Model optimization method and system based on collaborative intelligence of large and small models

Abstract

The present invention relates to the field of data processing technology, and specifically discloses a model optimization method and system based on large and small model collaborative intelligence, including: obtaining multi-source heterogeneous data, performing feature extraction on the multi-source heterogeneous data based on a large language model, and obtaining multiple key feature information; obtaining entity information based on the key feature information, and performing large model training on all the entity information to obtain a large model training knowledge graph; obtaining a first entity evaluation indicator and a first change trend corresponding to each entity information based on the large model training knowledge graph, and screening the entity information based on the first entity evaluation indicator to obtain key feature entities. The present invention obtains multi-source heterogeneous data, uses a large language model to extract key feature information from unstructured and structured data, constructs a knowledge graph, and performs entity relationship reasoning, information completion, and semantic understanding, thereby improving the comprehensiveness and accuracy of large and small models.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a model optimization method and system based on collaborative intelligence of large and small models. Background Art

[0002] In modern data-driven intelligent systems, model optimization is a key link to improve system performance and reliability. With the development of big data and artificial intelligence technologies, various application scenarios have higher and higher requirements for models, especially in the fields of mobile network optimization, smart city management, and industrial production optimization.

[0003] Existing model optimization methods usually rely on a single data source, such as using only structured data (database records, sensor data, etc.) or unstructured data (text, images, etc.). The use of such a single data source is prone to missing important information, resulting in insufficient comprehensiveness and accuracy of the model. For example, existing feature extraction methods have difficulty processing complex unstructured data such as text and images, and are prone to losing key information. These methods usually only focus on evaluation indicators of a single dimension, making it difficult to fully evaluate the importance and impact of entities. Summary of the invention

[0004] The purpose of the present invention is to provide a model optimization method and system based on large and small model collaborative intelligence to solve the technical problems raised in the above background technology.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] Model optimization method based on collaborative intelligence of large and small models, including:

[0007] Obtain multi-source heterogeneous data, extract features from the multi-source heterogeneous data based on a large language model, and obtain multiple key feature information;

[0008] Acquire entity information according to the key feature information, and perform large model training on all the entity information to obtain a large model training knowledge graph;

[0009] Obtaining a first entity evaluation index and a first change trend corresponding to each entity information according to the large model training knowledge graph, and screening the entity information according to the first entity evaluation index to obtain a key feature entity;

[0010] Performing collaborative training of a small model on each of the key feature entities to obtain a second entity evaluation index and a second change trend corresponding to each of the key feature entities;

[0011] Acquire a performance evaluation value according to the second entity evaluation index and the second change trend;

[0012] Determine whether the performance evaluation value is greater than or equal to a preset threshold;

[0013] If it is greater than or equal to, it is determined that the multi-source heterogeneous data has been optimized;

[0014] If it is less than, it is judged that the multi-source heterogeneous data has not been optimized, and the artificial graph optimization index corresponding to the second entity evaluation index is obtained, and the artificial graph optimization index, the second entity evaluation index and the second change trend are re-trained in the large model until the multi-source heterogeneous data are all optimized.

[0015] Preferably, the step of extracting features from multi-source heterogeneous data based on a large language model to obtain multiple key feature information includes:

[0016] Classifying the multi-source heterogeneous data to obtain unstructured data and structured data;

[0017] Performing data preprocessing on the unstructured data to obtain first data to be processed;

[0018] Normalizing the structured data to obtain second data to be processed;

[0019] Extracting key text and image information from the first data to be processed based on the large language model to obtain key text and image information;

[0020] Extracting multi-dimensional key information from the second data to be processed based on the large language model to obtain multi-dimensional key information;

[0021] The key text image information and the multi-dimensional key information are integrated to obtain key feature information.

[0022] Preferably, the step of acquiring entity information according to the key feature information and performing large model training on all the entity information to obtain a large model training knowledge graph includes:

[0023] Performing entity feature extraction on the key feature information to obtain multiple entity information;

[0024] Obtaining entity relationships between each entity feature information;

[0025] Establishing a graph node according to the entity information;

[0026] Establish a knowledge graph based on the graph nodes and entity relationships;

[0027] Based on the big model training, the knowledge graph is trained in entity relationship thrust, information completion and semantic understanding to obtain the big model training knowledge graph.

[0028] Preferably, the step of obtaining a first entity evaluation index and a first change trend corresponding to each entity information according to the large model training knowledge graph, and screening the entity information according to the first entity evaluation index to obtain a key feature entity includes:

[0029] Obtaining the entity type corresponding to each of the entity information;

[0030] Obtaining dimension evaluation indicators according to the entity type;

[0031] Acquire a first entity evaluation indicator corresponding to each entity information based on the dimension evaluation indicator;

[0032] Acquire historical data corresponding to each entity information based on the large model training knowledge graph;

[0033] Acquire a change direction based on the historical data, and use the change direction as a first change trend;

[0034] Determining whether the first entity evaluation index is greater than or equal to a preset value;

[0035] If it is less than, it is determined that the entity information corresponding to the first entity evaluation indicator does not meet the conditions, and the entity information that does not meet the conditions is not used as a key feature entity;

[0036] If it is greater than or equal to, it is determined that the entity information corresponding to the first entity evaluation indicator meets the conditions, and the entity information that meets the conditions is used as the key feature entity.

