Industrial knowledge fiberization reconfiguration meta extraction method, device, equipment and medium

Abstract

The application discloses an industrial knowledge fiberization reconstruction meta-extraction method and device, equipment and medium, and relates to the technical field of artificial intelligence, which comprises the following steps: obtaining a unified knowledge vector by performing multi-modal semantic isomorphic mapping on multi-source heterogeneous industrial information, obtaining a fine-grained industrial knowledge roughcast through semantic chain construction and orthogonal decoupling guided by a one-hot mask and parameterized mapping, and then reconstructing the roughcast into candidate industrial meta-knowledge by a large model generation, and after verification, the candidate industrial meta-knowledge is issued to an industrial regulation terminal to regulate an industrial production process. The application realizes unified representation and unified granularity expression of multi-source heterogeneous industrial knowledge, ensures that the extraction result is formal, consistent in granularity and reliable in quality, and provides high-quality knowledge support for industrial production regulation and decision-making.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method, apparatus, equipment and medium for the fiber-based reconstruction of industrial knowledge. Background Technology

[0002] Existing industrial knowledge extraction methods mainly fall into three categories: one category dynamically corrects the results of industrial text knowledge extraction by integrating knowledge credibility assessment and expert collaborative review; another category builds a knowledge extraction system based on a large language model and improves extraction quality by dynamically calling external knowledge bases; and the third category improves the accuracy and efficiency of fault knowledge extraction by collaboratively calling multiple knowledge extraction tools and determining extraction strategies based on preset task templates and historical cases.

[0003] However, most of the above methods are geared towards single-source industrial texts or specific domains, making it difficult to adapt to multi-source heterogeneous industrial data such as text, tables, images, and production databases, and the knowledge extraction results are difficult to represent uniformly. At the same time, industrial knowledge is distributed in complex, multi-level production systems, with close cross-level connections and implicit semantic relationships, resulting in inconsistent extraction granularity. In addition, existing methods either rely on historical data review by experts, external knowledge bases, and prior templates, which have high labor costs and limited generalization ability, and generally lack a unified and quantifiable quality evaluation mechanism, making it difficult to conduct systematic autonomous evaluation of the accuracy, completeness, and consistency of the extraction results.

[0004] Therefore, how to uniformly represent and extract industrial production information from diverse sources and in different forms, and how to conduct autonomous quality inspection of the extraction results to obtain standardized and reliable industrial knowledge to support the control and decision-making of industrial production processes, has become an urgent problem to be solved. Summary of the Invention

[0005] The main objective of this application is to provide a method, apparatus, equipment, and medium for the fiber-based reconstruction of industrial knowledge, aiming to solve the technical problem of how to obtain standardized and reliable industrial knowledge to support the control and decision-making of industrial production processes.

[0006] To achieve the above objectives, this application proposes a meta-extraction method for the fiber-based reconstruction of industrial knowledge, comprising: Acquire multi-source heterogeneous industrial information, which includes industrial text information, industrial equipment runtime sequence data, and industrial image information; Multimodal semantic isomorphic mapping is performed on the multi-source heterogeneous industrial information to obtain a unified knowledge vector; The unified knowledge vector is subjected to semantic chain construction and one-hot mask-guided orthogonal decoupling to obtain multi-class decoupled industrial information semantic chains. The decoupled industrial information semantic chains are parameterized and mapped to obtain a fine-grained industrial knowledge rough embryo. By generatively reconstructing the fine-grained industrial knowledge embryo using a large model, candidate industrial meta-knowledge is obtained. The candidate industrial meta-knowledge is verified, and the candidate industrial meta-knowledge that passes the verification is taken as the target industrial meta-knowledge. The target industrial meta-knowledge is then sent to the industrial control terminal so that the industrial control terminal can control the industrial production process according to the target industrial meta-knowledge.

[0007] Furthermore, to achieve the above objectives, this application also proposes a meta-extraction device for the fiber-based reconstruction of industrial knowledge, the meta-extraction device comprising: The acquisition module is used to acquire multi-source heterogeneous industrial information, which includes industrial text information, industrial equipment runtime sequence data, and industrial image information. The mapping module is used to perform multimodal semantic isomorphic mapping on the multi-source heterogeneous industrial information to obtain a unified knowledge vector; The decoupling module is used to construct semantic chains and perform orthogonal decoupling guided by one-hot masks on the unified knowledge vector to obtain multiple types of decoupled industrial information semantic chains. The parameterization module is used to perform parameterized mapping on the multi-class decoupled industrial information semantic chain to obtain a fine-grained industrial knowledge rough embryo, wherein the fine-grained industrial knowledge rough embryo consists of knowledge objects, attribute parameters, association relationships and constraints. The reconstruction module is used to perform generative reconstruction of the fine-grained industrial knowledge embryo through a large model to obtain candidate industrial meta-knowledge, wherein the candidate industrial meta-knowledge includes a knowledge parameter source, a set of meta-knowledge instances evolved from the knowledge parameter source, and a source mapping relationship between the set of meta-knowledge instances and the knowledge parameter source. The verification module is used to verify the candidate industrial meta-knowledge, take the verified candidate industrial meta-knowledge as the target industrial meta-knowledge, and send the target industrial meta-knowledge to the industrial control terminal so that the industrial control terminal can control the industrial production process according to the target industrial meta-knowledge.

[0008] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the meta-extraction method for industrial knowledge fiber reconstruction as described above.

[0009] In addition, to achieve the above objectives, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the meta-extraction method for industrial knowledge fiber reconstruction as described above.

[0010] This application obtains a unified knowledge vector by performing multimodal semantic isomorphic mapping on multi-source heterogeneous industrial information. Through semantic chain construction and orthogonal decoupling guided by one-hot masks, and parameterized mapping, a fine-grained industrial knowledge rough embryo is obtained. This embryo is then reconstructed into candidate industrial meta-knowledge using a large model generative approach. After verification, it is distributed as target industrial meta-knowledge to the industrial control terminal to regulate industrial production processes. This application achieves a unified representation and uniform granularity of multi-source heterogeneous industrial knowledge, ensuring that the extracted results are standardized in form, consistent in granularity, and reliable in quality, providing high-quality knowledge support for industrial production control decisions. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a flowchart illustrating the first embodiment of the meta-extraction method for industrial knowledge fiber reconstruction in this application. Figure 2 This is a flowchart illustrating the second embodiment of the meta-extraction method for industrial knowledge fiber reconstruction in this application. Figure 3 This is a schematic diagram of the module structure of the meta-extraction device for the industrial knowledge fiber reconstruction of this application. Figure 4 This is a schematic diagram of the device structure of the hardware operating environment involved in the meta-extraction method for industrial knowledge fiber reconstruction in the embodiments of this application.

[0013] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0014] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0015] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0016] Existing industrial knowledge extraction methods are ill-suited to multi-source heterogeneous industrial data such as text, time-series data, and images. The extraction results are inconsistent in representation and granularity, and rely on expert experience, external knowledge bases, or prior templates. They lack an autonomous quality evaluation mechanism and cannot provide standardized and reliable knowledge support for the control of industrial production processes.

[0017] To address the aforementioned issues, this application proposes a method for extracting meta-information from the fiberized reconstruction of industrial knowledge: Multimodal semantic isomorphic mapping is performed on multi-source heterogeneous industrial information to obtain a unified knowledge vector. Semantic chain construction and orthogonal decoupling guided by one-hot masks yield multi-class decoupled industrial information semantic chains. Then, fine-grained industrial knowledge roughness is formed through parameterized mapping. Subsequently, candidate industrial meta-knowledge is obtained through generative reconstruction of a large model. After verification, the target industrial meta-knowledge is distributed to the industrial control terminal, enabling the terminal to regulate the industrial production process based on the target meta-knowledge. This achieves unified representation, unified granularity extraction, and autonomous quality verification of multi-source heterogeneous industrial knowledge.

[0018] Based on the above, this application also provides a meta-extraction method for industrial knowledge fiber reconstruction, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the method for extracting meta-extraction from the fiberized reconstruction of industrial knowledge in this application.

[0019] In this embodiment, the meta-extraction method for industrial knowledge fiber reconstruction includes steps S10 to S60: Step S10: Obtain multi-source heterogeneous industrial information.

[0020] It should be noted that multi-source heterogeneous industrial information refers to a collection of industrial information from different sources and in different data formats. Multi-source heterogeneous industrial information includes industrial text information, industrial equipment runtime sequence data, and industrial image information. Industrial text information is written data recording process rules, operating procedures, and fault phenomena. Industrial equipment runtime sequence data is a sequence of process parameters collected chronologically during the operation of production equipment. Industrial image information is image data describing the production process and equipment structure.

[0021] Furthermore, taking the alumina evaporation production process as an example, the industrial text information is first read from the production site's document system. This includes written materials such as process specifications, operation records, and fault reports. These materials are usually stored in document form and can be broken down into chapters and entries after being read, resulting in several text fragments.

[0022] Secondly, the operation sequence data of industrial equipment is collected through sensors deployed on the production site. The data includes process parameters such as temperature, pressure, liquid level, steam flow rate, mother liquor concentration, and heat transfer temperature difference of the evaporator. Each sensor collects data according to its preset sampling period. For example, temperature and pressure are collected once every 1 second, and mother liquor concentration is collected once every 60 seconds. Each data collection record is accompanied by a timestamp of the time of collection.

