A cement plant technical equipment question and answer system and method based on a large language model

Abstract

The application provides a cement plant technical equipment question and answer system and method based on a large language model, which comprises a front-end interaction module, receives a cement equipment document to be processed and an input question and answer request; a back-end service module, performs permission verification and format checking on the received request; a knowledge base docking module, analyzes the cement equipment document and stores it into a vector knowledge base, and performs context retrieval according to the question and answer request; a data storage module, stores cement equipment basic data and context data; and a large model interaction module, fuses the context data and the cement equipment basic data, constructs a cement industry standardized prompt word, and calls a preset large language model to generate a question and answer result. The cement equipment document is automatically processed in a full process from uploading to intelligent question and answer, which can greatly shorten the period for obtaining professional knowledge of cement equipment, improve the precision of professional term retrieval, and ensure that the generated question and answer result meets the rigorous requirements of the industry.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and industrial knowledge management, and in particular relates to a question-and-answer system and method for cement plant technical equipment based on a large language model. Background Technology

[0002] The technical equipment system of cement plants is complex, encompassing dozens of core equipment types such as vertical mills, rotary kilns, and grate coolers. In actual production and operation, relevant technical manuals, equipment basic parameter documents, fault handling plans, and on-site maintenance procedures have long existed in the form of paper documents, scattered electronic files, and personal experience notes, lacking unified and efficient management and utilization methods. With the popularization of artificial intelligence technology, some industrial sectors have begun to introduce knowledge base question-answering systems based on large language models to improve retrieval efficiency and the level of intelligence in technical guidance.

[0003] However, existing general question-and-answer systems have significant limitations when dealing with specialized applications in the cement equipment field. Traditional document parsing schemes typically employ a uniform, fixed-length, coarse-grained text truncation method, failing to consider the layout characteristics of cement equipment documents, which are characterized by "dense technical parameters, numerous cross-page tables, and strong correlation between parameters and fault cases." This easily leads to a rigid separation between complete sets of equipment technical parameters and corresponding fault handling solutions, resulting in logical gaps in the context of the knowledge entered into the database.

[0004] Furthermore, existing retrieval mechanisms are mostly based on general semantic logic for preliminary matching, failing to accurately capture specialized parameters and technical terms with strong vertical domain characteristics, such as "grate gap in grate cooler" or "rotary kiln support roller temperature threshold," resulting in extremely low retrieval recall rates for core technical parameters. At the same time, the responses output by large language models often do not follow the common fault-handling logic steps in the cement industry, nor do they accurately label key operating parameters, making it difficult to provide readily usable professional guidance to on-site technicians. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention aims to propose a question-and-answer system and method for cement plant technical equipment based on a large language model, in order to solve the problems of easy context breakage and low accuracy of technical term retrieval when the existing general question-and-answer system processes cement equipment documents.

[0006] To achieve the above objectives, the technical solution of the present invention is implemented as follows:

[0007] A question-and-answer system for cement plant technical equipment based on a large language model, comprising: The front-end interaction module is used to receive cement equipment documents to be processed and input Q&A requests, and to display the processed Q&A results. The backend service module is connected to the frontend interaction module and is used to perform permission verification and format validation on the received question and answer requests, and to act as a data relay node between the modules to forward data. The knowledge base docking module is connected to the backend service module and is used to parse the cement equipment documents and store them in the vector knowledge base, as well as to perform context retrieval based on the question and answer request. The data storage module is used to store basic data of cement equipment, as well as context data retrieved by the knowledge base docking module; The large model interaction module is connected to the knowledge base interface module, data storage module and backend service module respectively. It is used to integrate the context data with the basic data of cement equipment, construct standardized prompt words for the cement industry, call the preset large language model to generate question and answer results, and feed them back to the backend service module.

[0008] Furthermore, the knowledge base docking module includes a document parsing submodule, which performs text segmentation using a preset character length and a preset overlap rate when processing the cement equipment document; The preset character length is 490 to 510 characters, and the preset overlap rate is 8% to 12%.

[0009] Furthermore, for the table content in the cement equipment document, the document parsing submodule identifies the table boundaries and hierarchical relationships through structured parsing technology, converts the table content into structured data, and performs dimensional segmentation according to equipment parameter categories or fault types. At the same time, it binds a unified table identifier to all segments belonging to the same table.

[0010] Furthermore, for multi-page or complex parameter tables, numerical parameters in the table are extracted as feature words, and the feature words are fused with the table text vector before being stored in the database.

[0011] Furthermore, the knowledge base docking module also includes a context retrieval submodule, which performs retrieval based on a preset semantic similarity threshold and returns a preset number of document fragment contexts carrying document source identifiers in descending order of similarity. The semantic similarity threshold is greater than or equal to 0.8.

