A technology big data evaluation system and method based on ensemble learning

Abstract

The present invention discloses a technology big data evaluation system and method based on integrated learning. The system includes: a technology data source integration module, an information hierarchy analysis module, and a multi-modal hybrid expert evaluation module. The technology data source integration module realizes the chained production of multi-source technology index data, text data, and image data by constructing a technology big data production workflow, and realizes the multi-instance parallel execution of the technology big data production workflow by constructing a task model integrated scheduling platform; the information hierarchy analysis module forms technology index evaluation data sets of different modes and dimensions by constructing an information hierarchy structure system; the multi-modal hybrid expert evaluation module constructs an expert model for three modal data of indicator data, text data, and image data, and integrates the expert model evaluation results of different modalities and different dimensional data through a gated network model to realize the comprehensive evaluation of the technology index. The present invention improves the accuracy of technology evaluation.

Description

Technical Field

[0001] The present invention relates to the field of scientific and technological big data evaluation, and in particular to a scientific and technological big data evaluation system and method based on ensemble learning. Background Art

[0002] With the acceleration of scientific and technological progress and digital transformation, big data in science and technology has become an important resource for innovation and development. These data are not only huge in volume, but also come from a wide range of sources and in a variety of modalities, including structured indicator data, semi-structured text, and unstructured image data. The heterogeneity, high dimensionality, and nonlinear characteristics of big data in science and technology pose great challenges to traditional evaluation methods. Existing big data evaluation methods mainly include rule-based analysis methods, statistical models, and single machine learning models. However, these methods have many shortcomings in dealing with the complexity and diversity of big data in science and technology. Rule-based analysis methods rely on pre-set rules, have poor flexibility, and are difficult to adapt to implicit relationships in data. Although statistical models can provide judgments on overall trends, they are often greatly affected by outliers when the data volume is large and there is a lot of noise data, making it difficult to ensure the accuracy of the analysis. In addition, single machine learning models such as decision trees and support vector machines, although they can adapt to certain nonlinear patterns, are prone to overfitting or underfitting when the data volume is large, and have insufficient generalization capabilities, making it difficult to meet the complex needs in practical applications. At the same time, although there are some big data collection and analysis systems, these systems focus more on generality and lack professional solutions for the characteristics and needs of data in the field of science and technology. Faced with the complex and diverse, time-sensitive, large-scale and highly professional characteristics of data in the field of science and technology, existing systems are difficult to effectively integrate and analyze, thus affecting the comprehensive grasp and in-depth mining of scientific and technological innovation. In order to better utilize scientific and technological big data to promote innovation and development, it is necessary to design customized solutions specifically for the data characteristics and needs in the field of science and technology to achieve accurate and efficient data integration analysis and application. In summary, the relevant fields currently lack an effective evaluation solution and its system to integrate and analyze these massive multi-source scientific and technological big data and provide valuable insights. Summary of the invention

[0003] The purpose of the present invention is to propose a scientific and technological big data evaluation system and method based on ensemble learning.

[0004] The technical solution to achieve the purpose of the present invention is: a science and technology big data evaluation system based on integrated learning, including a science and technology data source integration module, an information hierarchy analysis module, and a multimodal hybrid expert evaluation module, wherein:

[0005] The science and technology data source integration module is used to collect and integrate various science and technology data sources, and realize the linked production of multi-source science and technology indicator data, text data and image data by building a science and technology big data production workflow. By building a task model integration scheduling platform, it realizes the parallel execution of multiple instances of the science and technology big data production workflow, and outputs structured indicators, text summaries and standardized image modality data;

[0006] The information hierarchy analysis module further subdivides the structured indicators, text summaries and standardized image modal data from five dimensions, namely, the science and technology development environment, science and technology development resources, science and technology innovation output, science and technology and industrial security, and science and technology comprehensive contribution, to construct an information hierarchy system and form a science and technology index evaluation data set of different modalities and dimensions;

[0007] The multimodal hybrid expert evaluation module constructs an expert model for three modal data: indicator data, text data, and image data. It evaluates the science and technology index from five dimensions: science and technology development environment, science and technology development resources, science and technology innovation output, science and technology and industrial security, and comprehensive science and technology contribution. It also integrates the evaluation results of data in different modalities and dimensions through the corresponding gating network to obtain a comprehensive science and technology index.

[0008] Furthermore, in the science and technology data source integration module, the science and technology big data production workflow includes: science and technology data extraction model, science and technology data conversion model, and science and technology data loading model, among which:

[0009] The scientific and technological data extraction model is composed of a data source connector, a data collector, a data verifier and a data encapsulator in series, which is used for the connection and structured encapsulation of original scientific and technological data. The data source connector establishes a multi-channel communication pipeline with various scientific and technological data sources such as web pages, tables, databases, and API interfaces through a multi-source scientific and technological data connection pool; the data collector collects original scientific and technological data through the communication pipeline, and calls the corresponding verifier to perform consistency checks on the data format and integrity according to the type of scientific and technological data source; the data verifier encapsulates the original scientific and technological data that has passed the consistency check into a message and sends it to the distributed message queue to achieve asynchronous decoupling from the data encapsulator; the data encapsulator receives the message from the distributed message queue, and further processes it into the form of DATA containing a globally unique identifier INDEX and original data and metadata, and sends the processing result in<INDEX,DATA> Structured storage;

[0010] The scientific and technological data conversion model consists of a modal identifier and a modal processor, which is used to convert the structured data output by the scientific and technological data extraction model into structured indicator data, text summary data and standardized image modal data according to the data modality. The modal processor includes three types: indicator processor, text processor and image processor. The modal identifier identifies the data modality and distributes it to the corresponding modal processor; the indicator processor converts the numerical data into structured<KEY, VALUE, TIMSTAMP> In this format, KEY is the indicator name, VALUE is the indicator value, and TIMSTAMP is the timestamp. The text processor uses a large language model to extract the text summary TEXT and organizes the results into<KEY, TEXT, TIMSTAMP> The image processor scales the original image into a standard image FIGURE of size 256×256, and organizes the result into<KEY, FIGURE, TIMSTAMP> form; ultimately, the multi-source scientific and technological data are further processed into structured indicators, text summaries, and standardized image modality data;

