on the source data obtained from each data source. The source data includes structured data, semi-structured data, and unstructured data. Determine multi-source heterogeneous data based on the source data of all data sources after data preprocessing.

3. The method for automatic optimization of industry large models based on artificial intelligence according to claim 2 is characterized in that: Determine the data source entity set and industry knowledge graph based on multi-source heterogeneous data, and determine the training and test datasets for the target industry, including: Performing entity extraction on the source data of each data source in the preprocessed multi-source heterogeneous data to determine a data source sub-entity set of each data source, wherein the data source sub-entity set includes multiple sub-entity objects and sub-entity attributes of each sub-entity object; Perform entity disambiguation and entity deduplication on the data source entity set of all data sources to determine the data source entity set of all data sources, wherein the data source entity set includes multiple entity objects and entity attributes of each entity object; Extract relationships from all source data based on the data source sub-entity set of each data source and the data source entity set of all data sources, determine the logical relationship between every two entity objects in the data source entity set of all data sources, and determine the data source relationship set of all data sources; Build an industry knowledge graph based on the data source entity set and data source relationship set of all data sources; Based on the industry knowledge graph, stratified sampling is performed on multi-source heterogeneous data to determine the training data set and test data set.

4. The method for automatic optimization of industry large models based on artificial intelligence according to claim 1 is characterized in that: Determine the target industry's task requirements based on the industry macro model based on multi-source heterogeneous data and data source entity sets, including: Analyze multi-source heterogeneous data and the data source entity set of all data sources to determine the first task requirement set of all data sources in the target industry, where the first task requirement set includes multiple first subtask requirements and the priority of each first subtask requirement; Performing requirement preprocessing on the task requirement set to determine a second task requirement set, wherein the requirement preprocessing includes at least requirement conflict resolution and requirement coordination and integration, and the second task requirement set includes a plurality of second subtask requirements and a priority of each second subtask requirement; Analyze the second task requirement set and determine the consistency of every two second subtask requirements in the second task requirement set; Based on the priorities of all second subtask requirements in the second task requirement set and the consistency of all second subtask requirements, the industry task requirement set of the target industry based on the industry big model is determined, wherein the industry task requirement set includes multiple industry subtask requirements and the priority of each industry subtask requirement.

5. The method for automatic optimization of industry large models based on artificial intelligence according to claim 4 is characterized in that: Determine the initial large model based on task requirements and industry knowledge graph, including: Map each industry subtask requirement in the industry task requirement set to all entity objects in the industry knowledge graph and the logical relationships between all entity objects, and determine the subtask entity set and subtask relationship set of each industry subtask requirement in the industry task requirement set; Determine the subtask entity set of each industry subtask requirement in the industry task requirement set as the model input of each initial sub-model, and determine each industry subtask requirement in the industry task requirement set as the model output of the corresponding initial sub-model; Determine an initial sub-model for each industry sub-task requirement in the industry task requirement set based on the model input and model output of each initial sub-model; Based on the initial sub-models of all industry sub-task requirements in the industry task requirement set, the initial large model of the target industry is determined.

6. The method for automatic optimization of industry large models based on artificial intelligence according to claim 5 is characterized in that: Train the initial large model based on the training data set to determine the industry large model of the target industry, including: Based on the model input and model output of the industry subtask requirements of each initial sub-model in the initial large model, the training data set is extracted to determine the sub-model training set of the initial sub-model required by each industry subtask; Based on the sub-model training set required by each industry sub-task, the corresponding initial sub-model is trained to determine the industry sub-model of the target industry based on the industry sub-task requirements; Based on the industry sub-models of all industry sub-task requirements in the industry task requirement set, determine the industry macro-model of the target industry.