[0012] Furthermore, the large model interaction module includes a data splicing submodule, which is used to standardize the retrieved context data and cement equipment basic data, and prioritize the fault threshold and core technical parameters in the prompt words according to the professional characteristics of cement equipment. The standardized prompt template should include at least: contextual basis, basic equipment parameters, user questions, and industry-specific answer requirements.

[0013] Furthermore, the industry-specific answer requirements include: for fault-related questions, the large language model is required to answer according to the logical steps of "shutdown, inspection, handling, and trial operation"; for parameter-related questions, the data must be accurately cited and the corresponding physical quantity units must be labeled.

[0014] Furthermore, the data storage module includes a relational database submodule and a non-relational database submodule; The relational database submodule is used to store basic data of cement equipment; The non-relational database submodule serves as a caching layer, used to store intermediate context data, temporary prompts, and user identity tokens corresponding to the question-and-answer request, with each type of data having a different caching duration.

[0015] Furthermore, the backend service module includes a permission verification submodule, which is used to perform hierarchical permission verification on users based on user identity tokens, so as to grant different system operation permissions according to different roles.

[0016] A question-answering method for cement plant technical equipment based on a large language model includes the following steps: Receive user-inputted Q&A requests and / or pending cement equipment documents, and verify the identity and permissions and format validity of the Q&A requests; Based on the user's identity and permissions, determine whether to parse and process the cement equipment document and store it in the vector knowledge base; If the identity and permission verification and format validity verification pass, then semantic retrieval is performed in the vector knowledge base according to the question and answer request to obtain context data, and pre-stored cement equipment basic data is obtained simultaneously. The acquired contextual data is integrated with the basic data of cement equipment to construct standardized prompts for the cement industry; a preset large language model is invoked to generate corresponding question-and-answer results based on the standardized prompts for the cement industry, and finally fed back to the front-end interactive module for visualization.

[0017] Compared with existing technologies, the question-answering system and method for cement plant technical equipment based on a large language model described in this invention have the following advantages: This invention uses a backend service module as a relay node to deeply decouple and connect the frontend's unstructured document reception, the knowledge base's vectorized storage, and the large model's logical reasoning. This changes the current situation where each link operates independently and the process has poor connectivity in existing technologies, and realizes fully automated processing from document upload to intelligent question answering, greatly shortening the path for technicians to acquire knowledge about cement equipment.

[0018] By leveraging the large model interaction module, the real-time contextual data retrieved by the knowledge base docking module is deeply integrated with the pre-stored structured basic data in the data storage module. Based on this, standardized prompt words for the cement industry are constructed. This effectively solves the illusion problem and the lack of professional context that exist when the general large language model deals with complex equipment in the vertical field of cement (such as vertical mills, rotary kilns, etc.). It ensures that the generated question and answer results are both industry-relevant and have extremely high technical reference value.

[0019] By setting a text segmentation mechanism with preset character length and preset overlap rate through the document parsing submodule, and combining it with the structured parsing and dimensional segmentation of the table content in cement equipment documents, the segmented text and table fragments can fully retain the contextual relevance of individual technical parameters and fault handling logic, avoiding knowledge context breaks caused by improper segmentation, and significantly improving the semantic integrity and usability of subsequent retrieval fragments.

[0020] By configuring a data splicing submodule in the large model interaction module, fault thresholds and core technical parameters are prioritized in the prompts based on the professional characteristics of cement equipment. A standardized prompt template containing contextual basis, basic equipment parameters, user questions, and industry-specific answer requirements is adopted to constrain the large language model to answer fault-related questions according to the logical steps of "shutdown, inspection, handling, and trial operation" and to answer parameter-related questions with accurate annotation of physical quantity units. This ensures that the system-generated answers directly conform to the professional usage habits of cement industry technicians and can be used for on-site operation guidance without secondary processing. Attached Figure Description

[0021] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a system control block diagram provided in an embodiment of the present invention; Figure 2 This is an example of the core table structure design for a relational database submodule provided in an embodiment of the present invention; Figure 3 This is a design example of the core cache for a non-relational database submodule provided in an embodiment of the present invention. Detailed Implementation

[0022] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0023] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, are used only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0024] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0025] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0026] like Figures 1 to 3 As shown, a question-and-answer system for cement plant technical equipment based on a large language model includes: The front-end interaction module is used to receive cement equipment documents to be processed and input Q&A requests, and to display the processed Q&A results. The backend service module is connected to the frontend interaction module and is used to perform permission verification and format validation on the received question and answer requests, and to act as a data relay node between the modules to forward data. The knowledge base docking module is connected to the backend service module and is used to parse the cement equipment documents and store them in the vector knowledge base, as well as to perform context retrieval based on the question and answer request. The data storage module is used to store basic data of cement equipment, as well as context data retrieved by the knowledge base docking module; The large model interaction module is connected to the knowledge base interface module, data storage module and backend service module respectively. It is used to integrate the context data with the basic data of cement equipment, construct standardized prompt words for the cement industry, call the preset large language model to generate question and answer results, and feed them back to the backend service module.

