A method and device for intelligent decision-making of large model inference and silt coal separation and recovery

Abstract

The application discloses a kind of big model reasoning's stone coal separation and recovery intelligent decision-making method and device, belong to industrial intelligent decision-making field.Through the construction of the field knowledge base containing vector database and relational database;Using low rank adaptation technology fine-tuning general large language model, obtain domain-specific model;When reasoning, receive question and working condition data, retrieve relevant text fragments from knowledge base and combine into enhanced prompts, input domain-specific model to generate decision text containing parameter adjustment suggestions;Safety filtering and boundary clamping are carried out on the decision text, and the structured decision suggestion is output;Record the adoption feedback and operation effect, automatically filter high-quality cases to supplement to the knowledge base, and trigger model incremental fine-tuning.The application solves the problem of low efficiency of artificial experience decision and difficulty in continuous accumulation of knowledge, realizes intelligent decision and model dynamic optimization.

Description

Technical Field

[0001] This invention relates to the field of industrial intelligent decision-making, and in particular to an intelligent decision-making method and apparatus for stone and coal sorting and recycling based on large-scale model reasoning. Background Technology

[0002] In the operation of a coal and stone sorting and recovery system, when abnormal operating conditions such as abnormal fan speed, valve blockage, or fluctuations in conveying pressure occur, or when external conditions such as sudden changes in coal quality or load adjustments occur, front-line operators find it difficult to quickly make optimal parameter adjustment decisions. Existing technologies mainly offer the following solutions: First, traditional manual operation relies entirely on human experience, resulting in inconsistent decision-making quality and poor response timeliness; second, rule-based engine-based decision support systems can only handle limited scenarios within preset thresholds and cannot cope with complex and changing conditions; third, remote telephone guidance from experts has long response times and limited expert resources; fourth, general industrial knowledge base systems require manual retrieval and cannot provide personalized decision-making suggestions based on real-time operating conditions. These solutions collectively lead to a timeframe of more than 15 minutes from anomaly detection to operational execution, and also suffer from problems such as reliance on individual operational experience, difficulty in knowledge transfer, poor decision consistency, and a lack of automatic verification of safety boundaries. Regarding artificial intelligence applications, while large language models have been applied in fields such as medicine and law, they are currently lacking in the specific process scenario of coal and stone sorting and recovery in thermal power plants. General-purpose models lack expertise in this area and cannot be directly used for process decision support. Summary of the Invention

[0003] The present invention aims to at least partially solve one of the technical problems in the related art.

[0004] To address this issue, this invention proposes an intelligent decision-making method for stone and coal sorting and recycling based on large-scale model reasoning. This method constructs a domain knowledge base containing vector and relational databases by acquiring unstructured documents and structured operational data. Low-rank adaptation technology is used to fine-tune a general-purpose large-scale language model, resulting in a domain-specific model. During reasoning, questions from operators and real-time operational data are received, and relevant text fragments are retrieved from the knowledge base and combined to create enhanced prompts. These prompts are then input into the domain-specific model to generate decision text containing parameter adjustment suggestions. The decision text undergoes security filtering and boundary constraints, outputting structured decision suggestions. Adoption feedback and operational results are recorded, and high-quality cases are automatically selected to supplement the knowledge base, triggering incremental model fine-tuning. This invention solves the problems of low efficiency and difficulty in continuous knowledge accumulation associated with relying on human experience-based decision-making, achieving intelligent decision-making and dynamic model optimization.

[0005] Another objective of this invention is to propose an intelligent decision-making device for the sorting and recycling of stones and coal based on large-scale model reasoning.

[0006] To achieve the above objectives, this invention proposes a large-scale model-based intelligent decision-making method for the separation and recovery of stone and coal, comprising:

[0007] Unstructured documents and structured operational data in the field of stone and coal sorting and recycling are obtained. The unstructured documents are vectorized and stored in a vector database, and the structured operational data is stored in a relational database to construct a domain knowledge base. Low-rank adaptation technology is used to efficiently fine-tune the parameters of a general large language model to obtain a domain-specific model. During inference, natural language questions from operators and real-time operating data are received. Operation procedure texts and historical case fragments related to the questions are retrieved from the vector database. The retrieved text fragments, real-time operating data and questions are combined to form enhanced prompts, which are then input into the domain-specific model to generate decision text containing parameter adjustment suggestions. The decision text is subjected to security filtering and structured output. The suggested parameter values ​​in the decision text are extracted and compared with the preset multi-level security boundary table. Parameter values ​​that exceed the hard boundary are forcibly clamped to boundary values ​​and a warning is added. The output is a structured decision suggestion containing operation steps and the basis for the suggestion. Record the operators' feedback on the adoption of the structured decision suggestions and the subsequent operational effects, and generate an interaction log; automatically select high-quality cases from the interaction log according to preset quality conditions, and after review, add high-quality cases to the domain knowledge base, and trigger incremental fine-tuning updates of the domain-specific model based on the accumulated feedback data.