[0023] Next, acquire industrial image information. This includes evaporation flow charts, equipment structure diagrams, and instrument images captured during inspections. The evaporation flow charts and equipment structure diagrams can be retrieved from the production system's design drawing library, while the instrument images can be captured by a camera installed in front of the instrument panel at a preset sampling period.

[0024] Then, the three types of information collected are preprocessed. For the industrial equipment running sequence data, the minimum preset sampling period of each sensor is used as the time reference. Data collected at different times are time-aligned, and missing values ​​in the sequence are filled with values ​​from adjacent collection times. For industrial text information, blank pages and irrelevant headers and footers are removed. For industrial image information, blurry and incomplete images are removed.

[0025] Finally, the preprocessed text fragments, aligned process parameter sequences, and filtered images are categorized according to their respective production processes and organized into multi-source heterogeneous industrial information. This is for use in subsequent mapping processing.

[0026] Step S20: Perform multimodal semantic isomorphic mapping on multi-source heterogeneous industrial information to obtain a unified knowledge vector.

[0027] Specifically, industrial text information is segmented and semantically encoded to obtain text semantic descriptions; industrial equipment runtime sequence data is discretized and dimensionality reduced to obtain temporal semantic descriptions; and industrial image information is localized and attribute extracted to obtain image semantic descriptions. Semantic encoding can use BERT (Bidirectional Encoding Representation Model, a text encoding model based on an attention mechanism), and image feature extraction can use ResNet (Residual Network, a convolutional neural network). Subsequently, the three semantic descriptions are projected onto a shared semantic space and concatenated and aligned to obtain a unified knowledge vector.

[0028] It should be noted that multimodal semantic isomorphic mapping refers to the process of converting data from different modalities into vector representations of the same dimension within the same semantic space. The isomorphic semantic space refers to the vector space shared by data from different modalities after projection. A unified knowledge vector refers to the vector representation obtained by fusing semantic descriptions from text, time series, and images.

[0029] Furthermore, taking the alumina evaporation production process as an example, step S20 includes: First, performing text structure segmentation and functional semantic segmentation on the industrial text information to obtain multiple semantic blocks. Specifically, the processing of the industrial text information is carried out in two steps. First, text structure segmentation is performed, dividing the text into several paragraph segments according to the chapter titles and paragraph marks of the process specifications and operation records. Then, functional semantic segmentation is performed, identifying the functional semantics expressed by each paragraph segment, such as "heating operation", "feed control", and "salt discharge treatment". Continuous paragraphs expressing the same functional semantics are merged into one semantic block, and paragraphs expressing different functional semantics are assigned to different semantic blocks, ultimately obtaining multiple semantic blocks. Each semantic block contains a complete functional semantic such as evaporator operation, mother liquor treatment, or scale prevention and control.

[0030] Secondly, a semantic encoding network is used to encode features of multiple semantic blocks to obtain a text semantic description. Specifically, the semantic encoding network encodes features of each semantic block one by one. The semantic encoding network uses BERT. Each semantic block input yields a fixed-dimensional feature vector. The feature vectors of all semantic blocks together form the text semantic description, expressed by the following formula: in, Represents the semantic description of the text; Feature encoding functions representing text modalities; This represents the number of semantic blocks after segmentation. For the first Each feature vector contains specific functional semantics.

[0031] Next, semantic object recognition, temporal variation discretization, and high-frequency variation feature dimensionality reduction are performed on the industrial equipment runtime sequence data to obtain a temporal semantic description. Furthermore, regional semantic localization and extraction of regional attributes and numerical information are performed on the industrial image information to obtain an image semantic description. Specifically, in processing the industrial equipment runtime sequence data, semantic object recognition is first performed to identify the physical objects corresponding to parameters such as temperature, pressure, liquid level, steam flow rate, and mother liquor concentration from the process parameter sequence. Then, temporal variation discretization is performed, dividing the continuous values ​​of each parameter into several intervals according to the variation amplitude. For example, the heat transfer temperature difference is divided into three intervals: low temperature difference, medium temperature difference, and high temperature difference. The sequence is converted into a discrete state sequence, with the number of intervals ranging from 3 to 5, determined based on the historical fluctuation range of the parameters. Next, high-frequency variation feature dimensionality reduction is performed, using wavelet transform to separate the high-frequency fluctuation components in the sequence, and principal component analysis to compress the dimensionality of the high-frequency components, retaining the components with the highest variance contribution rate, thus obtaining a temporal semantic description. The specific formula is expressed as follows: in, Represents temporal semantic description; This represents the time-series mapping function, which outputs the denoised, standardized time-series feature matrix. This represents the discretization operation of time-series changes; This represents a dimensionality reduction operation for high-frequency features.

[0032] In processing industrial image information, ResNet is used to perform region-level semantic localization on evaporation flow diagrams and equipment structure diagrams, selecting areas such as evaporators, heating chambers, circulating pumps, and pipeline connections in the images. Then, the attribute and numerical information of each region is extracted, such as the equipment name, connection direction, and instrument readings labeled in the region. The visual semantic features of each region are combined to form an image semantic description, expressed by the following formula: in, Represents the semantic description of an image; Represents a visual feature encoding function; Indicates the number of semantic regions located; Indicates the first A visual semantic feature vector containing regional attributes and quantized values.

[0033] Next, using the transformation matrices corresponding to each modality, the text semantic description, temporal semantic description, and image semantic description are projected onto a shared isomorphic semantic space to obtain corresponding same-dimensional projection features. Specifically, using the transformation matrices corresponding to each modality, the text semantic description, temporal semantic description, and image semantic description are projected onto the shared isomorphic semantic space respectively. Each of the three modalities uses an independent transformation matrix, resulting in three same-dimensional projection features after projection. The parameters of the transformation matrices can be pre-trained through contrastive learning, using different modal data of the same device object as positive sample pairs during training.

[0034] Next, based on the semantic overlap of the corresponding same-dimensional projection features in terms of equipment objects, operating parameters, and process relationships, the cross-modal alignment weights between the same-dimensional projection features of any two modalities are calculated, and a cross-modal alignment weight matrix is ​​constructed based on these weights. Specifically, the cross-modal alignment weights between the same-dimensional projection features of any two modalities are calculated by statistically analyzing the semantic overlap of the projection features of the two modalities in terms of equipment objects, operating parameters, and process relationships. The cosine similarity between the projection features can be used as a measure of the degree of overlap. The higher the similarity, the more parts of the two modalities describe the same process content, and the larger the alignment weight value. The alignment weights between any two modalities are then row-normalized to form a cross-modal alignment weight matrix.

[0035] Finally, the corresponding same-dimensional projection features are concatenated to obtain concatenated features. Then, the concatenated features are weighted and fused according to the cross-modal alignment weight matrix to obtain a unified knowledge vector. Specifically, the three same-dimensional projection features are concatenated in the order of text, time series, and image to obtain concatenated features. Then, the modal-specific parts of the concatenated features are weighted and fused according to the cross-modal alignment weight matrix. The modal part with a larger alignment weight occupies a larger proportion in the fusion. The fusion result is used as the unified knowledge vector. The specific formula is as follows: in, This represents the unified knowledge vector of the final output. These are the transformation matrices for the projection of text modality, temporal structured data modality, and image modality onto the shared space, respectively. Indicates feature concatenation operation; This represents the weight matrix for cross-modal alignment.

[0036] Step S30 involves constructing semantic chains and performing orthogonal decoupling guided by one-hot masks on the unified knowledge vector to obtain multi-class decoupled industrial information semantic chains.

[0037] Specifically, a unified knowledge vector is input into a weakly supervised semantic perception network for serialization and topological mapping to obtain a continuous industrial information semantic chain. This continuous industrial information semantic chain includes information features, mechanism features, empirical features, and causal features from the unified knowledge vector. Then, industrial mechanism knowledge, industrial empirical knowledge, and industrial causal knowledge are fused into the network as prior knowledge. Under the guidance of the prior knowledge, the continuous industrial information semantic chain is orthogonally decoupled and projected to obtain multiple independent orthogonal industrial knowledge views. Each view is converted into a one-hot mask and then element-wise multiplied with the continuous industrial information semantic chain to obtain multiple decoupled industrial information semantic chains.

[0038] It should be noted that the industrial information semantic chain refers to a chain-like structure that organizes industrial information according to semantic association. Weakly supervised semantic perception networks are semantic extraction networks trained using a small amount of labeled data combined with prior knowledge. Industrial prior knowledge includes industrial mechanism knowledge, industrial experience knowledge, and industrial causal knowledge. An orthogonal industrial knowledge view refers to the knowledge feature representation obtained by projecting semantic features onto mutually independent subspaces. A one-hot mask is a mask vector where only the target category position has a value of 1, and all other positions have a value of 0. Element-wise multiplication refers to the operation of multiplying corresponding elements of two vectors of the same dimension.