[0027] This invention is deployed in the industrial intranet environment of a cement plant. Through the above-mentioned modular architecture, it integrates functions such as document parsing, vector retrieval, structured data retrieval, large language model interaction, and result display into a closed-loop system, realizing full-process automation from cement equipment knowledge management to intelligent question answering.

[0028] The following section, with reference to the accompanying drawings and specific examples, provides a detailed explanation of the implementation details of each functional module.

[0029] 1. Front-end interaction module The front-end interaction module is developed based on the Vue3 framework and uses the Element Plus component library to build the user interface. Specifically, the front-end interaction module includes a document upload submodule, a question-and-answer input submodule, and a result display submodule.

[0030] The document upload submodule supports receiving cement equipment documents in formats such as doc, pdf, and txt, and is configured to automatically verify the integrity of the document format and the file size before uploading. In this embodiment, the size of a single uploaded file is limited to no more than 100MB. Multi-part uploads are implemented using the axios library, and the upload progress is fed back to the front-end interface in real time via a ProgressEvent event object.

[0031] The question-and-answer input submodule provides a text input box component for users to input question-and-answer requests in natural language. In this embodiment, the question-and-answer input submodule also has a built-in keyword association function. Its association word library comes from the cement equipment keyword table preset in the data storage module. When the user inputs characters such as "rotary kiln support roller", it automatically suggests relevant professional terms such as "support roller temperature threshold" and "support roller grease replenishment" to improve the user's input efficiency and accuracy.

[0032] The results display submodule presents the responses returned by the large model interaction module in rich text format. During the display, key technical parameters and their units are highlighted. The results display submodule also provides export functionality and source document tracing capabilities. For example, users can click the document source identifier at the end of the response to directly jump to the online preview page of the corresponding document; they can also click the "Export Results" button to save the response as a docx file for offline viewing and archiving. The front-end interaction module adopts a responsive layout design, adapting to both personal computer terminals and industrial tablet terminals.

[0033] 2. Backend service module The backend service module, serving as the core control and relay node of the system, is deployed on a Linux industrial server and developed using the .NET 9 architecture. The backend service module includes a request receiving submodule, a data forwarding submodule, and an authorization verification submodule.

[0034] The request receiving submodule receives requests from the front-end interaction module via the HTTP / 2 protocol, using JSON as the data transmission format. This submodule integrates parameter validation logic, implemented using the FluentValidation library, to check the field names, data types, and numerical ranges contained in the request message. When parameter validation fails, the request receiving submodule generates a standardized JSON response containing an error code and description, directly returning it to the front-end interaction module to prevent invalid requests from entering subsequent processing.

[0035] The data forwarding submodule is used to implement asynchronous data interaction between modules. In this embodiment, the data forwarding submodule integrates a RabbitMQ message queue, publishing verified request messages to the message queue. Asynchronous forwarding using the message queue effectively avoids front-end interface lag caused by long subsequent processing times, improving the smoothness of user operation. Each piece of data forwarded via the message queue is assigned a globally unique UUID identifier, facilitating the tracking of the entire data flow process in the system logs. The system logs are segmented by date and stored in a designated directory on the Linux server.

[0036] The permission verification submodule is used to perform hierarchical permission verification for users based on user identity tokens, granting different system operation permissions according to different roles. Specifically, the permission verification submodule generates and issues a JWT-formatted user identity token when a user logs in. The token contains the user identifier and its corresponding role information. In this embodiment, three roles are preset: administrator, technician, and general employee. The administrator role is granted full operation permissions for document upload, deletion, modification, and user management; the technician role is granted operation permissions for document upload and question-and-answer retrieval; and the general employee role is only granted operation permissions for question-and-answer retrieval. Upon receiving a request, the permission verification submodule verifies the validity and permission scope of the user identity token carried in the request header, thereby ensuring the security of knowledge base data access.

[0037] 3. Knowledge Base Integration Module The knowledge base integration module, developed in Python, serves as a bridge between this system and an external vector knowledge base platform. This module includes a document parsing submodule and a context retrieval submodule, both of which utilize the RAGFlow platform V3.0 interface and have been customized and optimized for the specific characteristics of the cement industry.