[0011] The scientific and technological data loading model consists of two parts: a data router and a heterogeneous data storage engine, which are used to persistently store structured indicator data, text summary data, and standardized image modality data. The data router will receive the received data according to the data modality recognition results.<KEY, VALUE, TIMSTAMP> The indicator data in the format is distributed to the indicator data storage pipeline.<KEY, TEXT, TIMSTAMP> The text summary data in the format of<KEY, FIGURE, TIMSTAMP> The standardized image data in the format is distributed to the image data storage pipeline; the heterogeneous data storage engine adopts different storage strategies according to the data modality. For the indicator data storage pipeline, the relational database PostgreSQL is selected as the storage system, and the indicator index is established according to KEY_TIMSTAMP for storage; for the text data storage pipeline, the NoSQL database Elasticsearch is selected as the storage system, and the text index is established according to KEY_TIMSTAMP, and TEXT is used as the text content for storage; for the image data storage pipeline, the distributed file system HDFS is selected as the storage system, and the image index is established according to KEY_TIMSTAMP, and FIGURE is used as the graphic content for storage.

[0012] Furthermore, in the technology data source integration module, the task model integration scheduling platform uses the CeleryExecutor mode of Airflow to deploy the Airflow scheduler, web server and worker node in the Kubernetes cluster to achieve the integrated scheduling management of technology big data production workflow. The specific method is as follows:

[0013] The Web server registers three types of task models: technology data extraction, technology data conversion, and technology data loading, and builds task model container images to support cross-platform deployment;

[0014] DAG is used to define the dependencies, data transfer methods, and concurrent execution strategies of the three types of task models. The Airflow scheduler checks the status and dependencies of the task models in the DAG, and uses the CeleryExecutor mode to determine the execution order and number of instances of the task models.

[0015] The working node calls the Kubernetes API to create a corresponding number of Pods in sequence according to the execution order and number of instances of the task model to realize the parallel execution of scientific and technological big data production workflow tasks. In this process, Kubernetes selects the running node for the Pod through the built-in scheduling algorithm, and the task model container image is instantiated and executed on the running node.

[0016] Furthermore, the information hierarchy analysis module constructs an information hierarchy system from five dimensions: science and technology development environment, science and technology development resources, science and technology innovation output, science and technology and industrial security, and science and technology comprehensive contribution, to further subdivide the structured indicators, text summaries, and standardized image modal data output by the science and technology data source integration module. The formal description is as follows:

[0017] definition Represents the complete set of scientific and technological data, including , , Represent the index data set, text summary data set, and standardized image data set respectively. The data set after hierarchical analysis is expressed as:

[0018] (1)

[0019] (2)

[0020] (3)

[0021] in They represent the science and technology development environment indicator dataset, science and technology development resource indicator dataset, science and technology innovation output indicator dataset, science and technology and industrial security indicator dataset, and science and technology comprehensive contribution indicator dataset respectively; They represent the text dataset of science and technology development environment, the text dataset of science and technology development resources, the text dataset of science and technology innovation output, the text dataset of science and technology and industrial security, and the text dataset of comprehensive science and technology contribution; They respectively represent the science and technology development environment image dataset, the science and technology development resource image dataset, the science and technology innovation output image dataset, the science and technology and industrial security image dataset, and the science and technology comprehensive contribution image dataset.

[0022] Furthermore, in the multimodal hybrid expert evaluation module, the expert model includes an indicator data expert model, a text data expert model and an image data expert model, wherein the indicator data expert model realizes the technology index evaluation of the indicator modality through multi-criteria decision-making from the perspective of quantitative analysis; the text data expert model realizes the technology index evaluation of the text modality through the prompt template function, the text encoder and the text evaluation network; the image data expert model realizes the technology index evaluation of the image modality through the image encoder and the image evaluation network;

[0023] Through a gating network model Fusion indicator data expert model , Text Data Expert Model , Image Data Expert Model Evaluation results in different modal data dimensions to achieve the science and technology index The comprehensive evaluation of is shown in formula (4):

[0024] (4)

[0025] in,

[0026] Indicator Data Expert Model Technology Index for Indicator Modes Specifically, the Science and Technology Development Environment Indicators dataset is used. , Science and Technology Development Resource Indicator Dataset , Science and Technology Innovation Output Index Dataset , Science and Technology and Industrial Security Index Dataset , Science and Technology Comprehensive Contribution Index Dataset The technology index is evaluated separately and the gated network model is used Gating in the middle indicator dimension The fusion is expressed as

[0027] (5)

[0028] in , , , , Represents the indicator data expert model The science and technology index is evaluated based on five indicator data sets: science and technology development environment, science and technology development resources, science and technology innovation output, science and technology and industrial security, and comprehensive science and technology contribution;

[0029] Expert Model for Text Data Technology index for text mode Specifically, the Science and Technology Development Environment Text Dataset is used , Science and Technology Development Resources Text Dataset , Science and Technology Innovation Output Text Dataset , Science and Technology and Industrial Security Text Dataset , Science and Technology Comprehensive Contribution Text Dataset The technology index is evaluated separately and the gated network model is used Chinese text modal gating The fusion is expressed as

[0030] (6)

[0031] in , , , , Represents the text data expert model The science and technology index is evaluated on five text data sets: science and technology development environment, science and technology development resources, science and technology innovation output, science and technology and industrial security, and comprehensive science and technology contribution;

[0032] Image Data Expert Model Technology Index for Imaging Modalities Specifically, we use the Science and Technology Development Environment Image Dataset , Science and Technology Development Resource Image Dataset , Technological innovation output image dataset , Science and Technology and Industry Security Image Dataset , Science and Technology Comprehensive Contribution Image Dataset The technology index is evaluated separately and the gated network model is used Image modality gating The fusion is expressed as

[0033] (7)

[0034] in , , , , Represents the image data expert model A science and technology index evaluated on five image datasets: science and technology development environment, science and technology development resources, science and technology innovation output, science and technology and industrial security, and comprehensive science and technology contribution.

