Knowledge and data driven model construction method for health management of abrasive grinding equipment

Abstract

The application provides a knowledge and data double-driven broken mill equipment health management industry model construction method, and belongs to the technical field of artificial intelligence.In the application, firstly, broken mill equipment multi-source knowledge and operation and maintenance data are acquired, knowledge extraction, sequence reconstruction and unified vectorization processing are performed, a mixed vector with semantic alignment is formed, differences of heterogeneous data are eliminated, and standardized input is provided for a model;then, a neural engine is constructed based on a Transformer, a symbolic reasoning engine is constructed in combination with an expert rule base and forward reasoning, an interactive checkable double-engine collaborative reasoning architecture is formed, data generalization and mechanism constraint are considered;next, real-time operation and maintenance data and query input architecture are constructed, a result is generated by the neural engine, and the symbolic engine is checked, iteratively modified, and a reliable reasoning sample is obtained;finally, a health management instruction set is constructed, a low-rank adaptive fine-tuning model is adopted, RAG retrieval and thinking chain prompting are fused, and finally integrated and highly reliable broken mill equipment health management industry model construction is completed.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, specifically relating to a method for constructing an industry model for health management of grinding and crushing equipment driven by both knowledge and data. Background Technology

[0002] With the development of industrial intelligence and large-scale modeling technology, industry-specific large-scale models for health management of industrial equipment are gradually becoming the core support for intelligent operation and maintenance of equipment. These models enable integrated analysis of fault diagnosis, health assessment, lifespan prediction, and operation and maintenance decisions, effectively improving the reliability and efficiency of critical equipment operation. Building specialized industry-specific large-scale models has become an important development direction for the intelligent upgrading of process equipment such as crushing and grinding.

[0003] Existing industrial equipment health management models mostly adopt general large models combined with domain data for fine-tuning. They mainly rely on a single data-driven mode to complete reasoning, and some only achieve domain adaptation through simple knowledge injection. They lack deep integration of structured knowledge and operation and maintenance data, and have not established an iteratively verifiable reasoning mechanism. They are difficult to fully adapt to the strong mechanistic constraints of equipment health management, are prone to domain illusions and logical contradictions, and have low sample reliability and poor model interpretability. They are difficult to meet the practical application requirements of high reliability, strong mechanism and traceability in the health management of grinding and crushing equipment. Summary of the Invention

[0004] The technical problem to be solved by this invention is how to improve the output reliability of the constructed industry model for health management of grinding and crushing equipment. In view of the shortcomings of the existing technology, this invention provides a method for constructing an industry model for health management of grinding and crushing equipment driven by both knowledge and data.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: This invention provides a method for constructing an industry model for health management of grinding and crushing equipment driven by both knowledge and data, comprising: S1. Obtain multi-source knowledge and operation and maintenance data in the field of grinding and crushing equipment, and sequentially perform knowledge extraction, data sequence reconstruction and unified vectorization processing on the multi-source knowledge and operation and maintenance data to obtain hybrid vector data; S2. Construct a neural engine based on the Transformer architecture model and the hybrid vector data, construct a symbolic reasoning engine based on the expert rule base and the forward inference engine, establish a data interaction interface between the neural engine and the symbolic reasoning engine, and form a dual-engine collaborative reasoning architecture. S3. Input the real-time equipment operation and maintenance data and user query requests into the dual-engine collaborative reasoning architecture, generate preliminary reasoning results through the neural engine, verify the preliminary reasoning results through the mechanism rules by the symbolic reasoning engine, and perform a preset number of iterations and corrections through the cooperation of the neural engine and the symbolic reasoning engine to obtain credible reasoning results and reasoning process, and form a credible reasoning sample training set. S4. Based on the entire health management process of the grinding and crushing equipment field, construct a health management instruction set for grinding and crushing equipment. Based on the trusted inference sample training set and the health management instruction set for grinding and crushing equipment, perform domain adaptation fine-tuning on the Transformer architecture model through the low-rank adaptive fine-tuning method, and integrate RAG retrieval enhancement generation and mind chain prompts to complete the construction of the industry model for health management of grinding and crushing equipment.

[0006] Compared to existing technologies, the beneficial effects of this invention include: First, acquiring multi-source knowledge and operation and maintenance data in the field of grinding and crushing equipment, and sequentially performing knowledge extraction, data sequence reconstruction, and unified vectorization processing on this type of data, converting text, knowledge, and time-series data into semantically aligned hybrid vector data, eliminating heterogeneous data differences through a unified representation method, providing a standardized data foundation for subsequent model input and inference, and avoiding feature learning bias caused by inconsistent data formats; Subsequently, building a neural engine with strong feature learning capabilities based on the Transformer architecture model, and simultaneously building a symbolic inference engine with logical constraints based on an expert rule base and a forward inference engine, and establishing a data interaction interface between the two, forming a dual-engine collaborative inference architecture that can mutually verify and iteratively optimize, enabling the model to simultaneously possess data generalization ability and mechanism constraint ability, structurally solving the problem of unreliable reasoning by a single large model and the tendency to experience domain illusion; Next, inputting real-time equipment operation and maintenance data and user query requests into the dual-engine collaborative inference architecture, first by the neural engine... The engine generates preliminary inference results based on data features, and then the symbolic inference engine verifies and judges the results according to the equipment mechanism rules. Results that do not conform to the rules are marked as contradictory and sent back to the neural engine for re-inference. The correction is completed within a maximum of preset iterations. Finally, a reliable inference result and a complete inference process are output, forming a high-quality training sample set. The reliability of the samples is significantly improved through cross-validation of the dual engines, providing high-quality supervision signals for subsequent model fine-tuning. Finally, a dedicated instruction set is built around the entire health management process of grinding and crushing equipment. Based on the reliable inference samples and instruction set, the Transformer model is adapted to the domain using a low-rank adaptive fine-tuning method. The RAG retrieval enhancement and mind chain prompting mechanism are integrated into the model. The professionalism and interpretability of the model are further enhanced through external knowledge retrieval and step-by-step reasoning guidance. Finally, the industry model for health management of grinding and crushing equipment with data learning ability, domain knowledge constraints and reasoning traceability is completed, realizing the integrated and reliable output of fault diagnosis, health assessment, life prediction and operation and maintenance decision-making.