7. The method for automatic optimization of industry large models based on artificial intelligence according to claim 6 is characterized in that: Based on the test training set and the industry model, determine the evaluation value of each industry sub-model in the industry model, and determine the model evaluation data of the industry model, including: Based on the model input of the industry subtask requirements of each initial sub-model in the initial large model, the test data set is extracted to determine the sub-model test set of the industry sub-model required by each industry subtask; Inputting the sub-model test set of each industry sub-task requirement in the industry task requirement set into the corresponding industry sub-model, and determining the prediction result of the industry sub-model of each industry sub-task requirement in the industry task requirement set based on the output result of the industry sub-model, wherein the prediction output includes at least one or more prediction sub-results; Based on the model output of the industry subtask requirements of each initial sub-model in the initial large model, extract the test data set to determine the actual result of each industry subtask requirement, wherein the actual result includes at least one or more actual sub-results; Determine the evaluation value of the industry sub-model of each industry sub-task requirement in the industry task requirement set based on the sub-task entity set, the sub-task relationship set, and the prediction results of the industry sub-model of each industry sub-task requirement in the industry task requirement set; Based on the evaluation values of the industry sub-models of all industry sub-task requirements in the industry task requirement set, the model evaluation data of the industry large model is determined.

8. The method for automatic optimization of industry large models based on artificial intelligence according to claim 7 is characterized in that: Determine the evaluation value of the industry sub-model for each industry sub-task requirement in the industry task requirement set, including: Among them, A k represents the industry sub-model evaluation value of the kth industry sub-task requirement in the industry task requirement set, p kj represents the jth prediction sub-result in the prediction results of the industry sub-model of the kth industry sub-task requirement in the industry task requirement set, a kj represents the jth actual sub-result in the actual results of the kth industry sub-task requirement in the industry task requirement set, d(p kj ,a kj ) represents the j-th prediction result p kj and the jth actual sub-result a kj The difference measurement function between kj represents the weight of the jth actual sub-result in the actual results of the kth industry sub-task requirement in the industry task requirement set, kN4 represents the number of actual sub-results in the actual results of the kth industry sub-task requirement in the industry task requirement set, E k The impact value of the subtask entity set and subtask relationship set of the kth industry subtask requirement in the industry task requirement set on the industry sub-model evaluation, E1 k The first sub-influence value of the kth industry sub-task requirement in the industry task requirement set on the corresponding industry sub-model, E2 k Pr represents the second sub-influence value of the kth industry sub-task requirement in the industry task requirement set on the corresponding industry sub-model, k Indicates the priority of the kth industry subtask requirement in the industry task requirement set, En k The subtask entity set representing the kth industry subtask requirement in the industry task requirement set, En i Represents the subtask entity set of the i-th industry subtask requirement in the industry task requirement set, Re k Represents the subtask relationship set of the kth industry subtask requirement in the industry task requirement set, Re i represents the subtask relationship set of the i-th industry subtask requirement in the industry task requirement set, represents the number of logical relationships between the ath entity object in the subtask entity set of the kth industry subtask requirement in the industry task requirement set and all other entity objects in the subtask entity set, kN1 represents the number of entity objects in the subtask entity set of the kth industry subtask requirement in the industry task requirement set, DE represents the data source entity set, DR represents the data source relationship set, Nu1 DE Indicates the number of entity objects in the data source entity collection, Nu2 DR It represents the number of logical relationships in the data source relationship set, α represents the coordination parameter, λ represents the density parameter, and N3 represents the number of industry subtask requirements in the industry task requirement set.

9. An industry large model automatic optimization system based on artificial intelligence, characterized by: The method for automatically optimizing an industry large model based on artificial intelligence according to any one of embodiments 1 to 8 comprises: Acquisition module: Acquires multi-source heterogeneous data of the target industry, determines the data source entity set and industry knowledge graph based on the multi-source heterogeneous data, and determines the training data set and test data set of the target industry; Determination module: Determines the target industry's industry task requirement set based on the industry big model based on multi-source heterogeneous data and data source entity sets, and determines the initial big model based on the industry task requirement set and the industry knowledge graph; Evaluation module: Trains the initial large model based on the training data set, determines the industry large model of the target industry, determines the evaluation value of each industry sub-model in the industry large model based on the test training set and the industry large model, and determines the model evaluation data of the industry large model; Optimization module: Iteratively optimizes the industry big model based on the industry task requirement set and model evaluation data of the industry big model.

Images