[0037] Preferably, the step of performing small model collaborative training on each of the key feature entities to obtain a second entity evaluation index and a second change trend corresponding to each of the key feature entities includes:

[0038] Acquire representative type data according to the key feature entity;

[0039] Acquire and configure a small model according to the representative type data;

[0040] Dividing the representative type data to obtain a plurality of batch training data;

[0041] Training each batch of training data to obtain batch evaluation indicators;

[0042] Acquire a second entity evaluation indicator corresponding to the key feature entity according to each of the batch evaluation indicators;

[0043] Obtain prediction result information according to each of the batch training data;

[0044] A second change trend is obtained based on the prediction result information.

[0045] Preferably, the step of acquiring a performance evaluation value according to the second entity evaluation index and the second change trend includes:

[0046] Obtaining a second entity evaluation indicator;

[0047] Obtaining a first evaluation coefficient according to the second entity evaluation indicator;

[0048] Obtaining a change rate according to a second change trend;

[0049] Acquire a second evaluation coefficient according to the change rate;

[0050] A performance evaluation value is calculated based on the second entity evaluation index, the first evaluation coefficient, the change rate, and the second evaluation coefficient.

[0051] The present invention also discloses a model optimization system based on large and small model collaborative intelligence, including:

[0052] The first acquisition module is used to acquire multi-source heterogeneous data, perform feature extraction on the multi-source heterogeneous data based on a large language model, and obtain multiple key feature information;

[0053] A second acquisition module is used to acquire entity information according to the key feature information, and perform large model training on all the entity information to obtain a large model training knowledge graph;

[0054] A third acquisition module is used to obtain a first entity evaluation index and a first change trend corresponding to each entity information according to the large model training knowledge graph, and filter the entity information according to the first entity evaluation index to obtain a key feature entity;

[0055] A fourth acquisition module is used to perform small model collaborative training on each of the key feature entities to obtain a second entity evaluation index and a second change trend corresponding to each of the key feature entities;

[0056] A fifth acquisition module, configured to acquire a performance evaluation value according to the second entity evaluation indicator and the second change trend;

[0057] A judgment module is used to judge whether the performance evaluation value is greater than or equal to a preset threshold;

[0058] If it is greater than or equal to, it is determined that the multi-source heterogeneous data has been optimized;

[0059] If it is less than, it is judged that the multi-source heterogeneous data has not been optimized, and the artificial graph optimization index corresponding to the second entity evaluation index is obtained, and the artificial graph optimization index, the second entity evaluation index and the second change trend are re-trained in the large model until the multi-source heterogeneous data are all optimized.

[0060] Preferably, the second acquisition module includes:

[0061] A first acquisition unit, used for classifying the multi-source heterogeneous data to obtain unstructured data and structured data;

[0062] A second acquisition unit, configured to perform data preprocessing on the unstructured data to obtain first data to be processed;

[0063] A third acquisition unit is used to perform normalization processing on the structured data to obtain second data to be processed;

[0064] A fourth acquisition unit is used to extract key text and picture information from the first data to be processed based on the large language model to obtain key text and picture information;

[0065] a fifth acquisition unit, configured to extract multi-dimensional key information from the second data to be processed based on the large language model to obtain the multi-dimensional key information;

[0066] The sixth acquisition unit is used to integrate the key text image information and the multi-dimensional key information to obtain key feature information.

[0067] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the above-mentioned model optimization method based on collaborative intelligence of large and small models when executing the computer program.

[0068] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-mentioned model optimization method based on collaborative intelligence of large and small models.

[0069] The beneficial effects of the present application are as follows: the present invention obtains multi-source heterogeneous data, uses a large language model to extract key feature information from unstructured and structured data, constructs a knowledge graph, and performs entity relationship reasoning, information completion, and semantic understanding, thereby improving the comprehensiveness and accuracy of large and small models. Through multi-dimensional evaluation indicators and change trend analysis, key feature entities are screened, redundant information is reduced, and processing efficiency is improved. Small model collaborative training runs efficiently on terminal devices, reduces resource consumption, and improves the pertinence and adaptability of the model. Through comprehensive evaluation and dynamic adjustment of model parameters, explainable performance evaluation results are provided, and automation and continuous optimization of model optimization are achieved. These improvements significantly improve the comprehensiveness, accuracy, and optimization effect of the model, and solve the problems of single data source, difficult feature extraction, single evaluation indicator, high resource consumption, and low optimization efficiency in existing methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 The figure is a schematic diagram of a method flow of an embodiment of the present application.

[0071] Figure 2 A schematic diagram of the system structure of an embodiment of the present application.

[0072] The purpose, features and advantages of this application will be further described in conjunction with the embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0073] It should be understood that the specific embodiments described herein are only used to explain the present invention, and are not used to limit the present invention.

[0074] like Figure 1 As shown, the present application provides a model optimization method based on large and small model collaborative intelligence, including:

[0075] S1. Obtain multi-source heterogeneous data, extract features from the multi-source heterogeneous data based on a large language model, and obtain multiple key feature information;

[0076] S2. Acquire entity information according to the key feature information, and perform large model training on all the entity information to obtain a large model training knowledge graph;

[0077] S3. Obtaining a first entity evaluation index and a first change trend corresponding to each entity information according to the large model training knowledge graph, and screening the entity information according to the first entity evaluation index to obtain a key feature entity;

[0078] S4, performing small model collaborative training on each of the key feature entities to obtain a second entity evaluation index and a second change trend corresponding to each of the key feature entities;

[0079] S5. Obtaining a performance evaluation value according to the second entity evaluation index and the second change trend;

[0080] S6. Determine whether the performance evaluation value is greater than or equal to a preset threshold;

[0081] If it is greater than or equal to, it is determined that the multi-source heterogeneous data has been optimized;

[0082] If it is less than, it is judged that the multi-source heterogeneous data has not been optimized, and the artificial graph optimization index corresponding to the second entity evaluation index is obtained, and the artificial graph optimization index, the second entity evaluation index and the second change trend are re-trained in the large model until the multi-source heterogeneous data are all optimized.