[0007] Optionally, the multi-source knowledge and operation and maintenance data in S1 include structured knowledge, unstructured knowledge, and equipment time-series monitoring data. The structured knowledge includes equipment structural entities, fault mode entities, component relationships, and fault causal relationships. The unstructured knowledge includes equipment operation manuals, fault repair cases, industry standards and specifications, and operation and maintenance text logs. The equipment time-series monitoring data includes vibration data, temperature data, pressure data, and current data.

[0008] Optionally, the knowledge extraction in S1 specifically includes: S11. Identify and extract knowledge entities and relationships between knowledge entities from the structured knowledge, assign a unique identifier to each knowledge entity and relationship between knowledge entities, and form structured knowledge text; S12. Perform word segmentation, stop word removal, and part-of-speech tagging on the unstructured knowledge text, extract professional and effective information, and convert it into an operable data format to form unstructured knowledge text, which is then combined with structured knowledge text to form text data.

[0009] Optionally, the data sequence reconstruction and unified vectorization processing in S1 specifically includes: S13. Identify knowledge entities in the text data, and replace entity words in the text data with corresponding entity special lexical units; S14. Add special word elements representing the relationships between the knowledge entities to the end of the text data sequence to form a mixed input sequence that integrates the text information and the knowledge information; S15. The mixed input sequence is transformed into a fixed-dimensional vector through the Embedding layer of the Transformer architecture model; S16. Perform time-domain feature extraction and frequency feature extraction on the device time-series monitoring data, concatenate them to obtain a high-dimensional feature vector, perform normalization and dimension mapping on the high-dimensional feature vector in sequence, obtain a feature vector with the same dimension as the fixed-dimensional vector, and form the hybrid vector data with the fixed-dimensional vector.

[0010] Optionally, the neural engine constructed based on the Transformer architecture model and the hybrid vector data in S2 specifically includes: S21. Based on the Transformer architecture model, perform global feature modeling on the hybrid vector data, and learn the implicit correlation patterns in the hybrid vector data through a self-attention mechanism. The implicit correlation patterns include failure modes, health status, and lifespan trends. S22. Based on the implicit correlation rules, a task output layer is built on the backbone network of the Transformer architecture model to construct the neural engine.

[0011] Optionally, the construction of the symbolic reasoning engine based on the expert rule base and the forward inference engine in S2 specifically includes: S23. Collect rules compiled by domain experts based on the equipment mechanism, fault diagnosis and operation and maintenance constraints knowledge of the crushing and grinding equipment field, and store the rules in a structured and indexed manner to form the expert rule base. The expert rule base includes equipment mechanism rules, fault diagnosis rules and operation and maintenance constraint rules. S24. Build a forward inference engine and configure rule matching, logic verification and conflict resolution mechanisms to construct the symbolic inference engine.

[0012] Optionally, S3 includes: S31. Input the real-time operation and maintenance data of the equipment and the user query request into the neural engine. Based on the implicit correlation law, generate the preliminary reasoning result through the neural engine. The preliminary reasoning structure includes key data features of fault hypothesis, health assessment result, decision suggestion and reasoning basis. S32. The preliminary reasoning result is transmitted to the symbolic reasoning engine through the neural engine, the symbolic reasoning engine calls the expert rule base for rule matching, and the preliminary reasoning result is logically verified based on the hybrid vector data. S33. If the logic verification passes, the preliminary reasoning result is marked as the reliable reasoning result by the symbolic reasoning engine, and the reliable reasoning result and the corresponding reasoning process are output. S34. If the logical verification fails, the symbolic reasoning engine locates the contradiction in the preliminary reasoning result, transmits the contradiction information to the neural engine based on the conflict resolution mechanism, and allows the neural engine to reanalyze the real-time equipment operation and maintenance data and adjust the preliminary reasoning result based on the contradiction information. The adjusted preliminary reasoning result is then transmitted back to the symbolic reasoning engine. When the logical verification of the symbolic reasoning engine passes or the preliminary reasoning result has been iterated and corrected three times, the last preliminary reasoning result is marked as the reliable reasoning result, and the reliable reasoning result and the corresponding reasoning process are output. S35. Based on the credible reasoning results and the reasoning process, form the credible reasoning sample training set.

[0013] Optionally, the construction of a health management instruction set for grinding equipment based on the entire health management process in the grinding equipment field in step S4 specifically includes: S41. Based on the entire health management process in the field of grinding and crushing equipment, determine the scope of instructions around the core tasks of health management in the field of grinding and crushing equipment. The core tasks of health management include fault diagnosis, health assessment, remaining life prediction and operation and maintenance decision-making. S42. Based on the core equipment in the grinding and crushing equipment field, typical working conditions and operation and maintenance scenarios are sorted out, and a unified instruction template is adopted to generate a preset scale of domain instructions according to the scenario and task type. After being reviewed, deduplicated and verified by domain experts, the health management instruction set of the grinding and crushing equipment is formed. The unified instruction template is a four-segment structure template consisting of instruction description, input data example, output result example and evaluation standard.

[0014] Optionally, the step S4, which involves performing domain adaptation fine-tuning of the Transformer architecture model based on the trusted inference sample training set and the grinding equipment health management instruction set using a low-rank adaptive fine-tuning method, specifically includes: S43. Introduce a low-rank decomposition matrix into the projection layer of the self-attention module of the Transformer architecture model and complete parameter initialization; S44. Based on the trusted inference sample training set and the grinding equipment health management instruction set, divide the training set and the validation set, and configure the training hyperparameters, including the learning rate, batch size, training rounds, gradient pruning and weight decay. S45. Input the hybrid vector data, perform inference training of the Transformer architecture model based on the training set and the training hyperparameters, and output the inference result; S46. Combining the equipment mechanism rules in the field of grinding equipment, calculate the task loss and mechanism consistency constraint loss based on the reasoning results, and update the parameters of the low-rank decomposition matrix; S47. After iterating the inference training until the validation set index is optimal, the weights of the low-rank decomposition matrix and the original weights of the Transformer architecture model are fused to complete the initial construction of the grinding equipment health management industry model.

