Oxygen-enriched side-blown copper smelting process control system based on LLM and RAG

Abstract

The invention relates to an LLM and RAG-based oxygen-enriched side-blown copper smelting process control system. The system comprises a parameter configuration module, an expert experience and literature management module, a vector database module, an RAG module, a cue word module and an LLM module. The parameter configuration module is used for realizing parameter configuration of a key process; the expert experience and literature management module is used for configuring key parameters, mapping words into a low-dimensional dense vector space, and converting the words into semantic vectors so as to facilitate efficient numerical calculation; the vector database module is used for carrying out structured storage on the converted semantic vectors; the RAG module is used for retrieving and introducing latest and related data from a knowledge base in real time; the cue word module is used as the input of the large model to generate a recommendation result which better meets the expectation; the LLM module is used for generating a recommendation result according to inference and analysis of the cue word; the LLM and RAG technologies are applied to the oxygen-enriched side-blown copper smelting process control system, and real-time and accurate control over the smelting process is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of oxygen-enriched side-blown copper smelting production, specifically an oxygen-enriched side-blown copper smelting process control system based on LLM and RAG. Background Art

[0002] In the field of copper smelting, the oxygen-enriched side-blown smelting technology is an important smelting process. In the actual production process, the operating conditions of oxygen-enriched side-blown copper smelting are complex and diverse, and changes in parameters such as raw material ratio, oxygen concentration, air volume, and fuel supplement amount will all have a significant impact on the smelting process. The literature "Design and Application of the Oxygen-Enriched Side-Blown Copper Smelting Process Control System" was published in Southern Metals Magazine. This system realizes control by building a mathematical model of the side-blown smelting process and a metallurgical thermodynamics database, mainly relying on mechanism models such as material balance models and heat balance models. These models often need to simplify and assume the complex smelting process during establishment, resulting in a certain deviation between the calculation results and actual production.

[0003] In order to solve the problems existing in the prior art and improve the control accuracy and stability of the oxygen-enriched side-blown copper smelting process, it is necessary to introduce more advanced technical means. In recent years, large language models (LLMs) and retrieval-augmented generation (RAG) technologies have made remarkable progress in the field of natural language processing. Their powerful language generation ability and dynamic knowledge supplementation ability provide new ideas for the control of complex industrial processes.

[0004] Therefore, how to reduce the usage threshold of control system maintenance and update and make the control system more generalizable has become an urgent problem to be solved. Summary of the Invention

[0005] The main purpose of the present invention is to solve the problems existing in the above-mentioned prior art, and provide an oxygen-enriched side-blown copper smelting process control system based on LLM and RAG. Introducing LLM and RAG technologies into the oxygen-enriched side-blown copper smelting process control system can significantly improve the control accuracy and stability of the control system, realize real-time and precise regulation of the smelting process, improve production quality and efficiency, reduce production costs, and is of great significance for promoting the intelligent development of the copper smelting industry.

[0006] The specific solution of the present invention is as follows: An oxygen-enriched side-blow copper smelting process control system based on LLM and RAG is divided into an offline part and an online part. The offline part includes a parameter configuration module, an expert experience and literature management module, and a vector database module. The online part includes a RAG module, a prompt module, and an LLM module. The parameter configuration module is used to configure the parameters of key processes. The expert experience and literature management module is used to map words into a low-dimensional dense vector space according to the key parameters configured in the parameter configuration module for the expert experience described in natural language and the content of literature materials, and convert them into semantic vectors for efficient numerical calculation. The vector database module is used to structurally store the converted semantic vectors. The RAG module is used to combine the large model with the relevant information of expert experience and literature materials in the external vector database, and retrieve and introduce the latest and relevant literature materials from the knowledge base in real time. The prompt module is used to integrate the abnormal state description of real-time data with the expert experience and literature information selected by vector similarity as the input of the large model, and adjust the prompt words to make the large model generate more satisfactory recommended results. The LLM module is used to perform reasoning and analysis according to the prompt words and generate recommended results.

[0007] Further, the key process parameter names, point addresses, upper and lower limits, and threshold information for trend judgment configured by the parameter configuration module are used for the automatic parsing and structural storage of expert experience and the function module for converting real-time data into abnormal state descriptions.

[0008] Further, for relevant process documents, the system first performs OCR recognition on pdf files, converts the pdf-format materials into txt-format literature materials, and then segments the txt-format literature materials to form literature paragraph corpora. For relevant process operation expert experience, the system automatically slices the expert experience into multiple pieces of expert experience first, then performs semantic extraction through a large language model, converts the expert experience into a parameter description form, and then through a semantic model, the extracted structured expert experience original text is mapped into a low-dimensional dense vector space through the word embedding technology of the semantic model, and converted into semantic vectors and stored in the corresponding expert experience vector database.