[0008] The intelligent decision-making method for stone and coal sorting and recycling based on large-model inference in an embodiment of the present invention may also have the following additional technical features: In one embodiment of the present invention, the step of acquiring unstructured documents and structured operational data in the field of stone and coal sorting and recycling, storing the unstructured documents in a vector database after vectorization, and storing the structured operational data in a relational database to construct a domain knowledge base includes: The unstructured document is segmented, and each text segment is converted into a dense vector using a Chinese semantic embedding model. The dense vectors are then stored in a vector database using the HNSW algorithm. Structured operational data, including timestamps, equipment numbers, process parameter values, and operating condition labels, is stored in a relational database.

[0009] In one embodiment of the present invention, vectorizing an unstructured document includes: A sliding window strategy is used to segment unstructured documents, with a window size of 512 tokens and a step size of 256 tokens. For key clauses in the safety operating procedures, overlapping segments with a step size smaller than the window size are used to preserve semantic coherence.

[0010] In one embodiment of the present invention, efficient parameter fine-tuning includes: Using LoRA technology, the weight update of the general large language model is defined as... ,in For frozen pre-trained weights, and It is a learnable low-rank matrix; the data used during training includes question-and-answer pairs constructed from operating procedures, historical fault handling cases, and condition-parameter adjustment suggestion pairs.

[0011] In one embodiment of the present invention, the retrieval employs a hybrid retrieval strategy, including: The semantic score is obtained by calculating the cosine similarity between the user query vector and the document vector. The keyword matching score is calculated, and the two are weighted and summed to obtain the final ranking score. The top K text fragments containing operation procedure fragments and historical case fragments are returned in descending order according to the final ranking score.

[0012] In one embodiment of the present invention, the cosine similarity calculation formula is as follows: The filtering threshold is set to 0.7; the weighted summation formula is... ,in If K is 5, it returns 2 operation procedure fragments and 3 historical case fragments.

[0013] In one embodiment of the present invention, security filtering includes: Set hard lower limits for each process parameter and hard limit For any extracted suggested parameter value Output clamp value ;like If the value exceeds the recommended range but does not reach the hard boundary, a warning mark is generated without forcing a change; if the hard boundary is reached, the value is forcibly corrected to the boundary value, a warning mark is added, and the original recommended value is retained.

[0014] In one embodiment of the present invention, the structured decision suggestion includes: current working condition judgment, recommended parameter combination, step-by-step operation steps, and suggestion basis; wherein, the suggestion basis is referenced from the retrieved operation procedure clauses or historical case numbers, and includes a description of the parameter derivation logic.

[0015] In one embodiment of the present invention, the preset quality condition is: Define adoption flag The indicator that the operating parameters are within the normal range for two consecutive hours after adoption. A system confidence level greater than 0.8 is a key indicator. The formula for judging case quality is: If and only if When the corresponding interaction record is determined to be a high-quality case, it is submitted to the review queue and updated after expert confirmation.

[0016] To achieve the above objectives, another aspect of the present invention proposes an intelligent decision-making device for stone and coal sorting and recycling based on large-scale model reasoning, comprising: The knowledge base construction module is used to acquire unstructured documents and structured operational data in the field of stone and coal sorting and recycling. The unstructured documents are vectorized and stored in a vector database, and the structured operational data is stored in a relational database to construct a domain knowledge base. The intelligent reasoning module is used to efficiently fine-tune the parameters of a general large language model using low-rank adaptation technology to obtain a domain-specific model. During reasoning, it receives natural language questions from operators and real-time operating data, retrieves relevant operating procedure texts and historical case fragments from the vector database, combines the retrieved text fragments, real-time operating data and questions into enhanced prompts, inputs them into the domain-specific model, and generates decision text containing parameter adjustment suggestions. The security filtering and output module is used to perform security filtering and structured output on the decision text, extract the suggested parameter values ​​in the decision text, compare them with the preset multi-level security boundary table, forcibly clamp the parameter values ​​that exceed the hard boundary to the boundary value and add a warning, and output structured decision suggestions containing operation steps and the basis for the suggestions. The co-evolution module is used to record the operators' adoption feedback on the structured decision suggestions and the subsequent operational effects, and generate interaction logs. High-quality cases are automatically selected from the interaction logs according to preset quality conditions. After review, high-quality cases are added to the domain knowledge base, and incremental fine-tuning updates of the domain-specific model are triggered based on the accumulated feedback data.

