Large model architecture in iron and steel industry

Abstract

The invention provides a large model architecture in the iron and steel industry. The large model architecture comprises a data access layer, a data fusion layer, a cognitive intelligent layer, a large model layer and an application display layer. The data access layer is used for receiving steel production line data and outputting the data to the data fusion layer; the data fusion layer is used for constructing a knowledge base based on the steel production line data and the steel knowledge data; the cognitive intelligent layer is used for analyzing and processing the steel production line data to obtain initial data; the large model layer is used for obtaining a steel industry large model based on the initial data and the knowledge base; the application display layer is used for obtaining target data based on the steel industry large model. Through the five-layer large model architecture of the data access layer, the data fusion layer, the cognitive intelligent layer, the large model layer and the application display layer, convenient utilization of the large model is realized, the actual problems in the iron and steel industry are effectively solved, and the customer experience is improved.

Description

Technical Field

[0001] The present disclosure relates to the technical field of the steel industry, and particularly to a large model architecture in the steel industry. Background Art

[0002] Currently, large models are widely used in the rail transit industry and the medical industry, generally only involving natural language processing scenarios. However, in the steel industry, there are multiple application scenarios, and currently, there is no effective large model application architecture, resulting in problems such as poor large model application effects and poor customer experience. Summary of the Invention

[0003] The technical problem to be solved by the present disclosure is to overcome the defects in the prior art that there is no effective large model application architecture in the steel industry, resulting in poor large model application effects and poor customer experience, and to provide a large model architecture in the steel industry.

[0004] The present disclosure solves the above technical problem through the following technical solutions:

[0005] The present disclosure provides a large model architecture in the steel industry, and the large model architecture includes a data access layer, a data fusion layer, a cognitive intelligence layer, a large model layer, and an application display layer;

[0006] The data access layer is used to receive steel production line data and output it to the data fusion layer;

[0007] The data fusion layer is used to construct a knowledge base based on the steel production line data and steel knowledge data;

[0008] The cognitive intelligence layer is used to analyze and process the steel production line data to obtain initial data;

[0009] The large model layer is used to obtain a large model for the steel industry based on the initial data and the knowledge base;

[0010] The application display layer is used to obtain target data based on the large model for the steel industry.

[0011] Optionally, the data access layer is further used to receive the steel production line data from at least one of a steel production line, a inspection and testing system, a process control level system, and a production control level system.

[0012] Optionally, the data fusion layer is further used to adopt knowledge graph technology to fuse and process the steel production line data and the steel knowledge data to construct the knowledge base.

[0013] Optionally, the steel knowledge data includes at least one of coal blending knowledge data, continuous casting quality data, surface defect data, energy management data, visual data, and scheduling data.

[0014] Optionally, the cognitive intelligence layer includes a knowledge reasoning unit;

[0015] The knowledge reasoning unit is used to reason about the steel production line data to obtain the initial data;

[0016] And / or,

[0017] The cognitive intelligence layer includes a few-shot generation unit;

[0018] The few-shot generation unit is used to generate the initial data based on the steel production line data by using few-shot learning techniques;

[0019] And / or,

[0020] The cognitive intelligence layer includes a multimodal understanding unit;

[0021] The multimodal understanding unit is used to understand the steel production line data in multiple modalities to obtain the initial data;

[0022] And / or,

[0023] The cognitive intelligence layer includes a cognitive decision-making unit;

[0024] The cognitive decision-making unit is used to perform decision-making processing on the steel production line data to obtain the initial data;

[0025] And / or,

[0026] The cognitive intelligence layer includes an intelligent planning unit;

[0027] The intelligent planning unit is used to perform planning processing on the steel production line data to obtain the initial data.

[0028] Optionally, the large model layer is further used to fuse a general model and a special model based on the initial data and the knowledge base to obtain the steel industry large model.

[0029] Optionally, the general model includes at least one of DeepSeek, Qwen, Wenxin Yiyan, and LIAMA3 (a large language model developed by Meta);

[0030] And / or,

[0031] The special model includes at least one of a process optimization large model, a time series prediction large model, a visual detection large model, and a natural language processing large model;

[0032] Among them, the process optimization large model is used to optimize the coal blending process, the time series prediction large model is used to predict the quality of continuous casting ladles, the visual inspection large model is used to detect surface defects of products, and the natural language processing large model is used to parse natural language.