[0035] Furthermore, the indicator data expert model Using a multi-criteria evaluation method, for indicators of different dimensions, the distance between the science and technology index decision vector between the evaluated object and the "most ideal" and "least ideal" objects is calculated. , to quantify the technology index of the corresponding dimension indicators;

[0036] (8)

[0037] (9)

[0038] (10)

[0039] in, and Respectively represent the objects being evaluated The distance between the technology index decision vector and the "most ideal" and "least ideal" objects, which are selected from all The best and worst objects among the evaluated objects are virtualized as the "boundary line" for evaluation comparison; and The kth element value of the technology indicator decision vector of the “most ideal” and “least ideal” objects, respectively. is the number of science and technology indicators, Indicates the corresponding weight coefficient of the kth scientific and technological indicator; Indicates the object being evaluated The kth element value of the technology indicator decision vector, and They respectively represent the operations of obtaining the maximum and minimum values ​​of scientific and technological indicators.

[0040] Furthermore, the text data expert model First, through the prompt template function Normalize the input text summary and then use the text encoder The normalized text is embedded to obtain a 128-dimensional text feature vector, and finally the text evaluation network is used to evaluate the text feature vector. Output the technology index of the corresponding dimension text, where:

[0041] Tip Template Function It consists of a series of template rules, which are used to normalize the original text paragraphs into the target text space. The template rules specifically include synonym replacement rules, vocabulary standardization rules, abbreviation expansion rules, date format unification rules, punctuation normalization rules, and text format unification rules;

[0042] Text Encoder A pre-trained bidirectional Transformer encoder is used to embed the normalized text into a high-order text feature vector, which is then reduced to a 128-dimensional text feature vector through a linear layer;

[0043] Text Evaluation Network A multi-layer perceptron network with two hidden layers is used to map the 128-dimensional text feature vector to the technology index of the text of the corresponding dimension.

[0044] Furthermore, the image data expert model First, we use the image encoder The normalized image is embedded to obtain a 128-dimensional image feature vector, which is then passed through the image evaluation network Output the technology index of the corresponding dimension image, where:

[0045] Image Encoder ResNet-18 is used as the backbone network, and the last classification layer is removed. At the same time, a global average pooling layer is added at the end of the backbone network to extract global features, a fully connected layer is used for feature mapping, and a batch normalization layer outputs a 128-dimensional image feature vector;

[0046] Image Evaluation Network A fully connected network with three hidden layers is used to map the 128-dimensional image feature vector to the technology index of the image of the corresponding dimension.

[0047] Furthermore, the gated network model A linear projection layer is used to merge the feature representations of indicators, texts, and images into a multimodal space. Then a 15-dimensional fully connected layer and a corresponding Softmax layer are used to output the weight coefficients of the indicator data expert model, text data expert model, and image data expert model in five dimensions, namely, scientific and technological development environment, scientific and technological development resources, scientific and technological innovation output, scientific and technological and industrial security, and comprehensive scientific and technological contribution. The comprehensive scientific and technological index is obtained by weighting scientific and technological indicators of different modalities and dimensions.

[0048] An evaluation method based on ensemble learning utilizes the science and technology big data evaluation system based on ensemble learning to achieve evaluation based on ensemble learning.

[0049] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the scientific and technological big data evaluation method based on integrated learning is implemented to achieve evaluation based on integrated learning.

[0050] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the scientific and technological big data evaluation method based on integrated learning to achieve evaluation based on integrated learning.

[0051] Compared with the existing technology, the present invention has the following significant advantages: 1) It proposes an information hierarchy system for science and technology big data evaluation and an evaluation method based on ensemble learning, which changes the traditional single evaluation model. 2) It constructs a theoretically complete and scientifically reasonable science and technology index evaluation model, that is, it realizes multi-dimensional science and technology index fusion evaluation through a multi-modal hybrid expert model combined with a gated network model, which improves the accuracy of science and technology evaluation. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 This is an architecture diagram of the scientific and technological big data evaluation based on integrated learning in the present invention.

[0053] Figure 2 This is the architecture diagram of the scientific and technological data source integration module.

[0054] Figure 3 It is the architecture diagram of the information hierarchy analysis module.

[0055] Figure 4 It is the architecture diagram of the multimodal hybrid expert evaluation model.

[0056] Figure 5 It is the indicator data expert model flow chart.

[0057] Figure 6 It is a flow chart for quantifying the weights of the science and technology index. DETAILED DESCRIPTION

[0058] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0059] The overall structure of the present invention is as follows Figure 1 As shown. Through the construction of data production workflow and model integration scheduling, the extraction, conversion and loading of science and technology big data are realized; by taking science and technology input-output as the evaluation and analysis idea, an information hierarchy system is constructed to provide science and technology index evaluation data sets of different modes and dimensions; by constructing a multimodal hybrid expert model based on ensemble learning (including indicator data expert model, text data expert model, image data expert model), the evaluation of science and technology big data of different modes and dimensions is realized, and by designing a gated network model to fuse the evaluation results of different modes and dimensions, a comprehensive evaluation of the science and technology index is realized.

[0060] The following is combined with Figure 2-Figure 6The composition and function of each module are described in detail.