[0015] Optionally, the integration of RAG retrieval enhancement generation and mind chain hints in S4 specifically includes: S48. Construct a vector retrieval library for the grinding and crushing equipment field based on the hybrid vector data, and integrate a RAG retrieval enhancement generation mechanism into the initially constructed grinding and crushing equipment health management industry model, so as to call knowledge fragments of the vector retrieval library during the reasoning process through the RAG retrieval enhancement generation mechanism. S49. Construct a thought chain prompt template and integrate the thought chain prompt template into the initially constructed health management industry model for grinding and crushing equipment, so as to provide step-by-step reasoning guidance during the reasoning process through the thought chain prompt template. Attached Figure Description

[0016] The present invention will now be described in further detail with reference to the accompanying drawings.

[0017] Figure 1 The flowchart of the industry model construction method for health management of grinding and crushing equipment driven by both knowledge and data in this embodiment of the invention. Detailed Implementation

[0018] To better understand the present invention, the following embodiments further illustrate the content of the invention, but the scope of protection of the present invention is not limited to the following embodiments. Numerous specific details are set forth in the following description to provide a more thorough understanding of the invention. However, it will be apparent to those skilled in the art that the present invention can be practiced without one or more of these details.

[0019] It should be noted that the Z-axis in the attached figures represents the vertical direction, i.e., the up-down position, with the positive direction of the Z-axis representing upward and the negative direction representing downward; the Y-axis in the attached figures represents the horizontal direction and is designated as the front-back position, with the positive direction of the Y-axis representing the front side and the negative direction representing the back side; the X-axis in the attached figures represents the left-right position, with the positive direction of the X-axis representing the right side and the negative direction representing the left side. It should also be noted that the aforementioned representations of the Z, Y, and X axes are only for the convenience of describing the present invention and for simplifying the description, and are not intended to 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 limiting the present invention.

[0020] The term "comprising" and its variations as used herein are open-ended, meaning "including but not limited to"; the term "based on" means "at least partially based on"; the term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments"; and the term "optionally" means "optional embodiments". Definitions of other terms will be given in the following description. It should be noted that the concepts of "first," "second," etc., mentioned in this invention are used only to distinguish different devices, modules, or units, and are not intended to limit the order of functions performed by these devices, modules, or units or their interdependencies.

[0021] It should be noted that the terms "a" and "a plurality of" used in this invention are illustrative rather than restrictive. Those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0022] An embodiment of the present invention provides a method for constructing an industry model for health management of grinding and crushing equipment driven by both knowledge and data, comprising: S1, acquiring multi-source knowledge and operation and maintenance data in the field of grinding and crushing equipment, and sequentially performing knowledge extraction, data sequence reconstruction, and unified vectorization processing on the multi-source knowledge and operation and maintenance data to obtain hybrid vector data; S2, constructing a neural engine based on the Transformer architecture model and the hybrid vector data, constructing a symbolic reasoning engine based on an expert rule base and a forward inference engine, establishing a data interaction interface between the neural engine and the symbolic reasoning engine, and forming a dual-engine collaborative reasoning architecture; S3, inputting real-time operation and maintenance data of the equipment and user query requests into the dual-engine collaborative... The inference architecture generates initial inference results through a neural engine, verifies these results using a symbolic inference engine based on mechanistic rules, and then iterates and corrects them a predetermined number of times through the cooperation of the neural engine and the symbolic inference engine to obtain reliable inference results and processes, forming a reliable inference sample training set. S4: Based on the entire health management process in the grinding and crushing equipment field, a health management instruction set for grinding and crushing equipment is constructed. Based on the reliable inference sample training set and the instruction set, the Transformer architecture model is fine-tuned for domain adaptation using a low-rank adaptive fine-tuning method, and RAG retrieval enhancement generation and mind chain prompts are integrated to complete the construction of the industry model for health management of grinding and crushing equipment.

[0023] Specifically, the dual-driven knowledge and data approach refers to simultaneously utilizing knowledge from the grinding and crushing equipment domain and actual operation and maintenance data to drive model training and inference. Multi-source knowledge and operation and maintenance data refer to the collection of equipment domain knowledge and operational monitoring data. Knowledge extraction is used to extract professional knowledge units from multi-source information; data sequence reconstruction is used to fuse text and knowledge into a unified input; unified vectorization refers to mapping heterogeneous data to the same vector space to achieve a fused representation of knowledge and data. The Transformer architecture is a deep neural network architecture based on a self-attention mechanism, possessing global feature modeling capabilities. The neural engine is a data-driven inference unit based on the Transformer architecture, while the symbolic inference engine is a logical verification unit based on expert rules and forward inference. Iterative correction refers to repeatedly verifying and optimizing the inference results within a preset number of times (3 times). Low-rank adaptation fine-tuning (LoRA) is used for fine-tuning. RAG stands for Retrieval Augmented Generation, which is used to introduce external credible knowledge. Mind chain hints guide the model's step-by-step inference and improve interpretability.

[0024] In this embodiment, firstly, as... Figure 1As shown in S1, multi-source knowledge and operation and maintenance data in the field of grinding and crushing equipment are acquired. Knowledge extraction, data sequence reconstruction, and unified vectorization processing are then performed on this data sequentially. Text, knowledge, and time-series data are converted into semantically aligned hybrid vector data. A unified representation method eliminates heterogeneous data differences, providing a standardized data foundation for subsequent model input and inference, and avoiding feature learning bias caused by inconsistent data formats. Subsequently, as... Figure 1 As shown in S2, a neural engine with strong feature learning capabilities is built based on the Transformer architecture model. Simultaneously, a symbolic reasoning engine with logical constraints is built based on an expert rule base and a forward inference engine. A data interaction interface is established between the two, forming a dual-engine collaborative reasoning architecture that can mutually verify and iteratively optimize. This enables the model to possess both data generalization ability and mechanistic constraint ability, structurally solving the problems of unreliable reasoning and domain illusion caused by single large models. Then, as... Figure 1 As shown in S3, real-time equipment operation and maintenance data and user query requests are input into a dual-engine collaborative inference architecture. First, the neural engine generates preliminary inference results based on data features. Then, the symbolic inference engine verifies and judges the results according to equipment mechanism rules. Results that do not conform to the rules are marked as contradictory and sent back to the neural engine for re-inference. Corrections are completed within a maximum of preset iterations. Finally, a reliable inference result and a complete inference process are output, forming a high-quality training sample set. Cross-validation by the dual engines significantly improves sample reliability, providing high-quality supervision signals for subsequent model fine-tuning. Finally, as... Figure 1 As shown in S4, a dedicated instruction set is constructed around the entire health management process of grinding and crushing equipment. Based on the reliable inference samples and instruction set, a low-rank adaptive fine-tuning method is used to adapt the Transformer model to the domain. The RAG retrieval enhancement and mind chain prompting mechanism are integrated into the model. The professionalism and interpretability of the model are further enhanced through external knowledge retrieval and step-by-step reasoning guidance. Finally, the industry model for health management of grinding and crushing equipment with data learning ability, domain knowledge constraints and reasoning traceability is completed, realizing the integrated and reliable output of fault diagnosis, health assessment, life prediction and operation and maintenance decision-making.