[0033] Optionally, the large model for the iron and steel industry adopts at least one hybrid neural network architecture among the Transformer architecture (a deep learning model architecture provided by Google), convolutional neural network, and recurrent neural network.

[0034] Optionally, the application display layer is further configured to input actual data in an actual application scenario into the large model for the iron and steel industry to output the target data.

[0035] Optionally, the actual application scenario includes at least one of an intelligent coal blending scenario, a production quality control scenario, a surface defect detection scenario, an energy management scenario, a process unmanned scenario, and an intelligent shift scheduling scenario.

[0036] Based on common general knowledge in the art, the above preferred conditions can be combined arbitrarily to obtain various preferred examples of the present disclosure.

[0037] The positive and progressive effects of the present disclosure are as follows:

[0038] Through the five-layer large model architecture of the data access layer, data fusion layer, cognitive intelligence layer, large model layer, and application display layer, the present disclosure realizes the convenient utilization of the large model, effectively solves practical problems in the iron and steel industry, and improves the customer experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 It is a module schematic diagram of the large model architecture in the iron and steel industry according to Embodiment 1 of the present disclosure;

[0040] Figure 2 It is a module schematic diagram of the large model architecture in the iron and steel industry according to Embodiment 2 of the present disclosure;

[0041] Figure 3 It is a specific example diagram of the large model architecture in the iron and steel industry according to Embodiment 2 of the present disclosure. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0042] The present disclosure will be further described below by way of examples, but the present disclosure is not limited thereto.

[0043] In the embodiments of the present disclosure, prefix words such as "first" and "second" are only used to distinguish different described objects, and have no restrictive effect on the position, order, priority, quantity, content, etc. of the described objects. In the embodiments of the present disclosure, the use of prefix words such as ordinal numbers for distinguishing described objects does not constitute a limitation on the described objects. For the statements of the described objects, refer to the descriptions in the context of the embodiments, and no redundant limitations should be formed due to the use of such prefix words. In addition, in the description of this embodiment, unless otherwise specified, the meaning of "a plurality" is two or more.

[0044] In the embodiments of the present disclosure, the processing of the collection, storage, use, processing, transmission, provision, and disclosure of the user's personal information involved all complies with the provisions of relevant laws and regulations and does not violate public order and good customs.

[0045] Embodiment 1

[0046] This embodiment provides a large model architecture in the steel industry, as Figure 1 shown, the large model architecture includes a data access layer 1, a data fusion layer 2, a cognitive intelligence layer 3, a large model layer 4, and an application display layer 5;

[0047] The data access layer 1 is used to receive steel production line data and output it to the data fusion layer;

[0048] The data fusion layer 2 is used to construct a knowledge base based on the steel production line data and steel knowledge data;

[0049] The cognitive intelligence layer 3 is used to analyze and process the steel production line data to obtain initial data;

[0050] The large model layer 4 is used to obtain a large model for the steel industry based on the initial data and the knowledge base;

[0051] The application display layer 5 is used to obtain target data based on the large model for the steel industry.

[0052] In this embodiment, through the five-layer large model architecture of the data access layer, data fusion layer, cognitive intelligence layer, large model layer, and application display layer, the convenient utilization of the large model is realized, effectively solving the actual problems in the steel industry and improving the customer experience.

[0053] Embodiment 2

[0054] This embodiment provides a large model architecture in the steel industry, which is a further improvement on Embodiment 1.

[0055] In an implementable solution, the data access layer 1 is further used to receive steel production line data from at least one of a steel production line, a inspection and testing system, a process control level system, and a production control level system.

[0056] Specifically, the data access layer is used to access basic automation data from the steel production line and related systems, achieve data docking of the production line, and connect to the existing production line. In addition, it is also necessary to dock with the LIMS (Laboratory Information Management System) inspection and testing system, L2 process control level system, and L3 production control level system during the production line process, and access the data one by one through various communication protocols.