[0039] Furthermore, taking the alumina evaporation production process as an example, step S30 includes: First, inputting the unified knowledge vector into a weakly supervised semantic perception network for sequential topological mapping to obtain a continuous industrial information semantic chain. Specifically, the unified knowledge vector is input into the weakly supervised semantic perception network, which integrates prior industrial knowledge. Includes industrial mechanism knowledge Industrial experience and knowledge and industrial causal knowledge The network first performs topological mapping on the various features in the vector according to the semantic sequence, arranging the features describing the surface content, physical principles, operational rules, and causal relationships of the same process in sequence, forming a continuous industrial information semantic chain. For example, during the alumina evaporation process, the chain sequentially includes features such as the description of the feed, the change in heat transfer temperature difference, the heat transfer mechanism, the experience of heating operation, and the causal description of scaling caused by the increase in temperature difference.

[0040] Secondly, the weakly supervised semantic perception network is jointly constrained by directed acyclic graph structure constraint loss, mechanism consistency constraint loss, and time constraint loss, and orthogonal decoupled projection is performed on the continuous industrial information semantic chain to obtain multiple candidate orthogonal industrial knowledge views. Specifically, the network is jointly constrained by directed acyclic graph structure constraint loss, mechanism consistency constraint loss, and time constraint loss. The three losses are weighted and summed according to a preset ratio, which can be 1:1:1. The three losses respectively ensure that the causal direction is not circular, the extracted features do not violate the heat transfer mechanism, and the cause precedes the effect in time. Under the joint constraints, the continuous industrial information semantic chain is orthogonally decoupled and projected to obtain multiple candidate orthogonal industrial knowledge views. The specific formula is expressed as follows: in, This represents the i-th candidate orthogonal industrial knowledge view; i represents the view number, for example, when i is 1, 2, or 3, it corresponds to the mechanism-based, experience-based, and causal-based knowledge views, respectively. This represents the orthogonal decoupling projection function; Represents a semantic chain of continuous industrial information; This represents the i-th type of industrial prior knowledge, which is the prior knowledge that guides the generation of the i-th view. For example, when i is 1, it is industrial mechanism knowledge; when i is 2, it is industrial experience knowledge; and when i is 3, it is industrial causal knowledge. This indicates that the constraint condition is that the total loss tends to be minimized.

[0041] Next, the feature correlation between any two candidate orthogonal industrial knowledge views is calculated. The correlation can be represented by the canonical correlation coefficient of the feature matrices of the two views, with a value ranging from 0 to 1. The larger the value, the more the features of the two views overlap.

[0042] Then, when the feature correlation exceeds a preset orthogonality threshold, the corresponding candidate orthogonal industrial knowledge views are re-orthogonally decoupled and projected until the feature correlation between any two candidate orthogonal industrial knowledge views does not exceed the preset orthogonality threshold, resulting in multiple independent orthogonal industrial knowledge views. Specifically, when the feature correlation between two candidate orthogonal industrial knowledge views exceeds the preset orthogonality threshold, it indicates that the two views are not completely decoupled. The two views are then re-orthogonally decoupled and projected. During the re-projection, the weight of the orthogonality penalty term between the two views is increased to promote their separation. After the re-projection, the feature correlation is calculated again until the feature correlation between any two candidate orthogonal industrial knowledge views does not exceed the preset orthogonality threshold. At this point, the views are considered as multiple independent orthogonal industrial knowledge views.

[0043] Finally, multiple independent orthogonal industrial knowledge views are transformed into multiple corresponding one-hot masks. Then, element-wise decoupling of the continuous industrial information semantic chain is performed based on these one-hot masks, resulting in multiple types of decoupled industrial information semantic chains. Specifically, each independent orthogonal industrial knowledge view is transformed into a corresponding one-hot mask. , This represents a one-hot encoding operation. The mask has the same length as the continuous industrial information semantic chain. The position corresponding to the category of interest in the view is set to 1, and the other positions are set to 0. Multiple one-hot masks are element-wise multiplied with the continuous industrial information semantic chain. Each mask filters out the corresponding category fragment from the semantic chain, resulting in a multi-class decoupled industrial information semantic chain. The specific formula is as follows: in, This indicates an element-wise multiplication operation; Indicates the first Decoupled semantic chains obtained under orthogonal industrial knowledge view constraints, such as mechanism-based semantic chains, experience-based semantic chains, and causal semantic chains.

[0044] Step S40: Perform parameterized mapping on the decoupled industrial information semantic chains of multiple types to obtain a fine-grained industrial knowledge rough embryo.

[0045] Specifically, for each type of decoupled industrial information semantic chain, the knowledge objects within it are identified, such as evaporators, heating chambers, mother liquor, and steam pipelines. The attribute parameters of the knowledge objects are extracted, such as heat transfer temperature difference, steam flow rate, mother liquor concentration, evaporation pressure, and liquid level. The process sequence relationship, causal relationship, and other correlation relationships between the knowledge objects, as well as the applicable constraints, are also extracted. The above four parts are assembled into parameterized knowledge units. All types of parameterized knowledge units are aggregated to form an industrial knowledge parameter base space, resulting in a fine-grained industrial knowledge rough embryo.

[0046] It should be noted that parametric mapping refers to the process of converting semantic knowledge into structured units composed of four elements: knowledge objects, attribute parameters, relationships, and constraints. A parametric knowledge unit refers to a knowledge representation unit composed of these four elements. The industrial knowledge parameter base space refers to the parametric knowledge set formed by the aggregation of all parametric knowledge units. Fine-grained industrial knowledge embryos refer to the industrial knowledge set obtained through parametric mapping that has not yet undergone reconstruction.

[0047] Further, taking the alumina evaporation production process as an example, step S40 includes: First, identifying knowledge objects for each of the multiple decoupled industrial information semantic chains to obtain knowledge objects. It should be understood that knowledge objects refer to entities that carry knowledge in industrial production; in this example, they include equipment, processes, materials, and quality indicators. Specifically, for each of the multiple decoupled industrial information semantic chains, knowledge objects are identified one by one. For example, equipment objects such as evaporators and heating chambers, and the evaporation process, are identified from the mechanism-based semantic chain; mother liquor materials and the quality indicator of scaling are identified from the experience-based semantic chain. The identification results are used as knowledge objects.

[0048] Secondly, extract the attribute parameters of knowledge objects and the relationships between them. It's important to understand that attribute parameters are quantifiable parameters describing the state or characteristics of a knowledge object; in this example, they include temperature, pressure, thickness, and velocity. Relationships refer to the connections between knowledge objects; in this example, they include process sequence relationships, causal relationships, and constraint relationships. Process sequence relationships refer to the order in which each process step is executed; causal relationships refer to the relationship where a change in one object causes a change in another; and constraint relationships refer to the relationship where the value of one object is restricted by another. Specifically, extract the attribute parameters for each knowledge object, such as the temperature and pressure attributes of the evaporator, the steam flow rate attributes of the steam pipeline, and the thickness attributes of the heating chamber pipe wall. Simultaneously, extract the relationships between knowledge objects, such as the process sequence relationship between the feeding and discharging processes, the causal relationship where an increase in heat transfer temperature difference leads to increased scaling, and the constraint relationship where the steam flow rate is limited by the evaporation pressure.

[0049] Next, based on the category of the decoupled industrial information semantic chain to which the knowledge object belongs, the corresponding constraints are determined. It's important to understand that constraints refer to the applicable rules that the knowledge object must follow in production. Specifically, based on the category of the decoupled industrial information semantic chain to which the knowledge object belongs, the constraints corresponding to that knowledge object are determined. For example, in the mechanism-based semantic chain, the evaporator is subject to heat transfer mechanism rules, and its constraint could be that the heat transfer temperature difference must not exceed a preset upper limit, which can be 20 to 40 degrees Celsius, determined according to the rated operating conditions in the evaporator design manual. In the experience-based semantic chain, the salt removal operation is subject to operational experience rules, and its constraint could be that salt removal is performed when the mother liquor concentration reaches a preset concentration value.

[0050] Next, the knowledge object, attribute parameters, relationships, and constraints are assembled into parameterized knowledge units. Specifically, each knowledge object, its attribute parameters, relationships, and constraints are assembled into a parameterized knowledge unit. This assembly process is completed through a knowledge parameterization mapping function, expressed by the following formula: in, Indicates the first Parametric industrial knowledge units can be further represented as , It represents industrial knowledge objects, such as equipment, processes, materials, or quality indicators; Represents object attributes and their parameter values, such as temperature, pressure, thickness, and speed; It indicates the relationships between objects, such as the sequence of processes, causal relationships, or constraint relationships; This indicates the constraints or applicable rules corresponding to the knowledge unit; This represents a knowledge parameterization mapping function, which transforms semantic knowledge chains into structured knowledge units. A unit fully describes a knowledge object: "what it is, how many parameters it has, who it is associated with, and what rules it follows."

[0051] Then, when there is a conflict between the attribute parameters and constraints in the parameterized knowledge unit, the attribute parameters are corrected according to the association relationship to obtain the verified parameterized knowledge unit. Specifically, after assembly, the parameterized knowledge unit is checked for conflicts by comparing the attribute parameters in the unit with the constraints. When the value of the attribute parameter conflicts with the constraints, for example, if the evaporation pressure attribute parameter of a certain unit exceeds the pressure upper limit specified by the constraints, it is corrected according to the association relationship. Specifically, the constraint relationship between evaporation pressure and steam flow rate is used to bring the steam flow rate attribute parameter associated with the pressure back to the range allowed by the constraints, and the value of evaporation pressure is corrected simultaneously. After correction, the comparison is repeated until there is no conflict, and the verified parameterized knowledge unit is obtained.