[0009] Further, the prompt module combines the real-time production process data converted into abnormal state descriptions through the data processing module with the expert experience and literature information selected by the vector similarity algorithm to form a prompt word indicating the abnormal state description, relevant expert experience, and relevant literature materials of the input information, and indicates the dictionary format of the output requirements.

[0010] Furthermore, the present invention also provides an oxygen-enriched side-blow copper smelting process control method based on LLM and RAG. Using the above system, the operation steps are as follows: S1. Configure parameters, and configure the point table addresses, Chinese names, upper and lower limits, trend judgment thresholds, and related information of the key parameters that need to be monitored and paid attention to for abnormalities during the oxygen-enriched side-blow copper smelting process; S2. Upload expert experience. Upload the expert experience sorted out at the oxygen-enriched side-blow copper smelting production site or the standard operation guide of expert experience refined by process personnel to the SOP management in the oxygen-enriched side-blow copper smelting large model system. The system extracts each item structually according to the measured points such as the SO2 concentration at the outlet of the waste heat boiler and the melting slag temperature configured in the parameter configuration. After being reviewed by the user, it is converted into a vector through semantic model embedding technology and stored in the SOP vector database dedicated to oxygen-enriched side-blow copper smelting; S3. Upload literature materials. Upload the literature materials in pdf format related to oxygen-enriched side-blow copper smelting to the oxygen-enriched side-blow copper smelting large model system. The system automatically performs OCR recognition and paragraph segmentation, and converts the segmented natural language description into a vector through semantic model embedding technology and stores it in the literature vector database dedicated to oxygen-enriched side-blow copper smelting; S4. Online real-time inference. The data processing module collects multi-source heterogeneous data in real time, effectively monitors and judges the over-limit situation and fluctuation situation of multi-source heterogeneous abnormal data, and converts it into a description of the abnormal state of the key parameters during the oxygen-enriched side-blow copper smelting process; S5. LLM+RAG real-time recommendation. Through RAG technology, combine the information of the SOP vector database dedicated to oxygen-enriched side-blow copper smelting and the relevant literature vector database to form a prompt word, and request the interface of the oxygen-enriched side-blow copper smelting large model inference, and then finally output the recommended adjustment result of the control variable in the oxygen-enriched side-blow copper smelting production.

[0011] Furthermore, in S4, a script is separately made for the part with complex processing logic, and the description of the abnormal state of the key parameters is in the form of natural language.

[0012] The present invention has the following advantages compared with the prior art: Applying LLM and RAG technologies to the oxygen-enriched side-blow copper smelting process control system can achieve real-time and precise control of the smelting process. Through RAG technology, the system can obtain the latest smelting data and expert knowledge from the external vector database, making up for the problems of outdated or insufficient knowledge of LLM; avoiding the need to modify the code when there is a need to modify or supplement the expert experience of the process; and avoiding the problem that the control system needs to be modified or even rewritten when changing to a new business scenario. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 is the schematic diagram of the system architecture of the present invention; Figure 2It is a schematic diagram of the offline part of the present invention; Figure 3 It is a schematic diagram of the online part of the present invention; Figure 4 It is a flowchart of the control method of the present invention. Detailed implementation manners

[0014] Refer to Figures 1-3 , this embodiment is an oxygen-enriched side-blow copper smelting process control system based on LLM and RAG, which is divided into an offline part and an online part. The offline part includes a parameter configuration module, an expert experience and literature management module, and a vector database module. The online part includes a RAG module, a prompt module, and an LLM module; the parameter configuration module of the offline part configures the names of key process parameters, point addresses, as well as information such as upper and lower limits and thresholds for trend judgment, which will be used for function modules such as automatic parsing and structured storage of expert experience, and conversion of real-time data into abnormal state descriptions.

[0015] For the expert experience and literature management module and the vector database module of the offline part, for relevant process documents, the system will first perform OCR recognition on pdf files, convert the pdf-format materials into txt-format literature materials. Then, the txt-format literature materials will be segmented to form literature paragraph corpora. Using natural language processing technology, the original text of the literature paragraph corpora will be mapped into a low-dimensional dense vector space through the word embedding technology of the semantic model, converted into semantic vectors of each literature paragraph, and stored in the corresponding vector database. For relevant expert experience of process operations, the system will automatically split the expert experience into multiple pieces of expert experience, and then perform semantic extraction through a large language model, converting each piece of expert experience into a parameter description form such as "parameter 1 is high, parameter 2 is low, parameter 3 has an upward trend, parameter 4 has a downward trend". Then, through the semantic model, the extracted structured expert experience original text will be mapped into a low-dimensional dense vector space through the word embedding technology of the semantic model, converted into semantic vectors, and stored in the corresponding expert experience vector database.