[0083] As described in steps S1-S6 above, existing model optimization methods usually rely on a single data source, such as using only structured data (database records, sensor data, etc.) or unstructured data (text, images, etc.). The use of such a single data source is prone to missing important information, resulting in insufficient comprehensiveness and accuracy of the model;

[0084] By acquiring multi-source heterogeneous data, the present invention can obtain rich information from multiple angles and sources, and improve the comprehensiveness and accuracy of large and small models. By utilizing the powerful natural language processing capabilities of the large language model, key information can be extracted from unstructured data, such as key features in text and pictures, and multi-dimensional key information can also be extracted from structured data. In addition, the unstructured data is preprocessed and the structured data is normalized to ensure the consistency and comparability of the data, improve the efficiency and accuracy of subsequent processing, and solve the problem that the existing data processing methods usually rely on a single data source, which is easy to miss important information and affect the comprehensiveness and accuracy of the model. The existing feature extraction method is difficult to process complex unstructured data, such as text and pictures, and is easy to lose key information; then, through the extraction of key feature information, each entity and its attributes can be accurately identified, providing a basis for the subsequent construction of the knowledge graph. Obtaining the relationship between each entity and constructing the association graph between entities helps to understand the complex relationship between entities. Based on the knowledge graph trained by the large model, entity relationship reasoning, information completion and semantic understanding can be performed, the integrity and accuracy of the knowledge graph can be improved, and the existing entity recognition method may miss or misdetect, affecting the accuracy of subsequent analysis. Existing relationship extraction methods may not be able to capture complex entity relationships, resulting in incompleteness of the knowledge graph. Then, by defining multi-dimensional evaluation indicators for each entity, the importance and impact of the entity can be fully evaluated. By analyzing historical data, the change trend of each entity is obtained, which helps to understand the change of entities over time. According to the evaluation indicators and change trends, key feature entities are screened out to reduce redundant information, improve the efficiency and accuracy of subsequent processing, and solve the problem that existing evaluation methods may only focus on indicators of a single dimension and it is difficult to fully evaluate the importance and impact of entities. Existing data analysis methods may not be able to capture the change trend of entities, affecting subsequent decision-making and optimization. Then, through the collaborative training of small models, it can run efficiently on terminal devices, reduce resource consumption, and improve the deployment and reasoning efficiency of models. Small models can be personalized according to the characteristics of key feature entities to improve the pertinence and adaptability of the model. Through small model training, the second entity evaluation indicator and the second change trend of each key feature entity are obtained, the evaluation indicators are updated and optimized, and the accuracy and reliability of the model are improved. The existing large-scale models consume large resources when running on terminal devices and are difficult to apply on a large scale. Existing models may lack personalized training capabilities and are difficult to adapt to the needs of different scenarios. Through the second entity evaluation indicator and the second change trend, the performance of the model is comprehensively evaluated to provide a comprehensive evaluation result. According to the performance evaluation value, the model parameters and training strategy are dynamically adjusted to improve the optimization effect of the model. Through detailed evaluation indicators and change trends, explainable performance evaluation results are provided to facilitate understanding and optimization, solving the problem that existing performance evaluation methods may only focus on a single indicator and are difficult to comprehensively evaluate the performance of the model.Existing model optimization methods may lack dynamic adjustment capabilities, which affects the optimization effect. Finally, through the judgment of performance evaluation values, the model optimization is automated to improve the optimization efficiency and effect. According to the performance evaluation results, iterative optimization is performed to gradually improve the performance of the model. When the performance evaluation value is lower than the preset threshold, it indicates that the multi-source heterogeneous data has not been optimized. At this time, it is necessary to obtain the artificial graph optimization index corresponding to the second entity evaluation index. The specific steps are as follows: First, the second entity evaluation index of each key feature entity is deeply analyzed to identify the deficiencies of the model in entity recognition, relationship extraction or semantic understanding. Based on this analysis, domain experts formulate targeted artificial optimization plans to solve the identified problem points. According to the plan, optimization indicators that clearly point out the direction of improvement and can be quantified are generated, such as improving the accuracy of specific entity classification or enhancing the accuracy of relationship prediction. Subsequently, these artificial graph optimization indicators are integrated with the original second entity evaluation indicators and change trends to form a new training data set to ensure that the newly added optimization part can guide the subsequent large model training. Finally, the large model is retrained using the updated data set to gradually improve the previous problems and improve the overall performance until the optimization requirements are met, forming a closed-loop feedback mechanism to continuously optimize the model. Solve the problem that existing model optimization methods may require manual adjustment of parameters, resulting in low optimization efficiency. Existing optimization methods may lack iterative optimization capabilities, affecting optimization results. Existing optimization methods may lack a closed-loop feedback mechanism, making it difficult to continuously optimize model points.

[0085] In one embodiment, the step of extracting features from multi-source heterogeneous data based on a large language model to obtain multiple key feature information includes:

[0086] S101, classifying the multi-source heterogeneous data to obtain unstructured data and structured data;

[0087] S102, performing data preprocessing on the unstructured data to obtain first data to be processed;

[0088] S103, normalizing the structured data to obtain second data to be processed;

[0089] S104, extracting key text and image information from the first data to be processed based on the large language model to obtain key text and image information;

[0090] S105, extracting multi-dimensional key information from the second data to be processed based on the large language model to obtain multi-dimensional key information;

[0091] S106: Integrate the key text image information and the multi-dimensional key information to obtain key feature information.