[0025] Optionally, the multi-source knowledge and operation and maintenance data in S1 include structured knowledge, unstructured knowledge, and equipment time-series monitoring data. The structured knowledge includes equipment structural entities, fault mode entities, component relationships, and fault causal relationships. The unstructured knowledge includes equipment operation manuals, fault repair cases, industry standard specifications, and operation and maintenance text logs. The equipment time-series monitoring data includes vibration data, temperature data, pressure data, and current data.

[0026] Specifically, structured knowledge refers to directly computable domain knowledge organized in the form of entities and relationships, with the aim of providing clear domain logic for models; unstructured knowledge refers to natural language data, with the aim of covering experiential knowledge in real operation and maintenance scenarios; and equipment time-series monitoring data reflects the real-time operating status of equipment.

[0027] In this optional embodiment, multi-source knowledge and operation and maintenance data consist of structured knowledge, unstructured knowledge, and equipment time-series monitoring data. Structured knowledge exists in standardized forms such as equipment components, fault modes, and relationships, and can be directly used for knowledge graph construction and rule reasoning. Unstructured knowledge exists in natural language text form, supplementing a wealth of practical operation and maintenance experience and industry standards. Time-series monitoring data can reflect the real-time operating status and abnormal changes of equipment. During the data processing stage, structured knowledge provides direct evidence for knowledge entity and relationship extraction, unstructured knowledge is preprocessed into computable text, and time-series data provides dynamic features for health status assessment. These three types of data complement and corroborate each other, enabling the model to possess both static domain knowledge and dynamic operational awareness. This setup, by introducing multiple types and dimensions of data sources, effectively avoids the knowledge bias problem caused by a single data type, enhances the model's adaptability to complex operating conditions and variable faults, and provides a rich and reliable input foundation for subsequent knowledge extraction, sequence reconstruction, and vectorization processing, improving the accuracy and robustness of the final industry model in the health management tasks of grinding and crushing equipment.

[0028] Optionally, the knowledge extraction in S1 specifically includes: S11, identifying and extracting knowledge entities and relationships between knowledge entities from structured knowledge, assigning a unique identifier to each knowledge entity and relationship between knowledge entities to form structured knowledge text; S12, performing word segmentation, stop word removal, and part-of-speech tagging preprocessing on unstructured knowledge, extracting professional and effective information and converting it into an operable data format to form unstructured knowledge text, and combining it with structured knowledge text to form text data.

[0029] Specifically, knowledge entities refer to core conceptual units in fields such as equipment, components, faults, and phenomena; relationships between knowledge entities refer to logical connections such as composition, causality, and representation between entities; unique identifiers are used to distinguish different entities and relationships, achieving standardized knowledge labeling; word segmentation, stop word removal, and part-of-speech tagging are natural language preprocessing operations, the purpose of which is to eliminate redundant information and retain professional and effective content.

[0030] In this optional embodiment, the knowledge extraction process is divided into two parts: structured knowledge processing and unstructured knowledge processing. In the structured knowledge processing, structured knowledge is first identified, extracting knowledge entities such as equipment structure and fault modes, as well as relationships between entities such as component associations and fault causality. A unique identifier is assigned to each type of entity and relationship, transforming the structured knowledge into standardized structured knowledge text, enabling unified knowledge expression and rapid retrieval. In the unstructured knowledge processing, unstructured knowledge undergoes text preprocessing, including word segmentation, removal of meaningless stop words, and part-of-speech tagging to filter professional terms. Effective information related to equipment operation and maintenance and fault diagnosis is extracted and converted into a data format that the model can process, forming a regular unstructured knowledge text. Based on this, the two types of knowledge text are combined to form unified text data, preserving the rigorous logic of structured knowledge while absorbing the rich contextual information of unstructured knowledge. This achieves standardized fusion of multi-source knowledge, providing clear and accurate processing objects for knowledge entity replacement and relationship labeling in subsequent data sequence reconstruction, improving knowledge injection efficiency and the model's absorption of domain knowledge.

[0031] Optionally, the data sequence reconstruction and unified vectorization processing in S1 specifically includes: S13, identifying knowledge entities in the text data and replacing entity words in the text data with corresponding entity special words; S14, adding corresponding relationship special words of the relationships between the knowledge entities involved to the end of the text data sequence to form a mixed input sequence that integrates text information and knowledge information; S15, converting the mixed input sequence into a fixed-dimensional vector through the Embedding layer of the Transformer architecture model; S16, performing time-domain feature extraction and frequency feature extraction on the device time-series monitoring data, concatenating them to obtain a high-dimensional feature vector, performing normalization and dimension mapping on the high-dimensional feature vector in sequence, obtaining a feature vector with the same dimension as the fixed-dimensional vector, and forming a mixed vector data with the fixed-dimensional vector.

[0032] Specifically, knowledge entities are core professional terms in the domain; special tokens are special markers that the model can recognize, used to highlight knowledge entities and relationships; the embedding layer is a word embedding layer built into the Transformer model, used to map text symbols into continuous numerical vectors; time-domain features and frequency-domain features describe the changing patterns of time-series data from the time dimension and frequency dimension, respectively; normalization and dimensional mapping are used to make the time-series feature vector consistent with the text vector dimension.