[0052] Finally, the verified parameterized knowledge units corresponding to various decoupled industrial information semantic chains are aggregated to obtain a fine-grained industrial knowledge rough embryo, and an industrial knowledge parameter basis space is constructed. Specifically, the verified parameterized knowledge units corresponding to various decoupled industrial information semantic chains are aggregated to form a set of fine-grained industrial knowledge rough embryos. Each knowledge item in the rough embryo consists of a knowledge object, attribute parameters, association relationships, and constraints. The aggregation result is expressed by the following formula: in, It represents a coarse-grained industrial knowledge prototype, that is, a knowledge set formed by the aggregation of all parameterized knowledge units; This represents the total number of categories in the decoupled industrial information semantic chain. This set together constitutes the multidimensional industrial knowledge parameter base space, which serves as the input for subsequent generative reconstruction.

[0053] Step S50: Generatively reconstruct the fine-grained industrial knowledge embryo using a large model to obtain candidate industrial meta-knowledge.

[0054] Specifically, the attribute parameters in the fine-grained industrial knowledge embryo are summarized to obtain multiple knowledge parameter sources, such as heat transfer temperature difference, steam flow rate, mother liquor concentration, evaporation pressure, liquid level, and scaling state. Using a large model as the driving engine and knowledge parameter sources as anchors, the dispersed parameters, states, mechanistic relationships, and control rules in the fine-grained industrial knowledge embryo are organized and evolved through reasoning. This generates a set of meta-knowledge instances evolved from the knowledge parameter sources, and establishes a source-tracing mapping relationship between the meta-knowledge instance set and the knowledge parameter sources to obtain candidate industrial meta-knowledge. The large model can adopt Qwen (a kind of open-source large language model), and the reasoning process is completed by designing prompt words that contain the knowledge parameter sources and embryo content.

[0055] It should be noted that the knowledge parameter source refers to the attribute parameters that serve as the origin of knowledge evolution. The meta-knowledge instance set refers to the set of all meta-knowledge instances evolved from the same knowledge parameter source. The source mapping relationship refers to the correspondence between meta-knowledge instances and their evolutionary origins. Candidate industrial meta-knowledge refers to a structured knowledge set that is yet to be verified, consisting of the knowledge parameter source, the meta-knowledge instance set, and the source mapping relationship.

[0056] Furthermore, taking the alumina evaporation production process as an example, step S50 includes: First, summarizing the attribute parameters in the fine-grained industrial knowledge rough to obtain the overall parameter source space. Specifically, summarizing the attribute parameters in the fine-grained industrial knowledge rough, counting the frequency of each attribute parameter in all parameterized knowledge units and the number of times it is associated, and refining the attribute parameters whose frequency exceeds a preset frequency threshold and are associated by multiple knowledge units as knowledge parameter sources. The preset frequency threshold can be one-tenth to one-fifth of the total number of units, determined according to the rough size. The summarized knowledge parameter sources include heat transfer temperature difference, steam flow rate, mother liquor concentration, evaporation pressure, liquid level, and scaling state. All knowledge parameter sources constitute the overall parameter source space.

[0057] Secondly, based on the local triviality constraint, the local neighborhood of each knowledge parameter source in the overall parameter source space is determined, and the parameterized knowledge unit within the local neighborhood is decomposed into a direct product combination of the knowledge parameter part and the knowledge form part. It is important to understand that the local triviality constraint means that any parameterized knowledge unit within the local neighborhood can be decomposed into an independent combination of the knowledge parameter part and the knowledge form part, and there are no coupling terms between the decomposed knowledge parameter part and the knowledge form part; the local neighborhood refers to the range of parameterized knowledge units surrounding a certain knowledge parameter source and associated with that knowledge parameter source; the knowledge parameter part refers to the part of the parameterized knowledge unit composed of the knowledge parameter source itself; the knowledge form part refers to the part of the parameterized knowledge unit composed of knowledge objects, relationships, and constraints, excluding the knowledge parameter source; the direct product combination refers to the structural form in which the knowledge parameter part and the knowledge form part are independently combined; and the coupling term refers to the inseparable cross-component that depends on both the parameter part and the form part after decomposition. Specifically, for each knowledge parameter source in the overall parameter source space, its local neighborhood is determined. This is done by screening out all parameterized knowledge units whose attribute parameters contain that knowledge parameter source from the fine-grained industrial knowledge roughness, and assigning them to the local neighborhood of that knowledge parameter source. For example, units whose attribute parameters include heat transfer temperature difference are all assigned to the local neighborhood of heat transfer temperature difference. Then, based on the local triviality constraint, each parameterized knowledge unit within the local neighborhood is decomposed into a direct product of the knowledge parameter part and the knowledge form part. The specific formula is expressed as follows: in, Represents a local neighborhood in the source space of industrial knowledge parameters; It represents a specific form of knowledge, namely, fiber space; Indicates the source mapping relationship; express The inverse mapping is used to look up all the meta-knowledge instances generated by the evolution of the knowledge parameter source. Represents the local neighborhood under the source mapping All corresponding knowledge units; It represents a local homeomorphism, and its function is to decompose the knowledge unit in the local neighborhood into two parts: parameters and morphology. This indicates that the two are structurally equivalent; This represents the projection onto the first factor; Indicates the composition of mappings; Indicates prior to By disassembling and reprojecting to retrieve the parameter source, the source-tracing mapping relationship is obtained. This formula constrains the large model to decompose complex industrial knowledge into independent combinations of parameters and forms within a local space. For example, in the local neighborhood of the heat transfer temperature difference, the knowledge parameter part is the heat transfer temperature difference itself, and the knowledge form part is the heating chamber that carries the heat transfer temperature difference, the causal relationship between the heat transfer temperature difference and the scaling state, and the control rules. The decomposition requires that the two parts are independent of each other and that there are no coupling terms. If there are still coupling terms after decomposition, the coupling terms are further decomposed into the parameter part or the form part until there are no coupling terms.

[0058] Next, the large model performs evolutionary reasoning on the knowledge objects, attribute parameters, relationships, and constraints in the product combination, generating a set of meta-knowledge instances evolved from the knowledge parameter source. Specifically, using the knowledge parameter source as the anchor point, the large model performs evolutionary reasoning on the knowledge objects, attribute parameters, relationships, and constraints in the product combination. The large model can use Qwen. During reasoning, the content of the knowledge parameter source and the product combination is organized as prompt words and input into the large model, guiding the large model to organize the scattered parameters, states, mechanistic relationships, and control rules under the same anchor point. For example, it generates meta-knowledge instances such as "increased temperature difference causes aggravated scaling, requiring an increase in the circulating pump speed" around the heat transfer temperature difference. All generated instances constitute the set of meta-knowledge instances evolved from this knowledge parameter source.

[0059] Then, based on the fibrillation constraint, each meta-knowledge instance in the set of meta-knowledge instances is verified by tracing its origin, establishing a tracing mapping relationship between the meta-knowledge instance and its unique corresponding knowledge parameter source. It's important to understand that the fibrillation constraint ensures that any meta-knowledge instance has a unique knowledge parameter source as its evolutionary origin. When a meta-knowledge instance can be uniquely traced back to its knowledge parameter source, a tracing mapping relationship is established between the meta-knowledge instance and the knowledge parameter source; otherwise, evolutionary reasoning is re-performed through the larger model until the unique knowledge parameter source is found. Specifically, each meta-knowledge instance in the set of meta-knowledge instances is verified by tracing its origin based on the fibrillation constraint. Fibrillation means that every knowledge manifestation has a unique origin, evolving from a specific knowledge parameter source, and a clear mapping relationship can be established between knowledge and its source, expressed by the formula: in, This represents a specific knowledge parameter source at the underlying level. Indicated by The set of all knowledge generated from the source, i.e., the fibers on the source of this knowledge parameter; Represents a specific instance of meta-knowledge; Representing instances of meta-knowledge The source of knowledge parameters was found through source tracing and mapping; Indicates and The equivalent category is formed by all knowledge from the same source. This formula requires that each meta-knowledge instance can be traced back to a unique knowledge parameter source along the source mapping, and that instances from the same source can be classified into the same category. Based on this, it is checked whether each instance in the meta-knowledge instance set is evolved only from the corresponding knowledge parameter source. When a meta-knowledge instance can be uniquely traced back to the knowledge parameter source, a source mapping relationship between the meta-knowledge instance and the knowledge parameter source is established. When a meta-knowledge instance cannot be traced back to the knowledge parameter source or can be traced back to multiple knowledge parameter sources at the same time, it indicates that the evolution path of the instance is unclear. The instance is then re-inferred through evolution by the large model until it is uniquely traced back to the corresponding knowledge parameter source.