[0038] (1) Document parsing submodule The document parsing submodule is used to parse cement equipment documents and convert them into structured data that can be retrieved using vector retrieval.

[0039] For unstructured text content in the document, the document parsing submodule uses a preset character length and a preset overlap rate for text segmentation; the preset character length is 490–510 characters, and the segment overlap rate is 8%–12%. In a preferred embodiment, the preset character length is specifically set to 500 characters, and the segment overlap rate is specifically set to 10%. This parameter combination was determined after analyzing the sentence structure and logical paragraph distribution characteristics of cement equipment documents: the 500-character length ensures that most single technical parameter descriptions or complete single-step fault handling logic are fully contained within the same segment, while the 10% overlap rate ensures that there is an overlap area of ​​approximately 50 characters between adjacent segments. This overlap area usually contains connecting terms, thus maintaining the semantic coherence of cross-segment content during the vector retrieval stage.

[0040] For tables in cement equipment documentation, the document parsing submodule implements a specialized processing workflow. Specifically, the document parsing submodule uses structured parsing technology to identify table boundaries and hierarchical relationships. For complex parameter tables or fault case tables containing multi-page or multi-level row and column structures, it first calls an external table extraction interface (RAGFlow platform version V3.0). This interface uses OCR and table structure recognition algorithms to automatically detect the boundaries of multi-page tables and align the table headers, while parsing out the parent-child hierarchical relationships between multi-level table headers. After parsing, the table content is converted into structured data in standardized JSON format, and the equipment model to which the table belongs and its page number range in the original document are recorded.

[0041] Subsequently, the document parsing submodule performs dimensional segmentation on the aforementioned JSON structured data. Unlike segmentation based on character count, this segmentation is based on equipment parameter categories or fault types. For example, a parameter table containing "rotary kiln support roller bearing model, lubrication cycle, temperature threshold, and vibration limit" can be divided into multiple independent segments such as "temperature parameters" and "vibration parameters"; a fault handling table containing "support roller high temperature fault" and "support roller abnormal noise fault" can be segmented according to the fault phenomenon type. Each segment retains its corresponding equipment category and parameter / fault dimension information.

[0042] Crucially, the document parsing submodule binds a unified table identifier to all fragments belonging to the same table. This table identifier is a globally unique identifier, stored as a metadata field along with the fragment vectors in the vector knowledge base. During subsequent retrievals, the system can use this table identifier to associate and retrieve fragments from different dimensions within the same original table, ensuring that the contextual information provided to the large language model fully covers the associated parameters or similar fault handling logic carried by the table.

[0043] In a further optimized embodiment, for multi-page or complex parameter tables, the document parsing submodule also extracts numerical parameters from the table as feature words, and merges these feature words with the table text vector before storing them in the database. For example, numerical values ​​such as "38.5℃" and "500mL" and their unit combinations are extracted as independent feature words. The word vectors corresponding to these feature words are then weighted and concatenated or summed with the sentence vectors of the table text paragraphs to generate a comprehensive vector representation that is finally stored in the vector knowledge base. This processing enhances the retrieval and recall capabilities for specific numerical parameters, enabling the system to directly locate table fragments containing specific numerical values ​​when users search using keywords such as "38.5" or "roller temperature threshold".

[0044] (2) Contextual Retrieval Submodule The context retrieval submodule is used to perform semantic retrieval in the vector knowledge base based on question-answering requests. Specifically, the context retrieval submodule performs retrieval based on a preset semantic similarity threshold and returns a preset number of document fragment contexts carrying document source identifiers in descending order of similarity; wherein the semantic similarity threshold is greater than or equal to 0.8.

[0045] In this embodiment, the similarity threshold is set to 0.8, and the number of results returned is set to 5. The context retrieval submodule first calls the vectorization interface (the retrieval interface of RAGFlow platform version V3.0) to convert the user-input question-and-answer request text into a query vector in the same vector space. Only fragments with a similarity greater than or equal to 0.8 are retained, and they are sorted in descending order of similarity, returning the first 5 results. Each returned document fragment carries the document source identifier corresponding to that fragment in its context. This mechanism ensures the traceability of the answer results and also enables the subsequent large model interaction module to obtain high-quality reference data.

[0046] By setting a high similarity threshold, this invention effectively filters out background noise information with low semantic relevance. In particular, for long-tail professional terms in the cement industry such as "grate cooler grate gap" and "vertical mill grinding pressure", it can ensure that the search results are highly focused on core technical knowledge, thereby improving the hit rate of professional term retrieval.