[0017] This invention discloses an intelligent decision-making method and apparatus for stone and coal sorting and recovery based on large-model reasoning. By constructing a domain knowledge base containing vector and relational databases, and employing low-rank adaptation technology to fine-tune a general-purpose large language model to obtain a domain-specific model, relevant text fragments are retrieved and combined with real-time operating conditions during reasoning to generate enhanced prompts for parameter adjustment suggestions. Furthermore, the decision text is subjected to security filtering and boundary constraints. This effectively solves the problems of low decision-making efficiency and difficulty in continuous knowledge accumulation caused by reliance on human experience in existing technologies. It achieves integrated decision-making throughout the entire process, from knowledge acquisition, intelligent reasoning, safe output to feedback evolution, significantly improving the accuracy and operational safety of stone and coal sorting and recovery decisions, and enhancing the method's adaptability and engineering practical value under complex operating conditions.

[0018] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0019] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart of a large-model reasoning-based intelligent decision-making method for stone and coal sorting and recycling according to an embodiment of the present invention; Figure 2 This is a flowchart of a method according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of an intelligent decision-making device for large-scale model reasoning in stone and coal sorting and recycling according to an embodiment of the present invention. Detailed Implementation

[0020] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0021] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0022] The following description, with reference to the accompanying drawings, describes an intelligent decision-making method and apparatus for large-scale model reasoning in the sorting and recycling of gravel and coal according to an embodiment of the present invention.

[0023] The core idea of ​​this invention is to address the problems of low efficiency and difficulty in continuous knowledge accumulation in the process of stone and coal sorting and recycling by relying on manual experience-based decision-making. It solves the shortcomings of traditional experience-based decision-making, such as insufficient accuracy and lack of self-optimization, by constructing an intelligent decision-making method covering the entire process of "knowledge base construction—model fine-tuning—enhanced reasoning—safe output—feedback evolution." First, unstructured documents and structured operational data are acquired to construct a domain knowledge base containing vector and relational databases. Based on this, a low-rank adaptation technique is used to fine-tune a general-purpose large language model to obtain a domain-specific model. For the reasoning stage, questions from operators and real-time operating conditions are received, relevant text fragments are retrieved and combined into enhanced prompts, and input into the domain-specific model to generate decision text containing parameter adjustment suggestions. Then, the decision text is subjected to safe filtering and boundary constraints to output structured decision suggestions. Finally, the adoption feedback and operational effects are recorded, high-quality cases are automatically selected to supplement the knowledge base, and incremental model fine-tuning is triggered. By integrating the above-mentioned technical path of "knowledge modeling, intelligent reasoning, security constraints, and co-evolution", this invention transforms the traditional experience-based decision-making that relies on human experience into an intelligent decision-making system that can continuously learn, is safe and controllable, and adapts to complex working conditions, significantly improving the accuracy, safety, and engineering practical value of stone and coal sorting and recycling decisions.

[0024] The following description, with reference to the accompanying drawings, describes an intelligent decision-making method and apparatus for large-scale model reasoning in the sorting and recycling of gravel and coal according to an embodiment of the present invention.

[0025] like Figure 1 As shown, the intelligent decision-making method for stone and coal sorting and recycling based on large-scale model reasoning of the present invention includes the following steps: S1. Obtain unstructured documents and structured operational data in the field of stone and coal sorting and recycling. After vectorization processing, store the unstructured documents in a vector database and store the structured operational data in a relational database to construct a domain knowledge base.

[0026] In one embodiment of the present invention, the domain knowledge base serves as the foundational data support for the intelligent decision-making system. It requires the systematic collection and organization of various knowledge resources within the field of stone and coal sorting and recovery. The construction of the domain knowledge base includes: operating procedure documents: standard operating procedures, safe operating procedures, start-up and shutdown procedures, etc., for the stone and coal discharge system; equipment operation manuals: technical parameters, operating instructions, and maintenance requirements for various equipment (fans, valves, conveying pipelines, sorting devices, etc.); historical accident cases and handling records: various abnormal operating conditions, fault handling processes, and experience summaries since commissioning; expert interview experience: tacit knowledge and practical experience obtained through interviews with senior operators and equipment experts; and historical operating data: historical operating data since commissioning and its corresponding operating condition annotations.