[0060] Finally, based on the source mapping relationship, the knowledge parameter sources and the set of meta-knowledge instances are organized into a structured set to obtain candidate industrial meta-knowledge. Specifically, based on the source mapping relationship, the knowledge parameter sources and the corresponding set of meta-knowledge instances are organized into a structured set to obtain candidate industrial meta-knowledge. The form of the candidate industrial meta-knowledge is expressed by the formula: in, It represents the overall parameter source space, that is, the set of all knowledge parameter sources; This indicates the source of knowledge parameters involved in this reconstruction; Indicates the first One source of knowledge parameters; Indicates by the first A set of meta-knowledge instances generated by the evolution of knowledge parameter sources; Represents each instance in the collection The source mapping results are all The formula clarifies that candidate industrial meta-knowledge must contain three core elements: knowledge parameter source. Meta-knowledge instance set And the source mapping relationship between instances and parameter sources The final output of candidate industrial meta-knowledge is consistent in form and complete in evolution path.

[0061] Step S60: Verify the candidate industrial meta-knowledge, take the verified candidate industrial meta-knowledge as the target industrial meta-knowledge, and send the target industrial meta-knowledge to the industrial control terminal so that the industrial control terminal can control the industrial production process according to the target industrial meta-knowledge.

[0062] Specifically, candidate industrial meta-knowledge is verified from three dimensions: completeness of mechanism, information gain, and independence of knowledge. Candidate industrial meta-knowledge that passes all three verifications is used as target industrial meta-knowledge. If any one verification fails, the error signal is sent back and regenerated until the verification is passed. Then, the target industrial meta-knowledge is sent to the industrial control terminal. The industrial control terminal adjusts the operating parameters according to the mechanism relationship and control rules contained in the target industrial meta-knowledge. The preset thresholds involved in the verification can be set according to the statistical distribution of historical verification data, such as taking the preset quantile value of the historical data distribution.

[0063] It should be noted that the mechanism completeness test refers to the process of verifying whether meta-knowledge can be uniquely traced back to its knowledge parameter source. The information gain test refers to the process of verifying whether knowledge has reached an indivisible granularity. The knowledge independence test refers to the process of verifying whether the meta-knowledge corresponding to different knowledge parameter sources is independent of each other. Target industrial meta-knowledge refers to industrial meta-knowledge that has passed all verifications and can be used for industrial production control. The industrial control terminal can be a control device that adjusts the operating parameters of the production process.

[0064] Further, taking the alumina evaporation production process as an example, after step S60: First, the actual production process data after the industrial control terminal has regulated the industrial production process is collected, and the actual production process data is stored in the production process database. It is important to understand that the actual production process data refers to the operational data generated at the production site after the industrial control terminal performs regulation. In this example, it includes the collected values ​​of parameters such as temperature, pressure, liquid level, and steam flow rate of the evaporator after regulation. The production process database refers to a database that stores the actual production process data according to the regulation cycle. Specifically, after the industrial control terminal regulates parameters such as steam flow rate and evaporation pressure of the evaporation process according to the target industrial element knowledge, it collects the regulated actual production process data. Each collection record includes the regulation time and regulation cycle number, and is stored in the production process database according to the regulation cycle.

[0065] Secondly, based on the source mapping relationship in the target industrial meta-knowledge, the actual production process data corresponding to a single control is compared with the expected production state corresponding to the meta-knowledge instance in the target industrial meta-knowledge to obtain single-time deviation information. It is important to understand that the expected production state refers to the state that the production parameters should reach after the control rules are executed, as recorded in the meta-knowledge instance. Single-time deviation information refers to the difference between the actual production process data and the expected production state after a single control. The preset single-time deviation threshold is the boundary for determining whether a single control deviation needs to be addressed; it can be taken as 5% to 10% of the normal operating range of the corresponding parameter, determined based on the allowable fluctuation range of the process. Specifically, a single comparison is performed for each control. Based on the source mapping relationship in the target industrial meta-knowledge, the meta-knowledge instance corresponding to this control is found, and the expected production state recorded in the instance is read, such as the expected evaporation pressure falling back to the preset range. The actual collected value of the single control is subtracted from the expected production state, and the difference is used as the single-time deviation information.

[0066] Next, when the single deviation information exceeds a preset single deviation threshold, the corresponding knowledge parameter source is located based on the single deviation information. Then, based on the actual production process data from the single control corresponding to the knowledge parameter source, the parameterized knowledge unit corresponding to the knowledge parameter source in the fine-grained industrial knowledge rough is updated, resulting in the first updated fine-grained industrial knowledge rough. It is important to understand that the first updated fine-grained industrial knowledge rough refers to the fine-grained industrial knowledge rough obtained after updating the single control data. Specifically, when the single deviation information exceeds the preset single deviation threshold, it indicates that the current control has not achieved the expected results. Based on the single deviation information, the knowledge parameter source corresponding to the deviation is located along the source-tracing mapping relationship, for example, the evaporation pressure knowledge parameter source. Then, the attribute parameter values ​​in the parameterized knowledge unit corresponding to the evaporation pressure in the fine-grained industrial knowledge rough are replaced with the actual data collected during the current control, resulting in the first updated fine-grained industrial knowledge rough.

[0067] Next, according to the preset update cycle, the actual production process data accumulated in the production process database for multiple control cycles are compared in batches with the expected production state corresponding to the meta-knowledge instances in the target industrial meta-knowledge to obtain cumulative deviation information. Then, when the cumulative deviation information exceeds the preset cumulative deviation threshold, the corresponding knowledge parameter source is located based on the cumulative deviation information. Based on the actual production process data for multiple control cycles corresponding to the knowledge parameter source, the parameterized knowledge units corresponding to the knowledge parameter source in the fine-grained industrial knowledge rough embryo are updated in batches to obtain the second updated fine-grained industrial knowledge rough embryo. It is important to understand that the cumulative deviation information refers to the difference obtained by accumulating the deviations of multiple control cycles; the preset cumulative deviation threshold is the boundary for determining whether the cumulative deviation needs to be processed, which can be 3 to 5 times the preset single deviation threshold; the preset update cycle is the time interval for triggering batch comparison, which can be 1 to 7 days, determined based on the stability of the production conditions; and the second updated fine-grained industrial knowledge rough embryo is the fine-grained industrial knowledge rough embryo obtained after batch updates of data from multiple cycles. Specifically, after each preset update cycle, the actual production process data accumulated in the production process database for multiple control cycles are compared in batches. The collected values ​​of each cycle are compared with the expected production state of the corresponding meta-knowledge instance, and the absolute values ​​of each deviation are summed to obtain cumulative deviation information. When the cumulative deviation information is greater than the preset cumulative deviation threshold, it indicates that there is a persistent and trending deviation, and a single update is insufficient to correct it. Based on the cumulative deviation information, the corresponding knowledge parameter source is located, and the parameterized knowledge units corresponding to the knowledge parameter source in the fine-grained industrial knowledge rough are updated in batches using the actual data collected in multiple control cycles. For example, the attribute parameter values ​​and constraints of the unit corresponding to the knowledge parameter source are corrected together according to the statistical mean of the data from multiple cycles to obtain the second updated fine-grained industrial knowledge rough.

[0068] Then, the first or second updated fine-grained industrial knowledge embryo is generatively reconstructed and validated using a large model to obtain the updated target industrial meta-knowledge. Specifically, the first or second updated fine-grained industrial knowledge embryo is generatively reconstructed using a large model, with the reconstruction method consistent with the aforementioned generative reconstruction. The reconstruction result is then validated in three dimensions: mechanism completeness, information gain, and knowledge independence. After passing the validation, the updated target industrial meta-knowledge is obtained.

[0069] Finally, the updated target industrial metadata is sent to the industrial control terminal, enabling the terminal to regulate the industrial production process based on this updated metadata. Specifically, the updated target industrial metadata replaces the original metadata, and the terminal regulates subsequent industrial production processes accordingly. Completed regulations are not retrospectively adjusted.

[0070] This embodiment obtains a unified knowledge vector by performing multimodal semantic isomorphic mapping on multi-source heterogeneous industrial information. Through semantic chain construction and orthogonal decoupling guided by one-hot masks, and parameterized mapping, a fine-grained industrial knowledge rough embryo is obtained. This embryo is then reconstructed into candidate industrial meta-knowledge using a large model generative approach. After verification, it is distributed as the target industrial meta-knowledge to the industrial control terminal to regulate the industrial production process. This embodiment achieves a unified representation and uniform granularity of multi-source heterogeneous industrial knowledge, ensuring that the extracted results are standardized in form, consistent in granularity, and reliable in quality, providing high-quality knowledge support for industrial production control decisions.

[0071] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to that in Embodiment 1 above can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 2 The meta-extraction method S60 for the fiber-based reconstruction of industrial knowledge further includes steps S201 to S206: Step S201: Calculate the first similarity between the candidate industrial meta-knowledge and the knowledge parameter source corresponding to the candidate industrial meta-knowledge, and the second similarity between the candidate industrial meta-knowledge and any unrelated knowledge parameter source. Calculate the consistency score based on the first and second similarities. When the consistency score does not exceed the preset advantage threshold, the mechanism completeness verification of the candidate industrial meta-knowledge is passed.