[0061] A science and technology big data evaluation system based on integrated learning includes a science and technology data source integration module, an information hierarchy analysis module, and a multi-modal hybrid expert evaluation module, wherein:

[0062] The scientific and technological data source integration module is used to collect and integrate various scientific and technological data sources, and realize the linked production of multi-source scientific and technological indicator data, text data and image data by constructing a scientific and technological big data production workflow. By constructing a task model integrated scheduling platform, it realizes the parallel execution of multiple instances of the scientific and technological big data production workflow, and outputs structured indicators, text summaries and standardized image modality data. Figure 2 , the specific implementation steps are as follows:

[0063] (1) Construction of scientific and technological big data production workflow

[0064] Reference Figure 2 The scientific and technological big data production workflow includes a scientific and technological data extraction model, a scientific and technological data conversion model, and a scientific and technological data loading model, which realizes the linked production of multi-source scientific and technological indicator data, text data, and image data.

[0065] The technology data extraction model is used for the connection and structured packaging of original technology data. The technology data extraction model is composed of a data source connector, a data collector, a data verifier, and a data encapsulator in series. The specific implementation steps are as follows:

[0066] 1) The data source connector builds a multi-source scientific and technological data connection pool to establish multi-channel communication channels with various scientific and technological data sources such as web pages, tables, databases, and API interfaces;

[0067] 2) The data collector collects original scientific and technological data through the communication channel, and calls the corresponding verifier to perform data consistency check according to the type of scientific and technological data source, including format check and integrity check.

[0068] 3) The original scientific and technological data that has passed the consistency check is encapsulated into messages by the data verifier and sent to the distributed message queue, realizing asynchronous decoupling from the data encapsulator;

[0069] 4) The data encapsulator receives messages from the distributed message queue and further processes them into a DATA format containing a globally unique identifier INDEX and original scientific and technological data and metadata.<INDEX,DATA> Structured storage to support further processing of downstream task models.

[0070] The technology data conversion model is used to convert the output of the technology data extraction model<INDEX, DATA> The structured data is converted into structured index data, text summary data and standardized image modality data according to the data modality. The model is composed of a modality identifier and a modality processor, wherein the modality processor includes three types: an index processor, a text processor and an image processor. The specific implementation steps are as follows:

[0071] 1) Modal identifier based on<INDEX,DATA> The data modality of DATA in the input data is distributed to the corresponding modality processor, and the data modality includes three modalities: indicator, text, and image;

[0072] 2) The indicator processor converts numerical data into structured<KEY, VALUE> The text processor uses a large language model to extract text summaries TEXT and organizes the results into<KEY, TEXT> The image processor scales the original image into a standard image FIGURE of size 256×256 and organizes the result into<KEY, FIGURE> Finally, the multi-source scientific and technological data are further processed into structured indicators, text summaries, and standardized image modality data.

[0073] The technology data loading model is used to persistently store the structured indicator data, text summary data, and standardized image modality data output by the technology data conversion model, providing a data basis for technology evaluation. The technology data loading model consists of two parts: a data router and a heterogeneous data storage engine. The specific implementation steps are as follows:

[0074] 1) Data routing will receive the received<KEY, VALUE> The indicator data in the format is distributed to the indicator data storage pipeline.<KEY, TEXT> The text summary data in the format of<KEY, FIGURE> Distribute standardized image data in the format to the image data storage pipeline;

[0075] 2) Heterogeneous data storage engines use different storage strategies according to data modalities. For the indicator data storage pipeline, the relational database PostgreSQL is selected as the storage system, and indexes are established according to the KEY value to support high-frequency indicator queries and updates. For the text data storage pipeline, the NoSQL database Elasticsearch is selected as the storage system, and the KEY value is used as the document ID and the TEXT value is used as the document content to support full-text retrieval of large-scale text data. For the image data storage pipeline, the distributed file system HDFS is selected as the storage system, and the KEY value is used as the file name and the FIGURE value is used as the file content. At the same time, the data sharding strategy is combined to take into account the efficiency and security of image data storage.

[0076] (2) Integrated Scheduling of Technology Game Task Model

[0077] Reference Figure 2 This system builds a task model integrated scheduling platform based on Airflow to realize multi-instance execution of science and technology big data production workflows, and realizes the parallel production of multi-source science and technology indicator data, text data, and image data. The task model integrated scheduling platform uses Airflow's CeleryExecutor mode to deploy Airflow's scheduler, web server, and worker nodes in the Kubernetes cluster, and uses the Kubernetes container orchestration management engine to realize the scheduling and execution of science and technology big data production workflows. The specific implementation steps are as follows:

[0078] 1) The Web server registers three types of task models: scientific and technological data extraction, scientific and technological data conversion, and scientific and technological data loading, and builds task model container images to support cross-platform deployment;

[0079] 2) Use DAG (directed acyclic graph) to define the dependencies, data transfer methods, and concurrent execution strategies of the three types of task models. The Airflow scheduler checks the status and dependencies of the task models in the DAG, and uses the CeleryExecutor mode to determine the execution order and number of instances of the task models;

[0080] 3) Based on the execution order and number of instances of the task model, the working node calls the Kubernetes API to create a corresponding number of Pods (encapsulating the task model container image) in order to realize the parallel execution of the scientific and technological big data production workflow tasks. In this process, Kubernetes selects the appropriate running node for the Pod through the built-in scheduling algorithm, and the task model container image is instantiated and executed on the running node.

[0081] The information hierarchy analysis module uses the science and technology input-output as the evaluation and analysis idea, and further subdivides the structured indicators, text summaries and standardized image modal data output by the science and technology data source integration module from five dimensions, namely, the science and technology development environment, science and technology development resources, science and technology innovation output, science and technology and industrial security, and science and technology comprehensive contribution, based on the principles of scientificity, relative independence, total and per capita consideration, and combination of quantitative and qualitative. It constructs an information hierarchy system and provides science and technology index evaluation data sets of different modalities and dimensions. The specific implementation steps are as follows:

[0082] Reference Figure 3 ,The construction of the information hierarchical analysis module system of this system follows the following four basic principles:

[0083] First, scientificity. The selection of evaluation data should be consistent with the connotation and logical system of the science and technology index.