[0027] Furthermore, differentiated storage strategies are adopted for different types of knowledge resources: Unstructured document processing workflow: Unstructured documents such as operating procedures, equipment manuals, historical cases, and expert experience are preprocessed with text, including noise reduction, segmentation, and format standardization. Then, the text is converted into a high-dimensional vector representation through an embedding model and stored in a vector database, supporting semantic similarity retrieval; Structured data processing workflow: Historical operating data (such as process parameters such as temperature, pressure, flow rate, and speed, and their corresponding operating condition labels) are stored in a relational database, supporting precise querying and statistical analysis.

[0028] S2 employs low-rank adaptation technology to efficiently fine-tune the parameters of a general-purpose large language model, obtaining a domain-specific model. During inference, it receives natural language questions from operators and real-time operating data, retrieves relevant operating procedure texts and historical case fragments from the vector database, combines the retrieved text fragments, real-time operating data, and questions into enhanced prompts, inputs them into the domain-specific model, and generates decision text containing parameter adjustment suggestions.

[0029] In one embodiment of the present invention, a domain-specific intelligent agent is the core technical module, aiming to enable a general-purpose large language model to understand and process specialized problems in the field of stone and coal sorting and recovery. The construction of the domain-specific intelligent agent includes: Base model selection: using an open-source large language model with at least 70 bytes of parameters (e.g., ChatGLM, Qwen, etc.) as the base model. Models with a high parameter count have stronger reasoning and knowledge understanding capabilities, and can better handle complex process decision-making problems; Efficient parameter fine-tuning technology: using Low-Rank Adaptation (LoRA) technology to adapt the base model to the domain. The core idea of ​​LoRA technology is to capture domain-specific knowledge by adding a small number of learnable low-rank matrices while keeping the pre-trained model parameters unchanged. Compared to full parameter fine-tuning, LoRA only needs to update about 0.1% of the model's parameters to achieve good domain adaptation, significantly reducing computational resource consumption and overfitting risk, while maintaining the model's original capabilities.

[0030] Furthermore, for the pre-trained weight matrix The weight update of LoRA can be formally represented as:

[0031] in , ,rank , This is the scaling factor. During training, only A and B are optimized, while maintaining... The freeze remains unchanged.

[0032] Retrieval Enhancement Generation (RAG) Mechanism: Deploying a retrieval enhancement generation mechanism during the inference phase improves the accuracy and timeliness of responses. The specific process is as follows: Whenever an operator raises a question, the system first converts the question into a query vector, then retrieves the most relevant operational procedures and historical case fragments from the vector database. The search results are then concatenated with real-time DCS data to create enhanced prompts, which are finally fed into a finely tuned large model to generate the response. Through the RAG mechanism, the system can combine the latest real-time data with historical experience to provide more accurate and targeted decision-making suggestions.

[0033] S3, perform security filtering and structured output on the decision text, extract the suggested parameter values ​​in the decision text, compare them with the preset multi-level security boundary table, force the parameter values ​​that exceed the hard boundary to be bounded as boundary values ​​and add warnings, and output structured decision suggestions that include operation steps and the basis for the suggestions.

[0034] In one embodiment of the present invention, to ensure the safety and reliability of the decision recommendations output by the intelligent agent, the present invention designs a strict safety constraint layer and an interpretable output mechanism. The safety constraints and interpretable output include: a safety rule filter: before the reasoning opinions generated by the large model are output to the operators, they must be reviewed by the safety rule filter. The safety rule filter compares the recommended parameters (such as fan speed, valve opening, temperature, pressure, etc.) in the generated content with preset safety boundaries item by item, automatically intercepting and correcting any recommendations that exceed the safety range. The safety boundaries are divided into at least a recommended range and a hard upper / hard lower limit according to severity, with the hard boundary being an absolutely insurmountable physical safety limit. Let the recommended parameter value be x, and its hard lower limit be... The hard limit is Then output the clamp value. for:

[0035] If x is only outside the recommended range but does not touch the hard boundary, an orange alert is generated but no modification is forced; if it touches the hard boundary, it is automatically corrected to the boundary value and a warning mark is added, while the original recommended value is retained for expert reference.

[0036] Explainable output mechanism: To achieve traceable and explainable human-machine collaborative decision-making, this invention requires that the output content must include a "basis for recommendation" paragraph, which fully explains the source of the recommendation—whether it is cited from a certain operating procedure clause or analogous to the handling experience of a certain historical case, to ensure that operators can understand the logical basis behind the recommendation and enhance their confidence in decision-making.