[0072] It should be noted that the first similarity refers to the semantic similarity between the candidate industrial meta-knowledge and its corresponding knowledge parameter source, while the second similarity refers to the semantic similarity between the candidate industrial meta-knowledge and any unrelated knowledge parameter source. Unrelated knowledge parameter sources are those that do not have a source-tracing mapping relationship with the candidate industrial meta-knowledge. The consistency score is a score value representing the clarity of the source-tracing of the candidate industrial meta-knowledge, obtained by combining the first and second similarities. The lower the score, the clearer the source-tracing. The preset advantage threshold is the consistency score boundary for passing the completeness verification of the judgment mechanism, which can be taken from 0.01 to 0.1, and is determined based on the score distribution of samples confirmed to have correct source-tracing in historical verifications.

[0073] Specifically, the first and second similarities are calculated using a continuous mapping consistency test method based on contrastive learning. Meta-knowledge instances and knowledge parameter sources in the candidate industrial meta-knowledge are encoded into instance vectors and source vectors respectively by the same encoder. The cosine similarity between the instance vector and the source vector of the corresponding knowledge parameter source is calculated as the first similarity. Then, the cosine similarity between the instance vector and the source vector of any unrelated knowledge parameter source in the overall parameter source space is calculated, and the maximum value is taken as the second similarity.

[0074] Then, a consistency score is calculated based on the first similarity and the second similarity. The consistency score is calculated using the form of contrastive loss. The greater the weight of the first similarity relative to the similarity index of all source vectors, the smaller the consistency score. For example, when the similarity between the meta-knowledge instance and the corresponding knowledge parameter source is significantly higher than the similarity with unrelated knowledge parameter sources such as steam flow rate and mother liquor concentration other than heat transfer temperature difference, the consistency score is smaller.

[0075] Next, the consistency score is compared with a preset dominance threshold, as expressed by the following formula: in, This represents a meta-knowledge instance among the candidate industrial meta-knowledge; Representing instances of meta-knowledge Corresponding knowledge parameter source; Represents any knowledge parameter source in the overall parameter source space; This represents the mapping consistency evaluation function, used to calculate the similarity between meta-knowledge instances and knowledge parameter sources; The first similarity; The maximum value in the range is the second similarity; This represents the temperature coefficient hyperparameter, used to adjust the smoothness of the similarity distribution; Indicates exponentiation; Represents logarithmic operations; This represents the summation of all knowledge parameter sources in the overall parameter source space; Indicates the consistency score; This represents the set minimum threshold, also known as the preset dominance threshold. When the consistency score does not exceed the preset dominance threshold, it indicates that the correlation between the candidate industrial meta-knowledge and its corresponding knowledge parameter source is significantly higher than its correlation with other knowledge parameter sources, the tracing relationship is clear, and the completeness verification of the judgment mechanism is passed. When the consistency score exceeds the preset dominance threshold, it indicates that the candidate industrial meta-knowledge has experienced tracing confusion or cannot find the corresponding knowledge parameter source, the mechanism chain is broken, and the judgment test is failed.

[0076] Step S202: Split the knowledge parameter source corresponding to the candidate industrial meta-knowledge to obtain several sub-state sets, and calculate the information gain of the information entropy before and after the split. When the information gain is less than the preset gain threshold, the information gain of the candidate industrial meta-knowledge passes the verification.

[0077] It should be noted that the sub-state set refers to the subset obtained by further dividing the knowledge parameter source according to its value range or state category; information entropy is a measure of the uncertainty of the knowledge parameter source's values; and information gain is the difference in information entropy before and after splitting, used to measure the amount of new information brought about by further subdivision. The preset gain threshold is the information gain boundary for determining whether the knowledge granularity can be further subdivided. It can be a small value close to 0, such as 0.01 to 0.05, determined based on the splitting gain distribution of knowledge sources that have been confirmed to have reached the meta-granularity in historical data.

[0078] Specifically, the knowledge parameter source corresponding to the candidate industrial meta-knowledge is first split again. For example, the heat transfer temperature difference is further divided into several sub-state sets according to the value range. The information entropy of the knowledge parameter source before splitting and the total information entropy of each sub-state set after splitting are calculated according to probability weights. The information gain is obtained by subtracting the two. The specific formula is as follows: in, This represents the knowledge parameter source corresponding to the candidate industrial meta-knowledge; The first division represents the result of the split. A set of sub-states; The information entropy of the knowledge source before the split. The source of knowledge parameters is taken from the first... The probability of each state; For the first after the split The probability distribution of a set of substates; For the first Information entropy of a set of sub-states; Indicates information gain.

[0079] Next, when the information gain is less than the preset gain threshold, it means that no further subdivision of the knowledge parameter source can bring about effective new information, and the knowledge has reached the indivisible meta-granularity. The information gain test is passed. When the information gain is not less than the preset gain threshold, it means that the knowledge still contains decoupled redundant content. The test is failed.

[0080] Step S203: Calculate the mutual information between the sets of meta-knowledge instances corresponding to any two different knowledge parameter sources. When the mutual information is less than the preset independence threshold, the knowledge independence verification of the candidate industrial meta-knowledge passes.

[0081] It should be noted that mutual information is a measure of the degree of dependency between sets of meta-knowledge instances corresponding to two different knowledge parameter sources. The higher the value, the more information is shared between the two sets. The preset independence threshold is the mutual information boundary for determining whether meta-knowledge is mutually independent. It can be set to 0.1 to 0.3 and is determined based on the mutual information distribution of mutually independent meta-knowledge combinations confirmed in historical data.

[0082] Specifically, the co-occurrence of instances in the meta-knowledge instance sets corresponding to any two different knowledge parameter sources is statistically analyzed. The mutual information between the two sets is calculated based on the marginal probability distribution and joint probability distribution of the instances. The mutual information of each combination of knowledge parameter sources in the overall parameter source space is obtained by iterating through the combinations one by one. The specific calculation formula is as follows: in, and They represent the first The and the first A different source of knowledge parameters; and They represent respectively by and The reconstructed set of meta-knowledge instances; and They represent and Meta-knowledge instances in two candidate meta-knowledge sets; and Representing instances of meta-knowledge and The marginal probability distribution; This represents the joint probability distribution of the two, used to characterize the co-occurrence or dependency relationships between meta-knowledge instances under different parameter sources; This represents the mutual information between two sets of meta-knowledge instances; This represents a preset independence threshold. When the mutual information of each combination is less than the preset independence threshold, it indicates that there is no feature sharing or cross-derivative logic between the meta-knowledge corresponding to different knowledge parameter sources, and the knowledge independence check is passed. When the mutual information of any combination is not less than the preset independence threshold, it indicates that the features of the meta-knowledge have been mapped to other unrelated knowledge parameter sources, resulting in cross-source drift, and the check is failed.

[0083] Step S204: If any one of the following tests fails: completeness of mechanism test, information gain test, or knowledge independence test, an error signal is generated based on the failed test. The error signal is then fed back to the large model so that the large model can perform parameter iterative updates and regenerate candidate industrial meta-knowledge. The regenerated candidate industrial meta-knowledge is then verified again until all tests pass.

[0084] Specifically, if any of the three tests fails, an error signal is generated based on the type of the failed test. For example, if the mechanism completeness test fails, an error signal indicating source confusion is generated. The error signal is then fed back to the large model. The large model adjusts the generation parameters based on the error signal and regenerates the candidate industrial meta-knowledge. The regenerated candidate industrial meta-knowledge is then tested again in all three dimensions. If there are still failed items, the process of feeding back and regenerating is repeated until all tests pass.

[0085] Step S205: Select the candidate industrial meta-knowledge that passes all tests as the target industrial meta-knowledge. Step S206: Send the target industrial meta-knowledge to the industrial control terminal so that the industrial control terminal can control the industrial production process based on the target industrial meta-knowledge.

[0086] It should be noted that the target industrial knowledge is sent to the industrial control terminal, which reads the mechanism and control rules, such as "the increase in heat transfer temperature difference causes more scale to form, so the speed of the circulating pump needs to be increased", and adjusts the operating parameters such as steam flow and evaporation pressure in the alumina evaporation process accordingly.

[0087] This embodiment quantifies and verifies candidate industrial meta-knowledge from three dimensions: mechanistic completeness, information gain, and knowledge independence. It eliminates three types of defects: source confusion, granular redundancy, and cross-source drift. By using error signal feedback and large model parameter iteration to form a closed loop of generation and verification, unqualified knowledge is corrected through repeated iterations. Only meta-knowledge that meets all three dimensions can be issued as target industrial meta-knowledge to the industrial control terminal. Overall, it establishes an autonomous quality evaluation mechanism that does not require manual review, reduces the manual cost of knowledge verification, and improves the accuracy and reliability of extraction results, providing reliable knowledge support for industrial production process control.

[0088] Based on the first embodiment of this application, this application also provides a meta-extraction device for industrial knowledge fiber reconstruction, please refer to... Figure 3 The device includes: The acquisition module is used to acquire multi-source heterogeneous industrial information, including industrial text information, industrial equipment runtime sequence data, and industrial image information.

[0089] The mapping module is used to perform multimodal semantic isomorphic mapping on multi-source heterogeneous industrial information to obtain a unified knowledge vector.

[0090] The decoupling module is used to construct semantic chains and perform orthogonal decoupling guided by one-hot masks on a unified knowledge vector, resulting in multiple types of decoupled industrial information semantic chains.

[0091] The parameterization module is used to perform parameterized mapping on the semantic chains of decoupled industrial information from multiple categories to obtain a fine-grained industrial knowledge rough embryo, which consists of knowledge objects, attribute parameters, relationships, and constraints.