[0016] For the RAG module of the online part, using the vector database mentioned above as the knowledge base, the retrieval module of RAG technology will quickly find the literature materials most relevant to the parameters in the current abnormal state description in the knowledge base. This step usually involves efficient vectorized retrieval and similarity calculation to ensure that relevant information can be found quickly and accurately. The retrieved literature materials will be used as the context input for the large model to generate recommended parameters.

[0017] The prompt word module of the online part converts real-time production process data into abnormal state descriptions through a data processing module, and combines expert experience and literature information screened by vector similarity (vector similarity algorithms such as cosine similarity, Euclidean distance, etc.) to form prompt words, indicating that the input information has abnormal state descriptions, relevant expert experience, and relevant literature materials, and indicating the dictionary format of the output requirements. Guide the large model to generate recommended results that meet expectations.

[0018] In the LLM module of the online part, after receiving the prompt words, the large model will use its powerful natural language processing ability and knowledge reserve to conduct reasoning and analysis. According to the information in the prompt words, combined with its own knowledge base and logical reasoning ability, it will finally give a return value in dictionary format. After the output module parses the return body accordingly, the recommended result can be obtained.

[0019] See Figure 4 For the method of controlling the oxygen-enriched side-blown copper smelting process using the above control system, the following steps are included: S1: Configure parameters, configure the point table addresses (such as tagname in DCS), Chinese names, upper and lower limits, trend judgment thresholds, etc. of the key parameters (such as SO2 concentration at the outlet of the waste heat boiler, average temperature difference of the vertical water jacket, etc.) that need to be monitored and concerned about whether they are abnormal during the oxygen-enriched side-blown copper smelting process. Prepare for later semantic extraction of expert experience and real-time data processing for side-blown smelting control.

[0020] S2: Upload expert experience. Upload the expert experience sorted out at the oxygen-enriched side-blown copper smelting production site or the standard operation guide of expert experience refined by process personnel to the SOP management in the oxygen-enriched side-blown copper smelting large model system. The system extracts each item of the SO2 concentration at the outlet of the waste heat boiler, melting slag temperature, etc. measured points configured in the parameter configuration in a structured manner, and after user review, converts it into a vector through semantic model embedding technology and stores it in the SOP vector database dedicated to oxygen-enriched side-blown copper smelting.

[0021] S3: Upload literature materials. Upload the literature materials in pdf format related to oxygen-enriched side-blown copper smelting to the oxygen-enriched side-blown copper smelting large model system. The system automatically performs OCR recognition and paragraph segmentation, and converts the segmented natural language description into a vector through semantic model embedding technology and stores it in the literature vector database dedicated to oxygen-enriched side-blown copper smelting.

[0022] After completing the offline operation of the above oxygen-enriched side-blown copper smelting large model system, online reasoning of key control variables (such as oxygen-fuel ratio setting, instantaneous coal flow) can be carried out.

[0023] S4: Online real-time inference part. The real-time data processing module collects multi-source heterogeneous data (DCS process data, fluorescence assay data, manual input, etc.) in real time. Therefore, a separate script needs to be written for parts with complex processing logics (such as data processing of SO2 concentration data at the outlet of the waste heat boiler, where methods like median filtering need to be used). For the over-limit and fluctuation conditions of multi-source heterogeneous abnormal data, effective monitoring and judgment are carried out, and they are converted into descriptions of abnormal states of key parameters in the copper smelting process with oxygen-enriched side blowing.

[0024] S5: LLM + RAG real-time recommendation. After forming the description of the abnormal state of the key parameters in the copper smelting process with oxygen-enriched side blowing in step 4, prompt words will be formed by combining the information of the SOP vector database dedicated to the copper smelting process with oxygen-enriched side blowing and the relevant literature vector database through the RAG technology, and the interface for the inference of the large model for the copper smelting process with oxygen-enriched side blowing will be requested. Furthermore, the recommended adjustment results of the control variables in the copper smelting production with oxygen-enriched side blowing (such as oxygen-to-fuel ratio setting, instantaneous coal flow rate, instantaneous solvent flow rate, primary air tuyere) will be finally output.