[0047] 4. Large Model Interaction Module The large model interaction module is developed based on Python and includes a data stitching submodule, a large model calling submodule, and a result processing submodule.

[0048] (1) Data splicing submodule The data splicing submodule is used to merge the retrieved context data and cement equipment basic data to construct standardized prompts for the cement industry.

[0049] First, the data splicing submodule receives the document fragment context returned by the knowledge base integration module, extracts the core text content of each fragment, and labels it with its corresponding document source identifier. Simultaneously, the data splicing submodule sends a request to the data storage module to retrieve pre-stored basic data on cement equipment related to the current question-and-answer request. This basic data is stored in a relational database and includes, but is not limited to, equipment model, rated operating parameters, alarm thresholds, lubrication specifications, and key component models. The data splicing submodule then organizes this basic data into a standardized parameter list, removing redundant fields irrelevant to the current question.

[0050] Secondly, based on the specialized characteristics of cement equipment, the data splicing submodule prioritizes fault thresholds and core technical parameters in the prompts. For example, when a user inquires about excessively high rotary kiln support roller temperature, the information "Support roller bearing temperature alarm threshold: 38.5℃" is placed at the top of the parameter list to draw the full attention of the large language model.

[0051] Finally, the data splicing submodule is filled using a pre-defined standardized prompt word template. This standardized prompt word template includes at least the following four core sections: Contextual basis: Fill in the contextual text of the document fragments sorted by similarity and source-annotated.

[0052] Equipment basic parameters: Fill in the standardized parameter list after it has been organized.

[0053] User Question: Enter the question and answer request text entered by the user as is.

[0054] Industry-specific answer requirements: Fill in the answer guidelines specifically for the cement equipment industry.

[0055] The industry-specific answer requirements include: for fault-related questions, the large language model must answer according to the logical steps of "shutdown, inspection, handling, and trial operation"; for parameter-related questions, accurate data citation and corresponding physical quantity units must be provided. In addition, answer requirements may also include constraints such as overall answer length limits and standardized use of technical terminology.

[0056] After the data filling is completed, the data splicing submodule performs format validation on the generated complete prompt words, including checking whether each section is complete and not empty, and whether the total character length exceeds the maximum input length limit of the large language model. If the limit is exceeded, the content in the context-based sections is trimmed in order of similarity from low to high, and a concise explanation is added. The standardized prompt words for the cement industry generated in this step will be used as direct input to the large language model.

[0057] (2) Large model calls sub-module The large model invocation submodule is used to invoke a preset large language model and generate corresponding question-and-answer results based on the standardized prompts for the cement industry. In this embodiment, the preset large language model is the Tongyi Thousand Questions and Thousand Inquiries version, which is used to interact with its provided API endpoint. Regarding the configuration of invocation parameters, the temperature parameter is set to a low value of temperature=0.3 to ensure the certainty, accuracy, and reproducibility of the answer results, avoiding creative deviations from the facts by the large language model; the maximum number of output tokens is set to max_tokens=2000 to adapt to the length of the prompts and the expected length of the answer; and the request timeout is set to timeout=10 seconds. To improve system robustness, the large model invocation submodule is also configured with an exception retry mechanism: when the call fails, it automatically retryes after a few seconds, with a maximum of one retry.

[0058] (3) Result Processing Submodule The results processing submodule receives the raw answer text returned by the large language model and formats it according to industry standards. Specifically, the results processing submodule removes redundant tags from the answer text and restructures it according to the logic set in the industry-specific answer requirements: for fault-related questions, steps are explicitly numbered with first-level headings such as "Shutdown," "Inspection," "Handling," and "Trial Run"; for parameter-related questions, all numerical values ​​are followed by their corresponding physical units, and the numerical values ​​and units are highlighted together. The processed question-and-answer results are encapsulated into a JSON response body and sent to the backend service module, where they are then visualized by the frontend interaction module.

[0059] 5. Data storage module The data storage module adopts a hybrid storage architecture, including a relational database submodule and a non-relational database submodule.

[0060] The relational database submodule is used to store basic data for cement equipment. In this embodiment, a MySQL database is used, deployed on a dedicated data server equipped with a large-capacity SSD. The database has a standardized table structure, such as: a basic equipment information table, containing fields such as equipment ID, equipment model, rated parameters, alarm threshold, lubrication information, etc.; a user information table, storing user login name, encrypted password hash value, role ID, etc.; and a question and answer history table, storing question text, answer text, cited document source identifier, timestamp, etc. Using a relational database ensures strong consistency of structured data and high efficiency of relational queries.