[0033] In this optional embodiment, based on the knowledge extraction, data sequence reconstruction and unified vectorization processing are performed sequentially. First, knowledge entities in the text data are identified, and the original entity words are replaced with special entity terms, enabling the model to directly perceive key domain concepts during the input stage. Then, special terms corresponding to the relationship are appended to the end of the sequence, fusing text information and knowledge relationship information into a hybrid input sequence, realizing knowledge-enhanced input representation, allowing the model to learn features by combining domain knowledge during the encoding stage. Next, the hybrid input sequence is transformed into a fixed-dimensional vector through the embedding layer of the Transformer model, realizing semantic numerical representation of text and knowledge. At the same time, time-domain features and frequency-domain features are extracted from time-series monitoring data such as vibration and temperature, respectively, concatenated to form a high-dimensional feature vector, and then normalized. The dimension is then mapped to maintain the same dimension as the text vector, and finally, together with the time-series feature vector, they constitute hybrid vector data. Through the above processing, text semantics, domain knowledge, and runtime features are uniformly encoded into jointly computeable vectors, eliminating representation barriers between heterogeneous data, providing a unified data foundation for neural engine feature learning and dual-engine collaborative reasoning, and significantly improving the model's ability to fuse multimodal information and its reasoning accuracy.

[0034] Optionally, the neural engine construction based on the Transformer architecture model and hybrid vector data in S2 specifically includes: S21, using the Transformer architecture model as the base, performing global feature modeling on the hybrid vector data, and learning the implicit correlation patterns in the hybrid vector data through a self-attention mechanism. The implicit correlation patterns include failure modes, health states, and lifespan trends; S22, based on the implicit correlation patterns, building a task output layer on the backbone network of the Transformer architecture model to construct the neural engine.

[0035] Specifically, global feature modeling refers to performing overall correlation analysis on mixed vector data; self-attention mechanism is used to capture long-distance dependencies between data; implicit correlation patterns include potential patterns that are difficult to observe directly, such as failure modes, health status, and lifespan trends; and the task output layer is an output structure adapted to health management tasks.

[0036] In this optional embodiment, the neural engine is built upon the Transformer architecture. First, mixed vector data is input into the Transformer model, which utilizes its self-attention mechanism to perform global feature modeling on text, knowledge, and temporal fusion features. This automatically mines implicit correlations such as fault association patterns, health decay laws, and lifespan change trends hidden in the data, achieving accurate perception of complex operating states without relying on manual feature design. Subsequently, a task output layer adapted for fault diagnosis, health assessment, lifespan prediction, and maintenance decision-making is built after the Transformer backbone network, enabling the model to directly output structured reasoning results that meet industry requirements. This configuration allows the neural engine built in this way to retain the powerful semantic understanding and generalization capabilities of the pre-trained model while specifically adapting to the health management needs of grinding and crushing equipment. It can efficiently process multi-source fusion data and generate reasonable fault hypotheses and state assessment results, providing an efficient data-driven reasoning front-end for dual-engine collaborative reasoning, and simultaneously providing a basic model structure for subsequent domain-specific fine-tuning.

[0037] Optionally, the construction of the symbolic reasoning engine based on the expert rule base and forward inference engine in S2 specifically includes: S23, collecting rules organized by domain experts based on the equipment mechanism, fault diagnosis and operation and maintenance constraints knowledge in the field of grinding and crushing equipment, and storing the rules in a structured and indexed manner to form an expert rule base, which includes equipment mechanism rules, fault diagnosis rules and operation and maintenance constraint rules; S24, building a forward inference engine and configuring rule matching, logic verification and conflict resolution mechanisms to construct the symbolic reasoning engine.

[0038] Specifically, the expert rule base consists of equipment mechanism, fault diagnosis, and operation and maintenance constraint rules compiled by domain experts, including equipment mechanism rules, fault diagnosis rules, and operation and maintenance constraint rules; the forward inference engine is a reasoning execution module that derives conclusions from known conditions; rule matching is used to compare the reasoning results with the rules; logic verification is used to determine whether the results conform to domain logic; and the conflict resolution mechanism is used to handle decision selection when multiple rules conflict.

[0039] In this optional embodiment, during the construction of the symbolic reasoning engine, professional rules formed by domain experts based on the physical mechanism of grinding equipment, typical fault logic, and operation and maintenance safety specifications are first collected. These rules are then structured, indexed, and stored to construct an expert rule base containing equipment mechanism rules, fault diagnosis rules, and operation and maintenance constraint rules. This ensures that domain knowledge exists in the form of directly callable rules, guaranteeing the professionalism and rigor of the reasoning logic. Subsequently, a forward inference engine is built and configured with rule matching, logic verification, and conflict resolution mechanisms. The inference engine can automatically match relevant rules and execute logical judgments based on input information, and make reasonable decisions according to preset strategies when multiple rules conflict. The symbolic reasoning engine constructed in this way can independently carry out logical reasoning and result verification, strictly judge the compliance of fault assumptions and evaluation conclusions output by the neural engine, promptly identify erroneous results that violate the mechanism, and constrain model behavior from the logical level. It complements the data-driven neural engine, together forming a dual-engine collaborative reasoning architecture that combines flexibility and reliability.

[0040] Optionally, S3 includes: S31, inputting real-time equipment operation and maintenance data and user query requests to the neural engine, generating preliminary inference results through the neural engine based on implicit association patterns, the preliminary inference structure including fault hypotheses, health assessment results, decision suggestions, and key data features of the inference basis; S32, transmitting the preliminary inference results to the symbolic inference engine through the neural engine, calling the expert rule base for rule matching through the symbolic inference engine, and performing logical verification on the preliminary inference results based on hybrid vector data; S33, if the logical verification passes, marking the preliminary inference results as credible inference results through the symbolic inference engine, and outputting the credible inference results and corresponding inferences. S34. If the logic verification fails, the symbolic reasoning engine locates the contradiction in the preliminary reasoning result. Based on the conflict resolution mechanism, the contradiction information is transmitted to the neural engine. The neural engine then re-analyzes the real-time operation and maintenance data of the equipment and adjusts the preliminary reasoning result based on the contradiction information. The adjusted preliminary reasoning result is then transmitted back to the symbolic reasoning engine. When the logic verification of the symbolic reasoning engine passes or the preliminary reasoning result has been iterated and corrected 3 times, the last preliminary reasoning result is marked as a reliable reasoning result, and the reliable reasoning result and the corresponding reasoning process are output. S35. Based on the reliable reasoning result and the reasoning process, a reliable reasoning sample training set is formed.