[0057] In this solution, by receiving steel production line data from at least one of the steel production line, inspection and testing system, process control level system, and production control level system, the types of steel production line data are enriched, and the universality of the large model architecture is improved.

[0058] In an implementable solution, the data fusion layer 2 is also used to fuse and process steel production line data and steel knowledge data by using knowledge graph technology to construct a knowledge base.

[0059] Specifically, the data fusion layer is the knowledge and technology fusion layer. The knowledge and technology fusion layer is one of the key components in the large model architecture, aiming to deeply integrate the professional knowledge of the steel industry with actual production data to form a unified knowledge base and provide strong knowledge support for upper-layer applications. Professional knowledge is steel knowledge data, and actual production data is steel production line data. This layer covers knowledge and data in multiple fields, including intelligent coal blending, continuous casting quality production, surface defect detection, energy management, production line process, and shift scheduling.

[0060] First of all, the knowledge of intelligent coal blending involves the characteristics of various coals, such as ash content, sulfur content, etc., as well as rich coal blending experience knowledge. Combining this knowledge with real-time data can optimize the coal blending plan and improve production efficiency. The knowledge of the continuous casting quality production line includes process parameters, quality control standards, and expert experience. The fusion of these data helps to achieve accurate quality prediction and control. The knowledge of surface defect detection covers the characteristics of various defects and detection algorithms. Combining with image data can effectively identify and classify surface defects. The knowledge of energy management involves energy consumption models and optimization strategies. Combining with real-time energy consumption data can achieve efficient energy utilization. The knowledge of the production line process includes process flow and equipment operation rules, providing a basis for optimizing the production process. The knowledge of shift scheduling involves scheduling rules and resource allocation. Combining with the production plan can achieve reasonable shift scheduling.

[0061] In terms of the fusion method, the knowledge and technology fusion layer uses knowledge graph technology to represent the knowledge in different fields as a graph structure and establish the association relationships between them. For example, the relationship between the characteristics of coal and the coal blending plan, the relationship between process parameters and quality indicators, etc. In addition, machine learning methods are used to combine data and knowledge, train models, and achieve intelligent applications.

[0062] The fused knowledge base provides strong support for upper-layer applications. The intelligent coal blending application can obtain the characteristic data of coal and the knowledge of coal blending experience from the knowledge base, and combine real-time data for optimized coal blending; the continuous casting quality prediction application uses the process parameters and quality control standards in the knowledge base, and combines real-time data for quality prediction; the surface defect detection application uses the defect feature data in the knowledge base, and combines image data for defect detection; the energy management application uses the energy consumption model and optimization strategy in the knowledge base, and combines real-time energy consumption data for energy optimization; the shift scheduling and production scheduling application uses the scheduling rules and resource allocation knowledge in the knowledge base, and combines the production plan for shift scheduling and production scheduling.

[0063] To ensure the accuracy and timeliness of the knowledge base, a data update and maintenance mechanism has also been established. Through regular data updates and feedback mechanisms, the knowledge base is continuously optimized to adapt to changes in the production environment.

[0064] In this solution, by using knowledge graph technology to fuse steel production line data and steel knowledge data, a knowledge base is constructed to ensure the accuracy and reliability of the knowledge base.

[0065] In an implementable solution, the steel knowledge data includes at least one of coal blending knowledge data, continuous casting quality data, surface defect data, energy management data, visual data, and shift scheduling and production scheduling data.

[0066] In this solution, through various types of steel knowledge data, the diversity of data in the knowledge base is ensured.

[0067] In an implementable solution, as Figure 2 shown, the cognitive intelligence layer 3 includes a knowledge reasoning unit 31;

[0068] The knowledge reasoning unit 31 is used to reason about the steel production line data to obtain initial data.

[0069] Specifically, the cognitive intelligence layer is the cognitive intelligence engine layer. The knowledge reasoning unit covers a distributionally robust optimization algorithm framework, which is developed based on the theory of nonlinear expectation. The difference between this framework and the classical machine learning framework is that it does not assume that the samples satisfy the independent and identically distributed condition. In many practical application scenarios, especially in scenarios involving cognitive reasoning, the independent and identically distributed assumption usually does not hold. Therefore, based on this, a set of loss functions in the non-independent and identically distributed case is reconstructed. The reconstructed loss function solves the problem of non-smoothness with respect to parameters. For this class of loss functions, a complete set of computable algorithm theories is developed. The original optimization objective is improved using robust optimization, and a more stable solution can be given based on the intelligent coal blending scenario, providing support for the cognitive engine.