[0061] The non-relational database submodule serves as a caching layer, storing intermediate context data, temporary prompts, and user identity tokens corresponding to the question-and-answer requests. Different caching durations are set for each data type. In this embodiment, Redis is used as the caching database and deployed on a separate caching server. The specific caching strategy is as follows: The intermediate context data retrieved by the knowledge base integration module is cached using a specific key name format for a duration of one hour. Within one hour, if the system receives identical or highly similar question-and-answer requests, it can directly read the context data from the cache without needing to call the vector retrieval interface again.

[0062] Temporary hints generated by the data splicing submodule are cached using a specific key name format, with a cache duration of 30 minutes. This is primarily for debugging and auditing scenarios, and also facilitates quick reproduction of the answer to the same problem within a short period of time.

[0063] User identity tokens issued and verified by the authorization verification submodule are cached using a specific key name format, with the cache duration consistent with the token's expiration time, such as 2 hours. Cached token verification results avoid the need for database queries or encryption / decryption operations on every request.

[0064] By leveraging the collaborative work of relational and non-relational databases, this invention places frequently accessed hot data in a memory-level cache, significantly reducing the overhead of repeated calls to the underlying database and external knowledge base interfaces. Tests show that from the time a user initiates a question-and-answer request to the front-end interface rendering a complete answer, the average end-to-end response time can be controlled within 2.3 seconds, fully meeting the stringent real-time requirements of cement plant on-site operations and maintenance.

[0065] 6. Communication logic of each module The front-end interaction module initiates a document upload / Q&A request → the back-end service module verifies user permissions, performs format validation on the request, and then asynchronously forwards it via RabbitMQ → the knowledge base integration module completes document parsing and database entry / contextual precise retrieval, and the data storage module synchronously retrieves the corresponding cement equipment basic parameters → the large model interaction module concatenates the context data and basic parameters into cement industry-specific prompts, calls the Tongyi Thousand Questions large model to generate answers, and after formatting, feeds them back to the back-end service module → the back-end module forwards the final result to the front-end interaction module for display, completing the entire closed loop.

[0066] Based on the above system architecture, this invention also provides a question-and-answer method for cement plant technical equipment based on a large language model. The following detailed explanation of the method's steps is illustrated with a specific question-and-answer example—a technician inquiring about "how to handle the excessively high temperature (42℃) of the Φ4.8×74m rotary kiln support rollers."

[0067] S1: Receive user-inputted Q&A requests and / or pending cement equipment documents, and verify the identity and permissions and format validity of the Q&A requests.

[0068] First, users with administrator privileges select and upload a cement equipment document named "Φ4.8×74m Rotary Kiln Technical Manual.doc" through the document upload submodule of the front-end interaction module. The front-end interaction module validates the format and size of the uploaded file. After successful validation, the file is uploaded in chunks to the back-end service module.

[0069] The backend service module's permission verification submodule extracts the user's identity token from the request header, determines that the user has administrator privileges, and allows the document upload operation. The request receiving submodule validates the request parameters. After successful validation, the data forwarding submodule asynchronously forwards the document processing request to the knowledge base integration module via a message queue.

[0070] S2: Based on the user's identity and permissions, determine whether to parse and process the cement equipment document and store it in the vector knowledge base.

[0071] After receiving the document processing message, the document parsing submodule of the knowledge base integration module initiates the parsing and database entry process. The document parsing submodule segments the document text according to the aforementioned 500-character length and 10% overlap rate, while simultaneously identifying the "Rotary Kiln Support Roller Bearing Technical Parameter Table" and "Common Faults and Handling Measures Table" contained within the document. For these tables, the document parsing submodule calls the table extraction interface for structured parsing, converting it into JSON data, and then segmenting it dimensionally according to parameter categories and fault types, binding a unified table identifier to all segments originating from the same table. After parsing, the vectorization interface is called to convert each segment into vector data, which, along with metadata, is stored in the vector knowledge base. Upon successful database entry, the knowledge base integration module returns a success status code and the generated document ID to the backend service module, and finally, the frontend interaction module displays a "Successfully uploaded and parsed" message to the user.

[0072] S3: If the identity and permission verification and format validity verification pass, then perform semantic retrieval in the vector knowledge base according to the question and answer request to obtain context data, and simultaneously obtain the pre-stored cement equipment basic data.

[0073] Subsequently, a user with technical personnel privileges entered the question in the question-and-answer input submodule of the front-end interaction module: "How to deal with the excessively high temperature (42℃) of the Φ4.8×74m rotary kiln support roller?" The front-end interaction module then sent a question-and-answer request containing the question text and the user's identity token to the back-end service module.