[0037] Standardized Output Format: The system's output decision recommendations adopt a unified structured format, containing four core parts: a. Current Operating Condition Judgment: A description of the operating condition based on real-time data analysis; b. Recommended Parameter Combinations: The parameters to be adjusted and their recommended values ​​are presented in tabular form; c. Operation Steps: Specific step-by-step operation instructions, clarifying the operation sequence and precautions; d. Basis for Recommendations: A detailed explanation of the source of the recommendations and a list of supporting evidence.

[0038] S4. Record the operators' adoption feedback on the structured decision suggestions and the subsequent operational effects, and generate an interaction log; automatically select high-quality cases from the interaction log according to preset quality conditions, and after review, add high-quality cases to the domain knowledge base, and trigger incremental fine-tuning updates of the domain-specific model based on the accumulated feedback data.

[0039] In one embodiment of the present invention, to achieve continuous system optimization and self-evolution, the present invention designs a comprehensive human-machine collaborative interaction mechanism and a knowledge closed-loop update process. Human-machine collaboration and knowledge closed-loop include: Human-machine collaborative interaction: Operators have the final decision-making power over the suggestions of the intelligent agent, and can choose to adopt, modify, or reject them. A complete record of each interaction—including the original question, system suggestion, the operator's final operation, and subsequent feedback on the operational effect—is stored in the interaction log database with timestamps, forming a complete decision-making trajectory traceability chain; Knowledge base self-evolution mechanism: The system automatically extracts high-quality cases from the interaction logs periodically. Represented by Boolean variables: A = Adoption flag (operator clicks "Adopt"), B = Operating parameters are within the preset normal range for two consecutive hours thereafter, C = System output confidence score > 0.8, then the quality judgment of the case is: If and only if When the value is 1, the interaction record is considered a high-quality case and automatically enters the review queue. After review and confirmation by domain experts, these cases are automatically added to the knowledge base, expanding the system's knowledge coverage. For cases that are rejected or adopted but have poor results, the system automatically records them as "to be optimized" feedback for subsequent model iteration and optimization; continuous model optimization: based on the accumulated feedback data, regular model fine-tuning updates are performed, enabling the model to continuously learn from practical experience and constantly improve the accuracy and practicality of decision-making suggestions.

[0040] In summary, the intelligent decision-making method for stone and coal sorting and recovery proposed in this invention utilizes a large-scale model inference. Its core is to construct a dedicated intelligent agent for the field of stone and coal sorting and recovery by combining a general large-scale language model with efficient parameter fine-tuning, retrieval-enhanced generation (RAG), and a security constraint layer. The system input consists of natural language questions from operators and real-time DCS data context; the output includes directly executable parameter adjustment suggestions and step-by-step operation guidance.

[0041] Furthermore, such as Figure 2As shown, the specific steps implemented by the method of the embodiment of the present invention are as follows: 1. Specific construction process of the domain knowledge base.

[0042] Document preprocessing and segmentation strategy: Standardize the collected unstructured documents (operating procedures, equipment manuals, historical cases, etc.), including: unifying the text encoding format, removing irrelevant noise content (such as headers, footers, watermarks, etc.), standardizing term expressions, etc. The sliding window strategy is adopted for text segmentation, with the window size set to 512 tokens and the step size set to 256 tokens. Let the total text length be L (counted by token), the window size W=512, and the step size S=256, then the i-th text segment is:

[0043] For important clauses (such as key provisions of safety operating procedures), overlapping segmentation (i.e., S<W) is adopted to preserve the semantic coherence between paragraphs and ensure retrieval completeness.

[0044] Embedding model selection: It is recommended to adopt Chinese semantic embedding models such as text2vec-base-chinese or m3e-base to convert text into 768-dimensional or 1024-dimensional dense vectors. This model performs excellently on Chinese semantic understanding tasks and can accurately capture the professional terms and business semantics in the stone coal field.

[0045] Vector database selection: Adopt open-source vector databases such as Milvus or Chroma to support millisecond-level similarity retrieval. The database is configured with separated primary replicas and read-only replicas to ensure retrieval performance and availability. The vector index adopts the HNSW (Hierarchical Navigable Small World) algorithm to realize efficient approximate nearest neighbor search while ensuring retrieval accuracy.

[0046] Structured data storage: Historical operating data is stored in a MySQL relational database, and the table structure includes fields such as timestamp, equipment number, process parameter name, parameter value, working condition label, and remarks. The data retention period is set to 3 years, supporting multi-dimensional queries by time range, equipment type, working condition type, etc.