[0070] In this solution, the knowledge reasoning unit infers the steel production line data to obtain the initial data, ensuring the accuracy and reliability of the initial data.

[0071] In an implementable solution, the cognitive intelligence layer 3 includes a small sample generation unit 32;

[0072] The small sample generation unit 32 is used to generate initial data based on the steel production line data by using small sample learning technology.

[0073] Specifically, the small sample learning technology is implemented based on the distributionally robust optimization framework and the method of knowledge and data fusion. The distributionally robust optimization framework can greatly improve the utilization rate of samples. At the same time, knowledge in fields such as natural science and cognitive science is integrated to depict complex phenomena and behaviors. Through the application of AIGC (Artificial Intelligence Generated Content), corresponding scenario generation is provided for surface defects and intelligent coal blending.

[0074] In this solution, the utilization rate of the steel production line data is improved by obtaining the initial data by using small sample learning technology.

[0075] In an implementable solution, the cognitive intelligence layer 3 includes a multimodal understanding unit 33;

[0076] The multimodal understanding unit 33 is used to understand the steel production line data in multiple modalities to obtain the initial data.

[0077] Specifically, aiming at the multi - morphological characteristics of data and the characteristics of human cognitive cross - domain learning, for various multimodal learning technologies, including multimodal fusion, multimodal alignment, and multimodal pre - training, etc., a cognitive computing framework is constructed based on this. For multimodal inputs (text, speech, video, etc.), a complete set of recognition and understanding methods are developed by using multimodal feature extraction and fusion technologies, which can effectively complete tasks such as object detection, object recognition, and semantic segmentation. In the field of generative artificial intelligence, the base model is trained by using multimodal alignment pre - training technology to improve the fusion adaptability of different modal information, and the effect of obtaining high - quality alignment generation through fine - tuning with a small number of samples is achieved. Provide relevant engine support for surface defect detection and process unmanned scenarios.

[0078] In this solution, the utilization rate of the steel production line data is improved by understanding the steel production line data in multiple modalities to obtain the initial data.

[0079] In an implementable solution, the cognitive intelligence layer 3 includes a cognitive decision - making unit 34;

[0080] The cognitive decision - making unit 34 is used to perform decision - making processing on the steel production line data to obtain the initial data.

[0081] Specifically, the cognitive decision-making unit utilizes cognitive reasoning and decision-making technologies that combine few-shot learning techniques with domain knowledge. Based on the combination of industrial domain knowledge, few-shot learning, and data models, cognitive reasoning and decision-making technologies are adopted to optimize existing quality prediction, quality traceability, instrument control, production line management, etc. in industrial manufacturing, assisting the intelligent development of enterprises. The method based on knowledge-data fusion can effectively alleviate the problem of scarce effective samples in industrial scenarios and greatly reduce the dependence on data. The models formed for process planning, ratio decision-making, fault diagnosis, defect detection, etc. can effectively realize production intelligence. At the same time, by adaptively adjusting parameters and pre-judging quality, production costs can be effectively reduced and production defects can be avoided in a timely manner.

[0082] In this solution, by making decision processing on the steel production line data, initial data is obtained, ensuring the accuracy and reliability of the initial data.

[0083] In an implementable solution, the cognitive intelligence layer 3 includes an intelligent planning unit 35;

[0084] The intelligent planning unit 35 is used to perform planning processing on the steel production line data to obtain initial data.

[0085] Specifically, the intelligent planning unit adopts a probabilistic knowledge graph. Since the knowledge in the cognitive scenario is not static, nor is it "yes or no" deterministic knowledge, but a kind of dynamic knowledge, this dynamic knowledge can more effectively depict real-world data. At the same time, people's representation of knowledge is not a "yes or no" deterministic fact, but contains a large number of uncertain facts classified as "possibly". Therefore, a probabilistic knowledge graph is proposed and constructed to represent dynamic knowledge, realize the depiction of non-deterministic facts, and perform effective reasoning based on this, and conduct intelligent planning for industrial scenarios.