[0084] The second is relative independence. On the basis of meeting the dimensions of the information system, select data with weak correlation in statistical representation to effectively improve the independence of data dimensions.

[0085] The third is to take into account both total volume and per capita. Per capita can take into account the differences in scientific and technological volume, while scientific and technological breakthroughs in core key areas care more about the difference between "0" and "1".

[0086] Fourth, combine quantitative and qualitative methods. Use authoritative data to characterize the "hard power" of scientific and technological development, and qualitatively describe the "soft power" of scientific and technological innovation that cannot or cannot be fully reflected by quantitative methods.

[0087] Based on the above four basic principles, this system constructs an information hierarchy system from five dimensions: science and technology development environment, science and technology development resources, science and technology innovation output, science and technology and industrial security, and comprehensive science and technology contribution. It further subdivides the structured indicators, text summaries, and standardized image modality data output by the science and technology data source integration module. The formal description is as follows:

[0088] definition Represents the complete set of scientific and technological data, including , , Represent the original indicator dataset, text summary dataset, and standardized image dataset respectively, then the dataset after hierarchical analysis is expressed as

[0089] (1)

[0090] (2)

[0091] (3)

[0092] in They represent the science and technology development environment indicator dataset, science and technology development resource indicator dataset, science and technology innovation output indicator dataset, science and technology and industrial security indicator dataset, and science and technology comprehensive contribution indicator dataset respectively; They represent the text dataset of science and technology development environment, the text dataset of science and technology development resources, the text dataset of science and technology innovation output, the text dataset of science and technology and industrial security, and the text dataset of comprehensive science and technology contribution; They respectively represent the science and technology development environment image dataset, the science and technology development resource image dataset, the science and technology innovation output image dataset, the science and technology and industrial security image dataset, and the science and technology comprehensive contribution image dataset.

[0093] The multimodal hybrid expert module constructs expert models from three modal data: indicator data, text data, and image data. The technology index is evaluated from five dimensions within the modality: technology development environment, technology development resources, technology innovation output, technology and industrial security, and comprehensive technology contribution. The evaluation results of different modalities are fused through corresponding gating networks. Among them, the indicator data expert model realizes the technology index evaluation of the indicator modality through multi-criteria decision-making from the perspective of quantitative analysis. The text data expert model realizes the technology index evaluation of the text modality through the prompt template function, text encoder, and text evaluation network. The image data expert model realizes the technology index evaluation of the image modality through the image encoder and image evaluation network. Refer to Figure 1 , Figure 4-Figure 6 , the specific implementation steps are as follows:

[0094] Reference Figure 4 The multimodal hybrid expert evaluation model is used for comprehensive science and technology index evaluation and science and technology competitiveness situation analysis. The multimodal hybrid expert evaluation model is based on the gated network model. Fusion indicator data expert model , Text Data Expert Model , Image Data Expert Model Evaluation results in different modal data dimensions to achieve the science and technology index The comprehensive evaluation is as shown in formula (4).

[0095] (4)

[0096] Indicator Data Expert Model Technology Index for Indicator Modes Specifically, the science and technology index is evaluated from five dimensions: science and technology development environment, science and technology development resources, science and technology innovation output, science and technology and industrial security, and comprehensive science and technology contribution. Gating in the middle indicator dimension The fusion is expressed as

[0097] (5)

[0098] in , , , , Represents the indicator data expert model In the Science and Technology Development Environment Indicators Dataset , Science and Technology Development Resource Indicator Dataset , Science and Technology Innovation Output Index Dataset , Science and Technology and Industrial Security Index Dataset , Science and Technology Comprehensive Contribution Index Dataset A total of 5 indicator data sets are used to evaluate the technology index.

[0099] Specifically, refer to Figure 5 , the indicator data expert model A multi-criteria evaluation method is used to calculate the distance between the technology index decision vector of the evaluated object and the "most ideal" and "least ideal" objects from different dimensions of technology indicators. , in order to quantify the science and technology index of the corresponding dimension. The larger the number, the higher the technology index of the corresponding dimension, and vice versa.

[0100] (6)

[0101] (7)

[0102] (8)

[0103] in, and Respectively represent the objects being evaluated The distance between the technology index decision vector and the "most ideal" and "least ideal" objects, which are selected from all The best and worst objects among the evaluated objects are virtualized as the "boundary line" for evaluation comparison; and The kth element value of the technology indicator decision vector of the “most ideal” and “least ideal” objects, respectively. is the number of science and technology indicators, Indicates the corresponding weight coefficient of the kth scientific and technological indicator; Indicates the object being evaluated The kth element value of the technology indicator decision vector, and They respectively represent the operations of obtaining the maximum and minimum values ​​of scientific and technological indicators.

[0104] Reference Figure 4 , the text data expert model Technology index for text mode Specifically, the science and technology index is evaluated from five dimensions: science and technology development environment, science and technology development resources, science and technology innovation output, science and technology and industrial security, and comprehensive science and technology contribution. Chinese text modal gating The fusion can be expressed as

[0105] (9)

[0106] in , , , , Represents the text data expert model In the context of technological development text dataset , Science and Technology Development Resources Text Dataset , Science and Technology Innovation Output Text Dataset , Science and Technology and Industrial Security Text Dataset , Science and Technology Comprehensive Contribution Text Dataset The technology index is evaluated on a total of 5 text datasets.

[0107] Specifically, the text data expert model First, through the prompt template function Normalize the input text summary and then use the text encoder The normalized text is embedded to obtain a 128-dimensional text feature vector, and finally the text evaluation network is used to evaluate the text feature vector. Output the technology index of the corresponding dimension.