[0047] 2. LoRA parameter efficient fine-tuning.

[0048] The core parameter settings of LoRA fine-tuning are shown in Table 1: Table 1

[0049] The weight update follows the aforementioned formula , during training, only the gradients of B and A are calculated, and the original weights remain unchanged.

[0050] Training Data Construction: Construct a domain-specific fine-tuning training dataset, including the following types: Question-Answer Pair Dataset: Extract key clauses from the operating procedures to construct "question-procedure clause" question-answer pairs, training the model's ability to understand and reference the procedures; Case Dataset: Organize historical fault handling cases into structured text, training the model to learn problem diagnosis and solutions from the cases; Parameter Inference Dataset: Construct "operating condition description-parameter adjustment suggestion" data pairs, training the model's ability to infer the optimal parameters based on the operating conditions.

[0051] 3. Implementation of Retrieval Enhancement Generation (RAG).

[0052] Search Strategy: A hybrid search strategy is employed, combining vector semantic search and keyword precision search. Semantic search captures the question intent, while keyword search ensures a match between proper nouns and precise parameter values.

[0053] Similarity calculation: Semantic retrieval uses cosine similarity to calculate the query vector q and document vector d. i The similarity is calculated using the following formula:

[0054] The threshold is set to 0.7; search results below this threshold will be filtered.

[0055] Hybrid score fusion: The final ranking score is composed of semantic similarity. Keyword matching score Weighted summation yields:

[0056] The keyword matching score can be a normalized value of TF-IDF or BM25, with λ=0.7 recommended to emphasize the role of semantic similarity in decision-making.

[0057] Top-K selection: Each search is sorted in descending order of fusion score, and the top-K (K=5) most relevant document fragments are returned, including 2 operational procedure fragments and 3 historical case fragments, to ensure that the answer has sufficient reference basis.

[0058] Prompt word template design: The enhanced prompt word template includes the following components: system role settings, DCS real-time data context, retrieved relevant procedural clauses, retrieved historical cases, and system output format requirements. Template-based design ensures consistency and controllability of the output format.

[0059] 4. Specific implementation of security rule filters.

[0060] Safety boundary parameter table: Establish a multi-level safety boundary table containing the following parameters, at least distinguishing between recommended ranges and hard boundaries, as shown in Table 2: Table 2

[0061] Comparison and Correction Algorithm: The security filter extracts all numerical parameters from the generated text, uses regular expressions to match parameter name-value pairs, and compares them item by item with the security boundary table. For each extracted suggested value x, the hard-boundary-constrained output value is calculated using the following formula:

[0062] If the original suggested value x is within the recommended range, it will be output as is; if it exceeds the recommended range but does not reach the hard boundary, the system will retain the original suggested value but add an orange warning mark; if it reaches the hard boundary, x′ will be forcibly clamped to the nearest boundary value, and the reason for the correction will be noted in the output. At the same time, the original suggested value will be retained for expert reference to ensure that valuable information is not lost due to automatic correction.

[0063] 5. Interpretable output.

[0064] The generation of the reference paragraph employs a citation mechanism: the system records document fragments and their sources (document titles, chapter numbers, etc.) encountered during the RAG search process, and automatically inserts citation marks when generating the response. The generated recommendation paragraph structure includes: a working condition matching description (indicating the similarities between the current working condition and the reference case / procedure), parameter derivation logic (explaining the reasoning process from historical experience / procedure clauses to specific recommended parameters), and a source list (completely listing the procedure clause numbers and case numbers cited in this recommendation).

[0065] 6. Design of human-computer collaborative interaction interface.

[0066] The interactive interface adopts a dialog box format and mainly includes: a problem input area (supporting natural language description of the current problem), a DCS data display area (displaying key parameters related to the problem in real time), an intelligent suggestion display area (presenting the system's decision suggestions in a structured format), an operation feedback area (providing "Accept", "Accept after modification", and "Reject" option buttons), and a history interaction area (displaying the complete interaction record of the current session).

[0067] 7. The specific process of case extraction, review, and storage in the knowledge closed loop.

[0068] The system automatically performs a case extraction task every day at midnight, filtering cases marked as "adopted" from the interaction logs within the past 24 hours. The filtering uses the quality judgment formula defined above. In other words, the following conditions must be met simultaneously: the operator clicked the "Accept" button (A=1), the operating data returned to the normal range for two consecutive hours thereafter (B=1), and the system confidence score was greater than 0.8 (C=1). The selected cases will automatically enter the waiting queue for review.