[0092] As described in the above steps S101-S106, the present invention can effectively extract key feature information from multi-source heterogeneous data through the above steps, and the specific steps include data classification, data preprocessing, normalization, feature extraction and information integration. These steps not only solve the various defects existing in the existing methods, but also significantly improve the efficiency and accuracy of data processing, enhance the comprehensiveness and expression ability of the model, and improve the training effect and performance of the model. First, by classifying multi-source heterogeneous data, unstructured data and structured data can be clearly distinguished, the efficiency and accuracy of data processing can be improved, and improper processing caused by data mixing can be avoided. Secondly, data preprocessing of unstructured data can remove noise and irrelevant information in unstructured data, improve the quality and readability of data, and make data easier to be understood and processed by the model. Next, normalization of structured data can unify the numerical range of data, eliminate the influence of data of different scales, improve the training effect and convergence speed of the model, and reduce the risk of overfitting. Then, based on the large language model, the text and image key information of the first data to be processed is extracted. The powerful natural language processing capabilities and computer vision technology of the large language model can be used to extract deep semantic information and key features from the text and image data, realize the fusion of multimodal information, and improve the comprehensiveness and accuracy of feature extraction. Next, based on the large language model, multi-dimensional key information is extracted from the second data to be processed. Multi-dimensional key information can be extracted from structured data, the correlation between different features can be analyzed, complex patterns and laws can be identified, and the prediction and classification capabilities of the model can be improved. Finally, the extracted key text image information and multi-dimensional key information are integrated to form comprehensive key feature information, which can better express the intrinsic characteristics and relationships of the data, enhance the expression ability of the model, and improve the training effect and performance of the model. Through these steps, the present invention not only solves the problems existing in the existing methods, such as data mixing, single processing method, data noise, inconsistent format, poor expression ability and low model performance, but also obtains multi-source heterogeneous data, and uses a large language model to extract key feature information from unstructured and structured data, constructs a knowledge graph and performs entity relationship reasoning, information completion and semantic understanding, thereby significantly improving the comprehensiveness and accuracy of large and small models. Specifically, this method not only optimizes the capabilities of the large model, enabling it to more efficiently process complex multi-source data and mine valuable information from it, but also enhances the operating efficiency and pertinence on the terminal device through the collaborative training of small models, and reduces resource consumption. For the large model, the present invention reduces redundant information and improves the quality and accuracy of the knowledge graph trained by the large model by screening out key feature entities through multi-dimensional evaluation indicators and change trend analysis. At the same time, through continuous large model training and iterative optimization, its ability to understand and adapt to complex data environments is ensured.As for the small model, the present invention realizes efficient operation on the terminal device, and improves the model deployment and reasoning efficiency by personalizing the small model to adapt to the needs of different scenarios. In addition, the small model collaborative training process based on key feature entities further improves the pertinence and adaptability of the model, ensures the performance of the model in practical applications, improves the efficiency and accuracy of data processing, enhances the comprehensiveness and expression ability of the model, and improves the training effect and performance of large and small models.

[0093] In one embodiment, the step of acquiring entity information according to the key feature information and performing large model training on all the entity information to obtain a large model training knowledge graph includes:

[0094] S201, extracting entity features from the key feature information to obtain multiple entity information;

[0095] S202, obtaining the entity relationship between each entity feature information;

[0096] S203, establishing a graph node according to the entity information;

[0097] S204, establishing a knowledge graph based on the graph nodes and entity relationships;

[0098] S205. Based on the big model training, the knowledge graph is trained in entity relationship thrust, information completion and semantic understanding to obtain a big model training knowledge graph.