[0092] The reconstruction module is used to generatively reconstruct fine-grained industrial knowledge embryos through a large model to obtain candidate industrial meta-knowledge. The candidate industrial meta-knowledge includes knowledge parameter sources, a set of meta-knowledge instances evolved from the knowledge parameter sources, and a source mapping relationship between the set of meta-knowledge instances and the knowledge parameter sources.

[0093] The verification module is used to verify candidate industrial meta-knowledge, and to take the verified candidate industrial meta-knowledge as the target industrial meta-knowledge. The target industrial meta-knowledge is then sent to the industrial control terminal so that the industrial control terminal can control the industrial production process according to the target industrial meta-knowledge.

[0094] The industrial knowledge fiber reconstruction meta-extraction device provided in this application, employing the industrial knowledge fiber reconstruction meta-extraction method in the above embodiments, can solve the technical problem of how to obtain standardized and reliable industrial knowledge to support the control decisions of industrial production processes. Compared with the prior art, the beneficial effects of the industrial knowledge fiber reconstruction meta-extraction device provided in this application are the same as those of the industrial knowledge fiber reconstruction meta-extraction method provided in the above embodiments, and other technical features in the industrial knowledge fiber reconstruction meta-extraction device are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0095] This application provides a meta-extraction device for industrial knowledge fiber reconstruction. The meta-extraction device for industrial knowledge fiber reconstruction includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the meta-extraction method for industrial knowledge fiber reconstruction in the above embodiment 1.

[0096] The following is for reference. Figure 4This document illustrates a structural schematic of a meta-extraction device suitable for implementing the industrial knowledge fiber reconstruction embodiments of this application. The meta-extraction device for industrial knowledge fiber reconstruction in these embodiments may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 4 The illustrated meta-extraction device for industrial knowledge fiber reconstruction is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0097] like Figure 4 As shown, the meta-extraction device for industrial knowledge fiber reconstruction may include a processing unit (e.g., a central processing unit, a graphics processing unit, etc.) that can perform various appropriate actions and processes based on a program stored in read-only memory (ROM) or a program loaded from a storage device into random access memory (RAM). The RAM also stores various programs and data required for the operation of the meta-extraction device for industrial knowledge fiber reconstruction. The processing unit, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus. Typically, the following can be connected to the I / O interface: input devices including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices including, for example, magnetic tapes, hard disks, etc.; and communication devices. The communication device allows the meta-extraction device for industrial knowledge fiber reconstruction to communicate wirelessly or wiredly with other devices to exchange data. Although the figures show various industrial-grade fiber reconstruction meta-extraction devices, it should be understood that implementation or possession of all of them is not required. More or fewer may be implemented alternatively.

[0098] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from a storage device, or installed from a ROM. When the computer program is executed by a processing device, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0099] The industrial knowledge fiber reconstruction meta-extraction device provided in this application, employing the industrial knowledge fiber reconstruction meta-extraction method in the above embodiments, can solve the technical problem of how to obtain standardized and reliable industrial knowledge to support the control decisions of industrial production processes. Compared with the prior art, the beneficial effects of the industrial knowledge fiber reconstruction meta-extraction device provided in this application are the same as the beneficial effects of the industrial knowledge fiber reconstruction meta-extraction method provided in the above embodiments, and other technical features in this industrial knowledge fiber reconstruction meta-extraction device are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0100] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0101] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0102] This application provides a computer-readable medium having computer-readable program instructions (i.e., a computer program) stored thereon, the computer-readable program instructions being used to perform the meta-extraction method for industrial knowledge fiber reconstruction in the above embodiments.

[0103] The computer-readable medium provided in this application may be, for example, a USB flash drive, but is not limited to electrical, magnetic, optical, electromagnetic, infrared, or semiconductor devices, or any combination thereof. More specific examples of computer-readable media may include, but are not limited to: electrical connections with one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable medium may be any tangible medium containing or storing a program that can be executed by instructions, used by a device, or used in conjunction with it. The program code contained on the computer-readable medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0104] The aforementioned computer-readable medium may be included in the meta-extraction device for industrial knowledge fiber reconstruction; or it may exist independently and not be assembled into the meta-extraction device for industrial knowledge fiber reconstruction.

[0105] The aforementioned computer-readable medium carries one or more programs that, when executed by a meta-extraction device for industrial knowledge fiber reconstruction, enable the device to write computer program code for performing the operations of this application in one or more programming languages ​​or a combination thereof. These programming languages ​​include object-oriented programming languages—such as Java, Smalltalk, and C++—and conventional procedural programming languages—such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0106] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of methods and computer program products according to various embodiments of this application. In this regard, all blocks in the flowcharts or block diagrams may represent a module, segment, or portion of code containing one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that all blocks in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using dedicated hardware-based implementations that perform the specified functions or operations, or using a combination of dedicated hardware and computer instructions.

[0107] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0108] The readable medium provided in this application is a computer-readable medium that stores computer-readable program instructions (i.e., a computer program) for executing the above-described method for fiber-based reconstruction of industrial knowledge, thereby solving the technical problem of how to obtain standardized and reliable industrial knowledge to support the control decisions of industrial production processes. Compared with the prior art, the beneficial effects of the computer-readable medium provided in this application are the same as those of the method for fiber-based reconstruction of industrial knowledge provided in the above embodiments, and will not be repeated here.

[0109] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the meta-extraction method for industrial knowledge fiber reconstruction as described above.

[0110] The computer program product provided in this application can solve the technical problem of how to obtain standardized and reliable industrial knowledge to support the control and decision-making of industrial production processes. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as the beneficial effects of the meta-extraction method for industrial knowledge fiber reconstruction provided in the above embodiments, and will not be repeated here.

[0111] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A method for extracting meta-extraction from the fiber-based reconstruction of industrial knowledge, characterized in that, The method includes: Acquire multi-source heterogeneous industrial information, which includes industrial text information, industrial equipment runtime sequence data, and industrial image information; Multimodal semantic isomorphic mapping is performed on the multi-source heterogeneous industrial information to obtain a unified knowledge vector; The unified knowledge vector is subjected to semantic chain construction and one-hot mask-guided orthogonal decoupling to obtain multi-class decoupled industrial information semantic chains. The decoupled industrial information semantic chains are parameterized and mapped to obtain a fine-grained industrial knowledge rough embryo. By generatively reconstructing the fine-grained industrial knowledge embryo using a large model, candidate industrial meta-knowledge is obtained. The candidate industrial meta-knowledge is verified, and the candidate industrial meta-knowledge that passes the verification is taken as the target industrial meta-knowledge. The target industrial meta-knowledge is then sent to the industrial control terminal so that the industrial control terminal can control the industrial production process according to the target industrial meta-knowledge.

2. The method as described in claim 1, characterized in that, The step of performing multimodal semantic isomorphic mapping on the multi-source heterogeneous industrial information to obtain a unified knowledge vector includes: The industrial text information is subjected to text structure segmentation and functional semantic segmentation to obtain multiple semantic blocks; The semantic blocks are feature-encoded using a semantic coding network to obtain a text semantic description; Semantic object recognition, temporal variation discretization, and high-frequency variation feature dimensionality reduction are performed on the runtime sequence data of the industrial equipment to obtain a temporal semantic description. Regional semantic localization and regional attribute and numerical information extraction are performed on the industrial image information to obtain an image semantic description. By using the transformation matrices corresponding to each modality, the text semantic description, the temporal semantic description, and the image semantic description are projected onto a shared isomorphic semantic space to obtain the corresponding same-dimensional projection features; Based on the semantic overlap of the corresponding same-dimensional projection features in terms of equipment objects, operating parameters, and process relationships, calculate the cross-modal alignment weight between the same-dimensional projection features of any two modalities, and construct a cross-modal alignment weight matrix based on the cross-modal alignment weight. The corresponding same-dimensional projection features are spliced ​​together to obtain spliced ​​features; The spliced ​​features are weighted and fused according to the cross-modal alignment weight matrix to obtain a unified knowledge vector.

3. The method as described in claim 1, characterized in that, The steps of constructing semantic chains and performing orthogonal decoupling guided by one-hot masks on the unified knowledge vector to obtain multi-class decoupled industrial information semantic chains include: The unified knowledge vector is input into a weakly supervised semantic perception network for serialization and topological mapping to obtain a continuous industrial information semantic chain, wherein the continuous industrial information semantic chain includes information features, mechanism features, empirical features and causal features in the unified knowledge vector; The weakly supervised semantic perception network is jointly constrained by directed acyclic graph structure constraint loss, mechanism consistency constraint loss and time constraint loss, and the continuous industrial information semantic chain is orthogonally decoupled and projected to obtain multiple candidate orthogonal industrial knowledge views. The weakly supervised semantic perception network integrates industrial prior knowledge, which includes industrial mechanism knowledge, industrial experience knowledge and industrial causal knowledge. Calculate the feature correlation between any two candidate orthogonal industrial knowledge views; When the feature correlation exceeds the preset orthogonality threshold, the corresponding candidate orthogonal industrial knowledge view is re-orthogonally decoupled and projected until the feature correlation between any two candidate orthogonal industrial knowledge views does not exceed the preset orthogonality threshold, thus obtaining multiple independent orthogonal industrial knowledge views. Multiple independent orthogonal industrial knowledge views are transformed into multiple one-hot masks, and the continuous industrial information semantic chains are decoupled by element-wise product based on the multiple one-hot masks to obtain multiple types of decoupled industrial information semantic chains.