[0108] Furthermore, the prompt template function It consists of a series of template rules, which are used to normalize the original text paragraphs into the target text space. The template rules specifically include synonym replacement rules, vocabulary standardization rules, abbreviation expansion rules, date format unification rules, punctuation normalization rules, and text format unification rules.

[0109] Furthermore, the text encoder A pre-trained bidirectional Transformer encoder is used to embed the normalized text into a high-order text feature vector, which is then reduced to a 128-dimensional text feature vector through a linear layer.

[0110] Furthermore, the text evaluation network A multi-layer perceptron (MLP) network with two hidden layers is used to calculate the technology index of the corresponding dimension from the 128-dimensional text feature vector.

[0111] Reference Figure 4 , the image data expert model Technology Index for Imaging Modalities Specifically, the science and technology index is evaluated from five dimensions: science and technology development environment, science and technology development resources, science and technology innovation output, science and technology and industrial security, and comprehensive science and technology contribution. Image modality gating The fusion can be expressed as

[0112] (10)

[0113] in , , , , Represents the image data expert model In the technology development environment image dataset , Science and Technology Development Resource Image Dataset , Technological innovation output image dataset , Science and Technology and Industry Security Image Dataset , Science and Technology Comprehensive Contribution Image Dataset The technology index is evaluated on a total of 5 text datasets.

[0114] Specifically, the image data expert model First, we use the image encoder The normalized image is embedded to obtain a 128-dimensional image feature vector, which is then passed through the image evaluation network Output the technology index of the corresponding dimension.

[0115] Furthermore, the image encoder A lightweight ResNet-18 is used as the backbone network, and the last classification layer is removed. At the same time, a global average pooling layer is added at the end of the backbone network to extract global features, a fully connected layer is used for feature mapping, and a batch normalization layer is used to output a 128-dimensional image feature vector.

[0116] Furthermore, the image evaluation network A fully connected network with three hidden layers is used to calculate the technology index of the corresponding dimension from the 128-dimensional image feature vector.

[0117] In particular, the image evaluation network A Dropout (0.2) regularization layer was added after each hidden layer to randomly drop neurons with a probability of 20% during training to enhance the robustness and generalization ability of the model.

[0118] Reference Figure 4 , the gating network model It is used for output control of expert models of different modes, and adaptively learns weights of different dimensions of different modes to control the role of different expert models in the comprehensive evaluation of science and technology indexes.

[0119] Specifically, the gating network model A linear projection layer is used to merge the feature representations of indicators, texts, and images into a multimodal space. Then a 15-dimensional fully connected layer and a corresponding Softmax layer are used to output the weight coefficients of the indicator data expert model, text data expert model, and image data expert model in five dimensions, namely, scientific and technological development environment, scientific and technological development resources, scientific and technological innovation output, scientific and technological and industrial security, and comprehensive scientific and technological contribution. The comprehensive scientific and technological index is obtained by weighting scientific and technological indicators of different modalities and dimensions.

[0120] In particular, refer to Figure 6 , the gating network model The output weights can also support domain experts to make adjustments based on their respective domain experience. Domain experts determine the relative importance of the indexes by constructing a pairwise comparison matrix of the 15 technology indexes. Assume that the relative importance of comparing the i-th technology index with the j-th technology index is quantified as , then the pairwise comparison matrix can be expressed as This system is based on The pairwise comparison matrix given by experts is used to calculate the weight coefficient of the science and technology index using the geometric analysis method, as shown in formula (11).

[0121] (11)

[0122] (12)

[0123] in, Technology Index The weight coefficient of =15 is the technology index dimension; represents the technology index determined by the expert p's score The weight coefficient of .

[0124] The comprehensive science and technology index not only reflects the overall science and technology index level of the evaluation object, but also determines the relative science and technology ranking of the evaluation object by comprehensively ranking different evaluation objects according to the size of their science and technology index.

[0125] Furthermore, by ranking the five dimensions of science and technology development environment, science and technology development resources, science and technology innovation output, science and technology industry security, and comprehensive science and technology contribution, a panoramic overview of scientific and technological strength is presented, effectively revealing the advantages and shortcomings of the evaluation objects in scientific and technological competition.

[0126] Furthermore, through the inflection point analysis of the technology index curve, you can hover the mouse to view the growth rate of the technology index before and after the inflection point; through horizontal / vertical comparative analysis of the technology index, it can provide valuable insights.

[0127] The present invention also proposes an evaluation method based on ensemble learning, and utilizes the science and technology big data evaluation system based on ensemble learning to implement evaluation based on ensemble learning.

[0128] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the science and technology big data evaluation system based on integrated learning is implemented to realize the evaluation based on integrated learning.

[0129] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the science and technology big data evaluation system based on integrated learning and realizes evaluation based on integrated learning.

[0130] The technical features of the above embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0131] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A scientific and technological big data evaluation system based on ensemble learning, characterized in that: It includes the science and technology data source integration module, information hierarchy analysis module, and multi-modal hybrid expert evaluation module, among which: The science and technology data source integration module is used to collect and integrate various science and technology data sources, and realize the linked production of multi-source science and technology indicator data, text data and image data by building a science and technology big data production workflow. By building a task model integration scheduling platform, it realizes the parallel execution of multiple instances of the science and technology big data production workflow, and outputs structured indicators, text summaries and standardized image modality data; The information hierarchy analysis module further subdivides the structured indicators, text summaries and standardized image modal data from five dimensions, namely, the science and technology development environment, science and technology development resources, science and technology innovation output, science and technology and industrial security, and science and technology comprehensive contribution, to construct an information hierarchy system and form a science and technology index evaluation data set of different modalities and dimensions; The multimodal hybrid expert evaluation module builds an expert model for three modal data: indicator data, text data, and image data. It evaluates the science and technology index from five dimensions: science and technology development environment, science and technology development resources, science and technology innovation output, science and technology and industrial security, and comprehensive science and technology contribution. It also integrates the evaluation results of data in different modalities and dimensions through the corresponding gated network to obtain a comprehensive science and technology index. In the multimodal hybrid expert evaluation module, the expert models include indicator data expert model, text data expert model and image data expert model. The indicator data expert model realizes the science and technology index evaluation of the indicator modality through multi-criteria decision-making from the perspective of quantitative analysis; the text data expert model realizes the science and technology index evaluation of the text modality through prompt template function, text encoder and text evaluation network; the image data expert model realizes the science and technology index evaluation of the image modality through image encoder and image evaluation network.