[0086] In this solution, by performing planning processing on the steel production line data to obtain initial data, the effectiveness and reliability of the initial data are ensured.

[0087] In an implementable solution, the large model layer 4 is also used to fuse a general model and a specialized model based on the initial data and the knowledge base to obtain a large model for the steel industry.

[0088] Specifically, the large model layer is the General-Specialized Large Model Layer, which is the core component in the large model architecture except for the cognitive intelligence engine layer. This layer provides efficient and accurate solutions for steel production by integrating the advantages of general models and specialized models.

[0089] The architecture design of the general and specialized large model layer is based on the deep integration of general models and industry-specific models, combined with multi-modal data processing capabilities, to form an intelligent solution applicable to the steel industry.

[0090] The general model and the specialized model are seamlessly integrated into the large model architecture in the steel industry. By combining general AI (Artificial Intelligence) capabilities and industry expertise, it significantly improves production efficiency, product quality, and operational effectiveness, driving the digital transformation and sustainable development of the steel industry.

[0091] In this solution, by integrating the general model and the specialized model, a large model for the steel industry is obtained, ensuring the reliability and practicality of the large model for the steel industry.

[0092] In an implementable solution, the general model includes at least one of DeepSeek, Qwen, Wenxin Yiyan, and Llama 3.

[0093] Specifically, DeepSeek: DeepSeek is renowned for its powerful data processing and intelligent decision-making capabilities and is suitable for scenarios such as production process optimization and equipment maintenance prediction in the steel industry. By integrating DeepSeek, the large model architecture can handle complex production data and provide efficient decision support.

[0094] Qwen: As a large language model developed by Alibaba Tongyi, Qwen performs excellently in natural language processing. It can be used for tasks such as parsing technical documents, generating production reports, and handling customer inquiries, improving the processing efficiency of text data.

[0095] Wenxin Yiyan: The Wenxin Yiyan model launched by Baidu is based on the ERNIE (Enhanced Representation through Knowledge Integration, a pre-trained language model architecture based on knowledge enhancement) architecture and has excellent semantic understanding and generation capabilities. In the steel industry, it can be used in scenarios such as quality control, product description generation, and customer relationship management to provide accurate language support.

[0096] Llama 3: Llama 3 is a large model developed by Meta and has advantages in fields such as materials science research and process optimization. By integrating Llama 3, the large model architecture can deeply analyze complex problems in steel production and provide a scientific basis.

[0097] In this solution, by integrating multiple types of general models, the reliability and practicality of the large model for the steel industry are ensured.

[0098] In an implementable solution, the dedicated model includes at least one of a process optimization large model, a time series prediction large model, a visual inspection large model, and a natural language processing large model;

[0099] Among them, the process optimization large model is used to optimize the coal blending process, the time series prediction large model is used to predict the quality of continuous casting ladles, the visual inspection large model is used to detect surface defects of products, and the natural language processing large model is used to parse natural language.

[0100] Specifically, the process optimization large model is applied to intelligent coal blending, optimizing the coal mixing ratio through deep learning technology to improve production efficiency and reduce energy consumption. The time series prediction large model is used for continuous casting quality determination, predicting quality results based on historical data to ensure product standards. The visual inspection large model realizes surface defect detection, automatically identifying product flaws and reducing manual inspection. The natural language processing large model processes language scenarios such as text parsing, customer consultation, and report generation to improve information processing efficiency.

[0101] In this solution, through various types of dedicated models, the reliability and practicality of the large model in the steel industry are ensured.

[0102] In an implementable solution, the large model in the steel industry adopts at least one hybrid neural network architecture of the Transformer architecture, convolutional neural network, and recurrent neural network.

[0103] Specifically, the Transformer architecture: is used to process text data such as production reports, technical documents, etc., providing efficient semantic understanding and generation capabilities.

[0104] CNN (Convolutional Neural Network): is used to process image data such as surface defect detection, equipment status monitoring, etc., providing efficient image recognition capabilities.