[0041] Specifically, the preliminary inference results include fault hypotheses, health assessments, decision recommendations, and key data features; rule matching refers to comparing the inference results with the expert rule base; logical verification is used to determine whether the results conform to the domain mechanism; contradiction points refer to the locations where the inference results conflict with the rules; the conflict resolution mechanism is used to handle rule contradictions and provide correction directions; iterative correction is set to a maximum of 3 iterations, balancing accuracy and efficiency; the credible inference sample training set is a high-quality fine-tuned sample verified by the dual engines.

[0042] In this optional embodiment, after the dual-engine collaborative reasoning architecture is built, reasoning verification and sample generation are implemented. First, real-time operation and maintenance data and user queries are input into the neural engine. The neural engine generates preliminary reasoning results based on learned implicit correlation patterns, including fault hypotheses, health assessments, decision suggestions, and supporting data features, completing the data-driven prediction process. Then, the preliminary reasoning results are transmitted to the symbolic reasoning engine. The engine calls the expert rule base for rule matching and performs logical verification using hybrid vector data to determine whether the results conform to the equipment mechanism and operation and maintenance specifications. If the verification passes, the result is directly marked as a reliable reasoning result and the complete reasoning process is output. If the verification fails, the symbolic reasoning engine locates the contradiction point and generates correction information through a conflict resolution mechanism, which is then sent back to the neural engine. The neural engine re-analyzes the data and adjusts the reasoning results based on the contradiction information, then sends them back to the symbolic reasoning engine for verification until the verification passes or the iteration count reaches 3, at which point the last result is determined as a reliable result. Finally, a reliable reasoning sample training set is constructed based on all reliable results and corresponding reasoning processes. This setup, through multiple collaborative iterative corrections, effectively eliminates unreliable outputs such as model illusions and logical errors, significantly improving sample quality and consistency. It provides high-quality supervisory data for subsequent model fine-tuning, thereby enhancing the reliability and interpretability of industry models from the source.

[0043] Optionally, the construction of the grinding equipment health management instruction set based on the entire health management process in the grinding equipment field in S4 specifically includes: S41, based on the entire health management process in the grinding equipment field, determining the instruction coverage scope around the core health management tasks in the grinding equipment field. The core health management tasks include fault diagnosis, health assessment, remaining life prediction, and operation and maintenance decision-making; S42, based on the core equipment in the grinding equipment field, sorting out typical working conditions and operation and maintenance scenarios, and using a unified instruction template, generating a preset scale of domain instructions according to the scenario and task type, and forming a grinding equipment health management instruction set after being reviewed, deduplicated, and verified by domain experts. The unified instruction template is a four-segment structure template of instruction description, input data example, output result example, and evaluation criteria.

[0044] Specifically, the entire health management process covers four core tasks: fault diagnosis, health assessment, remaining life prediction, and operation and maintenance decision-making; typical working conditions and operation and maintenance scenarios refer to the actual working scenarios of equipment under different operating states; the unified instruction template adopts a four-segment structure of instruction description, input data example, output result example, and evaluation criteria; the preset scale domain instructions are sufficient instruction samples generated according to the scenario; expert review, deduplication, and verification are used to ensure instruction quality.

[0045] In this optional embodiment, during the construction of a dedicated health management instruction set, the process first focuses on the entire health management process of grinding and crushing equipment, clarifying four core tasks: fault diagnosis, health assessment, remaining life prediction, and operation and maintenance decision-making. This determines the functional scope the instruction set needs to cover, ensuring that the instructions fully support actual operation and maintenance needs. Then, taking core equipment such as ball mills and crushers as examples, typical operating conditions and operation and maintenance scenarios are analyzed. A unified four-part template—instruction description, input data examples, output result examples, and evaluation criteria—is used to generate a pre-set scale of domain instructions in batches according to scenario and task type. These instructions are then reviewed by domain experts to remove duplicates and correct unreasonable content, ultimately forming a high-quality, high-coverage health management instruction set for grinding and crushing equipment. This setup transforms abstract health management tasks into concrete instruction forms that the model can learn, making fine-tuning goals clearer and output formats more standardized. Combined with trusted inference samples, it constitutes a complete fine-tuning dataset, significantly improving the model's understanding and execution capabilities of actual operation and maintenance instructions. This allows the final industry model to directly respond to user queries and output standardized, actionable operation and maintenance conclusions.

[0046] Optionally, the domain adaptation fine-tuning of the Transformer architecture model in S4, based on the trusted inference sample training set and the grinding and crushing equipment health management instruction set, using the low-rank adaptive fine-tuning method, specifically includes: S43, introducing a low-rank decomposition matrix into the projection layer of the self-attention module of the Transformer architecture model and completing parameter initialization; S44, dividing the training set and validation set based on the trusted inference sample training set and the grinding and crushing equipment health management instruction set, and configuring training hyperparameters, including learning rate, batch size, training epochs, gradient pruning, and weight decay; S45, inputting mixed vector data, performing inference training of the Transformer architecture model based on the training set and training hyperparameters, and outputting inference results; S46, combining the equipment mechanism rules in the grinding and crushing equipment domain, calculating the task loss and mechanism consistency constraint loss based on the inference results, and updating the parameters of the low-rank decomposition matrix; S47, after iteratively training the inference to the optimal validation set index, fusing the weights of the low-rank decomposition matrix and the original weights of the Transformer architecture model to complete the initial construction of the grinding and crushing equipment health management industry model.

[0047] Specifically, the low-rank decomposition matrix is ​​the small-dimensional matrix used for parameter updates in the LoRA method; the training set and validation set are used for model training and performance evaluation; hyperparameters include learning rate, batch size, training epochs, gradient clipping, and weight decay; task loss is the error between the model output and the label; mechanism consistency constraint loss is a penalty term that forces the model output to conform to the device mechanism; weight fusion refers to merging the low-rank matrix with the original model weights.