2. The technology big data evaluation system based on ensemble learning according to claim 1 is characterized in that: In the science and technology data source integration module, the science and technology big data production workflow includes: science and technology data extraction model, science and technology data conversion model, and science and technology data loading model, among which: The scientific and technological data extraction model is composed of a data source connector, a data collector, a data verifier and a data encapsulator in series, which is used for the connection and structured encapsulation of original scientific and technological data. The data source connector establishes a multi-channel communication pipeline with various scientific and technological data sources such as web pages, tables, databases, and API interfaces through a multi-source scientific and technological data connection pool; the data collector collects original scientific and technological data through the communication pipeline, and calls the corresponding verifier to perform consistency checks on the data format and integrity according to the type of scientific and technological data source; the data verifier encapsulates the original scientific and technological data that has passed the consistency check into a message and sends it to the distributed message queue to achieve asynchronous decoupling from the data encapsulator; the data encapsulator receives the message from the distributed message queue, and further processes it into the form of DATA containing a globally unique identifier INDEX and original data and metadata, and sends the processing result in<INDEX,DATA> Structured storage; The scientific and technological data conversion model consists of a modal identifier and a modal processor, which is used to convert the structured data output by the scientific and technological data extraction model into structured indicator data, text summary data and standardized image modal data according to the data modality. The modal processor includes three types: indicator processor, text processor and image processor. The modal identifier identifies the data modality and distributes it to the corresponding modal processor; the indicator processor converts the numerical data into structured<KEY,VALUE, TIMSTAMP> In this format, KEY is the indicator name, VALUE is the indicator value, and TIMSTAMP is the timestamp. The text processor uses a large language model to extract the text summary TEXT and organizes the results into<KEY, TEXT, TIMSTAMP> The image processor scales the original image into a standard image FIGURE of size 256×256, and organizes the result into<KEY,FIGURE, TIMSTAMP> form; ultimately, the multi-source scientific and technological data are further processed into structured indicators, text summaries, and standardized image modality data; The scientific and technological data loading model consists of two parts: a data router and a heterogeneous data storage engine, which are used to persistently store structured indicator data, text summary data, and standardized image modality data. The data router will receive the received data according to the data modality recognition results.<KEY, VALUE, TIMSTAMP> The indicator data in the format is distributed to the indicator data storage pipeline.<KEY, TEXT, TIMSTAMP> The text summary data in the format of<KEY,FIGURE, TIMSTAMP> The standardized image data in the format is distributed to the image data storage pipeline; the heterogeneous data storage engine adopts different storage strategies according to the data modality. For the indicator data storage pipeline, the relational database PostgreSQL is selected as the storage system, and the indicator index is established according to KEY_TIMSTAMP for storage; for the text data storage pipeline, the NoSQL database Elasticsearch is selected as the storage system, and the text index is established according to KEY_TIMSTAMP, and TEXT is used as the text content for storage; for the image data storage pipeline, the distributed file system HDFS is selected as the storage system, and the image index is established according to KEY_TIMSTAMP, and FIGURE is used as the graphic content for storage.

3. The technology big data evaluation system based on ensemble learning according to claim 1 is characterized in that: In the science and technology data source integration module, the task model integration scheduling platform uses the CeleryExecutor mode of Airflow to deploy the Airflow scheduler, web server and worker node in the Kubernetes cluster to achieve the integrated scheduling management of science and technology big data production workflow. The specific method is as follows: The Web server registers three types of task models: technology data extraction, technology data conversion, and technology data loading, and builds task model container images to support cross-platform deployment; DAG is used to define the dependencies, data transmission methods, and concurrent execution strategies of three types of task models. The Airflow scheduler checks the status and dependencies of the task models in the DAG, and uses the CeleryExecutor mode to determine the execution order and number of instances of the task models. The working node calls the Kubernetes API to create a corresponding number of Pods in sequence according to the execution order and number of instances of the task model to realize the parallel execution of scientific and technological big data production workflow tasks. In this process, Kubernetes selects the running node for the Pod through the built-in scheduling algorithm, and the task model container image is instantiated and executed on the running node.

4. The technology big data evaluation system based on ensemble learning according to claim 1 is characterized in that: The information hierarchy analysis module constructs an information hierarchy system from five dimensions: science and technology development environment, science and technology development resources, science and technology innovation output, science and technology and industrial security, and science and technology comprehensive contribution. It further subdivides the structured indicators, text summaries, and standardized image modal data output by the science and technology data source integration module. The formal description is as follows: definition Represents the complete set of scientific and technological data, including , , Represent the index data set, text summary data set, and standardized image data set respectively. The data set after hierarchical analysis is expressed as: （1）； （2）； （3）； in They represent the science and technology development environment indicator dataset, science and technology development resource indicator dataset, science and technology innovation output indicator dataset, science and technology and industrial security indicator dataset, and science and technology comprehensive contribution indicator dataset respectively; They represent the text dataset of science and technology development environment, the text dataset of science and technology development resources, the text dataset of science and technology innovation output, the text dataset of science and technology and industrial security, and the text dataset of comprehensive science and technology contribution; They respectively represent the science and technology development environment image dataset, the science and technology development resource image dataset, the science and technology innovation output image dataset, the science and technology and industrial security image dataset, and the science and technology comprehensive contribution image dataset.