[0099] As described in the above steps S201-S205, through the above steps, the present invention can effectively extract entity information from key feature information and construct a complete large-model training knowledge graph. The above steps not only solve the various defects existing in the existing methods, but also significantly improve the efficiency and accuracy of data processing, enhance the comprehensiveness and expression ability of the knowledge graph, and improve the quality and reliability of the knowledge graph. First, by extracting entity features from key feature information, each entity and its attributes can be accurately identified, providing a basis for subsequent knowledge graph construction. This step solves the problems of inaccurate entity recognition and incomplete coverage in existing methods, and improves the efficiency and accuracy of data processing. Secondly, obtaining the entity relationship between each entity feature information can clearly show the association between each entity, which helps to understand the complex relationship between entities. This step solves the problems of fuzzy relationships and poor structuring in existing methods, and enhances the understanding of data. Next, we establish graph nodes based on entity information (entity information refers to objects or concepts extracted from multi-source heterogeneous data that have physical or business logic meanings. These objects can be individuals, events, locations, organizations, etc. that actually exist in a specific application field, or they can be abstract concepts such as relationships, attributes, or states). In the application scenario of image recognition, entity information may refer to people, objects, landmark buildings, etc. in the image; in the scenario of financial risk assessment, entity information may be customer information, transaction records, market trends, etc. For example, mobile network optimization: entity information can be the location information of user devices, network connection quality, traffic usage, etc.; smart city management: entity information covers urban infrastructure (such as traffic lights, public facilities), citizen activity patterns, environmental monitoring data, etc.; industrial production optimization: entity information includes status parameters of production equipment, product defect detection results on the production line, raw material inventory levels, etc. After feature extraction of multi-source heterogeneous data based on a large language model, we can identify and obtain these entity information with practical application scenario meanings, and further analyze and process them, so that each entity can be clearly represented, which is convenient for subsequent knowledge graph construction and management. This step solves the problems of node confusion and inconvenient management in existing methods, making the management of entity information more convenient. Subsequently, a knowledge graph is established based on graph nodes and entity relationships, which can fully display the relationships between entities and improve data availability and query efficiency. This step solves the problems of incomplete structure and inconvenient query in existing methods and enhances the scalability of the knowledge graph. Finally, based on large model training, the knowledge graph is trained for entity relationship thrust, information completion, and semantic understanding. First, it is necessary to build an initial graph and prepare data. Each entity and its attributes are established by identifying key feature information, and manual annotations are added to some entities and relationships as the basis for supervised learning.Next, select or develop a pre-trained large language model (such as BERT, T5, GPT series) and fine-tune it on the knowledge graph task in a specific field. Use the large model to analyze the potential associations between entities, automatically discover implicit relationship types, and ensure the accuracy of newly added relationships through logical consistency checks. For incomplete or missing information, use the contextual understanding and reasoning capabilities of the large model to supplement it, and verify it in combination with authoritative data sources. In terms of semantic understanding, deeply parse the text content and enhance the semantic hierarchy in the knowledge graph, including tasks such as named entity recognition and sentiment analysis. At the same time, train the model to understand the meaning of different entities in various contexts. The whole process is continuously iterated and improved, external authoritative data sources are introduced as references, and an evaluation mechanism is set up to regularly check the integrity, consistency and timeliness of the knowledge graph. Finally, after a series of operations, a large model training knowledge graph rich in semantic information, comprehensive and accurate is constructed, which can discover potential associations, automatically complete missing information, enhance the semantic understanding ability of the knowledge graph, and improve the overall quality of the knowledge graph. This step solves the problems of shallow relationships, incomplete information, poor semantic understanding and low quality in existing methods, making the knowledge graph more reliable and useful.

[0100] In one embodiment, the step of obtaining a first entity evaluation index and a first change trend corresponding to each entity information according to the large model training knowledge graph, and screening the entity information according to the first entity evaluation index to obtain a key feature entity includes:

[0101] S301, obtaining the entity type corresponding to each entity information;

[0102] S302, obtaining a dimension evaluation index according to the entity type;

[0103] S303: Obtain a first entity evaluation indicator corresponding to each entity information based on the dimension evaluation indicator;

[0104] S304: Acquire historical data corresponding to each entity information based on the large model training knowledge graph;

[0105] S305, obtaining a change direction based on the historical data, and taking the change direction as a first change trend;

[0106] S306: Determine whether the first entity evaluation index is greater than or equal to a preset value;

[0107] If it is less than, it is determined that the entity information corresponding to the first entity evaluation indicator does not meet the conditions, and the entity information that does not meet the conditions is not used as a key feature entity;

[0108] If it is greater than or equal to, it is determined that the entity information corresponding to the first entity evaluation indicator meets the conditions, and the entity information that meets the conditions is used as the key feature entity.

[0109] As described in the above steps S301-S306, through the above steps, the present invention can effectively obtain the first entity evaluation index and the first change trend of each entity from the large model training knowledge graph, and screen the entity information according to the evaluation index to obtain the key feature entity. The above steps not only solve the various defects existing in the existing methods, but also significantly improve the efficiency and accuracy of data processing, enhance the comprehensiveness and scientificity of the evaluation, and improve the accuracy and effectiveness of the screening. First, by obtaining the type of each entity, different types of entities can be clearly distinguished, providing a basis for subsequent evaluation and screening, and solving the problems of unclear classification and inconvenient management in the existing methods. Then, by obtaining multi-dimensional evaluation indicators according to the entity type, the importance and impact of the entity can be comprehensively evaluated, the flexibility and accuracy of the evaluation are improved, and the problems of single evaluation and poor flexibility in the existing methods are solved. Then, based on the multi-dimensional evaluation indicators, the first entity evaluation indicators of each entity are obtained (the first entity evaluation indicators represent the multi-dimensional quantitative evaluation of each entity in its field or application scenario. These indicators comprehensively reflect the importance, influence, performance and other aspects of the entity, and provide a basis for the subsequent screening of key feature entities. In terms of importance, the relative importance of the entity in the entire system or application is measured. For example, in an intelligent transportation system, vehicle flow information at certain key intersections may be more important than other secondary roads. In terms of influence, the scope and intensity of the entity's influence on other related entities are evaluated. For example, in industrial production, the state change of a core equipment may have a significant impact on the entire production line. In terms of performance, the operating efficiency or quality level of the entity in its current state is described. For example, in financial risk assessment, customer credit scores directly affect the speed and security of the loan approval process), ensuring the comprehensiveness and scientificity of the evaluation results, and solving the problems of one-sided and subjective evaluation in existing methods. Subsequently, the historical data of each entity is obtained from the large model training knowledge graph, which can obtain rich historical information, improve the accuracy and reliability of the evaluation results, and solve the problems of insufficient data and low accuracy in existing methods. Next, based on historical data, the change direction is obtained and used as the first change trend, which can clarify the change direction of the entity, improve the prediction ability and decision support of the model, and solve the problems of unclear trend and poor predictability in existing methods. Finally, judging whether the entity meets the conditions based on the first entity evaluation index and screening out key feature entities can reduce redundant information, improve the efficiency of subsequent processing and the scientific nature of decision-making, and solve the problems of inaccurate screening and excessive redundant information in existing methods.