[0048] In this optional embodiment, during the low-rank adaptive fine-tuning process, a low-rank decomposition matrix is ​​first introduced into the projection layer of the self-attention module of the Transformer model and initialized. Simultaneously, the backbone model parameters are frozen, and only a small number of low-rank parameters are trained, significantly reducing computational load and training cost while ensuring fine-tuning effectiveness. Then, the training and validation sets are divided based on reliable samples and instruction sets, and hyperparameters such as learning rate, batch size, and gradient clipping are configured to ensure stable training convergence. Next, mixed vector data is input into the model, and inference training is performed on the training set to output prediction results. Then, the task loss and mechanism consistency constraint loss are calculated simultaneously based on device mechanism rules, using the joint loss to guide gradient descent. Only the low-rank decomposition matrix parameters are updated, ensuring that the model output satisfies both task accuracy and physical mechanism, suppressing illusions during the training phase. Finally, training is iterated until the validation set metrics reach their optimal state. The trained low-rank weights are then merged with the original Transformer model weights to complete the initial construction of the industry model. This setup, through the combination of LoRA's efficient fine-tuning and mechanistic constraints, allows the model to quickly adapt to the grinding equipment field while maintaining strong robustness and high inference accuracy, providing a high-performance foundational model for the subsequent integration of RAG and thought chains.

[0049] Optionally, the integration of RAG retrieval enhancement generation and mind chain prompts in S4 specifically includes: S48, constructing a vector retrieval library for the grinding and crushing equipment field based on hybrid vector data, and incorporating a RAG retrieval enhancement generation mechanism into the initially constructed grinding and crushing equipment health management industry model, so as to call knowledge fragments from the vector retrieval library during the reasoning process through the RAG retrieval enhancement generation mechanism; S49, constructing a mind chain prompt template, and incorporating a mind chain prompt template into the initially constructed grinding and crushing equipment health management industry model, so as to provide step-by-step reasoning guidance during the reasoning process through the mind chain prompt template.

[0050] Specifically, the vector retrieval library is constructed from hybrid vector data to store domain knowledge vectors; RAG retrieval enhancement generation dynamically recalls relevant knowledge fragments during reasoning to supplement model knowledge; the thought chain prompt template is a fixed structure to guide the model's step-by-step reasoning; the step-by-step reasoning guidance is used to enable the model to deduce step by step according to the logical chain, improving interpretability.

[0051] In this optional embodiment, in integrating RAG and the mind chain mechanism, a dedicated vector retrieval library for the grinding and crushing equipment field is first constructed based on hybrid vector data. The RAG retrieval enhancement generation mechanism is then integrated into the initially constructed industry model, enabling the model to recall relevant equipment knowledge, fault cases, and maintenance experience from the retrieval library in real time based on input queries during the reasoning process. This dynamically injects externally reliable knowledge into the reasoning process, compensating for the limited knowledge and timeliness of model parameters. Subsequently, a mind chain prompt template adapted to health management tasks is designed, constructing a step-by-step guided structure according to the logical links of status analysis, fault location, cause deduction, and maintenance suggestions, and integrating it into the model's reasoning process. During actual reasoning, the model first obtains external knowledge support through RAG, and then gradually conducts reasoning according to the mind chain template, ensuring that the output results are not only accurate and reliable but also possess a clear logical chain and interpretable basis. This setup, through the dual enhancement of RAG and the mind chain, results in a grinding and crushing equipment health management industry model that is knowledge-rich, logically rigorous, process-traceable, and produces low-illusion output. It can be directly applied to actual engineering maintenance scenarios, providing intelligent decision support for the safe and stable operation of equipment.

[0052] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for constructing an industry model for health management of grinding and crushing equipment driven by both knowledge and data, characterized in that, include: S1. Obtain multi-source knowledge and operation and maintenance data in the field of grinding and crushing equipment, and sequentially perform knowledge extraction, data sequence reconstruction and unified vectorization processing on the multi-source knowledge and operation and maintenance data to obtain hybrid vector data; S2. Construct a neural engine based on the Transformer architecture model and the hybrid vector data, construct a symbolic reasoning engine based on the expert rule base and the forward inference engine, establish a data interaction interface between the neural engine and the symbolic reasoning engine, and form a dual-engine collaborative reasoning architecture. S3. Input the real-time equipment operation and maintenance data and user query requests into the dual-engine collaborative reasoning architecture, generate preliminary reasoning results through the neural engine, verify the preliminary reasoning results through the mechanism rules by the symbolic reasoning engine, and perform a preset number of iterations and corrections through the cooperation of the neural engine and the symbolic reasoning engine to obtain credible reasoning results and reasoning process, and form a credible reasoning sample training set. S4. Based on the entire health management process of the grinding and crushing equipment field, construct a health management instruction set for grinding and crushing equipment. Based on the trusted inference sample training set and the health management instruction set for grinding and crushing equipment, perform domain adaptation fine-tuning on the Transformer architecture model through the low-rank adaptive fine-tuning method, and integrate RAG retrieval enhancement generation and mind chain prompts to complete the construction of the industry model for health management of grinding and crushing equipment.

2. The method for constructing an industry model for health management of grinding and crushing equipment driven by both knowledge and data as described in claim 1, characterized in that, The multi-source knowledge and operation and maintenance data in S1 include structured knowledge, unstructured knowledge, and equipment time-series monitoring data. The structured knowledge includes equipment structural entities, fault mode entities, component relationships, and fault causal relationships. The unstructured knowledge includes equipment operation manuals, fault repair cases, industry standards and specifications, and operation and maintenance text logs. The equipment time-series monitoring data includes vibration data, temperature data, pressure data, and current data.

3. The method for constructing an industry model for health management of grinding and crushing equipment driven by both knowledge and data as described in claim 2, characterized in that, The knowledge extraction in S1 specifically includes: S11. Identify and extract knowledge entities and relationships between knowledge entities from the structured knowledge, assign a unique identifier to each knowledge entity and relationship between knowledge entities, and form structured knowledge text; S12. Perform word segmentation, stop word removal, and part-of-speech tagging on the unstructured knowledge text, extract professional and effective information, and convert it into an operable data format to form unstructured knowledge text, which is then combined with structured knowledge text to form text data.