Claims

1. An oxygen-enriched side-blowing copper smelting process control system based on LLM and RAG, characterized in that: It is divided into an offline part and an online part. The offline part includes a parameter configuration module, an expert experience and literature management module, and a vector database module. The online part includes a RAG module, a prompt module, and an LLM module. The parameter configuration module is used to configure the parameters of key processes. The expert experience and literature management module is used to map words into a low-dimensional dense vector space according to the key parameters configured in the parameter configuration module for the expert experience and literature content described in natural language, and convert them into semantic vectors for efficient numerical calculation. The vector database module is used for structured storage of the converted semantic vectors. The RAG module is used to combine the large model with the relevant information of expert experience and literature in the external vector database, and retrieve and introduce the latest and relevant literature from the knowledge base in real time. The prompt module is used to integrate the abnormal state description of real-time data with the expert experience and literature information selected by vector similarity as the input of the large model, and adjust the prompt words to make the large model generate more satisfactory recommendation results. The LLM module is used to perform reasoning and analysis according to the prompt words and generate recommendation results.

2. The oxygen-enriched side-blowing copper smelting process control system based on LLM and RAG according to claim 1, characterized in that: The key process parameter names, point addresses, upper and lower limits, and threshold information for trend judgment configured by the parameter configuration module are used for the automatic parsing and structured storage of expert experience, and the function module for converting real-time data into abnormal state descriptions.

3. The oxygen-enriched side-blowing copper smelting process control system based on LLM and RAG according to claim 1, characterized in that: For the relevant process documents, the system first performs OCR recognition on the pdf files, converts the pdf-format materials into txt-format literature materials, and then segments the txt-format literature materials to form literature paragraph corpora. For the relevant expert experience of process operations, the system will first split the expert experience into multiple pieces of expert experience, then perform semantic extraction through the large language model, convert the expert experience into a parameter description form, and then through the semantic model, map the words into a low-dimensional dense vector space through the word embedding technology of the semantic model, convert them into semantic vectors and store them in the corresponding expert experience vector database.

4. The oxygen-enriched side-blowing copper smelting process control system based on LLM and RAG according to claim 1, characterized in that: The prompt module combines the real-time production process data converted into abnormal state descriptions through the data processing module with the expert experience and literature information selected by the vector similarity algorithm to form a prompt word indicating that the input information has abnormal state descriptions, relevant expert experience, and relevant literature materials, and indicating the dictionary format of the output requirements.

5. The control method for the oxygen-enriched side-blow copper smelting process based on LLM and RAG is characterized in that: Using the system according to any one of claims 1-4, the operation steps are as follows: S1. Configure parameters, and configure the point table address, Chinese name, upper and lower limits, trend judgment threshold, and related information of the key parameters that need to be monitored and concerned about whether they are abnormal during the oxygen-enriched side-blown copper smelting process. S2. Upload expert experience. Upload the expert experience sorted out at the site of oxygen-enriched side-blow copper smelting production or the standard operation instruction manual of expert experience refined by process personnel to the SOP management in the oxygen-enriched side-blow copper smelting large model system. The system extracts each item structurally according to the measured points such as the SO2 concentration at the outlet of the waste heat boiler and the smelting slag temperature configured in the parameter configuration. After being reviewed by the user, it is converted into a vector through semantic model embedding technology and stored in the SOP vector database dedicated to oxygen-enriched side-blow copper smelting; S3. Upload literature materials. Upload the literature materials in pdf format related to oxygen-enriched side-blow copper smelting to the oxygen-enriched side-blow copper smelting large model system. The system automatically performs OCR recognition and paragraph segmentation, and converts the segmented natural language description into a vector through semantic model embedding technology and stores it in the literature vector database dedicated to oxygen-enriched side-blow copper smelting; S4. Online real-time reasoning. The data processing module collects multi-source heterogeneous data in real time. For the over-limit situation and fluctuation situation of multi-source heterogeneous abnormal data, effective monitoring and judgment are carried out, and it is converted into a description of the abnormal state of key parameters in the oxygen-enriched side-blow copper smelting process; S5. LLM+RAG real-time recommendation. Through the RAG technology, a prompt is formed by combining the information in the SOP vector database dedicated to oxygen-enriched side-blow copper smelting and the relevant literature vector database, and the interface for reasoning of the oxygen-enriched side-blow copper smelting large model is requested, and finally the recommended adjustment result of the control variables in the oxygen-enriched side-blow copper smelting production is output.

6. The method for controlling the oxygen-enriched side-blowing copper smelting process based on LLM and RAG according to claim 5, characterized in that: In S4, a script is separately made for the part with complex processing logic, and the description of the abnormal state of key parameters is in the form of natural language.

Images