[0110] In one embodiment, the step of performing small model collaborative training on each of the key feature entities to obtain the second entity evaluation index and the second change trend corresponding to each of the key feature entities includes:

[0111] S401, acquiring representative type data according to the key feature entity;

[0112] S402, acquiring a small model according to the representative type data;

[0113] S403, dividing the representative type data to obtain a plurality of batch training data;

[0114] S404, training each batch of training data to obtain batch evaluation indicators;

[0115] S405, obtaining a second entity evaluation indicator corresponding to the key feature entity according to each of the batch evaluation indicators;

[0116] S406, obtaining prediction result information according to each batch of training data;

[0117] S407: Obtain a second change trend based on the prediction result information.

[0118] As described in the above steps S401-S407, the present invention can effectively train each key feature entity with a small model. In order to obtain and configure a small model based on representative type data, it is first necessary to deeply analyze the characteristics, distribution and application requirements of these data. For example, in image recognition tasks, the resolution and color mode are understood, while in text analysis, the text length and language type are focused on. Based on this analysis, the most suitable small model architecture is selected: for image data, a convolutional neural network (CNN) such as ResNet or EfficientNet is selected; for text or sequence data, a recurrent neural network (RNN), a long short-term memory network (LSTM), a gated recurrent unit (GRU) or a transformer is considered. After clarifying the input and output formats of the model, the necessary hyperparameters are set, including the learning rate, batch size, number of training rounds and optimizer type. Select an appropriate loss function (such as cross entropy loss for classification tasks) and evaluation indicators (such as accuracy, F1 score) to monitor the model performance during the training process. In the process of small model collaborative training, first select representative type data according to the characteristics of the key feature entity to ensure the representativeness of the training data. Next, we configure a suitable small model architecture for this data, such as a convolutional neural network (CNN) for image recognition or a transformer for text analysis. Then, we divide the data into multiple batch training sets to reduce memory usage and improve training efficiency. Each batch training set is trained independently, and batch evaluation indicators are obtained to monitor model performance in real time. The second entity evaluation indicators are obtained by combining all batch evaluation results to ensure comprehensive and accurate evaluation. For example, in an intelligent traffic management system, the goal is to optimize the traffic flow of urban roads and reduce congestion. To this end, we select "vehicle flow at major intersections" as the key feature entity for small model collaborative training. From historical surveillance camera and sensor data, we select vehicle flow records of multiple typical intersections in different time periods (such as peak hours and off-peak hours). These data include not only the number of vehicles, but also information such as vehicle type and driving direction to ensure that the training samples can fully reflect the actual traffic conditions. According to the characteristics of vehicle flow data, we select a long short-term memory network (LSTM), a recurrent neural network (RNN) that is particularly suitable for time series prediction, and divide the obtained data set into several small batches in chronological order, such as hourly or daily batches. Doing so can simulate the dynamic changes of the real world, while reducing the pressure of single training, improving training efficiency, and independently training each small batch to calculate evaluation indicators such as mean square error (MSE) and mean absolute error (MAE) to measure the prediction accuracy of the model. In this way, hyperparameters can be adjusted in real time during the training process to optimize model performance, and then the second entity evaluation indicators and second change trends corresponding to each key feature entity can be obtained.The above steps not only solve the various defects existing in the existing methods, but also significantly improve the efficiency and accuracy of data processing, and enhance the generalization ability, stability, prediction ability and decision support ability of the model. First, by obtaining representative type data, the representativeness of the training data can be ensured, the generalization ability of the model can be improved, and the problems of poor representativeness and low data quality in the existing methods can be solved. Then, selecting a suitable small model according to the representative type data can improve resource utilization, reduce computing costs, improve the adaptability and flexibility of the model, and solve the problems of poor adaptability and low resource utilization in the existing methods. Then, training the representative type data in batches can improve training efficiency, reduce memory usage, and improve the stability of model training, solving the problems of low training efficiency and poor stability in the existing methods. Subsequently, training each batch of training data and obtaining batch evaluation indicators can comprehensively evaluate the performance of the model on different data batches, provide real-time feedback, help adjust the training strategy, and solve the problems of one-sided evaluation and delayed feedback in the existing methods. Then, by combining multiple batch evaluation indicators, the performance of key feature entities can be comprehensively evaluated, ensuring the accuracy and reliability of the evaluation results, and solving the problems of incomplete evaluation and poor scientificity in the existing methods. Then, by obtaining prediction result information from batch training data, the prediction accuracy of the model can be improved, real-time prediction results can be provided, and prediction problems of the model can be discovered early, solving the problems of inaccurate prediction and poor real-time performance in existing methods. Finally, by obtaining the second change trend based on the prediction result information, the change direction of key feature entities can be clarified, the prediction ability of the model can be improved, and strong support can be provided for decision-making, solving the problems of unclear trends and weak decision support in existing methods.

[0119] In one embodiment, the step of acquiring a performance evaluation value according to the second entity evaluation index and the second change trend includes:

[0120] S501, obtaining a second entity evaluation index;

[0121] S502, obtaining a first evaluation coefficient according to the second entity evaluation indicator;

[0122] S503, obtaining a change rate according to a second change trend;

[0123] S504, obtaining a second evaluation coefficient according to the change rate;

[0124] S505: Calculate a performance evaluation value according to the second entity evaluation index, the first evaluation coefficient, the change rate, and the second evaluation coefficient, wherein the calculation formula is:

[0125] ;

[0126] in, represents the warning assessment value, represents the first evaluation coefficient, represents the second entity evaluation index, represents the second evaluation coefficient, Indicates the rate of change.