[0069] Domain experts can view cases awaiting review through the review interface and choose to "Add to Knowledge Base" (confirm case quality and add to the knowledge base), "Pending Optimization" (mark as a case that needs improvement), or "Discard" (consider the case to be of no reference value). Cases that pass review are automatically updated with the historical case index in the vector database and the case tag table in the relational database.

[0070] 8. Triggering conditions and execution methods for model iteration updates.

[0071] The model iteration and update employs a dual-trigger mechanism: Regular triggering: A model fine-tuning update is triggered after every 500 newly added, approved, high-quality cases are added to the database; Anomaly triggering: When the system detects that the adoption rate for a certain scenario has continuously decreased by more than 20%, it determines that the model's ability to handle that scenario is insufficient and immediately triggers targeted fine-tuning. Model fine-tuning uses an incremental training strategy, loading new training data onto the previous version of the model to avoid forgetting existing capabilities due to training from scratch. After fine-tuning, the effect is verified on the test set before deployment.

[0072] The method of this invention effectively solves the problems of low efficiency and difficulty in continuous knowledge accumulation when relying on human experience for decision-making. It realizes integrated decision-making throughout the entire process from knowledge construction, intelligent reasoning, safe output to feedback evolution, and significantly improves the accuracy and engineering practicality of stone and coal sorting and recycling decisions.

[0073] To achieve the above invention, such as Figure 3 As shown, this embodiment also provides a large-scale model inference-based intelligent decision-making device 10 for stone and coal sorting and recycling. The device 10 includes: The knowledge base construction module 100 is used to acquire unstructured documents and structured operational data in the field of stone and coal sorting and recycling, store the unstructured documents in a vector database after vectorization, and store the structured operational data in a relational database to construct a domain knowledge base.

[0074] The intelligent reasoning module 200 is used to efficiently fine-tune the parameters of a general-purpose large language model using low-rank adaptation technology to obtain a domain-specific model. During reasoning, it receives natural language questions from operators and real-time operating data, retrieves relevant operating procedure texts and historical case fragments from the vector database, combines the retrieved text fragments, real-time operating data, and questions into enhanced prompts, inputs them into the domain-specific model, and generates decision text containing parameter adjustment suggestions.

[0075] The security filtering and output module 300 is used to perform security filtering and structured output on the decision text, extract the suggested parameter values ​​in the decision text, compare them with the preset multi-level security boundary table, forcibly clamp the parameter values ​​that exceed the hard boundary to the boundary value and add a warning, and output structured decision suggestions containing operation steps and the basis for the suggestions.

[0076] The co-evolution module 400 is used to record the feedback from operators on the adoption of the structured decision suggestions and the subsequent operating effects, and generate an interaction log. High-quality cases are automatically selected from the interaction log according to preset quality conditions. After review, the high-quality cases are added to the domain knowledge base, and the incremental fine-tuning and updating of the domain-specific model is triggered based on the accumulated feedback data.

[0077] The device in this invention effectively solves the problems of low efficiency and difficulty in continuous knowledge accumulation when relying on human experience for decision-making. It realizes integrated decision-making throughout the entire process from knowledge construction, intelligent reasoning, safe output to feedback evolution, and significantly improves the accuracy and engineering practicality of stone and coal sorting and recycling decisions.

[0078] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0079] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

Claims

1. A large-scale model-based intelligent decision-making method for the separation and recovery of gravel and coal, characterized in that, include: Unstructured documents and structured operational data in the field of stone and coal sorting and recycling are obtained. The unstructured documents are vectorized and stored in a vector database, and the structured operational data is stored in a relational database to construct a domain knowledge base. Low-rank adaptation technology is used to efficiently fine-tune the parameters of a general large language model to obtain a domain-specific model. During inference, natural language questions from operators and real-time operating data are received. Operation procedure texts and historical case fragments related to the questions are retrieved from the vector database. The retrieved text fragments, real-time operating data and questions are combined to form enhanced prompts, which are then input into the domain-specific model to generate decision text containing parameter adjustment suggestions. The decision text is subjected to security filtering and structured output. The suggested parameter values ​​in the decision text are extracted and compared with the preset multi-level security boundary table. Parameter values ​​that exceed the hard boundary are forcibly clamped to boundary values ​​and warnings are added. The output is a structured decision suggestion containing operation steps and the basis for the suggestion. Record the operators' feedback on the adoption of the structured decision suggestions and the subsequent operational effects, and generate an interaction log; automatically select high-quality cases from the interaction log according to preset quality conditions, and after review, add high-quality cases to the domain knowledge base, and trigger incremental fine-tuning updates of the domain-specific model based on the accumulated feedback data.