4. The method for constructing an industry model for health management of grinding and crushing equipment driven by both knowledge and data as described in claim 3, characterized in that, The data sequence reconstruction and unified vectorization processing in S1 specifically includes: S13. Identify knowledge entities in the text data, and replace entity words in the text data with corresponding entity special lexical units; S14. Add special word elements representing the relationships between the knowledge entities to the end of the text data sequence to form a mixed input sequence that integrates the text information and the knowledge information; S15. The mixed input sequence is transformed into a fixed-dimensional vector through the Embedding layer of the Transformer architecture model; S16. Perform time-domain feature extraction and frequency feature extraction on the device time-series monitoring data, concatenate them to obtain a high-dimensional feature vector, perform normalization and dimension mapping on the high-dimensional feature vector in sequence, obtain a feature vector with the same dimension as the fixed-dimensional vector, and form the hybrid vector data with the fixed-dimensional vector.

5. The method for constructing an industry model for health management of grinding and crushing equipment driven by both knowledge and data as described in any one of claims 1 to 4, characterized in that, The neural engine constructed based on the Transformer architecture model and the hybrid vector data in S2 specifically includes: S21. Based on the Transformer architecture model, perform global feature modeling on the hybrid vector data, and learn the implicit correlation patterns in the hybrid vector data through a self-attention mechanism. The implicit correlation patterns include failure modes, health status, and lifespan trends. S22. Based on the implicit correlation rules, a task output layer is built on the backbone network of the Transformer architecture model to construct the neural engine.

6. The method for constructing an industry model for health management of grinding and crushing equipment driven by both knowledge and data as described in claim 5, characterized in that, The construction of the symbolic reasoning engine based on the expert rule base and the forward inference engine in S2 specifically includes: S23. Collect rules compiled by domain experts based on the equipment mechanism, fault diagnosis and operation and maintenance constraints knowledge in the field of grinding and crushing equipment, and store the rules in a structured and indexed manner to form the expert rule base. The expert rule base includes equipment mechanism rules, fault diagnosis rules and operation and maintenance constraint rules. S24. Build a forward inference engine and configure rule matching, logic verification and conflict resolution mechanisms to construct the symbolic inference engine.

7. The method for constructing an industry model for health management of grinding and crushing equipment driven by both knowledge and data as described in claim 6, characterized in that, S3 includes: S31. Input the real-time operation and maintenance data of the equipment and the user query request into the neural engine. Based on the implicit correlation law, generate the preliminary reasoning result through the neural engine. The preliminary reasoning structure includes key data features of fault hypothesis, health assessment result, decision suggestion and reasoning basis. S32. The preliminary reasoning result is transmitted to the symbolic reasoning engine through the neural engine, the symbolic reasoning engine calls the expert rule base for rule matching, and the preliminary reasoning result is logically verified based on the hybrid vector data. S33. If the logic verification passes, the preliminary reasoning result is marked as the reliable reasoning result by the symbolic reasoning engine, and the reliable reasoning result and the corresponding reasoning process are output. S34. If the logical verification fails, the symbolic reasoning engine locates the contradiction in the preliminary reasoning result, transmits the contradiction information to the neural engine based on the conflict resolution mechanism, and allows the neural engine to reanalyze the real-time equipment operation and maintenance data and adjust the preliminary reasoning result based on the contradiction information. The adjusted preliminary reasoning result is then transmitted back to the symbolic reasoning engine. When the logical verification of the symbolic reasoning engine passes or the preliminary reasoning result has been iterated and corrected three times, the last preliminary reasoning result is marked as the reliable reasoning result, and the reliable reasoning result and the corresponding reasoning process are output. S35. Based on the credible reasoning results and the reasoning process, form the credible reasoning sample training set.

8. The method for constructing an industry model for health management of grinding and crushing equipment driven by both knowledge and data as described in any one of claims 1 to 4, characterized in that, The construction of a health management instruction set for grinding equipment based on the entire health management process in the grinding equipment field in S4 specifically includes: S41. Based on the entire health management process in the field of grinding and crushing equipment, determine the scope of instructions around the core tasks of health management in the field of grinding and crushing equipment. The core tasks of health management include fault diagnosis, health assessment, remaining life prediction and operation and maintenance decision-making. S42. Based on the core equipment in the grinding and crushing equipment field, typical working conditions and operation and maintenance scenarios are sorted out, and a unified instruction template is adopted to generate a preset scale of domain instructions according to the scenario and task type. After being reviewed, deduplicated and verified by domain experts, the health management instruction set of the grinding and crushing equipment is formed. The unified instruction template is a four-segment structure template consisting of instruction description, input data example, output result example and evaluation standard.

9. The method for constructing an industry model for health management of grinding and crushing equipment driven by both knowledge and data as described in any one of claims 1 to 4, characterized in that, The step S4, which involves performing domain adaptation fine-tuning of the Transformer architecture model based on the trusted inference sample training set and the grinding equipment health management instruction set using a low-rank adaptive fine-tuning method, specifically includes: S43. Introduce a low-rank decomposition matrix into the projection layer of the self-attention module of the Transformer architecture model and complete parameter initialization; S44. Based on the trusted inference sample training set and the grinding equipment health management instruction set, divide the training set and the validation set, and configure the training hyperparameters, including the learning rate, batch size, training rounds, gradient pruning and weight decay. S45. Input the hybrid vector data, and perform inference training of the Transformer architecture model based on the training set and the training hyperparameters, and output the inference result; S46. Combining the equipment mechanism rules in the field of grinding equipment, calculate the task loss and mechanism consistency constraint loss based on the reasoning results, and update the parameters of the low-rank decomposition matrix; S47. After iterating the inference training until the validation set index is optimal, the weights of the low-rank decomposition matrix and the original weights of the Transformer architecture model are fused to complete the initial construction of the grinding equipment health management industry model.

10. The method for constructing an industry model for health management of grinding and crushing equipment driven by both knowledge and data as described in any one of claims 1 to 4, characterized in that, The integration of RAG retrieval enhancement generation and mind chain prompts in S4 specifically includes: S48. Construct a vector retrieval library for the grinding and crushing equipment field based on the hybrid vector data, and integrate a RAG retrieval enhancement generation mechanism into the initially constructed grinding and crushing equipment health management industry model, so as to call knowledge fragments of the vector retrieval library during the reasoning process through the RAG retrieval enhancement generation mechanism. S49. Construct a thought chain prompt template and integrate the thought chain prompt template into the initially constructed health management industry model for grinding and crushing equipment, so as to provide step-by-step reasoning guidance during the reasoning process through the thought chain prompt template.

Images