[0127] As described in the above steps S501-S505, the present invention not only solves the various defects existing in the existing methods, but also significantly improves the efficiency and accuracy of data processing, enhances the comprehensiveness and scientificity of the evaluation, and improves the accuracy and reliability of the performance evaluation. First, by obtaining the second entity evaluation index, the performance of each key feature entity after the small model training can be comprehensively evaluated to ensure the comprehensiveness and accuracy of the evaluation results, and solve the problems of one-sided evaluation and strong subjectivity in the existing methods. Then, according to the second entity evaluation index, the first evaluation coefficient is obtained (the importance of each second entity evaluation index is determined by expert scoring, historical data statistics or machine learning methods. For example, assuming that the importance scores of evaluation index A and change rate B are 8 and 5 respectively, the importance of each evaluation index and change rate is standardized to ensure that the sum of all weights is 1. Calculate the total score: 8 + 5 = 13; calculate the standardized weight: A's weight is 8 / 13 ≈0.615, B's weight is 5 / 13 ≈0.385), the weights of different evaluation indicators can be reasonably allocated to ensure the rationality of the evaluation results, and solve the problems of unreasonable weights and poor flexibility in the existing methods. Then, by obtaining the change rate according to the second change trend, the change direction and speed of the key feature entity can be clarified, the dynamic adaptability and predictive ability of the model can be improved, and the problems of unclear trend and poor predictability in the existing methods can be solved. Subsequently, the second evaluation coefficient is obtained according to the change rate (the degree of influence of each change rate on the performance evaluation is determined through historical data statistics or machine learning methods), and the weight of the evaluation index can be dynamically adjusted to ensure the dynamic adaptability of the evaluation result, which solves the problems of static evaluation and poor flexibility in the existing methods. Finally, by comprehensively considering the second entity evaluation index, the first evaluation coefficient, the change rate and the second evaluation coefficient, the performance of the model can be comprehensively evaluated to ensure the accuracy and reliability of the evaluation results, which solves the problems of incomplete evaluation and poor scientificity in the existing methods. The calculation process of the performance evaluation value is transparent, the results are interpretable, easy to understand and optimize, and solves the problem of poor interpretability in the existing methods.

[0128] like Figure 2 As shown, the present invention also provides a model optimization system based on large and small model collaborative intelligence, including:

[0129] The first acquisition module 1 is used to acquire multi-source heterogeneous data, extract features from the multi-source heterogeneous data based on a large language model, and obtain multiple key feature information;

[0130] The second acquisition module 2 is used to acquire entity information according to the key feature information, and perform large model training on all the entity information to obtain a large model training knowledge graph;

[0131] The third acquisition module 3 is used to obtain the first entity evaluation index and the first change trend corresponding to each entity information according to the large model training knowledge graph, and screen the entity information according to the first entity evaluation index to obtain the key feature entity;

[0132] A fourth acquisition module 4 is used to perform small model collaborative training on each of the key feature entities to obtain a second entity evaluation index and a second change trend corresponding to each of the key feature entities;

[0133] A fifth acquisition module 5, configured to acquire a performance evaluation value according to the second entity evaluation index and the second change trend;

[0134] A judgment module 6 is used to judge whether the performance evaluation value is greater than or equal to a preset threshold;

[0135] If it is greater than or equal to, it is determined that the multi-source heterogeneous data has been optimized;

[0136] If it is less than, it is judged that the multi-source heterogeneous data has not been optimized, and the artificial graph optimization index corresponding to the second entity evaluation index is obtained, and the artificial graph optimization index, the second entity evaluation index and the second change trend are re-trained in the large model until the multi-source heterogeneous data are all optimized.

[0137] In one embodiment, the first acquisition module 1 includes:

[0138] A first acquisition unit, used for classifying the multi-source heterogeneous data to obtain unstructured data and structured data;

[0139] A second acquisition unit, configured to perform data preprocessing on the unstructured data to obtain first data to be processed;

[0140] A third acquisition unit is used to perform normalization processing on the structured data to obtain second data to be processed;

[0141] A fourth acquisition unit is used to extract key text and picture information from the first data to be processed based on the large language model to obtain key text and picture information;

[0142] a fifth acquisition unit, configured to extract multi-dimensional key information from the second data to be processed based on the large language model to obtain the multi-dimensional key information;

[0143] The sixth acquisition unit is used to integrate the key text image information and the multi-dimensional key information to obtain key feature information.

[0144] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the above-mentioned model optimization method based on collaborative intelligence of large and small models when executing the computer program.

[0145] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-mentioned model optimization method based on collaborative intelligence of large and small models.

[0146] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, value library or other media provided in this application and used in the embodiments may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0147] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, device, article or method including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, device, article or method. In the absence of further restrictions, an element defined by the sentence "includes a ..." does not exclude the presence of other identical elements in the process, device, article or method including the element.