5. The technology big data evaluation system based on ensemble learning according to claim 1 is characterized in that: The multimodal hybrid expert evaluation module is implemented through a gating network model Fusion indicator data expert model , Text Data Expert Model , Image Data Expert Model Evaluation results in different modal data dimensions to achieve the science and technology index The comprehensive evaluation of is shown in formula (4): （4）； in, Indicator Data Expert Model Technology Index for Indicator Modes Specifically, the Science and Technology Development Environment Indicators dataset is used. , Science and Technology Development Resource Indicator Dataset , Science and Technology Innovation Output Index Dataset , Science and Technology and Industrial Security Index Dataset , Science and Technology Comprehensive Contribution Index Dataset The technology index is evaluated separately and the gated network model is used Gating in the middle indicator dimension The fusion is expressed as （5）； in , , , , Represents the indicator data expert model The science and technology index is evaluated based on five indicator data sets: science and technology development environment, science and technology development resources, science and technology innovation output, science and technology and industrial security, and comprehensive science and technology contribution; Text Data Expert Model Technology index for text mode Specifically, the Science and Technology Development Environment Text Dataset is used , Science and Technology Development Resources Text Dataset , Science and Technology Innovation Output Text Dataset , Science and Technology and Industrial Security Text Dataset , Science and Technology Comprehensive Contribution Text Dataset The technology index is evaluated separately and the gated network model is used Chinese text modal gating The fusion is expressed as （6）； in , , , , Represents the text data expert model The science and technology index is evaluated on five text data sets: science and technology development environment, science and technology development resources, science and technology innovation output, science and technology and industrial security, and comprehensive science and technology contribution; Image Data Expert Model Technology Index for Imaging Modalities Specifically, we use the Science and Technology Development Environment Image Dataset , Science and Technology Development Resource Image Dataset , Technological innovation output image dataset , Science and Technology and Industry Security Image Dataset , Science and Technology Comprehensive Contribution Image Dataset The technology index is evaluated separately and the gated network model is used Image modality gating The fusion is expressed as （7）； in , , , , Represents the image data expert model A science and technology index evaluated on five image datasets: science and technology development environment, science and technology development resources, science and technology innovation output, science and technology and industrial security, and comprehensive science and technology contribution.

6. The technology big data evaluation system based on ensemble learning according to claim 5 is characterized in that: Indicator Data Expert Model Using a multi-criteria evaluation method, for indicators of different dimensions, the distance between the science and technology index decision vector between the evaluated object and the "most ideal" and "least ideal" objects is calculated. , to quantify the technology index of the corresponding dimension indicators; （8）； （9）； （10）； in, and Respectively represent the objects being evaluated The distance between the technology index decision vector and the "most ideal" and "least ideal" objects, which are selected from all The best and worst objects among the evaluated objects are virtualized as the "boundary line" for evaluation comparison; and The kth element value of the technology indicator decision vector of the "most ideal" and "least ideal" objects, respectively. is the number of science and technology indicators, Indicates the corresponding weight coefficient of the kth scientific and technological indicator; Indicates the object being evaluated The kth element value of the technology indicator decision vector, and They respectively represent the operations of obtaining the maximum and minimum values ​​of scientific and technological indicators.

7. The technology big data evaluation system based on ensemble learning according to claim 5 is characterized in that: Text Data Expert Model First, through the prompt template function Normalize the input text summary and then use the text encoder The normalized text is embedded to obtain a 128-dimensional text feature vector, and finally the text evaluation network is used to evaluate the text feature vector. Output the technology index of the corresponding dimension text, where: Tip Template Function It consists of a series of template rules. The template rules are used to normalize the original text paragraphs into the target text space. The template rules specifically include synonym replacement rules, vocabulary standardization rules, abbreviation expansion rules, date format unification rules, punctuation normalization rules, and text format unification rules. Text Encoder A pre-trained bidirectional Transformer encoder is used to embed the normalized text into a high-order text feature vector, which is then reduced to a 128-dimensional text feature vector through a linear layer; Text Evaluation Network A multi-layer perceptron network with two hidden layers is used to map the 128-dimensional text feature vector into the technology index of the text of the corresponding dimension.

8. The technology big data evaluation system based on ensemble learning according to claim 5 is characterized in that: Image Data Expert Model First, we use the image encoder The normalized image is embedded to obtain a 128-dimensional image feature vector, which is then passed through the image evaluation network Output the technology index of the corresponding dimension image, where: Image Encoder ResNet-18 is used as the backbone network, and the last classification layer is removed. At the same time, a global average pooling layer is added at the end of the backbone network to extract global features, a fully connected layer is used for feature mapping, and a batch normalization layer outputs a 128-dimensional image feature vector; Image Evaluation Network A fully connected network with three hidden layers is used to map the 128-dimensional image feature vector to the technology index of the image of the corresponding dimension.

9. The technology big data evaluation system based on ensemble learning according to claim 5 is characterized in that: Gated Network Model A linear projection layer is used to merge the feature representations of indicators, texts, and images into a multimodal space. Then a 15-dimensional fully connected layer and a corresponding Softmax layer are used to output the weight coefficients of the indicator data expert model, text data expert model, and image data expert model in five dimensions, namely, scientific and technological development environment, scientific and technological development resources, scientific and technological innovation output, scientific and technological and industrial security, and comprehensive scientific and technological contribution, for a total of 15 scientific and technological indexes. The comprehensive scientific and technological index is obtained by weighting scientific and technological indicators of different modalities and dimensions.

10. An evaluation method based on ensemble learning, characterized in that: The scientific and technological big data evaluation system based on integrated learning is utilized as described in any one of claims 1 to 9 to realize evaluation based on integrated learning.

Images