[0105] RNN (Recurrent Neural Network): is used to process time series data such as sensor data, production process data, etc., providing dynamic data analysis capabilities.

[0106] In this solution, through various types of hybrid neural network architectures, the large model in the steel industry is obtained, ensuring the reliability and practicality of the large model in the steel industry.

[0107] In an implementable solution, the application display layer 5 is also used to input actual data in the actual application scenario into the large model in the steel industry to output target data.

[0108] Specifically, the application display layer, namely the application and display layer in the steel industry, is the top layer of the entire large model architecture, aiming to transform advanced model and data analysis capabilities into actual industrial applications and present them to users in an intuitive way. This layer is a bridge connecting technology and users, ensuring that complex data processing and model reasoning can be effectively utilized to solve practical problems.

[0109] In this solution, by inputting the actual data in the actual application scenario into the large model of the steel industry to output the target data, it effectively solves the actual problems in the actual application scenario of the steel industry and improves the customer experience.

[0110] In an implementable solution, the actual application scenario includes at least one of the intelligent coal blending scenario, production quality control scenario, surface defect detection scenario, energy management scenario, process unmanned scenario, and intelligent shift scheduling scenario.

[0111] Specifically, in the intelligent coal blending scenario, the process optimization large model and the intelligent coal blending large model are used to provide the optimal coal blending plan according to real-time data and historical performance, reducing costs and improving efficiency;

[0112] In the production quality control scenario, the time series prediction large model and the quality large model are adopted to predict quality problems and achieve proactive adjustment in the production process to ensure product quality;

[0113] The surface defect detection scenario, namely the continuous casting, hot rolling, and cold rolling surface inspection scenarios, uses the visual detection large model and the surface defect detection large model to automatically identify surface defects, replacing manual inspection and improving detection efficiency and accuracy;

[0114] In the energy management scenario, the time series prediction large model and the energy management large model are used to predict energy demand, optimize energy distribution and use, reduce costs and improve environmental benefits;

[0115] In the process unmanned scenario, the visual large model is used to realize automatic equipment monitoring, support unmanned production, and improve safety and efficiency;

[0116] In the intelligent shift scheduling scenario, the process optimization, time series prediction large model, and shift scheduling large model are combined to optimize the production plan, meeting the demand while reducing costs and resource consumption.

[0117] In this solution, by inputting the actual data of multiple actual application scenarios into the large model of the steel industry to output the target data, the universality of the large model of the steel industry is ensured and the customer experience is improved.

[0118] The following combines specific examples to illustrate the working principle of the large model architecture in the steel industry of this embodiment, such as Figure 3As shown in the figure, the large model architecture in the steel industry includes a data access layer, a knowledge and technology integration layer, a cognitive intelligence engine layer, a general and specialized large model layer, and an application and display layer for the steel industry. The data access layer is used to receive at least one of the basic automation data of the steel production line, the L2 process control level system data, the L3 production control level system data, and the MES (Manufacturing Execution System) system data. The knowledge base of the knowledge and technology integration layer includes at least one of coal blending knowledge data, continuous casting quality data, surface defect data, energy management data, visual data, and shift scheduling data. The cognitive intelligence engine layer includes five major capabilities: knowledge reasoning, few-shot generation, multimodal understanding, cognitive decision-making, and intelligent planning, which are respectively implemented through a knowledge reasoning unit, a few-shot generation unit, a multimodal understanding unit, a cognitive decision-making unit, and an intelligent planning unit. The general and specialized large model layer includes at least one general model such as DeepSeek, LIAMA3, Qwen, Wenxin Yiyan, and GPT4 (Generative Pre-trained Transformer 4), uses instruction tuning to enhance the model's ability to understand and follow human instructions, and adopts the Transformer architecture. The general and specialized large model layer includes at least one specialized model such as a process optimization large model, a time series prediction large model, a visual inspection large model, a natural language processing large model, an intelligent coal blending large model, a continuous casting, hot rolling, and cold rolling quality large model, a surface defect detection large model, an energy management large model, a process inspection large model, and a shift scheduling large model. The actual application scenarios of the application and display layer for the steel industry include at least one of the intelligent coal blending scenario, the production line quality control scenario, the continuous casting, hot rolling, and cold rolling surface inspection scenarios, the energy management scenario, the process unmanned scenario, and the intelligent shift scheduling scenario.