[0074] After the backend service module's permission verification submodule verifies the technical personnel's permissions, it forwards the question-and-answer request to the knowledge base integration module. Simultaneously, it initiates a query to the relational database submodule of the data storage module to retrieve pre-stored basic cement equipment data related to the "Φ4.8×74m rotary kiln". This basic data includes: the temperature alarm threshold of the roller bearing, the recommended grease type, and the standard grease injection amount for the bearing cavity, etc.

[0075] After receiving a question-and-answer request, the context retrieval submodule of the knowledge base integration module converts the text into a query vector and performs a search in the vector knowledge base based on the conditions of a semantic similarity threshold ≥ 0.8 and a return quantity of 5. The retrieval process utilizes the vectors generated in the previous document parsing phase, which incorporate numerical parameter feature words. Ultimately, the context retrieval submodule returns 5 of the most relevant document fragment contexts, covering the cause analysis, inspection steps, and handling measures for the overheating fault of the roller. Each context is labeled with its source document ID. This intermediate context data is simultaneously cached in the non-relational database submodule for 1 hour.

[0076] S4: The acquired context data is integrated with the basic data of cement equipment to construct standardized prompts for the cement industry; a preset large language model is called to generate corresponding question and answer results based on the standardized prompts for the cement industry, and finally fed back to the front-end interaction module for visualization.

[0077] After receiving the context data and basic cement equipment data, the data splicing submodule of the large model interaction module begins constructing standardized prompts for the cement industry. It extracts the core text from the context fragments, prioritizes alarm thresholds as core technical parameters in the parameter list, and populates the list using a standardized prompt template. The industry-specific response requirements explicitly instruct the large language model to organize its response according to the logical steps of "shutdown, inspection, processing, and trial operation." After format validation, the spliced ​​prompts are passed to the large model calling submodule.

[0078] The large model call submodule calls the preset large language model API, passing in prompt words and preset parameters, and successfully obtains the original answer text generated by the large language model. The result processing submodule formats and organizes the text, clearly dividing the fault handling process into four numbered steps: "Shutdown → Inspection → Processing → Trial Run", and highlights the parameters and their units.

[0079] The compiled Q&A results are forwarded from the backend service module to the frontend interaction module. The results display submodule of the frontend interaction module displays the answer in a rich text area, along with a source document identifier. Users can click on the identifier to view the specific page number of the source document online using the integrated document preview component; users can also click the "Export Results" button to export the complete answer as a docx file for saving.

[0080] This completes a full-process closed loop execution, from document entry to intelligent question answering.

[0081] Example of output result: The following is a method for handling the excessively high temperature (42℃) of the Φ4.8×74m rotary kiln support roller in the knowledge vector base: a. When the temperature of the rotary kiln support roller exceeds 38.5℃, the machine must be stopped immediately and the bearing lubrication must be checked. If the grease is insufficient, 3# lithium-based grease must be added, and the amount added should be 1 / 2 to 2 / 3 of the bearing cavity volume. b. The rotary kiln support roller bearing model is 23160CA / W33, with a rated dynamic load of 1800kN and a rated static load of 2800kN; Based on the following basic equipment parameters (equipment model: Φ4.8×74m rotary kiln): Roller temperature threshold: 38.5℃ Bearing grease type: #3 lithium-based grease Bearing cavity volume: 500mL Answering the question from a cement industry professional's perspective: How to handle the excessively high temperature (42℃) of the support rollers in a Φ4.8×74m rotary kiln? Requirements: 1. For fault-related issues, specific handling steps must be provided; 2. For parameter-related issues, data must be accurately cited; 3. Use concise language and avoid redundant information.

[0082] a. Immediate shutdown: Because the current temperature (42℃) exceeds the threshold of 38.5℃, the rotary kiln must be stopped to avoid damage to the bearings; b. Check lubrication: Remove the end cap of the roller bearing and check the remaining amount of No. 3 lithium-based grease. If it is insufficient, add 1 / 2 to 2 / 3 of the bearing cavity volume (500mL). c. Bearing inspection: After replenishing grease, test run. If the temperature still exceeds the threshold, check if the bearing is worn (refer to the bearing's rated dynamic load of 1800kN and rated static load of 2800kN). If the wear exceeds the standard, the bearing needs to be replaced (model: 23160CA / W33).

[0083] In one specific embodiment of the present invention, the system deployed and configured as described above was tested in the industrial intranet environment of a cement plant. The test selected 100 sets of question-and-answer requests covering typical scenarios such as parameter querying, fault diagnosis, and maintenance consultation. The test results are shown in Table 1. Table 1 Test Results

[0084] Test data shows that the system and method provided by this invention have fully achieved the expected design goals. Compared with existing general industrial question-and-answer systems, they have significantly improved in terms of industry adaptability, retrieval accuracy, real-time response, and professionalism of answers.