2. The method as described in claim 1, characterized in that, The process involves acquiring unstructured documents and structured operational data in the field of stone and coal sorting and recycling, vectorizing the unstructured documents and storing them in a vector database, and storing the structured operational data in a relational database to construct a domain knowledge base, including: The unstructured document is segmented, and each text segment is converted into a dense vector using a Chinese semantic embedding model. The dense vectors are then stored in a vector database using the HNSW algorithm. Structured operational data, including timestamps, equipment numbers, process parameter values, and operating condition labels, is stored in a relational database.

3. The method as described in claim 1, characterized in that, Vectorization of unstructured documents includes: A sliding window strategy is used to segment unstructured documents, with a window size of 512 tokens and a step size of 256 tokens. For key clauses in the safety operating procedures, overlapping segments with a step size smaller than the window size are used to preserve semantic coherence.

4. The method as described in claim 1, characterized in that, Efficient parameter fine-tuning, including: Using LoRA technology, the weight update of the general large language model is defined as... ,in For frozen pre-trained weights, and It is a learnable low-rank matrix; the data used during training includes question-and-answer pairs constructed from operating procedures, historical fault handling cases, and condition-parameter adjustment suggestion pairs.

5. The method as described in claim 1, characterized in that, The search employs a hybrid search strategy, including: The semantic score is obtained by calculating the cosine similarity between the user query vector and the document vector. The keyword matching score is calculated, and the two are weighted and summed to obtain the final ranking score. The top K text fragments containing operation procedure fragments and historical case fragments are returned in descending order according to the final ranking score.

6. The method as described in claim 5, characterized in that, The formula for calculating cosine similarity is: The filter threshold is set to 0.7; The weighted summation formula is: ,in If K is 5, it returns 2 operation procedure fragments and 3 historical case fragments.

7. The method according to claim 1, characterized in that, Security filtering includes: Set hard lower limits for each process parameter and hard limit For any extracted suggested parameter value Output clamp value ;like If the value exceeds the recommended range but does not reach the hard boundary, a warning mark is generated without forcing a change; if the hard boundary is reached, the value is forcibly corrected to the boundary value, a warning mark is added, and the original recommended value is retained.

8. The method as described in claim 1, characterized in that, The structured decision recommendations include: current operating condition assessment, recommended parameter combinations, step-by-step operation procedures, and the basis for the recommendations; among which, the basis for the recommendations is cited from the retrieved operating procedure clauses or historical case numbers, and includes an explanation of the parameter derivation logic.

9. The method according to claim 1, characterized in that, The preset quality conditions are: Define adoption flag The indicator that the operating parameters are within the normal range for two consecutive hours after adoption. A system confidence level greater than 0.8 is a key indicator. The formula for judging case quality is: If and only if When the corresponding interaction record is determined to be a high-quality case, it is submitted to the review queue and updated after expert confirmation.

10. A large-scale model-based intelligent decision-making device for the separation and recovery of stones and coal, characterized in that, include: The knowledge base construction module is used to acquire unstructured documents and structured operational data in the field of stone and coal sorting and recycling. The unstructured documents are vectorized and stored in a vector database, and the structured operational data is stored in a relational database to construct a domain knowledge base. The intelligent reasoning module is used to efficiently fine-tune the parameters of a general large language model using low-rank adaptation technology to obtain a domain-specific model. During reasoning, it receives natural language questions from operators and real-time operating data, retrieves relevant operating procedure texts and historical case fragments from the vector database, combines the retrieved text fragments, real-time operating data and questions into enhanced prompts, inputs them into the domain-specific model, and generates decision text containing parameter adjustment suggestions. The security filtering and output module is used to perform security filtering and structured output on the decision text, extract the suggested parameter values ​​in the decision text, compare them with the preset multi-level security boundary table, forcibly clamp the parameter values ​​that exceed the hard boundary to the boundary value and add a warning, and output structured decision suggestions containing operation steps and the basis for the suggestions. The co-evolution module is used to record the operators' adoption feedback on the structured decision suggestions and the subsequent operational effects, and generate interaction logs. High-quality cases are automatically selected from the interaction logs according to preset quality conditions. After review, high-quality cases are added to the domain knowledge base, and incremental fine-tuning updates of the domain-specific model are triggered based on the accumulated feedback data.

Images