4. The method as described in claim 1, characterized in that, The step of parameterizing and mapping the decoupled industrial information semantic chains to obtain a fine-grained industrial knowledge rough embryo includes: Knowledge objects are identified for each of the decoupled industrial information semantic chains in the multi-type decoupled industrial information semantic chains to obtain knowledge objects, wherein the knowledge objects include equipment, processes, materials and quality indicators. Extract the attribute parameters of the knowledge object and the association between the knowledge object, wherein the attribute parameters include temperature, pressure, thickness and speed, and the association includes process sequence relationship, causal relationship and constraint relationship; Based on the category of the decoupled industrial information semantic chain to which the knowledge object belongs, determine the constraints corresponding to the knowledge object, wherein the constraints are the rules applicable to the knowledge object; The knowledge object, the attribute parameters, the association relationship, and the constraint conditions are assembled into a parameterized knowledge unit; When the attribute parameters in the parameterized knowledge unit conflict with the constraints, the attribute parameters are corrected according to the association relationship to obtain the verified parameterized knowledge unit. By aggregating various decoupled industrial information semantic chains and corresponding verified parameterized knowledge units, a fine-grained industrial knowledge embryo is obtained, and an industrial knowledge parameter base space is constructed. The fine-grained industrial knowledge embryo consists of knowledge objects, attribute parameters, associations, and constraints.

5. The method as described in claim 1, characterized in that, The step of generatively reconstructing the fine-grained industrial knowledge embryo using a large model to obtain candidate industrial meta-knowledge includes: The attribute parameters in the fine-grained industrial knowledge rough are summarized by parameter source induction to obtain an overall parameter source space, wherein the overall parameter source space includes multiple knowledge parameter sources; The local neighborhood of each knowledge parameter source in the overall parameter source space is determined based on the local triviality constraint, and the parameterized knowledge unit in the local neighborhood is decomposed into a direct product combination of the knowledge parameter part and the knowledge form part. The local triviality constraint means that any parameterized knowledge unit in the local neighborhood can be decomposed into an independent combination of the knowledge parameter part and the knowledge form part, and there are no coupling terms between the decomposed knowledge parameter part and the knowledge form part. By performing evolutionary reasoning on the knowledge objects, attribute parameters, relationships, and constraints in the direct product combination using a large model, a set of meta-knowledge instances evolved from the knowledge parameter source is generated; According to the fibrous constraint, each meta-knowledge instance in the meta-knowledge instance set is verified by tracing its origin, and a tracing mapping relationship between the meta-knowledge instance and the unique corresponding knowledge parameter source is established. The fibrous constraint means that any meta-knowledge instance has a unique knowledge parameter source as the evolutionary origin. When the meta-knowledge instance can be uniquely traced back to the knowledge parameter source, the tracing mapping relationship between the meta-knowledge instance and the knowledge parameter source is established. Otherwise, the evolutionary reasoning is re-performed through the large model until the knowledge parameter source is uniquely traced back. Based on the source mapping relationship, the knowledge parameter source and the meta-knowledge instance set are organized into a structured set to obtain candidate industrial meta-knowledge, wherein the candidate industrial meta-knowledge includes the knowledge parameter source, the meta-knowledge instance set evolved from the knowledge parameter source, and the source mapping relationship between the meta-knowledge instance set and the knowledge parameter source.

6. The method as described in claim 1, characterized in that, The steps of verifying the candidate industrial metadata, using the verified candidate industrial metadata as the target industrial metadata, and sending the target industrial metadata to the industrial control terminal so that the industrial control terminal can control the industrial production process according to the target industrial metadata include: Calculate the first similarity between the candidate industrial meta-knowledge and the knowledge parameter source corresponding to the candidate industrial meta-knowledge, and the second similarity between the candidate industrial meta-knowledge and any unrelated knowledge parameter source; A consistency score is calculated based on the first similarity and the second similarity. When the consistency score does not exceed a preset advantage threshold, the mechanism completeness verification of the candidate industrial meta-knowledge passes. The knowledge parameter source corresponding to the candidate industrial meta-knowledge is split to obtain several sub-state sets, and the information gain of the information entropy before and after the split is calculated. When the information gain is less than a preset gain threshold, the information gain verification of the candidate industrial meta-knowledge passes. Calculate the mutual information between the sets of meta-knowledge instances corresponding to any two different knowledge parameter sources; When the mutual information is less than a preset independence threshold, the knowledge independence verification of the candidate industrial meta-knowledge passes; If any one of the following tests fails: the mechanism completeness test, the information gain test, or the knowledge independence test, an error signal is generated based on the failed test, and the error signal is fed back to the large model so that the large model can perform parameter iterative updates and regenerate candidate industrial meta-knowledge. The regenerated candidate industrial meta-knowledge is then verified again until all tests pass. Candidate industrial meta-knowledge that passes all tests will be used as target industrial meta-knowledge. The target industrial metadata is sent to the industrial control terminal so that the industrial control terminal can control the industrial production process based on the target industrial metadata.

7. The method as described in claim 1, characterized in that, After the step of sending the target industrial metadata to the industrial control terminal so that the industrial control terminal can control the industrial production process according to the target industrial metadata, the following steps are included: The actual production process data after the industrial control terminal controls the industrial production process is collected, and the actual production process data is stored in the production process database; Based on the source mapping relationship in the target industrial meta-knowledge, the actual production process data corresponding to a single control is compared with the expected production state corresponding to the meta-knowledge instance in the target industrial meta-knowledge to obtain single deviation information. When the single deviation information is greater than the preset single deviation threshold, the corresponding knowledge parameter source is located according to the single deviation information, and the parameterized knowledge unit corresponding to the knowledge parameter source in the fine-grained industrial knowledge rough is updated according to the actual production process data of the single control corresponding to the knowledge parameter source, so as to obtain the first updated fine-grained industrial knowledge rough. According to a preset update cycle, the actual production process data accumulated in the production process database for multiple control cycles are compared in batches with the expected production state corresponding to the meta-knowledge instance in the target industrial meta-knowledge to obtain cumulative deviation information. When the cumulative deviation information is greater than the preset cumulative deviation threshold, the corresponding knowledge parameter source is located according to the cumulative deviation information, and the parameterized knowledge units corresponding to the knowledge parameter source in the fine-grained industrial knowledge rough are updated in batches according to the actual production process data of multiple control cycles corresponding to the knowledge parameter source, so as to obtain the second updated fine-grained industrial knowledge rough. The first or second updated fine-grained industrial knowledge rough model is regeneratively reconstructed and verified using a large model to obtain the updated target industrial meta-knowledge. The updated target industrial metadata is sent to the industrial control terminal so that the industrial control terminal can control the industrial production process based on the updated target industrial metadata.

8. A meta-extraction device for the fiber-based reconstruction of industrial knowledge, characterized in that, The apparatus is applied to the meta-extraction method for industrial knowledge fiber reconstruction as described in any one of claims 1 to 7, the apparatus comprising: The acquisition module is used to acquire multi-source heterogeneous industrial information, which includes industrial text information, industrial equipment runtime sequence data, and industrial image information. The mapping module is used to perform multimodal semantic isomorphic mapping on the multi-source heterogeneous industrial information to obtain a unified knowledge vector; The decoupling module is used to construct semantic chains and perform orthogonal decoupling guided by one-hot masks on the unified knowledge vector to obtain multiple types of decoupled industrial information semantic chains. The parameterization module is used to perform parameterized mapping on the multi-class decoupled industrial information semantic chain to obtain a fine-grained industrial knowledge rough embryo, wherein the fine-grained industrial knowledge rough embryo consists of knowledge objects, attribute parameters, association relationships and constraints. The reconstruction module is used to perform generative reconstruction of the fine-grained industrial knowledge embryo through a large model to obtain candidate industrial meta-knowledge, wherein the candidate industrial meta-knowledge includes a knowledge parameter source, a set of meta-knowledge instances evolved from the knowledge parameter source, and a source mapping relationship between the set of meta-knowledge instances and the knowledge parameter source. The verification module is used to verify the candidate industrial meta-knowledge, take the verified candidate industrial meta-knowledge as the target industrial meta-knowledge, and send the target industrial meta-knowledge to the industrial control terminal so that the industrial control terminal can control the industrial production process according to the target industrial meta-knowledge.

9. A meta-extraction device for the fiber-based reconstruction of industrial knowledge, characterized in that, The device includes: a memory, a processor, and a meta-extraction program for industrial knowledge fiber reconstruction stored on the memory and running on the processor, the meta-extraction program for industrial knowledge fiber reconstruction configured to implement the steps of the meta-extraction method for industrial knowledge fiber reconstruction as described in any one of claims 1-7.

10. A storage medium, characterized in that, The storage medium stores a meta-extraction program for industrial knowledge fiber reconstruction, which, when executed by a processor, implements the steps of the meta-extraction method for industrial knowledge fiber reconstruction as described in any one of claims 1-7.

Images