[0085] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A question-and-answer system for cement plant technical equipment based on a large language model, characterized in that, include: The front-end interaction module is used to receive cement equipment documents to be processed and input Q&A requests, and to display the processed Q&A results. The backend service module is connected to the frontend interaction module and is used to perform permission verification and format validation on the received question and answer requests, and to act as a data relay node between the modules to forward data. The knowledge base docking module is connected to the backend service module and is used to parse the cement equipment documents and store them in the vector knowledge base, as well as to perform context retrieval based on the question and answer request. The data storage module is used to store basic data of cement equipment, as well as context data retrieved by the knowledge base docking module; The large model interaction module is connected to the knowledge base interface module, data storage module and backend service module respectively. It is used to integrate the context data with the basic data of cement equipment, construct standardized prompt words for the cement industry, call the preset large language model to generate question and answer results, and feed them back to the backend service module.

2. The cement plant technical equipment question-and-answer system based on a large language model according to claim 1, characterized in that: The knowledge base docking module includes a document parsing submodule. When processing the cement equipment document, the document parsing submodule uses a preset character length and a preset overlap rate to perform text segmentation. The preset character length is 490 to 510 characters, and the preset overlap rate is 8% to 12%.

3. The cement plant technical equipment question-and-answer system based on a large language model according to claim 2, characterized in that: For the table content in the cement equipment document, the document parsing submodule identifies the table boundaries and hierarchical relationships through structured parsing technology, converts the table content into structured data, and performs dimensional segmentation according to equipment parameter categories or fault types. At the same time, it binds a unified table identifier to all segments belonging to the same table.

4. The cement plant technical equipment question-and-answer system based on a large language model according to claim 3, characterized in that: For multi-page or complex parameter tables, extract the numerical parameters in the table as feature words, and then fuse the feature words with the table text vector before storing them in the database.

5. The cement plant technical equipment question-and-answer system based on a large language model according to claim 1, characterized in that: The knowledge base docking module also includes a context retrieval submodule, which performs retrieval based on a preset semantic similarity threshold and returns a preset number of document fragment contexts carrying document source identifiers in descending order of similarity. The semantic similarity threshold is greater than or equal to 0.

8.

6. The cement plant technical equipment question-and-answer system based on a large language model according to claim 1, characterized in that: The large model interaction module includes a data splicing submodule, which is used to standardize the retrieved context data and cement equipment basic data, and prioritize the fault threshold and core technical parameters in the prompt words according to the professional characteristics of cement equipment. The standardized prompt template should include at least: contextual basis, basic equipment parameters, user questions, and industry-specific answer requirements.

7. The cement plant technical equipment question-and-answer system based on a large language model according to claim 6, characterized in that: The industry-specific answer requirements include: for fault-related questions, the large language model should answer according to the logical steps of "shutdown, inspection, handling, and trial operation"; for parameter-related questions, accurate data should be cited and the corresponding physical quantity units should be labeled.

8. The cement plant technical equipment question-and-answer system based on a large language model according to claim 1, characterized in that: The data storage module includes a relational database submodule and a non-relational database submodule; The relational database submodule is used to store basic data of cement equipment; The non-relational database submodule serves as a caching layer, used to store intermediate context data, temporary prompts, and user identity tokens corresponding to the question-and-answer request, with each type of data having a different caching duration.

9. The cement plant technical equipment question-and-answer system based on a large language model according to claim 1, characterized in that: The backend service module includes an access control submodule, which is used to perform hierarchical access control verification on users based on user identity tokens, so as to grant different system operation permissions according to different roles.

10. A question-answering method for cement plant technical equipment based on a large language model, characterized in that, The application of the cement plant technical equipment question-answering system based on a large language model as described in any one of claims 1-9 includes the following steps: Receive user-inputted Q&A requests and / or pending cement equipment documents, and verify the identity and permissions and format validity of the Q&A requests; Based on the user's identity and permissions, determine whether to parse and process the cement equipment document and store it in the vector knowledge base; If the identity and permission verification and format validity verification pass, then semantic retrieval is performed in the vector knowledge base according to the question and answer request to obtain context data, and pre-stored cement equipment basic data is obtained simultaneously. The acquired contextual data is integrated with the basic data of cement equipment to construct standardized prompts for the cement industry; a preset large language model is invoked to generate corresponding question-and-answer results based on the standardized prompts for the cement industry, and finally fed back to the front-end interactive module for visualization.

Images