[0148] The above description is only a preferred embodiment of the present invention, and does not limit the patent scope of the present invention. Any equivalent results or equivalent process changes made using the contents of the present invention description and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A model optimization method based on collaborative intelligence of large and small models, characterized in that: include: Acquire multi-source heterogeneous data, classify the multi-source heterogeneous data, and obtain unstructured data and structured data; Performing data preprocessing on the unstructured data to obtain first data to be processed; Normalizing the structured data to obtain second data to be processed; Extracting key text and image information from the first data to be processed based on the large language model to obtain key text and image information; Extracting multi-dimensional key information from the second data to be processed based on the large language model to obtain multi-dimensional key information; Integrate the key text image information and the multi-dimensional key information to obtain key feature information; Acquire entity information according to the key feature information, and perform large model training on all the entity information to obtain a large model training knowledge graph; Obtaining a first entity evaluation index and a first change trend corresponding to each entity information according to the large model training knowledge graph, and screening the entity information according to the first entity evaluation index to obtain a key feature entity; Performing collaborative training of a small model on each of the key feature entities to obtain a second entity evaluation index and a second change trend corresponding to each of the key feature entities; Acquire a performance evaluation value according to the second entity evaluation index and the second change trend; Determine whether the performance evaluation value is greater than or equal to a preset threshold; If it is greater than or equal to, it is determined that the multi-source heterogeneous data has been optimized; If it is less than, it is judged that the multi-source heterogeneous data has not been optimized, and the artificial graph optimization index corresponding to the second entity evaluation index is obtained, and the artificial graph optimization index, the second entity evaluation index and the second change trend are re-trained in the large model until the multi-source heterogeneous data are all optimized.

2. The model optimization method based on large and small model collaborative intelligence according to claim 1 is characterized in that: The step of acquiring entity information according to the key feature information, and performing large model training on all the entity information to obtain a large model training knowledge graph includes: Performing entity feature extraction on the key feature information to obtain multiple entity information; Obtaining entity relationships between each entity feature information; Establishing a graph node according to the entity information; Establish a knowledge graph based on the graph nodes and entity relationships; Based on the big model training, the knowledge graph is trained in entity relationship thrust, information completion and semantic understanding to obtain the big model training knowledge graph.

3. The model optimization method based on large and small model collaborative intelligence according to claim 1 is characterized in that: The step of obtaining a first entity evaluation index and a first change trend corresponding to each entity information according to the large model training knowledge graph, and screening the entity information according to the first entity evaluation index to obtain a key feature entity includes: Obtaining the entity type corresponding to each of the entity information; Obtaining dimension evaluation indicators according to the entity type; Acquire a first entity evaluation indicator corresponding to each entity information based on the dimension evaluation indicator; Acquire historical data corresponding to each entity information based on the large model training knowledge graph; Acquire a change direction based on the historical data, and use the change direction as a first change trend; Determining whether the first entity evaluation index is greater than or equal to a preset value; If it is less than, it is determined that the entity information corresponding to the first entity evaluation indicator does not meet the conditions, and the entity information that does not meet the conditions is not used as a key feature entity; If it is greater than or equal to, it is determined that the entity information corresponding to the first entity evaluation indicator meets the conditions, and the entity information that meets the conditions is used as the key feature entity.

4. The model optimization method based on large and small model collaborative intelligence according to claim 1 is characterized in that: The step of performing small model collaborative training on each of the key feature entities to obtain a second entity evaluation index and a second change trend corresponding to each of the key feature entities includes: Acquire representative type data according to the key feature entity; Acquire and configure a small model according to the representative type data; Dividing the representative type data to obtain a plurality of batch training data; Training each batch of training data to obtain batch evaluation indicators; Acquire a second entity evaluation indicator corresponding to the key feature entity according to each of the batch evaluation indicators; Obtain prediction result information according to each of the batch training data; A second change trend is obtained based on the prediction result information.

5. The model optimization method based on large and small model collaborative intelligence according to claim 1 is characterized in that: The step of acquiring a performance evaluation value according to the second entity evaluation index and the second change trend comprises: Obtaining a second entity evaluation indicator; Obtaining a first evaluation coefficient according to the second entity evaluation indicator; Obtaining a change rate according to a second change trend; Acquire a second evaluation coefficient according to the change rate; A performance evaluation value is calculated based on the second entity evaluation index, the first evaluation coefficient, the change rate, and the second evaluation coefficient.

6. Model optimization system based on collaborative intelligence of large and small models, characterized by: include: A first acquisition module is used to acquire multi-source heterogeneous data and classify the multi-source heterogeneous data to obtain unstructured data and structured data; Performing data preprocessing on the unstructured data to obtain first data to be processed; Normalizing the structured data to obtain second data to be processed; Extracting key text and image information from the first data to be processed based on the large language model to obtain key text and image information; Extracting multi-dimensional key information from the second data to be processed based on the large language model to obtain multi-dimensional key information; Integrate the key text image information and the multi-dimensional key information to obtain key feature information; A second acquisition module is used to acquire entity information according to the key feature information, and perform large model training on all the entity information to obtain a large model training knowledge graph; A third acquisition module is used to obtain a first entity evaluation index and a first change trend corresponding to each entity information according to the large model training knowledge graph, and filter the entity information according to the first entity evaluation index to obtain a key feature entity; A fourth acquisition module is used to perform small model collaborative training on each of the key feature entities to obtain a second entity evaluation index and a second change trend corresponding to each of the key feature entities; A fifth acquisition module, configured to acquire a performance evaluation value according to the second entity evaluation indicator and the second change trend; A judgment module is used to judge whether the performance evaluation value is greater than or equal to a preset threshold; If it is greater than or equal to, it is determined that the multi-source heterogeneous data has been optimized; If it is less than, it is judged that the multi-source heterogeneous data has not been optimized, and the artificial graph optimization index corresponding to the second entity evaluation index is obtained, and the artificial graph optimization index, the second entity evaluation index and the second change trend are re-trained in the large model until the multi-source heterogeneous data are all optimized.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 5 are implemented.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 5 are implemented.

Images