[0119] In this embodiment, through the five-layer large model architecture of the data access layer, the data fusion layer, the cognitive intelligence layer, the large model layer, and the application display layer, the convenient utilization of the large model is realized, effectively solving the actual problems in the steel industry and improving the customer experience.

[0120] Although the specific implementation manners of the present disclosure have been described above, those skilled in the art should understand that this is only an example, and the protection scope of the present disclosure is defined by the appended claims. Without departing from the principles and essence of the present disclosure, those skilled in the art can make various changes or modifications to these implementation manners, but these changes and modifications all fall within the protection scope of the present disclosure.

Claims

1. A large model architecture in the steel industry, characterized in that: The big model architecture includes a data access layer, a data fusion layer, a cognitive intelligence layer, a big model layer and an application display layer; The data access layer is used to receive steel production line data and output it to the data fusion layer; The data fusion layer is used to construct a knowledge base based on the steel production line data and the steel knowledge data; The cognitive intelligence layer is used to analyze and process the steel production line data to obtain initial data; The large model layer is used to obtain a large model of the steel industry based on the initial data and the knowledge base; The application display layer is used to obtain target data based on the steel industry big model.

2. The large model framework in the steel industry as claimed in claim 1, characterized in that: The data access layer is also used to receive the steel production line data from at least one of a steel production line, an inspection and testing system, a process control level system and a production control level system.

3. The large model framework in the steel industry as claimed in claim 1, characterized in that: The data fusion layer is also used to fuse the steel production line data and the steel knowledge data using knowledge graph technology to construct the knowledge base.

4. The large model framework in the steel industry as claimed in claim 3, characterized in that: The steel knowledge data includes at least one of coal blending knowledge data, continuous casting quality data, surface defect data, energy management data, visual data, and shift and production scheduling data.

5. The large model framework in the steel industry as claimed in claim 1, characterized in that: The cognitive intelligence layer includes a knowledge reasoning unit; The knowledge reasoning unit is used to reason about the steel production line data to obtain the initial data; and / or, The cognitive intelligence layer includes a small sample generation unit; The small sample generation unit is used to generate the initial data based on the steel production line data by using a small sample learning technology; and / or, The cognitive intelligence layer includes a multimodal understanding unit; The multimodal understanding unit is used to understand the steel production line data in multiple modes to obtain the initial data; and / or, The cognitive intelligence layer includes a cognitive decision-making unit; The cognitive decision unit is used to perform decision processing on the steel production line data to obtain the initial data; and / or, The cognitive intelligence layer includes an intelligent planning unit; The intelligent planning unit is used to plan and process the steel production line data to obtain the initial data.

6. The large model framework in the steel industry as claimed in claim 1, characterized in that: The large model layer is also used to fuse the general model and the special model based on the initial data and the knowledge base to obtain the steel industry large model.

7. The large model framework in the steel industry as claimed in claim 6, characterized in that: The general model includes at least one of DeepSeek, Qwen, Wenxinyiyan and LIAMA3; and / or, The dedicated model includes at least one of a process optimization model, a time series prediction model, a visual inspection model, and a natural language processing model; Among them, the process optimization large model is used to optimize the coal blending process, the time series prediction large model is used to predict the quality of the continuous casting ladle, the visual inspection large model is used to detect surface defects of the product, and the natural language processing large model is used to parse natural language.

8. The large model framework in the steel industry as claimed in claim 6, characterized in that: The steel industry large model adopts at least one hybrid neural network architecture among the Transformer architecture, convolutional neural network and recurrent neural network.

9. The large model framework in the steel industry as claimed in claim 1, characterized in that: The application display layer is also used to input actual data in actual application scenarios into the steel industry big model to output the target data.

10. The large model framework in the steel industry according to claim 9, characterized in that: The actual application scenarios include at least one of intelligent coal blending scenarios, production quality control scenarios, surface defect detection scenarios, energy management scenarios, unmanned process scenarios, and intelligent scheduling and production scheduling scenarios.

Images